Progress in Probability, Vol.

57, 135-150
© 2004 Birkhiiuser Verlag Basel/Switzerland

Intersection Exponents and the Multifractal

Spectrum for Measures on Brownian Paths
Peter Morters

Abstract. The aim of this paper is to give a survey of recent developments

in the multifractal analysis of measures arising in the study of Brownian mo-
tion. We first discuss the harmonic measure on a Brownian path, studied by
Lawler, and the intersection local time on the intersection of two independent
Brownian paths, studied by Klenke and Morters. In both examples the inter-
section exponents of Brownian motion play a crucial role in the multifractal
analysis. Whereas the first example is in line with the multifractal formal-
ism, the second example is not and more subtle effects come into the picture.
Then we discuss the occupation measure of Brownian motion, in which case
a refined multifractal analysis was carried out by Dembo, Peres, Rosen and
Zeitouni.
Mathematics Subject Classification (2000). Primary 60J65; Secondary 28A80.
Keywords. Brownian motion, occupation measure, harmonic measure, inter-
section local time, critical exponent, intersection exponent, multifractal spec-
trum, multifractal formalism, exceptional set, large deviation heuristic.

1. Introduction
An interesting fractal geometric characteristic for discrete and continuous models
of statistical physics is the multifractal spectrum, which originated in the study
of turbulence models, see e.g. [33]. Loosely speaking, the multifractal spectrum
evaluates the degree of variation in the local intensity of a spatial distribution,
or, in other words, it describes quantitatively how frequent a certain type of local
irregularity appears in a fractal measure, set or function.
There are many different ways in which this concept can be made precise,
see for example the contribution of Levy-Vehel in this volume. In this paper I
would like to focus on the multifractal spectrum as discussed e.g. in [43, 13, 40, 4],
where irregularities are described in terms of local dimensions and the size of
sets is merusured by the Hausdorff dimension. There is a host of literature on this
136 Peter Marters

particular type of spectrum spanning almost two decades, the reader is referred to
the survey [35] and the extensive list of references given there.
To give the definition of this type of multifractal spectrum, suppose that f. L
is a (fractal) measure on JRd. One can define the local dimension of the measure f. L
at the point x E JRd as
. 10gf..L(B(x, r))
11m -a (1.1)
rlO logr -,
if the limit exists. In words, this is the rate of decay of the measure f..L(B(x, r)) as
the concentric balls B(x, r) of radius r 1 0 zoom into the point x. The value f(a)
of the multifractal spectrum is the Hausdorff dimension of the set of points x with
local dimension a, i.e.
· {
d ImxElN.:lm
TTbd 1· log f..L(B(x, r)) } (1.2)
1 =a.
rlO ogr
In many cases of interest, the limit in (1.1) has to be replaced by liminf or lim-
sup to obtain an interesting nontrivial spectrum. A measure is called multifmctal
if f(a) > 0 for more than one value of a E JR, in any of these definitions.
Whereas much of the earlier literature on multifractal theory was focused on
the analysis of (mostly deterministic) objects with strong features of self-similarity,
and the verification of the multifractal formalism proposed, e.g. in [13], the interest
has shifted more recently towards the analysis of stochastic processes by means of
multifractal spectra. This trend can be seen in applied fields, like the analysis of
time series, see for example [15, 42], as well as in mathematical physics, see [10]
for a survey from a physics point of view. In the probability context new patterns
have emerged and a lot of interest is going towards measures where the multifractal
formalism fails, and towards refinements of the multifractal formalism.
In the present paper we shall review these developments by considering three
natural measures on the path of a Brownian motion in dimension d = 3. 1 These
examples are
• the harmonic measure on a Brownian path, studied by Lawler [24, 25, 26],
• the intersection local times on the intersection of two independent Brownian
path, studied by Klenke and Marters [18],
• the occupation measure on a Brownian path, studied by Dembo, Peres, Rosen
and Zeitouni [6, 7], see [5] for an introduction.
The main emphasis will be on the first two examples, as they give us also
the opportunity to highlight another important subject, which has emerged in
recent years in fractal geometry, namely the relation of Hausdorff dimension and
the critical exponents of statistical physics, in this case the intersection exponents.
We start the paper with an introduction of intersection exponents in the next
section, followed by three sections discussing our three examples. A final section
tries to wrap up the developments described in this paper, hinting also at some
expectations for future developments.

lThe focus is on d = 3 for better comparison, though some results extend to other dimensions.
 Measures on Brownian Paths 137

