Multifractal Analysis of Weak Gibbs Measure

Ergod. Th. & Dynam. Sys.
(2003), 23, 17511784 c 2003 Cambridge University Press

DOI: 10.1017/S0143385703000051 Printed in the United Kingdom
Multifractal analysis of weak Gibbs measures
and phase transitionapplication to some
Bernoulli convolutions
DE-JUN FENG and ERIC OLIVIER
Department of Mathematical Sciences, Tsinghua University, Beijing 100084,
Peoples Republic of China
(e-mail: dfeng@math.tsinghua.edu.cn)
Department of Mathematics, The Chinese University of Hong Kong, Shatin,
New Territories, Hong Kong
(Received 6 June 2001 and accepted in revised form 23 January 2003)
Abstract. For a given expanding d-fold covering transformation of the one-dimensional
torus, the notion of weak Gibbs measure is dened by a natural generalization of the
classical Gibbs property. For these measures, we prove that the singularity spectrum
and the L
q
-spectrum form a Legendre transform pair. The main difculty comes from
the possible existence of rst-order phase transition points, that is, points where the
L
q
-spectrum is not differentiable. We give examples of weak Gibbs measure with phase
transition, including the so-called Erd os measure.
0. Introduction
The one-dimensional torus S
1
:= R/Z is endowed with the natural metric and dim
H
M
denotes the Hausdorff dimension of any M S
1
(by convention, dim
H
= ).
Let B
r
(x) be the closed ball of radius r > 0 centered at x S
1
; the local dimension
of a Borel probability measure at x is by denition
DIM
(x) := lim
r0
log (B
r
(x))
log r
, (1)
provided that the limit exists. The level set E() ( R) associated to is the set of points
x S
1
such that DIM
(x) exists and is equal to . The map dim

H
E() is called the
singularity spectrum of . Heuristic arguments using techniques of statistical mechanics
(see [16] for example) show that the singularity spectrum should be nite on a compact
interval denoted DOM() and is expected to be the Legendre transform conjugate of the
L
q
-spectrum associated to (see Denition 1.3); that is, for all DOM(),
dim
H
E() = inf{q (q); q R} =:
(). (2)
http://www.paper.edu.cn
1752 D.-J. Feng and E. Olivier
The multifractal analysis of a probability measure is concerned with rigorous arguments
ensuring that the Legendre transform formula (2) holds; the rst multifractal formalisms
were established for Gibbs [7, 36, 39], quasi-Bernoulli [4, 17] and self-similar measures
[6, 25, 33]: we refer to [34] for a general overview of the subject and complete references.
In this paper, we focus our attention on the notion of weak Gibbs measure dened in [46]
by a natural generalization of the classical Gibbs property (see Denition 1.1). Of special
interest are the g-measures, as considered in [22], which turn out to be weak Gibbs, as well
as a large class of the conformal measures studied in [19]. It follows from the variational
characterization of the g-measure proved in [26] that a non-ergodic g-measure cannot be
Gibbs. The existence of non-ergodic g-measures established in [3] shows that a weak
Gibbs measure need not be Gibbs.
Let T be an expanding d-fold covering transformation of S
1
(with d 2) so that there
exists a Markov partition M
1
of T , by d (semi-open) intervals; if M
n
(n 1) denotes the
partition of S
1
by the n-step basic intervals then M :=

n
M
n
is called the Markovian
net of T . From now on, we assume that is a weak Gibbs measure of a potential
dened on the symbolic space which codes the Markovian dynamics of T . Our approach
to the multifractal analysis of is classical; we consider the level sets E(|M) associated
to the Markovian local dimension DIM
(|M) and we give (Theorem A) an intermediate

Legendre transform formula, say, for any in the interior of DOM(),
dim
H
E(|M) =
(), (3)
where the concave map
is implicitly dened by a pressure equation. An important

and non-trivial step is to prove (Theorem B) that dim
H
E(|M) = dim
H
E(), when
is a weak Gibbs measure. The last step is achieved by proving (Theorem C) that
coincides with the L

q
-spectrum and we conclude (Theorem A
) that the Legendre

transform formula (2) holds, when is weak Gibbs.
These generalizations are relevant essentially because the L
q
-spectrum need not be
real-analytic/differentiable when is weak Gibbs. For the thermodynamic formalism on
lattices, a system is said to exhibit a phase transition when a thermodynamic function
displays a defect of analyticity at some critical value (see [40, ch. 5]). We shall say that the
real number q
c
is a phase transition point (respectively a rst-order phase transition point)
if is not real-analytic (respectively not differentiable) at q
c
: this will make sense, for we
shall prove (Theorem C) that coincides with a thermodynamic function determined by
a pressure equation. Let us denote by
(q
+
c
) (respectively
(q
c
)) the right (respectively
left) derivative of at q
c
; if
(q
+
c
) <
(q
c
) then the mass distribution principle does
not apply to give the desired lower bound of dim
H
E(|M) when
(q
+
c
) < <
(q
c
).
Our argument depends on the tangency property of the topological pressure, which yields a
thermodynamic characterization of
(q
+
) and
(q
) for any q R (Lemma 3.2) and on

a formula which gives the Billingsley dimension of the generic points of a (not necessarily
ergodic) shift-invariant measure.
In 1 we describe the framework of the expanding d-fold covering transformation of the
one-dimensional torus and the thermodynamic formalismof the equilibriumstate. Then we
give a denition of the weak Gibbs measure (Denition 1.1) making possible a rigorous
statement of Theorems A, B, C and A
; the proofs of these theorems are given in 3.

Multifractal analysis of weak Gibbs measures 1753
Section 2 is devoted to an illustration of the previous results through the analysis
of two examples of Bernoulli convolutions. We rst study (2.1) the so-called
(2, 3)-Bernoulli convolution: this measure is proved to be weak Gibbs (Theorem 2.4) and
we show that its L
q
-spectrum displays a phase transition point (Theorem 2.5). The case of
the (2, 3)-Bernoulli convolution is closely related to our main application, concerned with
the multifractal analysis of the Erd os measure. The Bernoulli convolution
(1 < < 2)
dened in 2.2 is a non-atomic probability measure supported by the unit interval which is
either continuous or purely singular (see [21]). Erd os proved in [10] that
is purely
singular when is a Pisot number (i.e. an algebraic integer whose conjugates have
modulus less than 1). When = (1 +

5)/2 the measure :=
is called [42] the

Erd os measure. The multifractal analysis of has been partially studied in [13, 24, 27];
we prove (Theorem 2.9) that is a weak Gibbs measure (but not Gibbs) with respect
to a suitable 3-fold covering transformation (the potential of is dened by means of
continued fractions), so that the full multifractal formalism is completely established
(our contribution is concerned with the decreasing part of the singularity spectrum of ).
In [13], the rst author completes a result in [24] by giving an explicit formula for the
L
q
-spectrum of the Erd os measure, proving in addition that there exists a negative q
c
such
that: (i) (q) = q log 2/log for any q q
c
; (ii) is innitely differentiable at any
q > q
c
; and (iii) is not differentiable at q
c
. The weak Gibbs property of makes
possible (Theorem C) the use of the thermodynamic formalism and one may interpret q
c
as a rst-order phase transition. The variational principle allows an alternative approach
to (i) (Theorem 2.10), which is actually enough to ensure that is not real-analytic at
q
c
< 0; the fact that is not differentiable at q
c
is more difcult to establish: Appendix A
is devoted to a self-contained proof of this result based on the original argument in [13].
We point out that the multifractal formalism is valid for
n
, when
n
is the Pisot number
such that
n
n
=
n1
n
+ +
n
+ 1 (n 3) [13, 32], the corresponding L
q
-spectrum
being differentiable on the whole real line [13]. Even if some partial results can be
achieved (see e.g. [14]), the general case of a Pisot number seems to remain a difcult
problem.
1. Multifractal formalism of weak Gibbs measures
1.1. General framework. Let T be an expanding continuous transformation of S
1
such
that T
1
{x} is of cardinality d > 1 for any x S
1
. Assuming that T (0) = 0, there
exist d points x
d
= 0 = x
0
, x
1
, . . . , x
d1
such that T is a one-to-one, onto mapping
from [i] := [x
i
, x
i+1
[ to S
1
. The bilateral restrictions T
i
:]x
i
, x
i+1
[ S
1
\{0} are
diffeomorphisms with H older-continuous derivative (the transformation T may not be
differentiable at the points x
i
): we say that T is a regular d-fold covering transformation
(d-f.c.t.) of S
1
. Let
d
be the one-sided direct product of an innite number of copies of
the alphabet A
d
:= {0, . . . , d 1}, i.e.
d
:=

0
A
d
; we assume that
d
is endowed
with the product topology and :
d

d
denotes the (one-sided) shift transformation.
Each x S
1
is associated to a unique point (x) := (
k
)
d
such that T
k
(x) [
k
],
for any integer k 0, and : S
1

d
is a one-to-one map such that the following
diagram commutes
S
1
T
//
S
1
//

d
(we usually call the coding map of T ). For any x S
1
such that = (x), we
denote by I
n
(x) = [
0

n1
] the n-step basic interval about x, that is, the set of
points y S
1
such that T
k
(y) [
k
], for k = 0, . . . , n 1. Let [[w]] be the set of
the
d
such that
0

n1
= w, which is usually called the n-step cylinder set about
w (clearly,
1
([[w]]) = [w]). If M
0
:= {S
1
} and if M
n
(n > 0) is the collection of
the n-step basic intervals then M := {M}
n=0
is by denition the Markovian net of T .
The H older-continuous function
0
:
d
R such that
0
() = log |T
(x)| when
= (x) (and extended to
d
by continuity) is called the volume-derivative potential of
T with respect to M. By a classical application of the Mean Value Theorem, there exists
a constant K > 1 such that, for any (S
1
) and any integer n > 0,
1
K
|[
1

n
]|
exp(S
n
0
())
K (4)
(|J| stands for the length of any interval J S
1
and S
n
0
() :=

n1
k=0

0
(
k
)).
Denition 1.1. The measure dened on S
1
is said to be a weak Gibbs measure of the
potential :
d
R, if there exists a sub-exponential sequence of real numbers
K(n) > 1 (i.e. lim
n
(1/n) logK(n) = 0) such that, for any n > 0, and any (S
1
),
1
K(n)
[
0

n1
]
exp(S
n
())
K(n); (5)
without loss of generality, we assume that K(n) increases with n.
If K(n) is constant, one recovers the classical notion of Gibbs measures [2];
accordingly, by (4), the Lebesgue measure is a Gibbs measure of the volume-derivative
potential
0
.
Suppose that the probability measure is fully supported by S
1
; for (S
1
), we set
1
() := log [
0
] and for any n > 1,
n
() := log
[
0

n1
]
[
1

n1
]
. (6)
Using the density of (S
1
) in
d
, one extends
n
to a continuous map dened on the whole
of
d
and called the n-step potential of . The following lemma provides a useful way to
prove that is weak Gibbs (the proof is left to the reader).
LEMMA 1.2. Let
n
be the n-step potential of a fully supported probability measure ; if
n
converges uniformly to a potential then is a weak Gibbs measure of .
Even if a weak Gibbs measure need not be invariant under the dynamics of T , the notion
is closely related to the theory of equilibriumstates. The topological pressure of a potential
:
d
R (simply assumed to be continuous) is, by denition,
P() := lim
n
1
n
log
wA
n
d
exp(S
n
[[w]]), (7)
where S
n
[[w]] := max{S
n
(); [[w]]} (a sub-additive argument ensures that the limit
in (7) does exist). The variational principle of Walters [45] asserts that, for any -invariant
probability measure of metric entropy h
(), one has h
() + () P(), equality
being obtained when is an equilibrium state of ; the set of equilibrium states of is a
non-empty weak- compact convex set (in fact a Choquet simplex), whose extreme points
are -ergodic measures. We shall use the basic properties of the topological pressure listed
in [45, Theorem 9.7].
It is worth noting that the weak Gibbs property is satised by a g-measure in the sense of
Keane [22] as well as the so-called conformal measures. More precisely, given a potential
:
d
R, the probability measure dened on S
1
is said to be e
-conformal, if for
any Borel set A, with A [i] for some i A
d
, one has
T (A) =
_
A
e
(x)
d(x). (8)
Under the condition that is fully supported by S
1
, it is easily seen that (8) implies the
uniform convergence of the n-step potentials of , ensuring by Lemma 1.2 that is weak
Gibbs. However, the converse is not true in general, even if there exists a partial reciprocal;
to see this, we notice that if in addition to being of full support and e
-conformal,
one also assumes that is T -invariant, then, according to the previous remark, is a
weak Gibbs measure of , but it is also necessary that is normalized in the sense that
=
e
()
1 (the n-step potentials are trivially normalized and, using the uniform
convergence, one deduces that is also normalized). Now, if one assumes that is a
T -invariant weak Gibbs measure of the normalized potential , then it is easily seen that
is an equilibrium state of and by the variational principle in [26, Th eor` eme 1], one
deduces that is e
-conformal. When the potential is normalized one usually writes

g = e
so that the T -invariant e
-conformal measures are exactly the g-measures.

