You are on page 1of 108

Resistance forms, quasisymmetric maps and heat

kernel estimates
Jun Kigami
Author address:
Graduate School of Informatics, Kyoto University, Kyoto 606-
8501, Japan
E-mail address: kigami@i.kyoto-u.ac.jp
Contents
1. Introduction 1
Part 1. Resistance forms and heat kernels 7
2. Topology associated with a subspace of functions 8
3. Basics on resistance forms 10
4. the Green function 14
5. Topologies associated with resistance forms 17
6. Regularity of resistance forms 21
7. Annulus comparable condition and local property 22
8. Trace of resistance form 25
9. Resistance forms as Dirichlet forms 28
10. Transition density 30
Part 2. Quasisymmetric metrics and volume doubling measures 39
11. Semi-quasisymmetric metrics 40
12. Quasisymmetric metrics 43
13. Relations of measures and metrics 45
14. Construction of quasisymmetric metrics 50
Part 3. Volume doubling measures and heat kernel estimates 55
15. Main results on heat kernel estimates 56
16. Example: the -stable process on R 61
17. Basic tools in heat kernel estimates 64
18. Proof of Theorem 15.6 68
19. Proof of Theorems 15.10, 15.11 and 15.13 71
Part 4. Random Sierpinski gaskets 75
20. Generalized Sierpinski gasket 76
21. Random Sierpinski gasket 81
22. Resistance forms on Random Sierpinski gaskets 83
23. Volume doubling property 87
24. Homogeneous case 92
25. Introducing randomness 97
Bibliography 99
Assumptions, Conditions and Properties in Parentheses 101
List of Notations 102
Index 104
v
Abstract
Assume that there is some analytic structure, a dierential equation or a
stochastic process for example, on a metric space. To describe asymptotic behaviors
of analytic objects, the original metric of the space may not be the best one. Every
now and then one can construct a better metric which is somehow intrinsic with
respect to the analytic structure and under which asymptotic behaviors of the
analytic objects have nice expressions. The problem is when and how one can nd
such a metric.
In this paper, we consider the above problem in the case of stochastic processes
associated with Dirichlet forms derived from resistance forms. Our main concerns
are following two problems:
(I) When and how can we nd a metric which is suitable for describing asymptotic
behaviors of the heat kernels associated with such processes?
(II) What kind of requirement for jumps of a process is necessary to ensure good
asymptotic behaviors of the heat kernels associated with such processes?
Note that in general stochastic processes associated with Dirichlet forms have
jumps, i. e. paths of such processes are not continuous.
The answer to (I) is for measures to have volume doubling property with respect
to the resistance metric associated with a resistance form. Under volume doubling
property, a new metric which is quasisymmetric with respect to the resistance
metric is constructed and the Li-Yau type diagonal sub-Gaussian estimate of the
heat kernel associated with the process using the new metric is shown.
About the question (II), we will propose a condition called annulus comparable
condition, (ACC) for short. This condition is shown to be equivalent to the existence
of a good diagonal heat kernel estimate.
As an application, asymptotic behaviors of the traces of 1-dimensional -stable
processes are obtained.
In the course of discussion, considerable numbers of pages are spent on the
theory of resistance forms and quasisymmetric maps.
1991 Mathematics Subject Classication. Primary 30L10, 31E05, 60J35; Secondary 28A80,
43A99, 60G52.
Key words and phrases. resistance form, Green function, quasisymmetric map, volume dou-
bling property, jump process, heat kernel.
vi
1. INTRODUCTION 1
1. Introduction
Originally, the main purpose of this paper is to give answers to the following two
questions on heat kernels associated with Dirichlet forms derived from resistance
forms. Such Dirichlet forms roughly correspond to Hunt processes for which every
point has positive capacity.
(I) When and how can we nd metrics which are suitable for describing asymptotic
behaviors of heat kernels?
(II) What kind of requirement for jumps of processes and/or Dirichlet forms is
necessary to ensure good asymptotic behaviors of associated heat kernels?
Eventually we are going to make these questions more precise. For the moment,
let us explain what a heat kernel is. Assume that we have a regular Dirichlet form
(E, D) on L
2
(X, ), where X is a metric space, is a Borel regular measure on X, E
is a nonnegative closed symmetric form on L
2
(X, ) and D is the domain of E. Let
L be the Laplacian associated with this Dirichlet form, i.e. Lv is characterized
by the unique element in L
2
(X, ) which satises
E(u, v) =
_
X
u(Lv)d
for any u F. A nonnegative measurable function p(t, x, y) on (0, +) X
2
is
called a heat kernel associated with the Dirichlet form (E, D) on L
2
(X, ) if
u(t, x) =
_
X
p(t, x, y)u(y)(dy)
for any (t, x, y) (0, +) X
2
and any initial value u L
2
(X, ), where u(t, x) is
the solution of the heat equation associated with the Laplacian L:
u
t
= Lu.
The heat kernel may not exist in general. However, it is know to exist in many
cases like the Brownian motions on Euclidean spaces, Riemannian manifolds and
certain classes of fractals.
If the Dirichlet form (E, D) has the local property, in other words, the corre-
sponding stochastic process is a diusion, then one of the preferable goals on an
asymptotic estimate of a heat kernel is to show the so-called Li-Yau type (sub-
)Gaussian estimate, which is
(1.1) p(t, x, y)
c
1
V
d
(x, t
1/
)
exp
_
c
2
_
d(x, y)

t
_
1/(1)
_
,
where d is a metric on X, V
d
(x, r) is the volume of a ball B
d
(x, r) = {y|d(x, y) < r}
and 2 is a constant. It is well-known that the heat kernel of the Brownian
motion on R
n
is Gaussian which is a special case of (1.1) with d(x, y) = |xy|, = 2
and V
d
(x, r) = r
n
. Li and Yau have shown in [42] that, for a complete Riemannian
manifold with non-negative Ricci curvature, (1.1) holds with = 2, where d is the
geodesic metric and V
d
(x, r) is the Riemannian volume. In this case, (1.1) is called
the Li-Yau type Gaussian estimate. Note that V
d
(x, t
1/
) may have inhomogeneity
with respect to x in this case. For fractals, Barlow and Perkins have shown in [9]
that the Brownian motion on the Sierpinski gasket satises sub-Gaussian estimate,
that is, (1.1) with d(x, y) = |x y|, = log 5/ log 2 and V
d
(x, r) = r

, where
2
= log 3/ log 2 is the Hausdor dimension of the Sierpinski gasket. Note that
V
d
(x, r) is homogeneous in this particular case. Full generality of (1.1) is realized,
for example, by a certain time change of the Brownian motion on [0, 1], whose heat
kernel satises (1.1) with > 2 and inhomogeneous V
d
(x, r). See [38] for details.
There have been extensive studies on the conditions which are equivalent to
(1.1). For Riemannian manifolds, Grigoryan [23] and Salo-Coste [47] have in-
dependently shown that the Li-Yau type Gaussian estimate is equivalent to the
Poincare inequality and the volume doubling property. For random walks on
weighted graphs, Grigoryan and Telcs have obtained several equivalent conditions
for general Li-Yau type sub-Gaussian estimate, for example, the combination of
the volume doubling property, the elliptic Harnack inequality and the Poincare in-
equality in [25, 26]. Similar results have been obtained for diusions. See [31] and
[10] for example.
The importance of the Li-Yau type (sub-)Gaussian estimate (1.1) is that it
describes asymptotic behaviors of an analytical object, namely, p(t, x, y) in terms
of geometrical objects like the metric d and the volume of a ball V
d
(x, r). Such
an interplay of analysis and geometry makes the study of heat kernels interesting.
In this paper, we have resistance forms on the side of analysis and quasisymmetric
maps on the side of geometry. To establish a foundation in studying heat kernel
estimates, we rst need to do considerable works on both sides, i.e. resistance forms
and quasisymmetric maps. Those two subjects come to the other main parts of this
paper as a consequence.
The theory of resistance forms has been developed to study analysis on low-
dimensional fractals including the Sierpinski gasket, the 2-dimensional Sierpinski
carpet, random Sierpinski gaskets and so on. Roughly, a symmetric non-negative
denite quadratic form E on a subspace F of real-valued functions on a set X is
called a resistance form on X if it has the Markov property and
min{E(u, u)|u F, u(x) = 1 and u(y) = 0}
exists and is positive for any x = y X. The reciprocal of the above minimum,
denoted by R(x, y), is known to be a metric (distance) and is called the resistance
metric associated with (E, F). See [36] for details. Note that unlike the Dirichlet
forms, a resistance form is dened without referring to any measure on the space
and hence it is not necessarily a Dirichlet form as it is. In Part 1, we are going to
establish fundamental notions on resistance forms, for instance, the existence and
properties of the Green function with an innite set as a boundary, regularity of
a resistance form, traces, construction of a Dirichlet form from a resistance form,
the existence and continuity of heat kernels. Assume that (X, R) is compact for
simplicity and let be a Borel regular measure on (X, R). In Section 9, a regular
resistance form (E, F) is shown to be a regular Dirichlet form on L
2
(X, ). We
also prove that the associated heat kernel p(t, x, y) exists and is continuous on
(0, +) X X in Section 10. Even if (X, R) is not compact, we are going to
obtain a modied versions of those statements under mild assumptions. See Part I
for details.
The notion of quasisymmetric maps has been introduced by Tukia and Vaisala
in [51] as a generalization of quasiconformal mappings in the complex plane. Soon
its importance has been recognized in wide areas of analysis and geometry. There
have been many works on quasisymmetric maps since then. See Heinonen [32] and
Semmes [48] for references. In this paper, we are going to modify the resistance
1. INTRODUCTION 3
metric R quasisymmetrically to obtain a new metric which is more suitable for
describing asymptotic behaviors of a heat kernel. The key of modication is to
realize the following relation:
(1.2) Resistance Volume (Distance)

,
where Volume is the volume of a ball and Distance is the distance with respect
to the new metric. With (1.5), we are going to show that the mean exit time from
a ball is comparable with r

and this fact will lead us to Li-Yau type on-diagonal


estimate of the heat kernel described in (1.6). Quasisymmetric modication of
a metric has many advantages. For example, it preserves the volume doubling
property of a measure. In Part 2, we will study quasisymmetric homeomorphisms on
a metric space. In particular, we are going to establish relations between properties
such as (1.2) concerning the original metric D, the quasisymmetrically modied
metric d and the volume of a ball V
d
(x, r) = (B
d
(x, r)) and show how to construct
a metric d which is quasisymmetric to the original metric D and satisfy a desired
property like (1.2).
Let us return to question (I). We will conne ourselves to the case of diusion
processes for simplicity. The lower part of the Li-Yau type (sub-)Gaussian estimate
(1.1) is known to hold only when the distance is geodesic, i.e. any two points are
connected by a geodesic curve. This is not the case for most of metric spaces. So,
we use an adequate substitute called near diagonal lower estimate, (NDL)
,d
for
short. We say that (NDL)
,d
holds if and only if
(1.3)
c
3
V
d
(x, t
1/
)
p(t, x, y)
for any x, y which satisfy d(x, y)

c
4
t. For upper estimates, the Li-Yau type
(sub-)Gaussian upper estimate of order , (LYU)
,d
for short, is said to hold if and
only if
(1.4) p(t, x, y)
c
5
V
d
(x, t
1/
)
exp
_
c
6
_
d(x, y)

t
_
1/(1)
_
.
Another important property is the doubling property of a heat kernel, (KD) for
short, that is,
(1.5) p(t, x, x) c
7
p(2t, x, x).
Note that p(t, x, x) is monotonically decreasing with respect to t. It is easy to see
that the Li-Yau type (sub-)Gaussian heat kernel estimate together with the volume
doubling property implies (KD). Let p(t, x, y) be the heat kernel associated with a
diusion process. Now, the question (I) can be rephrased as follows:
Question When and how can we nd a metric d under which p(t, x, y) satises
(LYU)
,d
, (NDL)
,d
and (KD)?
In Corollary 15.12, we are going to answer this if the Dirichlet form associated
with the diusion process is derived from a resistance form. Roughly speaking, we
obtain the following statement.
Answer The underlying measure has the volume doubling property with respect
to the resistance metric R if and only if there exist > 1 and a metric d which is
quasisymmetric with respect to R such that (LYU)
,d
, (NDL)
,d
and (KD) hold.
4
Of course, one can ask the same question for general diusion process with a heat
kernel. Such a problem is very interesting. In this paper, however, we only consider
the case where the process is associated with a Dirichlet form induced by a resistance
form.
Next, we are going to explain the second problem, the question (II). Recently,
there have been many results on asymptotic behaviors of a heat kernel associated
with a jump process. See [12, 15, 16, 5] for example. They have dealt with a
specic class of jump processes and studied a set of conditions which is equivalent
to certain kind of (o-diagonal) heat kernel estimate. For example, in [15], they
have shown the existence of jointly continuous heat kernel for a generalization of
-stable process on an Ahlfors regular set and given a condition for best possible
o-diagonal heat kernel estimate. In this paper, we will only consider the following
Li-Yau type on-diagonal estimate, (LYD)
,d
for short,
(1.6) p(t, x, x)
1
V
d
(x, t
1/
)
which is the diagonal part of (1.1). Our question is
Question When and how can we nd a metric d with (LYD)
,d
for a given (jump)
process which possesses a heat kernel?
In this case, the when part of the question includes the study of the requirement
on jumps. In this paper, again we conne ourselves to the case where Dirichlet
forms are derived from resistance forms. Our proposal for a condition on jumps is
the annulus comparable condition, (ACC) for short, which says that the resistance
between a point and the complement of a ball is comparable with the resistance
between a point and an annulus. More exactly, (ACC) is formulated as
(1.7) R(x, B
R
(x, r)
c
) R(x, A
R
(x, r, (1 +)r))
on (x, r) X (0, +) for some > 0, where R is a resistance metric, B
R
(x, r)
is a resistance ball and A
R
(x, r, (1 +)r) = B
R
(x, (1 +)r)\B
R
(x, r) is an annulus.
(B
R
(x, s) is the closure of a ball B
R
(x, r) with respect to the resistance metric.) If
the process in question has no jump, i.e. is a diusion process, then the quantities
in the both sides of (1.7) coincide and hence (ACC) holds. As our answer to the
above question, we obtain the following statement in Theorem 15.11:
Theorem 1.1. The following three conditions are equivalent:
(C1) The underlying measure has the volume doubling property with respect to
R and (ACC) holds.
(C2) The underlying measure has the volume doubling property with respect to
R and the so-called Einstein relation:
Resistance Volume Mean exit time
holds for the resistance metric.
(C3) (ACC) and (KD) is satised and there exist > 1 and a metric d which is
quasisymmetric with respect to R such that (LYD)
,d
holds.
See [26, 50] on the Einstein relation, which is known to be implied by the
Li-Yau type (sub-)Gaussian heat kernel estimate.
Our work on heat kernel estimates is largely inspired by the previous two papers
[6] and [41]. In [6], the strongly recurrent random walk on an innite graph has
1. INTRODUCTION 5
been studied by using two dierent metrics, one is the shortest path metric d on
the graph and the other is the resistance metric R. In [6], the condition R(), that
is,
R() R(x, y)V
d
(x, d(x, y)) d(x, y)

