1411 2312

HAUSDORFF SPECTRUM OF HARMONIC MEASURE
RYOKICHI TANAKA
arXiv:1411.2312v2 [math.PR] 7 May 2015
Abstract. For every non-elementary hyperbolic group, we show that for every
random walk with finitely supported admissible step distribution, the associated
entropy equals the drift times the logarithmic volume growth if and only if the
corresponding harmonic measure is comparable with Hausdorfff measure on the
boundary. Moreover, we introduce one parameter family of probability measures
which interpolates a Patterson-Sullivan measure and the harmonic measure, and
establish a formula of Hausdorff spectrum (multifractal spectrum) of the harmonic
measure. We also give some finitary versions of dimensional properties of the
harmonic measure.
1. Introduction
Let Γ be a finitely generated group. For a probability measure µ on it, we obtain

a random walk on Γ by multiplying from right independent random elements with
the law µ, and the distribution of the random walk at the time n is given by the
n-th convolution power µ∗n . There are several important quantities which capture
the asymptotic behaviours of the random walks. Define the entropy h and the drift
l (also called the rate of escape, or the speed) by
1 X ∗n 1X
h := lim − µ (x) log µ∗n (x), l := lim |x|µ∗n (x),
n→∞ n x∈Γ n→∞ n
x∈Γ
where |·| denotes the word norm associated with a finite symmetric set of generators
of Γ. It is known that the entropy, introduced by Avez [Ave], is equal to 0 if and only
if all bounded µ-harmonic functions on Γ are constants ([Der], [KV]). The entropy
and the drift are connected via the logarithmic volume growth v of the group which
is defined by ev := limn→∞ |Bn |1/n , where |Bn | denotes the cardinality of the set Bn
of words of length at most n, by the inequality
h ≤ lv, (1)
as soon as all those quantities are well-defined ([Gui], see also e.g., [BHM1] and
[Ver]). The inequality (1) is also called the fundamental inequality. In [Ver], Vershik
proposed to study the equality case of (1). In this paper, we focus on hyperbolic
groups in the sense of Gromov and characterise the equality of (1) in terms of the
boundary behaviours of the random walks. For every hyperbolic group Γ, one can
define the geometric boundary ∂Γ, which is compact and admits a metric dε with
a parameter ε > 0. The harmonic measure ν is defined by the hitting distribution
of the random walk starting from the identity on the boundary ∂Γ, corresponding
to the step distribution µ on Γ. The boundary ∂Γ has the Hausdorff dimension
Date: July 9, 2018. Supported by JSPS Grant-in-Aid for Young Scientists (B) 26800029.
1
2 RYOKICHI TANAKA
D = v/ε and the D-Hausdorff measure HD is finite and positive on ∂Γ [Coo]. Here
the D-Hausdorff measure HD is a natural measure to compare with the harmonic
measure ν. We call a probability measure µ on the group Γ admissible if the support
of µ generates the whole group Γ as a semigroup. In the present paper, µ is always
finitely supported and admissible unless stated otherwise.
Theorem 1.1. For every finitely supported admissible probability measure µ on
every finitely generated non-elementary hyperbolic group Γ equipped with a word
metric, it holds that h = lv if and only if the corresponding harmonic measure ν and
the D-Hausdorff measure HD are mutually absolutely continuous and their densities
are uniformly bounded from above and from below.
In Section 5, we prove this result in Theorem 5.8. See Remark 5.10, for another
equivalent condition to the equality case. Blachère, Haissinsky and Mathieu estab-
lished this result for every finitely supported admissible and symmetric probability
measure µ [BHM2, Corollary 1.2, Theorem 1.5]. We extend it to non-symmetric mea-
sures from a completely different approach as we describe later. Recently, Gouëzel,
Mathéus and Maucourant have proven that for a non-elementary hyperbolic group
Γ which is not virtually free and equipped with a word metric, for every finitely sup-
ported admissible probability measure µ, the equality h = lv never holds [GMM2].
Together with their results, one concludes that in this setting, the harmonic mea-
sure ν and the D-Hausdorff measure HD are always mutually singular. Connell and
Muchnik proved that for an infinitely supported probability measure µ on Γ, the
D-Hausdorff measure HD (and also a Patterson-Sullivan measure) and the harmonic
measure for µ can be equivalent ([CM1, Remark 0.5] and [CM2]). On the other hand,
Le Prince showed that for every finitely generated non-elementary hyperbolic group
Γ, there exists a finitely supported admissible and symmetric probability measure µ
such that the corresponding harmonic measure ν and the D-Hausdorff measure HD
are mutually singular [LeP]. Ledrappier proved the corresponding result to Theo-
rem 1.1 for non-cyclic free groups for every finitely supported admissible probability
measure µ [Led, Corollary 3.15] . For free groups, it is straightforward to see that
if µ depends only on the word length associated with the standard symmetric gen-
erating set, then the corresponding harmonic measure coincides with the Hausdorff
measure (the uniform measure on the boundary) up to a multiplicative constant.
Let us also mention some related works. In [GMM1], Gouëzel et al. studied a vari-
ant of the fundamental inequality (1) and also obtained some rigidity results for
the equality case. Apart from Cayley graphs of groups, Lyons extensively studied
the equivalence of the harmonic measure and the Patterson-Sullivan measure for
universal covering trees of finite graphs [Lyo].
A novel feature of our approach is to introduce one parameter family of probability
measures µθ , which interpolates a Patterson-Sullivan measure and the harmonic
measure on the boundary ∂Γ. Let us consider for every θ ∈ R,
1 X
β(θ) := lim log G(1, x)θ ,
n→∞ n
x∈S
n
where G(x, y) is the Green function associated with µ, we denote by 1 the identity
of the group, and by Sn the set of words of length n. The limit exists by the Ancona
HAUSDORFF SPECTRUM OF HARMONIC MEASURE 3
inequality (Lemma 3.1), and we show that β is convex, in fact, analytic except for at
most finitely many points and continuously differentiable at every point (Proposition
4.8 and Corollary 4.11). Theorem 1.1 is deduced via the following dimensional
properties of the harmonic measure ν.
Theorem 1.2. Let Γ and µ be as in Theorem 1.1, and ν be the corresponding
harmonic measure on the boundary ∂Γ. It holds that
log ν (B(ξ, r)) h
lim = , ν-a.e. ξ. (2)
r→0 log r εl
Define the set
log ν (B(ξ, r))
Eα = ξ ∈ ∂Γ lim =α ,
r→0 log r
then the Hausdorff dimension of the set Eα is given by the Legendre transform of β,
i.e.,
αθ + β(θ)
dimH Eα = ,
ε
for every α = −β ′ (θ), where β is continuously differentiable on the whole R, and
B(ξ, r) denotes the ball of radius r centred at ξ.
In Section 5, we prove this result in Theorem 5.2 and Theorem 5.4 in a more
detailed form. The measure µθ is constructed by the Patterson-Sullivan technique
(Theorem 3.3). As θ = 0, the measure µ0 is a Patterson-Sullivan measure, and
thus comparable with the D-Hausdorff measure HD , and as θ = 1, the measure
µ1 is the harmonic measure ν (Corollary 3.4). The probability measure µθ satisfies
that µθ (Eα ) = 1 for α = −β ′ (θ). We call dimH Eα as a function in α the Haus-
dorff spectrum of the measure ν. To determine the Hausdorff spectrum is called
multifractal analysis which has been extensively studied in fractal geometry. In
fact, there are many technical similarities to analyse harmonic measures and self-
conformal measures ([Fen] and [PU]). For the backgrounds on this topic, see also
[Fal] and references therein. We show that the measure µθ satisfies a Gibbs-like
property with respect to β(θ), where β(θ) is an analogue of the pressure (Corollary
3.5). This measure µθ is also characterised by the eigenmeasures of certain trans-
fer operator built on a symbolic dynamical system associated with an automatic
structure of the group (Theorem 4.9). To study the measure µθ , we employ the
results about the Martin boundary of a hyperbolic group by Izumi, Neshveyev and
Okayasu [INO] and a generalised thermodynamic formalism due to Gouëzel [Gou1].
Note that the formula (2) is proved for every non-elementary hyperbolic group and
for every finitely supported symmetric probability measure µ in [BHM2, Theorem
1.3], for every non-cyclic free groups and for every probability measure µ of finite
first moment in [Led, Theorem 4.15], for a general class of random walks on trees in
[Kai1] and for the simple random walks on the Galton-Watson trees in [LPP].
The following result is a finitary version of Theorem 1.2, inspired by the corre-
sponding results for the Galton-Watson trees by Lyons, Pemantle and Peres [LPP].
Theorem 1.3. Let Γ and µ be as in Theorem 1.1, and consider the associated
random walk starting at the identity on Γ. For every a ∈ (0, 1), there exists a subset
4 RYOKICHI TANAKA
Γa ⊂ Γ such that the random walk stays in Γa for every time with probability at least
1 − a, and
lim |Γa ∩ Sn |1/n = eh/l .
n→∞
In Section 6, we prove this result in Theorem 6.2. In particular, if h < lv, then the
random walk is confined in an exponentially small part of the group with positive
probability. This can be compared with [Ver, p.669], where a random generation of
group elements which is called the Monte Carlo method is discussed. For example,
the random generation of group elements according to a random walk does not
produce the whole data of the group in this case; see also [GMM2].
Let us return to Theorem 1.1. For a symmetric probability measure µ, i.e., µ(x) =
µ(x−1 ) for every x ∈ Γ, one can define the Green metric dG (x, y) = − log F (x, y),
where F (x, y) denotes the probability that the random walk starting at x ever
reaches y, and show that Γ is hyperbolic with respect to dG according to the Ancona
inequality [BHM2]. The Green metric dG is not geodesic; nevertheless, one can use
approximate trees argument and most of common techniques for the geodesic case
work. The harmonic measure ν is actually a quasi-conformal (in fact, conformal)
measure with respect to the metric induced in the boundary ∂Γ by the Green met-
ric dG , and this fact plays an essential role to deduce Theorem 1.1 and that the
local dimension of the harmonic measure ν is h/(εl) for ν-a.e. in Theorem 1.2 in
the symmetric case. The symmetry of µ is required to make the Green metric dG a
genuine metric in Γ; otherwise it is not clear that one can discuss about the hyper-
bolicity for a non-symmetric metric. Here our alternative is to introduce a measure
µθ , whose construction is actually inspired by a recent work of Gouëzel on the local
limit theorem on a hyperbolic group [Gou1], and to obtain the Hausdorff spectrum
of the harmonic measure ν. In many cases, a description of the Hausdorff spectrum
is a main purpose on its own right, especially, when it is motivated by a problem in
statistical physics. A fundamental observation in this paper, however, is rather the
converse; namely, we use the multifractal analysis to compare the harmonic measure
with a natural reference measure which is the D-Hausdorff measure on the boundary
of the group Γ. More precisely, we shall see that the function β is affine on R if
and only if those two measures are mutually absolutely continuous (Theorem 5.8).
Furthermore, the description of the Hausdorff spectrum implies that the harmonic
measure has a rich multifractal structure as soon as it is singular with respect to
the D-Hasudorff measure (Theorem 5.4). In particular, the range of the Hausdorff
spectrum contains the interval [h/(εl), v/ε].
Let us mention about an extension to a step distribution µ of unbounded support.
The arguments in the present paper work once we have the Ancona inequality and
its strengthened one (Theorem 2.5). Gouëzel has proven those for every admissible
probability measure µ with a super-exponential tail [Gou2]. At the same time, he
has also proven a failure of description of the Martin boundary in the usual sense
for an admissible probability measure with an exponential tail [ibid]. Hence the
results in this paper are extended to every admissible step distribution µ with a
super-exponential tail, but it is obscure whether one could extend to µ with an
exponential tail in the present approach. We shall also mention about an extension
to a left invariant metric which is not induced by a word length in Γ. For example,
one is interested in the setting where Γ acts cocompactly on the hyperbolic space Hn
and the metric in Γ is defined by d(x, y) := dHn (xo, yo) for a reference point o in Hn .
In fact, in [BHM2], they proved Theorem 1.1 for symmetric µ with every metric
d which is hyperbolic and quasi-isometric to a word metric in Γ (not necessarily
geodesic). Some of our results still hold for such a metric d, and in fact, most of the
results are expected to remain valid; but we do not proceed to this direction in the
present paper for the simplification of the proofs.
Finally, we close this introduction by pointing out some related problems in a
continuous setting. On the special linear groups, the regularity problem of the har-
monic measures is proposed by Kaimanovich and Le Prince [KL], and they showed
that there exists a finitely supported symmetric probability measure (the support
can generate a given Zariski dense subgroup) on SL(d, R) (d ≥ 2) such that the
corresponding harmonic measure is non-atomic singular with respect to a natural
smooth measure class on the Furstenberg boundary. They proved this result via
the dimension inequality of the harmonic measure. Bourgain constructed a finitely
supported symmetric probability measure on SL(2, R) such that the corresponding
harmonic measure is absolutely continuous with respect to Lebesgue measure on
the circle [Bou]. Brieussel and the author proved analogous results in the three
dimensional solvable Lie group Sol [BT], namely, they showed that random walks
with finitely supported step distributions on it can produce both absolutely con-
tinuous and singular harmonic measures with respect to Lebesgue measure on the
corresponding boundary. The dimension inequality is also used there to prove the
existence of a finitely supported probability measure whose harmonic measure is sin-
gular. We point out, however, the dimension equality of type (2) is still missing in
both cases, and in general, in the Lie group settings to the extent of our knowledge.
The organisation of this paper is the following: In Section 2, we recall some known
results about hyperbolic groups, random walks on it, and the Martin boundary. In
Section 3, we construct the measure µθ , and show several properties of µθ including
the Gibbs property (Corollary 3.5) and the ergodicity (Corollary 3.7). In Section
4 we review a generalised thermodynamic formalism associated with an automatic
structure of the group due to Gouëzel [Gou1] and prove the regularity of β. In
Section 5, we prove Theorem 1.2 in Theorem 5.2 and in Theorem 5.4, and then we
deduce Theorem 1.1 in Theorem 5.8. In Section 6, we show finitary results, following
the Galton-Watson trees cases [LPP], and prove Theorem 1.3 in Theorem 6.2.
Notation 1.4. Throughout this article, we use C, C ′ , . . . , to denote absolute con-

