Brownian Motion

Based on previous notes by Oleg Zaboronski,

Stefan Grosskinsky, Roger Tribe, Jon Warren, . . .

Tommaso Rosati, MA4F7/ST403

Contents

1 Construction of Brownian motion 4

1.1 Stochastic processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Gaussian processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.1 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Construction of Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2 Simple properties of Brownian motion . . . . . . . . . . . . . . . . . 11
1.3.3 Kolmogorov’s continuity criterion . . . . . . . . . . . . . . . . . . . . 11

2 Universality and functional CLTs 14

2.1 Random walks and weak convergence . . . . . . . . . . . . . . . . . . . . . . 14
2.1.1 The Wiener measure on the space of paths. . . . . . . . . . . . . . . 15
2.1.2 Convergence in distribution (or weak convergence) . . . . . . . . . . 15
2.1.3 Compactness in the canonical space of paths . . . . . . . . . . . . . 17
2.2 Donsker’s invariance principle . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 BM as a strong Markov process 23

3.1 The Markov property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Blumenthal’s 0-1 law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 The strong Markov property . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 The reflection principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.5 Markov processes related to Brownian motion . . . . . . . . . . . . . . . . . 33

4 More path properties of Brownian motion 36

4.1 Modulus of continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2 Variation of Brownian paths . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3 Large-time behaviour and law of the iterated logarithm . . . . . . . . . . . 40

5 Brownian motion as a martingale 43

5.1 Basic properties of continuous martingales . . . . . . . . . . . . . . . . . . . 43
5.2 Wald’s lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.3 From martingales to harmonic functions . . . . . . . . . . . . . . . . . . . . 49

6 Brownian motion in higher dimensions 52

6.1 Harmonic functions and Brownian motion . . . . . . . . . . . . . . . . . . . 52
6.1.1 Connection to Brownian Motion . . . . . . . . . . . . . . . . . . . . 55
6.2 The Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.3 Recurrence and transience of BM . . . . . . . . . . . . . . . . . . . . . . . . 58
6.3.1 Occupation measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.4 The Feynman-Kac formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

2
CONTENTS 3

7 Hausdorff dimension of the range of Brownian motion 66

7.1 Lower bounds on the Hausdorff dimension . . . . . . . . . . . . . . . . . . . 68
7.2 Upper bounds on the Hausdorff dimension . . . . . . . . . . . . . . . . . . . 69
Chapter 1

Construction of Brownian motion

1.1 Stochastic processes

Let (Ω, F, P) be a probability space, and (S, S) a measurable space. A random variable
with values in S is a measurable map

X: Ω → S .

This course, instead of random variables, will focus on stochastic processes.

Definition 1.1.1. Let I ⊆ [0, ∞) be an index set. Then a stochastic process over I with
values in S is a collection {Xt }t∈I of random variables. Namely, for every t ∈ I, Xt is a
measurable map
Xt : Ω → S .
If I = N we refer to {Xn }n∈N as a discrete time stochastic process. If I = [0, ∞) we refer
to {Xt }t∈[0,∞) as a continuous time stochastic processes.

For example, let {ηi }i∈N be a collection of i.i.d. real valued random variables. Then
the random walk {Sn }n∈N
Xn
Sn = ηi
i=0

is a discrete time stochastic process. Of course, it can also be seen as a continuous time
stochastic processes, by defining St = S⌊t⌋ for any t > 0.

Kolmogorov’s extension theorem

A stochastic process can be seen as a random variable taking values in a larger state
space. Indeed, if {Xt }t∈I is a stochastic
Q process, then for every ω ∈ ΩI the (infinite) vector
I
(Xt (ω))t∈I is an element of S = t∈I S. On the product space S one has coordinate
maps πt : S I → S for any t ∈ I, defined by

πt ((sr )r∈I ) = st .

Then the space S I is a measurable space, when enriched with the product σ-field S I . This
is the smallest sigma-field that makes finitely many coordinate maps measurable:

S I = σ(πt : t ∈ I) .

4
1.1. STOCHASTIC PROCESSES 5

Let us denote with T the set of ordered times t = (t1 , · · · , tn ) with n arbitrary and
t1 < · · · < tn . Then we can equivalently characerise S I as the smallest σ-field that
contains all cylinder sets of the form
Ct,A = {(st )t∈I ∈ S I : (st1 , · · · , stn ) ∈ A} , (1.1)
for any choice of A ∈ S t and any t ∈ T .
Theorem 1.1.2 (Dynkin Systems). Let D be a collection of subsets of a nonempty set Ω
such that
1. Ω ∈ D.
2. A, B ∈ D and B ⊆ A imply A \ B ∈ D.
S∞
3. {An }∞
n=1 ⊆ D and A1 ⊆ A2 ⊆ · · · An ⊆ · · · imply n=1 Ai ∈ D.
If C is any other collection of subsets and C ⊆ D, then also σ(C) ⊆ D.
Exercise 1.1.3. Any stochastic process can be viewed as a measurable map X : Ω → S I
(use cylinder sets and the Dynkin System theorem).
Proof. Let D be the collection of subsets A ⊆ S I such that X −1 (A) is measurable. Then
D is a Dynkin system containing cylinder sets.

Now consider a stochastic process X and a set of times t = (t1 , · · · , tn ) such that
t1 < · · · < tn and tk ∈ I for all k ∈ {1, · · · , n}. To any such choice of t we can associate the
finite dimensional distributions (FDDs) of X a times t. Namely, we obtain the probability
measure µt on S n characterised by
µt (A1 × · · · × An ) = P(Xt1 ∈ A1 , · · · , Xtn ∈ An ) , ∀Ai ⊆ S .
The family {µt }t∈T of FDDs of a stochastic process is consistent, in the following sense.
Definition 1.1.4. A collection {µt }t∈T of probability measures such that µt is a measure
on S n (where t = (t1 , · · · , tn )) is said to be consistent if for any t = (t1 , · · · , tn ) ∈ T and
any k ∈ {1, · · · , n} it holds that
µt̂k (A1 × · · · × Ak−1 × Ak+1 × · · · × An ) = µt (A1 × · · · × Ak−1 × S × Ak+1 × · · · × An ) ,

where t̂k = (t1 , · · · , tk−1 , tk+1 , · · · tn ).

The law µ of a stochastic process X is uniquely characterised by collection {µt }t∈T of
all FDDs: this is the content of Kolmogorov’s extension theorem.
Theorem 1.1.5 (Kolmogorov extension). Let (S, S) be (Rd , B(Rd )) for some d ∈ N. Let
{µt }t∈T be a consistent family of measures. Then there exists a unique probability measure
on (S I , S I ) such that {µt }t∈T is the set of its FDDs.
Before we pass to the proof of this result, we must recall another extension theorem.
Theorem 1.1.6 (Caratheodory extension). Let (S, S) be a measurable space and C ⊆ S
an algebra of sets (closed under complements, finite unions and finite intersections) such
that σ(C) = S. Let µ beSa map µ : C → [0, 1] such that for any disjoint collection of sets
{An }n∈N ⊆ C satisfying n∈N An ∈ C, it holds that
!
[ X
µ An = µ(An ) .
n∈N n∈N

Then there exists a unique measure µ : S → [0, 1] such that µ = µ when restricted to C.
6 CHAPTER 1. CONSTRUCTION OF BROWNIAN MOTION

We omit the proof of Caratheodory’s theorem, and pass to a proof of Kolmogorov’s

theorem.

Proof of Theorem 1.1.5. For any t ∈ T and A ∈ S I , let Ct,A be the Cylinder set as in
(1.1). Then define
µ(Ct,A ) = µt (A) .
Since there can exists (t, A), (t′ , A′ ) such that Ct,A = Ct′ ,A′ , the definition of µ makes
sense only if for any such couple we have that µt (A) = µt (A′ ). It suffices to prove the
claim for product sets, namely assuming that A = A1 × · · · × An and A′ = A′1 × · · · × A′n ,
we must show that
µt (A1 × · · · × An ) = µt′ (A′1 × · · · A′n′ ) .
This follows from the consistency of the measures and the observation that if t ∈ t \ t′
(when viewed as sets of times), then necessarily At = Rd , since Ct,A = Ct′ ,A′ .
Exercise 1.1.7. Prove that the collection of cylinder sets {Ct,A }t,A is an algebra, namely
that it is closed under taking complements, finite unions and finite intersections.
In view of this exercise, the existence of the measure µ (that is the extension of its
definition to the entirety of S I ) follows from Theorem 1.1.6, if we show that the measure
is countably
S additive. Namely, let Cn , n ∈ N be a disjoint collection of sets in C such that
C = n∈N Cn ∈ C. We would like to prove that
X
µ(C) = µ(Cn ) .
n∈N

This is equivalent to proving that

[
lim µ(Qn ) = 0 , with Qn = Cm .
n→∞
m⩾n

Since µ(Qn ) is decreasing in n its limit exists. Assume therefore that

lim µ(Qn ) = ε > 0 ,

n→∞

and let us find a contradiction. Let us assume in addition that for each n ∈ N there exist
times (t1 , · · · , tn ) and a set An ∈ B(Rd×n ) such that

Qn = {(st1 , · · · , stn ) ∈ An } .

Exercise 1.1.8. How can we always reduce ourselves toTthis setting,T up to choosing a
slightly different sequence of decreasing sets Q′n satisfying n∈N Q′n = n∈N Qn ?
If the sets An were compact we could now complete the proof. Indeed, since µ(Qn ) > ε
for every n ∈ N the sets Qn cannot be empty and we would find a point (snt1 , · · · , sntn ) ∈ Qn .
Since the sets are also decreasing, the sequence (snt1 )n∈N ∈ A1 , which is compact, so there
exists a limit point rt1 ∈ A1 . Similarly for (snt1 , snt2 )n∈N , which must admit a limit point
T of
the form (rt1 , rt2 ) ∈ A2 . Iterating this procedure we can construct a point in p ∈ n Qn
satisfying
πtk (p) = rk ,
T
which contradicts the fact that n Qn = ∅.
In general the sets An are not compact, but this problem can be overcome as follows.
For each n ∈ N one can find a compact subset Kn ⊆ An such that µtn (An \Kn ) ⩽ 2εn (Borel
1.2. GAUSSIAN PROCESSES 7

sets in Rd are inner regular). Then define Q

e n = {(st , · · · , stn ) ∈ Kn } and Qn = T
1 m⩽n Qm .
e
In this way Qn is again a decreasing sequence of sets. In addition, we can lower bound

µ(Qn ) = µ(Qn ) − µ(Qn \ Qn )

!
\
= µ(Qn ) − µ Qn \ Q
en
m⩽n
n
X  
⩾ µ(Qn ) − µ Qm \ Q
em
m=1
n
X ε ε
⩾ε− ⩾ .
2n 2
m=1

Hence, we can apply the reasoning above to the sets Q

e n and the proof is complete.

1.2 Gaussian processes

Now let us focus on a particular class of stochastic processes, namely Gaussian processes.
These are uniquely characterised by their mean and covariance structure.

Definition 1.2.1. A real vector-valued

P random variable (X1 , . . . Xn ) is a multivariate
Gaussian if all linear combinations i ai Xi , for any ai ∈ R are univariate Gaussians.

Recall that the probability density function (PDF) fX of a (univariate) Gaussian X ∼

N (µ, σ 2 ) with mean µ ∈ R and variance σ 2 > 0 is given by
1 2 /(2σ 2 )
fX (x) = √ e−(x−µ) .
2πσ 2
Before we proceed, let us make some remarks.

• If X1 , . . . , Xn are independent Gaussians, then (X1 , . . . , Xn ) is multivariate Gaus-

sian.

• Definition 1.2.1 is stronger than the statement that each Xi is Gaussian. For exam-
ple, let X ∼ N (0, 1) and

+X , if |X| ≤ 1
Y = .
−X , if |X| > 1

Then also Y ∼ N (0, 1). But since |X + Y | ≤ 2 and X + Y is not constant, X + Y

is not Gaussian and so (X, Y ) is not bivariate Gaussian.

• Suppose X = (X1 , . . . , Xn ) is Gaussian. Then for any matrix A ∈ Rm×n the random
vector Y = AX ∈ Rm is Gaussian.

As we mentioned, the distribution of a multivariate Gaussian is uniquely determined by

its mean and its covariance structure. Let us recall this result in finite dimensions.

Proposition 1.2.2. The distribution of a Gaussian
X = (X1 , . . . , Xn ) is fully determined

by its mean vector µ = E[X] = E[X1 ], . . . , E[Xn ] and its covariance matrix

Σ = (σij )i,j=1,...,n where σij = E ((Xi − µi )(Xj − µj ))) = Cov[Xi , Xj ] .

8 CHAPTER 1. CONSTRUCTION OF BROWNIAN MOTION

Σ is always positive semi-definite. If it is also positive definite, X has a PDF which is

given by  
−1/2 1 T −1
fX (x) = det(2πΣ) exp − (x − µ) Σ (x − µ) . (1.2)
2
Finally, for every mean vector µ ∈ Rd and positive semi-definite matrix Σ ∈ Rd×d there
exists a unique multivariate Gaussian distribution N (µ, Σ).

The previous proposition allows us to define Gaussian stochastic processes based solely
on their mean and covariance structure.

Definition 1.2.3. A continuous-time stochastic process (Xt )t⩾0 with state space S = R
is a Gaussian process if all FDDs are
Gaussian, i.e. if for all n ∈ N and all t1 , . . . , tn ≥ 0
the random vector X(t1 ), . . . , X(tn ) has a multivariate Gaussian distribution.

The definition extends analogously to discrete time stochastic processes. From Propo-
sition 1.2.2 we deduce that the following holds.

Corollary 1.2.4. Let Xt be a Gaussian process. Then its law on (Rd )[0,∞) is uniquely
determined by its mean and covariance functions:
   
µ(t) := E X(t) and σ(s, t) := Cov X(s), X(t) , s, t ≥ 0 . (1.3)

Moreover, for any two functions t 7→ µt and (s, t) 7→ σs,t , such that for any t = (t1 , . . . , tn ) ∈
T the matrix (σti ,tj )i,j is positive semidefinite there exists a unique Gaussian process with
mean function µ and covariance function σ.

Example 1.2.5. Let X(t) : t ∈ N) be a family of i.i.d. (independent, identically dis-

tributed) standard Gaussians X(t) ∼ N (0, 1). Then X(t) : t ∈ N) is a Gaussian process
with mean µ(t) ≡ 0 and covariance

1 , if s = t
σ(s, t) = .
0 , if s ̸= t

1.2.1 Brownian motion

We are now ready to introduce (one-dimensional) Brownian motion as a continuous time
stochastic process on R[0,∞) .

Definition 1.2.6. A Brownian motion (Bt )t⩾0 started in B0 = 0 is a stochastic process

satisfying

1. (Stationary increments) For any t ⩾ s ⩾ 0 the increment Bt − Bs has the same

distribution as Bt−s .

2. (Independent increments) For any t ⩾ s ⩾ r ⩾ u the increments Bt −Bs and Br −Bu

are independent.

3. (Variance and mean) For any t ⩾ 0 we have Var(Bt ) = t, and E[Bt ] = 0. Or better
Bt ∼ N (0, t).

4. (Continuous sample paths) Almost surely, sample paths are continuous:

P ({ω : t 7→ Bt (ω) is continuous on [0, ∞)}) = 1 .

1.2. GAUSSIAN PROCESSES 9

Observe that it is yet unclear that Brownian motion exists, as we have no tools to
prove that the fourth property holds. But the tools we have introduced so far allow us to
construct a unique (in law, on R[0,∞) ) stochastic process satisfying properties (1) − (3).
For the previous definition to make sense (in view of Proposition 1.2.2 and Kol-
mogorov’s consistency criterion), we observe that for ordered times 0 ⩽ t1 < . . . < tn the
FDDs of a stochastic process (Xt )t⩾0 at times t = (t1 , . . . , tn ) are fully characterized by the
initial distribution of X0 and distribution of the increments ∆(ti−1 , ti ) := X(ti ) − X(ti−1 )
(with the convention that t0 = 0), since we can write:
 n
X 


X(t1 ), . . . , X(tn ) = X(0) + ∆(t0 , t1 ), ∆(t0 , t1 ) + ∆(t1 , t2 ), . . . , ∆(ti−1 , ti ) . (1.4)
i=1


Proposition 1.2.7. The following are equivalent for a stochastic process X(t) : t ≥ 0
on R with a fixed initial condition X(0) = x.


(a) X(t) : t ≥ 0 has stationary independent increments with
X(t) − X(s) ∼ N (0, t − s) for all t > s ≥ 0 .


(b) X(t) : t ≥ 0 is a Gaussian process with constant mean E[X(t)] = x and covariance
Cov[Xs , Xt ] = s ∧ t (:= min{s, t}) . (1.5)

Proof. Suppose (a) holds. Pick n ≥ 1, a1 , . . . , an ∈ R and t0 = 0 < t1 < · · · < tn

which are (wlog) ordered. Then, using (1.4), we find bk such that
n
X n
X

ak X(tk ) = b1 x + bk X(tk ) − X(tk−1 ) . (1.6)
k=1 k=1
The r.h.s. is a sum of independent
Gaussians, and hence Gaussian (see the remarks after
Definition 1.2.1), so X(t) : t ≥ 0 is a Gaussian process. With X(t) − X(s) ∼ N (0, t − s)
we obviously have   
E X(t) = x + E X(t) − X(0)] = x ,
and to compute covariances pick s < t and write
  (I) 


Cov X(s), X(t) = E X(s) − x X(t) − x


 
2  (II)
= E X(s) − x X(t) − X(s) + E X(s) − x = 0+s=s∧t .
Equality (I) is due to the invariance of Cov w. r. t. a constant shift, equality (II) is due
to the independence of increments. The final formula holds for any s, t ≥ 0 due to the
symmetry of covariance, Cov[Xt , Xs ] = Cov[Xs , Xt ]. Now assume (b) holds. Then for all
s < t, X(t) − X(s) is Gaussian with mean 0 and
    
Var X(t) − X(s) = Var X(t)] + Var X(s)] − 2Cov X(t), X(s)] = t + s − 2(s ∧ t) = t − s ,
so the process has stationary increments. To check independence of increments, take
0 ≤ t1 < · · · < tn and note that (1.6) implies that increments


X(t2 ) − X(t1 ), . . . , X(tn ) − X(tn−1 )
are Gaussian, so by Proposition 1.2.2 it is enough to show that increments are uncorrelated.
To do so, take u < v ≤ s < t and compute
 
Cov X(v) − X(u), X(t) − X(s) = v ∧ t − v ∧ s − u ∧ t + u ∧ s = v − v − u + u = 0 .
□
10 CHAPTER 1. CONSTRUCTION OF BROWNIAN MOTION

1.3 Construction of Brownian motion

While the FDDs fix the distributional properties of a process via its law on the product
space (S I , S I ), we might be interested in path properties of Brownian motion. For example
we might wonder whether the typical path t 7→ Bt (ω) is continuous (or Hölder continu-
ous, or even differentiable). For all these purposes, the infinite produce space R[0,∞) is
insufficient and we have to consider the law of B on the space of continuous functions.

Exercise 1.3.1. The set C = {t 7→ ω(t) : ω is continuous} is not an element of B(R[0,∞) ).

1.3.1 Brownian motion

The problem in constructing Brownian motion is not just the technicality observed in
Exercise 1.3.1. Indeed, even if we were to assign a reasonable “pretend-probability” to the
event C, it would have to be 0 rather than 1. Assume that P(C) ∈ [0, 1] exists and define
for each k ∈ N 
B(t) , if t ̸= τk
Bk (t) :=
B(t) + 1 , if t = τk


with independent random variables τk ∼ U [k − 1, k) . Then the processes Bk = Bk (t) :
t ≥ 0 are defined on (R[0,∞) , B(R[0,∞) )) and have the same FDDs as B since

 
P Bk (t) = B(t) = P[τk ̸= t] = 1 for all t ≥ 0 , (1.7)

due to the continuous distribution of the τk . Since via Theorem 1.1.5 the FDDs uniquely
determine the law of a process on (R[0,∞) , B(R[0,∞) )), we must have P[Ck ] = P[C] for all
k ≥ 1 where Ck = {Xk is continuous}. Since the Ck ⊆ Ω0 are mutually disjoint this
implies P[C] = 0.
Instead, we shall prove that it is always possible to modify Brownian motion in such
a way that the FDDs are not affected, but also so that its sample paths are continuous in
time.
 
Definition 1.3.2. Y is a modification of X, if P X(t) = Y (t) = 1 for all t ≥ 0.

Lemma 1.3.3. If Y is a modification of X, then the laws of Y and X on R[0,∞) coincide.

Proof. Since a finite union of null-sets is still a null-set, we have that for any t1 < t2 <
· · · < tn
P(Yt1 = Xt1 , Yt2 = Xt2 , . . . , Ytn = Xtn ) = 1 .
The result follows.

In the above construction the processes B and Bk , k ≥ 1 are all modifications of

Brownian motion, but at most one of them can have continuous paths. The construction
in Theorem 1.1.5 does not deliver a continuous modification of BM: this is a different
result.

Theorem 1.3.4. Brownian motion exists.

Before we prove this result, let us consider a few simple properties of Brownian motion
(assuming that the process exists).
1.3. CONSTRUCTION OF BROWNIAN MOTION 11

1.3.2 Simple properties of Brownian motion

Proposition 1.3.5.
Suppose B(t) : t ≥ 0 is a BM started in x = 0, then so is the
process X(t) : t ≥ 0 provided that

(a) X(t) := B(t + s) − B(s) for any fixed s ≥ 0 , (time translation)

(b) X(t) := −B(t) , (reflection)

1
(c) X(t) := √ B(ct) for any fixed c > 0 , (scaling)
c

0 , for t = 0
(d) X(t) := , (time inversion)
tB(1/t) , for t > 0

(e) X(t) := B(1) − B(1 − t) for t ∈ [0, 1] . (time reversal)

The proof is left as an exercise.