1.2. Statement of the multifractal theorems. The classical starting point of the
multifractal analysis is to make possible an application of the ShannonMcMillan
Theorem, by the introduction of the Markovian local dimension
DIM
(x|M) := lim
n
log (I
n
(x))
log |I
n
(x)|
, (9)
provided that the limit exists; then the -level set corresponding to this local dimension is
by denition
E(|M) := {x S
1
; DIM
(x|M) = }. (10)
From now on, suppose that the probability measure dened on S
1
is a weak Gibbs
measure of the negative potential :
d
R. For any (q, t ) R R we consider the
partition function
Z
n
(q, t ) :=
wA
n
d
[w]
q
|[w]|
t
. (11)
Using the Gibbs property (4) of the Lebesgue measure and the weak Gibbs property (5)
of , it follows from the denition given in (7) that
P(q t
0
) = lim
n
1
n
log Z
n
(q, t ) =: P
(q, t ). (12)
The convex property of P implies that P
is a convex map on R R; for any q, q
R
with q
> q, one has

q
t
0
(q
q) sup() +(q t
0
),
where sup() := sup{();
d
}; hence
P(q
t
0
) (q
q) sup() +P(q t
0
),
and thus
P
(q, t ) P
(q
, t )
q q
sup(). (13)
Since is supposed continuous and negative, sup() < 0, and one deduces from (13)
that the convex map q P
(q, t ) decreases from + to . Using the assumption

that sup(
0
) < 0, the same argument shows that, for any q R xed, the convex map
t P
(q, t ) increases from to +. Therefore, there exists an increasing concave

map
: R R dened by the implicit equation P
(q,
(q)) = 0. The underlying

thermodynamic formalism related to the convex map P
leads to a characterization of the

points where
is not differentiable (i.e. the rst-order phase transition points): actually,

according to Lemma 3.2 proved in 3.1,
(q
) = sup
_
()
(
0
)
_
and
(q
+
) = inf
_
()
(
0
)
_
,
where the inmum and the supremum are respectively taken over the probability measures
, equilibriumstates of the potential q
(q)
0
. Lemma 3.2 also provides a description
of the behavior of
about + and : more precisely we shall prove that

inf
_
()
(
0
)
_
= lim
q+
(q)
q
=: and sup
_
()
(
0
)
_
= lim
q
(q)
q
=: ,
where the inmum and the supremum are respectively taken over the probability measures
which are -invariant on
d
.
It is now possible to state the multifractal formalism of a weak Gibbs measure with
respect to the Markovian local dimension.
THEOREM A. Let Mbe the Markovian net of a d-f.c.t. T and be a weak Gibbs measure
of a negative potential :
d
R. The set of all points with E( |M) = is the
interval [, ] and dim
H
E( |M) =
(), for any < < .

When T is differentiable on the whole torus and is a Gibbs measure associated to a
potential which is continuous with respect to the natural topology on S
1
, it is well known
[36] that E() = E(|M). We emphasize that, in our framework, the transformation
T (respectively the potential of the weak Gibbs measure) may not be differentiable
(respectively continuous) for the natural topology on S
1
: in that case one may have
E() = E(|M). The following theorem is a crucial point of our multifractal analysis.
THEOREM B. Let Mbe the Markovian net of a d-f.c.t. T and be a weak Gibbs measure.
Then, dim
H
E(|M) = dim
H
E(), for any R.
Denition 1.3. The L
q
-spectrum of is the concave map : R R {} such that
(q) := liminf
r0
log inf
_
i
(B
i
)
q
; {B
i
}
i
_
log r
,
where {B
i
}
i
runs over the family of covers of S
1
by closed balls of radius r.
THEOREM C. Let T be a d-f.c.t. and be a weak Gibbs measure of a negative potential
:
d
R. Then,
(q) = (q), for any q R.

Finally, we can state the strong version of Theorem A.
THEOREM A
. Let T be a d-f.c.t. and be a weak Gibbs measure of a negative potential

:
d
R. The domain of the nite values of dim
H
E() is DOM() = [, ] and
dim
H
E() =
(), for any < < .

2. Bernoulli convolutions
For practical reasons, we shall need basic notions about the set of words on an alphabet.
Given A := {0, . . . , s 1} (s 2) a nite alphabet, each element in A
n
(n 1) is denoted
by a string of n letters/digits in A that we call a word; by convention A
0
is reduced to
the empty word . By denition, A
is the set of words on A, that is, A
:=

n=0
A
n
.
We denote by wm the concatenation of the two words w and m so that A
, endowed with
the concatenation, is a monoid with unit element . Whenever x
0
, . . . , x
s1
are s elements
of a monoid (X, ) with identity element e, we denote x
:= e and x
w
:= x
0
x
n1
,
for any word w =
0

n1
A
.
2.1. The (2,3)-Bernoulli convolution. Let b and d be two integers with 2 b < d (b is
for basis and d is for digit). The (uniform) (b, d)-Bernoulli convolution is dened as the
probability distribution of the random variable X :
d
R such that
X() =
1
b
b 1
d 1
n=0
n
b
n
,
for any = (
i
)
i=0

d
, when
d
is endowed with the equidistributed Bernoulli
probability. The measure is non-atomic, is supported by the whole unit interval
I = [0, 1], and is either absolutely continuous or purely singular (see [21, Theorem 35]).
In this section we focus our attention on the (2, 3)-Bernoulli convolution , when
3
is
endowed with the equidistributed Bernoulli measure
3
(by [8, Propositions 5.2 and 5.3],
is known to be purely singular). The measure turns out to be self-similar; to see this,
notice that for M I one has X() M if and only if
0
/4 + X()/2 M, that is,
S
0
X () M, where S
0
(x) := x/2 +
0
/4; since
3
is a Bernoulli measure, for
= 0, 1 or 2 one obtains
3
([[]] {S
X () M}) =
3
[[]]
3
{S
X () M}
=
1
3
3
{S
X() M} =
1
3
(S
1
(M)).
Finally, one deduces that satises the following self-similarity equation:
=
1
3
S
1
0
+
1
3
S
1
1
+
1
3
S
1
2
. (14)
Each dyadic sub-interval of I is coded by a word w {0, 2}
and in what follows we

denote [w] := S
w
(I). From the self-similarity property of the measure given in (14) and
the fact that S
1
1
S
00
(M) = {0} and S
1
2
S
00
(M) = , for any M I, one obtains
[00w] =
1
3
(S
1
0
[00w]) +
1
3
(S
1
1
[00w]) +
1
3
(S
1
2
[00w])
=
1
3
(S
1
0
S
0
[0w]) +
1
3
(S
1
1
S
00
[w]) +
1
3
(S
1
2
S
00
[w]) =
1
3
([0w]).
Likewise, using the identity S
12
= S
20
and the fact that S
1
0
S
20
(M) {1}, for any M I,
[20w] =
1
3
(S
1
0
S
20
[w]) +
1
3
(S
1
1
S
20
[w]) +
1
3
(S
1
2
S
20
[w])
=
1
3
(S
1
1
S
12
[w]) +
1
3
[0w]
=
1
3
([2w]) +
1
3
([0w]),
and the following matricial identity holds:
_
[00w]
[20w]
_
=
1
3
P
0
_
[0w]
[2w]
_
, where P
0
:=
_
1 0
1 1
_
. (15)
Since S
10
= S
02
, one gets in the same way that
_
[02w]
[22w]
_
=
1
3
P
2
_
[0w]
[2w]
_
, where P
2
:=
_
1 1
0 1
_
. (16)
A simple induction using (15), (16) and the fact that [0] = [2] = 1/2 yields, for any
0

n1
{0, 2}
n
,
[
0

n1
] =
1
2
1
3
n1
t
V
0
P
n1
V, (17)
where
V
0
:=
_
1
0
_
, V
2
:=
_
0
1
_
, V :=
_
1
1
_
.
In order to t our framework, the measure is identied to a measure on the torus S
1
.
Let T
2
: S
1
S
1
be the multiplication by 2 (mod 1); it is a 2-f.c.t. coded by the full shift
: where :=
0
{0, 2} and the volume-derivative potential is
0
: R
such that
0
log 2. A dyadic interval [
0

n1
] becomes an n-step basic interval,
i.e. [0] = [0, 1/2[, [2] = [1/2, 1[ and x [
0

n1
] if and only if T
k
2
(x) [
k
]
whenever 0 k n 1.
Denote by 0 := (
i
= 0)
i=0
; then, for any potential : R,
exp(S
n
(20)) =
exp((20))
exp((0))
{exp((0))}
n
.
A direct computation using (17) gives [20
n1
] = n/(2 3
n1
), for any n > 0; if is a
Gibbs measure of , then there exists a constant K > 1 such that, for any n 1,
1
K
1
n
{3 exp((0))}
n
K;
this is impossible and one concludes that is not a Gibbs measure. Our aim is to
prove that satises the weak Gibbs property with respect to some potential to be
identied. We consider the probability measure
dened on S
1
by setting, for any word
0

n1
{0, 2}
n
,
[
0

n1
] =
1
2
1
3
n
t
VP
n1
V. (18)
It is clear that
is T
2
-invariant and the next proposition shows how it is related to .
PROPOSITION 2.1. For any , and any integer n 1,
3
n +2

[
0

n1
]
[
0

n1
]
3.
Proof. Let w {0, 2}
n
; it is easily seen that [w]/
[w] 3 and thus it remains to prove

that [w]/
[w] 3/(n +2). Assume that w = 0

a
1
2
a
2

a
k
, with {0, 2}, k 2 and
a
1
, . . . , a
k
> 0 (the cases a
1
= 0 or n are similar); if we denote
P
a
2
1
2
P
a
k
V =:
_
p
q
_
,
then
[w]
[w]
=
3(1 1)
_
p
q
_
(1 1)P
a
1
0
P
2
_
p
q
_ =
3(p + q)
(1 +a
1
)(p + q) +q
3
a
1
+2
.
Since n = a
1
+ + a
k
, it follows that a
1
n and thus [w]/
[w] 3/(n +2). 2

In order to dene the potential associated to
in Theorem 2.2 below (as well as for the