stants whose exact values may change from line to line, and also use them with
subscripts, for instance, Cδ to specify its dependency only on δ. For functions f and
g, we write f ≍ g if there exist some constants c1 , c2 > 0 such that c1 g ≤ f ≤ c2 g,
and also write f ≍δ g if the implied constants depend only on δ. Furthermore, for
measures ν1 and ν2 , we say that those are comparable if ν1 and ν2 are mutually ab-
solutely continuous and their densities are bounded from above and from below, and
also write ν1 ≍ ν2 . We denote by f (n) = O(g(n)) if there exists a constant C > 0
independent of n such that f (n) ≤ Cg(n), by f (n) = o(g(n)) if f (n)/g(n) → 0 as
n → ∞, and by f (n) ∼ g(n) if f (n)/g(n) → 1 as n → ∞.
6 RYOKICHI TANAKA
2. Preliminary
2.1. Hyperbolic groups. Let Γ be a finitely generated group with a finite sym-
metric set S of generators. We denote by |x| the word length of x in Γ associated
with S, and by d(x, y) = |x−1 y| the corresponding word metric which is left invariant
under the action of Γ. (We do not specify S.) The Gromov product is defined by
d(x, z) + d(y, z) − d(x, y)
(x|y)z = ,
2
for x, y, and z in Γ. We say that the metric space (Γ, d) is δ-hyperbolic, if there
exists a constant δ ≥ 0 such that for all x, y, z and w in Γ, it holds that
(x|y)w ≥ min ((x|z)w , (y|z)w ) − δ.
If (Γ, d) is δ-hyperbolic, then every geodesic triangle is 4δ-slim, i.e., each side of the
triangle is within the 4δ-neighbourhood of the union of the other two sides. The
notion of the hyperbolicity, which means that it is δ-hyperbolic for some δ ≥ 0, is
invariant under quasi-isometry between two geodesic metric spaces, hence it does
not depend on the choice of the finite symmetric set of generators. More details
about the hyperbolic groups, see [GdlH].
The boundary ∂Γ of Γ is defined in the following way: Fix a base point the
identity 1 of Γ and write (x|y) for (x|y)1. We say that a sequence {xn }∞n=0 converges
to infinity if (xn |xm ) → ∞ as n and m tend to infinity, and two sequences {xn }∞ n=0
and {yn }∞ n=0 converging to infinity are equivalent if (xn |ym ) → ∞ as n and m tend
to infinity. We define the geometric boundary ∂Γ as the set of sequences converging
to infinity modulo the equivalence relation. Since (Γ, d) is a proper geodesic space
(i.e., all closed balls are compact), the boundary ∂Γ is also identified with the set
of geodesic rays modulo bounded distances [GdlH, Chapitre 7, Proposition 4]. The
Gromov product is extended to Γ ∪ ∂Γ by setting

(ξ|η) = sup lim inf (xn |ym) {xn } ∈ ξ, {yn } ∈ η ,
n,m→∞
where the supremum is taken over all sequences {xn }∞ ∞

n=0 converging to ξ and {yn }n=0
converging to η. For the extended Gromov product, it holds that
lim inf (xn |ym ) ≤ (ξ|η) ≤ lim inf (xn |ym) + 2δ,
n,m→∞ n,m→∞
for every such sequence by the δ-hyperbolicity. The space Γ ∪ ∂Γ is compact,

equipped with the topology by bases of balls in Γ and sets of the form Ur (x) =
{ξ ∈ Γ ∪ ∂Γ (ξ|x) > r}. In fact, it admits a compatible metric dε with a parameter
ε > 0 small enough such that
C −1 e−ε(ξ|η) ≤ dε (ξ, η) ≤ Ce−ε(ξ|η) ,
for some constant C > 0. We fix a parameter ε once and for all so that all the
quantities associated with the metric in the boundary ∂Γ such as the Hausdorff
dimension are defined with respect to the metric dε .
Recall that the shadow S(x, R) for x in Γ and R ≥ 0 is the set of points ξ in
the boundary ∂Γ such that every geodesic ray issuing from 1 converging to ξ passes
through the ball of radius R centred at x. The shadows are compared with balls.
Lemma 2.1. For every R ≥ 4δ and for every ξ in ∂Γ, it holds that for every point
x on every geodesic ray starting from 1 converging to ξ,
B(ξ, C −1eεR e−ε|x| ) ⊂ S(x, R) ⊂ B(ξ, CeεR e−ε|x| ).
One is able to show the lemma by the δ-hyperbolicity (e.g. [Cal]). The shadows
are used to describe measures on the boundary ∂Γ.
2.2. Random walks on groups. For a countable group Γ, consider a probability

measure µ on it. Here we always assume that the measure µ is admissible, i.e.,
the support of µ generates the whole group Γ as a semigroup. The random walk
{xn }∞
n=0 on Γ starting at the identity 1 is defined by xn := X1 · · · Xn and x0 = 1,
where X1 , X2 , . . . , is a sequence of independent random elements with the common
distribution µ. We denote by Ω = ΓZ+ the space of paths and by P the distribution
of the random walk on Ω given by the push-forward measure of the product measure
µZ+ induced by the map {Xn }∞ ∞
n=1 7→ {xn }n=0 . The distribution of xn is given by the
n-th convolution power of µ, namely, P (xn = x) = µ∗n (x). Let us define the Green
function by
X∞
G(x, y) = pn (x, y),
n=0
−1
where pn (x, y) = P(xn = x y). Recall that the Green function has finite values
if and only if the random walk is transient, and it is the case as soon as Γ is a
non-amenable group. Henceforth, we assume that Γ is a non-elementary hyperbolic
group, i.e., non-amenable, and µ is a finitely supported admissible probability mea-
sure on Γ. In general, we do not assume that µ is symmetric, i.e., µ(x) = µ(x−1 ).
The following is a special case of the theorem by Kaimanovich [Kai2, Theorem 7.4].
We give a sketch of the proof for the sake of convenience.
Theorem 2.2 (Kaimanovich). The random walk {xn }∞ n=0 converges to a point in
the boundary ∂Γ almost surely in P. Moreover, there exists a positive constant l > 0
such that
d(1, xn ) = ln + o(n),
P-almost surely, and for P-almost every sample path, there exists a unit speed geo-
desic ray γω emanating from the identity 1 such that
d(xn , γω (⌊ln⌋)) = o(n).
Here ⌊a⌋ denotes the integer part of a. Moreover, one can take the map from Ω to
the space of geodesic rays endowed with the pointwise convergence topology: ω 7→ γω
as a Borel measurable map.
A sketch of the proof. For an admissible probability measure µ, the drift l is positive
since the Poisson boundary of (Γ, µ) is non-trivial. The Kingman subadditive ergodic
theorem implies that d(1, xn ) = ln+o(n) for P-a.s. Together with d(xn , xn+1 ) = o(n)
(in fact, bounded in our case), the Gromov product grows lineally: (xn |xn+1 ) =
ln + o(n), a.s. in P. It follows that the sequence {xn }∞n=0 is Cauchy with respect to
−ε(ξ|η)
the metric dε (ξ, η) ≤ Ce in Γ ∪ ∂Γ, and thus it converges to a point ξ in the
boundary. For a unit speed geodesic ray γω converging to ξ from 1, one can deduce
that each xn is within the distance o(n) from γω (⌊ln⌋). The map ω 7→ γω is taken to
8 RYOKICHI TANAKA
be Borel measurable by choosing the lexicographically minimum geodesic ray. More

details, see Section 7 in [Kai2].
The harmonic measure is the hitting distribution of the random walk on the
boundary.
Definition 2.3. We call the distribution of the point to which the random walk
{xn }∞
n=0 converges, the harmonic measure ν corresponding to µ.
The harmonic measure ν is a unique stationary measure on the boundary ∂Γ,

i.e.,
X
ν= µ(g)g∗ν, (3)
g∈Γ
where x∗ ν is the push-forward of ν by the left multiplication of x, or equivalently
the hitting distribution of the random walk starting from x with the same step
distribution µ; see e.g. [Kai2, Theorem 2.4].
2.3. Martin boundary. The Green function is factorized as

G(x, y) = F (x, y)G(y, y),
where F (x, y) is the probability that the random walk starting at x ever hits
y, and G(y, y) = G(1, 1) by the left invariance by Γ of the transition probabil-
ity. Associated with the Green function G(x, y), the Martin kernel is defined by
K(x, y) := G(x, y)/G(1, y). The Martin compactification Γ of Γ is the smallest
compactification of Γ where Γ is equipped with the discrete topology, such that for
every x, the Martin kernel K(x, ·) extends continuously on Γ. The Martin boundary
is the part which is newly added to Γ, i.e., ∂M Γ = Γ \ Γ. For every finitely sup-
ported admissible µ on Γ, the Martin boundary ∂M Γ is identified with the geometric
boundary ∂Γ by Ancona [Anc] (see also [INO]). The Poisson boundary (∂Γ, ν) is
realised in the Martin boundary, and it holds that for ν-a.e. ξ,
dx∗ ν
(ξ) = K(x, ξ). (4)
dν
More details, see, e.g. [Woe, Chapter IV]. Let us recall the so-called Ancona inequal-
ity which plays an important role to identify the Martin boundary with ∂Γ.
Theorem 2.4 ([Anc], Theorem 3.3, and [INO], Proposition 2.1). There exists a
constant C ≥ 1 such that for every geodesic segment connecting x and y, and every
point z on it,
C −1 G(x, z)G(z, y) ≤ G(x, y) ≤ CG(x, z)G(z, y). (5)
For an extension to probability measures µ of super-exponential tails, see [Gou2].

Note that the first inequality is clear since F (x, y) ≥ F (x, z)F (z, y). The Ancona
inequality implies that the Green function G(x, y) is almost multiplicative along
geodesics. We also frequently use the Harnack inequality: there exists a constant
C ′ ≥ 1 such that for x, y, z, and z ′ in Γ,
′ ′
G(x, z) ≤ C ′d(z,z ) G(x, z ′ ) and G(z, y) ≤ C ′d(z,z ) G(z ′ , y). (6)
These follow since the Green function is a super-harmonic function. Combining (5)
with (6), we deduce that for every geodesic segment connecting x and y, and every
point z within the distance r from the geodesic segment,
G(x, y) ≤ CC ′2r G(x, z)G(z, y).
One of important consequences of (5) is that the Martin kernel K(x, ·) is Hölder
continuous on ∂Γ for each x [INO, Theorem 3.3]. In this paper, we use a bit stronger
form of this fact.
Theorem 2.5. There exist constant C > 0 and ρ < 1 such that if two geodesic
segments connecting x and y, and x′ and y ′ , respectively, have a common geodesic
segment of length n within each of those 4δ-neighbourhoods, then
G(x, y)/G(x′ , y)
− 1 ≤ Cρn .
G(x, y ′)/G(x′ , y ′)
The proof follows from the same argument as in [GL, Theorem 4.6]. See also
an extension to probability measures µ of super-exponential tails [Gou2]. We note
that the result is deduced once the Ancona inequality (in fact, a relative version of
it given in [INO]) holds. In [GL] and [Gou1, Theorem 2.9], they proved the above
estimate for a symmetric µ to deal with the r-Green function Gr (x, y). Here we
discuss only G(x, y) which is enough for our purpose. From Theorem 2.5, for every
x and x′ = 1, and every sequence {yn }∞ n=0 along a geodesic ray from 1 converging to
ξ, the Martin kernel G(x, yn )/G(1, yn ) converges to the limit K(x, ξ). Furthermore,
log K(x, ξ) is Hölder continuous in ξ with an exponent independent of x.
In [INO], Izumi, Neshveyev and Okayasu proved a Gibbs-like property of a har-
monic measure in terms of G(x, y) or F (x, y) in the following sense:
Theorem 2.6 ([INO], Theorem 4.1). Let R ≥ 4δ. It holds that for every x in Γ,
ν(S(x, R)) ≍R G(1, x),
where the implied constants are independent of x.
Remark 2.7. Actually, they use the following notation in [INO]: Denote by U(ξ, m)
the set of all η in ∂Γ such that for arbitrary two geodesic rays γ from 1 converging to ξ
and γ ′ from 1 converging to η, respectively, one has limn→∞ (γ(n)|γ ′ (n)) > m, where
the sequence is nondecreasing, hence the limit exists. They have shown that for
every ξ in ∂Γ and every geodesic ray from 1 towards ξ, and for all m ≥ 0, it holds
that ν(U(ξ, m)) ≍ F (1, γ(m)). The above statement is deduced by the relation:
U(ξ, n − R + 4δ) ⊂ S(γ(n), R) ⊂ U(ξ, n − R), which follows by the δ-hyperbolicity,
and by the Harnack inequality (6).
3. Measure µθ
First, we recall the Busemann function on (Γ, d). For each ξ in ∂Γ, take a geodesic
ray γ converging to ξ. Let
bγ (x, y) := lim (d(x, γ(n)) − d(γ(n), y)) ,
n→∞
10 RYOKICHI TANAKA
for x, y in Γ, where the limit exists since a sequence d(x, γ(n)) − n is decreasing by
the triangular inequality. The Busemann function of the point ξ is
bξ (x, y) = sup {bγ (x, y) | geodesic rays γ towards ξ} .
By the δ-hyperbolicity, it holds that
bγ (x, y) ≤ bξ (x, y) ≤ bγ (x, y) + 32δ, (7)
for every geodesic ray γ towards ξ, regardless of the starting point [GdlH, Chapitre
8].
The function β defined in the following lemma will play an important role in this
paper.
Lemma 3.1. For every θ ∈ R, the limit
1 X
β(θ) := lim log G(1, x)θ
n→∞ n
x∈S n
exists, where Sn is the set of words of length n in Γ. Moreover, the function β is