Corollary 1.3.6. For a BM B(t) : t ≥ 0 we have B(t)/t → 0 almost surely as t → ∞.


Proof. Using X(t) : t ≥ 0 from Point (d) of Proposition 1.3.5, which has continuous
paths, we see that

B(t)/t = X(1/t) → X(0) = 0 a.s. as t → ∞ .

Note that this implies the following asymptotic behaviour: For a ∈ R we have
 B(t) 
 +∞ , a > 0
B(t) + at = t +a → almost surely as t → ∞ .
t −∞ , a < 0

1.3.3 Kolmogorov’s continuity criterion

We will now prove Theorem 1.3.4, which will follow from the following result.
Theorem 1.3.7 (Kolmogorov’s continuity criterion). For any continuous-time stochastic
process t 7→ Xt , suppose that there exist constants α, β, C, T > 0 such that

E|Xt − Xs |α ⩽ C|t − s|1+β , ∀0 ⩽ s ⩽ t ⩽ T .

Then there exists a continuous modification X

e of X such that
!
|X
et − X
es |
P sup γ
<∞ =1,
0⩽s⩽t⩽T |t − s|

for any γ ∈ (0, β/α).

In particular, the conclusion of Theorem 1.3.7 provides more than mere continuity: it
e for any γ ∈ (0, 1/2).
even guarantees γ−Hölder continuity for the modification X,
Exercise 1.3.8. Prove Theorem 1.3.4. Hint: first show that for every n ∈ N there exists
a Cn such that
E|Bt − Bs |2n = Cn |t − s|n .
12 CHAPTER 1. CONSTRUCTION OF BROWNIAN MOTION

Proof of Theorem 1.3.7. To prove this result, let us fix T = 1 without loss of generality,
and introduce dyadic points on [0, 1]:
  ∞
k [
Dn = : k = 0, . . . , 2n , D= Dn .
2n
n=1

We write tnk ∈ Dn for the point tnk = 2kn . We will then study the cointinuity properties
of X on the set of points D and construct a modification X e of X that coincides with the
latter on D.
Step 1: Continuity on D. For two successive points tnk , tnk+1 ∈ Dn we can use Markov’s
inequality to bound

P(|Xtnk − Xtnk+1 | ⩾ 2−γn ) ⩽ C2γαn 2−(1+β)n ,

and similarly  
−γn
P max |Xtnk − Xtnk+1 | ⩾ 2 ⩽ C2−(β−γα)n .
k=0,...,2n

By our assumption on α, β, γ and the Borel–Cantelli lemma, we conclude that

∞ \ ∞
!
[
P Am = 1 ,
n=1 m=n

with  
−γn
An = max |X tn
k
−X tn
k+1
|<2 .
k=0,...,2n

In words, all but finitely many of the collection of events {An }n∈N take place, almost
surely. Let us therefore define the set
∞ \
[ ∞
Λ= Am .
n=1 m=n

Next, we claim that for every ω ∈ Λ there exists a C(ω) > 0 such that

|Xt − Xs |
sup γ
⩽ C(ω) .
s,t∈D,s̸=t |t − s|

Indeed, suppose that ω ∈ ∞

T
m=n0 Am . Then we claim that for any m ⩾ n ⩾ n0 and
t, s ∈ Dm such that |t − s| < 2−n we have
m
X
|Xt (ω) − Xs (ω)| ⩽ 2 2−γj . (1.8)
j=n+1

We will prove (1.8) by induction. Suppose that m = n + 1 and w.l.o.g. consider the case
s < t. Then necessarily s = tm m m −m+1 ,
k , t = tk+1 for some k = 0, . . . , 2 − 1, since t − s < 2
and the claim follows from the definition of Am .
Next, suppose (1.8) is true for m and let us prove it for m + 1. There exist k < l ∈
{0, . . . , 2m+1 } such that s = tm+1
l , t = tm+1
k . Then we consider points s ⩽ s < t ⩽ t such
that in s, t ∈ Dm , constructed as follows
( (
tm+1
l−1 if l is odd, tm+1 if k is odd,
s = m+1 t = k+1
tl if l is even, tm+1
k if k is even.
1.3. CONSTRUCTION OF BROWNIAN MOTION 13

We can bound
|Xt − Xs | ⩽ |Xt − Xt | + |Xt − Xs | + |Xs − Xs |
Xm
−(m+1)γ
⩽ 22 +2 2−kγ
k=n+1
m+1
X
⩽2 2−kγ ,
k=n+1

as required. Now we can conclude that for any ω ∈ Λ there exists an n0 (ω) such that for
any m > n ⩾ n0 we have
4 8
|Xt − Xs | ⩽ −γ
2−γn ⩽ |t − s|γ , ∀t, s ∈ Dm : 2−n−1 ⩽ |t − s| ⩽ 2−n .
1−2 1 − 2−γ
We can conclude that for every ω ∈ Λ

|Xt − Xs |
sup γ
<∞. (1.9)
s,t∈D |t − s|

Step 2: Construction of X.
e Now we define

et (ω) = 0 , ∀ω ̸∈ Λ ,
X et (ω) = Xt (ω) , ∀ω ∈ Λ , t ∈ D .
X

And finally, for arbitrary t ∈ [0, 1], consider a sequence {tk }k∈N ⊆ D such that tk → t
as k → ∞. By (1.9), we have that limk→∞ Xtk (ω) exists for every ω ∈ Λ and does not
depend on the particular choice of converging sequence {tk }. Therefore, set

X
et (ω) = lim Xt (ω) ,
k
∀ω ∈ Λ .
k→∞

The proof is complete if we now prove that X(ω)

e is a modification of X. Clearly, for t ∈ D
we have P(Xt = X et ) = 1. Instead, for t ∈ [0, 1] \ D we make use of the following exercise
(hint: an argument very similar to the very first bound in Step 1).
Exercise 1.3.9. For any t ∈ [0, 1] \ D and {tk }k∈N ⊆ D such that tk → t as k → ∞ we
have that
Xtk → Xt , in probability.

We therefore have Xtk → Xt in probability and Xtk → X

et almost surely, so that
P(Xt = X
et ) = 1 follows.
Chapter 2

Universality and functional CLTs

2.1 Random walks and weak convergence

In this chapter we obtain a different construction of Brownian motion, as the functional
scaling limit of random walks. Let {Xi }∈∈N be a sequence of i.i.d. random variables with

E[Xi ] = 0 , E[Xi2 ] = 1 . (2.1)

We consider the associated random walk (RW)

n
X
Sn = Xi , ∀n ∈ N , S0 = 0 .
i=1

Interpolating linearly between the integer points we can define the continuous time process



St = S⌊t⌋ + t − ⌊t⌋ S⌊t⌋+1 − S⌊t⌋ , ∀t ⩾ 0 , (2.2)

which is a random function on the space of continuous paths C[0, ∞). We rescale space
and time in analogy to the CLT and set
Snt
Stn = √ for all t ⩾ 0 .
n

This defines a sequence of paths S 1 , S 2 , . . . ∈ C[0, ∞). Our aim in this chapter will be to
prove that this sequence converges to a Brownian motion in law (see the section below for
the definition of the concepts used in this result).

Theorem 2.1.1 (Donsker’s invariance principle). Consider a sequence of i.i.d. random

variables {Xi }i∈N such that (2.1) is satisfied, and define S n as in (2.2). Then the sequence
{S n }n∈N converges in distribution to a Brownian motion t 7→ Bt , on the space C([0, ∞); R)
of continuous functions with the topology of uniform convergence on compact sets.

This result is also called a functional central limit theorem, since it is a functional
analogue of the CLT. The term ’invariance principle’ pertains to the fact that BM as the
scaling limit does not depend on the details of the original RW (apart from the normali-
√
sation of mean and variance), and is itself invariant (in law) under scaling: Bt = λBt/λ .
In other words, BM is a universal continuous limit for normalised RW’s.
Taking the increments Xk to be normalised is not a restriction.
pFor a general RW Sn the
result would hold for the normalised RW Ŝn := Sn − nE[Xk ] / nVar[Xk ] in full analogy
to the standard CLT.

14
2.1. RANDOM WALKS AND WEAK CONVERGENCE 15

2.1.1 The Wiener measure on the space of paths.

In order to obtain the desired convergence result, we first have to study the law of Brownian
motion on the space of continuous functions. Let us denote with C([0, ∞); R) the space
of continuous functions ω : [0, ∞) → R with the topology induced by the metric
X 1
d(ω1 , ω2 ) = max {|ω1 (t) − ω2 (t)| ∧ 1} ,
2n 0⩽t⩽n
n∈N

which amounts to uniform convergence on all compact sets.

Exercise 2.1.2. Endowed with the metric d, the space C([0, ∞)) is complete and separable.

We can also characterise the Borel σ−field on C([0, ∞)) as follows.

Lemma 2.1.3. Let us define the cylinder sets (analogously to (1.1))

Ct,A = {ω ∈ C([0, ∞)) : ωt1 ∈ A1 , · · · , ωtn ∈ An } , (2.3)

for any choice of (At )t∈t ∈ B(R)t and any t ∈ T . Then

B(C([0, ∞))) = σ(Ct,A : t ∈ T , A ∈ B(R)t ) .

Proof. If A is closed, than Ct,A is closed and therefore lies in B(C([0, ∞))). If follows
that σ(Ct,A : t, A) ⊆ B(C([0, ∞))). For the converse inequality (say that we look only
at times in the interval [0, 1]), it suffices to show that an open ball Bf (δ), for any δ > 0
lies in σ(Ct,A : t, A). Here we observe that since we are considering continuous paths it
suffices to condition on all rational times.

It follows that Brownian motion as constructed in Theorem 1.3.4 induces a law on

C([0, ∞)).

Definition 2.1.4 (Canonical space). We say that (C([0, ∞)), B, P), with P being the law
of Brownian motion, is the canonical proability space for Brownian motion. We also refer
to P as the Wiener measure on the space of paths C([0, ∞); R).

For simplicity, we will later consider convergence only on a finite time horizon, that is
in the space C([0, 1]).

2.1.2 Convergence in distribution (or weak convergence)

Before we attack the proof of Theorem 2.1.1 we recall the definition and a main result
concerning the convergence in distribution on general metric spaces.

Definition 2.1.5. Let E be a metric space and {Xn }n∈N , X E-valued random variables.
d
Then Xn converges to X in distribution (or in law, or weakly), written Xn −→ X, if
   
E g(Xn ) → E g(X) as n → ∞

for all bounded continuous functions g : E → R.

Although we will not make use of this, weak convergence is associated to a topology
on the space of probability measures, the so-called weak-∗ topology, and therefore there
is exists a natural concept of compactness. For our purposes we will only need sequential
compactness, therefore we define the following.
16 CHAPTER 2. UNIVERSALITY AND FUNCTIONAL CLTS

Definition 2.1.6. Let (S, ϱ) be a metric space and Π a collection of probability measures
on (S, B(S)).
1. We say that Π is compact, if for any sequence {Pn }n∈N ⊆ Π there exists a subse-
quence {nk }k∈N and a probability measure P ∈ Π such that

Pn → P weakly for n → ∞ .

2. We say that Π is tight if for every ε ∈ (0, 1) there exists a compact set Kε ⊆ S such
that
inf P(Kε ) ⩾ 1 − ε .
P∈Π

Then we can characterise compactness in terms of tightness.

Theorem 2.1.7 (Prokhorov’s theorem). Let (S, ϱ) be a separable and complete metric
space and Π a collection of probability measures as above. Then Π is compact if and only
if Π is tight.
Idea of proof (See Billingsley, “Convergence of probability measures”, Chapter 1.5). Let us
consider the one-dimensional case S = R, and let us only prove the implication tightness
⇒ compactness. Also let us assume that Π = {Pn }n∈N , up to choosing a sequence in the
class.
We consider the CDF Fn associated to a measure Pn :

Fn (x) = Pn ((−∞, x]) .

Then let {xi }i∈N be a countable dense set in R. We can find a subsequence {n1k }k such
that
lim Fn1 (x1 ) → F (x1 ) ,
k→∞ k

for some F (x1 ) ∈ [0, 1]. Then iteratively define subsequences

{nℓk }k ⊆ {nℓ−1 1
k }k ⊆ · · · ⊆ {nk }k ,

such that for every ℓ ∈ N

lim Fnℓ (xℓ ) → F (xℓ ) ,
k→∞ k

for some xℓ ∈ R. Choosing the diagonal sequence {nkk }k∈N we obtain that

lim Fnk (xℓ ) = F (xℓ ) , ∀ℓ ∈ N .

k→∞ k

By construction the limiting F is defined only on {xℓ }ℓ∈N and is increasing. It can therefore
be extended in a unique way to a cadlag function on the entire line R.
Our aim is to prove that the sequence converges weakly to a probability measure, so
we have to make sure that
lim F (x) = 1 .
x→∞
From tightness we know that for every ε > 0 there exists a xε ∈ R such that

F (x) ⩾ 1 − ε , ∀x ⩾ xε .

Tightness links compactness for a set of measures to compactness on the underlying

space (S, S). Therefore, if we are to study weak convergence for probability measures on
C([0, ∞)) we have to understand compact sets in this space.
2.1. RANDOM WALKS AND WEAK CONVERGENCE 17

2.1.3 Compactness in the canonical space of paths

Compactness in C([0, 1]) is studied through the modulus of continuity of a function. We
write
m(ω, δ) = sup |ωr − ωs |
0⩽s,r⩽1
|s−r|⩽δ

Exercise 2.1.8. The modulus of continuity m is decreasing in δ with limδ↓0 m(ω, δ) = 0

and continuous in ω ∈ C([0, 1]).
Compact sets in C([0, 1]) are then characterised by the Ascoli–Arzelá theorem.
Theorem 2.1.9 (Ascoli–Arzelá). A set A ⊆ C([0, 1]) has compact closure iff

sup |ω(0)| < ∞ , lim sup m(ω, δ) = 0 .

ω∈A δ↓0 ω∈A

Proof. We only prove that the condition is sufficient. Since the space C([0, 1]) is a metric
space it suffices to check sequential compactness. Consider a sequence {ω n }n∈N ⊆ A. We
must prove that there exists a limit point ω ∈ C([0, 1]) for the sequence. By assumption
we find a constant C > 0 such that
sup |ωs | ⩽ C . (2.4)
0⩽s⩽1

From (2.4) we have that for every rational t ∈ Q ∩ [0, 1] the sequence ωtn is bounded. Let
{ti }i∈N = Q ∩ [0, 1] be an enumeration of the rationals.
Then there exists a subsequence {n1k }k∈N such that
n1
ωt1k → ωt1 ∈ R ,
for some ωt1 . We can refine the subsequence to obtain convergence also at t2 : there exists
{n2k }k ⊆ {n1k }k such that also
n2
ωtik → ωti , for i = 1, 2 .
Iterating the procedure we find nested sequences

{nℓk }k ⊆ {nℓ−1 1
k }k ⊆ · · · ⊆ {nk }k ,

such that
nℓ
ωtik → ωti , ∀i = 1, . . . , ℓ .
Choosing a diagonal sequence we obtain that
nk
ωtik → ωti , for k → ∞ , ∀i ∈ N .

Now ω is uniformly continuous on Q ∩ [0, 1], since it satisfies by construction

|ωt − ωs | ⩽ sup m(ω, δ) , ∀t, s ∈ Q ∩ [0, 1] , |t − s| ⩽ δ .

ω∈A

and since by assumption

lim sup m(ω, δ) = 0 .
δ↓0 ω∈A

Therefore, ω can be extended to a continuous function in C([0, 1]), and it can be checked
k
from the calculations above that the convergence ω nk → ω holds in C([0, 1]).
18 CHAPTER 2. UNIVERSALITY AND FUNCTIONAL CLTS

We can use this result to obtain a characterisation of tightness for measures on the
space of paths.

Theorem 2.1.10. A sequence {Pn }n∈N of probability measures on (C, B(C)) is tight if
and only if
lim sup Pn (ω : |ω0 | > λ) = 0 ,
λ↑∞ n∈N

lim sup Pn (ω : m(ω, δ) > η) = 0 , ∀η > 0 .

δ↓0 n∈N

Proof. As above, we prove only that the condition is sufficient, and we show the result
only on a compact time interval, say on C([0, 1]). Let us write m(ω, δ) for m1 (ω, δ). For
any ε > 0 we have to find a compact set Kε ⊆ C([0, 1]) such that inf n∈N Pn (Kε ) ⩾ 1 − 2ε.
Now by assumption, for any k ∈ N we can find a δk > 0 such that

sup Pn (m(ω, δk ) > k −1 ) ⩽ ε2−k

n∈N

c
Let us denote with Ak = {m(ω, δk ) > k −1 }, so that the set Aε =
S
k∈N Ak satisfies

P(A) ⩾ 1 − ε ,

and in addition, we have that by definition

sup m(ω, δk ) ⩽ k −1 ,
ω∈A

so that
lim sup m(ω, δ) = 0 .
δ↓0 ω∈A

In addition choose λε > 0 such that the set B = {|ω0 | > λ} satisfies

sup Pn (B) ⩽ ε .
n∈N

Then we obtain that by Theorem 2.1.9 the set A ∩ B c is compact and

inf Pn (A ∩ B c ) ⩾ P(A) − P(B) ⩾ 1 − 2ε ,

n∈N

which is sufficient to obtain our result.

Building on the previous result, we can for example obtain a criterion for tightness
that is linked to the Kolmogorov–Chentsov continuity criterion of Theorem 1.3.7.

Proposition 2.1.11. Suppose that a sequence of probability measures {Pn }n∈N on C([0, 1])
satisfies for some α, β, γ, C > 0

sup En |ω0 |ζ < ∞ ,

n∈N
sup En |ωt − ωs |α ⩽ C|t − s|1+β , ∀t, s ∈ [0, 1] .
n∈N

Then the sequence of probability measures {Pn }n∈N is tight.

The proof of this result is an exercise.

2.2. DONSKER’S INVARIANCE PRINCIPLE 19

2.2 Donsker’s invariance principle

The first step towards the proof of Theorem 2.1.1 is to show that the sequence S n is tight
in C([0, ∞)). Here we apply Proposition 2.1.11.

Proposition 2.2.1. Indicate with Pn the law associated to S n as in the statement of

Theorem 2.1.1. Then the sequence of probability measures {Pn }n∈N is tight.

Proof. Our aim is to apply Theorem 2.1.10. We therefore have to prove that

lim sup P (m(S n , δ) > η) = 0 , ∀η > 0 .

δ↓0 n

First, we observe that it suffices to prove that

lim lim sup P (m(S n , δ) > η) = 0 , ∀η > 0 . (2.5)

δ↓0 n→∞

Indeed (2.5) holds true, then for any ε > 0 there exists n(ε, δ) ∈ N, δε ∈ (0, 1) such that
 
sup P m(S k , δε ) > η ⩽ ε .
k⩾nε (ε,δε )

Since the left hand-side is decreasing in δ, we actually can conclude that

 
sup P m(S k , δ) > η ⩽ ε , ∀δ ∈ (0, δε ) .
k⩾nε (ε,δε )

Therefore
 
lim sup P (m(S n , δ) > η) ⩽ ε + lim sup P m(S k , δ) > η ⩽ ε ,
δ↓0 n δ↓0 k⩽n(ε,δε )

so that the claim follows. Therefore, it remains only to check (2.5). This will now follow
from Lemma 2.2.3 below. Indeed, recall that
tn
n 1 X
m(S , δ) = sup |Stn − Ssn | ⩽√ sup | Xj | ,
s,t∈[0,1], ,|t−s|⩽δ n 0⩽s⩽t⩽1, ,|t−s|⩽δ
j=sn

where
tn = ⌊nt⌋ + 1 , sn = ⌊sn⌋ .
The last quantity is furthermore upper bounded by
k+ℓ
1 X
√ sup | Xj | .
n k⩽n+1,ℓ⩽⌊nδ⌋+1
j=k

We are therefore exactly in the setting of Lemma 2.2.3.

In addition, we can control the convergence of FDDs.

Proposition 2.2.2. For any t = (t1 , . . . , tm ) ∈ T we have that the following convergence
holds true weakly:
(Stn1 , . . . , Stnm ) → (Bt1 , . . . , Btm ) .
20 CHAPTER 2. UNIVERSALITY AND FUNCTIONAL CLTS

Proof. Let us consider the vector of increments

(Stn1 , ∆nt2 ,t1 , . . . , ∆ntm ,tm−1 ) ,

where ∆nti ,ti−1 = Stni − Stni−1 . By definition we have that

⌊nti ⌋
1 X
∆nti ,ti−1 =√ Xj
n
j=⌊nti−1 ⌋+1
1 1
+ √ (nti − ⌊nti ⌋)X⌊nti ⌋+1 − √ (nti−1 − ⌊nti−1 ⌋)X⌊nti−1 ⌋+1 .
n n

Now observe that both rest terms

1 1
√ (nti − ⌊nti ⌋)X⌊nti ⌋+1 , √ (nti−1 − ⌊nti−1 ⌋)X⌊nti−1 ⌋+1
n n

converge to zero in probability. Moreover, if a sequence Xn → X weakly and Xn − X

en → 0
in probability, then also Xn → X weakly. Since from the CLT we know that
e

⌊nti ⌋
1 X
√ Xj → N (0, ti − ti−1 )
n
j=⌊nti−1 ⌋+1

we see that ∆nti ,ti−1 → ∆ti ,ti−1 = Bti − Bti−1 weakly. Since

⌊nti ⌋ ⌊ntk ⌋
1 X 1 X
√ Xj is independent of √ Xj
n n
j=⌊nti−1 ⌋+1 j=⌊ntk−1 ⌋+1

for i ̸= k, we also obtain the convergence of the vectors:

(Stni , ∆nt2 ,t1 , . . . , ∆ntn ,tn−1 ) → (Bt1 , ∆t2 ,t1 , . . . , ∆tn ,tn−1 ) ,

and the proof is complete.