has been shown to be essentially equivalent to the random walk version of (1.1).
Note the resemblance between (1.2) and R(). The metric d is however xed in
their case. In [41], Kumagai has studied the (strongly recurrent) diusion process
associated with a resistance form using the resistance metric R. He has shown
that the uniform volume doubling property with respect to R is equivalent to the
combination of natural extensions of (LYU)
,d
and (NDL)
,d
with respect to R. See
the remark after Theorem 15.10 for details. Examining those results carefully from
geometrical view point, we have realized that quasisymmetric change of metrics
(implicitly) plays an important role. In this respect, this paper can be though of
an extension of those works.
There is another closely related work. In [39], a problem which is very similar
to our question (I) has been investigated for a heat kernel associated with a self-
similar Dirichlet form on a self-similar set. The result in [39] is also quite similar
to ours. It has been shown that the volume doubling property of the underlying
measure is equivalent to the existence of a metric with (LYD)
,d
. Note that the
results in [39] include higher dimensional Sierpinski carpets where the self-similar
Dirichlet forms are not induced by resistance forms. The processes studied in [39],
however, have been all diusions
Finally, we present an application of our results to an -stable process on R for
(1, 2]. Dene
E
()
(u, u) =
_
R
2
(u(x) u(y))
2
|x y|
1+
dxdy
and F
()
= {u|E
()
(u, u) < +} for (1, 2) and (E
(2)
, F
(2)
) is the ordinary
Dirichlet form associated with the Brownian motion on R. Then (E
()
, F
()
) is a
resistance form for (1, 2] and the associated resistance metric is c|x y|
1
.
If = 2, then the corresponding process is not a diusion but has jumps. Let
p
()
(t, x, y) be the associated heat kernel. (We will show the existence of the heat
kernel p
()
(t, x, y) in Section 16.) It is well known that p
()
(t, x, x) = ct
1/
. Let
(E
()
|
K
, F
()
|
K
) be the trace of (E
()
, F
()
) onto the ternary Cantor set K. Let
p
()
K
(t, x, y) be the heat kernel associated with the Dirichlet form on L
2
(K, ) in-
duced by (E
()
|
K
, F
()
|
K
), where is the normalized Hausdor measure of K. By
Theorem 15.13, we conrm that (ACC) holds and obtain
p
()
K
(t, x, x) t

,
where =
log 2
(1) log 3+log 2
. See Section 16 for details.
This paper consists of four parts. In Part 1, we will develop basic theory of
resistance forms regarding the Green function, trace of a form, regularity and heat
kernels. This part is the foundation of the discussion in Part 3. Part 2 is devoted
to studying quasisymmetric homeomorphisms. This is another foundation of the
discussion in Part 3. After preparing those basics, we will consider heat kernel
estimates in Part 3. Finally in Part 4, we consider estimates of heat kernels on
random Sierpinski gaskets as an application of the theorems in Part 3.
6
The followings are conventions in notations in this paper.
(1) Let f and g be functions with variables x
1
, . . . , x
n
. We use f g for any
(x
1
, . . . , x
n
) A if and only if there exist positive constants c
1
and c
2
such that
c
1
f(x
1
, . . . , x
n
) g(x
1
, . . . , x
n
) c
2
f(x
1
, . . . , x
n
)
for any (x
1
, . . . , x
n
) A.
(2) The lower case c (with or without a subscript) represents a constant which is
independent of the variables in question and may have dierent values from place
to place (even in the same line).
Part 1
Resistance forms and heat kernels
In this part, we will establish basics of resistance forms such as the Green
function, harmonic functions, traces and heat kernels. In the previous papers [36,
34, 37], we have established the notions of the Green function, harmonic functions
and traces if a boundary is a nite set. One of the main subjects is to extend those
results to the case where a boundary is an innite set. In fact, we should determine
what kind of an innite set can be regarded as a proper boundary in the rst place.
To do so, we introduce a new topology determined by the domain of a resistance
form and show that the closed set with respect to this new topology can be regarded
as a boundary. Moreover, we will establish the existence of jointly continuous heat
kernel associated with the Dirichlet form derived from a resistance form under
several mild assumptions, which do not include the ultracontractivity. Recall that
if the transition semigroup associated with a Dirichlet form is ultracontractive, then
there exists a (pointwise) integral kernel of the transition semigroup, which will give
a heat kernel after necessary modications. See [18, Lemma 2.1.2], [22, Lemma 3.2]
and [5, Theorem 2.1] for details.
The followings are basic notations used in this paper.
Notation. (1) For a set V , we dene (V ) = {f|f : V R}. If V is a
nite set, (V ) is considered to be equipped with the standard inner product (, )
V
dened by (u, v)
V
=

pV
u(p)v(p) for any u, v (V ). Also |u|
V
=
_
(u, u)
V
for
any u (V ).
(2) Let V be a nite set. The characteristic function
V
U
of a subset U V is
dened by

V
U
(q) =
_
1 if q U,
0 otherwise.
If no confusion can occur, we write
U
instead of
V
U
. If U = {p} for a point p V ,
we write
p
instead of
{p}
. If H : (V ) (V ) is a linear map, then we set
H
pq
= (H
q
)(p) for p, q V . Then (Hf)(p) =

qV
H
pq
f(q) for any f (V ).
(3) Let (X, d) be a metric space. Then
B
d
(x, r) = {y|y X, d(x, y) < r}
for x X and r > 0.
2. Topology associated with a subspace of functions
In this section, we will introduce an operation B B
F
from subsets of a space
X to itself associated with a linear subspace F of real valued functions (X). This
operation will turn out to be essential in describing whether a set can be treated
as a boundary or not. More precisely, we will show in Theorems 4.1 and 4.3 that
a set B can be a proper boundary if and only if B
F
= B. Also the importance of
the condition that B
F
= B is revealed in Theorem 6.3 where we have equivalent
conditions for regularity of resistance forms.
Definition 2.1. Let F be a linear subspace of (X) for a set X. For a subset
B X, dene
F(B) = {u|u F, u(x) = 0 for any x B}.
and
B
F
=

uF(B)
u
1
(0)
2. TOPOLOGY ASSOCIATED WITH A SUBSPACE OF FUNCTIONS 9
The following lemma is immediate from the denition.
Lemma 2.2. Let F be a linear subspace of (X) for a set X.
(1) For any B X, B B
F
, F(B) = F(B
F
) and (B
F
)
F
= B
F
.
(2) X
F
= X.
(3)
F
= if and only if {u(x)|u F} = R for any x X.
The above lemma suggests that the operation B B
F
satises the axiom of
closure and hence it denes a topology on X. Indeed, this is the case if F is stable
under the unit contraction.
Definition 2.3. (1) For u : X R, dene u : X [0, 1] by
u(p) =
_

_
1 if u(p) 1,
u(p) if 0 < u(p) < 1,
0 if u(p) 0.
u is called the unit contraction of u.
(2) Let F be a linear subspace of (X) for a set X. F is said to be stable under
the unit contraction if and only if u F for any u F.
In the case of Dirichlet forms, the condition that the domain is stable under
the unit contraction is one of the equivalent conditions of the Markov property. See
[21, Section 1.1] for details. (In [21], their terminology is that the unit contraction
operates on F in place of that F is stable under the unit contraction.)
Theorem 2.4. Let F be a linear subspace of (X) for a set X. Assume that
{u(x)|u F} = R for any x X and that F is stable under the unit contraction.
Dene
C
F
= {B|B X, B
F
= B}.
Then C
F
satises the axiom of closed sets and it denes a topology of X. Moreover,
the T
1
-axiom of separation holds for this topology, i.e. {x} is a closed set for any
x X, if and only if, for any x, y X with x = y, there exists u F such that
u(x) = u(y).
As a topology given by a family of real-valued function, the notion of ne
topology has been introduced in classical axiomatic potential theory by Brelot [13].
In the case of resistance forms, we will see that the topology given by C
F
coincides
with the ne topology associated with the cone of nonnegative functions in the
domain of resistance form in Theorem 5.7. Our proof of Theorem 5.7, however,
depends essentially on Theorem 4.3, where the condition B
F
= B has already
played a crucial role. So the coincidence of C
R
and the ne topology does give us
small help in studying resistance forms. See the comments after Theorem 5.7 for
details.
The rest of this section is devoted to proving the above theorem.
Lemma 2.5. Under the assumptions of Theorem 2.4, if B C
F
and x X\B,
then there exists u F such that u F(B), u(x) = 1 and 0 u(y) 1 for any
y X.
Proof. Since B
F
= B, there exists v F(B) such that v(x) = 0. Let
u = v/v(x). Then u satises the required properties.
10
Proof of Theorem 2.4. First we show that C
F
satises the axiom of closed
sets. Since F(X) = {0}, X
F
= X. Also we have
F
= by Lemma 2.2-(3). Let
B
i
C
F
for i = 1, 2 and let x (B
1
B
2
)
c
, where A
c
is the complement of A in
X, i.e. A
c
= X\A. By Lemma 2.5, there exists u
i
F(B
i
) such that u
i
= u
i
and
u
i
(x) = 1. Let v = u
1
+ u
2
1. Then v(x) = 1 and v(y) 0 for any y B
1
B
2
.
If u = v, then u F(B
1
B
2
) and u(x) = 1. Hence B
1
B
2
C
F
. Let B

C
F
for any . Set B =

. If x / B, then there exists

such that
x / B

. We have u F(B

) F(B) satisfying u(x) = 0. Hence x / B


F
. This
shows B C
F
. Thus C
F
satises the axiom of closed sets.
Next dene U
x,y
=
_
_
f(x)
f(y)
_

f F
_
. We will show that U
x,y
= R
2
if there
exists u F such that u(x) = u(y). Suppose that u(x) = 0. Considering u/u(x),
we see that
_
1
a
_
U
x,y
, where a = 1. Since there exists v F with v(y) = 0, it
follows that
_
b
1
_
U
x,y
for some b R. Now we have ve cases.
Case 1: Assume that a 0. Considering the operation u u, we have
_
1
0
_
U
x,y
.
Also
_
b
1
_
U
x,y
. Since U
x,y
is a linear subspace of R
2
, U
x,y
= R
2
.
Case 2: Assume that b 0. By similar arguments as Case 1, we have U
x,y
= R
2
.
Case 3: Assume that b 1. The u-operation shows that
_
1
1
_
U
x,y
. Since
_
_
1
1
_
,
_
1
a
_
_
is independent, U
x,y
= R
2
.
Case 4: Assume that a (0, 1) and b (0, 1). Then
_
_
1
a
_
,
_
b
1
_
_
is independent.
Hence U
x,y
= R
2
.
Case 5: Assume that a > 1 and b (0, 1). The u-operation shows
_
1
1
_
U
x,y
.
Then
_
_
1
1
_
,
_
b
1
_
_
is independent and hence U
x,y
= R
2
.
Thus U
x,y
= R
2
in all the cases. Exchanging x and y, we also deduce the same
conclusion even if u(x) = 0. In particular, the fact that U
x,y
= R
2
implies that
y / {x}
F
. Hence if there exists u F such that u(x) = u(y) for any x, y X with
x = y, then {x} C
F
for any x X. The converse direction is immediate.
3. Basics on resistance forms
In this section, we rst introduce denition and basics on resistance forms.
Definition 3.1 (Resistance form). Let X be a set. A pair (E, F) is called a
resistance form on X if it satises the following conditions (RF1) through (RF5).
(RF1) F is a linear subspace of (X) containing constants and E is a non-negative
symmetric quadratic form on F. E(u, u) = 0 if and only if u is constant on X.
(RF2) Let be an equivalent relation on F dened by u v if and only if u v
is constant on X. Then (F/, E) is a Hilbert space.
(RF3) If x = y, then there exists u F such that u(x) = u(y).
(RF4) For any p, q X,
sup
_
|u(p) u(q)|
2
E(u, u)