Erd os measure in 2.2), we introduce some notations and ideas.
Given a sequence a
0
, a
1
, . . . of integers with a
0
0 and a
i
> 0 for i > 0, we denote
[a
0
; a
1
, . . . , a
k
] := a
0
+
1
a
1
+
1
.
.
.
+
1
a
k
=:
p
k
q
k
,
where the irreducible fraction p
k
/q
k
is the kth convergent of the continued fraction
[a
0
; a
1
, . . . ] = lim
k
[a
0
; a
1
, . . . , a
k
]. The integers p
k
and q
k
satisfy a well-known linear
recurrence, say
_
p
k
q
k
_
=
_
a
k
p
k1
+p
k2
a
k
q
k1
+q
k2
_
= Q
a
0
Q
a
k
_
1
0
_
where Q
a
i
:=
_
a
i
1
1 0
_
; (19)
we refer to [23] for a general presentation of continued fraction theory.
For {0, 2} and a any positive integer, we denote |a =
a
; given any sequence
a
1
, a
2
, . . . , a
k
(k 2) of positive integers, we dene |a
1
, a
2
, . . . , a
k
by means of the
induction formula |a
1
, a
2
, . . . , a
k
=
a
1
2 |a
2
, . . . , a
k
. Let 0 := (
i
= 0)
i=0
,
2 := (
i
= 2)
i=0
and :
3
N {} such that
() =
_
min{k 0;
k
[[1]] {0, 2}} if k 0,
k
[[1]] {0, 2};
if k 0,
k
/ [[1]] {0, 2};
hence the map :
3
N {} is the hitting time of [[1]] {0, 2}. Moreover, if
() = 0 then we set n() = 0, and if 0 < () < , there exists a unique nite
sequence of n() positive integers a
1
, . . . , a
n()
with a
1
+ + a
n()
= () and such
that,
0

()1
=
0
|a
1
, . . . , a
n()
; when () = we set n() := and
there exists a unique (innite) sequence of integers a
1
, a
2
, . . . with a
i
> 0 and such that
0

n1
=
0
|a
1
, . . . , a
k
, , where 0 < a
k+1
and a
1
+ +a
k
+ = n.
THEOREM 2.2. The T
2
-invariant probability measure
dened on S
1
by (18) is a
g-measure of the normalized potential : R such that
() =
_
_
log(1/3), if n() = () = 0,
log([1; a
1
, . . . , a
n()
]/3), if 0 < n(), () < ,
log([1; a
1
, . . . ]/3), if n() = () = .
The proof of Theorem 2.2 relies on the following lemma, which, according to (19),
makes the link between the matrix product formula in (18) dening
and the continued

fractions involved in the denition of the potential associated to
in Theorem 2.2.
LEMMA 2.3. For w = |a
1
, . . . , a
k
, one has
t
VP
w
V =
t
VQ
a
1
Q
a
k
V .
Proof. Notice that Q
0
Q
0
is the identity matrix and that P
a
2
Q
0
= Q
0
P
a
0
= Q
a
, for each
integer a 0; given any nite sequence of integers a
1
, . . . , a
k
with a
2
a
k1
> 0, one
has
t
VP
a
1
2
P
a
2
0
P
a
k
2
V =
t
V (P
a
1
2
Q
0
)(Q
0
P
a
2
0
) (P
a
k
2
Q
0
)Q
0
V
=
t
VQ
a
1
Q
a
2
Q
a
k
V. 2
Proof of Theorem 2.2. Assume that () = (the case () < being similar).
For n 1 one writes
0

n1
=
0
|a
1
, . . . , a
k
, , with n = a
1
+ +a
k
+ and
0 < a
k+1
; one considers x := [1; a
1
, . . . ] and for any k 1,
p
k
q
k
= [1; a
1
, . . . , a
k
] and
p
k
q
k
:= [1; a
1
, . . . , a
k
, +1].
By the denition of the n-step potential
n
of
, it follows from Lemma 2.3 that

3 exp(
n
()) =
t
VQ
a
1
Q
a
2
Q
a
k
Q
V
t
VQ
a
1
1
Q
a
2
Q
a
k
Q
V
=
(1 0)Q
1
Q
a
1
Q
a
k
Q
+1
t
(1 0)
(0 1)Q
1
Q
a
1
Q
a
k
Q
+1
t
(1 0)
=
(1 0)
t
(p
k
q
k
)
(0 1)
t
(p
k
q
k
)
=
p
k
q
k
.
Using classical inequalities of the theory of continued fractions,
x
p
k
q
x
p
k
q
k
p
k
q
k
k
q
1
q
k
q
k+1
+
1
q
k
q
1
q
k+1
+
1
q
k
.
Since n = a
1
+ + a
k
+ , one gets q
k+1
q
k
> n, which gives the following upper
bound:
|exp(()) exp(
n
())|
2
3n
.
The desired result is obtained by an application of Lemma 1.2. 2
From Proposition 2.1 and Theorem 2.2, one deduces the following.
THEOREM 2.4. The (2, 3)-Bernoulli convolution is a weak Gibbs measure of .
Since is a weak Gibbs measure of the potential , it satises the multifractal
formalism as stated in Theorem A
; moreover, by Theorem C, its L

q
-spectrum coincides
with the concave function
, a solution of the implicit equation P
(q,
(q)) = 0.
The following theorem proves that is not real-analytic at a critical point q
c
< 0, meaning
that q
c
is a phase transition point.
THEOREM 2.5. The L
q
-spectrum of the (2, 3)-Bernoulli convolution is a concave
function for which there exists q
c
< 0 such that (q) = q log 3/log 2 if and only if q q
c
.
The proof of Theorem 2.5 depends on the following lemma.
LEMMA 2.6. lim
q
k=1
a
1
,...,a
k
>0
(
t
VQ
a
1
Q
a
k
V )
q
= 0.
Proof. We rst prove that if a
1
, . . . , a
k
are positive integers then
t
VQ
a
1
Q
a
k
V
_
a
1
+
2
2
_

_
a
k
+
2
2
_
, (20)
where = (1 +

5)/2. It is readily checked that
t
VQ
a
V (a +
2
)/
2
when a > 0.
Given k positive integers a
1
, . . . , a
k
, one has, for any integer a > 0,
t
VQ
a
1
Q
a
k
Q
a
V
t
VQ
a
1
Q
a
k1
Q
a
k
1
_
a +1
a +1
_
.
If (20) is valid for rank 1 up to rank k 1, then (even when a
k
1 = 0),
t
VQ
a
1
Q
a
k
Q
a
V
_
a
1
+
2
2
_

_
a
k1
+
2
2
__
a
k
1 +
2
2
_
(a +1)
_
a
1
+
2
2
_

_
a
k1
+
2
2
__
a
k
+
2
2
__
a +
2
2
_
,
and (20) follows by induction. Using (20), one gets, for any q < 0,
k=1
a
1
,...,a
k
>0
(
t
VQ
a
1
Q
a
k
V )
q
k=1
a
1
,...,a
k
>0
__
a
1
+
2
2
_

_
a
k
+
2
2
__
q
k=1
_

n=1
_
n +
2
2
_
q
_
k
and the desired result holds. 2
Proof of Theorem 2.5. The Bernoulli convolution being a weak Gibbs measure of , one
deduces from Theorem C that (q) =
(q): this means that, for any q R,

(q) = max{t; P
(q, t ) 0} = min{t; P
(q, t ) 0}. (21)

Let
0
:= log 3/log 2; for 0 = (
i
= 0)
i=0
one has
0
(0) = log 2 and (0) = log3,
so that q(0)
0
q
0
(0) = 0, for any q R: it follows from the variational principle
that
0 = h
(
0
) +
_
{q()
0
q
0
()} d
0
() P(q
0
q
0
) =: P
(q,
0
q),
which by (21) implies that (q)
0
q. We now prove that (q)
0
q when q is
sufciently close to . Since
is also a weak Gibbs measure of , one can write

P
(q, t ) = lim
n
1
n
log Z
n
(q, t ), where Z
n
(q, t ) :=
w{0,2}
n
[w]
q
/|[w]|
t
.
By Lemma 2.3, one has, for any q < 0,
Z
n
(q,
0
q) :=
w{0,2}
n
(
[w]/|[w]|
0
)
q
=
1
2
q
w{0,2}
n
(
t
VP
w
V)
q
=
2
2
q
n
k=1
a
1
++a
k
=n
(
t
VQ
a
1
Q
a
k
V )
q
so that
Z
n
(q,
0
q)
2
2
q
k=1
a
1
,...,a
k
>0
(
t
VQ
a
1
Q
a
k
V)
q
.
Hence, by Lemma 2.6, there exists q
0
< 0 such that Z
n
(q, q) 1/2
q
for any q < q
0
and
each n 1. Therefore, P
(q,
0
q) 0 and (21) gives (q)
0
q when q q
0
; since
is a concave map with (0) = 1, there exists q
c
< 0 such that (q) =
0
q if and only if
q q
c
. 2
2.2. The Erd os measure. Suppose that 1 < < 2 and let := 1/(1). The Bernoulli
convolution
can be dened as the probability distribution of the random variable

Y
:
2
R such that Y
() = (1/)
k=0
k
/
k+1
, where
2
is endowed with
the equidistributed Bernoulli measure. As in the case of a (b, d)-Bernoulli convolution,
satises a self-similarity equation, say
=
1
2
S
1
0
+
1
2
S
1
1
, (22)
with S
(x) = (x +/)/ and {0, 1}. From now on, we assume that = (1 +
5)/2,
that is =
is the Erd os measure. The algebraic equation

2
= + 1 satised by
implies that 1/ = 1 =: , so that S
0
(x) = x and S
1
(x) = x +
2
. It is easily seen
that the intervals S
00
[0, 1[, S
100
[0, 1[ = S
011
[0, 1[ and S
11
[0, 1[ form a partition of [0, 1[;
this means that R
0
:= S
00
, R
1
:= S
100
= S
011
and R
2
:= S
11
dene a non-overlapping
system of afne contractions on [0, 1[ and for any word w on the alphabet {0, 1, 2}, we
denote [w] = R
w
(I). By symmetry [0] = [2] and according to the self-similarity
property (22) satised by , one gets [1] = ([0] +[2])/2, so that
[0] = [1] = [2] =
1
3
. (23)
Likewise, for any word w on the alphabet {0, 1, 2},
[1w] =
1
2
(S
1
0
[1w]) +
1
2
(S
1
1
[1w])
=
1
2
(S
1
0
S
011
[w]) +
1
2
(S
1
1
S
100
[w])
=
1
2
(R
2
[w]) +
1
2
(R
0
[w]) =
1
2
[2w] +
1
2
[0w],
which we can write in the following matricial way:
_
[1w]
[1w]
_
= P
1
_
[0w]
[2w]
_
where P
1
:=
_
1/2 1/2
1/2 1/2
_
. (24)
As initially noticed by Strichartz et al. [43], the identity S
100
= S
011
plays a crucial role in
the overlapping situation involved in the self-similarity of . It follows from the same kind
of elementary computation which yields (24) that
_
[01w]
[21w]
_
=
1
4
P
1
_
[0w]
[2w]
_
; (25)
_
[00w]
[20w]
_
=
1
4
P
0
_
[0w]
[2w]
_
where P
0
:=
_
1 0
1 1
_
; (26)
_
[02w]
[22w]
_
=
1
4
P
2
_
[0w]
[2w]
_
where P
2
:=
_
1 1
0 1
_
. (27)
By induction using (23), (24), (25), (26) and (27), one gets for any
0

n1
{0, 1, 2}
n
,
[
0

n1
] =
1
3
1
4
n1
t
V
0
P
0
P
n1
V, (28)
where
V
0
:=
_
1
0
_
, V
2
:=
_
0
1
_
, V
1
=
_
1/2
1/2
_
.
Similarly to the (2, 3)-Bernoulli convolution, the Erd os measure is considered as a
probability measure on S
1
; moreover, the transformations R
0
, R
1
and R
2
are identied to
the local inverses of a 3-f.c.t. of S
1
denoted T which is coded by the full shift :
3