Finally, we are able to prove Theorem 2.1.1.

Proof of Theorem 2.1.1. Our aim is to prove that Pn → P weakly (where P is the law of
Brownian motion on C[0, 1]). Suppose that this is not the case, then, since by Proposi-
tion 2.2.1 the sequence is tight, and hence by Theorem 2.1.7 precompact, there must exist
a law Q ̸= P and a subsequence {nk }k∈N such that Pnk → Q as k → ∞ (weakly). But the
law of a stochastic process is uniquely characterised by its finite-dimensional distributions.
By Proposition 2.2.2, Q must have the FDDs of a Brownian motion, therefore we have
found a contradiction.

Lemma 2.2.3. Under the assumptions of Theorem 2.1.1 we have that for every ε > 0
 
√
lim lim sup P  max |Sj+k − Sk | > ε n = 0 .
δ↓0 n→∞ 0⩽j⩽⌊δn⌋+1
0⩽k⩽n+1
2.2. DONSKER’S INVARIANCE PRINCIPLE 21

Proof. We split up the integer times up to n in bits of length ⌊nδ⌋ + 1. In particular, there
exists an m(δ, n) such that

(⌊nδ⌋ + 1)(m − 1) ⩽ n + 1 < (⌊nδ⌋ + 1)m ,

and there exists a C > 0 such that

C
m(δ, n) ⩽ , ∀n ∈ N , δ ∈ (0, 1) .
δ
In particular, for any k ∈ {1, . . . , n + 1}, there exists a p(k) ∈ {1, . . . , m} such that

(⌊nδ⌋ + 1)(p − 1) ⩽ k < (⌊nδ⌋ + 1)p .

And in addition, since j ⩽ ⌊nδ + 1⌋, either of the following two cases hold true

(⌊nδ⌋ + 1)(p − 1) ⩽ k + j < (⌊nδ⌋ + 1)p ,

(⌊nδ⌋ + 1)p ⩽ k + j < (⌊nδ⌋ + 1)(p + 1) .

Assume that j, k assume the maximum in the event whose probability we are attempting
to bound. In the first case we must have that either of the following two hold:
1 √ 1 √
|Sj+k − S(⌊nδ⌋+1)(p−1) | > ε n , or |Sk − S(⌊nδ⌋+1)(p−1) | > ε n .
2 2
In the second case, one of the following three has to hold:
1 √
|Sj+k − S(⌊nδ⌋+1)p | > ε n , or
3
1 √
|Sk − S(⌊nδ⌋+1)(p−1) | > ε n , or
3
1 √
|S(⌊nδ⌋+1)p − S(⌊nδ⌋+1)(p−1) | > ε n .
3
In particular, we can conclude that
 
m(n,δ)+1 
√  √
 [ 
max |S − Sk | > ε n ⊆ max |Sj+(⌊nδ⌋+1)(p−1) − S(⌊nδ⌋+1)(p−1) | > ε n .
0⩽j⩽⌊δn⌋+1 j+k  0⩽j⩽⌊δn⌋+1
0⩽k⩽n+1 p=1

Therefore, we can bound

 
√ √
 
P max |Sj+k − Sk | > ε n ⩽ mP max |Sj | > ε n
0⩽j⩽⌊δn⌋+1 0⩽j⩽⌊δn⌋+1
0⩽k⩽n+1
√
 
C
⩽ P max |Sj | > ε n .
δ 0⩽j⩽⌊δn⌋+1

Hence, the proof follows if we show that

√
 
1
lim lim sup P max |Sj | > ε n = 0 . (2.6)
δ↓0 δ n→∞ 0⩽j⩽⌊δn⌋+1

Therefore, our aim is now to prove (2.6). The idea is to use the CLT, which gaurantees
that (roughly)
√
√ ε2
lim P S⌊nδ⌋+1 ⩾ ε n ⩽ P(|Z| ⩾ δε) ≲ e− 2δ , (2.7)
n→∞
22 CHAPTER 2. UNIVERSALITY AND FUNCTIONAL CLTS

ε2
so that δ −1 e− δ → 0 as δ → 0. Of course, we first have to deal with the maximum over
the variable j. Let us introduce the stopping time
√
τ = inf{j : |Sj | > ε n} ,

and let us introduce a slack parameter

η ∈ (0, ε) .

Now we have the inclusion of events

√ √
 

max |Sj | > nε ⊆ |S⌊δn⌋+1 | > n(ε − η)
0⩽j⩽⌊δn⌋+1
[  √

|S⌊δn⌋+1 | < n(ε − η) ∩ {τ < ⌊δn⌋ + 1} .

Then we can bound

√ √
 
P max |Sj | > nε ⩽P(|S⌊δn⌋+1 | > n(ε − η))
0⩽j⩽⌊δn⌋+1
⌊δn⌋+1
X √
+ P(|S⌊δn⌋+1 | < n(ε − η)|τ = j)P(τ = j) .
j=1

Now, on the event √

{|S⌊δn⌋+1 | < n(ε − η)} ∩ {τ = j}
we have that √
|S⌊δn⌋+1 − Sj | ⩾ nη .
So that by Chebyshev’s ineqality

E |S⌊δn⌋+1 − Sj |2 |τ = j
 
√

P |S⌊δn⌋+1 | < n(ε − η)|τ = j ⩽
nη 2
⌊nδ⌋ + 1 1
⩽ E|X1 |2 2
⩽ ,
nη 2
where we could bound the expecation by using the independence of increments and the
last bound follows for n large by fixing
√
η = 3δ .

Overall, we have now obtained that

√ √
 
1
P max |Sj | > nε ⩽P(|S⌊δn⌋+1 | > n(ε − η)) + P(τ ⩽ ⌊nδ⌋) .
0⩽j⩽⌊δn⌋+1 2

This in turn implies

√ √
 
P max |Sj | > nε ⩽2P(|S⌊δn⌋+1 | > n(ε − η)) .
0⩽j⩽⌊δn⌋+1

Hence the desired (2.6) follows from (2.7).

Chapter 3

BM as a strong Markov process

3.1 The Markov property

For a stochastic process X on a probability space (Ω, F, P) we denote by

Px (·) := P[·|X(0) = x]

the law of the process started in x, and write Ex for the corresponding expectation.

Definition

3.1.1. [Filtration]A filtration on a probability space (Ω, F, P) is a family
Ft t⩾0 of σ-algebras such that F(s) ⊆ F(t) ⊆ F for all 0 ≤ s ≤ t.

(a) (Ω, F, (F(t))t⩾0 , P) is called a filtered probability space.

(b) A process X defined on a filtered probability space with filtration Ft t⩾0
is called
adapted if X(t) is F(t)-measurable for any t ≥ 0.

Lemma 3.1.2. A process

0

X on (Ω, F, P) taking values in (S, S) is adapted to the natural
filtration F (t) : t ≥ 0 , given by

F 0 (t) := σ X(s) : 0 ≤ s ≤ t = σ {ω : X(s, ω) ∈ A} for all A ∈ S and 0 ≤ s ≤ t .

The proof is immediate from the definition of F 0 (t). Intuitively, F 0 (t) contains all the
information from observing the process up to time t.

By analogy with Markov chains, we say that X possesses the Markov property if the
state of the process at time t ⩾ s, given its entire history up to time s, depends only on the
state Xs at time s. More precisely, an adapted process X : Ω → S is a Markov process
if
P[X(t) ∈ A | F(s)] = P[X(t) ∈ A | X(s)]
for all 0 ≤ s ≤ t and A ∈ S. Equivalently, we can define Markov property in terms of
transition kernels:

Definition 3.1.3. A function P : {0 ≤ s ≤ t < ∞} × S × S → [0, 1] is a Markov

transition kernel if

(a) P (·, ·, ·, A) is measurable as a function of (s, t, x) for each A ∈ S;

(b) P (s, t, x, ·) is a probability measure on (S, S) for all t ≥ s ≥ 0 and x ∈ S;

23
24 CHAPTER 3. BM AS A STRONG MARKOV PROCESS

(c) for all A ∈ S, x ∈ R and t > u > s > 0

Z
P (s, t, x, A) = P (u, t, y, A) P (s, u, x, dy) . (3.1)
S


Let X(t) : t ≥ 0 be an adapted process
on a filtered probability space (Ω, F, [F(t)], P)
with state space (S, S). X(t) :
t ≥ 0 is a Markov process with transition kernel P
w.r.t. the filtration F(t) : t ≥ 0 if for all t ≥ s and sets A ∈ S we have
 
P X(t) ∈ A F(s) = P (s, t, X(s), A) almost surely . (3.2)

Remarks.

• P (s, t, x, A) is the probability that the process takes a value in A at time t, provided
it is in state x at time s. Comparing the two equivalent definitions of a Markov
process, we find
 
P (s, t, x, A) = P X(t) ∈ A X(s) = x , t ≥ s ≥ 0 . (3.3)

• Property (c) is also known as the Chapman-Kolmogorov equations. It is just an

expression of consistency of the distributions of (X(s), X(tI ), X(t)) and (X(s), X(t))
for any intermediate time tI , s < tI < t.

• Note that for processes with continuous paths we have from (3.3)

1 , x∈A
P (t, t, x, .) = δx i.e. P (t, t, x, A) = .
0 , x ̸∈ A

• Markov process is called time-homogeneous if the transition kernel depends only

on time delay between transitions, P (s, t, x, A) = P (0, t − s, x, A) for all 0 ≤ s ≤ t,

A) ∈ S ×S. Informally, the time homogeneity property states that X(s+t) : t ≥
(x,
0 ∼ PX(s) for any fixed s ≥ 0, meaning that the time shifted process has the same
distribution as process restarted at the current position X(s). Whenever it does
not lead to confusion, we will denote the transition kernel for a time-homogeneous
Markov process using the same symbol P , P (s, t, ·, ·) → P (t − s, ·, ·). Therefore,
the Chapman-Kolmogorov equation (3.1) for a time-homogeneous Markov process
becomes Z
P (s + t, x, A) = P (s, dy) P (t, y, A).
S

• The definitions of conditional expectations and probabilities used in this Chapter

are recalled in handout 3.

• A Markov process w.r.t. its natural filtration is simply called a Markov process.

Definition 3.1.4. Two processes X and Y are called independent, if all finite dimen-
sional observations are independent, i.e. for all t1 , . . . , tn , n ∈ N and s1 , . . . , sm , m ∈ N



X(t1 ), . . . , X(tn ) ⊥ Y (s1 ), . . . , Y (sm ) .

The process X is called independent of a σ-algebra G if

X(t1 ), . . . , X(tn ) ⊥ 1A1 , . . . , 1Am for all A1 , . . . , Am ∈ G .
3.1. THE MARKOV PROPERTY 25

Remarks.

• For a BM B we have (S, S) = (R, B) and Px is the law of the process t 7→ x + B(t),
called BM started in x.

• If B and B ′ are independent BMs, then t 7→ x + σB(t) is a BM started in x with

diffusion coefficient σ 2 > 0, and t 7→ x + σB(t) + σ ′ B ′ (t) is a BM started in x
with diffusion coefficient σ 2 + σ ′2 .

Theorem 3.1.5. The following two hold true.

1. A Brownian motion B(t) : t ≥ 0 started at x is a time-homogeneous Markov
process (w.r.t. its natural filtration) with transition kernel density
Z
1 −(y−x)2 /(2t)
p(t, x, y) = √ e and transition kernel P (t, x, A) = p(t, x, y) dy .
2πt A
(3.4)


2. For any fixed s > 0, B(t + s) − B(s) : t ≥ 0 is a Brownian
motion started at the
origin, and it is independent of the process B(u) : 0 ≤ u ≤ s .

Remark 3.1.6. The Gaussian transition kernel density (or just density) is translation
invariant, i.e.
p(t, x, y) = p(t, 0, y − x) for all t ≥ 0, x, y ∈ R . (3.5)

Proof. We know that the
time-shifted process is still a Brownian motion, and independence
of B(u) : 0 ≤ u ≤ s follows directly from independence of increments of Brownian
motion, which proves (ii).
The full proof of part (i) can be based on finite-dimensional approximations of BM.
Here is a skeleton of a more direct proof: for any t, s ≥ 0, A ∈ B,
Z
0 0
P[B(t + s) ∈ A | F (s)] = E[1A (Bt+s ) | F (s)] = E[1B(t+s)∈A δx (B(s)) | F 0 (s)]dx
Z R

= E[1B(t+s)−B(s)∈A−x δx (B(s)) | F 0 (s)]dx

ZR
(a)
= E[1B(t+s)−B(s)∈A−x | F 0 (s)]δx (B(s))dx
ZR
(b)
= E[1B(t+s)−B(s)∈A−x ]δx (B(s))dx
R
Z Z 
(c) dy −y2 /(2t)
= √ e δx (B(s))dx
R A−x 2πt
Z
dy −(y−B(s))2 /(2t) (d)
= √ e = P(B(t + s) ∈ A | B(s)).
A 2πt

The equality (a) is due to the measurability of B(s) w. r. t. F 0 (s); (b) is due to the
independence of the increment B(t + s) − B(s) from F 0 (s); (c) is due to B(t + s) − B(s) ∼
N (0, t); (d) follows from (ii). The last equality proves that BM is a Markov process,
the penultimate inequality yields the transition kernel. To make the above rigorous one
needs to approximate δx with a sequence of bounded continuous functions and analyze
the sequence of corresponding expectations.
26 CHAPTER 3. BM AS A STRONG MARKOV PROCESS

Example 3.1.7. Brownian motion with drift t 7→ X(t) := B(t) + ct for some c ∈ R
is a also a time-homogeneous Markov process. Using the Markov property of BM we have



X(t+s)|F (s) = B(t+s)|B(s) +c(t+s) = B(t+s)−B(s) B(s) +ct+X(s) ∼ X(s)+N ct, t ,

which depends only on X(s) and t, so that X has transition density

1 −(y−x−ct)/(2t)
p(t, x, y) = √ e .
2πt
Now, the following generalisation of Theorem 3.1.5 is useful for finding transition
kernels explicitly:

Lemma 3.1.8. If (X(t), t ≥ 0) is a stochastic process with independent increments, then

X is a Markov process. If the distribution of (X(t) − X(s)) for t > s is continuous with
density p(s, t, ·), then the transition kernel is
Z
P (s, t, x, A) = p(s, t, y − x)dy.
A

The proof follows the same steps as the one of Theorem 3.1.5.

3.2 Blumenthal’s 0-1 law

0

a stochastic process X on (Ω, F, P) recall the notion of natural filtration F (t) : t ≥
For
0 as was introduced in Lemma 3.1.2. There is a second filtration that is canonically
associated to a stochastic process.

Definition 3.2.1. We define the augmented filtration F + (t) : t ≥ 0 by

\
F + (t) := F 0 (s) ⊇ F 0 (t) .
s>t

Remark 3.2.2. We observe the following.

1. The augmented filtration is larger (or finer) than the natural one, and F + (t) contains
infinitesimal information about the future, such as the event

ω : t 7→ X(t, ω) is differentiable at t .


2. By Theorem 3.1.5, for every s ≥ 0 the process B(t + s) − B(s) : t ≥ 0 is indepen-
dent of F 0 (s). The next result shows that this independence can be extended to the
augmented filtration.


Theorem 3.2.3. For every s ≥ 0 the process B(t + s) − B(s) : t ≥ 0 is independent of
F + (s).

Proof. By definition, (B(t+s)−B(s), s ≥ 0) is independent +

of F (s) if for all t+1 , t2 , . . . , tm ≥
0, the vector B(t1 + s) − B(s), . . . , B(tm + s) − B(s) is independent of F (s).
Therefore, to prove the desired independence, let sn ↘ s be a strictly decreasing
sequence converging to s. By continuity of paths,

B(t + sn ) − B(sn ) → B(t + s) − B(s) almost surely as n → ∞

3.2. BLUMENTHAL’S 0-1 LAW 27

By definition of F + (s), B(t1 + sn ) − B(sn ), . . . , B(tm + sn ) − B(sn ) is independent of

F + (s) for any n < ∞. Therefore, for any bounded continuous F : Rm → R,

E[F (B(t1 + s) − B(s), . . . , B(tm + s) − B(s)) | F + (s)]

= E[ lim F (B(t1 + sn ) − B(sn ), . . . , B(tm + sn ) − B(sn )) | F + (s)]
n→∞
= lim E[F (B(t1 + sn ) − B(sn ), . . . , B(tm + sn ) − B(sn )) | F + (s)]
n→∞
= lim E[F (B(t1 + sn ) − B(sn ), . . . , B(tm + sn ) − B(sn ))]
n→∞
= E[F (B(t1 + s) − B(s), . . . , B(tm + s) − B(s))].

Extending this result to all measurable integrable functions (measure theory details omit-
ted), we get
B(t1 + s) − B(s), . . . , B(tm + s) − B(s) ⊥ F + (s) ,

and thus the process B(t + s) − B(s) : t ≥ 0 is independent of F + (s).

0 (t) : t ≥ 0 ,


Definition 3.2.4. For a process X on (Ω, F, P) with natural filtration F
F + (0) = t>0 F 0 (t) is called the germ σ-algebra, and the tail σ-algebra is defined as
T

\

T := σ X(s) : s ≥ t .
t≥0

Theorem 3.2.5 (Blumenthal’s 0-1 law). Consider BM started in x ∈ R and A ∈ F + (0).

Then Px [A] ∈ {0, 1}.



Proof. For fixed B(0) = x ∈ R we have σ B(t) + x : t ≥ 0 = σ B(t) : t ≥ 0 which
is independent of F + (0) by

Theorem 3.2.3 above for s = 0. But since at the same time
F + (0) ⊆ σ B(t) : t ≥ 0 , any A ∈ F + (0) is independent of F + (0), so in particular
independent of A itself. Therefore, Px [A] = Px [A ∩ A] = Px [A]2 which implies the state-
ment.

Theorem 3.2.6. Consider BM started in x ∈ R and A ∈ T . Then Px [A] ∈ {0, 1}.

Proof. It suffices to look at the case x = 0. Consider

any probability space for BM , e. g.
the canonical model for BM B = B(t) : t ≥ 0 on (C0 , C, P), where C0 is the space of
continuous functions such that
lim ωt = 0 , lim t−1 ωt = 0 .
t→0 t→∞

Consider then the time inversion bijection ψ : C0 → C0 defined by ψω(t) = tω(1/t), t > 0
and ψω(0) = 0. In Proposition 1.3.5 we have shown that the process B◦ψ is again BM, and
by definition tail events are mapped to germ events and vice versa, so that ψT = F + (0).
So tail events for B are germ events for B ◦ ψ , and can therefore only have probability 0
or 1 according to Blumenthal’s 0-1 law.

Remark 3.2.7. We observe the following.

• Recall that for any index set Λ (finite or infinite, countable or uncountable) the
intersection ∩λ∈Λ Fλ of σ-algebras is a σ-algebra.

• In general, events in the tail

 σ-algebra can only depend on the limiting behaviour of
the process, involving e.g. limt→∞ X(t) exists . Unless the process
 is trivial (e.g.
X(t) ≡ X(0) for all t ≥ 0, where T = F0 ), events such as X(t) ≤ 0 for fixed
t > 0 are not tail events.
28 CHAPTER 3. BM AS A STRONG MARKOV PROCESS

• We will see in an exercise sheet that for tail events A ∈ T of Brownian motion,
Px [A] is in general independent of x, whereas for a germ event A ∈ F + (0) it may
depend on x.

As a first application we show that BM takes positive and negative values and zeros
in every small time interval to the right of 0, and we look at properties of local and global
extrema.

Theorem  3.2.8. For a Brownian motion B define τ := inf t > 0 : B(t) > 0 and
λ := inf t > 0 : B(t) = 0 . Then

P0 [τ = 0] = P0 [λ = 0] = 1 .

Proof. The event (exercise: check the measurability)

∞
\
there is 0 < ϵ < 1/n such that B(ϵ) > 0 ∈ F + (0) ,

{τ = 0} =
n=1

so by Theorem 3.2.5 we just have to show that the event in question has positive proba-
bility. This follows as
P0 [τ ≤ t] ≥ P0 [B(t) > 0] = 1/2
for all t > 0. Hence P0 [τ = 0] ≥ 1/2 and we have shown the first part. The same argument
works by replacing B(t) > 0 with B(t) < 0, and this implies also that P0 [λ = 0] = 1 using
continuity of paths and the intermediate value property.

Theorem 3.2.9. For a Brownian motion B, almost surely the following hold:
(a) Every local maximum is a strict local maximum.

(b) The set of times where the local maxima are attained is countable and dense.

(c) The global maximum is attained at a unique time.