u F, E(u, u) > 0
_
is nite. The above supremum is denoted by R
(E,F)
(p, q).
(RF5) u F and E( u, u) E(u, u) for any u F, where u is dened in Deni-
tion 2.3.
Note that the denition of resistance forms does not require any measure on
the space X at all. Being combined with a measure , a resistance form may induce
3. BASICS ON RESISTANCE FORMS 11
a Dirichlet form on L
2
(X, ) and the associated process, semigroup and Laplacian.
See Section 9.
By (RF3) and (RF5) along with Theorem 2.4, the axiom of closed sets holds
for C
F
and the associated topology satises the T
1
-separation axiom.
Proposition 3.2. Assume that u F for any u F. Then (RF3) in the
above denition is equivalent to the following conditions:
(RF3-1) F
F
= F for any nite subset F X.
(RF3-2) For any nite subset F X and any v (F), there exists u F such
that u|
F
= v.
Proof. (RF3) (RF3-1) By Theorem 2.4, (RF3) implies that {x}
F
= {x}
for any x X. Let F be a nite subset of X. Again by Theorem 2.4, F
F
=
(
xF
{x})
F
=
xF
{x}
F
= F.
(RF3-1) (RF3-2) Let F be a nite subset of X. Set F
x
= F\{x} for x F.
Since (F
x
)
F
= F
x
, there exists u
x
F such that u
x
|
F
x
0 and u
x
(x) = 1. For any
v (F), dene u =

xF
v(x)u
x
. Then u|
F
= v and u F.
(RF3-2) (RF3) This is obvious.
Remark. In the previous literatures [36, 34, 37], (RF3-2) was employed as a
part of the denition of resistance forms in place of the current (RF3).
By the results in [36, Chapter 2], we have the following fact.
Proposition 3.3. Let (E, F) be a resistance form on X. The supremum in
(RF4) is the maximum and R
(E,F)
is a metric on X.
Definition 3.4. Let (E, F) be a resistance form on X. R
(E,F)
is called the
resistance metric on X associated with the resistance form (E, F).
One of the most simple examples of resistance forms is weighted graphs.
Example 3.5 (Weighted graph). Let V be an (innite) countable set and let
H = {H
xy
}
x,yX
satisfy the following three conditions (WG1), (WG2) and (WG3).
(WG1) H
xy
= H
yx
0 and H
xx
= 0 for any x, y X.
(WG2) N(x) = {y|H
xy
> 0} is a nite set for any x X.
(WG3) For any x, y X, there exist x
1
, . . . , x
n
X such that x
1
= x, x
n
= y
and H
x
i
x
i+1
> 0 for any i = 1, . . . , n 1.
Then (V, H) is called a (locally nite irreducible) weighted graph. (The condi-
tion (WG3) is called the locally niteness and the condition (WG4) is called the
irreducibility.) Dene
F
(V,H)
=
_
u

u (V ),

x,yV
H
xy
(u(x) u(y))
2
< +
_
and, for u, v F
(V,H)
,
E
(V,H)
(u, v) =
1
2

x,yV
H
xy
(u(x) u(y))(v(x) v(y)).
Then (E
(V,H)
, F
(V,H)
) is a resistance form on V . There exists a random walk on V
associated with the weight graph (V, H). Namely dene
x
=

yN(x)
H
xy
and
P(x, y) = H
xy
/
x
. We give the transition probability from x to y in the unit time
by P(x, y). This random walk is called the random walk associated with (V, H).
12
Relations between (E
(V,H)
, F
(V,H)
) and the random walk associated with (V, H)
have been one of the main subjects in the theory of random walks and discrete
potential theory. See [52, 53], [43] and [49] for example. In particular, as we
mentioned in the introduction, asymptotic behaviors of the heat kernel associated
with the random walk has been studied in [6].
Note that in the above example, the set V is countable. This is not the case for
general resistance forms. For example, we have the resistance forms on R associated
with -stable process dened in the introduction and the resistance forms on the
(random) Sierpinski gaskets in Part 4.
If no confusion can occur, we use R to denote R
(E,F)
. By (RF4), we immediately
obtain the following fact.
Proposition 3.6. Let (E, F) be a resistance form on X and let R be the as-
sociated resistance metric. For any x, y X and any u F,
(3.1) |u(x) u(y)|
2
R(x, y)E(u, u).
In particular, u F is continuous with respect to the resistance metric.
Next we introduce the notion of Laplacians on a nite set and harmonic func-
tions with a nite set as a boundary. See [36, Section 2.1] for details, in particular,
the proofs of Proposition 3.8 and 3.10.
Definition 3.7. Let V be a non-empty nite set. Recall that (V ) is equipped
with the standard inner-product (, )
V
. A symmetric linear operator L : (V )
(V ) is called a Laplacian on V if it satises the following three conditions:
(L1) L is non-positive denite,
(L2) Lu = 0 if and only if u is a constant on V ,
(L3) L
pq
0 for all p = q V .
We use LA(V ) to denote the collection of Laplacians on V .
By [36, Proposition 2.1.3], we have the next proposition, which says that a
resistance form on a nite set corresponds to a Laplacian.
Proposition 3.8. Let V be a non-empty nite set and let L be a symmetric
linear operator form (V ) to itself. Dene a symmetric bilinear form E
L
on (V )
by E
L
(u, v) = (u, Lv)
V
for any u, v (V ). Then, E
L
is a resistance form on V
if and only if L LA(V ).
Using the standard inner-product (, )
V
, we implicitly choose the uniform dis-
tribution, i.e. the sum of all the Dirac masses on the space, as our measure on V .
This is why we may relate Laplacians with resistance forms on a nite set without
mentioning any measure explicitly. (Recall the comment after Denition 3.1.)
Definition 3.9. Let V be a nite set and let L LA(V ). The resistance form
(E
L
, (V )) on V is called the resistance form associated with L.
The harmonic function with a nite set as a boundary is dened as the energy
minimizing function.
Proposition 3.10. [36, Lemma 2.3.5] Let (E, F) be a resistance form on X
and let V be a nite subset of X. Let (V ). Then there exists a unique u F
such that u|
V
= and u attains the following minimum:
min{E(v, v)|v F, v|
V
= }.
3. BASICS ON RESISTANCE FORMS 13
Moreover, the map from to u is a linear map from (V ) to F. Denote this map
by h
V
. Then there exists a Laplacian L LA(V ) such that
(3.2) E
L
(, ) = E(h
V
(), h
V
()).
In Lemma 8.2, we are going to extend this proposition when V is an innite
set.
Definition 3.11. h
V
() dened in Proposition 3.10 is called the V -harmonic
function with the boundary value . Also we denote the above L LA(V ) in (3.2)
by L
(E,F),V
.
To construct concrete examples of resistance forms, we often start from a se-
quence of resistance forms on nite sets which has certain compatibility as follows.
Definition 3.12. Let {V
m
}
m0
be a sequence of nite sets and let L
m

LA(V
m
) for m 0. {(V
m
, L
m
)}
m0
is called a compatible sequence if and only if
L
(E
L
m+1
,(V
m+1
)),V
m
= L
m
for any m 0, i.e.
E
L
m
(u, u) = min{E
L
m+1
(v, v)|v (V
m+1
), u = v|
V
m
}
for any u (V
m
).
Combining the results in [36, Sections 2.2 and 2.3], in particular [36, Theo-
rems 2.2.6 and 2.3.10], we obtain the following theorem on construction of a resis-
tance form from a compatible sequence.
Theorem 3.13. Let {V
m
}
m0
be a sequence of nite sets and let L
m
LA(V
m
)
for m 0. Assume that S = {(V
m
, L
m
)}
m0
is a compatible sequence. Let V

m0
V
m
. Dene
F
S
= {u|u (V

), lim
m+
E
L
m
(u|
V
m
, u|
V
m
) < +}
and
E
S
(u, v) = lim
m+
E
L
m
(u|
V
m
, v|
V
m
)
for any u, v F
S
. Then (E
S
, F
S
) is a resistance form on V

and L
(E
S
,F
S
),V
m
=
L
m
for any m 0. Moreover, let R
S
be the resistance metric associated with
(E
S
, F
S
) and let (X, R) be the completion of (V

, R
S
). Then there exists a unique
resistance form (E, F) on X such that, for any u F, u is a continuous function
on X, u|
V

F
S
and E(u, u) = E
S
(u|
V

, u|
V

). In particular, R coincides with the


resistance metric associated with (E, F).
We are going to use this theorem to construct resistance forms in Example 5.5
and Section 22.
In contrast with the above theorem, if (X, R) is separable, the resistance form
(E, F) is always expressed as a limit of a compatible sequence. More precisely, the
following fact has been shown in [36, Section 2.3].
Theorem 3.14. Let (E, F) be a resistance form on a set X and let R be the
associated resistance metric. Let {V
m
}
m1
be an increasing sequence of nite sub-
sets of X. Assume that V

=
m1
V
m
is dense in X. Set L
m
= L
(E,F),V
m
where
L
(E,F),V
m
is dened in Denition 3.11. Then {(V
m
, L
m
)}
m1
is a compatible se-
quence. Moreover,
F = {u|u C(X, R), lim
m+
E
L
m
(u|
V
m
, u|
V
m
) < +},
14
where C(X, R) is the collection of real-valued continuous functions with respect to
the resistance metric, and
E(u, v) = lim
m+
E
L
m
(u|
V
m
, v|
V
m
)
for any u, v F.
By using this theorem, we have the following fact, which is used in Section 5.
Proposition 3.15. Let (E, F) be a resistance form on a set X and let R be
the resistance metric. Assume that (X, R) is separable. For u : X R, dene u
+
:
R [0, +) by u
+
(x) = max{u(x), 0}. Then u
+
F and E(u
+
, u
+
) E(u, u) for
any u F.
Proof. Since (X, R) is separable, we may choose a sequence {V
m
}
m0
of sub-
sets of X so that V
m
V
m+1
for any m 0 and V

=
m0
V
m
is dense in (X, R).
Let L
m
as in Theorem 3.14. Then E
L
m
(u, u) =

x,yV
m
(L
m
)
xy
(u(x) u(y))
2
/2
for any u (V
m
). Since (u
+
(x) u
+
(y))
2
(u(x) u(y))
2
, it follows that
E
L
m
(u
+
|
V
m
, u
+
|
V
m
) E
L
m
(u|
V
m
.u|
V
m
). Hence by Theorem 3.14, we see that u
+

F and E(u
+
, u
+
) E(u, u).
4. the Green function
In this section, we study the Green function associated with an innite set as
a boundary. In the course of discussion, we will show that a set B is a suitable
boundary if and only if B
F
= B. Conditions ensuring B
F
= B are given in the
next section. For example, if B is compact with respect to the resistance metric,
then B
F
will be shown to coincide with B.
Throughout this section, (E, F) is a resistance form on a set X and R is the
associated resistance metric. The next theorem establishes the existence and basic
properties of the Green function with an innite set as a boundary.
Theorem 4.1. Let B X be non-empty. Then (E, F(B)) is a Hilbert space
and there exists a unique g
B
: X X R that satises the following condition
(GF1):
(GF1) Dene g
x
B
(y) = g
B
(x, y). For any x X, g
x
B
F(B) and E(g
x
B
, u) = u(x)
for any u F(B).
Moreover, g
B
satises the following properties (GF2), (GF3) and (GF4):
(GF2) g
B
(x, x) g
B
(x, y) = g
B
(y, x) 0 for any x, y X. g
B
(x, x) > 0 if and
only if x / B
F
.
(GF3) Dene R(x, B) = g
B
(x, x) for any x X. If x / B
F
, then
R(x, B) =
_
min{E(u, u)|u F(B), u(x) = 1}
_
1
= sup
_
|u(x)|
2
E(u, u)

u F(B), u(x) = 0
_
.
(GF4) For any x, y, z X, |g
B
(x, y) g
B
(x, z)| R(y, z).
By (GF2), if B = B
F
, then g
x
B
0 for any x B
F
\B. Such a set B is not a
good boundary.
We will prove this and the next theorem at the same time.
Definition 4.2. The function g
B
(, ) given in the above theorem is called the
Green function associated with the boundary B or the B-Green function.
4. THE GREEN FUNCTION 15
As we have remarked after Denition 3.1, no measure is required to dene a
resistance form. Thus the denition of the Green function is also independent of a
choice of measures. After introducing a measure and constructing the Dirichlet
form on L
2
(X, ) induced by a resistance form in Section 8, we may observe that the
Green function g
B
(x, y) dened above coincides with the Green function associated
with the Dirichlet form. See Theorem 10.10 and Corollary 10.11 for details. For
example, g
B
(x, y) is shown to be the integral kernel of the nonnegative self-adjoint
operator associated with the Dirichlet form.
The next theorem assures another advantage of being B = B
F
. Namely, if
B = B
F
, we may reduce B to a one point, consider the shorted resistance
form (E, F
B
) and obtain a expression of the Green function (4.1) by the shorted
resistance metric R
B
(, ). In the case of weighted graphs, such an identication of
a subset of domain is called Rayleighs shorting method. See [19] for details.
Theorem 4.3. Let (E, F) be a resistance form and let B X be non-empty.
Suppose that B
F
= B. Set
F
B
= {u|u F, u is a constant on B}
and X
B
= {B}(X\B). Then (E, F
B
) is a resistance form on X
B
. Furthermore,
if R
B
(, ) is the resistance metric associated with (E, F
B
), then
(4.1) g
B
(x, y) =
R
B
(x, B) +R
B
(y, B) R
B
(x, y)
2
for any x, y X. In particular, R(x, B) = R
B
(x, B) for any x X\B.
Remark. In [45, Section 3], V. Metz has shown (4.1) in the case where B is a
one point.
The proofs of the those two theorems are divided into several parts.
Note that by Proposition 5.1 B is closed with respect to R if B
F
= B.
Proof of the first half of Theorem 4.1. Let x B and let F(x) =
F({x}). By (RF2), (E, F(x)) is a Hilbert space. Note that F(B) F(x). If
{u
m
}
m1
is a Cauchy sequence in F(B), there exists the limit u F(x). For
y B,
|u
m
(y) u(y)|
2
R(x, y)E(u
m
u, u
m
u).
Letting m +, we see that u(y) = 0. Hence u F(B). This shows that
(E, F(B)) is a Hilbert space. For any z X and any u F(B), |u(z)|
2