3
and associated to the volume-derivative potential
0
:
3
R, such that
0
() = log
2
1
[[0]]
() +log
3
1
[[1]]
() + log
2
1
[[2]]
().
We now consider the intervals [w] (w {0, 1, 2}
) as subsets of S
1
in such a way that
[0] := [0,
2
[, [1] := [
2
, [, [2] := [, 1[ (dening the one-step basic intervals of T ) and
for each word
0

n1
{0, 1, 2}
n
(n 1)
x [
0

n1
] T
k
(x) [
k
] k = 0, . . . , n 1,
(dening the n-step basic intervals of T ).
By (28) one gets that [10
n
] = (2n+4)/(34
n+1
) and thus is not a Gibbs measure with
respect to T . Following the same idea as in the case of the (2, 3)-Bernoulli convolution,
we introduce the T -invariant probability measure
, such that
[
0

n1
] =
1
2
1
4
n
t
VP
0
P
n1
V. (29)
The same argument leading to Proposition 2.1 gives the following.
PROPOSITION 2.7. For any
3
, and any integer n 1,
2
n +2

[
0

n1
]
[
0

n1
]
4.
In order to establish the following theorem, we use formula (29) dening
together
with Lemma 2.3 and we apply the argument of the uniform convergence of the n-step
potentials (Lemma 1.2) in a similar way leading to Theorem 2.2.
THEOREM 2.8. The T -invariant probability measure
dened on S
1
by (29) is a
g-measure of the normalized potential :
3
R such that
() =
_
_
log(1/4), if n() = () = 0,
log([1; a
1
, . . . , a
n()
+1]/4), if 0 < n(), () < ,
log([1; a
1
, . . . ]/4), if n() = () = .
As a corollary of Proposition 2.7 and Theorem 2.8, one has the following.
THEOREM 2.9. The Erd os measure is a weak Gibbs measure of .
The case of the Erd os measure is similar to the one of the (2, 3)-Bernoulli
convolution studied in 1.1; since is a weak Gibbs measure of the potential , it
satises the multifractal formalism as stated in Theorem A
and, by Theorem C, its

L
q
-spectrum coincides with the concave function
, a solution of the implicit equation

P
(q,
(q)) = 0. Theorem 2.10 below is the analog of Theorem 2.5: it proves that is
not real-analytic at a critical point q
c
< 0, meaning that q
c
is a phase transition point.
THEOREM 2.10. The L
q
-spectrum of the Erd os measure is a concave function for which
there exists a real number q
c
< 0 such that (q) = q log 2/log if and only if q q
c
.
Proof. The Erd os measure being a weak Gibbs measure of , one deduces from
Theorem C that (q) =
(q): this means that, for any q R,

(q) = max{t; P
(q, t ) 0} = min{t; P
(q, t ) 0}. (30)

For
0
:= log 2/log one can check that q(0)
0
q
0
(0) = 0, for any q R: it follows
from the variational principle that
0 = h
(
0
) +
_
{q()
0
q
0
()} d
0
() P(q
0
q
0
) =: P
(q,
0
q),
which implies that (q)
0
q. We now prove that (q)
0
q when q is sufciently
close to . Using the fact that
is also a weak Gibbs measure of , it is easily seen that

P
(q, t ) = lim
n
1
n
log

Z
n
(q, t ) where

Z
n
(q, t ) :=
w{0,1,2}
n
[w1]
q
|[w1]|
t
. (31)
Notice that any word w {0, 1, 2}
n
is associated to a unique sequence of (possibly empty)
words m
1
, . . . , m
k
(1 k n) on the alphabet {0, 2} such that w1 = m
1
1 m
k
1; since
[w1] =
[m
1
1]
[m
k
1] and |[w1]| = |[m
1
1]| |[m
k
1]|, one has
Z
n
(q, t )
k=1
m
1
,...,m
k
m
i
{0,2}
[m
1
1]
q
|[m
1
1]|
t
_

_
[m
k
1]
q
|[m
k
1]|
t
_
k=1
_

m{0,2}
[m1]
q
|[m1]|
t
_
k
(recall that {0, 2}
:= {}
n=1
{0, 2}
n
). However, according to Lemma 2.3, one gets
m{0,2}
[m1]
q
|[m1]|
0
q
= 2
q
+2
k=1
a
1
a
k
>0
(
t
VQ
a
1
Q
a
k
V )
q
,
which implies that

Z
n
(q,
0
q) is bounded for q < q
0
, with a small enough q
0
given by
Lemma 2.6: by (31) one has P
(q,
0
q) 0 and thus (q)
0
q when q q
0
; since
is concave with (0) = 1, there exists q
c
< 0 such that (q) =
0
q if and only if
q q
c
. 2
Remark 2.11. (1) Starting from Denition 1.3 of , it is proved in [13] that is not
differentiable at q
c
; according to our terminology and the fact that =
, this means
that q
c
is a critical value of a rst-order phase transition: we include in Appendix A a
proof of this result, based on the approach developed in [13].
(2) The phase transitions occurring in the multifractal formalism of the (2, 3)-Bernoulli
convolution and the Erd os measure are to be related to the problem of phase transition on
one-dimensional lattice systems and one-sided full-shift respectively studied in [9, 15] and
[18] (see also [19] and [30, 31] for the relationship with the multifractal formalism).
(3) We point out the similarity between the potentials associated with the (2, 3)-
Bernoulli convolution (Theorem 2.2), the Erd os measure (Theorem 2.8) and the potentials
considered in [41].
(4) The L
q
-spectra (for q > 0) and the sets of possible local dimensions for a special
class of self-similar measures with overlaps were studied in [12, 20]; the Hausdorff
dimension of the corresponding self-similar sets was determined in [38].
3. Proof of the multifractal theorems
3.1. Proof of Theorem A: lower bound. For any probability measure , -invariant
on
d
, we denote by G
() the set of -generic points of

d
, meaning that G
()
if and only if S
n
f ()/n tends to (f ), for any real-valued continuous function f dened
on
d
. Since the full-shift
d
satises the specication property, the set G
() is never
empty (especially when is not ergodic).
PROPOSITION 3.1. Let T be a d-f.c.t. of S
1
and : S
1

d
the associated coding map;
for any -invariant measure (not necessarily ergodic) on
d
, one has
dim
H

1
(G
()) =
h
()
(
0
)
.
Hint. For any Borel set M
d
, we denote by (M) the Lebesgue measure of
1
(M);
the probability is non-atomic and fully supported by
d
: by setting for any ,
d
,
d
(, ) =
_
_
_
1, if
0
=
0
,
inf{[[
0

n1
]]; [[
0

n1
]]}, if
0
=
0
,
one denes a metric d
compatible with the product topology on

d
. Let dim
be the
-Billingsley dimension [1], that is, the Hausdorff dimension on the metric space (
d
, d
).
Given a probability measure -invariant on
d
, we claim that
dim
() =
h
()
(
0
)
. (32)
When is ergodic, this formula can be deduced by a classical argument using the
ShannonMcMillan Theorem and the fact that (G
()) = 1. However, (32) is still

valid when is -invariant without being ergodic: this follows from [5, Theorems 7.1 and
7.2] and [29, Th eor` eme 2], where the fact that is a Gibbs measure is needed.
Actually, it is easily seen from(4) that is a Gibbs measure of
0
in the sense that there
exists a constant K > 1 such that, for any
d
and any integer n 1,
1
K
[[
0

n1
]]
exp(S
n
0
())
K;
then we can apply [37, Th eor` eme 1.2.2], ensuring that dim
H

1
(M) = dim
M, for any
M
d
: according to (32), one concludes that Proposition 3.1 holds.
We now prove a lemma that gives the characterization of the points where
is not
differentiable (i.e. the rst-order phase transition points); it is essentially a corollary of
the tangency property of the topological pressure [45, Theorem 9.15] saying that is an
equilibrium state of the potential :
d
R if and only if P( + ) P() (),
for any real-valued continuous function dened on
d
.
LEMMA 3.2. Given a negative potential :
d
R and
the concave map such that

P
(q,
(q)) = 0, the following two propositions hold:

(i) if I
q
is the simplex of the equilibrium states of q
(q)
0
then
(q
) = sup
I
q
_
()
(
0
)
_
and
(q
+
) = inf
I
q
_
()
(
0
)
_
;
(ii) if I
(
d
) is the simplex of the -invariant probability measures on
d
then
inf
I
(
d
)
_
()
(
0
)
_
= lim
q+
(q)
q
=:
and
sup
I
(
d
)
_
()
(
0
)
_
= lim
q
(q)
q
=: .
Proof. (i) Let q > 0; from the denition of the concave map
: R R one has
P(q
(q)
0
) = 0; similarly, for any h > 0
0 = P((q +h)
(q + h)
0
) = P(q
(q)
0
+h( (
(q
+
) +
h
)
0
)),
where we write
(q + h) =
(q) + h
(q
+
) + h
h
with
h
tending to 0 when h tends
to 0. In conclusion, one has the equation
P(q
(q)
0
+h( (
(q
+
) +
h
)
0
)) P(q
(q)
0
) = 0;
which, by the tangency property of the pressure, ensures that
() (
(q
+
) +
h
)(
0
) 0,
for every I
q
, that is,
(q
+
) ()/(
0
)
h
. When h tends to 0, one gets
(q
+
) ()/(
0
) and thus
(q
+
) inf
_
()
(
0
)
; I
q
_
.
By the same argument one gets
sup
_
()
(
0
)
; I
q
_

(q
).
It is clear that if
is differentiable at q then
(q) = ()/(
0
), for every I
q
.
Since
is not differentiable on an (at most) countable subset of R, there exists a sequence

of real numbers q
n
> q which tend to q and such that
is differentiable at q
n
for any n.
Let
n
I
q
n
and assume (by compactness) that
n
tends to a -invariant probability
measure in the weak- sense; since the potentials q
n

(q
n
)
0
converge uniformly
to q
(q)
0
, it follows from the variational principle and the upper semi-continuity
of the entropy that I
q
. Using the fact that
is concave,
(q
+
) = lim
n
(q
n
) = lim
n
n
()
n
(
0
)
=
()
(
0
)
,
and then
(q
+
) = inf{()/(
0
); I
q
}. The same argument applies to
(q
).
(ii) We prove the assertion for (the same argument applies to ). By the concavity of
it is clear that lim

q+

(q
+
); this fact together with part (i) yields
inf
_
()
(
0
)
; I
(
d
)
_
.
Moreover, if I
(
d
) then, by the variational principle, h
() + (q
(q)
0
)
P
(q,
(q)) = 0, for any q > 0. Since

0
is negative and uniformly bounded away
from 0,
(q)
q
()
q(
0
)
+
()
(
0
)
()
(
0
)
.
Therefore = lim
q+
(q)/q ()/(
0
), that is,
inf
_
()
(
0
)
; I
(
d
)
_
. 2
LEMMA 3.3. For any < < , let q
R such that
(q
+
) (q
); then there
exists I
q
such that ()/(

0
) = and for such a measure one has
()
(
0
)
= sup{q R; q
(q)} =:
().
Proof. We consider the map : ()/(
0
) which is continuous on the convex set
of probability measures on
d
, endowed with the weak- topology. Given < < ,
the set (I
q
) is a non-empty closed interval of R, for I

q
is a non-empty closed convex

subset of the -invariant measure. By Lemma 3.2, [
(q
+
), (q
)] = (I
q
) and thus
there exists I
q
(not necessarily ergodic) such that () = ()/(

0
) = .
We now apply the variational principle: on the one hand, for every q R,
h
() +(q (q)
0
) P(q (q)
0
) = 0;
on the other hand, since I
q
,
h
() + (q
(q
)
0
) = P(q
(q
)
0
) = 0.
Therefore, h
()/(
0
) = q
(q
) and h
()/(
0
) q (q), for any
q R, which means, by denition, that
() = h
()/(
0
). 2
We are now in a position to prove the lower bound
() dim
H
E(|M), for
< < . Given x S
1
with (x) = , the Gibbs property of the Lebesgue measure
(with respect to the volume-derivative potential
0
) ensures the existence of C > 1 such
that, for any integer n,
S
n
0
() log C log |I
n
(x)| S
n
0
() +log C.
Likewise, by the weak Gibbs properties of , there exists a sub-exponential sequence of
real numbers K(n) > 1 such that
S
n
() log K(n) log (I
n
(x)) S
n
() +log K(n).
The potential being negative, () < < 0 uniformly on
d
and S
n
()/n < for
any n; moreover, since (1/n) logK(n) tends to 0 when n goes to innity, one deduces
that S
n
() + log K(n) < 0, when n is sufciently large. For the same reason that
0
is
negative, one also has S
n
0
() +log C < 0, when n is sufciently large, and thus
S
n
() +log K(n)
S
n
0
() log C
log (I
n
(x))
log |I
n
(x))|