Proof. We first show that on any two fixed non-overlapping time intervals [a1 , b1 ] and
[a2 , b2 ] with b1 ≤ a2 the corresponding maxima m1 and m2 of BM on are different almost
surely: by Theorem 3.2.8 almost surely B(a2 ) < m2 , hence m2 agrees with the maximum
P∞the interval [a2 + 1/n, b2 ] for some n ∈ N. Therefore, P[max[a1 ,b1 ] B = max[a2 ,b2 ] B] ≤
on
n=1 P[max[a1 ,b1 ] B = max[a2 +1/n,b2 ] B]. To estimate the r. h. s. it is enough to consider a
pair of intervals: b1 < a2 . Using Markov property twice, we can write the event {m1 = m2 }
in terms of three independent random variables B(a2 ) − B(b1 ), m1 − B(b1 ), m2 − B(a2 )1 :


{m1 = m2 } = B(a2 ) − B(b1 ) = m1 − B(b1 ) − m2 − B(a2 )

Conditioning on the values of m1 − B(b1 ) and m2 − B(a2 ) the right-hand side is constant
while the left-hand side is a continuous random variable (recall that a2 > b1 ), which implies
that P[m1 = m2 ] = 0. Applying this result to a countable set of intervals with rational
endpoints, we find

P any two non-overlapping non-degenerate compact time intervals

with rational endpoints have different maxima = 1 . (3.6)
1
The second and the third variables are not Gaussian, as times of maxima are not stopping times, see
below.
3.3. THE STRONG MARKOV PROPERTY 29

(a) If BM had non-strict local maxima, then there would exist two such intervals where
BM has the same maximum, which is almost surely not the case.
(b) Theorem 3.2.8 implies that the maximum over any non-degenerate compact time
interval with rational endpoints is not attained at an endpoint. So each such interval
contains a local maximum, and the set of times ∆ ⊆ [0, 1] where they are attained is
dense. Since each local maximum is strict, ∆ has at most the cardinality of the set of such
intervals and is therefore countable.
(c) Due to (3.6), almost surely, for any q ∈ [0, 1] ∩ Q the maxima on [0, q] and [q, 1]
are different. If the global maximum were attained at two points t1 < t2 one could find
t1 < q < t2 for which the maxima on [0, q] and [q, 1] would agree, which has probability
zero.

3.3 The strong Markov property

Intuitively, the Markov property states that Brownian motion is started afresh at each
deterministic time instance. This also holds for an important class of random times,
which are introduced in the following.

Definition 3.3.1. A random variable T with values in [0, ∞], defined on a filtered prob-
ability
space (Ω, F, [F(t)], P) is called a stopping time with respect to the filtration F(t) :
t ≥ 0 , if {T ≤ t} ∈ F(t) for all t ≥ 0.

Remark 3.3.2 (Properties of stopping times). We observe the following.

• Every deterministic time t0 ≥ 0 is a stopping time w.r.t. every filtration, since of

course {t0 ≤ t} ∈ {∅, Ω} for all t ≥ 0.

• If (Tn : n ∈ N) is an increasing sequence of stopping times
w.r.t. F(t) : t ≥ 0 and

Tn ↗ T , then T is also a stopping time w.r.t. F(t) : t ≥ 0 , since

∞
\
{T ≤ t} = {Tn ≤ t} ∈ F(t) .
n=1

• On the other hand, if T is a stopping time w.r.t. F(t) : t ≥ 0 define for any n ≥ 1

Tn := (m + 1)2−n if m2−n ≤ T < (m + 1)2−n , (3.7)

i.e. we approximate the value of T by the
smallest dyadic integer k2−n > T . Then
Tn are stopping times w.r.t. F(t) : t ≥ 0 since, for t ∈ [m/2n , (m + 1)/2n )
∞
[
Tn ≤ t = T < m/2n = T ≤ m/2n − 1/q ∈ F(t) .
  
q=1

The sequence of stopping times (Tn ) converges to T , Tn ↘ T . It will be used below

for the proof of strong Markov property.

• Every stopping time T w.r.t. the natural filtration F
0 (t) : t ≥ 0 is also a stopping

time w.r.t. the augmented filtration F + (t) : t ≥ 0 , since F 0 (t) ⊆ F + (t) for all
t ≥ 0.
30 CHAPTER 3. BM AS A STRONG MARKOV PROCESS

An important property distinguishing the augmented from the natural filtration is the
right-continuity of the former:
Lemma 3.3.3. Consider a process with natural
filtration F 0 (t) : t ≥ 0 . Then the

corresponding augmented filtration F + (t) : t ≥ 0 is right-continuous, i.e.

\
F + (t) = F + (t + ϵ) for all t ≥ 0 .
ϵ>0

Suppose a random variable

T ∈ [0, ∞] satisfies {T < t} ∈ F(t) for every t ≥ 0 and the
filtration

F(t) : t ≥ 0 is right-continuous. Then T is a stopping time w.r.t. F(t) : t ≥
0 .
Proof. The first part follows directly from the Definition ?? via
\ ∞ \
\ ∞
F + (t + ϵ) = F 0 (t + 1/n + 1/k) = F + (t) .
ϵ>0 n=1 k=1


For the second part, given T as above and using right-continuity of F(t) : t ≥ 0 we get
∞
\ ∞
\
{T ≤ t} = {T < t + 1/n} ∈ F(t + 1/n) = F(t) .
n=1 n=1

The next proposition provides a large class of examples of stopping times.

Proposition 3.3.4. Consider BM started in x with natural and augmented filtrations.
For a set H ∈ B define the first hitting time as

TH := inf t ≥ 0 : B(t) ∈ H .

If H is closed (e.g. H = {h}), then TH is a stopping time w.r.t. F 0 (t) : t ≥ 0 .

+ (t) : t ≥ 0 ,


If G ⊆ R is open, then TG is a stopping time w.r.t. the augmented filtration F
but not necessarily w.r.t. F 0 (t) : t ≥ 0 .

Proof. Fix a closed H ⊆ R. Note that

TH (ω) ≤ t ⇔ ∃s ∈ [0, t] ∃x ∈ H : B(s)(ω) = x .

As usual we want to replace quantifiers by countable unions and intersections and write
∞
\ [ [ n  1 1 o
{TH ≤ t} = B(s) ∈ x − , x + ∈ F 0 (t) ,
n
n n
n=1 s∈Q∩[0,t] x∈Q∩H | {z }
∈F 0 (s)

where H n is the enlarged set

H n = {x ∈ R : (x − 1/n, x + 1/n) ∩ H ̸= ∅} .

Note that the equality above holds only because H is closed.

Now fix an open set G. By continuity of Brownian paths
\ \ [
{TG ≤ t} = {TG < s} = {B(r) ∈ G} ∈ F + (t),
| {z }
s>t s>t r∈Q∩(0,s)
∈F 0 (s)
3.3. THE STRONG MARKOV PROPERTY 31

which means that TG is a stopping time w.r.t. F + (t) : t ≥ 0 .

Finally, suppose that G is bounded and that the starting point x ̸∈ Ḡ. Then we can
fix a path γ : [0, t] → R with γ(0, 0
 t) ∩ Ḡ = ∅ and γ(t) ∈ ∂G. Then the σ-algebra F (t)
contains no non-trivial subset of B(s) = γ(s) for all 0 ≤ s ≤ t , i.e. no subset other than
the empty set and the set itself. If we had {TG ≤ t} ∈ F 0 (t) then

B(s) = γ(s) for all 0 ≤ s ≤ t, TG = t

would be in F 0 (t) and a non-trivial subset of this set, which is a contradiction.

From now on, we will silently assume that a stopping time is defined with respect
to a right continuous filtration such as an augmented filtration. Moreover, we will be
checking that T is a stopping time either using strict or non-strict inequalities, as for a
right continuous filtration {T < t} ∈ F(t) ⇐⇒ {T ≤ t} ∈ F(t).
Definition 3.3.5. For any stopping time T we define the σ-algebra

F + (T ) = A ∈ F : A ∩ {T ≤ t} ∈ F + (t) for all t ≥ 0 .



This means that the part of A that lies in {T ≤ t} should be measurable w.r.t. the
information available at time t. Heuristically, this is the collection of events that hap-

pened before the stopping time T . One can also check that the path B(t) : 0 ≤ t ≤ T
is F + (T )-measurable. Now we can state the strong Markov property for BM, which was
rigorously established by Hunt (1956) and Dynkin (1957).

Theorem 3.3.6 (Strong Markov property). For every almost surely finite stopping time
T , the process

B(T + t) − B(T ) : t ≥ 0 is a BM independent of F + (T ) .

Proof. Consider the stopping

times Tn as defined in (3.7) which approximate T from above.
Write Bk = Bk (t)
: t ≥ 0 for the BM defined by Bk (t) = B(t + k/2n ) − B(k/2n ) and
B∗ = B∗ (t) : t ≥ 0 for a process defined by B∗ (t) = B(t + Tn ) − B(Tn ).
Fix an arbitrary E ∈ F + (Tn ) and an event A ∈ C on path space (C0 , C). Then we have
∞
 X
P {Bk ∈ A} ∩ E ∩ {Tn = k2−n }
  
P {B∗ ∈ A} ∩ E =
k=0
∞
X
P[Bk ∈ A] P E ∩ {Tn = k2−n } ,
 
=
k=0

using that {Bk ∈ A} is independent of E ∩ {Tn = k2−n } ∈ F + (k2−n ) by Theorem ??.

With the Markov property of BM (Theorem 3.1.3), P[Bk ∈ A] = P[B ∈ A] does not
depend on k, and hence
∞
X
P E ∩ {Tn = k2−n } = P[B ∈ A]P[E] .
   
P {B∗ ∈ A} ∩ E = P[B ∈ A]
k=0

So B∗ is a BM independent of E, and hence of F + (Tn ).

For

a general stopping time T note that Tn ↘ T and for each n ≥ 0, B(Tn +s)−B(Tn ) :
s ≥ 0 is a BM independent of F (Tn ) ⊃ F + (T ). Thus the increments
+

B(s + t + T ) − B(t + T ) = lim B(s + t + Tn ) − B(t + Tn ) (3.8)

n→∞
32 CHAPTER 3. BM AS A STRONG MARKOV PROCESS

of the process B(T + u) − B(T ) : u ≥ 0 are independent and normally distributed
with mean zero and variance s. Since its paths are obviously almost surely continuous the
process is BM. Moreover, all increments (3.8), and hence the process itself, are independent
of F + (T ).

Remark 3.3.7. As an example, we observe that τ := inf{t ≥ 0 : B(t) = max0≤s≤1 B(s)}

is not a stopping time. Indeed, almost surely, τ < 1 by Theorem 3.2.9. Therefore, for all
sufficiently small t ≥ 0, B(τ + t) − B(τ ) < 0. If we assume τ is a stopping time, then due
to the strong Markov property B(τ + t) − B(τ ) is BM, and the above inequality contradicts
Theorem 3.3.6.

3.4 The reflection principle

The reflection principle is a particularly interesting application of the strong Markov prop-
erty, stating that BM reflected about the line ([T, ∞), B(T )) for some stopping time T is
still BM.


Theorem 3.4.1. If B(t) : t ≥ 0 is a BM and
T ∈ [0, ∞] is a stopping time, then the
Brownian Motion reflected at T , B ∗ (t) : t ≥ 0 defined by

B ∗ (t) = B(t)1{t≤T } + 2B(T ) − B(t) 1{t>T }

is also Brownian Motion.

Proof. If T = ∞ there is nothing to show. If T < ∞, by the strong Markov property and
symmetry of BM

t 7→ X(t) := − B(t + T ) − B(T )


is BM and independent of B(t) : 0 ≤ t ≤ T . Therefore, the continuous concatenation

B(t) , t≤T
t 7→ B ∗ (t) =
B(T ) + X(t − T ) , t > T

is Brownian Motion.

Remark 3.4.2. As a counter-example, the process reflected at τ defined in Remark 3.3.7

is not a Brownian motion: the reflected portion of the path will stay above B(τ ) in some
non-empty interval to the right of τ , which again contradicts the strong Markov property
and Theorem 3.2.8.
Now we apply the reflection principle to find the distribution of the

maximum process t 7→ M (t) := max B(t) . (3.9)

0≤s≤t


Proposition 3.4.3. Consider a Brownian Motion B(t) : t ≥ 0 and M (t) as defined in
(3.9). Then for any a > 0
     
P0 M (t) > a = 2P0 B(t) > a = P0 |B(t)| > a .

Proof. The hitting time Ta = inf{t
≥ 0 : B(t) = a is a stopping time (cf. Proposi-
tion 3.3.4). Then let B ∗ (t) : t ≥ 0 be BM reflected at Ta . The event {Mt > a} can be
written as a disjoint union:
  
M (t) > a = B(t) > a ∪ B(t) ≤ a, M (t) > a
3.5. MARKOV PROCESSES RELATED TO BROWNIAN MOTION 33

The second event coincides with B ∗ (t) > a , and hence the first statement follows from


the reflection principle. For the second statement we have the disjoint union
  
|B(t)| > a = B(t) > a ∪ B(t) < −a

which have the same probability by reflection symmetry of BM.

Remark 3.4.4. This result implies that the PDF fM for the maximum M (t) of BM up
to time t is given by
2 −x2 /(2t)
fM (x) = √ e for x > 0 .
2πt
In combination with Gaussian tail estimates, we obtain
√
2t 2
P0 M (t) > a ≤ √ e−a /(2t) for all a, t > 0 .
 
a π

Corollary 3.4.5. As a consequence of Proposition 3.4.3, we have that almost surely

Ta < ∞ , ∀a ∈ R .

And in addition almost surely

lim sup Bt = +∞ , lim inf Bt = −∞ .

t→∞ t→∞

Proof. For the first claim, we observe that

P(Ta < ∞) = P( lim Mt > a) = lim P(Mt > a) ,

t→∞ t→∞

since Mt is an increasing process, and therefore {Mt > a} an increasing sequence of sets.
Now
√
P(Mt > a) = P(|B1 | ⩾ a/ t) → 1 ,
as t → ∞. The nullset can be chosen uniformly over all a, by considering a countable set
of a’s. As for the second claim, we observe that

{ lim Mt = ∞} = {lim sup Bt = ∞} ,

t→∞ t→∞

together with the fact that P(limt→∞ Mt = ∞) = 1 by the previous calculation.

3.5 Markov processes related to Brownian motion

In
addition to Brownian Motion itself, reflected Brownian Motion defined as |B(t)| : t ≥
0 is a Markov process. The transition kernel P (t, x, .) is the law of |Y | with a Gaussian
Y with mean x and variance t, called the modulus or folded normal distribution with
parameters x and t. It has density

1  − (y−x)2 (y+x)2

p(t, x, y) = √ e 2t + e− 2t for x, y ⩾ 0 .
2πt

Next, we observe that the sequence of stopping times (Ta )a∈R forms itself a Markov process.
34 CHAPTER 3. BM AS A STRONG MARKOV PROCESS


Theorem 3.5.1. Consider a Brownian Motion B(t) : t ≥ 0
and define Ta := inf t ≥
0 : B(t) = a for any a ≥ 0. Then the process Ta : a ≥ 0 is an increasing Markov
process with transition kernel density

a  a2 
p(a, t, s) = p exp − 1 for all a > 0 .
2π(s − t)3 2(s − t) {s>t}

Proof. Fix a ⩾ b ⩾ 0 and note that for all t ⩾ 0 we have


{Ta ⩾ Tb + t} = B(Tb + s) − B(Tb ) < a − b for s < t .

By the strong Markov property for BM this event is independent F + (Tb ), and therefore
in particular of (Td : d ≤ b) ∈ F + (Tb ). This implies the Markov property for Ta : a ≥ 0 ,

and the transition kernel follows from the reflection principle:

h i
P[Ta − Tb ≤ t] = P[Ta−b ≤ t] = P max B(s) ≥ a − b
0≤s≤t
Z ∞
  1 −x2 /(2t)
= 2P B(t) ≥ a − b = 2 √ e dx
a−b 2πt
Z t
a − b −(a−b)2 /(2s)
= √ e ds ,
0 2πs3
p
where we use the substitution x = t/s(a − b) in the last step.

Remark 3.5.2. We observe the following:

• An analogous result holds for hitting times of −a ≤ 0 by symmetry of Brownian

Motion.

• The transition densities are homogeneous (notice that ’time’ and ’space’ for the pro-
cess T are interchanged

in comparison with Brownian Motion). Therefore the process
Ta : a ≥ 0 has stationary and independent increments. However, it’s paths are not
continuous.

• This process is called stable subordinator of index 21 , it is a member of the family

of Levy processes.

• The kernel density p(a, t, s) is an inverse Gamma distribution or Lévy distribution

in s, which is a stable distribution.

• Note that for t = 0 this result implies a heavy tail (with ’index’ 1/2)

a 2a
p(a, 0, s) ≃ s−3/2 √ P a, 0, [s, ∞) ≃ s−1/2 √


and as s → ∞ .
2π 2π

So the expected hitting time for BM is E[Ta ] = ∞ for all a ̸= 0 (using symmetry for
negative a). Intuitively, even though Ta < ∞ almost surely for all a, large positive
or negative excursion of BM lead to a diverging contribution to the expectation.

• Another simple but useful consequence for later is that monotonicity of the process
and independence of increments implies that Ta → ∞ as a → ∞ almost surely. We
summarize all useful direct consequences of Theorem 3.5.1 in the next result.
3.5. MARKOV PROCESSES RELATED TO BROWNIAN MOTION 35


Corollary 3.5.3. Consider BM B(t) : t ≥ 0 and let Ta := inf t ≥ 0 : B(t) = a for
a ∈ R. Then

Ta < ∞ almost surely for all a ∈ R ,

Ta → ∞ almost surely as |a| → ∞ ,
E[Ta ] = ∞ for all a ̸= 0 .


Theorem 3.5.4
(Lévy (1948)). Let M (t) : t ≥ 0 be the
maximum process for BM
B(t) : t ≥ 0 with M (t) := max0≤s≤t B(s). Then Y (t) : t ≥ 0 with Y (t) := M (t) − B(t)
is a reflected BM.


Idea of proof. The process Y (t) : t ≥ 0 has almost surely continuous paths. For fixed
s > 0, consider the process


B̂(t) : t ≥ 0 with B̂(t) := B(t + s) − B(t) ,

which is a BM. Since Y (s) is F + (s)-measurable, it suffices to show that conditional on

F + (s)
Y (s + t) F + (s) ∼ Y (s) + B̂(t) for all t ≥ 0 , (3.10)

which can be established using the reflection principle (see Q. 2.8 of the assignment for
the main computational step) and time reversibility of BM.

Remark 3.5.5. It is worth noticing that the maximum process itself is not Markov.
Chapter 4

More path properties of Brownian

motion

4.1 Modulus of continuity

By definition of Brownian Motion, sample paths are almost surely continuous, which
implies that on [0, 1], or any compact time interval, sample paths are uniformly continuous.
Uniform continuity coincides with the existence of a modulus of continuity of B : [0, 1] →
R: a (possibly random) function ϕ : [0, 1] → R with ϕ(h) → 0 as h ↘ 0 and ϕ(0) = 0 such
that |B(x) − B(y)| ≤ ϕ(|x − y|) for all x, y ∈ [0, 1]. The ’optimal’ modulus of continuity
satisfies

|B(t + h) − B(t)|
lim sup sup ≤1. (4.1)
h↘0 0≤t≤1−h ϕ(h)
p
Our first result shows that one can choose c > 0 so small that c h log(1/h) fails to be a
modulus of continuity.
√
Theorem 4.1.1. For every constant c < 2, almost surely, for every ϵ > 0 there exists
0 ≤ h ≤ ϵ and t ∈ [0, 1 − h] such that
p
B(t + h) − B(t) ≥ c h log(1/h) .
√
Proof. Let c < 2 and fix ϵ > 0. We have to show that P[Aϵ ] = 1 with
 p
Aϵ := ∃h ≤ ϵ ∃t ∈ [0, 1 − h] : B(t + h) − B(t) ≥ c h log(1/h) .
For integers k, n ≥ 0 define the events
n √ o
Ak,n := B (k + 1)e−n − B ke−n ≥ c n e−n/2 .

Choosing h = e−n with n > | log ϵ| we see that

∞ ⌊en −1⌋
[ [
Ak,n ⊆ Aϵ , (4.2)
n=| log ϵ| k=0

and in the following we can use a lower bound on P[Ak,n ] to show the the left-hand side
has probability 1. Using stationarity of increments, scaling and the Gaussian tail estimate
we get independently of k ≥ 0
√
−n √ −n/2  √  c n 1 2
√ e−c n/2 .


P[Ak,n ] = P B e > c ne = P B(1) > c n ≥ 2
c n + 1 2π

36
4.2. VARIATION OF BROWNIAN PATHS 37

√
Since c < 2 this implies that en P[Ak,n ] → ∞ as n → ∞. Therefore, using 1 − x ≤ e−x
for all x ∈ R, the probability that for given n none of the Ak,n occurs satisfies
n −1⌋
 ⌊e\ 

en
Ack,n = 1 − P[A0,n ] ≤ exp − en P[A0,n ] → 0


P as n → ∞ ,
k=0

where we used the independence of Brownian increments. This implies that

n −1⌋
 ⌊e[ 
P Ak,n → 1
k=0

and therefore the left-hand side of (4.2) occurs almost surely.

p
The next result goes in the opposite direction, establishing that φ(h) = C h log (1/h)
is indeed a modulus of continuity for C sufficiently large.

Theorem 4.1.2. There exists a constant C > 0 such that, almost surely, for all sufficiently
small h > 0 and all 0 ≤ t ≤ 1 − h,
p
B(t + h) − B(t) ≤ C h log(1/h) . (4.3)

The proof is omitted. One can determine the optimal constant c for the best possible
modulus of continuity which achieves
p equality in (4.1). √
The above lower bound on the
increments implies that ϕ(h) ≥ 2h log(1/h), and indeed 2 can be shown to be optimal.

|B(t + h) − B(t)|
Theorem 4.1.3. Almost surely, lim sup sup p =1.
h↘0 0≤t≤1−h 2h log(1/h)

Proof. See pp. 16-17 in [?]. A stronger statement with a lim replacing lim sup in the l. h.
s. of (4.1) can also be proved.