R(x, y)E(u, u). The map u u(z) is continuous linear functional and hence there
exists a unique
z
F(B) such that E(
z
, u) = u(z) for any u F(B). De-
ne g
B
(z, w) =
z
(w). Since E(
z
,
w
) =
z
(w) =
w
(z), we have (GF1) and
g
B
(z, w) = g
B
(w, z). If z B
F
, then u(z) = 0 for any u F(B). Hence
g
B
(z, z) = g
z
B
(z) = 0. Conversely, assume g
B
(z, z) = 0. Since g
B
(z, z) = E(g
z
B
, g
z
B
),
(RF1) implies that g
z
B
is constant on X. On the other hand, g
z
B
(y) = 0 for any
y B. Hence g
z
B
0. For any u F(B), u(z) = E(g
z
B
, u) = 0. Therefore,
z B
F
.
Definition 4.4. Let B X be non-empty. If x / B
F
, we dene
x
B
=
g
x
B
/g
B
(x, x).
Note that g
B
(x, x) > 0 if and only if x / B
F
by the above proof. Hence
x
B
is
well-dened.
16
Lemma 4.5. Let B X be non-empty and let x / B
F
. Then
x
B
is the unique
element which attains the following minimum:
min{E(u, u)|u F(B), u(x) = 1}.
In particular, (GF3) holds.
Proof. Let u F(B) with u(x) = 1. Since
E(u
x
B
,
x
B
) =
E(u
x
B
, g
x
B
)
g
B
(x, x)
=
(u(x) 1)
g
B
(x, x)
= 0,
we have
E(u, u) = E(u
x
B
, u
x
B
) +E(
x
B
,
x
B
)
Hence E(u, u) E(
x
B
,
x
B
) and if the equality holds, then u =
x
B
. Now,
E(
x
B
,
x
B
) =
E(g
x
B
, g
x
B
)
g
B
(x, x)
2
=
1
g
B
(x, x)
.
This suces for (GB3).
Lemma 4.6. Let B X be non-empty. Then g
B
(x, x) g
B
(x, y) 0 for any
x, y X.
Proof. If x B
F
, then g
x
B
0. Otherwise, dene v =
x
B
. Then by (RF5),
E(
x
B
,
x
B
) E(v, v). The above lemma shows that
B
x
= v. Hence 0
x
B
1.
Combining Lemma 4.6 and the results from Proof of the rst half of Theo-
rem 4.1, we have (GF2).
So far, we have obtained (GF1), (GF2) and (GF3). Before showing (GF4), we
prove Theorem 4.3.
Proof of Theorem 4.3. (RF1), (RF2) and (RF5) are immediate by the def-
inition of F
B
. To show (RF3), let x and y X with x = y. We may assume y = B
without loss of generality. Set B
x
= B {x}. Since (B
x
)
F
= B
x
, there exists
u F(B
x
) such that u(y) = 0. Hence we obtain (RF3). To see (RF4), note that
sup
_
|u(x) u(y)|
2
E(u, u)

u F
B
, E(u, u) > 0
_
R
(E,F)
(x, y)
because F
B
F. Hence we have (RF4). To prove (4.1), it is enough to show the
case where B is a one point. Namely we will show that
(4.2) g
{z}
(x, y) =
R(x, z) +R(y, z) R(x, y)
2
for any x, y, z X. We write g(x, y) = g
{z}
(x, y). The denition of R(, ) along
with Lemma 4.5 shows that g(x, x) = R(x, z). Also by Lemma 4.5, if u

(y) =
g(x, y)/g(x, x), then u

is the {x, z}-harmonic function whose boundary values are


u

(z) = 0 and u

(x) = 1. Let V = {x, y, z}. Then by Proposition 3.10, there exists


a Laplacian L LA(V ) with (3.2). Note that
E
L
(u

|
V
, u

|
V
) = min{E
L
(v, v)|v (V ), v(x) = 1, v(z) = 0}.
Therefore, (Lu

)(y) = 0. Set L = (L
pq
)
p,qV
. Hereafter we assume that L
pq
> 0
for any p, q V with p = q. (If this condition fails, the proof is easier.) Let
R
pq
= (L
pq
)
1
. Solving Lu

(y) = 0, we have
(4.3) u

(y) =
L
xy
L
xy
+L
yz
=
R
yz
R
xy
+R
yz
.
5. TOPOLOGIES ASSOCIATED WITH RESISTANCE FORMS 17
Dene R
x
= R
xy
R
xz
/R

, R
y
= R
yx
R
yz
/R

and R
z
= R
zx
R
zy
/R

, where R

=
R
xy
+R
yz
+R
zx
. By the -Y transform, R(p, q) = R
p
+R
q
for any p and q with
p = q. (See [36, Lemma 2.1.15] for -Y transform.) Hence
(4.4)
R(x, z) +R(y, z) R(x, y)
2
= R
z
Since g(x, x) = R(x, z), (4.3) implies
g(x, y) = g(x, x)u

(y) = R(x, z)u

(y) =
R
xz
(R
xy
+R
yz
)
R

(y) = R
z
.
By (4.4), we have (4.2).
Proof of (GF4) of Theorem 4.1. Let K = B
F
. Note that g
B
(x, y) =
g
K
(x, y). By (4.1),
|g
B
(x, y) g
B
(x, z)|
|R(y, K) R(z, K)| +|R
K
(x, y) R
K
(x, z)|
2
R
K
(y, z) R(y, z).

5. Topologies associated with resistance forms


We continue to assume that (E, F) is a resistance form on a set X and R is
the associated resistance metric. In the previous section, we have seen that the
condition for a subset B X being a good boundary is that B
F
= B. Note that
by Theorem 2.4, C
F
gives a topology on X which satises T
1
-axiom of separation.
Lemma 2.2 implies that B
F
= B if and only if B is a closed set with respect to
the C
F
-topology. Therefore, to see when B
F
= B occurs is to consider the relation
between the topologies induced by the resistance metric R and C
F
. Furthermore,
there exists a classical notion of ne topology associated with a cone of non-
negative functions introduced by Brelot in [13]. In this section, we will study
relations of those topologies on X. The topologies induced by the resistance metric
and C
F
are called the R-topology and the C
F
-topology respectively.
First we show that the C
F
-topology is coarser than the R-topology in general.
Proposition 5.1. B
F
is a closed set with respect to the resistance metric R.
In other word, the C
F
-topology is coarser (i.e. weaker) than that given by the R-
topology.
Proof. Let {x
n
}
n1
B
F
. Assume lim
n+
R(x, x
n
) = 0. If u F(B),
then u(x) = lim
n+
u(x
n
) = 0 for any u F(B). Hence x B
F
.
Later in Example 5.5, we have a resistance form on {0, 1, . . .} where there exists
an R-closed set which is not C
F
-closed.
Next, we study a sucient condition ensuring that B
F
= B.
Definition 5.2. (1) Let (X, d) be a metric space. For a non-empty subset of
B X, dene
N
d
(B, r) = min{#(A)|A B
yA
B
d
(y, r)}
18
for any r > 0, where #(A) is the number of the elements of A.
(2) Let (E, F) be a resistance form on X and let R be the associated resistance
metric. For any subsets U, V X, dene
R(U, V ) = inf{R(x, y)|x U, y V }.
Hereafter in this section, we use N(B, r) to denote N
R
(B, r).
The following theorem plays an important role in proving heat kernel estimates
in Part III.
Theorem 5.3. Let (E, F) be a resistance form on X. Let B be a non-empty
subset of X and let x X\B. If N(B, R(x, B)/2) < +, then x / B
F
and
R(x, B)
4N(B, R(x, B)/2)
R(x, B) R(x, B).
The key idea of the following proof has been extracted from [6, Lemma 2.4]
and [41, Lemma 4.1].
Proof. Write u
y
=
x
{y}
for any x, y X. Then, by (GF4),
u
y
(z) = u
y
(z) u
y
(y)
R(y, z)
R(y, x)
.
If y B, x X\B and z B
R
(x, R(x, B)/2), then u
y
(z) 1/2. Suppose
that n = N(B, R(x, B)/2) is nite. We may choose y
1
, . . . , y
n
B so that
B
n
i=1
B
R
(y
i
, R(x, B)/2). Dene v(z) = min
i=1,...,n
u
y
i
(z) for any z X. Then
v F, v(x) = 1 and v(z) 1/2 for any z B. Letting h = 2(v 1/2), we see that
0 h(z) 1 for any z X, h(x) = 1 and h F(B). Hence x / B
F
. Moreover,
E(h, h) 4E(v, v) 4
n

i=1
E(u
y
i
, u
y
i
) 4
n

i=1
1
R(x, y
i
)

4n
R(x, B)
.
Therefore,
R(x, B) =
_
min{E(u, u)|u F(B), u(x) = 1}
_
1

R(x, B)
4n
.

Corollary 5.4. Let (E, F) be a resistance form on X. If B is compact with


respect to the resistance metric associated with (E, F), then B
F
= B.
In general, even if (X, R) is locally compact, B
F
= B may happen for an
R-closed set B as you can see in the next example.
Example 5.5. Let X = N{0} and let V
m
= {1, . . . , m}{0}. Dene a linear
operator L
m
: (V
m
) (V
m
) by
(L
m
)
ij
=
_

_
2 if |i j| = 1 or |i j| = m,
1 if {i, j} = {0, k} for some k {1, . . . , m}\{1, m},
4 if i = j and i {1, m},
5 if i = j and i {1, . . . , m}\{1, m},
(m+ 2) if i = j = 0,
0 otherwise.
5. TOPOLOGIES ASSOCIATED WITH RESISTANCE FORMS 19
1 2 3 1 2 3 4
0 0
2 2 2 2 2
2 1 2 2 1 1 2
L
3
L
4
The italic number between vertices i and j is the value of (L
m
)
i,j
.
Figure 1. L
3
and L
4
See Figure 1 for L
3
and L
4
. L
m
is a Laplacian on V
m
and {(V
m
, L
m
)}
m1
is
a compatible sequence. Let S = {(V
m
, L
m
)}
m1
. Then by Theorem 3.13, we have
a resistance form (E
S
, F
S
) on V

=
m1
V
m
. Note that V

= X. We use E and
F instead of E
S
and F
S
respectively for ease of notation. Let R be the resistance
metric associated with (E, F). Hereafter in this example, we only consider the
topology induced by the resistance metric R. Using the fact that R(i, j) = R
m
(i, j)
for i, j V
m
, where R
m
is the resistance metric with respect to E
L
m
, we may
calculate R(i, j) for any i, j X. As a result,
_
R(0, j) =
1
3
for any j 1,
R(i, j) =
2
3
(1 2
|ij|
) if i, j 1.
Since 1/3 R(i, j) 2/3 for any i, j X with i = j, any one point set {x} is
closed and open. In particular, (X, R) is locally compact. Let B = N. Since B is
the complement of a open set {0}, B is closed. Dene (X) by (0) = 1 and
(x) = 0 for any x B. Since E
m
(|
V
m
, |
V
m
) = m+2 +as m +, we see
that / F. Therefore if u F(B), then u(0) = 0. This shows that B
F
= B{0}.
Now following Brelot [13], we introduce the notion of ne topology associated
with a cone of extended nonnegative valued functions.
Definition 5.6. (1) Dene R
+
+
= [0, +) {+} and f : R
+
+
[0, 1]
by f(x) = 1 (x + 1)
1
. (Note that f(+) = 1.) We give a metric d on R
+
+
by
d(a, b) = |f(x) f(y)|.
(2) Let X be a set and let {f|f : X R
+
+
}. Assume that is a cone, i.e.
a
1
f
1
+a
2
f
2
if a
1
, a
2
R and f
1
, f
2
and that + , where + means a
constant function whose value is +. (We use a convention that 0 += 0.) Let
O be a topology on X, i.e. O is a family of sets satisfying the axiom of open sets.
The coarsest topology which is ner than O and for which all the functions in are
continuous is called the ne topology associated with and O. We use O
F
(O, )
20
to denote the family of open sets with respect to the ne topology associated with
O and . In particular, if O = {, X}, then we write O
F
() = O
F
(O, ).
Considering the resistance form (E, F), the adequate candidate of is F
+
+
dened by
F
+
+
= {u|u F, u(x) 0 for any x X} {+}
Dene O
F
= {U|X\U C
F
}. Also dene O
R
as the collection of open sets with
respect to the R-topology. By Proposition 5.1, we have O
F
O
R
.
Theorem 5.7. Assume that (X, R) is separable.
(1) O
F
(O
R
, F
+
+
) = O
R
.
(2) O
F
(F
+
+
) = O
F
.
By Theorem 5.7, we realize an odd situation where the ne topology is coarser
than the R-topology. Because of this and the fact that we need Theorem 4.3, where
the condition B
F
= B has already been used, to prove Theorem 5.7, we can take
small advantage of the axiomatic potential theory developed in [13].
To prove the above theorem, we need the following lemma.
Lemma 5.8. Assume that (X, R) is separable. Then for any x X, there exists
u F such that u(x) = 0 and u(y) > 0 for any y X\{x}.
Proof. First note that by (RF2), (F
x
, E) is a Hilbert space, where F
x
=
{u|u(x) = 0, u F}.
For any y X, dene u
y
(z) = g
y
{x}
(z). Then by (GF2) and (GF4), if R(y, z) <
R(y, x), then
u
y
(y) u
y
(z) R(y, z) < R(x, y) = u
y
(y).
Hence u
y
(z) > 0 for any z B
R
(y, R(y, x)).
Now x > 0 and let {x
n
}
n1
be a dense subset of X\B
R
(x, ). Choose

