• No results found

A. Gomes, Charles

N/A
N/A
Protected

Academic year: 2020

Share "A. Gomes, Charles"

Copied!
67
0
0

Loading.... (view fulltext now)

Full text

(1)

Yuval Peres

Lecture notes edited by B

alint Vir

ag and Elchanan Mossel

(2)

Preface. These notes record lectures I gave at the Statistics Department, University of

California,BerkeleyinSpring1998. Iamgratefulto thestudentswhoattendedthecourse

and wrotetherstdraft ofthenotes: DiegoGarcia,Yoram Gat,Diogo A.Gomes,Charles

Holton, Frederic Latremoliere, Wei Li, Ben Morris, Jason Schweinsberg, Balint Virag, Ye

XiaandXiaowenZhou. ThedraftwaseditedbyBalintViragandElchananMossel. Ithank

Pertti Mattila for the invitation to lecture on thismaterial at the joint summerschool in

(3)

Chapter 1. Brownian Motion 1

1. Motivation {Intersectionof Brownianpaths 1

2. Gaussianrandomvariables 1

3. Levy's construction of Brownian motion 4

4. Basic properties ofBrownianmotion 6

5. Hausdor dimensionand Minkowskidimension 11

6. Hausdor dimensionof theBrownian pathand theBrowniangraph 13

7. Onnowheredierentiability 17

8. Strong Markovpropertyand thereectionprinciple 18

9. Localextrema ofBrownianmotion 19

10. Area ofplanarBrownianmotion paths 20

11. Zeros of theBrownian motion 21

12. Harris'Inequalityand itsconsequences 26

13. Pointsof increaseforrandomwalksand Brownian motion 29

14. Frostman's Lemma, energy,and dimensiondoubling 33

15. Skorokhod'srepresentation 38

16. Donsker'sInvariancePrinciple 43

17. Harmonicfunctions and Brownian motioninR d

46

18. Maximumprincipleforharmonicfunctions 49

19. The Dirichletproblem 50

20. Polar pointsand recurrence 51

21. Capacityand harmonicfunctions 53

22. Kaufman's theoremon uniformdimension doubling 56

23. Packingdimension 58

24. Hausdor dimensionand randomsets 59

25. An extensionof Levy's modulusof continuity 60

(4)

Brownian Motion

1. Motivation { Intersection of Brownian paths

ConsideranumberofBrownianmotionpathsstartedat dierentpoints. Saythatthey

intersect if there is a point which lies on all of the paths. Do the paths intersect? The

answerto thisquestiondepends onthedimension:

InR 2

,anynitenumberofpathsintersectwithpositiveprobability(thisisatheorem

of Dvoretsky,Erd}os,Kakutani inthe1950's),

InR 3

,two pathsintersect withpositiveprobability,butnotthree(thisis atheorem

of Dvoretsky,Erd}os,Kakutani andTaylor),

In R d

for d4,no pairof paths intersect withpositiveprobability.

The principle we willuse to establish these results is intersection equivalencebetween

BrownianmotionandcertainrandomCantor-typesets. Herewewillintroducetheconcept

for R 3

only. Partition the cube [0;1] 3

in eight congruent sub-cubes, and keep each of the

sub-cubeswithprobability 1

2

. Foreachcubethatremainedattheendofthisstage,partition

it into eight sub-cubes, keeping each of them with probability 1

2

, and so on. Let Q 3; 1

2

denote thelimitingset| that is,theintersection of thecubesremaining at allsteps. This

set is not empty with positive probability, since, if we consider that the remaining

sub-cubes of a given cube as its \children"in a branching process, then the expected number

of ospringsis four,sothisprocesshaspositiveprobabilitynotto dieout.

Onecanprove that,thereexiststwopositiveconstantsC

1 ;C

2

suchthat,ifisaclosed

subset of [0;1] 3

, and fB

t

g is a Brownian motion started at a point uniformly chosen in

[0;1] 3

,then:

C

1 P

Q

3; 1

2

\6=;

P( 9t0 B

t

2)C

2 P

Q

3; 1

2

\6=;

The motivation is that, though the intersection of two independent Brownian paths is

a complicatedobject, theintersection of two sets of theform Q 3; 1

2

is a set of the same

kind|namely,Q 3; 1

4

. The previouslydescribedbranchingprocessdiesoutassoon aswe

intersect more than two of these Cantor-type sets|hence the result about intersection of

paths inR 3

.

2. Gaussian random variables

Brownianmotionisat themeetingpointofthemostimportant categoriesofstochastic

processes: it is a martingale, a strong Markov process, a process with independent and

stationary increments, and a Gaussian process. We will construct Brownian motion as a

(5)

Definition 2.1. Areal-valuedrandomvariableXonaprobabilityspace(;F;P)has

a standard Gaussian (orstandard normal) distributionif

P(X>x)= 1

p

2 Z

+1

x e

u 2

=2

du

A vector-valued random variable X has an n-dimensional standard Gaussian

dis-tributionifits ncoordinatesare standardGaussianand independent.

A vector-valued random variable Y : ! R p

is Gaussian if there exists a

vector-valuedrandomvariableXhavingann-dimensionalstandardGaussiandistribution,apn

matrixA and ap-dimensionalvector b such that:

Y =AX+b (2.1)

We arenowreadyto denetheGaussianprocesses.

Definition 2.2. A stochasticprocess (X

t )

t2I

issaid to bea Gaussian processiffor

all k and t

1 ;:::;t

k

2I thevector (X

t1

;:::;X

t

k )

t

is Gaussian.

Recallthat thecovariance matrixof a randomvector isdened as

Cov(Y)=E

(Y EY)(Y EY) t

Then,bythelinearityofexpectation,theGaussian vector Y in(2.1) has

Cov(Y)=AA t

:

Recallthatan nnmatrixAis saidto beorthogonalifAA t

=I

n

. Thefollowinglemma

shows thatthedistributionofaGaussianvector isdeterminedbyits meanand covariance.

Lemma 2.3.

(i) If is an orthogonal nn matrix and X is an n-dimensional standard Gaussian

vector, then X is also an n-dimensionalstandard Gaussian vector.

(ii)If Y andZ areGaussianvectorsin R n

such that EY =EZ andCov(Y)=Cov(Z),

then Y and Z have the same distribution.

Proof.

(i) As the coordinates of X are independent standard Gaussian, X has density given

by:

f(x)=(2) n

2

e kxk=2

where k k denotes Euclidean norm. Since preserves this norm, the density of X is

invariant under.

(ii)ItissuÆcienttoconsiderthecasewhenEY =EZ =0. Then,usingDenition2.1),

there exist standardGaussianvectors X

1 ,X

2

andmatrices A;C sothat

Y =AX

1

and Z =CX

2 :

By adding some columnsof zeroesto A or C if necessary, we can assume that X

1 , X

2 are

bothk-vectors forsome k and A,C arebothnk matrices.

(6)

form a basis for the space A. For any matrix M let M

i

denote the ith row vector of M,

and denethelinear mapfrom A to C by

A

i =C

i

fori=1:::`.

Wewanttocheckthatisanisomorphism. Assumethatthereisavectorv

1 A

1

+:::+v

` A

`

whose imageis 0. Then the k-vector v =(v

1 ;v

2 ;:::;v

`

;0;:::;0) satises v t

C =0,and so

kv t

Ak 2

= v t

AA t

v = v t

CC t

v = 0, giving v t

A = 0. This shows that is one-to-one, in

particular dimA dimC. By symmetry A and C must have the same dimension, so is

an isomorphism.

AsthecoeÆcient(i;j)ofthematrixAA t

isthescalarproductofA

i andA

j

,theidentity

AA t

=CC t

impliesthat is an orthogonal transformation from A to C. We can extend

itto maptheorthocomplement ofA totheorthocomplementof Corthogonally,getting an

orthogonal map:R k

!R k

. Then

Y =AX

1

; Z =CX

2

=AX

2 ;

and (ii)follows from (i).

Thus,thersttwomomentsofaGaussianvectoraresuÆcienttocharacterizeits

distri-bution,hencetheintroductionofthenotationN(;)todesignatethenormaldistribution

withexpectation and covariance matrix. Ausefulcorollary of thislemmais:

Corollary 2.4. Let Z

1 ;Z

2

beindependent N(0; 2

) random variables. Then Z

1 +Z

2

and Z

1 Z

2

aretwo independent random variables havingthe same distributionN(0;2 2

).

Proof. 1

(Z

1 ;Z

2

) isa standardGaussianvector,and so,if:

= 1

p

2

1 1

1 1

thenis an orthogonal matrixsuch that

( p

2) 1

(Z

1 +Z

2 ;Z

1 Z

2 )

t

=

1

(Z

1 ;Z

2 )

t

;

and ourclaim followsfrom part (i)of theLemma.