2. Intersection exponents
The intersection exponents for random curves measure how likely these curves do
not intersect, and maybe should therefore be called non-intersection exponents.
They contain crucial information about the fractal geometry of the curves, and
substantial recent research is devoted to exploiting this information.
To define the intersection exponents for Brownian motion in d = 2,3, suppose
n, m ~ 1 are integers and let WI, ... , Wn +m are independent Brownian motions
in lR. d started uniformly on 8B(0, 1) and running up to the first exit times Ti(r)
from a large ball B(O,r). We divide the motions into two packets and look at the
union of the paths in each family,
Slh(r) = WI [0, TI(r)] U ... U Wn[O, Tn(r)] and
~2(r) = W n+1 [0, Tn+l(r)] U ... U Wn+m[O, Tn+m(r)].
The event that the two packets of Brownian paths fail to intersect has a decreasing
probability as r i 00. Indeed, it is easy, using subadditivity, to show that there
°
exists a constant < ~d(n, m) < 00 such that
lP{~I(r) n ~2(r) = 0} = r-~d(n,m)+o(1), as ri 00. (2.1)
The numbers ~d(n, m) are called the intersection exponents. It is not hard to
see that this definition is very robust under changes in the way the motions are
started, for example if one starts all motions in a packet in the same fixed point
(2.1) continues to hold if the starting points of the packets do not agree.

°
In fact, the approximation of non-intersection probabilities is stronger than
(2.1) suggests. Lawler [23] has shown that there exists constants < c < C < 00
such that
cr-Mn,m) :s: lP{~!h(r) n ~2(r) = 0} :s: Cr-~d(n,m) for all r > l.
The definition of intersection exponents can be extended naturally to one real
argument. For this purpose define the random variable
Zn(r) = lP{~I(r) n Wn+I[O, Tn+l(r)] = 01 WI' ... ' Wn }.
Again one can show that, for every >. > 0, there exists a constant ~d(n, >.) such
that
IE{Zn(r)A} = r-~d(n,A)+o(1), as r i 00. (2.2)

°
Clearly, this is consistent with the definition of ~d(n, >.) in the case of an integer >..
Finally, a natural way to extend the definition to the case>. = is by letting
lP{Zn(r) > O} = r-~d(n,O)+o(1), as r i 00.

°
Note that Zn (r) = is the event that the packet of n Brownian motion disconnects
the origin from infinity, so that any further Brownian motion cannot reach 8B(0, r)
without intersecting the packet. Therefore the exponents ~d(n, 0) are also called
disconnection exponents.
The first researchers interested in the values of the intersection exponents
were physicists. They have made the striking conjecture that in the plane many
138 Peter Morters

of these exponents are rational numbers, see e.g. Duplantier and Kwon [11] and
[10] for a survey. Very recently, Lawler, Schramm and Werner, in a seminal series
of papers [28, 29, 30] have been able to verify this rigorously, see [27] for a survey
over the series. They have shown that
(v24n + 1 + V240X + 1 - 2)2 - 4
6(n, oX) = 48 ' forn E N,oX ~ O. (2.3)
The proof of (2.3) is based on conformal invariance, and the stochastic Lowner
evolution (SLE). I shall not go further into the exciting recent developments about
SLE, as our examples are in d = 3, where no analogue of the stochastic Lowner
evolution exists.
Indeed, in d = 3 we do not know the values of the intersection exponents
explicitly, and it is very likely that they are typically irrational. I briefly review
what is currently known about the values of the intersection exponents ~d(n, oX).
The two values explicitly known in dimension d = 3 are
• 6(1,2) = 6(2,1) = 1. This follows from properties of the harmonic measure,
and an intuitive argument for this will be a spin-off of the next section. For
a rigorous and accessible argument see [20, Chapter 3] .
• 6(n,0) = 0 as no disconnection is possible.
Estimates for other values can be derived from this using (2.2), monotonicity
and concavity of oX f-4 6 (n, oX). A deep result of Lawler [23] shows that this function
is even strictly concave. As an example we get for 6(2,2) the estimates
6(2,2) > 6(2,1) = 1 by monotonicity and strict concavity,
and
6(2,2) < 26(2,1) = 2 by strict concavity.
The project of exploiting intersection exponents as a means to describe the fractal
geometry of Brownian motion and random walks was pioneered by Lawler [21, 22].
To name two examples, the set of cut points of a Brownian path B[O, 1] c ]Rd is the
set of points x = B(t) E B[O, 1] such that B[O, t) n B(t, 1] = 0. Lawler [21] showed
that the Hausdorff dimension of this set is 2 - ~d(1, 1). Intersection exponents
are also crucial in the solution of the famous Mandelbrot conjecture about the
Brownian frontier, the boundary of the unbounded component of the complement
of a planar Brownian curve. Lawler showed in [22] that the dimension of this set
is 2 - 6(2,0) and together with the value of this exponent obtained 5 years later
this shows that the dimension of the Brownian frontier is 4/3, as predicted by
Mandelbrot in 1983. Our first example below can be seen as a multifractal version
of this result (however presented here in d = 3).