n
> 0 so that

n1
_

n
R(x
n
, x) < +. Dene v
n
=

n
m=1

n
u
x
m
. Then

n1
_
E(v
n
v
n+1
, v
n
v
n+1
) =

n1
_

n+1
R(x
n+1
, x) < +. This im-
plies that {v
n
}
n1
is a Cauchy sequence in the Hilbert space (F
x
, E). Hence
v =

n1

n
u
x
n
F
x
. By the above argument, for any z X\B
R
(x, ), there
exists x
n
such that u
x
n
(z) > 0. Therefore, v(z) > 0 for any z X\B
R
(x, ). We
use v

to denote v. Let u
n
= v
1/n
. Choose
n
so that

n1
_

n
E(u
n
, u
n
) < +.
Then similar argument as

n1

n
u
x
n
, we see that

n1

n
u
n
belongs to F
x
. Let
u =

n1

n
u
n
. Then u(z) > 0 for any z X\{x}.
Proof of Theorem 5.7. First we show that
(5.1) O
F
(F
+
+
) O
F
O
R
.
Note that O
F
(F
+
+
) is generated by u
1
((, c)) and u
1
((c, +)) for any c R
and any u F
+
+
. Let u F and let c R. Set B = {x|u(x) c}. Dene
v(x) = max{u(x) c, 0}. Then Proposition 3.15 shows that v F. Furthermore,
v F(B) and v(x) > 0 for any x / B. Hence B
F
= B and so B C
F
. Therefore
u
1
((, c)) O
F
. Using similar argument, we also obtain that u
1
((c, +))
O
F
. Thus we have shown (5.1)
Since O
F
(O
R
, F
+
+
) is ner than O
R
, (5.1) implies O
F
(O
R
, F
+
+
) = O
R
.
Hence we have shown (1).
To prove (2), let B C
F
. By Theorem 4.3, (E, F
B
) is a resistance form on X
B
.
Let R
B
be the resistance metric associated with (E, F
B
) on X
B
. Then R
B
(x, y)
6. REGULARITY OF RESISTANCE FORMS 21
R(x, y) for any x, y X\B and R
B
(x, B) inf
zB
R(x, z). Therefore, since (X, R)
is separable, (X
B
, R
B
) is separable. Hence applying Lemma 5.8 to (E, F
B
) with
x = B, we obtain u F
B
which satises u(y) > 0 for any y X\B. Since u F
+
+
,
u
1
(0) = B is closed under O
F
(F
+
+
). This yields that O
F
O
F
(F
+
+
).
6. Regularity of resistance forms
Does a domain F of a resistance form E contain enough many functions? The
notion of regularity of a resistance form will provide an answer to such a question.
As you will see in Denition 6.2, a resistance form is regular if and only if the domain
of the resistance form is large enough to approximate any continuous function with
a compact support. It is notable that the operation B B
F
plays an important
role again in this section.
Let (E, F) be a resistance form on a set X and let R be the associated resistance
metric on X. We assume that (X, R) is separable in this section.
Definition 6.1. (1) Let u : X R. The support of u, supp(u) is dened
by supp(u) = {x|u(x) = 0}, where U is the closure of U X with respect to the
resistance metric. We use C
0
(X) to denote the collection of continuous functions
on X whose support are compact.
(2) Let K be a subset of X and let u : X R. We dene the supremum norm of
u on K, ||u||
,K
by
||u||
,K
= sup
xK
|u(x)|.
We write || ||

= || ||
,X
if no confusion can occur.
Definition 6.2. The resistance form (E, F) on X is called regular if and only
if F C
0
(X) is dense in C
0
(X) in the sense of the supremum norm || ||

.
The regularity of a resistance form is naturally associated with that of a Dirich-
let form. See Section 9 for details. The following theorem gives a simple criteria
for regularity.
Theorem 6.3. Assume that (X, R) is locally compact. The following conditions
are equivalent:
(R1) (E, F) is regular.
(R2) B
F
= B for any R-closed subset B. In other words, the R-topology coincides
with the C
F
-topology.
(R3) If B is R-closed and B
c
is R-compact, then B
F
= B.
(R4) If K is a compact subset of X, U is an R-open subset of X, K U and U
is R-compact, then there exists F such that supp() U, 0 (y) 1 for
any y X and |
K
1.
Combining the above theorem with Corollary 5.4, we obtain the following result.
Corollary 6.4. If (X, R) is R-compact, then (E, F) is regular.
In general, even if (X, R) is locally compact, (E, F) is not always regular. Recall
Example 5.5.
Hereafter, the R-topology will be always used when we will consider a resistance
form unless we say otherwise. For example, an open set means an R-open set.
To prove Theorem 6.3, we need the following lemma, which can be proven by
direct calculation.
22
Lemma 6.5. If u, v F C
0
(X), then uv F C
0
(X) and
E(uv, uv) 2||u||
2

E(v, v) + 2||v||
2

E(u, u).
Proof of Theorem 6.3. (R1) (R2) Let x / B. Choose r > 0 so that
B(x, r) is compact and B B(x, r) = . Then there exists f C
0
(X) such that
0 f(y) 1 for any y X, f(x) = 1 and supp(f) B(x, r). Since (E, F) is
regular, we may nd v F C
0
(X) such that ||v f||

1/3. Dene u = 3v 1.
Then u(x) = 1 and u|
B
0. Hence x / B
F
.
(R2) (R3) This is obvious.
(R3) (R4) By (R3), (U
c
)
F
= U
c
. Hence, for any x K, we may choose r
x
so
that B(x, r
x
) U and
x
U
c (y) 1/2 for any y B(x, r
x
). Since K is compact,
K
n
i=1
B(x
i
, r
x
i
) for some x
1
, . . . , x
n
K. Let v =

n
i=1

x
i
U
c . Then v(y) 1/2
for any y K and supp(v) U. If = 2v, then u satises the desired properties.
(R4) (R1) Let u C
0
(X). Set K = supp(u). Dene U
K
= {u|
K
: u
F C
0
(X)}. Then by (R4) and Lemma 6.5, we can verify the assumptions of the
Stone Weierstrass theorem for U
K
with respect to || ||
,K
. (See, for example,
[54] on the Stone Weierstrass theorem.) Hence, U
K
is dense in C(K) with respect
to the supremum norm on K. For any > 0, there exists u

F C
0
(X) such
that ||u u

||
,K
< . Let V = K {x : |u

(x)| < }. Suppose that x K and


that there exists {x
n
}
n=1,2,...
V
c
such that R(x
n
, x) 0 as x +. Then
|u

(x
n
)| for any n and hence |u

(x)| . On the other hand, since x


n
K
c
,
u(x
n
) = 0 for any n and hence u(x) = 0. Since x K, this contradict to the fact
that ||u u

||
,K
< . Therefore, V is open. We may choose a open set U so that
K U, U is compact and U V . Let be the function obtained in (R4). Dene
v

= u

. Then by Lemma 6.5, v

FC
0
(X). Also it follows that ||uv

||

.
This shows that F C
0
(X) is dense in C
0
(X) with respect to the norm || ||

.
7. Annulus comparable condition and local property
A resistance form may have a long-distance connection and/or a jump in gen-
eral. For instance, let us modify a given resistance form by adding a new resistor
between two separate points. The modied resistance form has a jump induced
by the added resistor. Such a jump naturally appears in the associated prob-
abilistic process. In this section, we introduce the annulus comparable condition,
(ACC) for short, which ensures certain control of such jumps, or direct connections
between two separate points. For example, Theorems in Section 15 will show that
(ACC) is necessary to get the Li-Yau type on-diagonal heat kernel estimate.
We need the following topological notion to state (ACC).
Definition 7.1. Let (X, d) be a metric space. (X, d) is said to be uniformly
perfect if and only if there exists > 0 such that B
d
(x, (1 + )r)\B
d
(x, r) = for
any x X and r > 0 with X\B
d
(x, r) = .
In [51], the notion of uniformly perfect is called homogeneously dense.
Note that if (X, d) is connected, then it is uniformly perfect. Uniformly per-
fectness ensures that widths of gaps in the space are asymptotically of geometric
progression. For example, the ternary Cantor set is uniformly perfect. In general,
any self-similar set where the contractions are similitudes is uniformly perfect as
follows.
7. ANNULUS COMPARABLE CONDITION AND LOCAL PROPERTY 23
Example 7.2. Let (X, d) be a complete metric space and let N 2. For any
i = 1, . . . , N, F
i
: X X is assumed to be a contraction, i.e.
sup
x=yR
n
d(F
i
(x), F
i
(y))
d(x, y)
< 1.
Then there exists a unique non-empty compact set K X which satises K =

N
i=1
F
i
(K). See [36] for details. K is called the self-similar set associated with
{F
i
}
i=1,...,N
. If (X, d) is R
n
with the Euclidean metric and every F
i
is a similitude,
i.e. F
i
(x) = r
i
A
i
x +a
i
, where r
i
(0, 1), A
i
is an orthogonal matrix and a
i
R
n
,
then K is uniformly perfect.
Next, we have an example which is complete and perfect but not uniformly
perfect.
Example 7.3. Let X = [0, 1]. If F
1
(x) = x
2
/3 and F
2
(x) = x/3 + 2/3, the
self-similar set associated with {F
1
, F
2
} is a (topological) Cantor set, i.e. complete,
perfect and compact. We denote the n-th iteration of F
1
by (F
1
)
n
, i.e. (F
1
)
0
(x) = x
for any x R and (F
1
)
n+1
= (F
1
)
n
F
1
. Dene a
n
= (F
1
)
n
(1/3) and b
n
=
(F
1
)
n
(2/3). Then by inductive argument, we see that (a
n
, b
n
) K = for any
n 0. Now, a
n
= 3e
2
n+1
log 3
and b
n
= 3e
2
n+1
(log 3(log 2)/2)
. Hence b
n
/a
n
=
(3/

2)
2
n+1
. This shows that K is not uniformly perfect.
In this section, (E, F) is a regular resistance form on X and R is the associated
resistance metric. We assume that (X, R) is separable and complete.
Definition 7.4. A resistance form (E, F) on X is said to satisfy the annulus
comparable condition, (ACC) for short, if and only if (X, R) is uniformly perfect
and there exists > 0 such that
(ACC) R(x, B
R
(x, r)
c
) R
_
x, B
R
(x, (1 +)r) B
R
(x, r)
c
_
for any x X and any r > 0 with B
R
(x, r) = X.
Remark. It is obvious that
R(x, B
R
(x, r)
c
) R
_
x, B
R
(x, (1 +)r) B
R
(x, r)
c
_
.
So the essential requirement of (ACC) is the opposite inequality up to a constant
multiplication.
The annulus comparable condition holds if (X, R) is uniformly perfect and
(E, F) has the local property dened below.
Definition 7.5. (E, F) is said to have the local property if and only if E(u, v) =
0 for any u, v F with R(supp(u), supp(v)) > 0.
Proposition 7.6. Assume that (E, F) has the local property and that B
R
(x, r)
is compact for any x X and any r > 0. If B
R
(x, (1 +)r) B
R
(x, r)
c
= , then
R(x, B
R
(x, r)
c
) = R
_
x, B
R
(x, (1 +)r) B
R
(x, r)
c
_
.
In particular, we have (ACC) if (X, R) is uniformly perfect.
Proof. Let K = B
R
(x, (1 +)r) B
R
(x, r)
c
. Recall that
x
K
(y) =
g
K
(x,y)
g
K
(x,x)
and that E(
x
K
,
x
K
) = R(x, K)
1
. By Theorem 6.3, there exists F such that
supp() B
R
(x, (1 +/2)r), 0 (y) 1 for any y X and (y) = 1 for any
24
y B
R
(x, r). By Lemma 6.5, if
1
=
x
K
and
2
=
x
K
(1 ), then
1
and
2
belong to F. Since supp(
2
) B
r
(x, (1 +)r)
c
, the local property implies
E(
x
K
,
x
K
) = E(
1
,
1
) +E(
2
,
2
) E(
1
,
1
).
Note that
1
(y) = 0 for any y B
R
(x, r)
c
and that
1
(x) = 1. Hence, E(
1
,
1
)
E(
x
B
,
x
B
), where B = B(x, r)
c
. On the other hand, since K B, E(
x
B
,
x
B
)
E(
x
K
,
x
K
). Therefore, we have
R(x, B)
1
= E(
x
B
,
x
B
) = E(
x
K
,
x
K
) = R(x, K)
1
.