S
n
() log K(n)
S
n
0
() +log C
. (33)
Now assume that = (x) G
(), where I
q
(given by Lemma 3.3), satises

()/(
0
) = ; by the denition of G
(), one has lim

n
S
n
()/n = () < 0 and
lim
n
S
n
0
()/n = (
0
) < 0, so that
lim
n
log (I
n
(x))
log |I
n
(x)|
=
()
(
0
)
= ;
the point x being arbitrarily taken in
1
(G
()), one deduces that
1
(G
()) E(|M).
Using successively Lemma 3.2 and Proposition 3.1, one concludes
() =
h
()
(
0
)
= dim
H

1
(G
()) dim
H
E(|M).
The required lower bound is proved.
3.2. Proof of Theorem A: upper bound. The following proof is essentially the argument
given by Brown et al. [4] that we adapt to our framework. Notice that the hypothesis
that is a weak Gibbs measure of the potential is implicitly needed to ensure that the
concave map
is well dened. We use the notions of -packing, box dimension (denoted

dim
B
) and packing dimension (denoted dim
P
), presented in Appendix B. We shall prove,
in Theorem 3.4 below, a stronger result than the upper-bound dim
H
E(|M)
()
involved in Theorem A; actually in place of the level sets E(|M), we shall deal with
what we call the fat level sets, dened by setting, for any R,
F
+
(|M) :=
_
x S
1
: limsup
n
log (I
n
(x))
log |I
n
(x)|

_
and
F
(|M) :=
_
x S
1
: liminf
n
log (I
n
(x))
log |I
n
(x)|

_
.
THEOREM 3.4. Let M be the Markov net of a regular d-f.c.t. T and be a weak Gibbs
measure of a negative potential :
d
R. Then, the following propositions hold:
(i) if <
(0
) then dim
P
F
+
( |M)
();
(ii) if
(0
+
) < then dim
P
F
( |M)
().
Proof. We prove part (i) while part (ii) can be handled in a similar way. See Figure 1
for the graph of
(q). To begin with, notice that, for
(0
+
)
(0
), one has
() =
(0) = 1, so that the upper bound dim

P
F
+
( |M)
() is trivial in that
case. We now consider that < <
(0
+
); given such that < <
(0
+
), we
dene
F
n
+
() := {x S
1
; |I
n
(x)|
(I
n
(x))},
for any n 0, so that
F
+
(|M)
_
m=0
_
n=m
F
n
+
(). (34)
Let m 0 and > 0 be arbitrarily chosen; we consider J =

n=m
J
n
, where J
n
A
n
d
and such that {[w]; w J} is an -packing of
_
n=m
F
n
+
(). By denition of an -packing,
it is clear that, for any word w J, there exists x
_
n=m
F
n
+
() such that [w] = I
n
(x)
for some n m, so that |[w]|
[w]. Accordingly, for any 0 and q > 0, one can

write
wJ
|[w]|
n=m
wJ
n
|[w]|
q
/|[w]|
q
n=m
wJ
n
[w]
q
/|[w]|
q
n=1
Z
n
(q, q ),
where Z
n
(, ) is as in (11). Since <
(0
+
), there exists q
0
> 0 with q
0

(q
0
)
=
(). When >
() one has q
0
<
(q
0
), which, by denition of
, implies
that
P
(q, q
0
) = P(q
0
(q
0
)
0
) < 0.
t =
(q)
q
0

()
(q
0
)
()
1
q
0
1
q
(q
0
, q
0
)
q
0

FIGURE 1. We represent here the graph of the concave map q
(q) (with a discontinuity of the derivative at

q = 0) and the gray area is the open domain of the (q, t ) RRfor which P
(q, t ) < 0; when < <
(0
+
),
the straight line with equation t = q
() is tangent to the graph of the map
at the point with abscissa

q
0
> 0, and P
(q
0
, q
0
) < 0 when >
().
Hence, there exists < 0 and a rank n
0
0 such that Z
n
(q
0
, q
0
) exp(n)
for any n n
0
; it follows that

wJ
[w]
< , for any >
(). This proves

that dim
B
_
n=m
F
n
+
()
(), for any m 0, and from (34) one concludes that

dim
P
F
+
(|M)
(). Since is an arbitrary real number such that < <
(0
+
),
the continuity of
at yields the desired upper bound stated in (i); the proof of the
theorem is complete. 2
3.3. Proof of TheoremB: upper bound. There exist conditions on the d-f.c.t. T ensuring
that E() = E(|M): for instance, this occurs when T is differentiable at the point 0 := 0
(mod 1). In our setting, the inclusion E() E(|M) does not hold in general; however,
we shall establish the upper bound dim
H
E()
()
= dim
H
E( | M) (see part (iii)
of Proposition 3.5 below), by means of the upper bounds in Theorem 3.4 involving the fat
level sets F
(|M) and F
+
(|M).
PROPOSITION 3.5. Let M be the Markov net of a regular d-f.c.t. T ; if is a Borel
probability measure then the following two properties hold:
(i) E() F
(|M);
(ii) dim
H
E() dim
H
F
+
( |M);
if in addition is a weak Gibbs measure, then one has:
(iii) dim
H
E()
()
.
In order to prove Proposition 3.5 we use ideas taken from the approach of Hofbauer
in [19]. Let dist(, ) be the usual distance on S
1
; for any x S
1
and 0 < d < 1/2, we
dene
N
d
(x) := {n N; dist(T
n
(x), 0) > d} and M
d
:= {x S
1
; #N
d
(x) < }
(#A stands for the cardinality of the countable set A.)
LEMMA 3.6. [19, Lemma 13] Suppose that T is a regular d-f.c.t. of the torus S
1
; then
lim
d0
dim
H
M
d
= 0.
We dene the n-step variation of the volume-derivative potential
0
, by setting
_
V
0
(
0
) := max{
0
()
0
(); ,
d
},
V
n
(
0
) := max{
0
()
0
();
0

n1
=
0

n1
} for n 1,
(35)
and we consider in addition
n
(
0
) := V
1
(
0
) + +V
n
(
0
). (36)
Notice that, since
0
is continuous on
d
, the variation V
n
(
0
) tends to 0 when n goes
to innity and, by a classical lemma on Ces` aro averages,
n
(
0
)/n tends to 0 as well.
We denote by d
n
(x) the distance of x to the boundary of the basic interval I
n
(x); the
following lemma shows that the ratio |I
n
(x)|/d
n
(x) is a sub-exponential sequence, when
x S
1
\M
d
, for some 0 < d < 1/2.
LEMMA 3.7. Given any x S
1
\M
d
, there exists a rank n
0
(depending on x) such that,
for any n n
0
,
0 < log
_
|I
n
(x)|
d
n
(x)
_

n
(
0
) + log d.
Proof. By an application of the Mean Value Theorem, it is clear that
log |I
n
(x)| sup
yI
n
(x)
{S
n
0
(y)};
using the fact that x / M
d
, another application of the Mean Value Theorem yields
log d
n
(x) inf
yI
n
(x)
{S
n
0
(y)} +log d,
and one concludes using the denition of
n
(
0
). 2
Proof of Proposition 3.5. (iii) When is a weak Gibbs measure, the upper bound
dim
H
E()
()
is a trivial consequence of Theorem 3.4 together with parts (i) and

(ii), which we now establish.
(i) Fix x S
1
and set
n
:= |I
n
(x)|, for any n 0. Then, for any rank n 0,
log (B
n
(x))
log
n
log (I
n
(x))
log |I
n
(x)|
,
ensuring that E() F
(|M).
(ii) Since T is regular and uniformly expanding,
0 < := inf
d
{exp(
0
())} sup
d
{exp(
0
())} =: < 1
and, by a classical application of the Mean Value Theorem,
n
|I
n
(x)|
n
, for any
x S
1
and any integer n 0. Let x S
1
\M
d
and denote by d
n
(x) the distance of
x to the boundary of the basic interval I
n
(x); it follows from the denition of M
d
that
B
d
n
(x)
(x) I
n
(x) and thus
log (B
d
n
(x)
(x))
logd
n
(x)
log (I
n
(x))
log d
n
(x)
=
log (I
n
(x))
log |I
n
(x)|
_
1 +
log(d
n
(x)/|I
n
(x)|)
log |I
n
(x)|
_
1
. (37)
Since one clearly has (1/n) log|I
n
(x)| log > 0, one deduces from Lemma 3.7
that
log(d
n
(x)/|I
n
(x)|)
log |I
n
(x)|
=
(1/n) log(|I
n
(x)|/d
n
(x))
(1/n) log |I
n
(x)|

(
n
(
0
) + log d)/n
log
,
which according to (37) yields
log (B
d
n
(x)
(x))
log d
n
(x)
log(I
n
(x))
log |I
n
(x)|
_
1
(
n
(
0
) +log d)/n
log
_
1
,
and thus E() (S
1
\M
d
) F
+
(|M), for
n
(
0
)/n tending to 0. Let > 0 be
arbitrarily given; by Hofbauers lemma there exists d > 0 such that dim
H
M
d
, so
that
dim
H
E() = dim
H
E() (S
1
\M
d
) +dim
H
E() M
d
dim
H
F
+
(|M) +.
Part (ii) is established since > 0 is arbitrarily chosen. 2
3.4. Proof of Theorem B: lower bound. In order to prove the lower bound
()

dim
H
E() we use the underlying Markov structure to construct a slim level set S(|M),
a subset of E() having the specied Hausdorff dimension, say
()
. Usually this is
achieved by constructing a Frostman measure on the level set E(); we shall use this
approach with the additional difculty that the measure to consider may not give a positive
measure to E() (this is related to the fact that may correspond to a rst-order phase
transition point).
The proof of the desired lower bound is a consequence of Theorem 3.8 below. Let us
split any
d
into an innite sequence of nite words, say
1
,
2
, . . . , so that
(by concatenation) =
1