3. The multifractal spectrum. of harmonic measure

Suppose A C ]R3 is a compact, nonpolar set. The harmonic measure on the set
A is defined as follows. Suppose B(y, R) is a ball such that A is contained inside
this ball. Let W : [0,00) ---+ ]Rd be a Brownian motion started in a point chosen
 Measures on Brownian Paths 139

uniformly from the sphere 8B(y, R) and let

T = inf{t > °:
Wt E A}
be the first hitting time of the set A by a Brownian motion. The law of W(T) can
be shown to be independent of the choice of B(y, R). The harmonic measure hA
on A is defined by
hA(B) = JP>{WT E B, T < oo} for all Borel sets B c A.
Note that the support of the harmonic measure is contained in the boundary of
the unbounded components of AC, called the frontier of A.
Now we look at A = B[0,1] C lR d , the path of a Brownian motion and
study the harmonic measure h = hA on this set. The following result about the
multifractal spectrum of the measure hA is due to Lawler [25, 26].
Theorem 3.1 (Lawler 1999). Define a convex function e as the Legendre transform
of the the strictly concave function A f--t 6(2, A), i.e.
C(a) = sup {6(2,A) - aA}.
A>O
Then, for every a such that 2 - e(a - 1) > 0, almost surely there exist points
xEAwith
lim logh(B(x,r)) = a.
r!O logr
Moreover, for these values of a, almost surely,
dim {x E A : lim log h(B(x, r)) = a} = 2 - C(a - 1).
r!O logr
Remark. The fact that this spectrum is nontrivial is due to the non-mean field
nature of the intersections, expressed in the strict concavity of 6(2, . ). 0

Following [26] I now give a very rough heuristic argument for Theorem 3.1,
which is based on the so-called multifractal formalism of [13]. Let

A(j,n) = B[j ~1, ~J, for 1:::; j:::; e 2n .

e e
Consider the quantity

Q(n,q) = Lh(A(j,n))q.
j=1
This is a random variable, which is intimately related to the multifractal spectrum.
Indeed, for large n, we can roughly assume that A(j, n) is the intersection of the
Brownian path A with a ball of radius e- n and if the local dimension of the points
in this ball is approximately a, then h(A(j,n)) ~ e- an . Hence, if we denote by
le(a) = {1 : :; j :::; e2n : h(A(j, n)) ~ e- an },
we can write
Q(n,q) = L#le(a)e- qan ,
a
140 Peter Morters

where the sum extends over an appropriate discretisation of the reals. The Laplace
method suggests that, for
. 1
f(a) = hm -log (#rt(a))
n
n~(X)

we have
Q(n,q) ~ exp (nsup{f(a) - qa}).
a
Suppose now that we can show that, for some function T,

Q(n, q) ~ exp(m(q)). (3.1)

Then we get,
sup {J(a) - qa} = T(q).
a

As the sets in rt( a) roughly cover the set of points of local dimension a, the quan-
tity f(a) is at least an upper bound for the dimension spectrum. If everything is
sufficiently regular, this bound is sharp. To identify f note that
T(q) = sup {J(a) - qa} = f*(q).
a

If f is concave, we may invert this Legendre transform and get

f(a) = T*(a) = inf {T(q) + qa}. (3.2)
q>O

Hence it boils down to finding the T in (3.1).

Here the intersection exponents are coming into play. We sketch the argument
for the upper bound for T, which is based on the first moment method. First fix
the set A = B[O, 1] and look at the harmonic measure of a small ball,
hA(B(x,r)) = P{W hits A first in B(x,r)} for x E A.
The probability that the path Whits B(x, r) is cr d - 2 = cr for some constant
c> O. Given that Whits B(x,r), we can look at the time-reversed path starting
on the boundary 8B(x, r) and stopped upon hitting 8B(y, R). Though this path
is strictly speaking no Brownian path, a sufficiently large part of it is equal to
a Brownian path conditioned to exit the annulus B(y, R) \ B(x, r) at 8B(y, R).
This is good enough for our purpose, and hence the reversed path heuristically
makes the first packet B~ (r) consisting of one Brownian motion. Considering A as
a random set, its distribution given x E A essentially coincides with a two-sided
Brownian motion started in x, forming the second packet B& (r) of two independent
Brownian motions. The Brownian motions in these packets are started on the
sphere 8B(x, r) and stopped when they leave the fixed sphere 8B(y, R), and by
Brownian scaling we easily map this onto the situation in the definition of the
intersection exponents, where motions are started on the unit sphere and stopped
upon leaving a large sphere of radius R/r. Hence we have
1E{ hA(B(x, r))} = cr p{ B~ (r) n B;(r) = 0} = r1+~3(2,1)+o(1),
 Measures on Brownian Paths 141

and more generally,

(3.3)
This gives
IEQ(n,q) r::d exp (n(2 - q - 6(2,q) + 0(1))).
By the Borel-Cantelli lemma we get from this that
Q(n, q) ::; exp (n(2 - q - 6(2, q) + 0(1)))
for all sufficiently large n. Hence
T(q) ::; 2 - q - 6(2, q).
To obtain the lower bound we have to add a second moment estimate to our first
moment estimate (3.3). This is considerably harder, and after substantial work
one arrives at
. 1
T(q) = hm -logQ(n,q)=2-q-6(2,q),
n--+oo n
and hence, from (3.2),
f(a) = 2 - sup {6(2, q) - q(a - I)}.