There are non-local resistance forms which satisfy (ACC), for example, the -
stable process on R and their traces on the Cantor set. See Sections 16. In the next
section, we will show that if the original resistance form has (ACC), then so do its
traces, which are non-local in general.
To study non-local cases, we need the doubling property of the space.
Definition 7.7. Let (X, d) be a metric space.
(1) (X, d) is said to have the doubling property or be a doubling space if and only
if
(7.1) sup
xX,r>0
N
d
(B
d
(x, r), r) < +
for any (0, 1), where N
d
(B, r) is dened in Denition 5.2. (2) Let be a Borel
regular measure on (X, d) which satises 0 < (B
d
(x, r)) < +for any x X and
any r > 0. is said to have the volume doubling property with respect to d or be
volume doubling with respect to d, (VD)
d
for short, if and only if there exists c > 0
such that
(VD)
d
(B
d
(x, 2r)) c(B
d
(x, r))
for any x X and any r > 0.
Remark. (1) It is easy to see that (7.1) holds for all (0, 1) if it holds for
some (0, 1). Hence (X, d) is a doubling space if (7.1) holds for some (0, 1).
(2) If is (VD)
d
, then, for any > 1, (B
d
(x, r)) (B
d
(x, r)) for any x X
and any r > 0.
One of the sucient condition for the doubling property of a space is the
existence of a measure which has the volume doubling property. The following
theorem is well-known. See [32] for example.
Proposition 7.8. Let (X, d) be a metric space and let be a Borel regular
measure on (X, R) with 0 < (B
R
(x, r)) < + for any x X and any r > 0. If
is (VD)
d
, then (X, d) has the doubling property.
The next proposition is straightforward from the denitions.
Proposition 7.9. If a metric space (X, d) has the doubling property, then any
bounded subset of (X, d) is totally bounded.
By the above proposition, if the space is doubling and complete, then every
bounded closed set is compact.
Now we return to (ACC). The following key lemma is a direct consequence of
Theorem 5.3.
8. TRACE OF RESISTANCE FORM 25
Lemma 7.10. Assume that (X, R) has the doubling property and is uniformly
perfect. Then, for some > 0,
(7.2) R
_
x, B
R
(x, r)
c
B
R
(x, (1 +)r)
_
r
for any x X and any r > 0 with B
R
(x, r) = X.
Proof. Set B = B
R
(x, (1 +)r) B
R
(x, r)
c
. Choose so that B = for any
x X and any r > 0 with B
R
(x, r) = X. Then, r R(x, B) (1 +)r. This and
the doubling property of (X, R) imply
N(B, R(x, B)/2) N(B, r/2) N(B
R
(x, (1 + 2)r), r/2) c

,
where c

is independent of x and r. Using Theorem 5.3, we see


r
8c

R(x, B) (1 +)r.

By the above lemma, (ACC) turns out to be equivalent to (RES) dened below
if (X, R) is a doubling space.
Definition 7.11. A resistance form (E, F) on X is said to satisfy the resistance
estimate, (RES) for short, if and only if
(RES) R(x, B
R
(x, r)
c
) r
for any x X and any r > 0 with B
R
(x, r) = X.
Theorem 7.12. Assume that (X, R) has the doubling property. Then (X, R)
is uniformly perfect and (RES) holds if and only if (ACC) holds.
Proof of Theorem 7.12. If (ACC) holds, then (7.2) and (ACC) immedi-
ately imply (RES). Conversely, (RES) along with (7.2) shows (ACC).
Corollary 7.13. If (E, F) has the local property, (X, R) has the doubling
property and is uniformly perfect, then (RES) holds.
8. Trace of resistance form
In this section, we introduce the notion of the trace of a resistance form on a
subset of the original domain. This notion is a counterpart of the notion of traces
in the theory of Dirichlet forms, which has been extensively studied in [21, Section
6.2], for example.
Throughout this section, (E, F) is a resistance form on X and R is the associated
resistance distance. We assume that (X, R) is separable and complete.
Definition 8.1. For a non-empty subset Y X, dene F|
Y
= {u|
Y
: u F}.
Lemma 8.2. Let Y be a non-empty subset of (X, R). For any u F|
Y
, there
exists a unique u

F such that u

|
Y
= u and
E(u

, u

) = min{E(v, v)|v F, v|
Y
= u}.
This lemma is an extension of Proposition 3.10. In fact, if Y is a nite set,
then we have F|
Y
= (Y ) by (RF3-2). Hence Lemma 8.2 holds in this case by
Proposition 3.10. The unique u

is thought of as the harmonic function with the


boundary value u on Y .
26
Proof of Lemma 8.2. Let p Y . Replacing u by u u(p), we may assume
that u(p) = 0 without loss of generality. Choose a sequence {v
n
}
n1
F so
that v
n
|
Y
= u and lim
n+
E(v
n
, v
n
) = inf{E(v, v)|v F, v|
Y
= u}. Let C =
sup
n
E(v
n
, v
n
). By Proposition 3.2, if v = v
n
, then
(8.1) |v(x) v(y)|
2
CR(x, y)
and
(8.2) |v(x)|
2
CR(x, p)
Let {V
m
}
m1
be an increasing sequence of nite subsets of X. Assume that V

m1
V
m
is dense in X. (Since (X, R) is separable, such {V
m
}
m1
does exists.) By
(8.1) and (8.2), the standard diagonal construction gives a subsequence {v
n
i
}
i1
which satises {v
n
i
(x)}
i1
is convergent as i + for any x V

=
m1
V
m
.
Dene u

(x) = lim
i+
v
n
i
(x) for any x V

. Since u

satises (8.1) and (8.2)


on V

with v = u

, u

is extended to a continuous function on X. Note that


the extended function also satises (8.1) and (8.2) on X with v = u

. Set E
m
=
E
L
(E,F),V
m
. Then, by Theorem 3.14,
(8.3) E
m
(v
n
, v
n
) E(v
n
, v
n
) C
for any m 1 and any n 1. Dene M = inf{E(v, v)|v F, v|
Y
= u}. For any
> 0, if n is large enough, then (8.3) shows E
m
(v
n
, v
n
) M + for any m 1.
Since v
n
|
V
m
u

|
V
m
as n +, it follows that E
m
(u

, u

) M + for any
m 1. Theorem 3.14 implies that u

F and E(u

, u

) M.
Next assume that u
i
F, u
i
|
Y
= u and E(u
i
, u
i
) = M for i = 1, 2. Since
E((u
1
+ u
2
)/2, (u
1
+ u
2
)/2) E(u
1
, u
1
), we have E(u
1
, u
2
u
1
) 0. Similarly,
E(u
2
, u
1
u
2
) 0. Combining those two inequalities, we obtain E(u
1
u
2
, u
1
u
2
) =
0. Since u
1
= u
2
on Y , we have u
1
= u
2
on X.
The next denition is an extension of the notion of h
V
dened in Denition 3.11.
Definition 8.3. Dene h
Y
: F|
Y
F by h
Y
(u) = u

, where u and u

are the
same as in Lemma 8.2. h
V
(u) is called the Y -harmonic function with the boundary
value u. For any u, v F|
Y
, dene E|
Y
(u, v) = E(h
Y
(u), h
Y
(v)). (E|
Y
, F|
Y
) is
called the trace of the resistance form (E, F) on Y .
Making use of the harmonic functions, we construct a resistance form on a
subspace Y of X, which is called the trace.
Theorem 8.4. Let Y be a non-empty subset of X. Then h
Y
: F|
Y
F is
linear and (E|
Y
, F|
Y
) is a resistance form on Y . The associated resistance metric
equals to the restriction of R on Y . If Y is closed and (E, F) is regular, then
(E|
Y
, F|
Y
) is regular.
We denote the restriction of R on Y Y by R|
Y
.
The following lemma is essential to prove the above theorem.
Lemma 8.5. Let Y be a non-empty subset of X. Dene
H
Y
= {u|u F, E(u, v) = 0 for any v F(Y )},
where F(Y ) is dened in Denition 2.1. Then, for any f F|
Y
, u = h
Y
(f) if
and only if u H
Y
and u|
Y
= f.
8. TRACE OF RESISTANCE FORM 27
By this lemma, H
Y
= Im(h
Y
) is the space of Y -harmonic functions and F =
H
Y
+ F(Y ), i.e. F is the direct sum of H
Y
and F(Y ). Moreover, E(u, v) = 0 for
any u H
Y
and any v F(Y ). The counterpart of this fact has been know for
Dirichlet forms. See [21, Section 6.2] for details. In the case of weighted graphs
(Example 3.5) such a decomposition is known as the Roydens decomposition. See
[49, Theorem (6.3)] for details.
Proof. Let f

= h
Y
(f). If v F and v|
Y
= f, then
E(t(v f

) +f

, t(v f

) +f

) E(f

, f

)
for any t R. Hence E(v f

, f

) = 0. This implies that f

H
Y
. Conversely
assume that u H
Y
and u|
Y
= f. Then, for any v F with v|
Y
= f,
E(v, v) = E((v u) +u, (v u) +u) = E(v u, v u) +E(u, u) E(u, u).
Hence by Lemma 8.2, u = h
Y
(f).
Proof of Theorem 8.4. By Lemma 8.5, if r
Y
: H
Y
F|
Y
is the restriction
on Y , then r
Y
is the inverse of h
Y
. Hence h
Y
is linear. The conditions (RF1)
through (RF4) for (E|
Y
, F|
Y
) follow immediately from the counterparts for (E, F).
About (RF5),
E|
Y
(u, u) = E(h
Y
(u), h
Y
(u)) E(h
Y
(u), h
Y
(u))
E(h
Y
(u), h
Y
(u)) = E|
Y
(u, u).
The rest of the statement is straightforward.
In the rest of this section, the conditions (ACC) and (RES) are shown to be
preserved by the traces under reasonable assumptions.
Theorem 8.6. Let (E, F) be a regular resistance form on X and let R be
the associated resistance metric. Assume that (E, F) satises (RES). If Y is a
closed subset of X and (Y, R|
Y
) is uniformly perfect, then (RES) holds for the trace
(E|
Y
, F|
Y
).
By Theorem 7.12, we immediately have the following corollary.
Corollary 8.7. Let (E, F) be a regular resistance form on X and let R be the
associated resistance metric. Assume that (X, R) has the doubling property. Let Y
be a closed subset of X and assume that (Y, R|
Y
) is uniformly perfect. If (ACC)
holds for (E, F), then so does for the trace (E|
Y
, F|
Y
).
Notation. Let (E, F) be a resistance form on X and let R be the associ-
ated resistance metric. For a non-empty subset Y of X, we use R
Y
to denote
the resistance metric associated with the trace (E|
Y
, F|
Y
) on Y . Also we write
B
Y
R
(x, r) = B
R
(x, r) Y for any x Y and r > 0.
Note that R
Y
(x, y) = R(x, y) for any x, y Y by Theorem 8.4. Hence R
Y
=
R|
Y
.
Proof of Theorem 8.6. Note that R
Y
(x, Y \B
Y
R
(x, r)) = R(x, B
R
(x, r)
c

Y ). Hence if (RES) holds for (E, F) then,


(8.4) R
Y
(x, Y \B
Y
R
(x, r)) R(x, B
R
(x, r)
c
) c
1
r.
28
On the other hand, since (Y, R
Y
) is uniformly perfect, there exists > 0 such that
B
Y
R
(x, (1 + )r)\B
Y
R
(x, r) = for any x Y and r > 0 with B
Y
R
(x, r) = Y . Let
y B
Y
R
(x, (1 +)r)\B
Y
R
(x, r). Then
(8.5) (1 +)r R
Y
(x, y) R
Y
(x, Y \B
Y
R
(x, r)).
Combining (8.4) and (8.5), we obtain (RES) for (E|
Y
, F|
Y
).
9. Resistance forms as Dirichlet forms
In this section, we will present how to obtain a regular Dirichlet form from a
regular resistance form and show that every single point has a positive capacity. As
in the previous sections, (E, F) is a resistance form on X and R is the associated
resistance metric on X. We continue to assume that (X, R) is separable, complete
and locally compact.
Let be a Borel regular measure on (X, R) which satises 0 < (B
R
(x, r)) <
+ for any x X and r > 0. Note that C
0
(X) is a dense subset of L
2
(X, ) by
those assumptions on .
Definition 9.1. For any u, v F L
2
(X, ), dene E
1
(u, v) by
E
1
(u, v) = E(u, v) +
_
X
uvd.
By [36, Theorem 2.4.1], we have the following fact.
Lemma 9.2. (F L
2
(X, ), E
1
) is a Hilbert space.
Since F C
0
(X) F L
2
(X, ), the closure of F C
0
(X) is a subset of
F L
2
(X, ).
Definition 9.3. We use D to denote the closure of F C
0
(X) with respect to
the inner product E
1
.
Note that if (X, R) is compact, then D = F.
Theorem 9.4. If (E, F) is regular, then (E|
DD
, D) is a regular Dirichlet form
on L
2
(X, ).
See [21] for the denition of a regular Dirichlet form.
For ease of notation, we write E instead of E
DD
.
Definition 9.5. The Dirichlet form (E, D) on L
2
(X, ) is called the Dirichlet
form induced by a resistance form (E, F).
Proof. (E, D) is closed form on L
2
(X, ). Also, since C
0
(X) is dense in
L
2
(X, ), the assumption that F C
0
(X) is dense in C
0
(X) shows that D is dense
in L
2
(K, ). Hence (E, D) is a regular Dirichlet form on L
2
(X, ) with a core
F C
0
(X).
Hereafter in this section, (E, F) is always assumed to be regular. Next we study
capacity of points associated with the Dirichlet form constructed above.
Lemma 9.6. Let x X. Then there exists c
x
> 0 such that
|u(x)| c
x
_
E
1
(u, u)
for any u D. In other words, the map u u(x) from D to R is bounded.
9. RESISTANCE FORMS AS DIRICHLET FORMS 29
Proof. Assume that there exists a sequence {u
n
}
n1
F such that u
n
(x) = 1
and E
1
(u
n
, u
n
) 1/n for any n 1. By (3.1),
|u
n
(x) u
n
(y)|
_
R(x, y)

n

_
R(x, y)
Hence u
n
(y) 1/2 for any y B
R
(x, 1/4). This implies that
||u
n
||
2
2