As a conclusion of this section, we state the following tail estimate for the standard

Gaussiandistribution:

Lemma 2.5. LetZ be distributed as N(0;1). Then for all x0:

x

x 2

+1 1

p

2 e

x 2

=2

P( Z >x) 1

x 1

p

2 e

x 2

=2

Proof. Theright inequalityisobtained bythe estimate:

P(Z >x) Z

+1

x u

x 1

p

2 e

u 2

=2

du

since, intheintegral, ux. The left inequalityisproved asfollows: Let usdene

f(x):=xe x

2

=2

(x 2

+1) Z

+1

e u

2

=2

(7)

We remarkthat f(0)<0 and lim

x!+1

f(x)=0:Moreover,

f 0

(x) = (1 x 2

+x 2

+1)e x

2

=2

2x Z

+1

x e

u 2

=2

du

= 2x

Z

+1

x e

u 2

=2

du 1

x e

x 2

=2

;

so theright inequality impliesf 0

(x)0 for all x 0. This impliesf(x)0, provingthe

lemma.

3. Levy's construction of Brownian motion

3.1. Denition. StandardBrownian motion on an intervalI =[0;a] orI =[0;1) is

denedbythefollowingproperties:

Definition 3.1. A stochastic process fB

t g

t2I

isa standard Brownian motionifit

isa Gaussianprocess suchthat:

(i) B

0 =0,

(ii) 8k naturaland 8t

1

<:::<t

k

in I: B

t

k B

t

k 1

;:::;B

t

2 B

t

1

areindependent,

(iii) 8t;s2I with t<s B

s B

t

hasN(0;s t) distribution.

(iv) Almost surely,t7!B

t

is continuouson I.

Asa corollaryof thisdenition,one can already remarkthat forallt;s2I:

Cov(B

t ;B

s

)=s^t:

Indeed, assume that t s. Then Cov(B

t ;B

s

) = Cov (B

t B

s ;B

s

) +Cov(B

s ;B

s ) by

bilinearityof the covariance. The rst term vanishes by the independence of increments,

andthesecondtermequalssbyproperties(iii)and(i). ThusbyLemma2.3wemayreplace

properties (ii)and (iii)inthedenitionby:

Forall t;s2I, Cov (B

t ;B

s

)=t^s.

Forall t2I,B

t

hasN(0;t) distribution.

orby:

Forall t;s2I with t<s, B

t B

s and B

s

areindependent.

Forall t2I,B

t

hasN(0;t) distribution.

Kolmogorov's extension theorem implies the existence of any countable time set

sto-chasticprocessfX

t

gifweknowitsnite-dimensionaldistributionsandtheyareconsistent.

Thus, standard Brownian motion could be easily constructed on any countable time set.

HoweverknowingnitedimensionaldistributionsisnotsuÆcienttogetcontinuouspaths,

asthefollowingexampleshows.

Example 3.2. Suppose that standard Brownian motion fB

t

g on [0;1] has been

con-structed, and consider an independent random variable U uniformlydistributed on [0;1].

Dene:

~

B

t =

B

t

ift6=U

0 otherwise

Thenite-dimensionaldistributionsoff ~

B

t

garethesameastheonesoffB

t

g. However,the

process f ~

(8)

3. LEVY'SCONSTRUCTION OFBROWNIANMOTION 5

Inmeasuretheory,oneoftenidentiesfunctionswiththeirequivalenceclassfor

almost-everywhere equality. As the above example shows, it is important not to make this

iden-tication in the study of continuous-time stochastic processes. Here we want to dene a

probabilitymeasureon theset ofcontinuous functions.

3.2. Construction. Thefollowingconstruction,duetoPaulLevy,consistofchoosing

the \right" values for the Brownian motion at each dyadic point of [0;1] and then

inter-polating linearly between these values. This construction is inductive, and, at each step,

a process is constructed, that hascontinuous paths. Brownian motion isthen theuniform

limitof these processes|hence its continuity. We willusethe followingbasiclemma. The

proof can be found,forinstance, inDurrett (1995).

Lemma 3.3 (Borel-Cantelli). Let fA

i g

i=0:::1

be a sequence of events, andlet

fA

i

i.o. g=limsup

i!1 A

i =

1

\

i=0 1

[

j=i A

j ;

where\i.o." abbreviates \innitely often".

(i) If P

1

i=0 P(A

i

)<1, then P(A

i

i.o. )=0.

(ii) If fA

i

g are pairwise independent, and P

1

i=0 P(A

i

)=1, then P(A

i

i.o.)=1.

Theorem 3.4 (Wiener1923). Standard Brownian motion on [0;1) exists.

Proof. (Levy 1948)

We rstconstructstandardBrownianmotionon[0;1]. Forn0,letD

n

=fk=2 n

:0

k 2 n

g, and let D = S

D

n

. Let fZ

d g

d2D

be a collection of independent N(0;1) random

variables. We willrst construct the valuesof B on D. Set B

0

=0, and B

1 = Z

1

. In an

inductive construction,foreach nwe willconstruct B

d

foralld2D

n

sothat

(i) Forallr <s<t inD

n

,theincrement B

t B

s

hasN(0;t s)distributionand is

independentofB

s B

r .

(ii) B

d

ford2D

n

areglobally independent oftheZ

d

ford2DnD

n .

These assertions hold for n = 0. Suppose that they hold for n 1. Dene, for all

d2D

n nD

n 1

arandomvariable B

d by

B

d =

B

d +B

d +

2

+ Z

d

2 (n+1)=2

(3.1)

where d +

= d+2 n

, and d = d 2 n

, and both are in D

n 1

. Since 1

2 [B

d

+ B

d ] is

N(0;1=2 n+1

)byinduction,and Z

d =2

(n+1)=2

isanindependentN(0;1=2 n+1

),theirsumand

theirdierence,B

d B

d

andB

d+ B

d

arebothN(0;1=2 n

)andindependentbyCorollary

2.4. Assertion(i) follows fromthisand theinductivehypothesis,and (ii)is clear.

HavingthuschosenthevaluesoftheprocessonD,wenow\interpolate"betweenthem.

Formally,let F

0

(x)=xZ

1

,and forn1,let let usintroducethefunction:

F

n (x)=

8

<

: 2

(n+1)=2

Z

x

forx2D

n nD

n 1 ;

0 forx2D

n 1 ;

linear betweenconsecutive pointsinD :

(9)

Thesefunctions arecontinuouson [0;1], andforall n andd2D n B d = n X i=0 F i (d)= 1 X i=0 F i (d): (3.3)

Thiscan be seenbyinduction. Supposethat itholdsforn 1. Letd2D

n D

n 1 . Since

for0in 1 F

i

islinear on [d ;d +

],weget

n 1 X i=0 F i (d)= n 1 X i=1 F i

(d )+F

i (d + ) 2 = B d +B d + 2 : (3.4) Since F n

(d)=2

(n+1)=2

Z

d

,comparing(3.1) and (3.4) gives(3.3).

Ontheother hand, we have,bydenitionof Z

d

and byLemma 2.5:

P jZ

d jc

p n exp c 2 n 2

so theseries P 1 n=0 P d2Dn P(jZ d jc

p

n) converges assoon asc> p

2log2. Fix such a c.

By the Borel-Cantelli Lemma 3.3 we concludethat there exists a random but niteN so

thatforall n>N and d2D

n

we have jZ

d j<c

p

n,and so:

kF n k 1 <c p n2 n=2 : (3.5)

This upper bound implies that the series P

1

n=0 F

n

(t) is uniformly convergent on [0;1],

and so it has a continuous limit, which we call fB

t

g. All we have to check is that the

increments of this process have the right nite-dimensional joint distributions. This is a

directconsequenceof thedensityof thesetDin[0;1] andthecontinuityof paths. Indeed,

let t 1 > t 2 > t 3

be in [0;1], then they are limits of sequences t

1;n , t 2;n and t 3;n in D,

respectively. Now

B t 3 B t 2 = lim k!1 (B t 3;k B t 2;k )

is a limit of Gaussian random variables, so itself is Gaussian with mean 0 and v

ari-ance lim n!1 ( t 3;n t 2;n ) = t

3 t

2

. The same holds for B

t2 B

t1

, moreover, these two

random variables are limit of independent random variables, since for n large enough,

t 1;n > t 2;n > t 3;n

. Applying this argument for any number of increments, we get that

fB

t

ghasindependentincrementssuchthatandforalls<tin[0;1]B

t B

s

hasN(0;t s)

distribution.