convex on R.
θ
P
Proof. Let Gθ (n) := x∈Sn G(1, x) . For every θ ∈ R, by the Ancona inequal-
ity (5), Gθ (n) is submultiplicative up to a bounded constant; hence the limit of
(1/n) log Gθ (n) exists. The convexity follows from G(1−t)θ0 +tθ1 (n) ≤ G(1−t)θ0 (n)Gtθ1 (n)
for 0 ≤ t ≤ 1 and θ0 , θ1 ∈ R.
The function β is convex on R, therefore it is continuous, and differentiable except

for at most countably many points. In fact, we will show that β is analytic except for
at most finitely many points and continuously differentiable at every point (Proposi-
tion 4.8 and Corollary 4.11). This is shown by identifying β(θ) with the pressure for
certain Ruelle-Perron-Frobenius transfer operator, based on an automatic structure
of the hyperbolic group Γ in Section 4.
Here we note that β has special values at θ = 0 and 1.
Lemma 3.2. It holds that β(0) = v, i.e., the logarithmic volume growth of (Γ, d) and
β(1) = 0. Furthermore, β(θ) → −∞ as θ → +∞, and β(θ) → +∞ as θ → −∞.
Proof. First, it follows that β(0) is just the logarithmic volume growth v > 0 by
definition. Next, we show that β(1) = 0. Note that the shadows S(x, R) (x ∈ Sn )
cover the boundary ∂Γ for some fixed R > 4δ with bounded overlaps, i.e., there is
N such that every ξ in the boundary is included in at most N shadows of the form
S(x, R) with |x| = n [Coo, Lemme 6.5]. By Theorem 2.6, we have that
X X
1 = ν(∂Γ) ≤ ν(S(x, R)) ≤ C G(1, x).
x∈Sn x∈Sn
On the other hand,

X X
G(1, x) ≤ C ν(S(x, R)) ≤ CNν (∪x∈Sn S(x, R)) ≤ CN.
x∈Sn x∈Sn
Therefore, we obtain β(1) = 0.

Finally, let us show that β(θ) → −∞ (resp.+∞) as θ → +∞ (resp.−∞). For the

random walk (Γ, µ), the spectral radius ρ for µ is strictly less than 1 and the n-th
step transition probability satisfies that pn (x, y) ≤ ρn . For a finite range random
walk, it takes at least d(x, y)/r steps to reach y from x for some constant r > 0.
Therefore, the Green function is estimated by G(x, y) ≤ Ce−λd(x,y) for constants
C > 0 and λ > 0. For every positive θ, one has β(θ) ≤ v − θλ, hence β(θ) goes
to −∞ as θ tends to +∞. On the other hand, for every negative θ, we have the
converse inequality: β(θ) ≥ v −θλ, and thus β(θ) goes to +∞ as θ tends to −∞.
We introduce one parameter family of measures µθ on the boundary ∂Γ; it in-

terpolates the harmonic measure and a Patterson-Sullivan measure. Recall that a
Patterson-Sullivan measure µ0 is given as a limit point of
−s|x|
P
x∈Γ e δx
µs := P −s|x|
x∈Γ e
as s ↓ v on the compactification Γ ∪ ∂Γ. The measure µ0 is supported on the

boundary ∂Γ, and satisfies
dg∗ µ0
(ξ) ≍ e−vbξ (g,1)
dµ0
for every g in Γ and µ0 -almost every ξ in ∂Γ. One can show that all limit points are
comparable with the D-Hausdorff measure; and thus they form a unique measure
class. This construction corresponds to the case where θ = 0 in the following.
Theorem 3.3. For every θ ∈ R, there exists a probability measure µθ on ∂Γ such

that
dg∗ µθ
(ξ) ≍ K(g, ξ)θ e−β(θ)bξ (g,1) (8)
dµθ
for every g in Γ and µθ -almost every ξ in ∂Γ.
Proof. The construction is analogous to the one of a Patterson-Sullivan measure on

∂Γ [Coo, Théorème 5.4]. Notice that β(θ) is the exponent of convergence for
X
G(1, x)θ e−s|x| ,
x∈Γ
i.e., it converges when s > β(θ) and diverges when s < β(θ). Define the measure on
Γ ∪ ∂Γ by
G(1, x)θ e−s|x| δx
P
µθ,s := Px∈Γ θ −s|x|
, (9)
x∈Γ G(1, x) e
where δx denotes the Dirac measure at x. First, suppose that x∈Γ G(1, x)θ e−s|x|
P
diverges as s ↓ β(θ). In this case, since µθ,s is supported on the compact space
Γ ∪ ∂Γ, as s ↓ β(θ), there is a subsequencial weak limit which we denote by µθ . The
limiting measure µθ is supported on ∂Γ as the denominator of (9) diverges. For the
12 RYOKICHI TANAKA
left multiplication by g, the measure µθ,s has the form:

−1
G(1, g −1x)θ e−s|g x|
g∗ µθ,s (x) = P θ −s|x|
x∈Γ G(1, x) e
G(1, g −1x)θ −s(|g−1 x|−|x|)
= e µθ,s (x).
G(1, x)θ
Here G(g, x)/G(1, x) goes to K(g, ξ) as x tends to ξ, and |g −1x| − |x| equals to
bξ (g, 1) up to a constant depending only on δ for every x in a neighbourhood of ξ in
Γ ∪ ∂Γ [Coo, Lemme 2.2]. Then the claim follows.
Next, we consider the case where x∈Γ G(1, x)θ e−s|x| does not diverge as s ↓ β(θ).
P
In this case, one can modify the series a little bit to make it diverge at the exponent
of convergence [Pat, Lemma 3.1]. There exists a positive increasing function k on
[0, ∞] such that
P for any ε > 0 there is b0 satisfying that k(ab) ≤ aε k(b) for b > b0 and
a > 1, and x∈Γ k(e|x| )G(1, x)θ e−s|x| diverges as s ↓ β(θ). We define the measure
µkθ,s with the modification by k as in (9). The rest follows in a similar way to the
above.
Corollary 3.4. For θ = 0, a measure µ0 is comparable with the D-Hausdorff mea-
sure HD on ∂Γ, where D = v/ε, i.e., µ0 ≍ HD . For θ = 1, a measure µ1 coincides
with the harmonic measure ν, i.e., µ1 = ν.
Proof. When θ = 0, a measure µ0 is nothing but a Patterson-Sullivan measure as we

see in the construction, and thus µ0 is comparable with the D-Hausdorff measure
([Coo, Corollaire 7.5]; see also [Cal, Corollary 2.5.10]). When θ = 1, in the proof of
Theorem 3.3, in fact we have
dg∗ µ1
(ξ) = K(g, ξ),
dµ1
since a parameter s converges from above to β(1) which is equal to 0 (Lemma
3.2). Recall that the harmonic measure ν is the unique stationary measure on the
boundary ∂Γ satisfying (3). By taking the PRadon-Nikodym derivative of g∗ ν with
respect to ν, noting (4), we have 1 = g∈Γ µ(g)K(g, ξ), for ν-a.e. ξ. By the
continuity of the Martin kernel K(g, ξ) in ξ, we deduce that a probability measure
µ1 is µ-stationary on ∂Γ, hence it has to coincide with the harmonic measure ν.
Corollary 3.5 (A Gibbs property of µθ ). For every θ ∈ R, let µθ be a probability
measure satisfying (8) in Theorem 3.3: Then there exists R0 such that for every
R ≥ R0 and for every g in Γ,
µθ (S(g, R)) ≍θ,R G(1, g)θ e−β(θ)|g| .
Proof. For every shadow S(g, R), and for every ξ ∈ S(g, R), take an arbitrary geo-
desic ray γ from 1 converging to ξ. By the Ancona inequality (5), for |g| < m, we
have
G(1, γ(m)) ≍ G(1, γ(|g|))G(γ(|g|), γ(m)).
Since d(g, γ(|g|)) ≤ 2R, by the Harnack inequality (6), it follows that G(1, γ(m)) ≍R
G(1, g)G(g, γ(m)). Hence by letting γ(m) → ξ, we have G(1, g)K(g, ξ) ≍R 1. On
the other hand, one obtains bξ (g, 1) = −|g| + O(R, δ) for all ξ ∈ S(g, R) by the
δ-hyperbolicity and by (7). Since µθ satisfies (8) in Theorem 3.3,
dg∗ µθ
(ξ) ≍θ,R G(1, g)−θ eβ(θ)|g| ,
dµθ
for all ξ ∈ S(g, R). This estimate yields
g∗ µθ (S(g, R)) ≍R G(1, g)−θ eβ(θ)|g| µθ (S(g, R)) . (10)
The rest is proceeded as in the proof of [Coo, Proposition 6.1]. Let 0 ≤ m0 < 1 be
the largest µθ mass of the point in ∂Γ, and fix m such that m0 < m < 1. By the
compactness of ∂Γ, there exists r0 > 0 small enough such that every ball of radius
r0 has the measure µθ less than m. Then there exists R0 depending on r0 such that
for every R ≥ R0 and for every g, the diameter of ∂Γ \ g −1 S(g, R) is less than r0
by [Coo, Lemme 6.3]. Since 0 < 1 − m ≤ µθ (g −1 S(g, R)), by (10), it holds that
µθ (S(g, R)) ≍R G(1, g)θ e−β(θ)|g| . Thus the corollary follows.
The following is suggested by Gouëzel [Gou3]. In particular, probability measures

satisfying (8) form a unique measure class.
Theorem 3.6. A probability measure satisfying (8) in Theorem 3.3 for some θ
is doubling. Arbitrary two probability measures satisfying (8) for the same θ are
mutually absolutely continuous and their densities are uniformly bounded from above
and from below.
Proof. Let µθ be a probability measure satisfying (8). We note that µθ is doubling

on shadows in the sense that: µθ (S(g, 2R)) ≍θ,R µθ (S(g, R)) for all g in Γ and for a
fixed R ≥ R0 by Corollary 3.5. By using this estimate iteratively (if necessary) and
by Lemma 2.1, one shows that µθ is doubling, i.e., there exists a constant C = Cθ,R
such that µθ (B(ξ, 2r)) ≤ Cµθ (B(ξ, r)) for all ξ in ∂Γ and for all r ≥ 0.
In the same way, since µθ (S(g, R)) ≍θ µ′θ (S(g, R)) for arbitrary two probability
measures µθ and µ′θ satisfying (8), two such measures are comparable on balls, i.e.,
µθ (B(ξ, r)) ≍θ µ′θ (B(ξ, r)). For every Borel set A in ∂Γ, and for all small enough
r > 0, take a compact set K and an open set U such that K ⊂ A ⊂ U, and
µθ (U \ K) and µ′θ (U \ K) are less than r. Since K is compact, there exists r0 > 0
such that all balls of radius less than r0 centred in K are included in U. By the
doubling property of µθ , the Vitali covering theorem (e.g. [Hei, Theorem 1.6]) implies
S countable disjoint balls {Bi } of radius less than r0 centred in K such
that there exist
that µθ (K \ i Bi ) = 0. Therefore
X X
µθ (K) ≤ µθ (Bi ) ≤ C µ′θ (Bi ) = Cµ′θ (∪i Bi ) ≤ Cµ′θ (U).
i i
By letting r tend 0, one obtains µθ (A) ≤ Cµ′θ (A). The reverse inequality is also
obtained in the same way; we conclude the proof.
Let us recall that a measure on the boundary ∂Γ is called ergodic for the action
of Γ, if every Γ-invariant Borel set has the measure null or co-null.
Corollary 3.7. For every θ ∈ R, every probability measure µθ satisfying (8) in
Theorem 3.3 is ergodic for the action of Γ.
14 RYOKICHI TANAKA
Proof. Let A be a Borel set which is Γ-invariant in ∂Γ, i.e., gA = A for every g in Γ.
Suppose that µθ (A) > 0. Since the restriction µθ |A of µθ to the set A satisfies (8) for
the same θ, we have µθ |A ≍ µθ , by Theorem 3.6; this implies that µθ (∂Γ\A) = 0.
The Gibbs property of µθ provides the following strong estimate for x∈Sn G(1, x)θ .
P
Proposition 3.8. For every θ ∈ R,