_
B(x,1/4)
u(y)
2
d (B
R
(x, 1/4))/4 > 0
This contradicts the fact that E
1
(u
n
, u
n
) 0 as n +.
Lemma 9.7. If K is a compact subset of X, then the restriction map
K
:
D C(K) dened by
K
(u) = u|
K
is a compact operator, where D and C(K) are
equipped with the norms
_
E
1
(, ) and || ||
,K
respectively.
Proof. Set D = sup
x,yK
R(x, y). Let U be a bounded subset of D, i.e. there
exists M > 0 such that E
1
(u, u) M for any u U. Then by (3.1),
|u(x) u(y)|
2
R(x, y)M
for any x, y X and any u U. Hence U is equicontinuous. Choose x

K. By
Lemma 9.6 along with (3.1),
u(x)
2
2|u(x) u(x

)|
2
+ 2|u(x

)|
2
2DM + 2c
2
x

M
for any u U and any x K. This shows that U is uniformly bounded on K.
By the Ascoli-Arzel`a theorem, {u|
K
}
uU
is relatively compact with respect to the
supremum norm. Hence
K
is a compact operator.
Definition 9.8. For an open set U X, dene the E
1
-capacity of U, Cap(U),
by
Cap(U) = inf{E
1
(u, u)|u D, u(x) 1 for any x U}.
If {u|u D, u(x) 1 for any x U} = , we dene Cap(U) = +. For any
A X, Cap(A) is dened by
Cap(A) = inf{Cap(U)|U is an open subset of X and A U}.
Theorem 9.9. For any x X, 0 < Cap({x}) < +. Moreover, if K is a
compact subset of X, then 0 < inf
xK
Cap({x}).
The proof of Theorem 9.9 requires the following lemma.
Lemma 9.10. For any x X, there exists a unique g D such that
E
1
(g, u) = u(x)
for any u D. Moreover, let = g/g(x). Then, is the unique element in
{u|u D, u(x) 1} which attains the following minimum
min{E
1
(u, u)|u D, u(x) 1}.
Proof. The existence of g follows by Lemma 9.6. Assume that E
1
(f, u) = u(x)
for any u D. Since E
1
(f g, u) = 0 for any u D, we have f = g. Now, if u D
and u(x) = a > 1, then E
1
(u a, ) = u(x)/g(x) 1/g(x) = 0. Hence,
E
1
(u, u) = E
1
(u a, u a) +E
1
(a, a) E
1
(, ).
This immediately shows the rest of the statement.
30
Definition 9.11. We denote the function g and in Lemma 9.10 by g
x
1
and

x
1
respectively.
Proof of Theorem 9.9. Fix x X. By the above lemma, for any open set
U with x U,
Cap(U) = min{E
1
(u, u)|u D, u(y) 1 for any y U}
min{E
1
(u, u)|u D, u(x) 1} E
1
(
x
1
,
x
1
) =
1
g
x
1
(x)
.
Hence 0 < 1/g
x
1
(x) < Cap({x}) < Cap(U) < +.
Let K be a compact subset of X. By Lemma 9.7, there exists c
K
> 0 such that
||u||
,K
c
K
_
E
1
(u, u) for any u D. Now, for x K,
g
x
1
(x) = E
1
(g
x
1
, g
x
1
) = sup
uD,u=0
E
1
(g
x
1
, u)
2
E
1
(u, u)
= sup
uD,u=0
u(x)
2
E
1
(u, u)
(c
K
)
2
Hence Cap({x}) 1/g
x
1
(x) (c
K
)
2
.
Following [21, Section 2.1], we introduce the notion of quasi continuous func-
tions.
Definition 9.12. A function u : X R is called quasi continuous if and only
if, for any > 0, there exists V X such that Cap(V ) < and u|
X\V
is continuous.
Theorem 9.9 implies that every quasi continuous function is continuous.
Proposition 9.13. Any quasi continuous function is continuous on X.
Proof. Let u be a quasi continuous function. Let x X. Since (X, R) is
locally compact, B
R
(x, r) is compact for some r > 0. By Theorem 9.9, we may
choose > 0 so that inf
yB
R
(x,r)
Cap({y}) > . There exists V X such that
Cap(V ) < and u|
X\V
is continuous. Since V B
R
(x, r) = , u is continuous at
x. Hence u is continuous on X.
10. Transition density
In this section, without ultracontractivity, we establish the existence of jointly
continuous transition density (i.e. heat kernel) associated with the regular Dirichlet
form derived from a resistance form.
As in the last section, (E, F) is a resistance form on X and R is the associated
resistance metric. We assume that (X, R) is separable, complete and locally com-
pact. is a Borel regular measure on X which satises 0 < (B
R
(x, r)) < +
for any x X and any r > 0. We continue to assume that (E, F) is regular. By
Theorem 9.4, (E, D) is a regular Dirichlet form on L
2
(X, ), where D is the closure
of F C
0
(X) with respect to the E
1
-inner product.
Let H be the nonnegative self-adjoint operator associated with the Dirichlet
form (E, D) on L
2
(X, ) and let T
t
be the corresponding strongly continuous semi-
group. Since T
t
u D for any u L
2
(X, ), we always take the continuous version
of T
t
u. In other words, we may naturally assume that T
t
u is continuous for any
t > 0.
Let M = (, {X
t
}
t>0
, {P
x
}
xX
) be the Hunt process associated with the reg-
ular Dirichlet form (E, D) on L
2
(X, ). By [21, Chapter 7], such a Hunt process
is determined outside a set of capacity zero in general. Note that Cap(A) = 0 if
10. TRANSITION DENSITY 31
and only if A = by Theorem 9.9. Hence, the Hunt process M is determined for
every x X in our case. Moreover, by [21, Theorem 4.2.1], every exceptional set is
empty. (See [21, Section 4.1] for the denition of exceptional sets and its relation
to other notions like polar sets, semi-polar sets and negligible sets.) Let p
t
be the transition semigroup associated with the Hunt process M. For non-negative
-measurable function u,
(p
t
u)(x) = E
x
(u(X
t
))
for any x X. Let B be the Borel -algebra of (X, R). We say that u is Borel mea-
surable, if and only if u
1
((a, b]) B for any a, b R. Combining Proposition 9.13
and [21, Theorem 4.2.3], we have the following statement.
Proposition 10.1. For any nonnegative u L
2
(X, ), (p
t
u)(x) = (T
t
u)(x)
for any t > 0 and any x X.
Definition 10.2. Let U be an open subset of X. Dene D
U
= {u|u
D, u|
U
c 0}. Also we dene E
U
= E|
D
U
D
U
.
Note that if U is compact, then D
U
= F(U
c
).
Combining the results in [21, Section 4.4], we have the following facts.
Theorem 10.3. Let
U
be the restriction of on U, i.e.
U
(A) = (A U)
for any Borel set U. Then (E
U
, D
U
) is a regular Dirichlet form on L
2
(U,
U
).
Moreover, if M
U
= (
U
, X
U
t
, P
U
x
) is the associated Hunt process, then
P
U
x
(X
U
t
A) = P(X
t
A, t <
U
)
for any Borel set A and any x U, where
U
is the exit time of U dened by

U
() = inf{t > 0|X
t
() X\U}.
Moreover, if p
U
t
is the transition semigroup associated with M
U
, hen
(10.1) (p
U
t
u)(x) = E
U
x
(u(X
U
t
)) = E
x
(
{t<
U
}
u(X
t
))
for any non-negative measurable function u and any x X.
Remark. For a function u : U R, we dene
U
(u) : X R by
U
(u)|
U
= u
and
U
(u)|
U
c 0. Through this extension map, L
2
(U,
U
) is regarded as a subspace
of L
2
(X, ). Also, if u D
U
, then
U
(u|
U
) = u and hence we may think of D
U
as
a subset of C(X) through
U
. Hereafter, we always use these conventions.
Remark. By the same reason as in the case of M, the process M
U
is deter-
mined for every x U.
The existence and the continuity of heat kernels have been studied by several
authors. In [5], the existence of quasi continuous versions of heat kernel (i.e. tran-
sition density) has been proven under ultracontractivity. Grigoryan has shown the
corresponding result only assuming local ultracontractivity in [22]. In [17], the
existence of jointly continuous heat kernels has been shown for the case of Dirichlet
forms induced by resistance forms under ultracontractivity. The following theorem
establishes the existence of jointly continuous heat kernel for the Dirichlet form in-
duced by a resistance form without ultracontractivity and, at the same time, gives
an upper diagonal estimate of the heat kernel. The main theorem of this section is
the following.
32
Theorem 10.4. Assume that B
R
(x, r) is compact for any x X and r > 0. Let
U be a non-empty open subset of X. Then there exits p
U
(t, x, y) : (0, +XX
[0, +) which satises the following conditions:
(TD1) p
U
(t, x, y) is continuous on (0, +) X X. Dene p
t,x
U
(y) = p
U
(t, x, y).
Then p
t,x
U
D
U
for any (t, x) (0, +) X.
(TD2) p
U
(t, x, y) = p
U
(t, y, x) for any (t, x, y) (0, +) X X.
(TD3) For any non-negative (Borel)-measurable function u and any x X,
(10.2) (p
U
t
u)(x) =
_
X
p
U
(t, x, y)u(y)(dy)
(TD4) For any t, s > 0 and any x, y X,
(10.3) p
U
(t +s, x, y) =
_
X
p
U
(t, x, z)p
U
(s, y, z)(dz).
Furthermore, let A be a Borel subset of X which satises 0 < (A) < +. Dene
R(x, A) = sup
yA
R(x, y) for any x X. Then
(10.4) p
U
(t, x, x)
2R(x, A)
t
+

2
(A)
for any x X and any t > 0.
The proof of the upper heat kernel estimate (10.4) is fairly simple. Originally,
the same result has been obtained by more complicated discussion in [6] and [41].
A simplied argument, which is essentially the same as ours, for random walks can
be found in [8].
Remark. In fact, we have the following inequality which is slightly better than
(10.4). For any > 0,
(10.5) p
U
(t, x, x)
_
1 +
1

_
R(x, A)
t
+

1 +
(A)
.
This inequality implies that
lim
t+
p
U
(t, x, x)
1
(X)
.
for any x X.
Definition 10.5. p
U
(t, x, y) is called the transition density and/or the heat
kernel associated with the Dirichlet form (E
U
, D
U
) on L
2
(X, ).
Corollary 10.6. Assume that B
R
(x, r) is compact for any x X and any
r > 0. Let U be a non-empty open subset of X. Then
lim
t0
tp
U
(t, x, x) = 0
for any x X.
Proof. Choose A = B
R
(x, r). By (10.4), it follows that tp
U
(t, x, x) 3r for
suciently small t.
Now we begin our proof of Theorem 10.4 which consists of several lemmas.
First we deal with the case where U is compact.
Lemma 10.7. If U is compact, then we have p
U
(t, x, y) : (0, +) X X
which satises (TD1), (TD2), (TD3) and (TD4).
10. TRANSITION DENSITY 33
Proof. Let H
U
be the non-negative self-adjoint operator on L
2
(U,
U
) associ-
ated with the Dirichlet form (E
U
, D
U
). Then by Lemma 9.7, H
U
has compact resol-
vent. Hence, there exists a complete orthonormal system {
n
}
n1
of L
2
(U,
U
) and
{
n
}
n1
[0, +) such that
n
Dom(H
U
) D
U
, H
U

n
=
n

n
,
n

n+1
and lim
n+

n
= +.
Claim 1:
(10.6)

n1
1

n
+ 1
< +
Proof of Claim 1: By Lemma 9.6, for any x U, there exists g
x
1,U
D
U
such that
E
1
(g
x
1,U
, u) = u(x) for any u D
U
. Since

n
(x) = E
1
(g
x
1,U
,
n
) = (
n
+ 1)
_
U
g
x
1,U

n
d
U
,
we have g
x
1,U
=

n1

n
(x)

n
+1

n
in L
2
(U,
U
). Hence
(10.7) g
x
1,U
(x) = E
1
(g
x
1,U
, g
x
1,U
) =

n1

n
(x)
2

n
+ 1
On the other hand, by the same argument as in the proof of Theorem 9.9, there
exists c
U
> 0 such that
|E
1
(u, g
x
1,U
)| |u(x)| ||u||
,K
c
U
_
E
1
(u, u)
for any u D
U
, where K = U. This implies that E
1
(g
x
1,U
, g
x
1,U
) c
U
. Combining
this with (10.7), we see that g
x
1,U
(x) is uniformly bounded on U. Hence integrating
(10.7) with respect to x, we obtain (10.6) by the monotone convergence theorem.
Claim 2: ||
n
||