°
q>O

Remark. As the total harmonic measure is constant, we have T(l) = and hence
our heuristics gives that 6(2,1) 1. For a rigorous argument using this idea,
see [20]. <>

4. The multifractal spectrum of intersection local time

Let W 1 , W 2 be two independent Brownian motions in ]R3 with a joint starting point
and running for one unit of time. Let S = W1[0, 1] n W2 [0, 1] be the intersection
of the paths, which is a (random) set of Hausdorff dimension 1. This set can
be equipped with a finite measure, the intersection local time £, the (projected)
intersection local time, which can be described symbolically by the formula

£(A) = i dy }]
2
1 1
ds8y (Wj (s)), for A c]Rd Borel. (4.1)

Le Gall has shown in [31, 32] that £ is the 'P-Hausdorff measure on the set S
for the gauge function
'P(x) = cx loglog(1/x)2, for x> 0,
where c is an unknown deterministic constant. Alternative rigorous constructions
of the random measure £ are reviewed in [19]. It is well-known that the local
dimension of £-almost all points equals 1, which is also the Hausdorff dimension
of the intersection set S.
Applying a multifractal formalism as in the previous example suggests that
the measure £ is not multifractal, as the T-function is linear. Indeed, points where
the local dimension is strictly below the typical value 1 do not exist, even in the
142 Peter Morters

weaker sense of a liminf in (1.1). In other words there is no nontrivial multifractal

spectrum to the left of 1.
In a recent paper, Klenke and Morters [18] are investigating whether there
are points where the local dimension exceeds 1. It turns out that this is indeed
the case, if the local dimension is defined in the weaker sense of a limsup in (1.1).
These exceptional points arise, loosely speaking, from random fluctuations in the
intersection local time which are too subtle to be picked up by the multifractal
formalism.
Theorem 4.1 (Klenke and Morters 2003). Suppose £ is the intersection local time
of two Brownian motions in ~3. For any 1 ::; a ::; ~ / (~ - 1), almost surely, there
exist points xES such that
.
11m log£(B(x,r)) _
sup 1 - a,
dO og r
for ~ = 6(2,2). Moreover, for these values of a, almost surely,
dim {x E S : lim sup log£(B(x,r)) =
dO logr
a} = ~ + 1-~.
a
(4.2)

Remark. Here the multifractal spectrum is strictly convex, hence it cannot be

found by means of the multifractal formalism of the last section, which by its
nature always predicts a concave spectrum. 0

In the following I give a heuristic argument for this result. The crucial ingre-
dient in the proof is an investigation of the strategy used by Brownian motion in
order to accumulate only a very small amount of mass in a small ball centred in
an intersection point xES.
In order to find the probability of the event £(B(x,r)) ~ r a given xES first
note that it is intuitively clear that in a typical intersection point xES a pair
of two-sided Brownian motions are attached. As the motions start in the same
point they typically have a lot of intersections right at the start. In order to avoid
intersections to the extent that £(B(x,r)) ~ r a it is optimal for the Brownian
motions to use the following strategy:
(1) Run normally until you hit oB(x,ra). The typical intersection local time
until then is ~ r a , and now one has to act.
(2) Do not intersect before leaving B(x, r 1- E ). The probability that this happens
is, by scaling and the definition of the intersection exponents,
r-((l-E-a)+o(1) .

(3) Once you reach oB(x, r 1- E ) never return to B(x, r). The probability that
this happens is ~ 1 - rE.
The total probability (or "cost") of this strategy is of order r-((l-a)+o(l),
essentially all of which is incurred in the second step. Note that it leads to an
empty annulus B(x,r) \ B(x,ra). It can be shown that this strategy is optimal,
i.e. no other strategy to get £(B(x, r)) ~ r a has a significantly higher probability.
 Measures on Brownian Paths 143

We now sketch how the upper bound in our spectrum can be deduced from
this. As dim S = 1 we can cover the set S by approximately r- a balls of radius ra.
Using the modulus of continuity result for Brownian motion one can see that, in
a ball of radius r- a , the probability that the ball contains a point xES with
.e(B(x,r)) ~ r a is approximately the same as the probability that the centre of
the ball has this property. Hence, for each ball in our collection, the probability of
containing a point xES with .e(B(x, r)) ~ r a is approximately r- W - a ). Choosing
a covering of the set
.
{ xES : hm log.e(B(x,r)) }
sup ;::: a
rlO logr
consisting of all these balls for the radii r = 2- n , n ;::: m and m fixed, leads to a
covering