2
and with the additional condition that the length of
each word
n
(n 1) is exactly n. Then we dene the one-to-one map :
d

d
by
setting
() := 1
1
1
2
1 1
n
1
n+1
1 =:
= (
i
)
i=0
. (38)
THEOREM 3.8. Let be a -invariant measure on
d
(d > 2); then the following hold:
(i) G
() := (G
()) G
();
If T is a 2-f.c.t. one can consider T T .
(ii) dim
() = dim
();
(iii) for any x S
1
such that (x) (
d
) and any probability measure on S
1
,
DIM
(x |M) = DIM
(x) = .
We claim that, for each < < , the upper bound
()
dim
H
E() is a
consequence of Theorem 3.8. To see this, let q
R be such that
(q
+
(q
+
);
using Lemma 3.3, we consider an equilibrium state of the potential q
(q
)
0
, say
, for which
()/
(
0
) = so that
1
(G
)) E(|M) and
dim
) =
h
(
0
)
=
()
.
We dene the slim level set S(|M) :=
1
(G
)); by part (iii) of Theorem 3.8,

one has DIM
(x|M) = DIM
(x) for any x S(|M), and since

1
(G
))
E(|M), one deduces from part (i) of Theorem 3.8 that S(|M) E(); by part (ii)
of Theorem 3.8, one concludes that
()
= dim
) = dim
) = dim
H
S(|M) dim
H
E(),
completing the proof of the desired lower bound.
Therefore it remains to establish Theorem 3.8. To begin with we shall prove the
following lemma.
LEMMA 3.9. Let T be a regular d-f.c.t. of S
1
(d > 2) and a Borel probability measure.
Suppose that x is a point of S
1
such that T
m
k
(x) [1] (k 1), where m
1
, m
2
, . . . form a
strictly increasing sequence of integers; if lim
k
log |I
m
k
(x)|/log |I
m
k+1
(x)| = 1, then
DIM
(x |M) = DIM
(x) = .
Proof. Let x S
1
such that T
m
k
(x) [1] (k 1); for = (x), it is clear that
I
m
k
+1
(x) = [
0

m
k
1
1] I
m
k
(x) = [
0

m
k
1
].
The Lebesgue measure being Gibbs with respect to the Markov net associated to T , there
exists a constant 0 < c < 1 such that, for any i = 0, 1, 2,
|[
0

m
k
1
i]| c|[
0

m
k
1
]| =: r
k
,
ensuring that B
r
k
(x) I
m
k
(x). Moreover, I
m
k
+p
(x) B
r
k
(x) for some constant integer
p > 0 and thus the following sequence of inclusions arises:
I
m
k
+p
(x) B
r
k
(x) I
m
k
(x) B
r
k
/c
(x).
If DIM
(x|M) = then it follows that log (B

r
k
(x))/log r
k
tends to , when k goes to ;
notice that for r
k+1
r r
k
log r
k
log r
k+1
log (B
r
k
(x))
log r
k
log (B
r
(x))
log r
log r
k+1
log r
k
log (B
r
k+1
(x))
log r
k+1
,
and from the assumption that lim
k
log |I
m
k
(x)|/log |I
m
k+1
(x)| = 1, one gets DIM
(x) = .
Conversely, if DIM
(x) = , a similar argument proves that DIM
(x|M) = . 2
Here, we use again the fact [37, Th eor` eme 1.2.2] that dim
H

1
(M) = dim
M, for any M
d
.
Proof of Theorem 3.8. For any integer p 1, dene n
p
:=

p1
k=0
k = p(p 1)/2 and
n
p
:= n
p
+ p = n
p+1
. For
d
given, recall that () =
= (
i
)
i=0
and notice
that, by construction,
n
p
[[
p
]],
n
p
1
[[1]] and
n
[[
p
]].
(i) Assume that f is a real-valued continuous function dened on
d
. For any
G
() and any integers n, p with p 2 and 0 n < p, one has

S
n
p
+n
f (
) S
n
p
+n
f ()
=
p1
k=1
{f (
n
k
1
) +S
k
f (
n
) S
k
f (
n
k
)}
+{f (
n
p
1
) +S
n
f (
n
) S
n
f (
n
p
)} S
p
f (
n
p
+n
),
=
p1
k=1
{S
k
f (
n
) S
k
f (
n
k
)} + {S
n
f (
n
) S
n
f (
n
p
)}
+
p
k=1
f (
n
k
1
) S
p
f (
n
p
+n
),
which gives, with V
k
(f ) dened as in (35) and
n
(f ) as in (36),
1
n
p
+n
|S
n
p
+n
f (
) S
n
p
+n
f ()|
2
p(p +1)
_
p
k=1
k
j=1
V
j
(f ) +pV
0
(f )
_
2(
p
(f ) +V
0
(f ))
p + 1
.
With
k
(f ) :=
k
(f ) +V
0
(f ), a straightforward computation yields, for any n 1,
1
n
|S
n
f (
) S
n
f ()|
2
n
(f )
n
+1
, (39)
where
n
is the integral part of (
8n +1 1)/2. The variation V

n
(f ) tends to 0 when n
goes to innity, as well as both
n
(f )/n and
n
(f )/n; this completes the proof of (i).
(ii) Let n :=
n
(
n
1)/2 + r, where r is an integer such that 0 r <
n
and let
n
= n +
n
. Since the Lebesgue measure is a Gibbs measure of the volume-derivative
potential
0
, one gets
log |[
1
]| S
n

0
(
) log K
for the constant K > 1 given in (4). Thereafter, from (39) one deduces that
log |[
1
]| S
n
0
() +S
0
(
n
) 2(n +
n
)
n
(
0
)
n
1
log K
S
n
0
()
_
1

n
n sup(
0
)
2(n +
n
)
n sup(
0
)
n
(
0
)
n
1

log K
n sup(
0
)
_
,
with sup(
0
) := sup{
0
();
d
} < 0, that is
log |[
1
]| S
n
0
()(1 A
n
), (40)
where
A
n
:=

n
n sup(
0
)
+
2(n +
n
)
n sup(
0
)
n
(
0
)
n
1
+
logK
n sup(
0
)
is clearly a quantity that tends to 0 when n goes to innity. We use again the Gibbs property
of the Lebesgue measure, say,
log |[
0

n1
]| S
n
0
() + log K,
which yields the following sequence of inequalities:
log |[
0

n1
]|
S
n
0
()
1 +
logK
S
n
0
()
1 +
log K
n sup(
0
)
. (41)
Since
0
< 0, it is clear that B
n
:= log K/(n sup(
0
)) tends to 0 when n goes to innity
and from (41) (with n large enough),
S
n
0
() log |[
0

n1
]|/(1 +B
n
). (42)
It follows from (40) and (42) that there exists a sequence of real numbers b
1
, b
2
, . . . with
lim
k
b
k
= 0 and such that
log |[
1
]| log |[
0

n1
]|
_
1 +A
n
1 + B
n
_
. (43)
In conclusion, for any > 0, there exists > 0 such that
d
(, ) d
(, )
1+
d
), (44)
and one deduces that dim
() = dim
(), for any -invariant measure .

(iii) Let x S
1
such that (x) = (
d
). For m
k
:= n
k
1 one has T
m
k
(x) [1]
(i.e.
m
k
[[1]]) and lim
k
log |I
m
k+1
(x)|/log |I
m
k
(x)| = 1. The result follows by an
application of Lemma 3.9. 2
3.5. Proof of Theorem C. Let F
0
:= {S
1
} and suppose that F
n
(n > 0) is a partition of
S
1
by non-empty intervals. We say that F :=

n
F
n
is a net if any interval in F
n
is the
union of intervals in F
n+1
. For x S
1
we denote by I
n
(x) the interval in F
n
such that
x I
n
(x), and we suppose that lim
n
|I
n
(x)| = 0. Moreover, F is said to be regular if there
exists a constant c > 1 such that, for any J, J
with F
n+1
J
J F
n
(n 0), one has
|J|/|J
| c. Given 0 < r < 1 and x S

1
, let n
r
(x) be the rank such that |I
n
r
(x)
(x)| < r
and |I
n
r
(x)1
(x)| r; furthermore, one denes an equivalence relation on S
1
by setting
x
r
y if and only if y I
n
r
(x)
(x): the so-called r-Moran partition F
r
is the partition of S
1
generated by this equivalence relation. Notice that, under the condition of F to be regular,
any interval J F
r
has approximately length r, in the sense that
r/c |J| < r.
We refer to the book of Pesin [35] for a systematic presentation of the previous framework.
In order to prove Theorem C we rst establish the following theorem.
THEOREM 3.10. Let F be a regular net of S
1
and a probability measure on S
1
. If is
supposed to be of full support, then, for any q R,
(q) := liminf
r0
log
JF
r
(J)
q
log r
.
Proof. According to Denition 1.3, one has (q) = liminf
r0
log Z(r)/log r, with
Z(r) := inf
_
i
(B
i
)
q
; {B
i
}
i
_
,
the inmum being taken over the r-covers of S
1
; for any r > 0, dene
Z(r) :=
JF
r
(J)
q
.
Let m
r
be the integral part of 1/r and consider B
r
:= {B
i
}
m
r
i=0
where B
i
:= [ir, (i + 1)r],
(0 i < m
r
) and B
m
r
= [1 r, 1] (B
r
is an r/2-cover of S
1
).
First, suppose that q < 0; by denition, one has |J| < r/2 for any J F
r/2
and thus
each ball B
i
B
r
contains at least one interval in F
r/2
, say J
i
, so that
Z(r/2)
i
(B
i
)
q
2
JF
r/2
(J)
q
2

Z(r/2)
(the factor 2 appears because one may have J
m
r
1
= J
m
r
). By the regularity of F one
also has 2r |J|, for any J F
2cr
(where c > 1 stands for the constant involved in the
regularity property of F); hence, each J F
2cr
contains at least one B
i
B
r
, implying
that

Z(2cr) 2Z(r/2); one deduces that (q) is the lower limit of the ratio log

Z(r)/log r
when r tends to zero.
We now consider the case of q 0. By the regularity of F, one has r |J| for any
J F
cr
; hence, each B
i
B
r
intersects no more than two elements of F
cr
, one of them,
say J
i
, satisfying (B
i
) 2(J
i
). One may have J
i
= J
j
for i = j, but since each interval
in F
cr
intersects at most N elements of B
r
, for some N independent of r, one gets
i
(B
i
)
q
2
q
N
i
(J
i
)
q
2
q
N
JF
cr
(J)
q
, i.e.
1
2
q
N
Z(r/2)

Z(cr). (45)
Suppose that C is an arbitrary r-cover of S
1
. It follows fromBesicovitchs covering Lemma
(cf. [28, p. 30]) that there exists a sub-cover C
of C, such that each point x S

1
belongs
to at most three balls in C
. One can check that, for 0 i m

r
, the number of C C
which intersects B
i
is bounded by 6. Using again the fact that J F
cr
intersects at most
N elements in B
r
, it is clear that J intersects at most 6N elements of C
; since any ball in

C
intersects no more than three elements of F

cr
,
JF
cr
(J)
q
3(6N)
q
CC
(C)
q
3(6N)
q
CC
(C)
q
.
Taking the inmum over the r-cover C leads to

Z(cr) 2(6N)
q
Z(r). With (45), one
concludes that (q) is the lower limit of the ratio log

Z(r)/log r when r tends to zero. 2
We now turn to the proof of Theorem C itself, which we split into two steps (the rst
one being inspired by an argument in [36]).
First step. To begin with, suppose that is a Gibbs measure of a H older continuous
potential (which, according to our denition, implies that P() = 0). For any q R
and any
d
, the trivial identity
exp(S
n
q
()) =
exp(qS
n
())
exp(
(q)S
n
0
())
(46)
holds for
q
:= q
(q)
0
. Denote by
q
the unique T -ergodic Gibbs measure
associated to the H older continuous potential
q
; since and the Lebesgue measure are
respectively Gibbs measures of and
0
, the identity (46) gives, for any word w,
q
[w]/R

[w]
q
|[w]|
(q)
R
q
[w] (47)
for some constant R
> 1. The Markovian net M being regular, it is possible by

Theorem 3.10 to consider the L
q
-spectrum dened by the mean of the Moran partitions
M
r
(0 < r < 1). By a summation of (47) over the [w] M
r
, there exists a constant
R
> 1 such that

1/R

Z(r, q)/r
(q)
R
, where

Z(r, q) =
JM
r
(J)
q
;
taking the limit when r tends to 0, one concludes that (q) =
(q), for any q R.