We have thusconstructed Brownian motion on [0;1]. To conclude, if fB n

t g

t

forn 0

areindependentBrownianmotionson [0;1], then

B t =B btc t btc + X 0i<btc B i 1

meetsour denitionof Brownianmotionon [0;1).

4. Basic properties of Brownian motion

Let fB(t)g

t0

be a standard Brownian motion, and let a6= 0. The following scaling

relation isa simpleconsequenceof thedenitions.

(10)

Also,dene thetime inversion of fB

t gas

W(t)=

0 t=0;

tB( 1

t

) t>0.

We claim thatW isa standardBrownian motion. Indeed,

Cov(W(t);W(s))=tsCov(B( 1

t ;

1

s

))=ts( 1

t ^

1

s

)=t^s;

so W and B have the same nite dimensional distributions, and they have the same

dis-tributions as processes on the rational numbers. Since the paths of W(t) are continuous

exceptmaybe at 0,we have

lim

t#0

W(t)= lim

t#0;t2Q

W(t)=0 a.s.

sothepaths of W(t)are continuous on[0;1) a.s. As acorollary,weget

Corollary 4.1. [Law of Large Numbers for Brownian motion]

lim

t!1 B(t)

t

=0 a:s:

Proof. lim

t!1 B(t)

t

=lim

t!1 W(

1

t

)=0a.s.

Exercise 4.2. Provethisresultdirectly. UsetheusualLawofLargeNumbersto show

that lim

n!1 B(n)

n

= 0. Then show that B(t) does not oscillate too much between n and

n+1.

Remark. Thesymmetryinherentinthetimeinversionpropertybecomesmore

appar-ent ifone considerstheOrnstein-Uhlenbeckdiusion,whichis given by

X(t)=e t

B(e 2t

)

This is a stationary Markov chain with X(t) N(0;1) for all t. It is a diusion with

a drift toward the origin proportional to the distance from the origin. Unlike Brownian

motion, the Ornstein-Uhlenbeck diusion is time reversible. The time inversion formula

givesfX(t)g

t0 d

=fX( t)g

t0

. Fort near 1, X(t) relates to theBrownianmotion near

0,and fortnear 1,X(t)relates to theBrownianmotion near1.

OneoftheadvantagesofLevy's constructionofBrownianmotionisthatiteasilyyields

a modulusofcontinuityresult. FollowingLevy,we dened Brownianmotion asan innite

sum P

1

n=0 F

n

, where each F

n

is a piecewise linear function given in (3.2). Its derivative

existsalmost everywhere, andbydenitionand (3.5)

kF 0

n k

1

kF

n k

1

2 n

C

1 (!)+

p

n2 n=2

(4.1)

TherandomconstantC

1

(!)isintroducedtodealwiththenitelymanyexceptionsto(3.5).

Nowfort;t+h2[0;1], we have

jB(t+h) B(t)j X

n jF

n

(t+h) F

n (t)j

X

hkF 0

n k

1 +

X

2kF

n k

1

(11)

By (3.5) and (4.1) if`>N forarandomN,then theabove is boundedby

h(C

1 (!)+

X

n` c

p

n2 n=2

)+2 X

n>` c

p

n2 n=2

C

2 (!)h

p

`2 `=2

+C

3 (!)

p

`2 `=2

Theinequalityholdsbecauseeachseriesisboundedbyaconstanttimesitsdominantterm.

Choosing`=blog

2

(1=h)c, andchoosingC(!)to take careof thecases when`N,weget

jB(t+h) B(t)jC(!) r

hlog

2 1

h

: (4.3)

The result is a (weak) form of Levy's modulusof continuity. This is not enough to make

fB

t

g adierentiablefunctionsince p

hh forsmallh. But we stillhave

Corollary 4.3. For every < 1

2

, Brownian paths are - Holder a.s.

Exercise 4.4. Show thata Brownian motionisa.s. not 1

2

- Holder.

Remark. Theredoesexistat=t(!)suchthatjB(t+h) B(t)jC(!)h 1

2

foreveryh,

a.s. However, such thavemeasure 0. Thisistheslowestmovement that islocallypossible.

Solution. LetA

k ;n

betheeventthatB((k+1)2 n

) B(k2 n

)>c p

n2 n=2

. Then

P(A

k ;n

)=P(B(1)>c p

n ) c

p

n

c 2

n+1 e

c 2

n=2

:

If0<c< p

2log(2)thenc 2

=2<log(2)and2 n

P(A

k ;n

)!1. Therefore,

P( 2

n

\

k =1 A

c

k ;n

)=[1 P(A

k ;n )]

2 n

e 2

n

P(A

k ;n )

!0:

Thelastinequalitycomesfrom thefact that1 xe x

for all x. Byconsideringh=2 n

;one

canseethat

P

8h B(t+h) B(t)c p

hlog

2 h

1

=0 ifc< p

2log2 :

We remark that the paths are a.s. not 1

2

-Holder. Indeed, the log

2 (

1

h

) factor in (4.3) cannot be

ignored. WewillseelaterthattheHausdordimensionofthegraphofBrownianmotionis 3

2 a.s.

Having proven that Brownian paths are somewhat \regular", let us see why they are

\bizarre". One reason is that the paths of Brownian motion have no intervals of

mono-tonicity. Indeed, if [a;b] is an interval of monotonicity, then dividing it up into n equal

sub-intervals[a

i ;a

i+1

]eachincrement B(a

i

) B(a

i+1

) hasto have thesamesign. Thishas

probability22 n

, and taking n! 1 shows that theprobabilitythat [a;b]is an interval

of monotonicity must be 0. Taking a countable union gives that there is no interval of

monotonicitywith rationalendpoints, buteach monotoneinterval wouldhave a monotone

rationalsub-interval.

We willnowshowthatforanytimet

0

,Brownianmotionisnotdierentiableatt

0 . For

this, weneed asimpleproposition.

Proposition 4.5. A.s.

limsup B(n)

p

n

=+1; liminf

n!1 B(n)

p

n

(12)

Remark. Comparing thiswithCorollary 4.1, it isnatural to ask what sequence B(n)

shouldbe dividedby to geta limsupwhich isgreater than 0but lessthan1. An answer

isgiven bytheLawof theiterated logarithm inalater section.

The proof of (4.4) relieson the followingstandard fact, whose proof can be found, for

example,inDurrett (1995). Consideraprobabilitymeasureon thespace ofreal sequences,

andlet X

1 ;X

2

;::: bethesequenceof randomvariablesitdenes. An event,thatisa

mea-surablesetofsequences,AisexchangeableifX

1 ;X

2

;::: satisfyAimpliesthatX

1 ;X

2 ;:::

satisfy A for all nite permutations . Here nite permutation means that

n

=n for all

suÆcientlylargen.

Proposition 4.6 (Hewitt-Savage 0-1 Law). If A is an exchangeableevent for an i.i.d.

sequencethen P(A) is0 or 1.

Proof of Proposition 4.5.

P(B(n)>c p

n i.o. )limsup

n!1

P(B(n)>c p

n)

Bythescalingproperty,theexpressioninthelimsupequalsP(B(1)>c),whichispositive.

LettingX

n

=B(n) B(n 1)theHewitt-Savage 0-1lawgivesthatB(n)>c p

ninnitely

often. Takingtheintersectionoverallnaturalcgivestherstpart of(4.4),and thesecond

isproved similarly.

ThetwoclaimsofProposition4.5togethermeanthatB(t)crosses0forarbitrarilylarge

values of t. If we use time inversion W(t)=tB( 1

t

), we get that Brownian motion crosses

0 for arbitrarily small values of t. Letting Z

B

= ft : B(t) = 0g, this means that 0 is an

accumulation pointfrom theright forZ

B

. Butweget even more. Fora functionf,dene

theupperand lower right derivatives

D

f(t) = limsup

h#0

f(t+h) f(t)

h

;

D

f(t) = liminf

h#0

f(t+h) f(t)

h

:

Then

D

W(0)limsup W(

1

n

) W(0)

( 1

n )

limsup p

nW( 1

n

)=limsup B(n)

p

n

which is innite by Proposition 4.5. Similarly, D

W(0) = 1, showing that W is not

dierentiableat 0.

Corollary 4.7. Fix t

0

0. Brownian motion W a.s. satises D

W(t

0

) = +1,

D

W(t

0

) = 1, and t

0

is an accumulation point from the right for the level set fs :

W(s)=W(t

0 )g.

Proof. t!W(t

0

+t) W(t

0

)is a standardBrownian motion.