X
G(1, x)θ ≍θ enβ(θ) .
x∈Sn
Proof. The proof is similar to the one of Lemma 3.2 for β(1) = 0. The shadows
S(x, R) (x ∈ Sn and R > max{4δ, R0 }) cover ∂Γ with bounded overlaps N. From
Corollary 3.5, we have
X X
1 = µθ (∂Γ) ≤ µθ (S(x, R)) ≤ C G(1, x)θ e−nβ(θ) ,
x∈Sn x∈Sn
and, X X
G(1, x)θ ≤ C eβ(θ)n µθ (S(x, R)) ≤ CNenβ(θ) .
x∈Sn x∈Sn
Thus we obtain the claim.
4. Symbolic dynamics
4.1. Automatic structure. Suppose that S is a finite symmetric set of generators

in Γ. An automaton is a directed graph A = (V, E, s∗ ) with a distinguished vertex
s∗ called the initial state, and a labeling on edges by generators α : E → S. For
a path ω in the graph, a sequence of edges e0 , e1 , . . . , en−1 , where the terminus of
ei is the origin of ei+1 , we associate a path α(ω) in the Cayley graph of Γ starting
from the identity: 1, α(e0 ), α(e0 )α(e1 ), . . . , α(e0 )α(e1 ) · · · α(en−1 ). Denote by α∗ (ω)
the terminus of α(ω).
Definition 4.1. We say that a finitely generated group Γ has a strongly Markov
automatic structure if there exists an automaton A = (V, E, s∗ ) such that
(1) Every v in V can be reached by a path ω starting from the initial state s∗ .
(2) For every directed path ω in A, the path α(ω) is a geodesic in Γ.
(3) The map α∗ evaluating the terminus of the path is one-to-one and onto from
the set of paths starting from s∗ to Γ.
By definition, the map α∗ gives a bijection from the set of paths from s∗ of length
n to the sphere Sn (the set of words of length n) in Γ.
A theorem by Cannon ensures that every hyperbolic group has a strongly Markov
automatic structure [Can]. In the sequel, we fix an automaton A for Γ. Let us denote
by Σ∗ the set of finite paths in the graph A (not necessarily starting at s∗ ), by Σ the
set of semi-infinite paths and by Σ = Σ∗ ∪ Σ the union of those sets. We define the
metric dΣ (ω, ω ′) = 2−n in Σ, where ω and ω ′ coincide just until the n-th entry. To a
semi-infinite path in the graph A, we can associate a geodesic ray starting from the
identity in the group Γ, and a point in the boundary as the extreme point. The map
α∗ is extended from Σ to Γ ∪ ∂Γ in this way. We note that the map α∗ is continuous,

in fact, Hölder.
Lemma 4.2. The map α∗ : Σ → Γ ∪ ∂Γ is Hölder continuous.
Proof. This follows from dε (ξ, η) ≤ Ce−ε(ξ|η) .
Our analysis will be built on the structure of the graph A. A component is a

maximal induced subgraph of A where every vertex is accessible from every other
vertex by a finite directed path. In general, an automaton has several distinct
components which we have to deal with. A surface group admits an automaton
with rather simple structure: one large component and just single vertices. For
general hyperbolic groups, however, it is not the case, and it is necessary to capture
how different components interact each other.
4.2. Thermodynamic formalism. In this section, we show that β(θ) is the pres-
sure for the Ruelle-Perron-Frobenius transfer operator for an appropriate subshift of
finite type. This enables us to prove that β(θ) is analytic except for at most finitely
many points. The subshift of finite type is constructed on an automatic structure of
Γ, where the thermodynamic formalism is established by Gouëzel [Gou1] based on
the classical case. It is not necessarily topologically mixing, or not even recurrent,
either. Below we employ the generalised thermodynamic formalism due to Gouëzel,
following [Gou1, Section 3.3].
For a finite directed graph A and the set of finite or semi-infinite paths Σ in A,
the shift σ : Σ → Σ is defined by deleting the first edge of a path. Given every real-
valued Hölder continuous function ϕ : Σ → R (a potential), we define the transfer
operator Lϕ acting on continuous functions by
′
X
Lϕ f (ω) = eϕ(ω ) f (ω ′),
σ(ω ′ )=ω
where for the empty path ω = ∅, we understand that the preimages of the shift σ
are taken by the nonempty ones. The transfer operator encodes the Birkhoff sum
Sn ϕ(ω) in the form that
′
X
Lnϕ f (ω) = eSn ϕ(ω ) f (ω ′), where Sn ϕ(ω) := ϕ(ω) + ϕ(σω) + · · · + ϕ(σ n−1 ω).
σn (ω ′ )=ω
The most fundamental case is where the graph A is topologically mixing, i.e., every
vertex is accessible from every other vertex (then the graph is called recurrent),
and for arbitrary two vertices in the graph, there is a path connecting those two
vertices of length n for all large enough n. In this case, the Ruelle-Perron-Frobenius
theorem describes the spectra of Lϕ ([Bow, Theorem 1.7], [PP, Theorem 2.2], and
[Gou1, Theorem 3.6]). In the case where the graph A is just recurrent, there is a
period p > 1 such that every loop has the length proportional to p. In this case, the
set of vertices of A is decomposed into p distinct subsets Vj where an edge emanating
from a vertex in Vj has the terminus in Vj+1 for every j ∈ Z/pZ. This decomposition
is called a cyclic decomposition of V , and the restriction of σ p to each Vj yields a
subshift of finite type which is topologically mixing. Consider the case where the
16 RYOKICHI TANAKA
graph A is not even recurrent. Then we decompose the graph A into components.
To each component C, one can associate a transfer operator LC by restricting ϕ
to paths staying in C. Each transfer operator LC has finitely many eigenvalues of
maximal modulus eP rC (ϕ) , for some real number P rC (ϕ) (the pressure), and they
are all simple and isolated. Define by P r(ϕ) = maxC P rC (ϕ) the maximum of the
pressure over all components. A component C is called maximal if P rC (ϕ) = P r(ϕ).
The asymptotic behaviour of Lnϕ is controlled by maximal components.
Definition 4.3. The potential ϕ is called semisimple if there are no directed paths
between arbitrary two distinct maximal components.
In the case where the potential ϕ is semisimple, the dominating terms of Lnϕ are
fairly well decomposed as in the following: Let denote by Höl the space of Hölder
continuous functions with some fixed exponent with the norm k · k.
Theorem 4.4 ([Gou1], Theorem 3.8). Suppose that a potential ϕ is semisimple.
Denote by C1 , . . . , CI the maximal components,Fwith the corresponding period pi , and
take for each i a cyclic decomposition Ci = j∈Z/pi Z Ci,j . Then there exist Hölder
R
continuous functions hi,j and measures λi,j with hi,j dλi,j = 1 such that for every
Hölder continuous function f ,
XI pX
i −1 Z
n nP r(ϕ)
Lϕ f − e f dλi,(j−n mod pi ) hi,j ≤ Ckf ke−nε0 enP r(ϕ) .
i=1 j=0
1
Ppi −1
The probability measures µi = pi j=0 hi,j λi,j are invariant under σ and ergodic
(but not mixing in general).
Denote by C→,i,j the set of edges from which Ci,j is accessible with a path of length
proportional to pi , and by Ci,j,→ the set of edges which can be accessible from Ci,j
by a path of length proportional to pi . The function hi,j is bounded from below on
paths starting with edges in Ci,j,→ and the empty path, and vanishes elsewhere. The
support of the measure λi,j is the set of infinite paths starting with edges in C→,i,j
and eventually staying in Ci .
To each component C (not necessarily maximal), we associate the space of infinite

paths staying in C, which we denote by ΣC . There is the Gibbs measure µC on ΣC ,
i.e.,
µC ([ω0 , . . . , ωn−1 ]) ≍ e−nP rC (ϕ)+Sn ϕ(ω) (11)
PpC −1
for every ω in ΣC , where µC is σ-invariant, having the form (1/pC ) j=0 hC,j λC,j by
abusing the notation in Theorem 4.4. We formulate a variational principle on each
ΣC . For a finite Borel measurable P partition D of ΣC and a measure m on it, one
defines the entropy Hm (D) = − D∈D m(D) log m(D). Define the Borel measurable
partition of ΣC by U := {Ux }x∈C , where Ux := {ω ∈ ΣC | ω0 = x}. For a σ-invariant
probability measure m on ΣC , the entropy of m is defined by
1
hm (σ) = lim Hm U ∨ σ −1 U ∨ · · · ∨ σ −n+1 U ,

n→∞ n
where D1 ∨ D2 denotes the partition refined by the partitions D1 and D2 , and

the limit exists for every σ-invariant probability measure m [Bow, Lemma 1.19].
The following corresponds to [Bow, Theorem 1.22], where the claim is stated for a
topologically mixing subshift of finite type, but a recurrent case follows from the
same line of the proof.
Proposition 4.5 (Variational principle). For a Hölder continuous potential ϕ (not
necessarily semisimple), and for each component C, one has
Z
max hm (σ) + ϕdm = P rC (ϕ),
ΣC
where the maximum is taken over all σ-invariant probability measures m on ΣC and
attained by the Gibbs measure µC .
In particular, P r(ϕ) is attained by the Gibbs measure corresponding to a maximal

component. The following proposition describes how the transfer operator behaves
under perturbations of the potential.
Proposition 4.6 ([Gou1], Proposition 3.10). Assume that a potential ϕ in Höl is
semisimple, and C1 , . . . , CI are the corresponding maximal components as in Theorem
4.4. There exist ε0 > 0 and C > 0 such that for all ψ small enough in Höl, there
exist Hölder continuous functions hψi,j and measures λψi,j with the same support as
hi,j and λi,j , respectively, and real numbers P rCi (ϕ + ψ) satisfying that for every f
in Höl,
I pi −1 Z
X X
n
Lϕ+ψ f − enP rCi (ϕ+ψ) ψ
f dλi,(j−n mod pi ) hψi,j ≤ Ckf ke−nε0 enP r(ϕ) .
i=1 j=0
The maps ψ 7→ P rCi (ϕ + ψ), ψ 7→ hψi,j and ψ 7→ λψi,j are analytic in the norm sense
from a small ball centred at 0 in Höl to R, Höl and the dual of Höl, respectively.
Moreover, Z
P rCi (ϕ + ψ) = P r(ϕ) + ψdµi + O kψk2 ,

where µi = p1i pj=0

P i −1
hi,j λi,j .
The proof is based on the decomposition on each component. Note that the
pressures P rCi (ϕ + ψ) are possibly different. One can also show that ϕ + ψ is
semisimple for all small enough ψ, since the pressures of ϕ + ψ on the components
other than the maximal ones are bounded away from P r(ϕ). This implies that the
maximal components of ϕ + ψ appear within those of ϕ for all small enough ψ.
We will need the following lemma to determine whether the potential is semisim-
ple or not. Let 1[E∗ ] be the indicator function on Σ defined on the set of paths
starting at the Pinitial state s∗ . By the definition of the transfer operator, we have
Lnϕ 1[E∗ ] (∅) = eSn ϕ(ω) , where the summation is taken over all paths starting at s∗
of length n.
Lemma 4.7. Let P r(ϕ) = maxC P rC (ϕ). Suppose that there are L components in
a finite directed graph A. We have
Lnϕ 1[E∗ ] (∅) ≤ CnL enP r(ϕ) .
18 RYOKICHI TANAKA
On the other hand, if there is a path from the initial state s∗ to successively k different
maximal components, then we have
Lnϕ 1[E∗ ] (∅) ≥ C ′ nk−1 enP r(ϕ) .
Proof. This is [Gou1, Lemma 3.7], in particular, the latter is a special case of it.
We give the proof of the former for completeness. Observe that in the components
graph which is the directed graph obtained by identifying each component with a
point, there are no loops, and thus, there are only finitely many directed paths.
Decompose a path ω in the automaton: When a path ω starting from s∗ goes
through k different components C1 , . . . , Ck successively, we decompose a path in
the automaton by ω = (u0, v1 , u1, . . . , ui−1 , vi , ui , . . . , uk−1, vk , uk ), where each vi is a
path in Ci , and u0 , u1 , . . . , uk−1 are transient paths connecting components C1 , . . . , Ck
successively and uk is a path emanating from Ck , possibly empty. By the Hölder
continuity of ϕ, we get
X X
Sn ϕ(ω) ≤ Ck + S|vi | ϕ(vi , ui , vi+1 , . . . , uk ) ≤ C ′ k + S|vi | ϕ(vi ),
i i
and thus,
   