4.2 Variation of Brownian paths

Definition 4.2.1. Let Π = {t0 = 0 < t1 < . . . < tn = t} be a finite partition of the
interval [0, t]. We define the mesh size of the partition to be |Π| := max1≤j≤n (tj − tj−1 ).
For p > 0 we call
n
p
X
SpΠ (f ; t) := f (tj ) − f (tj−1 )
i=1

the p-variation sum for Π of a function f : [0, ∞) → R. The supremum over all finite
partitions is called the p-variation

VARp (f ; t) := sup SpΠ (f ; t) : Π finite partition of [0, t] ∈ [0, ∞] .


(4.4)

If VAR1 (f ; t) < ∞, f is said to be of bounded or finite (total) variation. For a given

sequence of finite partitions (Πn : n ≥ 0) of [0, t] with |Πn | → 0 as n → ∞ we denote by

varp (f ; t) := lim SpΠn (f ; t) if the limit exists . (4.5)

n→∞

Note that the limit may in principle depend on the choice of the partition.
38 CHAPTER 4. MORE PATH PROPERTIES OF BROWNIAN MOTION

Remark 4.2.2. Note that the total variation (4.4) is always defined, and if (4.5) exists
VARp (f ; t) ≥ varp (f ; t). If f is continuous it is no restriction to require t0 = 0 and tn = t,
and we can focus on partitions with ti ∈ Q or any other dense set for i < n. Indeed, for
all ϵ > 0 and every partition Π = (ti : P 0 ≤ i ≤ n) we can find Π′ = (qi : 0 ≤ i ≤ n) with
points from the dense subset such that ni=0 f (ti ) − f (qi ) < ϵ. Thus
′
SpΠ (f ; t) − SpΠ (f ; t) ≤ cp ϵ

with a constant cp depending only on p > 0. Since by definition

VARp (f ; t) := sup sup SpΠ (f ; t) : Π partition of [0, t] with #Π = n ,


n

it is enough to calculate VARp (f ; t) along rational points.

Theorem 4.2.3 (Quadratic variation of Brownian Motion). Let B(t) : t ≥ 0 be a
Brownian Motion and (Πn : n ≥ 0) a sequence of finite partitions of [0, t] with |Πn | → 0
as n → ∞. Then

2 
E S2Πn (B; t) − t

→ 0 as n → ∞ .

Remark 4.2.4. The 2-variation var2 (B; t) := limn→∞ S2Πn is called quadratic variation
of BM. The above theorem says that the sequence S2Πn (B; t) converges to t in L2 (hence
in probability), for any sequence of finite partitions with vanishing mesh.

Proof. Let Πn = {t0 = 0 < t1 < . . . < tn = t} be some partition of [0, t], then we have
n n
 X

2  X
E S2Πn (B; t) =
 
E B(tj ) − B(tj−1 = (tj − tj−1 ) = t .
j=1 j=1

Therefore,
n

2   X
2 
E S2Πn (B; t) − t = Var S2Πn (B; t) =
  
Var B(tj ) − B(tj−1 ) ,
j=1

2
where we used the independence of increments. Since B(tj ) − B(tj−1 ) ∼ B(tj − tj−1 )2
with mean tj − tj−1 , we get with the scaling property

2 
= (tj − tj−1 )2 E B(1)4 − 1 = (tj − tj−1 )2 (3 − 1) ,
  
Var B(tj ) − B(tj−1 )

where the last identity uses the value of the 4-th Gaussian moment (see problem sheet 1).
This implies
n
 Πn
2  X
E S2 (B; t) − t ≤ 2|Πn | (tj − tj−1 ) = 2|Πn |t → 0
j=1

as |Πn | → 0, which implies the result.

Corollary 4.2.5. Almost surely, Brownian motion paths are of infinite total variation.
In fact VARp (B; t) = ∞ almost surely for all p < 2.
4.2. VARIATION OF BROWNIAN PATHS 39

Proof. Let p = 2 − δ for some δ > 0 and Πn any sequence of finite partitions of [0, t] with
|Πn | → 0 as n → ∞. Then
X
2 δ
X 2−δ
B(tj ) − B(tj−1 ≤ max B(tj ) − B(tj−1 ) B(tj ) − B(tj−1
tj−1 ,tj ∈Πn
tj−1 ,tj ∈Πn tj−1 ,tj ∈Πn
δ
≤ max B(tj ) − B(tj−1 ) VAR2−δ (B; t) .
tj−1 ,tj ∈Πn

By Theorem 4.2.3 the left-hand side converges along a subsequence almost surely to t (a
sequence that converges in probability admits a subsequence that converges almost surely).
On the other hand,
δ
max B(tj ) − B(tj−1 ) →0 almost surely as n → ∞ ,
tj−1 ,tj ∈Πn

by uniform continuity of Brownian paths on [0, t]. This implies VAR2−δ (B; t) = ∞ almost
surely.

Under additional assumptions on the sequence of partitions one can establish almost
sure convergence for the quadratic variation (without passing to subsequences).


Theorem 4.2.6 (Almost sure convergence of quadratic variation). P∞Let B(t) : t ≥ 0 be
BM and (Πn : n ≥ 0) a sequence of finite partitions of [0, t] with n=1 |Πn | < ∞. Then

S2Πn (B; t) → var2 (B; t) = t almost surely as n → ∞ .

Proof. As in the proof of Theorem 4.2.3 we have

2 
E S2Πn (B; t) − t

≤ 2|Πn |t

Using Chebyshev’s inequality we get for all ϵ > 0

∞
X h ∞
i 2t X
P S2Πn (B; t) − t > ϵ ≤ 2 |Πn | < ∞ . (4.6)
ϵ
n=1 n=1

(n)
Let Aϵ = {ω : |S2Πn (B(ω); t) − t| > ϵ}. It follows from (4.6) and Borel-Cantelli lemma
that almost surely, only finitely many of the A(n)ε can kick in. Therefore, almost surely,
there exists Nϵ (ω): such that for any n > Nϵ (ω), we have that |S2Πn (B(ω); t) − t| ≤ ϵ. In
other words, P[Ωϵ ] = 1, where Ωϵ = {ω : lim supn→∞ |S2Πn (B(ω); t) − t| ≤ ϵ}.
The event of interest Ω0 := {ω : limn→∞ S2Πn (B(ω); t) = t} can be presented as the
limit Ω0 = ∩∞k=1 Ω1/k . Therefore,

∞
X
P[Ω0 ] = 1 − P[Ωc0 ] =1− P[∪∞ c
k=1 Ω1/k ] ≥1− P[Ωc1/k ] = 1.
k=1

Therefore, P[Ω0 ] = 1 and we are done.

Remark 4.2.7. A generic choice of Πn is to use nested partitions, i.e. Πn ⊆ Πn+1 ⊆ . . .,

e.g. with |Πn | = 2−n (dyadic nesting), which obviously satisfies the summability condition
above.
Using a martingale convergence result, it can also be shown that nestedness alone is
sufficient for almost sure convergence. Conversely, there is a sequence of finite partitions
with vanishing mesh along which the sequence S2Πn diverges (see [?], Section 1.3 for details).
40 CHAPTER 4. MORE PATH PROPERTIES OF BROWNIAN MOTION

4.3 Large-time behaviour and law of the iterated logarithm

At present, our best result concerning large time behavior of Brownian motion is a conse-
quence of time-inversion.

Corollary 4.3.1. We have that:

Bt
lim =0 almost surely for all α > 1/2 .
t→∞ tα

It is natural to ask what the asymptotic smallest envelope for BM is, namely what is
a deterministic function ψ : (0, ∞) → R such that lim supt→∞ B(t)/ψ(t) = 1? Such ψ, if
it exists, determines the almost sure asymptotic growth of a Brownian Motion.


Theorem 4.3.2 (Law of the iterated logarithm). For BM B(t) : t ≥ 0 we have, almost
surely,
B(t)
lim sup √ =1.
t→∞ 2t log log t
Remark 4.3.3. By symmetry of BM, it follows that, almost surely,

B(t)
lim inf √ = −1 .
t→∞ 2t log log t
Proof. We have to prove the following two almost sure statements:
p
∀ϵ > 0 ∃t0 > 0 : ∀t ≥ t0 B(t) ≤ (1 + ϵ) 2t log log t , (4.7)

and p
∀ϵ > 0 ∀t0 > 0 ∃t ≥ t0 : B(t) ≥ (1 − ϵ) 2t log log t . (4.8)
We will establish (4.7) and (4.8), using Borel
√ Cantelli Lemmas and scaling along a geo-
metric sequence of times q n . Define ψ(t) := 2t log log t.

Upper bound. Fix ϵ > 0 and q > 1, and define

n o
An := maxn B(t) ≥ (1 + ϵ)ψ(q n ) .
0≤t≤q

By Proposition 3.4.3, the maximum of BM up to time t has the same distribution as |B(t)|,
so we can write h |B(q n )| ψ(q n ) i
P[An ] = P √ n ≥ (1 + ϵ) √ n .
q q
n √ n
By scaling, |B(q )|/ q is standard Gaussian, and we can use the Gaussian tail estimate
2
from Lemma 4.3.4 (P[Z ⩾ x] ≤ x−1 e−x /2 ), to obtain
2 log log q n 2
P[An ] ≤ 2e−(1+ϵ) = 2(n log q)−(1+ϵ) .

This is summable in n, and the (first) Borel Cantelli Lemma implies that, almost surely,
only finitely many events An occur, i.e. P[An .i.o.] = 0. So there is Nω < ∞ a.s.: for all
n > Nω and t such that q n−1 ≤ t ≤ q n , we have

B(t) B(t) ψ(q n ) t qn

= ≤ (1 + ϵ)q ,
ψ(t) ψ(q n ) q n ψ(t) |{z}
t
| {z } | {z }
≤1+ϵ ≤1 ≤q
4.3. LARGE-TIME BEHAVIOUR AND LAW OF THE ITERATED LOGARITHM 41

√ √
since ψ(t)/t = 2 log log t/ t is decreasing in t. This implies
B(t)
≤ (1 + ϵ)q for all t > q Nω +1 ,
ψ(t)
and since this holds for all q > 1, (4.7) is proved.

Lower bound. In order to use the (second) Borel Cantelli Lemma in the other direction,
we have to construct a relevant sequence of independent events. Again, fix q > 1, and
define
Dn := B(q n ) − B(q n−1 ) ≥ ψ(q n − q n−1 ) .


From the tail estimate for a standard Gaussian Z we have a constant C > 0 such that
2
P[Z ≥ z] ≥ Ce−z /2 /z for large enough z. Using this estimate and monotonicity of log we
get
n n−1
h ψ(q n − q n−1 ) i Ce− log log(q −q )
P[Dn ] = P Z ≥ p ≥p
q n − q n−1 2 log log(q n − q n−1 )
Ce− log(n log q) C̃
≥p > √
2 log(n log q) n log n
P
for a further constant C̃ and all n large enough. Therefore n≥1 P[Dn ] = ∞ and so, by
the independence of Dn ’s and the (second) Borel Catelli Lemma we have, almost surely,
for infinitely many n ≥ 1

B(q n ) ≥ B(q n−1 ) + ψ(q n − q n−1 ) ≥ −2ψ(q n−1 ) + ψ(q n − q n−1 ) .

Here the second inequality follows from applying the upper bound established above to
−B(t), which implies in particular that B(q n ) ≥ −2ψ(q n ) for all n large enough. Therefore
we get, almost surely, for infinitely many n
B(q n ) −2ψ(q n−1 ) + ψ(q n − q n−1 ) −2 q n − q n−1 2 1
≥ ≥ √ + =1− √ − . (4.9)
ψ(q n ) ψ(q n ) q qn q q
To obtain the second inequality we used for the first term
√ n
ψ(q n−1 ) ψ(q n−1 ) q 1 1
n
= p n √ ≤√ ,
ψ(q ) q n−1 ψ(q ) q q
√
since ψ(t)/ t is increasing in t for large enough t. For the second term we have an
analogous argument using that ψ(t)/t is decreasing in t. Now, (4.9) implies that a. s.
there is a sequence of times tn → ∞
B(tn ) 2 1
≥1− √ − ≥1−ϵ ,
ψ(tn ) q q

for any ϵ > 0 and q large enough. The lower bound (4.8) is proved.

The proof of the result above relies on the following Gaussian tail estimate:
Lemma 4.3.4. Let X ∼ N (0, 1). Then there exist 0 < c1 ⩽ c2 for all x > 1
1 2 1 2
c1 e−x /2 ≤ P[X > x] ≤ c2 e−x /2 .
x x
42 CHAPTER 4. MORE PATH PROPERTIES OF BROWNIAN MOTION

Proof. We have Z ∞
1 y2
P(X > x) = √ e− 2 dy .
x 2π
Then Z ∞
2 /2 2 /2 2 /2−λ2 /2−λx
e−y dy ⩾ e−(x+λ) λ = e−x λ.
x
The lower bound follows by choosing λ = 1/x. For the upper bound
Z ∞ Z ∞
2 y −y2 /2 1 2
e−y /2 dy ⩽ e dy ⩽ e−x /2 .
x x x x

The proof is complete.

Corollary 4.3.1 implies that the growth of B(t) as t → ∞ is almost surely bounded by
tα for any α > 1/2, while the LIL implies that this is sharp and the lim sup growth of B(t)
is faster than tα for any α ≤ 1/2, which we summarize in a corollary.

Corollary 4.3.5. For any α ∈ (−∞, 1/2] we have

B(t) B(t)
almost surely, lim sup = +∞ and lim inf = −∞ .
t→∞ tα t→∞ tα

Proof. Since
B(t) 1/2−α B(t) p
= t √ log log t
tα t log log t
the statement for α = 1/2 implies the general case for α ≤ 1/2, and the latter follows
directly from the LIL and the fact that limn→∞ log log n = ∞.

In addition to the extreme case α = 1/2, the case α = 0 shows that BM paths them-
selves of course do not converge but keep taking values arbitrarily large in positive and
negative directions. Using the time inversion property we can use the LIL to understand
continuity properties of BM in typical time points, as opposed to the modulus of continu-
ity in Theorem 4.1.1 which concerns the corresponding lim sup on a compact time interval.

Corollary 4.3.6. Suppose B(t) : t ≥ 0 is a BM. Then for all t ≥ 0 we have, almost
surely,
|B(t + h) − B(t)|
lim sup p =1.
h↘0 2h log log(1/h)
The proof is left as an exercise.
Chapter 5

Brownian motion as a martingale

5.1 Basic properties of continuous martingales

Definition 5.1.1. A real-valued

stochastic process
 X(t) : t ≥ 0 is a martingale w.r.t.
a filtration F(t) : t ≥ 0 if it is adapted, E |X(t)| < ∞ for all t ≥ 0, and 0 ≤ s ≤ t
 
E X(t) F(s) = X(s) almost surely .

The process is called a submartingale if ≥ holds, and a supermartingale if ≤ holds in the

above formula.

Remark 5.1.2. We observe the following:

• Martingales are processes where the current state is always the best prediction for
the future, and their origin comes from modelling fair games. Martingales are not
necessarily Markov processes and vice versa.

• The martingale property with s = 0 includes in particular that E[X(t)] = E[E[X(t) |

F(0)]] = E[X(0)] for all t ≥ 0.

• As normal expectations, conditional expectations are linear, so if X and Y are two

martingales w.r.t. the same filtration, then any linear combination is a martingale.
This includes also the case where Y (t) ≡ y is a constant.

• If X is a submartingale, then −X is a supermartingale. If X is a martingale, then

in general |X| is not a martingale, but still a submartingale since
   
E |X(t)| F(s) ≥ E X(t) F(s) = X(s) .

We can make the following examples:

• Standard +



  BM B(t) : t ≥ 0 is a martingale w.r.t. F (t) : t ≥ 0 , since obviously
E |B(t)| < ∞ for any t ≥ 0 and for all 0 ≤ s ≤ t

E B(t) F + (s) = E B(t) − B(s) F + (s) + B(s)

   
 
= E B(t) − B(s) + B(s) = B(s) .

Here we use independence of increments and E[B(s)|F + (s)] = B(s) since B(s) is
F + (s)-measurable.

43
44 CHAPTER 5. BROWNIAN MOTION AS A MARTINGALE

• Exponential martingale. For
any θ ∈ R, t 7→ M θ (t) := exp θB(t) − tθ2 /2 is a

martingale w.r.t. F + (t) : t ≥ 0 , since

2
E M θ (t) F + (s) = M θ (s)e−θ (t−s)/2 E eθ(B(t)−B(s)) F + (s)
   
2
= M θ (s)e−θ (t−s)/2 E eθB(t−s) = M θ (s) ,
 

using independence
 of increments and exponential Gaussian moments, which also
imply E |M θ (t)| < ∞ for any t ≥ 0.


The next two results are useful basic facts about continuous martingales which have
almost surely continuous paths, such as Brownian Motion. The first result provides con-
ditions under which the defining property of martingales can be extended to stopping
times.


Theorem 5.1.3. Let X(t) : t ≥ 0 be a continuous

martingale, and 0 ≤ S ≤ T stopping
times,

all w.r.t. a filtration F(t) : t ≥ 0 . If the stopped process X(t ∧ T ) : t ≥
  by an integrable random variable X, i.e. X(t ∧ T ) ≤ X a.s. ∀t ≥
0 is dominated
0 and E |X| < ∞ , then
 
E X(T ) F(S) = X(S) almost surely .
Proof in a slightly simpler setting. We prove a simpler version where S = 0 and for sim-
plicity we will also assume that the process X itself is bounded, namely that
|X(t)| ≤ X, a.s. ∀t ≥ 0,
where X is an integrable random variable: E[X] < ∞.
First we establish that
 
E X(t ∧ T ) F(0) = X(0) for all t ≥ 0 , (5.1)
by using a discrete approximation of T for each n ∈ N,
T n = tnm := (m + 1)2−n if m2−n ≤ T < (m + 1)2−n for all m ≥ 0 .
These are also stopping times with T n ↘ T as n → ∞. Now fix n and t ≥ 0, and let
t̄n : min{tnm : tnm > t}. Then we can write
E X(t ∧ T n ) F(0) = E X(T n )1T n ≤t F(0) + E X(t)1Tn >t F(0)
     
X
E X(tnm )1T n =tnm F(0) + E X(t)1T n ≥t̄n F(0) .
   
=
m:tn
m <t̄
n

Using {T n = tnm } = {T n ≤ tnm } \ {T n ≤ tnm−1 }, we can set up a telescopic sequence for the
sum on the right-hand side
E X(tnm )1T n =tnm F(0) = E X(tnm )1T n ≤tnm F(0) − E X(tnm )1T n ≤tnm−1 F(0)
     

= E X(tnm )1T n ≤tnm F(0) − E X(tnm−1 )1T n ≤tnm−1 F(0) ,

   

where we used the tower property and E X(tnm ) F(tnm−1 ) = X(tnm−1 ) in the second line.
 

Since 1T n ≤0 (ω) ≡ 0, we are left with

E X(t ∧ T n ) F(0) = E X(t̄n − 2−n )1T n <t̄n F(0) + E X(t)1T n ≥t̄n F(0)
     

= E X(t) F(0) +E X(t̄n − 2−n ) − X(t) 1T n <t̄n F(0)

  

| {z }
=X(0)

= X0 ,
5.1. BASIC PROPERTIES OF CONTINUOUS MARTINGALES 45

by using the martingale property also to treat the last term. Now, since X has continuous
paths, X(t ∧ T n ) → X(t ∧ T ) almost surely. Next, since by assumption |X(t ∧ T n )| ≤ X,
by the dominated convergence theorem,

taking the n → ∞ limit leads to (5.1).
Finally, since X(t ∧ T ) : t ≥ 0 is also dominated by an integrable random variable
X we can use the dominated convergence theorem to take the limit t → ∞ to conclude
 
E X(T ) F(0) = X(0) . (5.2)
The proof of the full statement under the original set of assumptions can be found
in Moerters & Peres on pages 54 and 354 and is based on the theory of discrete-time
martingales.
Note that the condition on domination of the stopped process by an integrable random
variable could be replaced by requiring bounded stopping times, i.e. for some K > 0 we
have T ≤ K almost surely. This is particularly easy to see if E[X 2 (t)] < ∞ using Doob’s
maximal inequality, which we recall below.
Theorem 5.1.4 (Doob’s Lp maximal inequality). Suppose X(t) : t ≥ 0 is a continuous

martingale and let p > 1. Then for any t ≥ 0

h p   p p 
E |X(t)|p .

E sup X(s) ≤
0≤s≤t p−1
Proof in the Brownian case. The proof for general X involves discrete-time martingales
and can be found in Moerters & Peres pages 57 and 358. If X = B, a Brownian motion,
the reflection principle can be used to derive Doob’s inequality. First, let us notice that
 
sup |B(s)| = max sup B(s) , sup (−B(s)) .
0≤s≤t 0≤s≤t 0≤s≤t

Therefore
 p   p  p
E sup |B(s)| ⩽ E sup B(s) + E sup (−B(s))
0≤s≤t 0≤s≤t 0≤s≤t
 p
= 2E sup B(s) = 2E [|B(t)|p ] .
0≤s≤t
d
The last identity is due to sup0≤s≤t B(s) ∼ |B(t)| (reflection principle).
The next remark is so useful that we will state is as a lemma.


Lemma 5.1.5. Let X(t) : t ≥
0 be a continuous martingale and T a stopping time
w.r.t. a filtration F(t) : t ≥ 0 . Assume that E[X(t)2 ] < ∞ for all t ≥ 0. Then the
stopped process
t 7→ X(t ∧ T ) is a martingale .