Does this implythat each t

0

is an accumulation point from the right for the level set

fs:W(s)=W(t

0

)ga.s.? Certainlynot;consider,forexamplethelast0offB

t

gbeforetime

(13)

t

0

-smusthave measure0. This istrue ingeneral. SupposeAis ameasurable event (set of

paths)such that

8t

0

P(t!W(t

0

+t) W(t

0

)satisesA)=1:

Let

t

be the operator that shifts paths by t. Then P( T

t

0 2Q

t0

(A)) = 1: In fact, the

Lebesguemeasureofpointst

0

sothatW doesnotsatisfy

t0

(A)is0a.s. Toseethis,apply

Fubinito thedoubleintegral

Z Z

1

0

1(W 2=

t

0

(A))dt

0

dP(W)

Exercise 4.8. Showthat8t

0 P(t

0

isalocalmaximumforB)=0;buta.s.localmaxima

area countable dense setin(0;1).

NowheredierentiabilityofBrownianmotionthereforerequiresamorecarefulargument

thannon-dierentiabilityat axed point.

Theorem 4.9 (Paley,Wiener,Zygmund1933). A.s. Brownian motion is nowhere

dif-ferentiable. Furthermore, almost surely for all t either D

B(t)=+1 or D

B(t)= 1.

Remark. For local maxima we have D

B(t) 0, and for local minima, D

B(t)0,

soit isimportantto have theeither-or inthestatement.

Proof. (Dvoretsky, Erd}os, Kakutani 1961) Suppose that there is a t

0

2 [0;1] such

that 1<D

B(t

0 )D

B(t

0

)<1. Then forsome M we wouldhave

sup

h2[0;1] jB(t

0

+h) B(t

0 )j

h

M: (4.5)

If t

0

is containedinthe binary interval[(k 1)=2 n

;k=2 n

]forn>2, thenforall 1j n

thetriangleinequalitygives

jB((k+j)=2 n

) B((k+j 1)=2 n

) jM(2j+1)=2 n

: (4.6)

Let

n;k

be theevent that (4.6) holdsfor j=1,2,and 3. Then bythescaling property

P(

n;k )P

jB(1)j7M= p

2 n

3

;

which isat most(7M2 n=2

) 3

,sincethe normaldensity islessthan1/2. Hence

P 2

n

[

k=1

n;k !

2 n

(7M2 n=2

) 3

=(7M) 3

2 n=2

;

whosethe sumoverall nis nite. By theBorel-Cantellilemma:

P( (4.5)is satisfed)P 2

n

[

k=1

n;k

forinnitelymanyn !

=0:

That is,forsuÆcientlylargen, there arenothree goodincrementsinarow so(4.5) isnot

satised.

Exercise 4.10. Let>1=2. Showthata.s.forallt>0;9h>0suchthatjB(t+h)

(14)

Solution. Supposethatthereisat

0

2[0;1]suchthat

sup

h2[0;1]

B(t0+h) B(t0)

h

1; and

inf

h2[0;1] B(t

0

+h) B(t

0 )

h

1

Ift

0 2

k 1

2 n

; k

2 n

for n>2,thenthetriangleinequalitygives

jB( k+j

2 n

) B(

k+j 1

2 n

)j2( j+1

2 n

)

:

Fixl1=( 1

2

)andletn;k betheevent

jB( k +j

2 n

) B( k +j 1

2 n

)j2 h

(j+1)

2 n

i

for j=1,2 ::: l

Then

P(

n;k

)[P(jB(1)j2 n=2

2( l+1

2 n

)

] l

[2 n=2

2( l+1

2 n

)

] l

sincethenormaldensityislessthan1/2. Hence

P( 2

n

[

k =1

n;k )2

n

[2 n=2

2( l+1

2 n

)

] l

=C[2

(1 l( 1=2))

] n

;

whichsums. Thus

P(limsup

n!1 2

n

[

k =1

n;k)=0:

Exercise 4.11. (Hard.) A.s.ifB(t

0

)=max

0t1

B(t) thenD

B(t

0

)= 1.

5. Hausdor dimension and Minkowski dimension

Definition 5.1 (Hausdor1918). LetAbeasubsetofametricspace. For >0,and

a possiblyinnite >0 denea measure ofsizeforA by

H

()

(A)=inff X

j jA

j j

:A [

j A

j ;jA

j j<g;

wherejjappliedtoasetmeansitsdiameter. ThequantityH

(1)

(A)iscalledtheHausdor

content of A.Dene the-dimensionalHausdor measureof Aas

H

(A)=lim

#0 H

()

(A)=sup

>0 H

() (A):

H

() is a Borel measure. Sub-additivity is obvious since H

()

(A[D) H

() (A)+

H

()

(D). Countable additivitycan beshownwithmore eort. The graphof H

(A)versus

shows that there is a critical value of where H

(A) jumps from 1 to 0. This critical

valueis calledtheHausdor dimensionof A.

Definition 5.2. The Hausdor dimension ofA is

dim (A)=inff: H

(15)

Notethat given >>0and H

()

(A)<1,we have

H

()

(A) ( )

H

() (A):

where the inequality follows from the fact that jA

j j

( )

jA

j j

for a covering set A

j

of diameter . Since is arbitrary here, we see that if H

(A) < 1, then H

(A) = 0:

Therefore, we can also denetheHausdor dimensionof Aas

supf:H

(A)=1g:

Nowlet uslookat anotherkindof dimensiondened asfollows.

Definition 5.3. The upper and lower Minkowski dimension (also called\Box"

dimension)ofA is denedas

dim

M

(A) = lim

#0 logN

(A)

log (1=) ;

dim

M

(A) = lim

#0 logN

(A)

log (1=) ;

whereN

(A)istheminimumnumberofballsofradiusneededto coverA. Ifdim

M (A)=

dim

M

(A), we callthecommonvaluetheMinkowski dimensionofA,denoted bydim

M .

Exercise 5.4. Show thatdim

M ([0;1]

d

)=d.

Proposition 5.5. For any set A, we have dim

H

(A)dim

M (A).

Proof. Bydenition,H

(2) N

(A)(2)

. If >dim

M

(A),thenwehavelimN

(A)(2)

=

0. It thenfollows thatdim

H

(A)dim

M (A).

We nowlookat some examplestoillustratethedenitionsabove. LetA beall therational

numbersintheunit interval. Since forevery set Dwe have dim

M

(D)=dim

M

(D),we get

dim

M

(A) = dim

M

(A ) =dim

M

([0;1]) = 1. On the other hand, a countable set A always

hasdim

H

(A)=0. Thisisbecause forany;>0,wecan coverthe countable pointsinA

byballsof radius=2, =4, =8, and soon. Then

H

() (A)

X

j jA

j j

X

j (2

j

)

<1:

Anotherexampleistheharmonicsequence A=f0g S

f1=ng

n1

. Itis acountable set, soit

hasHausdordimension0. Itcan be shownthatdim

M

(A)=1=2. (ThosepointsinAthat

arelessthan p

can becovered by1= p

ballsofradius. Therest canbecoveredby2= p

ballsof radius, one oneach point.)

We haveshown inCorollary 4.3that Brownianmotionis -Holderforany<1=2a.s.

This will allow us to infer an upper bound on the Hausdor dimension of its image and

graph. Denethe graphG

f

ofa functionastheset ofpoints(t;f(t))astrangesoverthe

domainof f.

Proposition 5.6. Let f :[0;1]!R bea -Holder continuous function. Then

(16)

Proof. Since f is -Holder, there exists a constant C

1

such that, if jt sj , then

jf(t) f(s)jC

1

=C

1 (1=

1

). Hence, theminimumnumberof ballsof radiusto

coverG

f

satisesN

(G

f )C

2 (1=

2

)forsome other constant C

2 .

Corollary 5.7.

dim

H (G

B

)dim

M (G

B

)dim

M (G

B

)3=2: a.s.

A functionf :[0;1]! R is called\reverse" -Holderifthere exists aconstant C>0 such

that foranyinterval [t;s], there is asubinterval[t

1 ;s

1

][t;s],suchthat jf(t

1

) f(s

1 )j

Cjt sj

.

Proposition 5.8. Let f : [0;1] ! R be -Holder and \reverse" -Holder. Then

dim

M (G

f

)=2 .

Remark. It followsfrom thehypotheses oftheabovepropositionthatsuchafunction

f hasthe propertydim

H (G

f

)>1:(Przytycki-Urbansky,1989.)

Exercise 5.9. Prove Proposition5.8.

Example 5.10. TheWeierstrassnowheredierentiablefunction

W(t)= 1

X

n=0 a

n

cos(b n

t);

ab>1, 0<a<1 is -Holder and \reverse" -Holderforsome 0< <1. For example,if

a=1=2, b=4,then =1=2.

Lemma 5.11. Suppose X isa complete metric space. Let f :X !Y be -Holder. For

any A X, we have dim

H

f(A) dim

H

(A)=: Similar statements for dim

M

and dim

M

are also true.