X X X X
Lnϕ 1[E∗ ] (∅) ≤ eCk  eSa1 ϕ(v1 )  · · ·  eSak ϕ(vk )  ,
a1 +···+ak =n−b |v1 |=a1 |vk |=ak
P
where b = |ui | and the first summation is taken over all possible paths in the
components graph. For each component C, the eigenvalue of maximal modulus of
P r(ϕ)
, hence we obtain |vi |=ai eSai ϕ(vi ) ≤
P
Lϕ restricted on C, is not greater than e
Ceai P r(ϕ) . Therefore it holds that
X X
Lnϕ 1[E∗ ] (∅) ≤ eCk e(n−b)P r(ϕ) ≤ CnL enP r(ϕ) ,
a1 +···+ak =n−b
and the lemma follows.
Let A = (V, E, s∗ ) be a strongly Markov automaton associated with a non-

elementary hyperbolic group Γ, and K(x, y) be the Martin kernel corresponding
to a finitely supported admissible probability measure µ on Γ. We define the pres-
sure by
ϕ(ω) = − log K(α∗ ω0 , α∗ ω), (12)
where a path ω = (ω0 , ω1 , . . . ) is a sequence of edges in the graph A. By the cocycle
identity: K(gg ′, ξ) = K(g, ξ)K(g ′, g −1ξ), we have
G(1, α∗ ω̄n ) ≍ eSn ϕ(ω) (13)
for the path ω̄n cut out from ω at the length n, where we used |Sn ϕ(ω) −Sn ϕ(ω̄n )| ≤
C by the Hölder continuity of ϕ.
By Lemma 4.2, the map α∗ : Σ → Γ ∪ ∂Γ is Hölder continuous, and thus the
pressure (12) is also Hölder continuous function on Σ.
Proposition 4.8. Let ϕ be the potential defined in (12). For every θ ∈ R, it holds
that β(θ) = P r(θϕ). In particular, β is analytic except for at most finitely many
points. Moreover, θϕ is semisimple for every θ ∈ R.
Proof. For each component C, the pressure P rC (θϕ) is analytic in θ; hence P r(θϕ)
is analytic in θ except for at most finitely many points since it is defined as the
maximum of finitely many analytic functions. For a path ω of finite length, the
pressure has the following form:
G(α∗ ω0 , α∗ ω)
ϕ(ω) = − log .
G(1, α∗ ω)
Therefore we obtain
1 X
G(1, x)θ = Lnθϕ 1[E∗ ] (∅).
G(1, 1)θ x∈S
n
By Lemma 4.7 and Proposition 3.8, we have β(θ) = P r(θϕ). Observe that if there
is a path from a maximal component to the other maximal component, then con-
catenating a path emanating from s∗ , one can assume that the path is starting from
s∗ and going through two distinct maximal components. If there is a directed path
from s∗ passing through k > 1 different maximal components, then by the second
inequality of Lemma 4.7, one has enβ(θ) ≥ Cnk−1 enP r(θϕ) ; a contradiction.
Thermodynamic formalism gives some additional information about a probability

measure µθ constructed in Theorem 3.3. Henceforth, we consider the potential ϕ
defined in (12). Fix θ ∈ R, let p1 , . . . , pI be the corresponding periods of maximal
components for the pressure θϕ, and denote by p the least common multiple of
those periods. Define the measures m̃np+q , q = 0, . . . , p − 1, by the positive linear
np+q
functionals on Σ: m̃np+q (f ) = e−(np+q)P r(θ) Lθϕ (1[E∗ ] f )(∅). For each fixed q, the
measures m̃np+q converge to some measure m̃q according to Theorem 4.4. Then the
Cesàro average of the measures ( N
P PN
n=1 m̃n )/( n=1 m̃n (Σ)) converge to a probability
measure weakly. (First, those measures converge for each Hölder continuous function
f , then we deduce the weak convergence by approximating each continuous function
f by Hölder continuous ones on Σ.) We denote the weak limit of the measures
( N
P PN
n=1 m̃n (1[E∗ ] ·))/( n=1 m̃n (1[E∗ ] )) by m̃θ .
Theorem 4.9. The measures µθ,s defined by (9) in Theorem 3.3 converge to α∗ m̃θ
as s ↓ β(θ) without passing to a subsequence, where α∗ m̃θ denotes the push-forward
of m̃θ by the map α∗ . In particular, the measure α∗ m̃θ is comparable with µθ .
Proof. In the above construction of the measure m̃θ , take f ◦ α∗ for every continuous
function f on Γ ∪ ∂Γ. By the relation x∈Sn G(1, x)θ f (x) = G(1, 1)θ Lnθϕ (1[E∗ ] f ◦
P
α∗ )(∅), we deduce that µθ,s converges weakly to α∗ m̃θ as s ↓ β(θ). As we see in the
proof of Theorem 3.3, every weak limit of µθ,s as s ↓ β(θ) satisfies (8). Therefore
α∗ m̃θ is comparable with µθ by Theorem 3.6.
The ergodicity of µθ provides the differentiability of β at θ ∈ R.

Proposition 4.10. If µθ is ergodic for the action of Γ for some θ ∈ R, then the
corresponding β is differentiable at θ.
Proof. Recall that β(θ) = maxC P rC (θϕ) by Proposition 4.8. For each θ ∈ R,
let C1 , . . . , Ck be the corresponding maximal components for the potential θϕ. By
′
R
Proposition 4.6, for each Ci , we have P rCi (θϕ) = ϕdµi (the derivative at θ). We
20 RYOKICHI TANAKA
will show that it holds that ϕdµi = ϕdµi′ for i 6= i′ , and then the differentiability
R R
of β at θ will follow. This is nothing but [Gou1, Proposition 3.16]. (The argument
is well adapted without almost any modifications.) Below we describe the proof for
convenience, following
R [Gou1] and the idea of Calegari and Fujiwara [CF, Lemma
4.24]. Let ci := ϕdµi . We define Ei the set of points ξ in ∂Γ such that the sequence
log F (1, γ(n))/d(1, γ(n)) converges to ci along some geodesic γ towards ξ. Notice
that in the above, one can replace some geodesic by every geodesic towards ξ, and
in particular, the set Ei is Γ-invariant. We will show that µθ (Ei ) = 1. This will
imply that Ei ∩ Ei′ 6= ∅, and thus ci = ci′ . Denote by Oi ⊂ Σ the set of infinite
paths ω such that the Birkhoff averages (1/n)Sn ϕ(ω) have the limit ci . The Birkhoff
ergodic theorem implies that µi (Oi ) = 1. (Recall that µi is σ-invariant and ergodic
on Σ.) We have α∗−1 Ei = Oi ∩ [E∗ ], where [E∗ ] is the set of paths starting at s∗ ,
since FP(1, α∗ ω̄n ) ≍ eSn ϕ(ω) by (13). The measure µi is comparable with the measure
βi = j λi,j on the set of paths staying in the component Ci . This enables one
to show that βi (Oic ) = 0 [Gou1, p.916]. We will show that µθ (Ei ) > 0. Suppose
the contrary. Since the push-forward of βi (· ∩ [E∗ ]) by α∗ is absolutely continu-
ous with respect to α∗ m̃θ which is comparable with µθ by Theorem 4.9, we have
βi (Oi ∩ [E∗ ]) = 0. The set Oic has the βi -measure 0 by the above, hence βi ([E∗ ]) = 0
and this is a contradiction since the measure βi has a positive mass on [E∗ ] by The-
orem 4.4. Above all, we have µθ (Ei ) > 0, in fact, the measure equals 1 if µθ is
ergodic. Thus we conclude the proof.
Corollary 4.11. For every finitely supported admissible probability measure µ on
Γ, the function β(θ) is continuously differentiable at every θ ∈ R.
Proof. Recall that β is analytic except at most finitely many points by Proposition
4.8. For every θ ∈ R, a probability measure µθ is ergodic by Corollary 3.7; the claim
follows from Proposition 4.10.
5. Hausdorff spectrum
For every finitely supported admissible probability measure µ on a non-elementary

hyperbolic group Γ equipped with a word metric, let ν be the corresponding har-
monic measure on the geometric boundary ∂Γ. In this section, we consider the
dimensional properties of the harmonic measure ν. First, we discuss about a typical
pointwise behaviour of ν, and establish a formula on the Hausdorff dimension of ν
with the entropy and the drift of the random walk. Next, we consider an atypical
pointwise behaviour of ν, and introduce the Hausdorff spectrum (the multifractal
spectrum), which describes a variety of pointwise dimensions at ν-measure null. The
Hausdorff spectrum is given by the Legendre transform of β, and this allows us to
prove that the fundamental inequality becomes the genuine equality if and only if
the harmonic measure ν is comparable with the D-Hausdorff measure HD on the
boundary ∂Γ. This holds for every finitely supported admissible probability measure
µ on Γ.
5.1. Hausdorff spectrum of the harmonic measure. Let µ be an arbitrary

finitely supported admissible probability measure on Γ. First, we prove the following
lemma for µ not necessarily symmetric. Define dG (x, y) = − log F (x, y) for x, y in
Γ. Here dG satisfies the triangular inequality in the sense that dG (x, y) ≤ dG (x, z) +
dG (z, y) and the positivity dG (x, y) > 0 for x 6= y, but does not satisfy the symmetry.
Lemma 5.1. There exist constants L > 0 and C ≥ 0 such that for all x, y in Γ,
L−1 d(x, y) − C ≤ dG (x, y) ≤ Ld(x, y).
Proof. Suppose that d(x, y) = m. Let L := maxs∈S dG (1, s) > 0. By the triangular
inequality for dG , we have dG (x, y) ≤ Lm = Ld(x, y), and thus the second inequality
holds. To show the first inequality, it follows from the fact that the Green function
satisfies that G(x, y) ≤ Ce−λd(x,y) . See also [BHM2, Proposition 3.5].
Recall that the Hausdorff dimension of the measure ν on the metric space (∂Γ, dε )
is defined by
dimH ν = inf{dimH C | ν(C c ) = 0}.
We show that the so-called dimension, entropy and drift formula for the harmonic
measure ν.
every non-elementary hyperbolic group Γ equipped with a word metric, the corre-
sponding harmonic measure ν satisfies that
log ν (B(ξ, r)) h
lim = ,
r→0 log r εl
for ν-a.e. ξ. In particular, dimH ν = h/(εl).
Proof. Note that by Lemma 2.1, it is enough to show the claim for the shadows
S(g, R): Fix an R > 4δ and for ν-a.e. ξ in ∂Γ, take an arbitrary geodesic γ emanating
from 1 towards ξ, we shall show that −(1/n) log ν (S(γ(n), R)) ∼ h/l. Notice that
ν (S(γ(n), R)) ≍ G(1, γ(n)) by Theorem 2.6, and thus (1/n) log ν (S(γ(n), R)) ∼
(1/n) log G(1, γ(n)) for such a geodesic ray. By Theorem 2.2, for P-almost every
sample path of the random walk, there is a geodesic ray γω from 1 converging to the
point ξ in the boundary such that d(xn , γω (⌊ln⌋)) = o(n), where l > 0 is the drift.
The triangular inequality of dG shows that
|dG (1, xn ) − dG (1, γω (⌊ln⌋))| ≤ max{dG (xn , γω (⌊ln⌋)) , dG (γω (⌊ln⌋), xn )}.
The inequality dG (x, y) ≤ Ld(x, y) in Lemma 5.1 implies that (1/n)dG (1, γω (⌊ln⌋)) ∼
(1/n)dG (1, xn ). The drift of the random walk with respect to dG (the Green speed)
coincides with the entropy h [BHM1, Theorem 1.1]. Therefore we obtain
1
dG (1, γω (⌊ln⌋)) ∼ h (14)
n
for P-a.e. ω. Above all, it holds that −(1/⌊ln⌋) log ν (S(γω (⌊ln⌋), R)) ∼ h/l for
P-a.e. ω. The first assertion yields the second e.g., [Pes, Theorem 7.1, Chapter 2,
p.42], by adapting to the present setting.
Remark 5.3. In the above proof, it is essential that the Green speed coincides with
the entropy [BHM1]. We note that it is shown for µ not necessarily symmetric in
their paper.
22 RYOKICHI TANAKA
Let us define Hausdorff spectrum of the measure ν. For every real number α, we
consider the following set:

log ν (B(ξ, r)) α
Eα = ξ ∈ ∂Γ lim = .
r→0 log r ε
The set Eα is possibly empty for some α. We call dimH Eα as the function of α the
Hausdorff spectrum of the measure ν. Let αmin := − limθ→+∞ β ′ (θ) and αmax :=
− limθ→−∞ β ′ (θ). Denote by f (α) = inf θ∈R {αθ + β(θ)} the Legendre transform of β
which is concave and continuous on [αmin , αmax ]. Equivalently, f (α) = −θβ ′ (θ)+β(θ)
for every α = −β ′ (θ). Note that the interval [αmin, αmax ] is a singleton if and only if
β is affine.
every non-elementary hyperbolic group Γ equipped with a word metric, the Hausdorff
spectrum of the corresponding harmonic measure ν is given by
f (α)
dimH Eα = ,
ε
for every α ∈ (αmin, αmax ). Moreover, the interval [αmin, αmax ] is bounded in R>0
and Eα = ∅ for every α ∈/ [αmin , αmax ].
Proof. Recall that β is continuously differentiable on R by Corollary 4.11. For each

θ ∈ R, the derivative of β at θ has a strictly negative value since β is convex and
β(θ) → ±∞ (resp.) as θ → ∓∞ (resp.) by Lemma 3.2. Let us take α = −β ′ (θ) ∈
(αmin , αmax ) if αmin 6= αmax ; or, α = αmin (or αmax ) if αmin = αmax . First, we will
show that µθ (Eα ) = 1 for such a θ. Consider the following set: For an arbitrary small
enough s > 0 and an arbitrary large enough n, the set of points ξ in ∂Γ, to which
there exists a geodesic ray γ from 1 converging such that ν (S(γ(n), R)) ≥ e−n(α−s) .
The measure µθ of such a set is bounded from above by for every r > 0,
Z
−n(α−s)
en(α−s)r ν (S(γ(n), R))r dµθ (ξ).