Proof. Since T is a stopping time, t 7→ X(t ∧ T ) is adapted to F(t) : t ≥ 0 . As
E |X(t ∧ T )| ≤ E sup |X(s)| ≤ E[( sup X(s))2 ]1/2 ≤ 4E |X(t)|2 < ∞,
     
0≤s≤t 0≤s≤t

for all t ≥ 0, where we used Cauchy-Schwarz inequality at the second step and Doob’s
maximal inequality at the third. Therefore, the process X(t ∧ T ) is integrable.
For s < t we have
  
  
E X(t ∧ T ) F(s) =E X(t ∧ T ) 1T ≤s + 1T >s F(s) = E X(T )1T ≤s F(s)
 
+ E X(t ∧ T )1T >s F(s) = X(T )1T ≤s + X(s)1T >s = X(s ∧ T ) ,
where we have used optional stopping to simplify the second term. The use of optional
stopping is justified as the stopping time T ∧ t is bounded, see the remark above.
46 CHAPTER 5. BROWNIAN MOTION AS A MARTINGALE

5.2 Wald’s lemmas

We now use the martingale property and the optional stopping theorem to prove Wald’s
lemmas for Brownian Motion. These results identify the first and second moment of BM
at well behaved stopping times.

Lemma 5.2.1 (Wald’s first lemma). Let B(t) : t ≥ 0 be standard BM and T ⩾ S be
two stopping times such that either

(a) E[T ] < ∞ or

(b) B(t ∧ T ) : t ≥ 0 is dominated by an integrable random variable .

Then we have
 
E B(T )|FS = BS .

Furthermore (a) implies (b).

Proof. Under condition (b) we can directly apply the optional stopping theorem to get

E[B(T )|FS ] = BS .

It remains to show that (a) implies (b). Suppose E[T ] < ∞ and define

⌈T ⌉
X
Mk := max B(t + k) − B(k) and M = Mk .
0≤t≤1
k=0

Observe that {Mk }k∈N is a sequence of iid random variables. Then

⌈T ⌉
X  X∞ ∞ h 
 X i
E E 1T >k−1 Mk F + (k − 1) + E[M0 ]

E[M ] = E Mk = E 1T >k−1 Mk =
k=0 k=0 k=1
 ∞
X 


= E[M0 ] 1 + P[T > k] ≤ E[M0 ] 1 + E[T ] < ∞ ,
k=0

where we used the tower property, Mk ⊥ F(k − 1) and Mk ∼ M0 for all k ≥ 1 to get to the
penultimate inequality. The second equality is due to the monotone convergence theorem.
To show that E[M0 ] is finite, let’s use reflection principle:

E[M0 ] = E[ sup |B(t)|] ≤ E[ sup B(t) + sup (−B(t))] = 2E[|B(1)|] < ∞ .

0≤t≤1 0≤t≤1 0≤t≤1

P⌊T ∧t⌋
Finally, B(t ∧ T ) ≤ k=0 max0≤t≤1 B(k + t) − B(k) ≤ M , so that (b) holds.

Corollary 5.2.2. Let S ≤ T be stopping times and (Bt )t a Brownian motion with respect
to a common filtration (Ft )t⩾0 , and assume that E[T ] < ∞. Then

E B(T )2 = E B(S)2 + E (B(T ) − B(S))2 .

     

Let us observe that we do not claim that any of the quantities above are finite.
5.2. WALD’S LEMMAS 47

Proof. The tower property for conditional expectations implies

h i
E B(T )2 = E B(S)2 + 2E B(S)E B(T ) − B(S) FS + E (B(T ) − B(S))2
      
(5.3)

Note that using the tower property, E[T ] < ∞ implies E T −S F + (S) < ∞ almost surely.
 

Therefore Wald’s lemma implies

   
E B(T ) − B(S) FS = E B(T − S) = 0 almost surely ,

so that the middle term in (5.3) vanishes. Note that in order to use the tower property
(and hence Wald’s lemma) we must first make sure that

BS (BT − BS ) ∈ L1 ,

although proving that either the positive or the negative part lies in L1 would be sufficient.
To obtain the integrability we can follow the construction in the proof of Lemma 5.2.1,
and we would obtain two inependent positive random variables MS , MT such that

B S ⩽ MS , BT − BS ⩽ MT ,

and such that

E[MS ] + E[MT ] ≲ 1 + E[T ] .
Then the required bound follows (by using the independence of MS , MT ).


Lemma 5.2.3. Let B(t) : t ≥ 0 be a Brownian Motion. Then the process

t 7→ M2 (t) := B(t)2 − t is a martingale .

Here as usual we imply that the martingale property holds with respect to the filtration
F = (F + (t) : t ≥ 0).

Proof. The process is adapted to F + (t) : t ≥ 0 and E |B(t)2 − t| < 2t < ∞ for all

 

t ≥ 0. Furthermore,

E B(t)2 − t F + (s) = E (B(t) − B(s))2 F + (s) + 2E B(t)B(s) F + (s) − B(s)2 − t

     

= (t − s) + 2B(s)2 − B(s)2 − t = B(s)2 − s ,

which completes the proof.

Lemma 5.2.4 (Wald’s second lemma). Let B(t) : t ≥ 0 be standard BM and T a
stopping time such that E[T ] < ∞. Then

E B(T )2 = E[T ] .
 

Proof. Consider the quadratic martingale M2 (t) : t ≥ 0 and define the stopping time

Tn := inf t ≥ 0 : |B(t)| = n ,

so that M2 (t ∧ T ∧ Tn ) : t ≥ 0 is dominated by the integrable random variable n2 + T . By

optional stopping (Theorem 5.1.3) applied to the stopped quadratic martingale M2 (t ∧ Tn )

we get
E B(T ∧ Tn )2 = E[T ∧ Tn ] .
 
48 CHAPTER 5. BROWNIAN MOTION AS A MARTINGALE

Then, by Corollary 5.2.2 we have E B(T )2 ≥ E B(T ∧ Tn )2 since T ∧ Tn ≤ T . Hence,

   

by monotone convergence
E B(T )2 ≥ lim E B(T ∧ Tn )2 = lim E[T ∧ Tn ] = E[T ] ,
   
n→∞ n→∞

since Tn → ∞ and T ∧ Tn ↗ T as n → ∞ almost surely, using Corollary 3.5.3. On the

other hand, we have by Fatou’s lemma
E B(T )2 ≤ lim inf E B(T ∧ Tn )2 = lim inf E[T ∧ Tn ] = E[T ] .
   
n→∞ n→∞

Remark 5.2.5. We observe the following.

• The statements of Wald’s lemmas go wrong without some integrability conditions on
the stopping time. For example, the claim E[B(T )] = 0 for any  stopping time is
false. As a counter-example, consider the hitting time Ta := inf t ≥ 0 : B(t) = a
for any a ̸= 0. Then a = E[a] = E[B(Ta )] = 0 - a contradiction. The weakest
assumption on T under which the statement of the first Wald’s Lemma is true is
E[T 1/2 ] < ∞, see Thm. 2.50 of Moerters & Peres. Also, Wald lemmas are not
applicable to random times, which are not stopping times.
• For stopping times 0 ≤ S ≤ T with E[T ] < ∞ Wald’s lemmas, together with Corol-
lary 5.2.2 imply
E B(T ) = 0 and E B(S)B(T ) = E B(S)2 = E[S] ,
     

in analogy to the mean and covariance properties for BM as a Gaussian process.

A spectacular application of Wald’s lemmas is the computation of exit probabilities

and expected exit times of Brownian motion from bounded intervals containing the origin.


Corollary 5.2.6. Consider a < 0 < b and a Brownian  Motion B(t) : t ≥ 0 denote the
first exit time of (a, b) by T = T{a,b} := Ta ∧ Tb = inf t ≥ 0 : B(t) ∈ {a, b} . Then
    b
(a) P B(T ) = a = 1 − P B(T ) = b =
|a| + b
(b) E[T ] = |a| b .
Proof. Since B(t ∧ T ) ≤ |a| ∨ b Wald’s first Lemma 5.2.1 implies
     
0 = E B(T ) = aP B(T ) = a + bP B(T ) = b ,
   
which implies (a) with P B(T ) = a = 1 − P B(T ) = b . To use Wald’s second Lemma
we check that E[T ] < ∞, writing
Z ∞ Z ∞
 
E[T ] = P[T > t]dt = P B(s) ∈ (a, b) for all s ∈ [0, t] dt < ∞,
0 0

 
⌈t⌉
as the integrand P B(s) ∈ (a, b) for all s ∈ [0, t] ≤ maxx∈(a,b) Px [B(s) ∈ (a, b), ∀s ∈ [0, 1]] ,
which decays exponentially fast with time. (Use Markov property recursively to prove the
last estimate. ) Now, by Wald’s second Lemma 5.2.4 and (a) we obtain
b |a|
E[T ] = E B(T )2 = a2 + b2
 
= |a|b .
|a| + b |a| + b
5.3. FROM MARTINGALES TO HARMONIC FUNCTIONS 49

Remark 5.2.7. Using the exponential martingale M θ (t) : t ≥ 0 Wald’s lemmas can

also be used to infer the moment generating function for the law of exit times (this is on
a problem sheet).

5.3 From martingales to harmonic functions

In this section we want to construct further martingales related to Brownian motion. First
note that the density of the Gaussian transition kernel for BM

1 −(y−x)2 /(2t)
p(t, x, y) = √ e ,
2πt

which characterizes BM as a homogeneous Markov process according to Definition 3.1.3,

is the unique solution to the heat or diffusion equation:

∂ 1 ∂2 1
p(t, x, y) = p(t, x, y) = ∆y p(t, x, y) for all x ∈ R , (5.4)
∂t 2 ∂y 2 2

with initial condition p(0, x, dy) = δx (dy) concentrating in y = x. Using integration by

parts, we find for any twice continuously differentiable test function f ∈ C02 (R, R) with
compact support that
Z
d 
f (y) 12 ∆y p(t, x, y)dy = Ex 12 ∆f (B(t)) .
  
Ex f (B(t)) = (5.5)
dt R

This is also known as Dynkin’s formula or Kolmogorov forward equation, and the operator
1 2
2 ∆ on C0 (R, R) is also called the infinitesimal generator of Brownian motion.

More generally, in the following let (S, d) be a compact or locally compact, separable
metric space (e.g. S = Rd with L2 -metric), with Borel σ-algebra S induced by all open
sets w.r.t. the metric d. Denote by C(S, R) all continuous, real-valued functions on S that

f (xn ) → 0 for all sequences (xn : n ∈ N)
vanish at infinity in case S is not compact, i.e.
such that d(xn , 0) → ∞. Then C(S, R), ∥.∥∞ is a Banach space.

Proposition 5.3.1. Let X(t) : t ≥ 0 be a time-homogeneous Markov process with state
space S and transition kernel p(t, x, dy) as in Definition 3.1.3, and assume that for all
f ∈ C(S, R) also Pt f ∈ C(S, R), where
Z
 
(Pt f )(x) := f (y)p(t, x, dy) = Ex f (X(t)) for all t ≥ 0 .
S

The family of operators (Pt : t ≥ 0) forms a semi-group, i.e.

Pt+s = Pt Ps for all t, s ≥ 0 and P0 = I (identity map) .

If this is also continuous, i.e. ∥Pt f − f ∥∞ → 0 as t ↘ 0 for all f ∈ C(S, R), the process
is called a Feller Markov process. The infinitesimal generator L : D ⊆ C(S, R) → C(S, R)
of the process is then the operator

1

Lf (x) := lim Pt f (x) − f (x) with domain D ⊆ C(S, R) ,
t↘0 t
50 CHAPTER 5. BROWNIAN MOTION AS A MARTINGALE

such that the (point-wise) limit is well-defined.

For any f ∈ D, the Kolmogorov forward equation (or Dynkin’s formula) holds, i.e.
 
d d    
Pt f = Pt Lf Ex f (Xt ) = Ex Lf (Xt ) ,
dt dt
 
where (P0 f )(x) = Ex f (X0 ) = f (x) .
Proof idea. The proof consists of a rigorous verification of the following argument: (i)
derive the semi-group property from the Chapman-Kolmogorov equations for transition
kernels (3.1). (ii) Justify the following steps for any f ∈ D ⊆ C(S, R):
Pt+∆t − Pt P∆t − I
f = Pt f → Pt Lf ,
∆t ∆t
which implies the forward equation/Dynkin’s formula.

Remark 5.3.2. Brownian motion is a Feller process on R with generator L = 12 ∆ and

domain D = C02 (R, R). Note that this is not the maximal set for which L can be defined,
but it is sufficient to fully characterize Pt on C(S, R). See the problem sheets for the com-
putation of generators of other Feller Markov processes, such as the Ornstein-Uhlenbeck
process.
For more details including properties of Markov semi-groups and forward/backward
equations see e.g. Liggett, Chapter 3. Heuristically, the generator describes the expected
change of a test function f under the dynamics of the process, and fluctuations around
that are given by a martingale.


Theorem 5.3.3. Let X(t) : t ≥ 0 be a Feller process on S. Then it has generator L
with domain D if and only if Lf ∈ C(S, R) for all f ∈ D, and
Z t
f
t 7→ M (t) := f (X(t)) − Lf (X(s)) ds
0

is a martingale w.r.t. the natural filtration of the process.

For a proof and further references on the so-called Martingale problem see again
Liggett, Chapter 3. We will prove the forward implication for BM keeping the class
of test functions as general as possible.
Theorem 5.3.4. Let f ∈ C 2 (R, R) and B(t) : t ≥ 0 BM. Further suppose that for all

x ∈ R and t > 0 we have

Z t 
 
Ex f (B(t)) < ∞ and Ex ∆f (B(s))| ds < ∞ . (5.6)
0

Then the process

Z t
f 1
t 7→ M (t) := f (B(t)) − ∆f (B(s)) ds is a martingale .
2 0

Proof. For any 0 ≤ s < t

 1
Z s Z t−s
1
E M f (t) F(s) = EB(s) f (B(t − s)) −
    
∆f (B(u)) du − EB(s) ∆f (B(u)) du .
2 0 2 0
| {z }
:=(∗)
5.3. FROM MARTINGALES TO HARMONIC FUNCTIONS 51

Here we have used the Markov property and regularity assumptions to exchange expecta-
tion and integration. Now, using integration by parts and the diffusion equation (5.4) for
the transition kernel we get for the integrand of (∗)
Z
1   1
E ∆f (B(u)) = p(u, B(s), y)∆f (y) dy
2 B(s) 2 R
Z Z
1 ∂
= ∆y p(u, B(s), y)f (y) dy = p(u, B(s), y)f (y) dy .
2 R R ∂u

Using the regularity assumptions (5.6) to exchange integrals and the Markov property we
get
Z Z t−s
∂
(∗) = lim dy p(u, B(s), y)f (y)
ϵ↘0 R ϵ ∂u
Z Z
= dyp(t − s, B(s), y)f (y) − lim dyp(ϵ, B(s), y)f (y)
R ϵ↘0 R
 
= EB(s) f (B(t − s)) − f (B(s)) ,

which after cancellation of two terms implies E M f (t) F(s) = M f (s) as required.
 

Example 5.3.5. Using f (x) = x gives the obvious martingale M f (t) = B(t), and f (x) =
x2 the quadratic martingale M f (t) = B(t)2 − t. On a problem sheet we will see further
examples with monomials f (x) = xk .

Corollary 5.3.6. Suppose f ∈ C 2 (R, R) is harmonic, i.e. ∆f (x) ≡ 0. Then the process

t 7→ f (B(t)) is a martingale .

This follows immediately from Theorem 5.3.4. Harmonic functions in R are not too
exciting and simply linear of the form f (x) = a + bx, a, b ∈ R, the regularity assumptions
(5.6) are automatically fulfilled. But the same concept also applies in higher dimensions
where it can be applied in a useful way as we will see later.
Chapter 6

Brownian motion in higher

dimensions

In this chapter we will discuss connections between Brownian Motion and harmonic func-
tions and recurrence and transience properties of Brownian Motion depending on the
dimension d.

6.1 Harmonic functions and Brownian motion

Definition 6.1.1. Let U ⊆ Rd be a domain, i.e. an open and connected set, and denote
by ∂U its boundary. A function u : U → R is called harmonic (on U ) if u ∈ C 2 (U ) is
twice continuously differentiable on U and

d
X ∂2
∆u(x) := u(x) = 0 for all x ∈ U .
j=1
∂x2j

If instead we have ∆u(x) ≥ 0 for all x ∈ U , u is called sub-harmonic.

Remark 6.1.2.

• ∆u = ∇·(∇u), so u P is harmonic if and only if ∇u : U → Rd is a divergence-free field

on U (here ∇ · v = di=1 ∂xi vi for a vector-valued function v). This interpretation
is relevant in application areas of incompressible flows driven by a potential u, or
electrostatics where u can be interpreted as an electrical potential. Divergence-free
property means that there is no flow or charge source in the domain U and all flow
or field lines ∇u that enter U also have to exit.

• The latter conservation property is best appreciated using Green’s identity, which is
a special case of Stokes’ Theorem: For any (open) ball B(x, r) ⊆ U centered in x
with radius r, we have
Z Z
∂u
∆u(y) dy = (y) dσ(y) .
B(x,r) ∂B(x,r) ∂n

Here σ denotes the surface measure on the sphere ∂B(x, r) defined in the next bullet
∂u
point, and ∂n (y) = ∇u(y) · n(y) where n(y) is the unit outer normal vector (of
Euclidan length 1) at the point y ∈ ∂B(x, r).

52
6.1. HARMONIC FUNCTIONS AND BROWNIAN MOTION 53

• We will use generalized polar coordinates y − x = ρϕ with ϕ ∈ ∂B(0, 1) to evaluate

integrals in over balls B(x, r) ⊆ Rd . For measurable f : Rd → R we write
Z Z r Z Z r Z
d−1
f (y) dy = dρρ f (x + ρϕ)dϕ =: dρ f (y) dσ(y) ,
B(x,r) ρ=0 ∂B(0,1) 0 ∂B(x,ρ)

which defines the surface measure σ on ∂B(x, r).

• The volume (Lebesgue measure) of a ball B(0, r) ⊆ Rd is given by

π d/2
Z
dy = rd Vd



L B(0, r) = where Vd := L B(0, 1) = ,
B(0,r) Γ(d/2 + 1)

with the usual values V2 = π and V3 = 4π/3 for unit balls.

The surface area of the sphere ∂B(0, r) is then given by

2π d/2
σ(∂B(0, r)) = drd−1 Vd = rd−1 Sd−1 where Sd−1 = dVd = ,
Γ(d/2)

with the usual values S1 = 2π and S2 = 4π for unit spheres.

• Examples of harmonic functions include (for the last cases the domains do not con-
tain the origin)

u(x) = a · x + b with a ∈ Rd and b ∈ R for all d ≥ 1 ;

u(x) = log |x| for d = 2 ;
2−d
u(x) = |x| for d ≥ 3 .

Here and in the following we simply write |x| := ∥x∥2 for the Euclidean norm of
x ∈ Rd .

• By linearity of the Laplacian, linear combinations of harmonic functions on a domain

U are also harmonic on U .

As a first result we establish alternative characterizations of harmonic functions which

do not involve derivatives and are therefore often useful in practice.

Theorem 6.1.3 (Mean value property). Let U ⊆ Rd be a domain and u : U → R

measurable and locally bounded. Then the following are equivalent:

(a) u is harmonic.
Z
1
(b) For any ball B(x, r) ⊆ U we have u(x) = u(y) dy.
L(B(x, r)) B(x,r)
Z
1
(c) For any ball B(x, r) ⊆ U we have u(x) = u(y) dσ(y).
σ(∂B(x, r)) ∂B(x,r)

Sketch of proof. (b)⇒(c): Assume u : U → Rd satisfies (b) and fix x ∈ U . Then there
exists R > 0 such that B(x, R) ⊆ U and we define
Z
1−d
ψ(r) := r u(y) dσ(y)
∂B(x,r)
54 CHAPTER 6. BROWNIAN MOTION IN HIGHER DIMENSIONS

for all 0 < r ≤ R. Then we have

Z Z r
d
r Vd u(x) = u(y) dy = ψ(s) sd−1 ds .
B(x,r) 0

Differentiating w.r.t. r gives ψ(r) = dVd u(x) for almost all 0 < r ≤ R. Since drd−1 Vd =
σ(∂B(x, r)) this implies (c) for almost all 0 < r ≤ R.
It remains to show that u is differentiable so that (c) holds for all 0 ≤ r < R and we can
infer (a) as well. To Rsee this, pick a smooth function g : [0, ∞) → [0, ∞) with compact
support in [0, R) and [0,∞)d g(|x|) dx = 1. Multiply (c) by g(r) and integrate over r to get
Z R Z R Z
u(x) σ(∂B(x, r))g(r) dr = g(r) dr u(y) dσ(y)
|0 R
{z } |
0 ∂B(x,r)
{z }
= g(|x|) dx=1


Rd
R
= Rd
u(y) g |y−x| dy

to obtain Z


u(x) = u(y) g |y − x| dy .
Rd
With appropriately chosen g, u can be written as a convolution of itself with a smooth
function for all x ∈ U , which implies that u ∈ C ∞ (U ) (using partial integration on the
r.h.s. to compute derivatives).
(c)⇒(b): Multiply (c) with r and integrate over 0 < r < R to obtain (c) for any R > 0
such that B(x, R) ⊆ U .
(c)⇒(a): We established above that (c) implies u ∈ C ∞ (U ). Suppose that ∆u ̸= 0 in U ,
so by continuity of ∆u there exists a ball B(x, r) ⊆ U such that either ∆u(y) > 0 or < 0
for all y ∈ B(x, r). But with the notation from above (c) implies that
Z Z
′ 1−d ∂u
0 = ψ (r) = r (y) dσ(y) = ∆u(y) dy ̸= 0, (6.1)
∂B(x,r) ∂n B(x,r)

which is a contradiction. Here we have used Green’s identity for the second equality, and
the polar coordinate representation of the integral to compute ψ ′ (r).
(a)⇒(c): Suppose u is harmonic and B(x, R) ∈ U . Then (6.1) implies that ψ ′ (r) = 0 for
all r < R, so ψ is constant and given by

ψ(r) ≡ lim ψ(r) = u(x)Sd−1 ,

r↘0

which implies (c).