Proof. Ifdim

H

(A)< <1,thenthereexistsacoverfA

j

gsuchthatA S

j A

j and

P

j jA

j j

< . Then ff(A

j

)g is a cover for f(A), and jf(A

j

)j CjA

j j

, where C is the

constant from the-Holdercondition. Thus,

X

j jf(A

j )j

=

C =

X

j jA

j j

<C =

!0

as!0,and hence dim

H

f(A)=:

Corollary 5.12. For A[0;1), we have dim

H

B(A)2dim

H

(A)^1 a:s:

6. Hausdor dimension of the Brownian path and the Brownian graph

The nowhere dierentiability of Brownian motion established in the previous section

suggests that its graph has dimension higher than one. Recall that the graph G

f of a

function f is the set of points (x;f(x)) where x ranges over the set where f is dened.

Taylor(1953) showed thatthegraph ofBrownianmotion hasHausdor dimension3=2.

Denethed-dimensionalstandardBrownianmotionwhosecoordinatesareindependent

one dimensional standardBrownian motions. This measure is invariant under orthogonal

transformations of R d

(17)

transforma-We willseethat planarBrownian motionisneighborhoodrecurrent,that is,itvisitsevery

neighborhood in the plane innitely often. In this sense, the image of planar Brownian

motion is comparable to the plane itself; another sense in which this happens is that of

Hausdor dimension: the image of planar and higher dimensional Brownian motion has

Hausdor dimensiontwo. Summingup,we willprove

Theorem 6.1 (Taylor 1953). Let B be d dimensional Brownian motion dened on

the time set [0;1]. Then

dim

H G

B

=3=2 a.s.

Moreover, if d2, then

dim

H

B[0;1]=2 a.s.

HigherdimensionalBrownianmotion therefore doublesthedimension ofthe timeline.

Naturally, the question arises whether this holds for subsets of the time line as well. In

certain sense,thiseven holdsford=1: notethe\^d"inthe followingtheorem.

Theorem 6.2 (McKean1955). For every subset A of [0;1), the image of A under d

dimensional Brownian motionhas Hausdor dimension 2dim

H

A^d a.s.

Theorem 6.3 (UniformDimensionDoubling(Kaufman 1969)).

Let B beBrownian motionin dimensionat least 2 A.s, for any A[0;1), we have

dim

H

B(A)=2dim

H (A).

Notice the dierence between the last two results. Kaufman's theorem has a much

stronger claim: itstatesdimensiondoublinguniformlyforallsets. Forthistheorem,d 2

is a necessary condition: we will see later that the zero set of one dimensional Brownian

motion hasdimension half,whileitsimage is thesingle point 0. We willprove Kaufman's

theorem inalater section. ForTheorem(6.1) weneedthe followingfact.

Theorem 6.4 (Mass DistributionPrinciple). If A X supports a positive Borel

mea-suresuch that (D)CjDj

for anyBorelsetD,then H

1 (A)

(A)

C

. Thisimpliesthat

H

(A) (A)

C

, and hencedim

H

(A):

Proof. IfA S

j A

j ,then

P

j jA

j j

C 1

P

j (A

j )C

1

(A):

Example 6.5. Consider themiddlethirdCantor setK

1=3

. We willseethat

dim

M (K

1=3

)=dim

H (K

1=3

)=log2=log3:

At thenth levelof theconstruction of K

1=3

,wehave 2 n

intervalsof length3 n

. Thus,

if3 n

<3 n+1

we haveN

(K

1=3 )2

n

. Therefore: dim

M (K

1=3

)log2=log3. Onthe

otherhand,let =log2=log3and betheCantor-Lebesguemeasure,whichassignsmass

2 n

foranylevel ninterval. LetD bea subsetof K

1=3

,and let nbethe integer such that

3 n

jDj3 1 n

. TheneverysuchDiscoveredbyatmostfourshorterterracingintervals

of theform[k3 n

;(k+1)3 n

]forsome k. We thenhave

(D)42 n

4jDj

;

and so

H

(K

1=3 )H

1 (K

1=3

)1=4;

which impliesthat dim

H

(18)

Next,we introducethe energymethoddueto Frostman.

Theorem 6.6 (Frostman 1935). Givena metricspace(X;), ifisaniteBorel

mea-sure supported on AX and

E

()

def

= ZZ

d(x)d(y)

(x;y)

<1;

then H

1

(A)>0, andhencedim

H

(A).

It can be shown that under the above conditions H

(A) = 1. The converse of this

theorem is also true. That is, for any <dim

H

(A), there exists a measure on A that

satisesE

()<1.

Proof. Given a measure ,dene thefunction

(;x)

def

= Z

d(y)

(x;y)

;

sothatE

()=

R

(;x)d(x). LetA[M]denotethesubsetofAwhere

(;)isatmost

M. There existsa numberM such thatA[M]haspositive-measure, sinceE

()isnite.

Let denote the measure restricted to theset A[M]. Then forany x 2A[M], we have

( ;x)

(;x) M. Now let D be a bounded subset of X. If D\A[M] = ; then

(D) = 0. Otherwise, take x 2 D\A[M]. Let m be the largest integer such that D is

containedintheopen ballofradius 2 m

aboutx. Then

M

Z

d(y)

(x;y)

Z

D

d(y)

(x;y)

2 m

(D):

The last inequality comes from the fact that (x;y) 2 m

for each y in D. Thus, we

have (D) M2 m

M(2jDj)

. This also holds when D\A[M] = ;. By the Mass

DistributionPrinciple,we concludethatH

1

(A)>H

1

(A[M])>0.

Nowweare readyto prove thesecond partof Taylor'stheorem.

Proof of Theorem6.1, Part2. From Corollary 4.3 we have that B

d

is Holder

forevery <1=2 a.s. Therefore, Lemma5.11 impliesthat

dim

H B

d

[0;1]2:

a.s. Fortheotherinequality,wewilluseFrostman'senergymethod. Let<2. Weusethe

occupationmeasure

B def

= LB 1

,whichmeansthat

B

(A)=LB 1

(A), forall measurable

subsets Aof R d

,or, equivalently,

Z

f(x)d

B (x)=

Z

1

0 f(B

t )dt

forall measurablefunctions f. We want to showthat

E Z

R d

Z

R d

d

B (x)d

B (y)

jx yj

=E Z

1

0 Z

1

0

dsdt

jB(t) B(s)j

<1: (6.1)

Letusevaluatethe expectation:

EjB(t) B(s)j

=E(jt sj 1=2

jZj)

=jt sj =2

Z

R c

d

jzj

e jzj

2

=2

(19)

HereZ denotesthed-dimensionalstandardGaussianrandomvariable. Theintegralcanbe

evaluatedusingpolarcoordinates, butallweneedisthatitisaniteconstant cdepending

on d and only. Substitutingthisexpressioninto (6.1) and usingFubini'stheoremwe get

EE

(

B )=c

Z

1

0 Z

1

0

dsdt

jt sj =2

2c Z

1

0 du

u =2

<1: (6.2)

ThereforeE

(

B

)<1 a.s.

Remark. Levy showed earlierin1940 that, when d=2, we have H 2

(B[0;1])=0 a.s.

The statement isactuallyalso true foralld 2.

NowletusturntothegraphG

B

ofBrownianmotion. Wewillshowaproof oftherst

halfof Taylor'stheorem forone dimensionalBrownianmotion.

Proof of Theorem6.1, Part1. We have shown inCorollary5.7 that

dim

H G

B

3=2:

Fortheotherinequality,let<3=2 andletAbethesubsetofthegraph. Deneameasure

on thegraphusingprojectionto thetimes axis:

(A) def

= L(f0t1:B(t)2Ag):

Changing variables, the energyof can be writtenas:

ZZ

d(x)d(y)

jx yj

= Z

1

0 Z

1

0

dsdt

(jt sj 2

+jB(t) B(s)j 2

) =2

:

Boundingtheintegrand,taking expectations, and applyingFubiniwe getthat

EE

()2 Z

1

0 E

(t 2

+B(t) 2

) =2

dt: (6.3)

Letn(z) denotethe standardnormal density. By scaling,theexpectedvalue above can be

writtenas

2 Z

+1

0 (t

2

+tz 2

) =2

n(z)dz: (6.4)

Comparing the size of the summands in the integration suggests separating z p

t from

z> p

t. Thenwecan bound(6.4) above bytwice

Z p

t

0 (t

2

) =2

dz+ Z

1

p

t (tz

2

) =2

n(z)dz=t 1

2

+t =2

Z

1

p

t z

n(z)dz:

Furthermore, weseparate thelastintegralat 1. We get

Z

1

p

t z

n(z)dz c

+

Z

1

p

t z

dz:

The laterintegralis oforder t (1 )=2

. Substitutingthese resultsinto (6.3),wesee thatthe

expected energy is nite when < 3=2. The claim now follows form Frostman's Energy

(20)

7. On nowhere dierentiability

Levy (1954) askswhether itis truethat

P[ 8tD

B(t)2f1g]=1?