with expected s-value

L 2-nas2na2nW-a) ,
00 00

1E{ ~.)rf)S } =
i=l n=m
which is finite iff s > ~/a + 1-~. Note that this argument for the upper bound is
particularly simple as it uses only a first moment method.
It is more difficult to establish a lower bound, as second moments (or in-
dependence in some form) have to be used. A particularly powerful technique to
establish lower bounds is the stochastic co-dimension method. Based on earlier
ideas of Taylor these techniques were developed further, first in [1, 38J and later
more systematically by Peres and coauthors, see e.g. [17, 8J or [19J. Nowadays
the method represents an elegant and powerful tool in fractal geometry, further
evidence of this can be found in the contribution of Xiao in this volume.
The stochastic co-dimension approach suggests that the dimension of a set
A C lR. d is at least 'Y if it intersects a certain compact random set rbJ of dimension
d - 'Y with positive probability, and conversely if dim A > 'Y then A intersects the
test set rbJ with positive probability. Suitable test sets are, for example, the
percolation limit sets with retention parameter 2-'1, see [12J and also [37J.
In our proof we assume that the test sets are independent of our Brownian
motions and construct, for any'Y E (0,1), a compact random set S* C S n r['Y]'
which is nonempty with positive probability, such that
dim U n S* ;::: 1 - 'Y whenever U open and U n S* =1= 0.
Fix 8 > 0 and define open sets
U(m,8) = {x E S : .e(B(x,r)) < r a - o for some 0 < r < ~}.
The regularity of S* is then used to show that for 'Y = ~/a +1- ~ the sets
U(m,8) n S* are dense in S*. To get an impression of this, fix r = 2- n and
note that given S* =1= 0 by the regularity property, for every open set U which
intersects S* there is a packing of S* n U consisting of approximately r'Y- 1 balls
144 Peter Morters

of radius r. We now localise the problem and look at each ball in the packing
separately. For both motions let S be the last entry of the path into the ball
before hitting the centre and T the first exit from the ball afterwards. We denote
by £* the intersection local time of the paths in the time intervals [S, T]. Then the
probability that inside our ball there is some x E S* such that
l*(B(x,r)) < ra-I) (4.3)
is of order
r -W-aH) r b+2)(a-I)-1) r -3(a-I)-1) .
Here the first factor comes from the probability that a point x, and by continuity
every point in a ball around x ofradius ~ ra-I), satisfies (4.3), the second factor
comes from the requirement that this ball intersects S*, and the third factor ac-
counts for the number of such small balls in the fixed ball of radius r. Moreover,
thanks to the localisation these events are essentially independent for different
balls. Hence the probability that we do not find a point x E S* satisfying (4.3) for
r = 2- n is
2 n (1-1')
(1 - 2n(H-y-1)(1-aH)) ~ 0, as n ---+ 00.

To complete the argument one has to show that for most points the localised
intersection local time £* is not different from £. In d ~ 3 this can be shown using
the transience of Brownian motion
Now, as U(m, 8) n S*, for mEN, 8> 0, are a nested family of open dense
subsets of the complete metric space S*, by Baire's theorem, the intersection is
also dense, and hence nonempty if S* is nonempty. Hence the intersection set
T(a):= n
mEN
U(m, 8) = {x E S : lim sup 10g£(B(x,r))
riO logr
oJ..
~ a}

intersects r[-y] with positive probability. On the other hand almost surely T(a +
°
c:) n r[-y] = 0 for all c: > by the upper bound. This shows that (4.2) holds with
positive probability, and the full result follows from a zero-one law argument.

5. The refined multifractal spectrum of occupation measure

The most natural measure on the Brownian path {W(t) : t E [0, I]} is of course
the occupation measure J.1., given by

J.1.(A) = 11 lA(Ws ) ds.

We have neglected this measure so far in this paper, as it is not multifractal.

Indeed, one can show that, almost surely,
lim 10gJ.1.I(B(x,r)) = 2, for all x E W[O, 1].
r!D ogr
This has been known for some time, for example it is implicit in the work of
Kaufman [16]. As there is no variation in the local dimension of J.1. it is natural to
 Measures on Brownian Paths 145

look at a finer scale and ask for an almost sure upper and lower hull of J..l(B{x, r))
when the point x is varied along the support of J..l. More precisely, looking at a the
lower hull first, we are asking for a gauge function ¢ : (0,6") --+ (0,00) such that

0< inf liminf /-L(B(x, r)) < 00.

dO ¢(r)
°
xEW[O,lj

A point x E W[O, 1] is called a thin point with thinness a > if

· . f J..l(B(x,r)) -
1Imm
dO "'()
'l'r
- a.

Once ¢ is identified, the dimension spectrum of thin points is given by

f(a) = dim {x E W[0,1] : li~l~nf J..l(~~~) r)) = a}.

In [7] the correct gauge function is found, and the spectrum of thin points is
identified for the occupation measure of a Brownian path in d = 3.
Theorem 5.1 (Dembo, Peres, Rosen and Zeitouni 2000). Almost surely, we have
1 . f r . f /-L(B(x,r))
= XEW[O,lj I~l~n r 2 [log(1/r)]-1 '
and, for every a 2: 1, almost surely,
. { [].. J..l(B(x, r)) } 2
dIm x E W 0, 1 : h~l~nf r 2 [log(1/r)]-1 = a = 2 - ~.