Second step. We now turn to the general case. Let us consider a sequence of H older
continuous potentials
k
which are uniformly convergent to . Since P() = 0, one has
|P(
k
)|
k

and thus (
k
P(
k
))
2
k

; hence one can

assume that P(
k
) = 0. We denote by
k
the unique -ergodic Gibbs measure of
k
; by
the rst step of the proof,
k
(q) =
k
(q), for any q R (
k
denotes the L
q
-spectrum
of
k
). Moreover, it follows from the classical properties of the topological pressure that
k
tends to
uniformly on the compact intervals. It remains to prove the pointwise

convergence of
k
(q) to (q). Given > 0, one has
k

whenever n is large
enough; therefore, for any integer m > 0 and any word w of length m, one has
[w] K(m)C
k
exp(2m)
k
[w] (48)
where m K(m) (respectively C
k
) is the sub-exponential sequence (respectively
constant) which characterizes the weak Gibbs property of (respectively the Gibbs
property of
k
, for k 1). Let N
r
be the maximal length of the words w such that
[w] M
r
; for any k 1 we consider the partition function

Z
k
(r, q) :=

JM
r

k
(J)
q
,
so that from (48)
Z(r, q) K(N
r
)C
k
exp(2N
r
)

Z
k
(r, q),
or equivalently
1
log r
log

Z(r, q)
1
log r
log

Z
k
(r, q) +
_
N
r
log r
log K(N
r
)
N
r
+
C
k
log r
+
2N
r
log r
_
.
When r tends to 0, one gets (q)
k
(q) 2a, where a is a constant such that
N
r
/log(1/r) a for any 0 < r < 1. The symmetric argument yields (q)
k
(q) +2a
and the proof of Theorem C is complete.
Acknowledgements. De-Jun Feng was supported by an HK RGC grant and the Special
Funds for Major State Basic Research Projects and Eric Olivier was also supported by
an HK RGC grant. The second author would like to thank N. Sidorov for pointing out
the problem of the Gibbs properties of the Erd os measure and specially A. Thomas for
enlightening and substantial discussions on this subject; he is also grateful to A. H. Fan,
J. P. Kahane, F. Ledrappier, P. Liardet, J. Peyri` ere and J. P. Thouvenot. The paper was
written during a postdoctoral visit to the Chinese University of Hong Kong and both
authors would like to thank K. S. Lau for his kind invitation and valuable discussions.
A. Appendix. First-order phase transition
It was rst proved in [13] that the L
q
-spectrum of the Erd os measure is not differentiable
at the critical value q
c
dened in Theorem 2.10; the present appendix is devoted to a self-
contained proof of this result.
THEOREM A.1. The L
q
-spectrum is not differentiable at the critical value q
c
.
Any word w {0, 1, 2}
n
is associated to a unique sequence of (possibly empty) words
m
1
, . . . , m
k
(1 k n) on the alphabet {0, 2} such that w1 = m
1
1 m
k
1; from(29) one
has
[w1] =
[m
1
1]
[m
k
1]; since |[w1]| = |[m
1
1]| |[m
k
1]|, one deduces that
[m
1
1 m
k
1]
q
|[m
1
1 m
k
1]|
t
=
k
i=1
(
t
VP
m
i
V )
q
(
t
/2
q
)
2|m
i
|+3
. (49)
Recall that
P
(q, t ) = lim
n
1
n
log

Z
n
(q, t ) where

Z
n
(q, t ) :=
w{0,1,2}
n
[w1]
q
|[w1]|
t
,
so that, with z
n
(q) :=

w{0,2}
n(
t
VP
w
V )
q
, one gets
Z
n
(q, t ) =
n
k=1
a
1
,...,a
k
0
a
1
++a
k
=n+1k
k
i=1
z
a
i
(q)(
t
/2
q
)
2a
i
+3
. (50)
PROPOSITION A.2.
(i) q
c
< 2.25 and
(ii)

n=0
nz
n
(q) < for any q < 2.25.
We shall give the proof of Proposition A.2 at the end of the present appendix.
PROPOSITION A.3. The following three propositions hold:
(i) X(q) :=
(q)
/2
q
= sup
_
0 x 1;
i=0
z
i
(q)x
2i+3
1
_
;
(ii) q q
c

i=0
z
i
(q) 1;
(iii) q
c
q < 2.25

n=0
z
n
(q)X(q)
2n+3
= 1.
Proof. (i) For x > 0, dene t
q
(x) := q log 2/log +log x/log so that, for any q R,
X(q) = max{x; P
(q, t
q
(x)) 0} = min{x; P
(q, t
q
(x)) 0}.
By denition {0, 2}
0
= {} and P
is the identity matrix, so that z

0
(q) = (
t
VP
V )
q
= 2
q
.
For any (q, x) R [0, 1] we set F(q, x) :=

n
z
n
(q)x
2n+3
( [0, +]) and we
consider, for any q R,
x(q) := sup{x 0; F(q, x) 1} = sup{x 0; F(q, x) < 1}, (51)
the second equality being justied by Abels theorem, for x(q) is bounded by the radius of
convergence of the power series
i
z
i
(q) x
2i+3
. It follows from (50) that
Z
n
(q, t
q
(x)) =
n
k=1
a
1
,...,a
k
0
a
1
++a
k
=n+1k
k
i=1
z
a
i
(q)x
2a
i
+3
k=0
_

n=0
z
n
(q)x
2n+3
_
k
,
that is,
Z
n
(q, t
q
(x))
k=0
(F(q, x))
k
. (52)
Given any x 0 with F(q, x) < 1, the upper bound in (52) implies that

Z
n
(q, t
q
(x))
is bounded and thus P
(q, t
q
(x)) 0: this means that x X(q), and from the second
equality in (51), one deduces that x(q) X(q).
For the converse inequality, let n 1 and r 1 be given so that
nr
m=1
Z
m
(q, t )
m
1
,...,m
n
{0,2}
|m
1
|,...,|m
n
|<r
[m
1
1]
q
|[m
1
1]|
t

[m
n
1]
q
|[m
n
1]|
t
=
_

m{0,2}
,|m|<r
[m1]
q
|[m1]|
t
_
n
=
_
r1
i=0
m{0,2}
i
[m1]
q
|[m1]|
t
_
n
,
which yields
nr
k=1
Z
k
(q, t
q
(x))
_
r1
i=0
z
i
(q) x
2i+3
_
n
. (53)
Let x > x(q) and r
0
>0 such that z
0
:=

r
0
1
i=0
z
i
(q)x
2i+3
> 1; from (53), one deduces
that
lim
n
1
n
log
nr
0
k=1
Z
k
(q, t
q
(x)) log z
0
> 0.
This is inconsistent with the fact that Z
n
(q, t
q
(x)) e
n
for > 0 and any n large
enough and thus P
(q, t
q
(x)) 0 and X(q) x: one concludes that X(q) x(q), for x
can be taken arbitrarily close to x(q). Finally, since (q) q log 2/log , for any q R,
one necessarily has 0 X(q) 1, which complete the proof of (i).
(ii) According to Theorem 2.10, one has q q
c
if and only if
(q)
/2
q
= 1, which
leads to (ii) as a straightforward consequence of (i).
(iii) Notice that the map F : (q, x) F(q, x) is nite and continuous at any point
(q, x) ], 2.25[ [0, 1], for part (ii) of Proposition A.2 ensures that
0 F(q, x)
n
z
n
(q) < ,
whenever q < 2.25 and 0 x 1. We rst assume that q
c
< q < 2.25 (which is
consistent with respect to part (i) of Proposition A.2); it is clear that F(q, ) increases from
0 up to F(q, 1). By (ii), one has F(q, 1) > 1 and thus there exists a unique 0 < x
q
< 1
such that F(q, x
q
) = 1, which also satises x
q
= max{0 x 1; F(q, x) 1}: since
0 X(q) 1, one deduces from (i) that X(q) = x
q
.
We now consider the case when q = q
c
; since X(q
c
) = 1, we need to prove that
F(q
c
, 1) =

n
z
n
(q
c
) = 1. Part (ii) insures that F(q, 1) > 1 for any q
c
< q, which
implies that F(q
c
, 1) 1, by the fact that F is continuous at (q
c
, 1): since (ii) also implies
that F(q
c
, 1) 1, one concludes that F(q
c
, 1) = 1, completing the proof of (iii) and of
the proposition as well. 2
Proof of Theorem A.1. We shall prove that X is not differentiable at q
c
; since by
Proposition A.3, one has X(q) = 1 whenever q q
c
, this will be established if one shows
that X
(q
+
c
) < 0. According to part (iii) of Proposition A.3, for any q
c
q < 2.25, one
has the equation
n=0
z
n
(q
c
)X(q
c
)
2n+3
=
n=0
z
n
(q)X(q)
2n+3
,
or equivalently, using the fact that X(q
c
) = 1,
n=0
(z
n
(q
c
) z
n
(q)) =
n=0
z
n
(q)(X(q)
2n+3
1)
= (X(q) X(q
c
))
n=0
z
n
(q)
2n+2
i=0
X(q)
i
.
Finally one can write
X(q) X(q
c
)
q q
c
=
n=0
(z
n
(q) z
n
(q
c
))/(q q
c
)
n=0
z
n
(q)
2n+2
i=0
X(q)
i
. (54)
If q tends to q
c
with q
c
< q < 2.25, then

2n+2
i=0
X(q)
i
tends to 2n + 3 and since by
Proposition A.2 one has
n=0
z
n
(q
c
) (2n +3) < , one concludes from (54) that
X
(q
+
c
) =
n=0
w{0,2}
n(
t
VP
w
V)
q
c
log(
t
VP
w
V )
n=0
z
n
(q
c
)(2n +3)
< 0. 2
Proof of Proposition A.2. Given (a
n
)
n=1
a sequence of positive integers, it is easily
checked by induction that, for any k 1,
t
VQ
a
1
Q
a
k
V (1 +a
1
) (1 +a
k1
)(2 +a
k
); (55)
likewise, if A(a
1
, . . . , a
2k
) := (1+a
1
a
2
)(1+a
3
a
4
) (1+a
2k1
a
2k
) then, for = 0 or 1,
t
VQ
a
1
Q
a
2k+
V
_
A(a
1
, . . . , a
2k
) if = 0,
A(a
1
, . . . , a
2k
)(1 +a
2k+1
) if = 1.
(56)
In what follows, will stand for the Riemann zeta function (i.e. (x) =

n=1
1/n
x
,
x > 1).
(i) When q < 1 one has (q) 1 < 1 and the following inequalities hold:
n=0
z
n
(q) = 2
q
+2
k=1
a
1
,...,a
k
1
(
t
VQ
a
1
Q
a
k
V )
q
2
q
+2
k=1
a
1
,...,a
k
1
(1 +a
1
)
q
(1 +a
k1
)
q
(2 +a
k
)
q
= 2
q
+2
_

n=1
(n +2)
q
_

k=0
_

n=1
(n +1)
q
_
k
= 2
q
+
2((q) 1 2
q
)
2 (q)
.
Accordingly, one can use numerical computations to get the lower bound

n
z
n
(2.25) > 1, proving that q
c
< 2.25.
(ii) Let q < 0; from (56) and the fact that (1 +a
2k+1
)
q
a
q
2k+1
, one gets
n=0
nz
n
(q) 2
n=0
n(2 +n)
q
+2
k=1
a
1
,...,a
2k
1
(a
1
+ +a
2k
)A(a
1
, . . . , a
2k
)
q
+ 2
k=0
a
1
,...,a
2k+1
1
(a
1
+ +a
2k+1
)A(a
1
, . . . , a
2k
)
q
a
q
2k+1
= 2
n=0
n(2 +n)
q
+2
k=1
a
1
,...,a
2k
1
2ka
1
A(a
1
, . . . , a
2k
)
q
+ 2
k=1
a
1
,...,a
2k+1
1
2ka
1
A(a
1
, . . . , a
2k
)
q
a
q
2k+1
+ 2
k=1
a
1
,...,a
2k+1
1
a
2k+1
A(a
1
, . . . , a
2k
)
q
a
q
2k+1
= 2
n=0
n(2 +n)
q
+4
_

n,m1
n(1 +nm)
q
_

k=1
k
_

n,m1
(1 +nm)
q
_
k1
+ 4
_

n,m1
n(1 +nm)
q
__

n=1
n
q
_

k=1
k
_

n,m1
(1 +nm)
q
_
k1
+ 2
_

n=1
n
q+1
_

k=1
_

n,m1
(1 +nm)
q
_
k
.
On the one hand, for any q < 2 one has
n=0
n(2 +n)
q
< and
n,m1
n(1 +nm)
q
2
n=1
n(1 +n)
q
+
_