The followingpropositiongivesanegative answerto thisquestion.

Proposition 7.1. A.s there is an uncountable set of times t at which the upper right

derivative D

B(t) iszero.

We sketch a proof below. Stronger and more general results can be found in Barlow

and Perkins(1984).

(SKETCH). Put

I=

B(1); sup

0s1 B(s)

;

and denea functiong:I ![0;1]bysetting

g(x)=supfs2[0;1]:B(s)=xg:

It is easy to check that a.s. the interval I is non-degenerate, g is strictly decreasing, left

continuous and satises B(g(x)) = x. Furthermore, a.s. the set of discontinuitiesof g is

dense in I since a.s. B has no interval of monotonicity. We restrict our attention to the

event of probability1 on which these assertionshold. Let

V

n

=fx2I :g(x h) g(x)>nh forsome h2(0;n 1

)g:

Sincegisleftcontinuousandstrictlydeceasing,onereadilyveriesthatV

n

isopen;itisalso

denseinI aseverypointofdiscontinuityofg isalimitfromtheleftofpointsofV

n

:Bythe

Bairecategorytheorem,V := T

n V

n

isuncountableanddenseinI:Nowifx2V thenthere

is a sequence x

n

"x such that g(x

n

) g(x) >n(x x

n

):Setting t=g(x) and t

n =g(x

n )

we have t

n

#tand t

n

t>n(B(t) B(t

n

));fromwhichitfollowsthatD

B(t)0:Onthe

other handD

B(t)0 sinceB(s)B(t)forall s2(t;1);bydenitionof t=g(x):

Exercise 7.2. Letf 2C([0;1]):ProvethatB(t)+f(t)isnowheredierentiablealmost

surely.

Isthe\typical"functioninC([0;1]) nowhere dierentiable? Itis aneasy applicationof

theBaire category theoremto show thatnowheredierentiabilityisa genericpropertyfor

C([0;1]): This result leaves something to be desired, perhaps, as topological and measure

theoreticnotions ofa \large"set neednotcoincide. Forexample,the setof pointsin[0;1]

whosebinaryexpansionhaszeroswithasymptoticfrequency1=2 isameagerset, yetithas

Lebesgue measure 1. We consider a related idea proposed by Christensen (1972) and by

Hunt, Sauer and Yorke (1992). Let X be a separable Banach space. Say that A X is

prevalent if there exists a Borel probabilitymeasure on X such that (x+A)=1 for

every x2X. A set iscallednegligible ifits complement isprevalent.

Proposition 7.3. If A

1 ;A

2

;::: are negligible subsets of X then S

i1 A

i

(21)

negli-Proof. Foreachi1let

A

i

beaBorelprobabilitymeasuresatisfying

A

i (x+A

i )=0

for all x 2 X: Using separability we can nd for each i a ball D

i

of radius 2 i

centered

at x

i

2 X with

A

i (D

i

) > 0: Dene probabilitymeasures

i

; i 1; by setting

i (E) =

A

i

(E +x

i jD

i

) for each Borel set E; so that

i

(x+A

i

) = 0 for all x and for all i: Let

(Y

i

;i 0) be a sequence of independent random variables with dist(Y

i ) =

i

: For all i

we have

i [jY

i j 2

i

] = 1. Therefore, S = P

i Y

i

converges almost surely. Writing

for the distribution of S and putting

j

= dist(S Y

j

); we have =

j

j

; and hence

(x+A

j )=

j

j (x+A

j

)=0forallx andforallj:Thus(x+[

i1 A

i

)=0forallx:

Proposition 7.4. A subsetA of R d

is negligiblei L

d

(A)=0:

Proof. ()) Assume A is negligible. Let

A

be a (Borel) probability measure such

that

A

(x+A)=0 for all x 2 X:Since L

d

A =L

d

(indeed L

d

=L

d

for any Borel

probabilitymeasure on R d

)wehave 0=L

d

A

(x+A)=L

d

(x+A)forall x2X:

(() If L

d

(A)= 0 then therestriction of L

d

to the unit cube is a probability measure

which vanisheson every translateof A:

Remark. It follows from Exercise 7.2 that the set of nowhere dierentiable functions

isprevalentin C([0;1]):

8. Strong Markov property and the reection principle

For each t 0 let F

0

(t)=fB(s) :stg be the smallest-eld making every B(s);

st;measurable,and setF

+

(t)=\

u>t F

0

(u)(theright-continuousltration). Itisknown

(see, for example, Durrett (1996), theorem 7.2.4) that F

0

(t) and F

+

(t) have the same

completion. A ltration fF(t)g

t0

is a Brownian ltration if for all t 0 the process

fB(t+s)g

s0

isindependentofF(t)andF(t)F

0

(t):Arandomvariable isastopping

time fora Brownianltration fF(t)g

t0

iff tg2F(t) forall t: For any randomtime

we denethepre- -eld

F():=fA:8tA\f tg2F(t)g:

Proposition 8.1. (Markov property) For every t0 the process

fB(t+s) B(t)g

s0

is standard Brownian motionindependent of F

0

(t) andF

+ (t):

It is evident from independence of increments that fB(t+s) B(t)g

s0

is standard

Brownian motionindependent of F

0

(t): That thisprocess is independent of F

+

(t) follows

from continuity;see, e.g.,Durrett (1996, 7.2.1)for details.

Themainresultof thissectionisthestrongMarkov propertyforBrownianmotion,

establishedindependentlybyHunt (1956) and Dynkin(1957):

Theorem 8.2. Supposethat isastoppingtime forthe Brownian ltration fF(t)g

t0 .

Then fB(+s) B()g

s0

is Brownian motion independent of F():

Sketch of Proof. Supposerstthat isanintegervaluedstoppingtimewithrespect

to aBrownianltrationfF(t)g

t0

:Foreach integer j theevent f =jgisinF(j) andthe

process fB(t+j) B(j)g

t0

(22)

" > 0; and approximating by such discrete stopping times gives the conclusion in the

generalcase. See, e.g., Durrett (1996,7.3.7) formore details.

Oneimportant consequenceof thestrongMarkov propertyis thefollowing:

Theorem 8.3 (Reection Principle). If is a stopping time then

B

(t):=B(t)1

(t)

+(2B() B(t))1

(t>)

(Brownian motion reected attime ) isalso standard Brownian motion.

Proof. We shalluse anelementaryfact:

Lemma 8.4. LetX;Y;Z berandom variables withX;Y independent andX;Z

indepen-dent. If Y d

Z then (X;Y) d

(X;Z):

The strong Markov property states that fB( +t) B()g

t0

is Brownian motion

independentofF();andbysymmetrythisisalsotrueoff (B(+t) B())g

t0

:Wesee

from thelemma that

(fB(t)g

0t

;fB(t+) B()g

t0 )

d

(fB(t)g

0t

;f(B() B(t+))g

t0 );

and thereectionprinciplefollows immediately.

Remark. Consider = infft : B(t) = max

0s1

B(s)g: Almost surely fB( +t)

B()g

t0

isnon-positive on some right neighborhoodof t=0, and hence is not Brownian

motion. ThestrongMarkovpropertydoesnotapplyherebecause is notastoppingtime

for any Brownian ltration. We willlater see that Brownian motionalmost surelyhas no

point of increase. Since is a point of increase of the reected process fB

(t)g; it follows

thatthe distributionsof Brownianmotionand of fB

(t)gare singular.

Exercise 8.5. Prove thatifA isa closedset then

A

isa stoppingtime.

Solution. fAtg= T

n1 S

s2[0;t]\Q

fdist (B(s);A) 1

n

g2F0(t):

More generally, if A is a Borel set then the hitting time

A

is a stopping time (see Bass

(1995)).

SetM(t)= max

0st

B(s):Our next resultsays M(t) d

jB(t)j:

Theorem 8.6. If a>0 then P[ M(t)>a]=2P[ B(t)>a]:

Proof. Set

a

=minft0:B(t)=agand let fB

(t)g be Brownianmotion reected

at

a

: Then fM(t) > ag is the disjoint union of the events fB(t) > ag and fM(t) >

a;B(t)ag; and sincefM(t)>a;B(t)ag=fB

(t)agthe desiredconclusion follows

immediatelyfrom thestrongMarkov property.