Remark 6.1.4. A function u ∈ C 2 (U ) is subharmonic if and only if

Z
1
u(x) ≤ u(y) dy for any ball B(x, r) ⊆ U . (6.2)
L(B(x, r)) B(x,r)
This can be shown in an analogous way to Theorem 6.1.3.
Theorem 6.1.5 (Maximum principle). Suppose u : Rd → R is subharmonic on a domain
U ⊆ Rd . Then
(a) If u attains its maximum in U , then u is constant.
(b) If u is continuous on U and U is bounded, then

max u(x) = max u(x) .

x∈U x∈∂U
6.1. HARMONIC FUNCTIONS AND BROWNIAN MOTION 55

If u is harmonic, the theorem applies to both u and −u, so that all extrema on bounded
domains is attained at the boundary. For d = 1 this is consistent with harmonic functions
being affine.

Proof. (a) Let M := maxx∈U u(x) and V = {x ∈ U : u(x) = M , which is non-empty

by assumption. V relatively closed in U (due to the continuity of u, any sequence in V
converging in U, converges to a point in V ). On the other hand, for any x ∈ V there exists
a ball B(x, r) ⊆ U , and by the mean value property (6.2)
Z
1
M = u(x) ≤ u(y) dy ≤ M .
L(B(x, r)) B(x,r)

Therefore equality holds and u(y) = M for almost all y ∈ B(x, r), and by continuity of u
this implies that B(x, r) ⊆ V . So V is open. V ∈ U is both relatively closed and open,
V ̸= ∅. Since U is connected, V = U which implies that u ≡ M is constant on U .
(b) Since u is continuous on U which is closed and bounded, u attains its maximum on
Ū . With (a), the maximum has to be attained on ∂U .

Corollary 6.1.6. Suppose u1 , u2 : Rd → R are harmonic on a bounded domain U ⊆ Rd

and continuous on U . If u1 = u2 on ∂U , then u1 = u2 on U .

Proof. We apply Theorem 6.1.5 to u1 − u2 to get

sup u1 (x) − u2 (x) = sup u1 (x) − u2 (x) = 0 ,
x∈Ū x∈∂U

so u1 (x) ≤ u2 (x) for all x ∈ U . The same works for u2 − u1 which is also harmonic and
vanishes on ∂U , and we get u1 (x) = u2 (x) for all x ∈ U .

6.1.1 Connection to Brownian Motion

Before we state the fundamental result of this
first section, we must introduce Brownian
motion in d dimensions. Let { B k (t) : t ≥ 0 }k∈{1,...,d} be a collection of pairwise inde-
pendent Brownian motions. Then we define d-dimensional Brownian motion to be the
stochastic process
t 7→ B(t) = B 1 (t), . . . , B d (t) .




Note that B(t) : t ≥ 0 is a Gaussian process with continuous paths and a homogeneous
Markov process with stationary, independent increments (see problem sheets). Further-
more path properties hold, including 0-1 laws for germ and tail events, athe martingale
property and Donsker’s invariance principle. The next result establishes then the connec-
tion between harmonic functions and Brownian motion.

Theorem 6.1.7. Let U ⊆ Rd be a domain, B(t) : t ≥ 0 a Brownian motion started in

x ∈ U and
τ = τ (∂U ) := min{t ≥ 0 : B(t) ∈ ∂U }
the first hitting time of the boundary. Let ϕ : ∂U → R be measurable, and such that the
function u : U → R with
 
u(x) := Ex ϕ(B(τ )) 1τ <∞ for all x ∈ U ,

is locally bounded. Then u is harmonic.

56 CHAPTER 6. BROWNIAN MOTION IN HIGHER DIMENSIONS


Proof. For fixed x ∈ U choose a ball B(x, r) ∈ U and let τ̃ := inf t > 0 : B(t) ̸∈ B(x, r) .
Then the tower property and strong Markov property for BM imply
h  i
u(x) = Ex Ex ϕ(B(τ )) 1τ <∞ F + (τ̃ ) = Ex u(B(τ̃ )) .
 

R
By symmetry of BM the last expectation is simply given by ∂B(x,r) u(y) dσ(y)/σ(∂B(x, r)).
This implies the mean value property in Theorem 6.1.3 (c) and hence u is harmonic.

6.2 The Dirichlet problem

We can build on the previous result to solve the so-called Dirichlet problem.

Definition 6.2.1. Let U ⊆ Rd be a domain and ϕ : ∂U → R a continuous function.

A continuous function u : U → R such that u ∈ C 2 (U ; R) is a solution to the Dirichlet
problem with boundary value ϕ, if
(
∆u(x) = 0 , ∀x ∈ U ,
(6.3)
u(x) = ϕ(x) , ∀x ∈ ∂U .

While uniqueness of the solution to the Dirichlet problem follows directly from the
maximum principle for harmonic functions in Corollary 6.1.6, to ensure existence some
regularity of the domain is needed to avoid over-specification at the boundary.

Definition 6.2.2. A domain U ⊆ Rd satisfies the Poincaré cone condition at x ∈ ∂U if

there exists a cone Cx (α) in x with opening angle α > 0, and h > 0 such that Cx (α) ∩
B(x, h) ⊆ U c .

Theorem 6.2.3. Let U ⊆ Rd be a bounded domain that satisfies the cone condition in
every point x ∈ ∂U , and let ϕ : ∂U → R be a continuous function. Let τ := τ (∂U ) =
inf{t ≥ 0 : B(t) ∈ ∂U }, which is an almost surely finite stopping time. Then u : U → R
given by
 
u(x) := Ex ϕ(B(τ )) for all x ∈ Ū (6.4)
is the unique solution to the Dirichlet problem (6.3).

Remark 6.2.4.

• A cone in direction d ∈ ∂B(0, 1) can be written as

n o
C0 (α) := t d + tan α y) : t ≥ 0, y · d = 0, ∥y∥2 ≤ 1 ,

but we will not use any explicit representation in the following. The only important
property will be that Cx (α) ∩ B(x, 1) has positive Lebesgue measure for any α > 0.

• The Dirichlet problem (6.3) is linear, so if u1 and u2 are solutions on a domain U

with boundary values ϕ1 and ϕ2 on ∂U , then au1 + bu2 is a solution with boundary
values aϕ1 + bϕ2 .

• Most domains that come to mind satisfy the cone condition, examples include open
balls or any other regular shape. A typical example that does not satisfy the cone
condition is given below.
6.2. THE DIRICHLET PROBLEM 57

Example 6.2.5. Consider the solution v of the Dirichlet problem on the unit ball B(0, 1) ∈
Rd with constant boundary data ϕ(x) ≡ ϕ ∈ R, which implies that v(x) ≡ ϕ.
Now let U = B(0, 1) \ {0} be the punctured ball, which does obviously not satisfy the
cone condition at 0 ∈ ∂U . In dimension d = 1, U is not a domain (not connected), and
the Dirichlet problem happens to have a solution on U which is the linear interpolation
between ϕ and the boundary value ϕ(0) on both intervals [−1, 0] and [0, 1].
In higher dimensions d ≥ 2, U is a domain, and in general the Dirichlet problem on U
does not have a solution. Since BM in higher dimensions almost surely does not hit points
(see later) we have τ (∂U ) = τ (∂B(0, 1)) and (6.4) implies that the solution on U agrees
with u(x) = v(x) = ϕ for all x ̸= 0.

In order to prove Theorem 6.2.3 we require a technical Lemma, which asserts that if
you start Brownian Motion infinitesimally close to the base of a cone, you hit that cone
with probability 1 before you exit a ball of size 1.

Lemma 6.2.6. Let 0 < α < 2π and C0 (α) ⊆ Rd be a cone based at the origin with opening
angle α, and  
a := sup Px τ (∂B(0, 1)) < τ (C0 (α)) .
x∈B(0,1/2)

Then a < 1 and, for any integer k ≥ 1 and h′ > 0, we have

Px τ (∂B(z, h′ )) < τ (Cz (α)) ≤ ak for all x, z ∈ Rd with |x − z| < 2−k h′ .

 

Proof. It is easy to see from a picture (but a bit harder to prove, so we omit it) that a
lower bound for 1 − a is given by the probability of the event that BM started at x in the
interior of a cone of opening angle α/2, hits ∂B(x, 3/2) before leaving the cone, which has
positive probability. This implies that a < 1.
If x ∈ B(0, 2−k ) then by the strong Markov property

 k−1
Y
Px τ (∂B(0, 2−k+i+1 )) < τ (C0 (α)) = ak ,
  
Px τ (∂B(0, 1)) < τ (C0 (α)) ≤ sup
−k+i )
i=0 x∈B(0,2

which follows from a nested argument with BM hitting spheres of increasing radius 2−k+i
and then restarting. Therefore, for any integer k ≥ 1 and h′ > 0, scaling of BM and
translation to z ∈ Rd implies the statement.

Proof of Theorem 6.2.3. Uniqueness of the solution follows directly from Corollary 6.1.6.
By Theorem 6.1.7 u is bounded and harmonic on U , and it remains to show that the
Poincaré cone condition implies that u is continuous on the boundary ∂U . For all x ∈ Ū
and z ∈ ∂U we have
  h i
u(x) − u(z) = Ex ϕ(B(τ (∂U ))) − ϕ(z) ≤ Ex ϕ(B(τ (∂U ))) − ϕ(z) . (6.5)

To bound this from above, fix z ∈ ∂U , then there is a cone Cz (α) with Cz (α)∩B(z, h) ⊆ U c .
By Lemma 6.2.6, for any integer k ≥ 1 and h′ > 0, we have

Px τ (∂B(z, h′ )) < τ (Cz (α)) ≤ ak

 

for all x ∈ Rd with |x − z| < 2−k h′ . Also, since ϕ is continuous on ∂U , for given ϵ > 0
there exists 0 ≤ δ < h such that

ϕ(y) − ϕ(z) < ϵ for all y ∈ ∂U with |y − z| < δ .

58 CHAPTER 6. BROWNIAN MOTION IN HIGHER DIMENSIONS

Now if the BM started in x hits the cone Cz (α), which

is outside of U , before the sphere
∂B(z, δ), then z − B(τ (∂U )) < δ and ϕ B(τ (∂U )) is close to ϕ(z). The complement has
small probability by Lemma 6.2.6, and this leads to the following upper bound for (6.5),
   
u(x) − u(z) ≤ 2∥ϕ∥∞ Px τ (∂B(z, δ)) < τ (Cz (α)) + ϵPx τ (∂U ) < τ (∂B(z, δ))
≤ 2∥ϕ∥∞ ak + ϵ .

Since ϵ > 0 and k ≥ 1 can be chosen arbitrarily, this implies continuity of u on Ū , which
finishes the proof.

As an application we prove a classical result from harmonic analysis, Liouville’s theo-

rem, by probabilistic means.
Theorem 6.2.7 (Liouville’s theorem). Any bounded harmonic function on Rd is constant.
Proof. Let u : Rd → [−M, M ] be a harmonic function, x ̸= y two points in Rd and H the

hyperplane orthogonal to y − x so that reflection on H takes x into y. Let
B(t) : t ≥ 0
be BM started in x and τ (H) := min{t ≥ 0 : B(t) ∈ H}. Let B(t) : t ≥ 0 by the process
reflected in H which is started in y. By the reflection principle (reflecting the component
of B in the direction perpendicular to H) we then have that



B(t) : t ≥ τ (H) ∼ B(t) : t ≥ τ (H) . (6.6)

Recall from Corollary 5.3.6 that harmonicity of u implies that u(B(t)) is a martingale, so
that in particular
     
u(x) = Ex u(B(t)) = Ex u(B(t))1t<τ (H) + Ex u(B(t))1t≥τ (h) .

The same relationship holds for u(y) when B(t) is replaced by B(t) and with (6.6)
     
u(x) − u(y) = Ex u(B(t))1t<τ (H) − Ey u(B(t))1t<τ (h) ≤ 2M Px t < τ (H) → 0

as t → ∞. This holds since by isotropy (see problem sheet) the projection of d-dimensional
BM along any direction is one-dimensional BM, so that with Theorem 3.5.1 the hitting
times of hyperplanes are finite almost surely. Thus u(x) = u(y) and since x and y were
chosen arbitrarily, u must be constant.

6.3 Recurrence and transience of BM

In order to discuss recurrence and transience for Brownian motion in Rd we will make
crucial use of the solution of the following radially symmetric Dirichlet problem. Define
the annulus
A := x ∈ Rd : r < |x| < R

for 0 < r < R < ∞ . (6.7)
This is a bounded domain that satifies the cone condition in Definition 6.2.2, so that by
Theorem 6.2.3 the Dirichlet problem

 ∆u(x) = 0 , x ∈ A
u(x) = ϕ(r) , |x| = r (6.8)
u(x) = ϕ(R) , |x| = R


has a unique solution of the form

 

u(x) = Ex ϕ(B(τ )) = ϕ(r)Px [τr < τR ] + ϕ(R) 1 − Px [τr < τR ] .
6.3. RECURRENCE AND TRANSIENCE OF BM 59

Here B(t) : t ≥ 0 is BM in Rd , τr = inf t > 0 : |B(t)| = r is the hitting time of



the inner boundary of A with τR analogously for the outer boundary, and τ = τr ∧ τR is
hitting time of ∂A; ϕ(r) and ϕ(R) are constants, which determine spherically symmetric
boundary condtions on the inner and outer boundaries of A correspondingly. We know
from Theorem 3.5.1 that Brownian motion in R almost surely hits every point in finite
time, so considering only the first coordinate of B(t) with x ∈ A we have τr , τR < ∞
almost surely, leading to the above representation of u(x). Hence, we obtain

ϕ(R) − u(x)
Px [τr < τR ] = for all x ∈ Ā ,
ϕ(R) − ϕ(r)

so that by solving the Dirichlet problem (6.8) we can get the following result.

Theorem 6.3.1. Let B(t) : t ≥ 0 be a Brownian motion in Rd started in x ∈ A as

given in (6.7). Then

 R−|x|

 R−r , if d = 1
log(R/|x|)
Px [τr < τR ] = log(R/r) , if d = 2 . (6.9)
R2−d −|x|2−d

, if d ≥ 3

R2−d −r2−d

Letting R → ∞ using DCT this directly implies the following result.

Corollary 6.3.2. Suppose B(t) : t ≥ 0 is BM in Rd started in x ̸∈ B(0, r). Then

(
1 , if d ≤ 2
Px [τr < ∞] = rd−2 .
|x|d−2
, if d ≥ 3

Proof. Proof of Theorem 6.3.1 We simply have to solve the Dirichlet problem

(6.8), and
with rotation invariant boundary data we make the ansatz u(x) = ψ |x|2 for some C 2
function ψ : [0, ∞) → R. Then we get

∆u(x) = 4|x|2 ψ ′′ |x|2 + 2dψ ′ |x|2 = 0 ,

and with y = |x|2 > 0 this can be written as ψ ′′ (y) = −dψ ′ (y)/(2y), which in turn
reduces to
d 2
∂y log ψ ′ (y) = − ⇐⇒ ∂y log [(ψ ′ )− d (y)] = ∂y log (y)
2y
This is solved by ψ ′ (y) = const. y −d/2 leading to the radially symmetric harmonic functions

 |x| , if d = 1
u(x) = C1 + C2 log |x| , if d = 2 ,
|x|2−d , if d ≥ 3


as we have already seen earlier. Fixing the constants to satisfy the boundary conditions
leads to (6.9).

Definition 6.3.3. A Markov process X(t) : t ≥ 0 on Rd is called

• Point recurrent, if, almost surely, for all x ∈ Rd there exists a (random) sequence
Tn ↗ ∞ such that X(Tn ) = x for all n ∈ N,
60 CHAPTER 6. BROWNIAN MOTION IN HIGHER DIMENSIONS

• Neighbourhood recurrent, if, almost surely, for all x ∈ Rd and ϵ > 0 there exists a
(random) sequence Tn ↗ ∞ such that X(Tn ) ∈ B(x, ϵ) for all n ∈ N,

• Transient, if, almost surely, |X(t)| → ∞ as t → ∞.

Note that the events in the above definition are tail events, so by the 0-1 law for
BM (which also holds in Rd ) it follows that BM is either transient or recurrent. Point
recurrence trivially implies neighbourhood recurrence.

Theorem 6.3.4. Brownian motion in Rd is point recurrent for d = 1, neighborhood

recurrent for d = 2 and transient for d ≥ 3.

Proof. We treat every dimension separately.

d = 1: Theorem 3.5.1 guarantees that for all x, y ∈ R, Py [τx < ∞] = 1 i.e. BM hits all
points in finite time. So for x ∈ R we can define

T1 := τx , Tn+1 := inf{t > Tn + 1 : B(t) = x} for n ≥ 1 .

Then Tn ↗ ∞ and B(Tn ) = x for all n ∈ N, so Brownian motion is point recurrent.

d = 2: Fix x ∈ R2 and ϵ > 0. By Corollary 6.3.2 and shift invariance T1 := τ B(x, ϵ) <
∞ almost surely, and the same applies to

Tn+1 := inf t > Tn + 1 : B(t) ∈ B(x, ϵ) for all n ≥ 1 .

Again we have a sequence Tn ↗ ∞ as required and BM is neighbourhood recurrent. In

both cases d = 1 and d = 2 above results for arbitrary (fixed) points x ∈ Rd and ϵ > 0 can
be extended to almost sure results for all x ∈ Rd by taking intersections over a countable
dense set of points in Rd and using continuity
 of paths. 
d = 3: We consider the events An := |B(t)| > n for all t ≥ τn3 where τn3 := inf t >
0 : |B(t)| = n3 . For |B(0)| < n3 we have τn3 < ∞ by considering only the first coordinate
of Brownian motion and path continuity. So for given x ∈ Rd and any n > |x|1/3 we have
by the strong Markov property and Corollary 6.3.2
  n d−2
Px [Acn ] = Ex PB(τn3 ) [τn < ∞] = = n4−2d ,

n3
which is summable in n so the Borel-Cantelli Lemma implies that

P[Acn i.o.] = 0 .

That means that almost surely there exists N ∈ N such that for all n ≥ N , An occurs
which implies |B(t)| → ∞ as t → ∞.

Remark 6.3.5.

• For the range of BM

B[0, ∞) := B(t) : t ≥ 0 ⊆ Rd


Theorem 6.3.4 implies that B[0, ∞) = R for d = 1.

For d = 2 we have B[0, ∞) = R2 i.e. the range is dense in R2 . It can be shown that
for d = 2 BM is not point recurrent, and for any dimension d ≥ 2 the range is a
strict subset B[0, ∞) ⊊ Rd with vanishing Lebesgue measure.

• It can be shown that the range of BM is a (random) fractal with Hausdorff dimension
2 in any space dimension d ≥ 3.
6.3. RECURRENCE AND TRANSIENCE OF BM 61

• In analogy to BM, the SSRW on Zd is transient for d ≥ 3 and recurrent for

d = 1, 2 where the difference between point and neighbourhood recurrence does not
arise. There is an analogous method of proof using discrete harmonic functions,
which, however, do not take such a simple form as for BM. An alternative is to use
Donsker’s invariance principle (Theorem 2.1.1) in higher dimensions, which implies
that SSRW is transient (recurrent) if and only if BM is transient (recurrent).
in R2 . Then,


Theorem 6.3.6 (Lévy 1940). Let B(t) : t ≥ 0 be a Brownian motion 
almost surely, the Lebesgue measure (area) of the range vanishes, L2 B[0, ∞) = 0.
 
Proof. Denote by X := L2 B[0, 1] the area of B[0, 1]. By σ-additivity of the Lebesgue
measure it is enough to show that X = 0 almost surely. First, we check that E[X] <
∞. Note that X > a implies that for at least one coordinate B i of BM we have
√
maxt∈[0,1] B i (t) > a/2. So with Proposition 3.4.3 and a Gaussian tail bound we get for
all a > 1
√ √
P[X > a] ≤ 2P max |W (t)| > a/2 = 4P |W (t)| > a/2 ≤ 4e−a/8 ,
   
t∈[0,1]


where W (t) : t ≥ 0 is a one-dimensional Brownian motion. Hence
Z ∞ Z ∞
E[X] = P[X > a] da ≤ 1 + 4 e−a/8 da < ∞ .
0 1
√
By scaling applied to each coordinate, B(3t) and 3B(t) have the same distribution, and
hence    
E L2 [B[0, 3]] = 3E L2 [B[0, 1]] = 3E[X] . (6.10)
In general we have of course L2 [B[0, 3]] ≤ L2 [B[0, 1]]+L2 [B[1, 2]]+L2 [B[2, 3]] with equality
if and only if intersections have vanishing Lebesgue measure, i.e.
 
L2 B[i, i + 1] ∩ B[j, j + 1] = 0 for all 0 ≤ i < j ≤ 2 .
 
Since by homogeneity of BM E L2 [B[i, i + 1]] = E[X], (6.10) therefore implies that the
intersection
 of any two of the B[i, i + 1] has measure zero. In particular, L2 B[0, 1] ∩
B[2, 3] = 0 almost surely.