Passing to the upper hull, we are looking for a gauge function ¢ such that
. J..l(B(x, r))
0< sup hmsup < 00. (5.1)
¢(r)
°
xEW[O,l) riO

A point x E W[O, 1] is called a thick point with thickness a > if

. J..l{B(x,r))
hm sup o,,{) = a,
rlO 'I' r
and a dimension spectrum of thick points is defined accordingly. In [6] the correct
gauge function is found, and the spectrum of thick points is identified for the
occupation measure of a Brownian path in d = 3.
Theorem 5.2 (Dembo, Peres, Rosen and Zeitouni 2000). Almost surely,
. J..l(B(x, r)) 16
sup hmsup =- (5.2)
xEW[O,l) dO r [log(1/r)]
2 11"2'

and, for all °: ; a ::; !~, almost surely,

. { d. J..l(B{x,r)) } 11"2
(5.3)
dIm x E IR : llI~isOuP r 2 [log{1/r)] = a = 2 - 8 a.
146 Peter Morters

For the proof of these two theorems the authors developed substantial refine-
ments of the tools available for getting lower bounds for the Hausdorff dimension,
some of which we have already benefited from in the previous section.
We only indicate briefly the key point in the proof of Theorem 5.2, which,

°
as in Theorem 4.1, is the strategy used by the Brownian motion to accumulate
exceptionally large mass in a small ball. For any T > we look at the occupation
measures J.lT for Brownian motion running for T time units,

J.lT(A) = loT IA(Ws ) ds, for A C 1R3 Borel.

Classical results of Ciesielski and Taylor give the upper tail behaviour of the ran-
dom variable J.loo(B(O, 1)): For a positive constant c> 0,

JP>{J.loo(B(O, 1)) 2': x} rv ce- X7r2 / 8 , as xi 00, (5.4)

hence, by scaling, we obtain for small balls B(O, r) and fixed a > 0,

JP>{J.loo(B(O,r)) 2': ar 2 10g(l/r)} rv cra7r2 / 8 , as r! 0.

The principal question is, which strategy is used by the Brownian path to accu-
mulate an occupation time of order r 2 10g(l/r) in B(O, r). The answer found in
[6] is that this mass is accumulated in a very short time interval after the start
of the motion and eventual later returns to this ball do not contribute essentially.
Indeed, for h(r) = r 2 10g(l/r)6 they obtain

JP>{J.lh(r)(B(O,r)) 2': ar 2 10g(l/r)} rv cra7r2 / 8 , as r! 0. (5.5)

This localisation effect is the crucial effect which leads to Theorem 5.2, we
note here that the effect breaks down in the recurrent case of d = 2, and in this
dimension it is a complete different effect that leads to the dimension spectrum,
which in this case was found in [8].

It is also worthwhile to look at a refined multifractal spectrum, when the

measure under investigation has only a one-sided multifractal spectrum. This is
the case, for example, in the situation of the intersection local time, studied in the
previous section. Recall that in this case there exist points with local dimension
exceeding the typical value 1 (these points are exceptionally thin!) but no points
where the local dimension is smaller than 1. It therefore makes sense to ask for
a dimension spectrum of thick points as defined above. This question is answered
by Konig and Morters in [19].
Theorem 5.3 (Konig and Morters 2002). Almost surely,
. £(B(x,r)) 1
sup hm sup [1 (1/ )]2 = (}2' (5.6)
xES r!O r og r
 Measures on Brownian Paths 147

for a constant 0 given by

0= inf {11\7'ljJII~ : 'ljJ E C6(lR. d ), r

iR(O,I)
'ljJ4 = I}.

Moreover, for all 0:::; a:::; (l/4, almost surely,

. { . £(B(x,r)) } r::
dIm xES: hn;t~uP r[log(1/r)J2 = a = 1- yaO.

Finally, looking again at the occupation measure J.L, recall that our definition
of thickness requires large mass only along a sequence of radii r n ! O. This sequence
might be arbitrarily sparse. What happens if we replace the limsup in (5.1) by a
liminf? In other words, we are looking for consistently thick points. A first task is
to identify a gauge function <p : (0, E) ----) (0, E) such that

O< · . f J.L(B(x, r))

11m
sup III < 00.
xEW[O,I] dO <p(r)
Dembo, Peres, Rosen and Zeitouni have shown [6, Theorem 1.6] that the right
gauge function here is <p(r) = r2, which is different from the gauge function 'ljJ(r) =
r 2 1og(1/r) in (5.2), but the problem of the dimension spectrum of consistently
thick points is still open.