9. Local extrema of Brownian motion

Proposition 9.1. Almost surely,every local maximum of Brownian motionis a strict

(23)

Lemma 9.2. Given two disjoint closed time intervals, the maxima of Brownian motion

on them are dierent almost surely.

Proof. Fori=1;2let[a

i ;b

i ],m

i

,denotethelower,respectivelythehigherinterval,and

thecorrespondingmaximumof Brownian motion. Note thatB(a

2

) B(b

1

) isindependent

ofthepairm

1 B(b

1

)andm

2 B(a

2

). Conditioningonthevaluesoftherandomvariables

m

1 B(b

1

) and m

2 B(a

2

),theevent m

1 =m

2

can be writtenas

B(a

2

) B(b

1 )=m

1 B(b

1 ) (m

2 B(a

2 )):

Theleft handsidebeingacontinuousrandomvariable,and theright handsideaconstant,

we see thatthisevent hasprobability0.

We nowprove Proposition9.1.

Proof. Thestatementofthelemmaholdsjointlyforalldisjointpairsofintervalswith

rationalendpoints. Thepropositionfollows,sinceifBrownianmotionhadanon-strictlocal

maximum, thenthere were two disjoint rationalintervals where Brownian motion hasthe

same maximum.

Corollary 9.3. Theset M of times where Brownian motion assumes its local

maxi-mum is countable and dense almost surely.

Proof. Consider the function from the set of non-degenerate closed intervals with

rationalendpointsto R given by

[a;b]7!inf

ta:B(t)= max

asb B(s)

:

TheimageofthismapcontainsthesetM almostsurelybythelemma. ThisshowsthatM

iscountable almost surely. We already knowthatB hasnointervalofincreaseordecrease

almost surely. It follows that B almost surelyhasa localmaximum inevery intervalwith

rationalendpoints,implyingthecorollary.

10. Area of planar Brownian motion paths

We have seenthattheimageofBrownianmotionis always2dimensional,soone might

askwhat its 2dimensionalHausdormeasure is. Itturnsouttobe0 inalldimensions;we

willprove itforthe planarcase. We willneedthefollowinglemma.

Lemma 10.1. If A

1 ;A

2 R

2

are Borelsets withpositivearea then

L

2

(fx2R 2

:L

2 (A

1 \(A

2

+x))>0g)>0:

Proof. One proof of this fact relies on (outer) regularity of Lebesgue measure. The

proof belowis more streamlined.

We may assumeA

1

and A

2

arebounded. By Fubini'stheorem,

Z

R 2

1

A1 1

A2

(x)dx= Z

R 2

Z

R 2

1

A1 (w)1

A2

(w x)dwdx

= Z

R 2

1

A

1 (w)

Z

R 2

1

A

2

(w x)dx

dw

(24)

Thus1

A

1 1

A

2

(x)>0onasetofpositivearea. But1

A

1 1

A

2

(x)=0unlessA

1 \(A

2 +x)

haspositive area, sothisprovesthelemma.

ThroughoutthissectionB denoteplanarBrownianmotion. Wearenowreadytoprove

Levy's theorem on theareaof its image.

Theorem 10.2 (Levy). Almostsurely L

2

(B[0;1])=0:

Proof. LetX denote theareaofB[0;1], and M be itsexpectedvalue. Firstwecheck

thatM <1. If a1 then

P[X >a]2P[jW(t)j> p

a=2forsome t2[0;1]]8e a=8

whereW isstandardone-dimensionalBrownianmotion. Thus

M = Z

1

0

P[X >a]da8 Z

1

0 e

a=8

da+1<1:

Note thatB(3t)and p

3B(t)have thesame distribution,and hence

EL

2

(B[0;3])=3EL

2

(B[0;1])=3M:

Note that we have L

2

(B[0;3]) P

2

j=0 L

2

(B[j;j+1]) with equality if and only if for0

i<j 2 we have L

2

(B[i;i+1]\B[j;j+1])= 0. On the other hand, forj = 0;1;2; we

have EL

2

(B[j;j+1])=M and

3M =EL

2

(B[0;3]) 2

X

j=0 EL

2

(B[j;j+1])=3M;

whence the intersection of any two of the B[j;j+1] has measure zero almost surely. In

particular,L

2

(B[0;1]\B[2;3])=0 almost surely.

LetR (x)denotetheareaofB[0;1]\(x+B[2;3] B[2]+B(1)). Ifwe conditiononthe

valuesofB[0;1];B[2;3] B(2),theninordertoevaluatetheexpectedvalueofB[0;1] \ B[2;3]

we shouldintegrate R (x)wherex hasthedistributionof B(2) B(1). Thus

0=E[L

2

(B[0;1]\B[2;3])]=(2) 1

Z

R 2

e jxj

2

=2

E[R(x)]dx

where we are averaging with respect to the Gaussian distribution of B(2) B(1). Thus

R (x)=0 a.s. forL

2

-almost all x,or, byFubini, the areaof the setwhere R (x)is positive

isa.s. zero. Fromthe lemmaweget thata.s.

L

2

(B[0;1])=0 or L

2

(B[2;3])=0:

TheobservationthatL

2

(B[0;1])andL

2

(B[2;3])areidenticallydistributedandindependent

completes theproof thatL

2

(B[0;1])=0almost surely.

11. Zeros of the Brownian motion

Inthissection,westartthestudyofthepropertiesofthezerosetZ

B

ofonedimensional

Brownianmotion. Wewillprove thatthissetisan uncountable closedsetwithnoisolated

points. This is, perhaps, surprising since, almost surely, a Brownian motion has isolated

zeros from the left (for instance, the rst zero after 1=2) or from the right (the last zero

(25)

Theorem 11.1. Let B bea one dimensional Brownian motion and Z

B

beits zero set,

i.e.,

Z

B

=ft2[0;+1):B(t)=0g:

Then, a.s., Z

B

isan uncountable closed set withno isolated points.

Proof. Clearly, with probability one, Z

B

is closed because B is continuous a.s.. To

provethatnopointofZ

B

isisolatedweconsiderthefollowingconstruction: foreachrational

q 2[0;1)considertherst zero after q,i.e.,

q

=infft>q:B(t)=0g. Notethat

q <1

a.s. and, sinceZ

B

is closed,the inf is a.s. a minimum. Bythe strongMarkovpropertywe

havethatforeachq,a.s.

q

isnotanisolatedzerofromtheright. But,sincethere areonly

countablymanyrationals,weconcludethata.s.,forall q rational,

q

isnotanisolatedzero

from theright. Our next taskis to prove that theremaining pointsof Z

B

arenotisolated

from theleft. So we claimthat any0<t2Z

B

whichis dierent from

q

forall rationalq

is notan isolatedpoint from theleft. To see thistake a sequence q

n

"t. Dene t

n =

q

n .

Clearlyq

n t

n

<t(ast

n

isnotisolatedfromtheright)andsot

n

"t. Thustisnotisolated

from theleft.

Finally, recall (see, for instance, Hewitt-Stromberg, 1965) that a closed set with no

isolatedpointsis uncountable and thisnishestheproof.

11.1. General Markov Processes. In this section, we dene general Markov

pro-cesses. ThenweprovethatBrownianmotion,reectedBrownianmotionandaprocessthat

involvesthe maximumof Brownianmotion areMarkov processes.

Definition 11.2. A function p(t;x;A), p : R R d

B ! R, where B is the Borel

-algebra inR d

,isa Markov transitionkernel provided

1. p(;;A)is measurable asa functionof (t;x), foreach A2B,

2. p(t;x;) isa Borelprobabilitymeasure forall t2R and x2R d

,

3. 8A2B; x2R d

and t;s>0,

p(t+s;x;A)= Z

R d

p(t;y;A)p(s;x;dy):

Definition 11.3. A process fX(t)g isaMarkov process withkernelp(t;x;A)ifforall

t>sand Borelset A2B we have

P(X(t)2AjF

s

)=p(t s;X(s);A);

whereF

s

=(X(u); us).

The next two examples aretrivial consequences of the Markov Property for Brownian

motion.

Example 11.4. A d-dimensionalBrownian motion isa Markov process and its

transi-tion kernelp(t;x;) hasN(x;t) distributionineach component.

SupposeZ hasN(x;t) distribution. Dene jN(x;t)j to bethedistributionof jZj.

(26)

pro-Theorem 11.6 (Levy, 1948). Let M(t) be the maximum process of a one dimensional

Brownian motion B(t),i.e. M(t)=max

0st

B(s). Then, the process Y(t)=M(t) B(t)

is Markov and its transition kernel p(t;x;) has jN(x;t)j distribution.