µθ {ξ ∈ ∂Γ | ν (S(γ(n), R)) ≥ e } ≤
∂Γ
(15)
(The map ξ 7→ γ(n) has to be measurable; this is obtained as in Theorem 2.2.) The
right hand side of (15) is bounded from above by
X
en(α−s)r ν (S(x, R))r µθ (S(x, R)) .
x∈Sn
By the Gibbs property of µθ (Corollary 3.5), we have µθ (S(x, R)) ≍ G(1, x)θ e−β(θ)|x| .
Together with ν (S(x, R)) ≍ G(1, x) by Theorem 2.6, the ingredient of the above
n(α−s)r θ+r −β(θ)n θ+r
P
summation is bounded by Ce G(1, x) e . Since x∈Sn G(1, x) ≍
β(θ+r)n n(α−s)r (β(θ+r)−β(θ))n
e by Proposition 3.8, we eventually obtain a bound by Ce e .
Since β is differentiable at θ, we have β(θ+r)−β(θ) = −αr+o(r) as r ↓ 0. Therefore
for all small enough r > 0 there exists a constant c(s, r) > 0 such that
µθ {ξ ∈ ∂Γ | ν (S(γ(n), R)) ≥ e−n(α−s) } ≤ Ce−nc(s,r).

It follows that
log ν (S(γ(n), R))
lim inf ≥ α − s, for µθ -a.e. ξ.
n→∞ −n
In a similar way, by taking the reverse inequality ν (S(γ(n), R)) ≤ e−n(α+s) in the
beginning and replacing r by −r, we also get
log ν (S(γ(n), R))
lim sup ≤ α + s, for µθ -a.e. ξ.
n→∞ −n
We conclude that limn→∞ (−1/n) log ν (S(γ(n), R)) = α, for µθ -a.e. ξ. The local
dimension (times the parameter ε) of µθ is obtained by:
log µθ (S(γ(n), R)) θ log ν (S(γ(n), R)) − β(θ)n
lim = lim = θα + β(θ),
n→∞ −n n→∞ −n
for µθ -a.e. ξ. Therefore, comparing the shadows with the balls by Lemma 2.1, we
have ε dimH Eα = θα + β(θ). (On the estimate of Hausdorff dimension, see, e.g.
[Fal, Proposition 4.9, p.74], where the proof is written in the Rn case, but is also
adapted to compact metric spaces.)
Next, we shall show that [αmin , αmax ] is a bounded interval in R>0 . For each θ ∈ R,
denote by mθ the Gibbs measure supported on ΣC corresponding to a maximal
component C forRθϕ. Let m∞ be a weak limit of {mθ } as θ tends R to +∞. Since for
′
each θ, we have ϕdmθ = βR (θ) by Proposition 4.6, we obtain ϕdm+∞ = −αmin .
In the same way, we obtain ϕdm−∞ = −αmax , where m−∞ is a weak limit of {mθ }
as θ tends to −∞. By the variational principle Proposition 4.5, for arbitrary θ0 and
θ, we have Z
hmθ + θ0 ϕdmθ ≤ β(θ0 ).
R
Choose a positive θ0 such that β(θ0 ) < 0. Since hmθ ≥ 0, we have − ϕdmθ ≥
−β(θ0 )/θ0 > 0 for every θ, and thus we get αmin > 0. On the other hand,
Z
αmax = − lim ϕdmθj ≤ sup(−ϕ(ω)) < ∞.
θj →−∞ ω∈Σ
Finally, we shall show that Eα = ∅ for every α ∈ / [αmin , αmax ]. Suppose that for an
α∈ / [αmin , αmax ], there exists a ξ in ∂Γ such that limr→0 (log ν (B(ξ, r))) /(log r) =
α/ε. Then we have an infinite path ω starting at s∗ such that α∗ ω = ξ and
(1/n)Sn ϕ(ω) → −α as n tends to ∞. Notice that such an infinite path ω eventually
stays in some component C (not necessarily maximal for ϕ). Define a measure on Σ
by
n−1
1X
mω,n := δ k .
n k=0 σ ω
Let mω be a weak limit of {mω,n }. Note that the measure mω is σ-invariant and
supported in ΣC . The variational principle Proposition 4.5 implies that for every θ,
Z
hmω (σ) + θ ϕdmω ≤ β(θ),
i.e., the line defined by the left hand side lies under the convex curve defined by the
right hand side in the above inequality. Thus there exists a θ ∈ [−∞, +∞] such
that Z Z
ϕdmω = ϕdmθ , (16)
24 RYOKICHI TANAKA
where we have θ = −∞ or +∞ when the line is asymptotic to β(θ). Here the

′
Rright hand side of (16) equals to Rβ (θ) which is in [−αmax ,R−αmin ]. Notice that since
ϕdmω,nj = (1/nj )Snj ϕ(ω) → ϕdmω , we have −α = ϕdmω . Hence it has to
hold that −α ∈ [−αmax , −αmin ]; a contradiction. We complete the proof.
Remark 5.5. In fact, we are able to show that the set of irregular points Eirr =
{ξ ∈ ∂Γ | limr→0 (log ν (B(ξ, r))) /(log r) does not exist.} is not empty. Indeed, for
each component C (whose cardinality is larger than one), one can find a path ω in ΣC
such that the Birkhoff averages (1/n)Sn ϕ(ω) do not converge by a result of Barreira
and Schmeling [BS, Theorem 2.1]. It has been also shown that such paths form a
set of full topological entropy [ibid]. It is expected to hold that dimH Eα = f (a)/ε
for α = αmin and αmax as well, but those would require some additional argument
(e.g., [PU, Section 9.2]).
Remark 5.6. If an automaton has a unique component whose cardinality is larger
than one, then β is analytic by definition and thus the Legendre transform f (α)
is also analytic on (αmin , αmax ). This is the case for free groups and for surface
groups, i.e., the fundamental groups of closed Riemann surfaces of genus greater
than one, for example. The condition that an automaton has a single non-trivial
component corresponds to Assumption 1 in [HMM]. In general, we have shown
that β is not only continuously differentiable but also analytic except for at most
finitely many points, hence it would be reasonable to expect that β is also analytic
for every non-elementary hyperbolic group and every finitely supported admissible
probability measure µ on it.
5.2. Proof of Theorem 1.1. First, we prove the following proposition which is a
slight extension of [BHM2, Proposition 5.4]. Let µ̌ be the reflected measure of µ,
i.e., µ̌(x) = µ(x−1 ). Notice that if µ is finitely supported admissible on Γ, then the
reflected measure µ̌ is also the case. Denote by ν and ν̌ the corresponding harmonic
measures on ∂Γ for µ and µ̌, respectively.
Proposition 5.7. If the D-Hausdorff measure HD and the harmonic measure ν,
and HD and the reflected harmonic measure ν̌ are mutually absolutely continuous,
respectively, then their Radon-Nikodym derivatives are uniformly bounded from above
and from below, i.e., HD ≍ ν and HD ≍ ν̌.
Proof. Define for x, y in Γ,

G(x, y)
φ(x, y) := .
G(x, 1)G(1, y)
There is an extension of φ on ∂ 2 Γ := ∂Γ × ∂Γ \ diagonal, which is Hölder continuous
on each compact subset. (See [Led, Proposition 3.17]. It is written for free groups
case, but the argument applies to the present setting as well.) Indeed, take ξ and η in
∂Γ (ξ 6= η) and r > 0 such that B(ξ, r)∩B(η, r) = ∅ (the balls are defined in Γ∪∂Γ).
For every x ∈ B(ξ, r)∩Γ and every y ∈ B(η, r)∩Γ, the function φ(x, y) is bounded by
the Ancona inequality (5) and the Harnack inequality (6). It holds that uniformly
for a fixed x, the function y 7→ G(x, y)/G(1, y) is Hölder continuous on B(η, r),
and uniformly for a fixed y, the function x 7→ G(x, y)/G(x, 1) = Ǧ(y, x)/Ǧ(1, x)
is also Hölder continuous on B(ξ, r) by Theorem 2.5. Therefore the function φ is
extended on ∂ 2 Γ and is Hölder continuous on each compact subset. The measure

dν̃(ξ, η) := φ(ξ, η)d(ν̌ ⊗ ν)(ξ, η) on ∂ 2 Γ is σ-finite and invariant under the diagonal
action of Γ, since
φ(ξ, η) G(x, g)G(g, y) dg∗(ν̌ ⊗ ν)
−1 −1
= lim = Ǩ(g, ξ)K(g, η) = (ξ, η).
φ(g ξ, g η) x→ξ, y→η G(x, 1)G(1, y) d(ν̌ ⊗ ν)
Let µ0 be a Patterson-Sullivan measure on ∂Γ. Define the measure on ∂ 2 Γ by
d(µ0 ⊗ µ0 )(ξ, η)
dµ̃0 (ξ, η) := .
exp(−2v(ξ|η))
Here the measure µ̃0 is σ-finite and Γ-quasi-invariant on ∂ 2 Γ, i.e., there exists a
constant C ≥ 1 such that C −1 µ̃0 (A) ≤ µ̃0 (gA) ≤ C µ̃0 (A) for all g in Γ and all
Borel sets A in ∂ 2 Γ [Coo, Corollaire 9.4]. Suppose that µ0 and ν, and µ0 and ν̌ are
mutually absolutely continuous, respectively. We have dµ0 = f dν and dµ0 = fˇdν̌
for some positive Borel measurable functions f and fˇ on ∂Γ. We obtain dµ̃0 (ξ, η) =
F̃ (ξ, η)dν̃(ξ, η), where
exp(2v(ξ|η))
F̃ (ξ, η) := fˇ(ξ)f (η) .
φ(ξ, η)
First, we shall prove that F̃ is uniformly bounded from above and from below ν̃-
a.e. Define the set AC := {(ξ, η) ∈ ∂ 2 Γ | C −1 ≤ F̃ (ξ, η) ≤ C}, then there exists
a constant C ≥ 1 such that ν̃(AC ) > 0. The measure ν̃ is ergodic for the diagonal
action of Γ since ν̌ ⊗ ν is ergodic [Kai2, Theorem 6.3]. Therefore, for ν̃-a.e. (ξ, η) in
∂ 2 Γ, there exists g in Γ such that g(ξ, η) ∈ AC . Since µ̃0 is Γ-quasi-invariant and ν̃
is Γ-invariant, we have F̃ (gξ, gη) ≍ F̃ (ξ, η) for all g in Γ and for ν̃-a.e. (ξ, η). Hence
there exists a constant C ′ ≥ 1 such that C ′−1 ≤ F̃ (ξ, η) ≤ C ′ for ν̃-a.e. (ξ, η). Thus
we have
φ(ξ, η)
fˇ(ξ)f (η) ≍ , ν̃-a.e.
exp(2v(ξ|η))
Suppose that f is unbounded on B(η, r) for all small enough r > 0. Let ξ 6= η such
that fˇ(ξ) > 0. The right hand side φ(ξ, η ′)/ exp(2v(ξ|η ′)) is uniformly bounded
for η ′ ∈ B(η, r) for all small enough r, while the left hand side fˇ(ξ)f (η ′) is not; a
contradiction. This implies that f (and also fˇ) has to be bounded on ∂Γ, and the
same argument yields the boundedness of 1/f (and 1/fˇ). The claim follows.
The logarithmic volume growth v of (Γ, d) has an interpretation as the Hausdorff