6. Summary and Outlook

By looking at three recently studied examples we have illuminated the following
three important directions in modern fractal geometry.
• Multifractal geometry beyond the multifractal formalism. [9, 14,39, 18]
Interesting stochastic examples show that the multifractal formalism fails if
the mass in typical small balls is a heavy-tailed random variable. In some
instances we see a strictly convex spectrum, a behaviour which cannot be
picked up by the formalism.
• Refined multifractal spectrum. [41,6, 7, 8, 34, 19,5]
In some cases of measures which are not multifractal, or have only a one-
sided multifractal spectrum, one can study a refined spectrum and get very
detailed information about exceptional sets.
• Critical exponents and fractal geometry. [21, 25, 26, 27, 19]
Critical exponents like the intersection exponents of Brownian motion contain
crucial information about the fractal geometry of the models involved. In
some cases this information can be extracted in terms of Hausdorff dimension
spectra involving the exponents.
Trying to look into the future of the area, I believe that the following topics
deserve attention before the next "Fractal Geometry and Stochastics" conference
will reunite many researchers in several years time.
148 Peter Morters

• Fractal geometry of discrete models.

Almost all problems discussed here have a discrete analogue, and in some
cases this has already been made explicit, see e.g. [25] for a discrete version of
the multifractal spectrum of harmonic measure, and [8] for a discrete version
of the refined multifractal spectrum of the occupation measure of a Brownian
path, both in d = 2. In the latter case this discretisation solved a long-
standing conjecture of Erdos and Taylor about the simple random walk in
the plane. I believe that in the future fractal geometers will increasingly look
into "discrete fractals" and apply their techniques to the fractal phenomena
arising, for example, in interacting particle systems and statistical physics.
• Increasingly complex dimension spectra.
Looking at recent results and techniques, for example, of the papers [7, 8]
or [17], one cannot help noticing the enormous improvement in the technical
standard and the sophistication of the tools presently available. However, I
do not believe that an end to this development is reached yet, as important
challenges remain open. Among these is the problem of consistently thick
points for Brownian motion in d 2: 3, mentioned in Section 5, and the problem
of the dimension spectrum of local times for Brownian motion in the plane
raised in [2].
• Fractal geometry of nonintersecting curves.
The stochastic Lowner evolution introduces a new class of universal non-
intersecting random curves in the plane, which arise naturally as scaling limits
of important physical models. These objects have a fractal nature which needs
to be investigated. This programme has already started and recently Beffara
[3] has calculated the Hausdorff dimension of these curves. Still, the fractal
geometry of these models will capture some researcher's imaginations for a
good while in the future.

References
[1] M.T. Barlow and E. Perkins Levels at which every Brownian excursion is exceptional.
Seminar on probability XVIII, Springer LNM 1059 (1984) 1-28.
[2] R. Bass, K. Burdzy and D. Khoshnevisan. Intersection local time for points of infinite
multiplicity. Ann. Probab. 22 (1994), 566-625.
[3] V. Beffara. The dimension of the SLE curves. Preprint arXiv: Math.PR/0211322
(2003).
[4] R. Cawley and R. D. Mauldin. Multifractal decompositions of Moran fractals. Adv.
Math. 92 (1992), 196-236.
[5] A. Dembo. Favourite points, cover times and fractals. Lecture notes, Ecole d'ete de
St. Flour (2003), Springer, to appear.
[6] A. Dembo, Y. Peres, J. Rosen and O. Zeitouni. Thick points for spatial Brownian
motion: multifractal analysis of occupation measure. Ann. Probab. 28 (2000), 1-35.
[7] A. Dembo, Y. Peres, J. Rosen and O. Zeitouni. Thin points for Brownian motion.
Ann. Inst. Henri Poincare, Probab. Stat. 36 (2000), 749-774.
 Measures on Brownian Paths 149