Proof. For t > 0, consider the two processes ^

B(t) = B(s+t) B(s) and ^

M(t) =

max

0ut ^

B(u). Dene F

B

(s) = ( B(t);0ts). To prove the theorem it suÆces to

checkthat conditionalon F

B

(s)and Y(s)=y,we have Y(s+t) d

=jy+ ^

B(t)j.

To prove theclaim notethat M(s+t)=M(s)_(B(s)+ ^

M(t)), and sowe have

Y(s+t)=M(s)_(B(s)+ ^

M(t)) (B(s)+ ^

B(t)):

Usingthefactthat a_b c=(a c)_(b c),we have

Y(s+t)=Y(s)_ ^

M(t) ^

B(t):

To nish, it suÆces to check, for every y0, that y_ ^

M(t) ^

B(t) d

=jy+ ^

B(t)j. For any

a0 write

P(y_ ^

M(t) ^

B(t)>a)=I+II;

whereI =P(y ^

B(t)>a) and

II =P(y ^

B(t)aand ^

M(t) ^

B(t)>a):

Since ^

B d

= ^

B wehave

I =P(y+ ^

B(t)>a):

To studythesecond termis usefulto denethe"time reversed" Brownianmotion

W(u)= ^

B(t u) ^

B(t);

for0ut. Note that W is also a Brownian motion for0u t sinceit iscontinuous

and its nitedimensionaldistributionsare Gaussianwiththeright covariances.

Let M

W

(t)=max

0ut

W(u). Then M

W (t)=

^

M(t) ^

B(t). Since W(t)= ^

B(t), we

have:

II =P(y+W(t)aand M

W

(t)>a):

IfweusethereectionprinciplebyreectingW(u)atthersttimeithitsawegetanother

BrownianmotionW

(u). IntermsofthisBrownianmotionwehaveII =P(W

(t)a+y).

Since W

d

= ^

B, it follows II = P(y + ^

B(t) a). The Brownian motion ^

B(t) has

continuousdistribution,and so,byadding I and II,weget

P(y_ ^

M(t) ^

B(t)>a)=P(jy+ ^

B(t)j>a):

Thisprovestheclaim and,consequently,thetheorem.

Proposition 11.7. TwoMarkov processes in R d

withcontinuouspaths, withthe same

initial distribution andtransition kernel are identical in law.

Outline of Proof. The nitedimensionaldistributionsare thesame. From thiswe

deducethattherestrictionofbothprocessesto rationaltimesagreeindistribution. Finally

(27)

SincetheprocessY(t)iscontinuousandhasthesame distributionasjB(t)j(theyhave

the same Markov transition kernel and same initial distribution) this proposition implies

fY(t)g d

=fjB(t)jg.

11.2. Hausdor dimension of Z

B

. We already know that Z

B

is an uncountable

set with no isolated points. In this section, we will prove that, with probabilityone, the

Hausdor dimension of Z

B

is 1=2. It turns out that it is relatively easy to bound from

bellow the dimension of the zero set of Y(t) (also known as set of record values of B).

Then,bytheresultsinthelastsection,thisdimensionmustbethesame ofZ

B

sincethese

two (random) sets havethe same distribution.

Definition 11.8. A time t is a recordtime for B ifY(t)=M(t) B(t)=0, i.e., ift

isa globalmaximumfromthe left.

The next lemma givesa lower bound on the Hausdor dimension of the set of record

times.

Lemma 11.9. With probability 1, dimft2[0;1]:Y(t)=0g1=2.

Proof. SinceM(t)isanincreasingfunction,wecanregarditasadistributionfunction

ofameasure,with(a;b]=M(b) M(a). Thismeasureissupportedonthesetofrecord

times. Weknowthat, withprobabilityone,theBrownianmotionisHoldercontinuouswith

anyexponent<1=2. Thus

M(b) M(a) max

0hb a

B(a+h) B(a)C

(b a)

;

where < 1=2 and C

is some random constant that doesn't depend on a or b. By the

Mass DistributionPrinciple,we get that, a.s., dimft2[0;1]:Y(t)=0g. By choosing

a sequence

n

"1=2 we nishthe proof.

RecallthattheupperMinkowskidimensionofasetisanupperboundfortheHausdor

dimension. To estimatethe Minkowskidimensionof Z

B

we willneedto know

P(9t2(a;a+):B(t)=0): (11.1)

Thisprobabilitycan becomputed explicitlyand we willleave thisasan exercise.

Exercise 11.10. Compute(11.1).

Solution. ConditionalonB(a)=x>0wehave

P(9t2(a;a+):B(t)=0jB(a)=x)=P( min

ata+

B(t)<0jB(a)=x):

Buttherighthandsideisequalto

P(max

0<t<

B(t)>x)=2P(B()>x);

usingreectionprinciple.

Byconsideringalsothecase wherexisnegativeweget

P(9t2(a;a+):B(t)=0)=4 Z

1

0 Z

1

x e

y 2

2 x

2

2a

2 p

a dydx:

Computingthislastintegral explicitly,weget

P(9t2(a;a+):B(t)=0)= 2

arctan r

(28)

However, forour purposes,thefollowingestimate willsuÆce.

Lemma 11.11. For any a;>0 we have

P(9t2(a;a+):B(t)=0)C r

a+ ;

for some appropriate positive constant C.

Proof. ConsidertheeventA givenbyjB(a+)j p

. Bythescaling propertyofthe

Brownianmotion,we can give theupperbound

P(A)=P

jB(1)j r

a+

2 r

a+

: (11.2)

However, knowing that Brownian motion has a zero in (a;a+) makes the event A very

likely. Indeed, we certainly have

P(A)P(A and 02B[a;a+]);

and thestrongMarkov propertyimpliesthat

P(A)~c P(02B[a;a+]); (11.3)

where

~

c= min

ata+

P(AjB(t)=0):

Because theminimumisachieved when t=ahave

~

c=P(jB(1)j1)>0;

byusingthescalingpropertyof theBrownianmotion.

From inequalities(11.2) and (11.3), we conclude

P(02B[a;a+]) 2

~ c

r

a+ :

Forany,possiblyrandom, closedset A[0;1]denea function

N

m (A)=

2 m

X

k=1 1

fA\[ k 1

2 m

; k

2 m

]6=;g :

Thisfunctioncountsthenumberofintervals oftheform [ k 1

2 m

; k

2 m

]intersectedbytheset A

and soisanaturalobjectifwe wantto computeMinkowskidimension. In thespecialcase

whereA=Z

B

we have

N

m (Z

B )=

2 m

X

k=1 1

f02B[ k 1

2 m

; k

2 m

]g :

The next lemma shows that estimates on the expected value of N

m

(A) will give us

boundson theMinkowskidimension (andhenceon theHausdor dimension).

Lemma 11.12. Suppose A is a closed random subset of [0;1] such that

EN

m

(A)c2 m

;

(29)

Proof. Consider

E 1

X

m=1 N

m (A)

2 m(+)

;

for>0. Then, bythe monotoneconvergence theorem,

E 1

X

m=1 N

m (A)

2 m(+)

= 1

X

m=1 EN

m (A)

2 m(+)

<1:

Thisestimate impliesthat

1

X

m=1 N

m (A)

2 m(+)

<1 a.s.;

and so, withprobabilityone,

limsup

m!1 N

m (A)

2 m(+)

=0:

From thelastequation follows

dim

M

(A)+; a.s. :

Let!0 throughsome countable sequence to get

dim

M

(A); a.s. :

Andthiscompletesthe proofof thelemma.

To getan upperboundon theHausdor dimensionof Z

B

note that

EN

m (Z

B )C

2 m

X

k=1 1

p

k

~

C2 m=2

;

sinceP 9t2

k 1

2 m

; k

2 m

:B(t)=0

C

p

k

. Thus, by thelast lemma, dim

M (Z

B

) 1=2 a.s.

Thisimpliesimmediatelydim

H (Z

B

)1=2 a.s. Combiningthisestimatewith Lemma11.9

we have

Theorem 11.13. With probability one we have

dim

H (Z

B )=

1

2 :

From this proof we can also inferthat H 1=2

(Z

B

)<1 a.s. Later in thecourse we will

prove that H 1=2

(Z

B

) =0. However, it is possibleto denea more sophisticated Hausdor

measure forwhich,withprobabilityone, 0<H (Z

B )<1.

12. Harris' Inequality and its consequences

We beginthissection byprovingHarris'inequality.

Lemma 12.1 (Harris' inequality). Suppose that

1 ;::: ;

d

are Borel probability

mea-sures on R and=

1

2

:::

d

. Let f;g:R d

!R bemeasurablefunctions that are

nondecreasing in each coordinate. Then,

Z

d

f(x)g(x)d Z

d

f(x)d

Z

d

g(x)d

References

Related documents