dimension of the boundary ∂Γ, i.e., dimH ∂Γ = v/ε ([Coo, Corollaire 7.6] and [Cal,
Corollary 2.5.10]). This leads a natural implication of the fundamental inequality
h ≤ lv as the inequality of dimensions: dimH ν ≤ dimH ∂Γ, as it is pointed out by
Vershik [Ver, Introduction, p.670]. Below we show that ν is of maximal Hausdorff
dimension if and only if ν and HD are comparable. Recall that D = v/ε. The
following is a restatement of Theorem 1.1 from Section 1.
Theorem 5.8. For every finitely supported admissible probability measure µ on ev-
ery non-elementary hyperbolic group Γ equipped with a word metric, it holds that
h = lv, if and only if the corresponding harmonic measure ν and the D-Hausdorff
measure HD on the boundary ∂Γ are mutually absolutely continuous and their den-
sities are uniformly bounded from above and from below, i.e., ν ≍ HD .
26 RYOKICHI TANAKA
Remark 5.9. The analogy between the logarithmic volume growth v and the topo-
logical entropy is pointed out in [Ver, p.681].
Proof of Theorem 5.8. Since the inequality h ≤ lv is equivalent to the inequality

dimH ν ≤ dimH ∂Γ by Theorem 5.2, we have the equality h = lv, as soon as the
harmonic measure ν is comparable with the D-Hausdorff measure HD . We shall
prove the converse. Recall that β is differentiable at every θ ∈ R (Corollary 4.11).
In view of Theorem 5.4, the Hausdorff spectrum dimH Eα has the maximum v/ε
at α = −β ′ (0). On the other hand, dimH Eα = h/(εl) at α = −β ′ (1). Indeed,
since ε dimH Eα = α at α = −β ′ (1) and µ1 = ν, one has α = h/l (Corollary 3.4
and Theorem 5.2). The fundamental observation is that the inequality h/l ≤ v
is strict if and only if β is strictly convex on [0, 1]. Suppose that the equality
h = lv holds. Then β has to be affine on [0, 1] (in fact, on R). Let us recall
that β(θ) = P r(θϕ) = maxC P rC (θϕ) (Proposition 4.8). We adapt the notation
in Theorem 4.4. Denote by C1 , . . . , CI the maximal components at θ = 0. Every
P rCi (θϕ) coincides with −vθ + v on R, since each P rCi (θϕ) is convex and analytic.
Hence Ci , i = 1, . . . , I are also precisely maximal components for all θ ∈ [0, 1], in
particular, for θ = 1. Since we now have P rC′′i (θϕ) = 0 for all θ ∈ R, the potential ϕ
is cohomologous to a constant, i.e., ϕ = a + u ◦ σ − u for some constant a and some
Hölder continuous function u on some subset Σ′ of Σ, by [PP, Proposition 4.12].
(Although in [PP], the statement is written for a topologically mixing subshift of
finite type, it is adapted to our setting by decomposing into components.) Here the
domain Σ′ ofSu is the union of the sets Σi of infinite Ppaths staying in the component
Ci , i.e., Σ = Σi . We use the relation Lϕ 1[E∗ ] (∅) = eSn ϕ(ω) , where the summation
′ n
runs over the paths ω of length n starting at s∗ . We define the set of paths ω starting
at s∗ satisfying that for a fixed K, the path ω is in one of those components Ci for
all n ≥ K, i.e.,
I
[
ΣK := {ω ∈ Σ | ω starts at s∗ and ωn ∈ Ci for all n ≥ K}.
i=1
Let ∂ΓK := α∗ (ΣK ). Recall that the components Ci , i = 1, . . . , I, are maximal both
at θ = 0 and at θ = 1, and µθ ≍ α∗ m̃θ , where m̃
SθI for θ = 0 and 1 are supported in
the common set of paths eventually staying in i=1 Ci . Therefore we deduce that
[
∂Γ = ∂ΓK , HD -a.e. and also ν-a.e., (17)
K
since HD ≍ µ0 and ν = µ1 (Corollary 3.4). For every ω ∈ ΣK , we have

|Sn ϕ(ω) − an| ≤ CK , (18)
where CK is a constant depending on K. Indeed, let ω be any extension of ω
staying in some component Ci , it follows that |Sn ϕ(ω) − Sn ϕ(ω)| ≤ C by the Hölder
continuity of ϕ, and |Sn ϕ(ω) − an| ≤ CK , by using |Sn ϕ(ω) − (ϕ(σ K ω) + · · · +
ϕ(σ n−1 ω))| ≤ CK for all n ≥ K. Then we obtain
Lnϕ 1[E∗ ] (∅) ≥ CK e(a+v)n .

On the other hand, as in the proof of Lemma 4.7, we deduce that there exists some
ε0 > 0 such that
Lnϕ 1[E∗ ] (∅) ≤ CnL (e−nε0 + en(a+v) ),
for all n > 0, since the summation over paths which do not stay in any Ci contributes
as O(e−nε0 ) at θ = 1. By Proposition 3.8, Lnϕ 1[E∗ ] (∅) ≍ 1 due to β(1) = 0, we
conclude that a = −v. Therefore, by (18), there exists a constant CK > 0 such that
for every path ω ∈ ΣK and for every n ≥ K, we have
−1 −vn
CK e ≤ eSn ϕ(ω) ≤ CK e−vn .
Together with G(1, α∗ ω̄n ) ≍ eSn ϕ(ω) by (13), we deduce that G(1, α∗ ω̄n ) ≍K e−vn
for all paths ω ∈ ΣK and for every n ≥ K. Consider ∂ΓK such that ν(∂ΓK ) > 0.
For every ξ ∈ ∂ΓK , there exists a path ω ∈ ΣK such that α∗ ω = ξ. Therefore we
obtain that there exists a constant CK > 0 such that for every ξ ∈ ∂ΓK and for
every n ≥ K,
−1 −vn
CK e ≤ ν (S(α∗ ω̄n , R)) ≤ CK e−vn ,
−1 v/ε
by Theorem 2.6, we also have CK r ≤ ν (B(ξ, r)) ≤ CK r v/ε , for every ξ ∈ ∂ΓK
and for all small enough r > 0 by Lemma 2.1. A standard covering argument
yields ν( · ∩ ∂ΓK ) ≍K HD ( · ∩ ∂ΓK ) (e.g. [Cal, Corollary 2.5.10]). Since we have
(17), we conclude that HD and ν are mutually absolutely continuous. Moreover,
we shall prove that their Radon-Nikodym derivatives are uniformly bounded from
above and from below. Indeed, notice that the entropy ȟ and the drift ˇl for the
reflected measure µ̌ coincide with those for µ, respectively, since µ̌∗n (x) = µ∗n (x−1 ).
Thus h = lv is equivalent to ȟ = ˇlv. As above, we also conclude that HD and ν̌
are mutually absolutely continuous. Therefore by Proposition 5.7, in fact, we have
HD ≍ ν.
Remark 5.10. For a fixed R ≥ R0 and for every g in Γ, we have HD (S(g, R)) ≍ e−v|g|
by Corollary 3.5 (see also [Coo, Proposition 6.1]). If HD ≍ ν, then combining with
G(1, g) ≍ ν(S(g, R)) (Theorem 2.6), we obtain G(1, g) ≍ e−v|g| ; In other words,
there exists a constant C > 0 such that for every g in Γ, it holds that
|log G(1, g) + v|g|| ≤ C. (19)
The last estimate (19) implies that h = lv since the Green speed is the entropy
[BHM1, Theorem 1.1]. The equivalence between h = lv and (19) is proven in
[BHM2, Corollary 1.2 and Theorem 1.5] (see also [Hai]) for every finitely supported
admissible symmetric probability measure µ on Γ. Compare with results on the
rigidity of cocycles in [GMM2].
6. finitary results
We show finitary versions of Theorem 5.2, which are inspired by the results on
simple random walks on Galton-Watson trees by Lyons, Pemantle and Peres [LPP].
As a consequence, if h < lv, then the random walk is confined in a strictly smaller
region in Γ; but if h = lv, then such confinement does not happen. We make it more
precise. Denote by r the diameter of the support of the step distribution µ, and by
[Sn ] the r-neighbourhood of the sphere Sn in Γ (the annulus). We define the first
hitting time for the random walk {xn }∞ n=0 to the annulus [Sn ] by τn := min{k ≥
28 RYOKICHI TANAKA
0 | xk ∈ [Sn ]}, where τn is finite P-a.s., since the random walk is transient and the
annulus is thick enough for the range of the step. Let ν[n] be the distribution of xτn ,
which is supported on [Sn ].
Theorem 6.1. Let ĥ := h/l.

(1) For the sequence of the first hitting points {xτn }∞
n=0 ,
1
lim − log ν[n] (xτn ) = ĥ, P-a.s.
n→∞ n
(2) For every a > 0,

−n(ĥ+a) −n(ĥ−a)
lim ν[n] x ∈ [Sn ] e ≤ ν[n] (x) ≤ e = 1.
n→∞
(3) For every a ∈ (0, 1), define Kn (a) as the minimal number of points which
form a set carrying the ν[n] -measure at least a, then
1
lim log Kn (a) = ĥ.
n→∞ n
Proof. The proof is analogous to [LPP, Theorem 9.8]. Note that (1) implies (2)
since almost sure convergence yields convergence in probability, and (2) implies (3)
directly. We prove (1). Recall that the Green speed coincides with the entropy:
1
lim − log F (1, xn ) = h,
n→∞ n
P-a.s. by [BHM1, Theorem 1.1]. Since we have ν[n] (xτn ) ≤ F (1, xτn ), and τn /n → 1/l
as n tends to infinity P-a.s., it holds that
1 h
lim inf − log ν[n] (xτn ) ≥ = ĥ,
n→∞ n l
P-a.s. For every s > ĥ and every 0 < t < (s − ĥ)/2, define the set
An = {x ∈ [Sn ] | ν[n] (x) ≤ e−n(s−t) , F (1, x) ≥ e−n(ĥ+t) }.

P
We have x∈[Sn ] F (1, x) ≍ 1 by Proposition 3.8, and thus we deduce that ν[n] (An ) ≤
−n(s−ĥ−2t)
Ce . Therefore, by the Borel-Cantelli lemma, P-a.s., there exists N such
that for all n ≥ N the point xτn is not in An . Since we have F (1, xτn ) ≥ e−n(ĥ+t)
eventually P-a.s., we obtain ν[n] (xτn ) ≥ e−n(s−t) eventually P-a.s. Hence it holds that
1
lim sup − log ν[n] (xτn ) ≤ s − t,
n→∞ n
P-a.s., by taking s arbitrary close to ĥ and t arbitrary close to 0, the theorem
follows.
We show a “confinement” result of the random walk, stated in Theorem 1.3. The
following is known for almost every realisation of Galton-Watson tree conditioning
on having infinite trees ([LPP, Theorem 9.9] and [LP, Theorem 16.30]).
Theorem 6.2. For every a ∈ (0, 1), there exists a subset Γa in Γ such that
P (xn ∈ Γa for all n ≥ 0) ≥ 1 − a,
and
1
lim log |Γa ∩ Sn | = ĥ.
n→∞ n
In particular, if ĥ = h/l < v, then the random walk stays in an exponentially small
region Γa of logarithmic volume growth ĥ with positive probability, and moreover,
by Theorem 6.1, there is no region of logarithmic volume growth strictly less than
ĥ where the random walk stays with positive probability. The confinement of the
random walk occurs if and only if the inequality h < lv is strict as it is also mentioned
in [Ver, p.683].
Proof of Theorem 6.2. By Theorem 2.2, P-a.s., there exists a geodesic ray γω starting
at 1 such that (1/n)d (γω (⌊ln⌋), xn ) → 0 as n tends to infinity, where l > 0 is the
drift of the random walk. Take the map ω 7→ γω as a Borel measurable map as in
Theorem 2.2. As (14) in the proof of Theorem 5.2, we have
1
− log G(1, γω (⌊ln⌋)) → ĥ,
⌊ln⌋
as n tends to infinity, P-a.s. According to the Egorov theorem, one can make
the above two convergences uniform on an event with arbitrary high probability.
Namely, for every a ∈ (0, 1) there exists an event Aa ⊂ Ω such that P(Aa ) ≥ 1 − a
and for some sequence δn decreasing to 0, for all ω ∈ Aa ,
1
− log G(1, γω (⌊ln⌋)) − ĥ ≤ δ⌊ln⌋ and d (γω (⌊ln⌋), xn ) ≤ ⌊ln⌋δ⌊ln⌋ . (20)
⌊ln⌋
n o
Let Va := x ∈ Γ | G(1, x) ≥ e−|x|(ĥ+δ|x| ) . Define Γa := x∈Va B x, |x|δ|x| . For
S
every ω ∈ Aa and for every n ≥ 0, the point γω (⌊ln⌋) is in Va , and thus, for every
ω ∈ Aa , it holds that xn ∈ Γa for all n ≥ 0 by (20). Hence
P (xn ∈ Γa for all n ≥ 0) ≥ 1 − a.
We deduce that |Va ∩ Sn | ≤ Cen(ĥ+δn ) since x∈Sn G(1, x) ≍ 1 by Proposition

P
3.8. Note that for all small enough t > 0 and for all large enough n, we have
|B(1, nδn )| ≤ ent . Therefore for all large enough n,
X n
|Γa ∩ Sn | ≤ Cekt ek(ĥ+t) ≤ Ce 1−t (ĥ+2t) ,
n n
1+t
≤k≤ 1−t
and it follows that