[8] A. Dembo, Y. Peres, J. Rosen and O. Zeitouni. Thick points for planar Brownian
motion and the Erdos-Taylor conjecture on random walk. Acta Math. 186 (2001),
239-270.
[9] D. Dolgopyat and V. Sidorov. Multifractal properties of the sets of zeroes of Brownian
paths. Fund. Math. 147 (1995),157-171.
[10] B. Duplantier. Conformal fractal geometry and boundary quantum gravity.
arXiv:math-ph/0303034, 118pp. (2003).
[11] B. Duplantier and K. H. Kwon. Conformal invariance and intersections of random
walks. Phys. Rev. Lett. 61 (1988),2514-1517.
[12] J. Hawkes. Trees generated by a simple branching process. J. London Math. Soc. 24
(1981) 373-384.
[13] T. C. Halsey, M. N. Jensen, L. P. Kadanoff, 1. Procaccia and B. 1. Shraiman. Fractal
measures and their singularities. Phys. Rev. A33 (1986), 1141-115l.
[14] X. Hu and S. J. Taylor. The multifractal structure of stable occupation measure.
Stoch. Processes Appl. 66 (1997) 283-299.
[15] P. Ch. Ivanov, 1. A. N. Amaral, A. L. Goldberger, S. Havlin, M. G. Rosenblum,
Z. Struzik, and H. E. Stanley. Multifractality in human heartbeat dynamics. Nature
399 (1999),461-465.
[16] R. Kaufman. Une propriete metrique du mouvement Brownien. C. R. Acad. Sci.
Paris 268 (1969), 727-728.
[17] D. Khoshnevisan, Y. Peres, and Y. Xiao. Limsup random fractals. Electron. J.
Probab. 5, Paper No.4, (2000).
[18] A. Klenke and P. Morters. The multifractal spectrum of Brownian intersection local
times. Preprint, University of Bath (2004).
[19] W. Konig and P. Morters. Brownian intersection local times: upper tails and thick
points. Ann. Probab. 30 (2002), 1605-1656.
[20] G. F. Lawler. Intersections of random walks. Birkhiiuser, Basel (1991).
[21] G. F. Lawler. Hausdorff dimension of cut points for Brownian motion. Elect. Journal
Probab. 1, Paper No.2, (1996).
[22] G. F. Lawler. The dimension of the frontier of planar Brownian motion. Elect.
Comm. in Probab. 1 (1996),29-47.
[23] G. F. Lawler. Strict concavity of the intersection exponent for Brownian motion in
two and three dimensions. Math. Phys. Electron. Journal 4, Paper No.5, (1998).
[24] G. F. Lawler. Geometric and fractal properties of Brownian motion and random walk
paths in two and three dimensions. Bolyai Math. Soc. Studies, 9 (1999), 219-258.
[25] G. F. Lawler. Multifractal nature of two dimensional simple random walk paths.
Sympos. Math. XXXIX, 231-264, Cambridge Univ. Press, Cambridge, (1999).
[26] G. F. Lawler. The frontier of a Brownian path is multifractal. Unpublished (1999).
[27] G. F. Lawler, O. Schramm and W. Werner. The dimension of the Brownian frontier
is 4/3. Math. Res. Lett. 8 (2001), 401-411.
[28] G. F. Lawler, O. Schramm and W. Werner. Values of Brownian intersection expo-
nents I: Half-plane exponents. Acta Math. 187 (2001), 237-273.
[29] G. F. Lawler, O. Schramm and W. Werner. Values of Brownian intersection expo-
nents, II: Plane exponents. Acta Math. 187 (2001), 275-308.
150 Peter Morters

[30] G. F. Lawler, O. Schramm and W. Werner. Values of Brownian intersection expo-

nents. III: Two-sided exponents. Ann. Inst. Henri Poincare 38 (2002), 109-123.
[31] J.-F. Le Gall. The exact Hausdorff measure of Brownian multiple points I. In: Sem-
inar on Stochastic Processes 1986, 107-137, Birkhauser, Boston (1987).
[32] J.-F. Le Gall. The exact Hausdorff measure of Brownian multiple points II. In: Sem-
inar on Stochastic Processes 1988,193-197, Birkhauser, Boston (1989).
[33] B. B. Mandelbrot. Intermittent turbulence in self-similar cascades: Divergence of
high moments and dimension of the carrier. J. Fluid Mech. 62 (1974), 331-358.
[34] P. Morters and N. R. Shieh. Thin and thick points for branching measure on a
Galton- Watson tree. Stat. and Probab. Letters 58 (2002) 13-22.
[35] L. Olsen. Multifractal geometry. In: Fractal Geometry and Stochastics II. C. Bandt,
S. Graf, M. Zahle (Eds), Birkhauser, Basel (2000), pp. 3-37.
[36] S. Orey and S. J. Taylor. How often on a Brownian path does the law of the iterated
logarithm fail? Proc. London Math. Soc. 28 (1974), 174-192.
[37] Y. Peres. Remarks on intersection-equivalence and capacity-equivalence. Ann. Inst.
Henri Poincare (Physique TMorique) 64 (1996) , 339-347.
[38] E.A. Perkins and S.J. Taylor. Measuring close approaches on a Brownian path. Ann.
Probab. 16 (1988), 1458-1480.
[39] E.A. Perkins and S.J. Taylor. The multifractal structure of super-Brownian motion.
Ann. Inst. Henri Poincare. 34 (1998), 97-138.
[40] D. A. Rand. The singularity speetrum f(a) for cookie-cutters. Ergod. Th. & Dynam.
Sys. 9 (1989), 527-541.
[41] N. R. Shieh and S. J. Taylor. Logarithmic multifractal spectrum of stable occupation
measure. Stoch. Processes Appl. 75 (1998), 249-261.
[42] A. Thriel and C. Perez Vicente. Multifraetal geometry in stock market time series.
Physica A 322 (2003), 629-649.
[43] E. B. Vul, Va. G. Sinai and K. M. Khanin, Feigenbaum universality and the thermo-
dynamic formalism. Russian Math. Surveys 39 (1984), 1-40.
Acknowledgement
I would like to thank the organisers of the conference, Christoph Bandt, Umberto
Mosco, and Martina Zahle for their hospitality.

Peter Morters
Department of Mathematical Sciences, University of Bath
Claverton Down, Bath BA2 6PA, United Kingdom
e-mail: [email protected] . uk