Using the Markov property we can
define two independent BMs B1 (t) : t ∈ [0, 1] with
B1 (t) = B(t) and B2 (t) : t ∈ [0, 1] with B2 (t) = B(t + 2) − B(2) + B(1), and the variable
Y := B(2) − B(1) is independent of both BMs. For x ∈ R2 we define
 
R(x) := L2 B1 [0, 1] ∩ (x + B2 [0, 1]) ,

and note that R(x) is independent of Y for all x. Therefore we can compute
Z
h  i 1 2
e−|x| /2 E R(x) dx ,
   
0 = E L2 B[0, 1] ∩ B[2, 3] = E R(Y ) =
2π R2
where we integrate w.r.t. the Gaussian distribution of B(2) − B(1). Therefore, for L2 -
almost all x ∈ R2 we have R(x) = 0 almost surely, and therefore by Fubini’s theorem

L {x ∈ R2 : R(x) > 0} = 0 almost surely .

 

One can show (see Lemma 2.23 in Moerters & Peres) that  this property
 of translation

for general Borel sets implies that almost surely, either L2 B[0, 1] = 0 or L2 B[2, 3] =
0. Since both
 are identically distributed and independent implies that both vanish and
L2 B[0, 1] = 0 almost surely.
62 CHAPTER 6. BROWNIAN MOTION IN HIGHER DIMENSIONS

be Brownian motion in Rd . For any two points

Corollary 6.3.7. Let B(t) : t ≥ 0
d

x, y ∈ R , d ≥ 2 we have Px y ∈ B(0, ∞) = 0.
Proof. By projection onto the first two coordinates and σ-additivity of Px , it suffices to
prove Px y ∈ B[0, 1] = 0 for d = 2. Note that Theorem 6.3.6 holds for arbitrary starting
points y ∈ R2 . By Fubini’s theorem for any fixed y ∈ R2
Z
   
Py x ∈ B[0, 1] dx = Ey L2 [B[0, 1]] = 0 .
R2

So for L2 -almost all x ∈ R2 we have

   
0 = Py x ∈ B[0, 1] = Px y ∈ B[0, 1] ,

where the second identity follows from symmetry of BM. It remains to show that by
excluding time t = 0 from therange this in fact holds for all x ∈ R2 . For any ϵ > 0 we
thus have, almost surely, PB(ϵ) y ∈ B[0, 1] = 0. Hence, for any x ∈ R2
      
Px y ∈ B[0, 1] = lim Px y ∈ B[ϵ, 1] = lim Ex PB(ϵ) y ∈ B[0, 1 − ϵ] = 0 ,
ϵ↘0 ϵ↘0

where we have used the Markov property in the second step.

This result immediately implies that BM in 2 dimensions cannot be point recurrent,

since all hitting times of particular points are infinity almost surely. We formulate this as
our last result.

Corollary 6.3.8. Brownian motion in R2 is not point recurrent.

6.3.1 Occupation measures

Instead of wondering whether Brownian motion reaches a fixed point, we can also ask
ourselves how much time Brownian motion passes in a given region. This leads to the
study of occupation measures.
Theorem 6.3.9. Let (Bt )t⩾0 be 1D Brownian motion and define the (random) measure
Z t
µt (A) = 1A (Bs ) ds , ∀A ∈ B(R) .
0

Then, almost surely, for any t > 0 the measure µt is absolutely continuous with respect to
the Lebesgue measure.
Proof. To prove absolute continuity it suffices to prove that P−almost surely the following
holds:
µt (Bx,ε )
lim sup < ∞ , for µt − almost all x ∈ R .
ε→0 L1 (Bx,ε )
Therefore it suffices to prove that
Z 
µt (Bx,ε )
E lim sup dµt (x) < ∞ .
R ε→0 L1 (Bx,ε )

By Fatou’s lemma the above is bounded by

Z 
1
lim sup E µt (Bx,ε ) dµt (x) .
ε→0 2ε R
6.3. RECURRENCE AND TRANSIENCE OF BM 63

Now, the term inside the lim sup is given by

Z Z t  Z t Z t 
E 1Bx,ε (Bs ) ds dµt (x) = E 1B0,ε (Bs − Br ) ds dr
R 0 0 0
Z tZ t
= P(|Bs − Br | ⩽ ε) ds dr
0 0
Z tZ t
ε
⩽ p ds dr ,
0 0 |r − t|

where in the last step we used scale invariance of BM together with the fact that the PDF
of a standard Gaussian is bounded from above by 1. Since
Z tZ t
1
p dr ds < ∞ , ∀t > 0 ,
0 0 |r − s|

the result follows.

Now we pass to study occupation measures in higher dimension. In particular, we

concentrate on dimensions d ⩾ 3 and consider, instead of the original (random) occupation
measure, the expected occupation time measure

νx,t (A) = Ex (µt (A)) .

We observe that
Z t Z Z t
νx,t (A) = Px (Bs ∈ A) ds = ps (x, y) ds dy ,
0 A 0

so that the expected occupation time measure admits an explicit density with respect to
the Lebesgue measure:
Z t
dνx,t
(y) = ps (x, y) ds .
dL1 0

Depending on the dimenision, we see different behaviors of the expected occupation mea-
sure as t → ∞.

Theorem 6.3.10. If d ⩽ 2, then

lim µt (A) = +∞ , P − almost surely .

t→∞

Instead in d ⩾ 3 we have Z
lim νx,t (A) = g(x, y) dy ,
t→∞ A

with
Γ(d/2 − 1)
g(x, y) = d |x − y|2−d .
2π 2

The proof of this result is left as an exercise in a problem sheet. Next, we show that
the kernel g appearing in Theorem 6.3.10 is harmonic outside the diagonal x = y.

Theorem 6.3.11. Id d ⩾ 3, then for any y ∈ Rd , the function g(·, y) (defined as in

Theorem 6.3.10) is harmonic on Rd \ {y}.
64 CHAPTER 6. BROWNIAN MOTION IN HIGHER DIMENSIONS

Proof. Let us fix a point x ̸= y and a ball Bx,ε ⊆ Rd \ {y}. Then it suffices to check the
mean value property: Z
1
g(x, y) = g(z, y) dz .
|Bx,ε | Bx,ε
Indeed, if we denote with

τ = inf{t > 0 : |Bt − x| = ε} ,

we find that
Z ∞  Z Z ∞
1
g(x, y) = Ex δy (Bt ) dt = Ez δy (Bt ) dt dz
τ |Bx,ε | Bx,ε 0
Z
1
= g(z, y) dz ,
|Bx,ε | Bx,ε

where we used the strong Markov property and rotational symmetry.

For completeness, we observe that one can prove that g is a fundamental solution to
the Poisson problem on Rd , for d ⩾ 3:
1
∆x g(x, y) + δy (x) = 0 , ∀x ∈ Rd .
2

6.4 The Feynman-Kac formula

As we have seen in (5.4) in one dimension, the Gaussian transition kernel p(t, x, y) =
2
√ 1 e−(y−x) /(2t) of BM solves the heat equation
2πt

∂
p(t, x, y) = ∆y p(t, x, y) with initial condition p(0, x, dy) = δx (dy) .
∂t
The same is true in higher space dimensions d > 1 (see problem sheet 4), and since for
the Gaussian transition kernel ∆y p(t, x, y) = ∆x p(t, x, y) an analogous backward heat
equation holds
∂ 1
p(t, x, y) = ∆x p(t, x, y) .
∂t 2
Multiplying this equation with f (y) and integrating over y implies that
Z
 
u(t, x) := Ex f (B(t)) = p(t, x, y)f (y) dy
Rd
R
solves the heat equation with initial data u(0, x) = Rd f (y) δx (dy) = f (x). This can be
extended to more general versions of the heat equation with a dissipation term, as is made
precise in the following.

Definition 6.4.1. A twice differentiable function u : (0, ∞) × Rd → [0, ∞) is said to solve

the heat equation with dissipation rate V : Rd → R and initial condition f : Rd → [0, ∞)
if we have

(a) for all x0 ∈ Rd : u(t, x) → f (x0 ) as x → x0 and t ↘ 0,

(b) for all (t, x) ∈ (0, ∞) × Rd : ∂t u(t, x) = 12 ∆x u(t, x) + V (x)u(t, x) .

6.4. THE FEYNMAN-KAC FORMULA 65

Theorem 6.4.2 (Feynman–Kac
formula). Suppose that V : Rd → R and f : Rd → [0, ∞)

are bounded, and B(t) : t ≥ 0 is a BM on R . Then u : [0, ∞) × Rd → [0, ∞) defined by
d

h Z t i
u(t, x) := Ex f (B(t)) exp V (B(s)) ds (6.11)
0

solves the heat equation on Rd with dissipation rate V and initial condition f .

Proof. For V ≡ 0 this corresponds to the backward heat equation (5.4) discussed above,
and the easiest proof is a direct calculation. P Expanding the exponential in (6.11) in a
power series we get for the terms u(t, x) = ∞ n=0 an (t, x), a0 (t, x) = Ex [f (B(t))] and for
n≥1
Z t Z t
1 h i
an (t, x) := Ex f (B(t)) ··· V (B(t1 )) · · · V (B(tn )) dt1 . . . dtn
n! 0 0
h Z t Z t Z t i
= Ex f (B(t)) dt1 dt2 · · · dtn V (B(t1 )) · · · V (B(tn ))
0 t1 tn−1
Z Z Z t Z t n
Y n
Y
= dx1 · · · dxn dt1 · · · dtn V (xi ) p(ti − ti−1 , xi−1 , xi )
Rd Rd 0 tn−1 i=1 i=1
Z
dy p(t − tn , xn , y)f (y)
Rd
Z Z t
= dx1 V (x1 ) dt1 p(t1 , x, x1 )an−1 (t − t1 , x1 )
Rd 0

with the convention x0 = x and t0 = 0. Using the backward heat equation 12 ∆x p(t1 , x, x1 ) =
∂t1 p(t1 , x, x1 ) and integrating by parts we get
Z Z t
1
∆x an (t, x) = dx1 V (x1 ) dt1 ∂t1 p(t1 , x, x1 )an−1 (t − t1 , x1 )
2 Rd 0
Z Z t
=− dx1 V (x1 ) dt1 p(t1 , x, x1 )∂t1 an−1 (t − t1 , x1 ) − V (x)an−1 (t, x)
Rd 0
Z
+ V (x1 )p(t, x, x1 )an−1 (0, x1 ) dx1
R
= ∂t an (t, x) − V (x)an−1 (t, x) .
P
Since f and V are bounded the sum n an (t, x) is absolutely convergent for all t ≥ 0 and
x ∈ Rd and we can exchange differentiation with the summation. Summing up all terms
verifies the validity of the heat equation with dissipation for u(t, x). The initial condition
is of course  
u(0, x) = Ex f (B(0)) = f (x) .
Chapter 7

Hausdorff dimension of the range

of Brownian motion

The Hausdorff dimension and measure of a set are intrinsic notions for a metric space
(E, ϱ). For a set F ⊆ E we define its diameter to be
|F | = sup{ϱ(x, y) : x, y ∈ F } ,
and we denote with {Ei }i∈N a covering of E if
∞
[
E⊆ Ei .
i=1

Definition 7.0.1. For any α ⩾ 0 and any metric space E we define the α–Hausdorff
content of E as
(∞ )
X
α α
H∞ (E) = inf |Ei | : {Ei }i∈N is a covering of E .
i=1

Furthermore we define the Hausdorff dimension

α α
dim(E) = inf {α ⩾ 0 : H∞ (E) = 0} = sup {α ⩾ 0 : H∞ (E) > 0} .
Remark 7.0.2. If E ⊆ Rd , then dim(E) ⩽ d.
Proof. Consider an α > d and suppose that E is bounded. Then it suffices to prove that
d–dimensional cubes Qr = [−r, r]d satisfy
α
H∞ (Qr ) = 0 ∀r ∈ N .
We can rewrite
Qr ⊆ ∪x∈2−n Z∩Qr Qn,x ,
with
d
Y
Qn,x = [xi , xi + 2−n ] .
i=1
Then since
|2−n Z ∩ Qr | ⩽ (2n 2r)d
we obtain
α
H∞ (Qr ) ⩽ C(d)2nd 2−nα → 0 , for n → ∞ ,
since by assumption α > d.

66
 67

We can also construct a Hausdorff measure as follows:

Definition 7.0.3. For any α ⩾ 0 and δ > 0 we define

(∞ )
X
α α
Hδ (E) = inf |Ei | : {Ei }i∈N is a covering of E and |Ei | ⩽ δ .
i=1

Then we define
Hα (E) = sup Hδα (E) = lim Hδα (E) .
δ>0 δ↓0

Remark 7.0.4. The Hausdorff measure satisfies the following two properties:

1. Monotonicity: Hα (E) ⩽ Hα (D), for all D ⊆ E.

2. Subadditivity (for any sequence of sets Ei ):

!
[ X
Hα Ei ⩽ Hα (Ei ) .
i∈N i∈N

Proof. We only prove the second claim (the first one is obvious). Fix any ε ∈ (0, 1) and
choose δ(ε) ∈ (0, 1) such that
! !
[ [
α α
H Ei ⩽ Hδ Ei + ε .
i∈N i∈N

Then for any i ∈ N let {Fik }k∈N be a δ-covering of Ei such that

X ε
|Fik | ⩽ Hδα (Ei ) + .
2i
k∈N

Then by construction {Fik }i,k∈N is a covering of i Ei , and we find

S

! !
[ [
α α
H Ei ⩽ ε + Hδ Ei
i∈N i∈N
X X ε
⩽ε+ |Fik | ⩽ Hδα (Ei ) +
2i−1
i,k i
( )
X
⩽ Hα (Ei ) + 2ε ,
i

so that the result follows since ε is arbitrary.

Next, we can connect the Hausdorff measure to the Hausdorff dimension through the
following result.

Proposition 7.0.5. The following hold true for any α ⩾ 0:

1. Hα (E) = 0 ⇐⇒ H∞
α (E) = 0.

2. dim(E) = inf{α : Hα (E) = 0} = sup{α : Hα (E) > 0}.

3. dim(E) = inf{α : Hα (E) < ∞} = sup{α : Hα (E) = ∞}.

68CHAPTER 7. HAUSDORFF DIMENSION OF THE RANGE OF BROWNIAN MOTION

Proof. We prove the results in order.

α (E) ⩽ Hα (E) we must only check the implication
1. Since H∞

α
H∞ (E) = 0 ⇒ Hα (E) = 0 .

Now, for any ε > 0 there exists a covering {Ei }i∈N of E such that
X
|Ei |α ⩽ ε .
i

1
It therefore follows that |Ei | ⩽ ε α for all i ∈ N so that also Hα1 (Ei ) < ε. Sending
εα
ε → 0 we obtain the desired result.

2. The second claim follows now directly from the definition of Hausdorff dimension
and the first claim.

3. Here it suffices to show that

Hα (E) > 0 ⇒ Hβ (E) = ∞ , ∀β < α .

To prove this we observe that for any δ-covering {Ei }i of E we have

X X X
|Ei |α = |Ei |β+(α−β) ⩽ δ α−β |Ei |β .
i i i

Then the result follows since the right hand-side vanishes for δ → 0.

7.1 Lower bounds on the Hausdorff dimension

The next step is to understand the effect of Hölder-continuous functions on Hausdorff
dimensions. Let us first recall the definition of α–Hölder continuous functions.

Definition 7.1.1. A function f : (E1 , ϱ1 ) → (E2 , ϱ2 ), for two metric spaces (Ei , ϱi ) is
α–Hölder continuous with constant C ∈ (0, ∞) (for some α ∈ (0, 1]) if

ϱ2 (f (x), f (y))
sup =C.
x̸=y∈E1 ϱ2 (x, y)

We then start our analysis with the following observation.

Lemma 7.1.2. If f : (E1 , ϱ1 ) → (E2 , ϱ2 ) is a α–Hölder continuous function (for some

α ∈ (0, 1]), then
Hβ (f (E1 )) ⩽ C β Hα·β (E1 ) , ∀β ⩾ 0 ,

where C > 0 is the α–Hölder constant of the function f . In particular

1
dim(f (E1 )) ⩽ dim(E1 ) .
α
7.2. UPPER BOUNDS ON THE HAUSDORFF DIMENSION 69

Proof. For any δ, ε > 0 let {Fkε,δ }k∈N be a δ-covering of E1 such that

Hδαβ (E1 ) ⩾
X ε,δ
|Fk |αβ − ε .
k

It follows that Gε,δ ε,δ

k = f (Fk ) is a covering of f (E1 ) with diameter

|f (Fkε,δ )| ⩽ C|Fkε,δ |α .

Therefore
β
|Gε,δ αβ
X
β β
HCδ α (f (E1 )) ⩽ k | ⩽ C (Hδ (E1 ) + ε) .
k

Sending ε, δ → 0 we obtain the desired result. In particular, we obtain that

1
dim(f (E1 )) ⩽ dim(E1 )
α
follows from the previous proposition.

As a corollary of this result we obtain the following.

Corollary 7.1.3. It holds that

dim(B(A)) ⩽ 2dim(A) ∧ d ,

for any A ⊆ [0, ∞).

7.2 Upper bounds on the Hausdorff dimension

The aim in this section is to prove the following result.

Theorem 7.2.1. Let d ⩾ 2 and (Bt )t⩾0 be d-dimensional Brownian motion. Then

dim(B[0, ∞)) = 2 .

It follows from Corollary 7.1.3 that dim(B[0, ∞)) ⩽ 2 for d ⩾ 2, so the only thing
left to prove is a lower bound on the dimension. To obtain a lower bound we use a so-
called “energy method”, which is reminiscent of Kolmogorov’s continuity criterion. We
start with a simple variant of this energy method, which is dubbed the “mass distribution
principle”.

Theorem 7.2.2 (Mass distribution principle). Suppose that (E, ϱ) is a metric space and
µ is a Borel measure on E such that for constants C, δ > 0 and α ⩾ 0 it holds that

0 < µ(E) < ∞ ,

and
µ(V ) ⩽ C|V |α .
for all closed sets V with |V | ⩽ δ. Then

µ(E)
Hα (E) ⩾ >0,
C
and hence dim(E) ⩾ α.
70CHAPTER 7. HAUSDORFF DIMENSION OF THE RANGE OF BROWNIAN MOTION

Proof. Take a cover {Ui }i of E with sets |Ui | ⩽ δ. Then

X X
µ(E) ⩽ µ(Ui ) ⩽ C |Ui |α ,
i i

so that the result follows.

We will now find a more sophisticated criterion, which does not require the condition
µ(V ) ⩽ C|V |α , which is sometimes hard to check. We define the α-energy associated to a
measure µ by: Z
1
Iα (µ) = α
dµ(x) dµ(y) ∈ (0, ∞] .
E×E ϱ(x, y)
Then we can prove the following result.
Theorem 7.2.3 (Energy method). Let α ⩾ 0 and µ be a Borel measure such that

0 < µ(E) < ∞ .

Then for any ε > 0 we can bound:

µ(E)2
Hεα (E) ⩾ R R −α dµ(x) dµ(y)
.
ϱ(x,y)⩽ε ϱ(x, y)

In particular, if Iα (µ) < ∞, then Hα (E) = ∞ and, in particular, dim(E) ⩾ α.

Proof. Let {Ui }i∈N be a pairwise disjoint covering of the set E. Then we can lower bound
Z Z XZ Z
−α
ϱ(x, y) dµ(x) dµ(y) ⩾ ϱ(x, y)−α dµ(x) dµ(y)
ϱ(x,y)⩽ε i Ui ×Ui
X Z Z
⩾ |Ui |−α dµ(x) dµ(y)
i Ui Ui
X
= |Ui |−α µ(Ui )2 .
i

Now we want to compare the latter quantity with µ(E). Here by Cauchy–Schwartz we
have that
X X α µ(Ui )
µ(E) ⩽ µ(Ui ) = |Ui | 2 α
i i
|Ui | 2
!1 !1
X 2 X (µ(Ui ))2 2
⩽ |Ui |α .
|Ui |α
i i
We conclude that
X µ(E)2 µ(E)2
|Ui |α ⩾ P (µ(U ))2 ⩾ R R −α dµ(x) dµ(y)
,
ϱ(x,y)⩽ε ϱ(x, y)
i
i i |Ui |α

as required.

Now we are going to apply this method to study the range of Brownian motion in
d ⩾ 2. In particular, we prove the following proposition.
Proposition 7.2.4. For any d ⩾ 2 and (Bt )t⩾0 a d−dimensional Brownian motion we
have
dim(B[0, ∞)) ⩾ 2 .
7.2. UPPER BOUNDS ON THE HAUSDORFF DIMENSION 71

Note that this concludes the proof of Theorem 7.2.3.

Proof. Of course, we must define an appropriate measure. We consider here the (random)
occupation measure Z t
µB,t (A) = 1A (Bs ) ds .
0
Clearly
t = µB,t (B[0, t]) ∈ (0, ∞) ,
so there is hope to apply Theorem 7.2.3. In order to apply the theorem it suffices to prove
that for any α ∈ (0, 2) we have
Z
1
E α
dµB,t (x) dµB,t (y) < ∞ .
Rd ×Rd |x − y|

This can be rewritten as

Z tZ t Z tZ t
1 1 1
E α
du ds = α ds du · E α
.
0 0 |B s − B u | 0 0 |s − u| 2 |B 1|

Now, the last expectation is finite because

Z
1
α
dz < ∞ ,
B0,1 |z|

for all α < d. In addition, the first integral is finite because

Z t Z t
1 1
α ds ⩽ 2 α ds < ∞ .
0 |s − u| 2 0 |s| 2