1 ĥ + 2t
lim sup log |Γa ∩ Sn | ≤ .
n→∞ n 1−t
The above holds for all small enough t > 0, then we obtain the upper bound by ĥ.
The lower bound follows by |Γa ∩ [Sn ]| ≥ Kn (1 − a) and by Theorem 6.1 (3).
30 RYOKICHI TANAKA
Acknowledgements. The author would like to thank Sébastien Gouëzel for

pointing out some errors in the first draft, numerous valuable suggestions and com-
ments to improve the results in this paper, and also for sending him a preliminary
version of the paper [GMM2], Vadim A. Kaimanovich for his suggestion on a possible
connection to multifractal analysis, Pierre Mathieu for helpful comments and discus-
sion about the present results, Artëm Sapozhnikov for discussion on random walks
on hyperbolic groups, Takefumi Kondo for helpful comments, Jérémie Brieussel for
carefully reading an earlier version of this paper and for helpful comments to im-
prove the presentation, and Yuval Peres for helpful conversation and for introducing
to him many fractal people in Jerusalem.
References
[Anc] Ancona A., Positive harmonic functions and hyperbolicity, Potential Theory, Prague 1987,
Lecture Notes in Mathematics, vol. 1344, Springer, Berlin, 1988, 1-23.
[Ave] Avez A., Entropie des groupes de type fini, C. R. Acad. Sci. Paris Sér. A 275, 1363-1366
(1972).
[BS] Barreira L., Schmeling J., Sets of “non-typical” points have full topological entropy and full
Hausdorff dimension, Israel J. Math., 116, 29-70 (2000).
[BHM1] Blachère S., Haissinsky P., Mathieu P., Asymptotic entropy and Green speed for random
walks on countable groups, Ann. Prob., Vol. 36, No.3, 1134-1152 (2008).
[BHM2] Blachère S., Haissinsky P., Mathieu P., Harmonic measures versus quasiconformal mea-
sures for hyperbolic groups, Ann. Sci. Éc. Norm. Supér. (4) 44, 683-721 (2011).
[Bou] Bourgain J., Finitely supported measures on SL2 (R) which are absolutely continuous at
infinity, Geometric aspect of functional analysis (B. Klartag et al. eds.), 133-141, Lecture
Notes in Math. 2050, Springer-Verlag, Berlin Heidelberg, 2012.
[Bow] Bowen R., Equilibrium states and the ergodic theory of Anosov diffeomorphisms, Lecture
Notes in Mathematics 470, Second revised edition, Springer-Verlag Berlin Heidelberg, 2008.
[BT] Brieussel J., Tanaka R., Discrete random walks on the group Sol, arXiv:1306.6180v1
[math.PR] (2013), to appear in Israel J. Math.
[Cal] Calegari D., The ergodic theory of hyperbolic groups, Geometry and Topology Down Under,
Contemp. Math. 597, Amer. Math. Soc., Providence, RI, 2013, 15-52.
[CF] Calegari D., Fujiwara K., Combable functions, quasimorphisms, and the central limit theorem,
Ergodic Theory Dynam. Systems, 30, 1343-1369 (2009).
[Can] Cannon J W., The combinatorial structure of cocompact discrete hyperbolic groups, Geom.
Dedicata 16, 123-148 (1984).
[CM1] Connell C., Muchnik R., Harmonicity of quasiconformal measures and Poisson boundaries
of hyperbolic spaces, Geom. Funct. Anal. Vol. 17, 707-769 (2007).
[CM2] Connell C., Muchnik R., Harmonicity of Gibbs measures, Duke Math. J. Vol. 137, No. 3,
461-509 (2007).
[Coo] Coornaert M., Mesures de Patterson-Sullivan sur le bord d’un espace hyperbolique an sens
de Gromov, Pacific J. Math., Vol 159, No 2, 241-270 (1993).
[Der] Derriennic Y., Quelques applications du théorème ergodique sous-additif, Conference on Ran-
dom Walks (Kleebach, 1979), 183-201, Astérisque, 74, Soc. Math. France, Paris, 1980.
[Fal] Falconer K., Fractal Geometry, Mathematical Foundations and Applications, Third Edition,
John Wiley & Sons, Chichester 2014.
[Fen] Feng D-J., Gibbs properties of self-conformal measures and the multi fractal formalism, Er-
godic Theory Dynam. Systems 27, no.3, 787-812 (2007).
[GdlH] Ghys E., de la Harpe P., Sur les groupes hyperboliques d’après Mikhael Gromov, Progress
in Mathematics, Vol. 83, Birkhäuser, Boston 1990.
[Gou1] Gouëzel S., Local limit theorem for symmetric random walks in Gromov-hyperbolic groups,
J. Amer. Math. Soc. Vol 27, No 3, 893-928 (2014).
[Gou2] Gouëzel S., Martin boundary of measures with infinite support in hyperbolic groups,
preprint, arXiv:1302.5388v1 [math.PR] (2013).
[Gou3] Gouëzel S., Private communication, 2014.
[GL] Gouëzel S., Lalley S P., Random walks on co-compact Fuchsian groups, Ann. Sci. Éc. Norm.
Supér., (4) 46 no. 1, 129-173 (2013).
[GMM1] Gouëzel S., Mathéus F., Maucourant F., Sharp lower bounds for the asymptotic entropy
of symmetric random walks, preprint, arXiv:1209.3378v3 [math.PR] (2014).
[GMM2] Gouëzel S., Mathéus F., Maucourant F., Entropy and drift in word hyperbolic groups,
arXiv:1501.05082v1 [math.PR] (2015).
[Gui] Guivarc’h Y., Sur la loi des grands nombres et le rayon spectral d’une marche aléatoire,
Conference on Random Walks (Kleebach, 1979), 47-98, Astérisque, 74, Soc. Math. France,
Paris, 1980.
[Hai] Haissinsky P., Marches aléatoires sur les groupes hyperboliques, Géométrie Ergodique
(Dal’Bo-Milonet, ed.), Monographie de L’Enseignement Mathématique, 43, 199-265 (2013).
[HMM] Haissinsky P., Mathieu P., Müller S., Renewal theory for random walks on surface groups,
arXiv:1304.7625v1 [math.PR] (2013).
[Hei] Heinonen J., Lectures on Analysis on Metric Spaces, Springer-Verlag, New York 2001.
[INO] Izumi M., Neshveyev S., Okayasu R., The ratio set of the harmonic measure of a random
walk on a hyperbolic group, Israel J. Math. 163, 285-316 (2008).
[Kai1] Kaimanovich V A., Hausdorff dimension of the harmonic measure on trees, Ergodic Theory
Dynam. Systems, 18, 631-660 (1998).
[Kai2] Kaimanovich V A., The Poisson formula for groups with hyperbolic properties, Ann. Math.
Second series, Vol. 152, No. 3, 659-692 (2000).
[KL] Kaimanovich V A., Le Prince V.,Matrix random products with singular harmonic measure,
Geom. Dedicata 150 (2011) 257-279.
[KV] Kaimanovich V A., Vershik A M., Random walks on discrete groups: boundary and entropy,
Ann. Probab, Vol. 11, No. 3, 457-490 (1983).
[Led] Ledrappier F., Some asymptotic properties of random walks on free groups, Topics in proba-
bility and Lie groups: boundary theory, 117-152 , CRM Proc. Lecture Notes, 28, Amer. Math.
Soc., Providence, RI, 2001.
[LeP] Le Prince V., Dimensional properties of the harmonic measure for a random walk on a
hyperbolic group, Trans. Amer. Math. Soc., Vol. 359, No. 6, 2881-2898 (2007).
[Lyo] Lyons R., Equivalence of boundary measures on covering trees of finite graphs, Ergodic The-
ory Dynam. Systems, 14, 575-597 (1994).
[LP] Lyons R., Peres Y., Probability on Trees and Networks, Cambridge University Press. In prepa-
ration. Current version available at http://mypage.iu.edu/˜rdlyons/
[LPP] Lyons R., Pemantle R., Peres Y., Ergodic theory on Galton-Watson trees: speed of random
walk and dimension of harmonic measure, Ergodic Theory Dynam. Systems, 15, 593-619
(1995).
[PP] Parry W., Pollicott M., Zeta functions and the periodic orbit structure of hyperbolic dynamics,
Astérisque, 187-188, Société Mathématique de France (1990).
[Pat] Patterson S J., The limit set of a Fuchsian group, Acta Math., 136, 241-273 (1976).
[Pes] Pesin Y B., Dimension Theory in Dynamical Systems, The University of Chicago Press,
Chicago and London (1997).
[PU] Przytycki F., Urbański M., Conformal Fractals: Ergodic Theory Methods, London Mathe-
matical Society Lecture Note Series 371, Cambridge University Press, Cambridge 2010.
[Ver] Vershik A M., Dynamic theory of growth in groups: Entropy, boundaries, examples, Russian
Math. Surveys, 55:4, 667-733 (2000).
[Woe] Woess W., Random walks on infinite graphs and groups, Cambridge Tracts in Mathematics,
138, Cambridge University Press, Cambridge, 2000.
Tohoku University, 2-1-1 Katahira, Aoba-ku, 980-8577 Sendai, Japan

E-mail address: rtanaka@m.tohoku.ac.jp

1411 2312

Uploaded by

Copyright:

Available Formats

1411 2312

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1411 2312

Uploaded by

Copyright:

Available Formats

HAUSDORFF SPECTRUM OF HARMONIC MEASURE

Let Γ be a finitely generated group. For a probability measure µ on it, we obtain

Notation 1.4. Throughout this article, we use C, C ′ , . . . , to denote absolute con-

where the supremum is taken over all sequences {xn }∞ ∞

for every such sequence by the δ-hyperbolicity. The space Γ ∪ ∂Γ is compact,

2.2. Random walks on groups. For a countable group Γ, consider a probability

be Borel measurable by choosing the lexicographically minimum geodesic ray. More

The harmonic measure ν is a unique stationary measure on the boundary ∂Γ,

2.3. Martin boundary. The Green function is factorized as

For an extension to probability measures µ of super-exponential tails, see [Gou2].

exists, where Sn is the set of words of length n in Γ. Moreover, the function β is

The function β is convex on R, therefore it is continuous, and differentiable except

On the other hand,

Therefore, we obtain β(1) = 0.

Finally, let us show that β(θ) → −∞ (resp.+∞) as θ → +∞ (resp.−∞). For the

We introduce one parameter family of measures µθ on the boundary ∂Γ; it in-

as s ↓ v on the compactification Γ ∪ ∂Γ. The measure µ0 is supported on the

Theorem 3.3. For every θ ∈ R, there exists a probability measure µθ on ∂Γ such

Proof. The construction is analogous to the one of a Patterson-Sullivan measure on

left multiplication by g, the measure µθ,s has the form:

Proof. When θ = 0, a measure µ0 is nothing but a Patterson-Sullivan measure as we

The following is suggested by Gouëzel [Gou3]. In particular, probability measures

Proof. Let µθ be a probability measure satisfying (8). We note that µθ is doubling

Proposition 3.8. For every θ ∈ R,

4.1. Automatic structure. Suppose that S is a finite symmetric set of generators

α∗ is extended from Σ to Γ ∪ ∂Γ in this way. We note that the map α∗ is continuous,

Proof. This follows from dε (ξ, η) ≤ Ce−ε(ξ|η) .

Our analysis will be built on the structure of the graph A. A component is a

To each component C (not necessarily maximal), we associate the space of infinite

where D1 ∨ D2 denotes the partition refined by the partitions D1 and D2 , and

In particular, P r(ϕ) is attained by the Gibbs measure corresponding to a maximal

where µi = p1i pj=0

and the lemma follows.

Let A = (V, E, s∗ ) be a strongly Markov automaton associated with a non-

Thermodynamic formalism gives some additional information about a probability

The ergodicity of µθ provides the differentiability of β at θ ∈ R.

For every finitely supported admissible probability measure µ on a non-elementary

5.1. Hausdorff spectrum of the harmonic measure. Let µ be an arbitrary

Proof. Recall that β is continuously differentiable on R by Corollary 4.11. For each

where we have θ = −∞ or +∞ when the line is asymptotic to β(θ). Here the

Proof. Define for x, y in Γ,

extended on ∂ 2 Γ and is Hölder continuous on each compact subset. The measure

The logarithmic volume growth v of (Γ, d) has an interpretation as the Hausdorff

Proof of Theorem 5.8. Since the inequality h ≤ lv is equivalent to the inequality

since HD ≍ µ0 and ν = µ1 (Corollary 3.4). For every ω ∈ ΣK , we have

Lnϕ 1[E∗ ] (∅) ≥ CK e(a+v)n .

Theorem 6.1. Let ĥ := h/l.

An = {x ∈ [Sn ] | ν[n] (x) ≤ e−n(s−t) , F (1, x) ≥ e−n(ĥ+t) }.

We deduce that |Va ∩ Sn | ≤ Cen(ĥ+δn ) since x∈Sn G(1, x) ≍ 1 by Proposition

and it follows that

Acknowledgements. The author would like to thank Sébastien Gouëzel for

Tohoku University, 2-1-1 Katahira, Aoba-ku, 980-8577 Sendai, Japan

You might also like