n=2
nn
q
__

n=2
n
q
_
< . (57)
Thus, there exist three positive constants C, C
and C
such that
n=0
nz
n
(q) C +C
k=1
k
_

n,m1
(1 +nm)
q
_
k1
+C
k=1
_

n,m1
(1 +nm)
q
_
k
. (58)
On the other hand, one has for any q < 1,
n,m1
(1 +nm)
q
=
n,m>1
(1 +nm)
q
+2
n=1
(1 +n)
q
2
q
n,m>1
(nm)
q
+2
n=2
n
q
2
q
=
_

n=2
n
q
_
2
+2
n=2
n
q
2
q
= ((q) 1)
2
+2((q) 1) 2
q
.
For
0
> 0 such that
2
0
+2
0
1/8 = 1, one can check that (2.25) 1 <
0
, so that
nm1
(1 + nm)
2.25
<
2
0
+2
0
1/8 = 1. (59)
Therefore, by (58) and (59), one concludes that
n
nz
n
(q) < when q < 2.25. 2
B. Appendix. Box and packing dimensions
Let M be a subset of S
1
. An -packing of M (0 < < 1) is a collection {J
i
}
i
of mutually
disjoint intervals J
i
F with |J
i
| and which intersect M; for any 0 1 we set
P
(M) := lim
0
sup
_
i
|J
i
|
; {J
i
}
i
_
,
where {J
i
}
i
runs over the -packing of M (P
is not sigma-additive). The box dimension

of M is by denition
dim
B
M := inf{; P
(M) = 0} = sup{; P
(M) = };
in general, dim
B
n
M
n
= sup
n
dim
B
M
n
, whenever M
1
, M
2
, . . . forma countable family
of subsets of S
1
. There are many other equivalent denitions of the box dimension of M,
one of them using a regular net of S
1
: suppose that F :=

n
F
n
is a net of S
1
, then, for
any 0 1, we set
P
(M|F) := lim
0
sup
_
i
|J
i
|
; {J
i
}
i
_
,
where {J
i
}
i
runs over the -packing of M by intervals in F. Recall that F is said to be
regular if there exists a constant c > 1 such that for any n one has |J|/|J
| c, for any
J F
n
, J
F
n+1
and J
J; if F is regular then
dim
B
M := inf{; P
(M|F) = 0} = sup{; P
(M|F) = }
and one recovers the usual denition of the box dimension by using the dyadic net for
example. The notion of packing dimension has been introduced by Tricot in [44]; the
packing dimension of M is
dim
P
M := inf
_
sup
n
dim
B
M
n
; M
_
n
M
n
_
,
with the important property that dim
P
n
M
n
= sup
n
dim
P
M
n
; moreover the packing
dimension is a useful notion to get an upper bound of the Hausdorff dimension since
dim
H
M dim
P
M dim
B
M.
A systematic approach and detailed proof about fractal dimensions can be found in
[11, 28]; we also refer to [35] for a point of view related to dynamical systems.
REFERENCES
[1] P. Billingsley. Ergodic Theory and Information. John Wiley, New York, 1965.
[2] R. Bowen. Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms (Lecture Notes in
Mathematics, 470). Springer, Berlin, 1975.
[3] M. Bramson and S. A. Kalikow. Nonuniqueness in g-functions. Israel J. Math. 84 (1993), 153160.
[4] G. Brown, G. Michon and J. Peyri` ere. On the multifractal analysis of measures. J. Stat. Phys. 6 (1992),
775790.
[5] H. Cajar. Billingsley Dimension in Probability Spaces (Lecture Notes in Mathematics, 892). Springer,
Berlin, 1981.
[6] R. Cawley and R. D. Mauldin. Multifractal decomposition of Moran fractals. Adv. Math. 92 (1992), 196
236.
[7] P. Collet, J. L. Lebowitz and A. Porzio. The dimension spectrum of some dynamical systems. J. Stat.
Phys. 47 (1987), 609644.
[8] J. M. Dumont, N. Sidorov and A. Thomas. Number of representations related to a linear recurrent basis.
Acta Arith. 88 (1999), 391396.
[9] F. J. Dyson. Existence of phase-transition in a one-dimensional Ising ferromagnet. Comm. Math. Phys.
12 (1969), 91107.
[10] P. Erd os. On a family of symmetric Bernoulli convolutions. Amer. J. Math. 61 (1939), 974976.
[11] K. Falconer. Fractal GeometryMathematical Foundations and Applications. John Wiley, Chichester,
1990.
[12] A.-H Fan, K.-S Lau and S.-M Ngai. Iterated function systems with overlaps. Asian J. Math. 4 (2000),
527552.
[13] D.-J. Feng. The limit Rademacher functions and Bernoulli convolutions associated with Pisot numbers I,
II. Preprint, 1999.
[14] D.-J. Feng. Smoothness of L
q
-spectrum of self-similar measures with overlaps. J. Lond. Math. Soc. 68
(2003), 102118.
[15] M. Fisher. The theory of condensation and the critical point. Physics 3 (1967), 255283.
[16] T. Halsey, M. Jensen, L. Kadanoff, I. Procaccia and B. Shraiman. Fractal measures and their singularities:
the characterization of strange sets. Phys. Rev. A 33 (1986), 11411151.
[17] Y. Heurteaux. Estimation de la dimension inf erieure et de la dimension sup erieure des mesures. Ann. Inst.
Henri Poincar e 34 (1998), 309338.
[18] F. Hofbauer. Examples for the nonuniqueness of the equilibrium state. Trans. Amer. Math. Soc. 228
(1977), 223241.
[19] F. Hofbauer. Local dimension for piecewise monotonic maps on the interval. Ergod. Th. & Dynam. Sys.
15 (1995), 11191142.
[20] T.-Y. Hu and K.-S. Lau. Multifractal structure of convolution of the cantor measure. Adv. App. Math. 22
(2000), 116.
[21] B. Jessen and A. Wintner. Distribution functions and the Riemann zeta function. Trans. Amer. Math. Soc.
38 (1935), 4888.
[22] M. Keane. Strongly mixing g-measures. Invent. Math. 16 (1972), 309324.
[23] A. Khintchin. Continued Fractions. University of Chicago Press, Chicago, 1964.
[24] K.-S. Lau and S.-M. Ngai. L
q
-spectrum of the Bernoulli convolution associated with the golden ratio.
Stud. Math. 131 (1998), 1729.
[25] K.-S. Lau and S.-M. Ngai. Multifractal measures and a weak separation condition. Adv. Math. 141 (1999),
4596.
[26] F. Ledrappier. Principe variationnel et syst` emes dynamiques symboliques. Z. Wahr. Verw. Geb. 30 (1974),
185202.
[27] F. Ledrappier and A. Porzio. A dimension formula for Bernoulli convolutions. J. Stat. Phys. 76 (1994),
225251.
[28] P. Mattila. Geometry of Sets and Measures in Euclidean Spaces. Cambridge University Press, Cambridge,
1995.
[29] E. Olivier. Dimension de Billingsley des ensembles satur es. C. R. Acad. Sci. Paris 328 (1998), 1316.
[30] E. Olivier. Multifractal analysis in symbolic dynamic and distribution of pointwise dimension for
g-measures. Nonlinearity 12 (1999), 15711585.
[31] E. Olivier. Structure multifractale dune dynamique non expansive d enie sur un ensemble de Cantor.
C. R. Acad. Sci. Paris 331 (2000), 605610.
[32] E. Olivier, N. Sidorov and A. Thomas. On the Gibbs properties of some Bernoulli convolutions related to
-numeration. Preprint, 2002.
[33] L. Olsen. A multifractal formalism. Adv. in Math. 116 (1995), 82196.
[34] L. Olsen. Multifractal geometry. Progr. Probability 46 (1998), 338.
[35] Y. Pesin. Dimension Theory in Dynamical Systems. University of Chicago Press, Chicago, 1997.
[36] Y. Pesin and H. Weiss. The multifractal analysis of Gibbs measures: motivation, mathematical foundation,
and examples. Chaos 7 (1997), 89106.
[37] J. Peyri` ere. Calcul de dimensions de Hausdorff. Duke Math. J. 44 (1976), 591601.
[38] H. Rao and Z. Y. Wen. A class of self-similar fractals with overlap structures. Adv. Appl. Math. 20 (1998),
5072.
[39] D. A. Rand. The singularity spectrum f () for cookie-cutters. Ergod. Th. & Dynam. Sys. 9 (1989), 527
541.
[40] D. Ruelle. Statistical Mechanics. Rigorous Results. Benjamin, New York, 1969.
[41] O. Sarig. Thermodynamic formalism for countable Markov shifts. Ergod. Th. & Dynam. Sys. 19 (1999),
15651594.
[42] N. Sidorov and A. Vershik. Ergodic properties of the Erd os measure, the entropy of the golden shift and
related problems. Monatshefte Math. 126 (1998), 215261.
[43] R. S. Strichartz, A. Taylor and T. Zhang. Density of self-similar measures. Exp. Math. 4 (1995), 101128.
[44] C. Tricot. Two denitions of fractional dimension. Math. Proc. Cambr. Philos. Soc. 91 (1982), 5774.
[45] P. Walters. An Introduction to Ergodic Theory. Springer, Berlin, 1982.
[46] M. Yuri. Zeta functions for certain non-hyperbolic systems and topological Markov approximations.
Ergod. Th. & Dynam. Sys. 17 (1997), 9971000.

Multifractal Analysis of Weak Gibbs Measure

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Multifractal Analysis of Weak Gibbs Measure

Încărcat de

Drepturi de autor:

Formate disponibile

Ergod. Th. & Dynam. Sys.

(2003), 23, 17511784 c 2003 Cambridge University Press

(x) exists and is equal to . The map dim

(|M) and we give (Theorem A) an intermediate

is implicitly dened by a pressure equation. An important

coincides with the L

) that the Legendre

) for any q R (Lemma 3.2) and on

; the proofs of these theorems are given in 3.

is called [42] the

(), one has h

-conformal. When the potential is normalized one usually writes

so that the T -invariant e

-conformal measures are exactly the g-measures.

is a convex map on R R; for any q, q

> q, one has

(q, t ) decreases from + to . Using the assumption

(q, t ) increases from to +. Therefore, there exists an increasing concave

: R R dened by the implicit equation P

(q)) = 0. The underlying

leads to a characterization of the

is not differentiable (i.e. the rst-order phase transition points): actually,

about + and : more precisely we shall prove that

(), for any < < .

(q) = (q), for any q R.

. Let T be a d-f.c.t. and be a weak Gibbs measure of a negative potential

(), for any < < .

is the set of words on A, that is, A

and in what follows we

[w] 3 and thus it remains to prove

[w] 3/(n +2). Assume that w = 0

[w] 3/(n +2). 2

in Theorem 2.2 below (as well as for the

and the continued

, it follows from Lemma 2.3 that

; moreover, by Theorem C, its L

, a solution of the implicit equation P

(q): this means that, for any q R,

(q, t ) 0}. (21)

is also a weak Gibbs measure of , one can write

can be dened as the probability distribution of the random variable

satises a self-similarity equation, say

is the Erd os measure. The algebraic equation

and, by Theorem C, its

, a solution of the implicit equation

(q): this means that, for any q R,

(q, t ) 0}. (30)

is also a weak Gibbs measure of , it is easily seen that

() the set of -generic points of

compatible with the product topology on

()) = 1. However, (32) is still

the concave map such that

(q)) = 0, the following two propositions hold:

is not differentiable on an (at most) countable subset of R, there exists a sequence

it is clear that lim

(q)) = 0, for any q > 0. Since

such that ()/(

) is a non-empty closed interval of R, for I

is a non-empty closed convex

(not necessarily ergodic) such that () = ()/(

(given by Lemma 3.3), satises

(), one has lim

()), one deduces that

is well dened. We use the notions of -packing, box dimension (denoted