D
n
for any n 2, where D = sup
x,yU
R(x, y).
Proof of Claim 2: By (3.1),
(10.8) |
n
(x)
n
(y)|
2
E(
n
,
n
)R(x, y) =
n
R(x, y).
We have two cases. First if U = X, then
n
(y) = 0 for any y U
c
. Hence (10.8)
implies the claim. Secondly, if U = X, then (X, R) is compact. It follows that
1
=
0 and
1
is constant on X. Hence
_
X

n
(x)(dx) = 0 for any n 2. For any x X,
we may nd y X so that
n
(x)
n
(y) 0. Since |
n
(x)|
2
|
n
(x)
n
(y)|
2
,
(10.8) yields the claim.
Claim 3:

n1
e

n
t

n
(x)
n
(y) converges uniformly on [T, +)XX for any
T > 0.
Proof of Claim 3: Note that e
a
2/a
2
for any a > 0. This fact with Claim 2
shows that |e

n
t

n
(x)
n
(y)| 2/(
n
t
2
). Using Claim 1, we immediately obtain
Claim 3.
Now, let p
U
(t, x, y) =

n1
e

n
t

n
(x)
n
(y). By Claim 3, p is continuous
on (0, +) X X. Also, p
U
is the integral kernel of the strongly continuous
semigroup {T
U
t
}
t>0
associated with the Dirichlet form (E
U
, D
U
) on L
2
(U,
U
). Let
A be a Borel set. Then
(T
U
t

A
)(x) =
_
X
p
U
(t, x, y)
A
(y)(dy).
In particular T
U
t

A
is continuous on X. By [21, Theorem 4.2.3], p
U
t

A
is a quasi
continuous version of T
U
t

A
. Since any quasi continuous function is continuous
by Proposition 9.13, we have (T
U
t

A
)(x) = (p
U
t

A
)(x) for any x X. Letting
34
p
U
(t, x, y) = p
U
(t, x, y), we have (TD3). The rest of the desired properties are
straightforward.
The following facts are well-known in general setting. See [21] for example. In
this particular situation, they can be checked by the eigenfunction expansion of the
heat kernel above.
Lemma 10.8. Assume that U is compact.
(1) For any t > 0 and any x, y X,
p
U
t
(t, x, y) = E(p
t/2,x
U
, p
t/2,y
U
).
(2) For any t, s > 0 and any x X,
E(p
t,x
U
, p
s,x
U
)
2
t +s
p
U
_
t +s
2
, x, x
_
.
Lemma 10.9. If U is compact, then (10.4) holds for any Borel subset A of X
which satises 0 < (A) < +.
Proof. Since
_
A
p
U
(t, x, y)(dy)
_
X
p
U
(t, x, y)(dy) 1, there exists y


A such that p
U
(t, x, y

) 1/(A). By this fact along with Lemma 10.8-(2),


1
2
p
U
(t, x, x)
2
p
U
(t, x, y

)
2
+|p
U
(t, x, x) p
U
(t, x, y

)|
2

1
(A)
2
+R(x, A)E(p
t,x
U
, p
t,x
U
)
1
(A)
2
+
R(x, A)
t
p
U
(t, x, x)
Solving this with respect to p
U
(t, x, x), we have
p
U
(t, x, x)
R(x, A)
t
+
_
2
(A)
2
+
R(x, A)
2
t
2
_1
2

2R(x, A)
t
+

2
(A)
.

Remark. To get (10.5), we only need to use


p
U
(t, x, x)
2
(1 +)p
U
(t, x, y)
2
+
_
1 +
1

_
|p
U
(t, x, x) p
U
(t, x, y)|
2
in place of the inequality with = 1 in the above proof.
Thus we have shown Theorem 10.4 if U is compact.
Proof of Theorem 10.4. If U is compact, then we have completed the proof.
Assume that U is not compact. Fix x

X and set U
n
= B
R
(x

, n) U for any
n = 1, 2, . . .. Note that U
n
is compact. Write p
n
(t, x, y) = p
U
n
(t, x, y).
Claim 1 p
n
(t, x, y) p
n+1
(t, x, y) for any x, y X and any n 1.
Proof of Claim 1. Let
n
=
U
n
. Then
n

n+1
for any n. Hence
(p
U
n
t
u)(x) = E
x
(
t<
n
u(X
t
)) E
x
(
t<
n+1
u(X
t
)) = (p
U
n+1
t
u)(x)
for any non-negative measurable function u and any x X. By (TD3), we deduce
Claim 1.
10. TRANSITION DENSITY 35
Let A be a Borel subset of X which satises 0 < (A) < +. By (TD4) and
(10.4), we have
(10.9) p
n
(t, x, y)
_
p
n
(t, x, x)
_
p
n
(t, y, y)

_
2R(x, A)
t
+

2
(A)
_1
2
_
2R(y, A)
t
+

2
(A)
_1
2
for any x X, any t > 0 and any n. Hence p
n
(t, x, y) is uniformly bounded and
monotonically nondecreasing as n +. This shows that p
n
(t, x, y) converges
as n +. If p(t, x, y) = lim
n+
p
n
(t, x, y), then p(t, x, y) satises the same
inequality as (10.9). In particular (10.4) holds for p(t, x, x). Also, we immediately
verity (TD2) and (TD4) for p(t, x, y) from corresponding properties of p
n
(t, x, y).
About (TD3), let u be a non-negative Borel-measurable function. Then by (TD3)
for p
n
(t, x, y),
(p
U
n
t
u)(x) = E
x
(
{t<
n
}
u(X
t
)) =
_
X
p
n
(t, x, y)u(y)(dy)
for any x X. The monotone convergence theorem shows that
E
x
(
t<
U
u(X
t
)) =
_
X
p(t, x, y)u(y)(dy).
Since the left-hand side of the about equality equals (p
U
t
u)(x), we have (TD3).
Finally we show (TD1). Fix (t, x, y) (0, +) XX. Dene V = (t , t +
) B
R
(x, r) B
R
(y, r), where r > 0 and 0 < < t. (10.4) shows that
C = sup
(s,x

,y

)V
1
,n1
__
p
n
(s, x

, x

)
s
+
_
p
n
(s, y

, y

)
s
_
< +,
where V
1
= ((t )/2, t + ) B
R
(x, r) B
R
(y, r). By Lemma 10.8, for any
(s, a, b) V and any n 1,
|p
n
(t, x, y) p
n
(s, a, b)|
|p
n
(t, x, y) p
n
(t, x, b)| +|p
n
(t, x, b) p
n
(t, a, b)| +|p
n
(t, a, b) p
n
(s, a, b)|

_
E(p
t,x
U
n
, p
t,x
U
n
)R(y, b) +
_
E(p
t,b
U
n
, p
t,b
U
n
)R(x, a) +|t s||
p
n
t
(t

, a, b)|

_
p
n
(t, x, x)R(y, b)
t
+
_
p
n
(t, b, b)R(x, a)
t
+ 2|t s|
_
p
n
(t

/2, a, a)p(t

/2, b, b)
t

C
_
R(x, a) +C
_
R(y, b) +C
2
|t s|
where t

is a value between t and s. Letting n +, we have


|p(t, x, y) p(s, a, b)| C
_
R(x, a) +C
_
R(y, b) +C
2
|t s|.
Hence p(t, x, y) is continuous on (0, +) X X. By (TD4), p
t,x
U
L
2
(X, ) for
any t > 0 and any x X. Using (TD3) and (TD4), we see
p
t,x
U
= p
U
t/2
(p
t/2,x
U
) = T
t
(p
t/2,x
U
)
for any t > 0 and any x, y X, where {T
t
}
t>0
is the strongly continuous semigroup
associated with the Dirichlet form (E
U
, D
U
) on L
2
(X, ). Hence p
t,x
U
D for any
t > 0 and any x X.
36
At the end of this section, we give a fundamental relations between the Green
function, Laplacian and the transition density. By Theorem 10.3, (E
U
, D
U
) is a
regular Dirichlet form on L
2
(U, |
U
). (D
U
and E
U
are dened in Denition 10.2.)
Theorem 10.10. Assume that U X, U is compact and that U = X. Let
B = X\U. Then
(10.10) g
B
(x, y) =
_

0
p
U
(t, x, y)dt
for any x, y X. Moreover, let H
U
be the nonnegative self-adjoint operator
on L
2
(U,
U
) associated with the Dirichlet form (E
U
, D
U
). Dene G
U
f for f
L
2
(U,
U
) by
(G
U
f)(x) =
_
U
g
B
(x, y)f(y)(dy).
for any x U. Then Im(G
U
) = Dom(H
U
) and G
U
is the inverse of H
U
.
By the next corollary, we can identify g
B
(x, y) with the (0-order) Green function
associated with the Dirichlet form (E
U
, D
U
). See Bass [11, Sections II.3 & II.4].
Note that (10.10) and (10.11) are the counterparts of [11, II (4.5)Corollary and
II (3.14)] respectively.
Corollary 10.11. Under the same assumptions as in Theorem 10.10,
(10.11) E
x
(
t<
U
f(X
s
)) =
_
U
g
B
(x, y)f(y)(dy).
for any nonnegative Borel measurable function f : X R and any x X. In
particular,
(10.12) E
x
(
U
) =
_
U
g
B
(x, y)(dy).
(10.12) connects the mean exit time with the Green function. It plays an
important role in the heat kernel estimate, in particular, in the proof of Lemma 18.1.
Proof of Theorem 10.10. We use the same notations as in the proof of
Lemma 10.7. Since B = , (RF2) implies that E
U
(u, u) > 0 for any u D
U
.
Therefore,
1
> 0. Note that g
x
B
F(B) = D
U
. We have
(10.13)

n
(x) = E(g
x
B
,
n
) =
_
U
g
x
B
(y)(H
U

n
)(y)(dy) =
n
_
U
g
x
B
(y)
n
(y)(dy).
Hence g
x
B
=

n1
(
n
)
1

n
(x)
n
in L
2
(U,
U
). This implies
(10.14) g
B
(x, y) = E(g
x
B
, g
y
B
) =

n1

n
(x)
n
(y)

n
for any x, y X. (Note that if x / X, then both sides of the above equality is 0.)
By the proof of Lemma 10.7,
p
U
(t, x, y) =

n0
e

n
t

n
(x)
n
(y),
10. TRANSITION DENSITY 37
where the innite sum is uniformly convergent on [T, +) X X for any T > 0.
Moreover, we have |e

n
t

n
(x)
n
(y)| 2/(
n
t
2
) by the proof of the Claim 3 of
Lemma 10.7. This shows that
_

T
p
U
(t, x, y)dt =

n1
e

n
T

n
(x)
n
(y).
Hence

_

T
p
U
(t, x, y)dt

n1

n
(x)
n
(y)

n1
(1 e

n
T
)
n
(x)
2

n1
(1 e

n
T
)
n
(y)
2

n
By the monotone convergence theorem, letting T 0, we obtain (10.10).
Now, note that
Dom(H
U
) =
_
u

u L
2
(U,
U
), u =

n1

n
,

n1
(
n

n
)
2
< +
_
and that H
U
u =

n1

n
if u =

n1

n
. Let f =

n1
a
n

n

L
2
(U,
U
). Then
(G
U
f)(x) =
_
U
g
x
B
fd =

n1
a
n

n
(x).
By the above facts, we see that G
U
f Dom(H
U
) and H
U
G
U
f = f. Similarly, it
is easy to see that G
U
H
U
u = u for any u Dom(H
U
). Thus G
U
is the inverse of
G
U
.
Proof of Corollary 10.11. By (10.1), it follows that
_
U
p
U
(t, x, y)f(y)(dy) = E
x
(
t<
U
f(X
s
)).
Integrating this from 0 to + with respect to t and using the Fubini theorem, we
have (10.11). (10.12) follows by letting f = 1.
Part 2
Quasisymmetric metrics and
volume doubling measures
The main subject of this part is the notion of quasisymmetric maps, which has
been introduced in [51] as certain generalization of quasiconformal mappings of
the complex plane. The results in this part will play an indispensable role in the
next part, where we will modify the original resistance metric quasisymmetrically
to obtain a metric which is suitable for describing asymptotic behaviors of the
associated heat kernel.
At the rst section, we present several notions, whose combinations are shown
to be equivalent to being quasisymmetric in the second section, where we give
the precise denition of quasisymmetry in Denition 12.1. Establishing such an
equivalence, we resolve the notion of being quasisymmetric into geometric and
analytic components. In the latter two sections, we discuss relations between a
metric and a measure. Under the volume doubling property of a measure, we will
construct a quasisymmetric metric which satises prescribed relation between a
measure and a metric.
Since [51], quasisymmetric maps and related subjects have been studied deeply
by many authors. See Heinonen [32] for example. Some of the results in this section
may be included in the preceding articles. However, we give all the proofs since it
is dicult to nd an exact reference from such a huge literature.
11. Semi-quasisymmetric metrics
In this section, we introduce several notions associated with quasisymmetric
mappings and clarify their relations.
Notation. Let X be a set and let d be a distance on X. B
d
(x, r) is the
closed ball, i.e. B
d
(x, r) = {y|y X, d(x, y) r}. For any A X, diam(A, d)
is the diameter of A with respect to d dened by diam(A, d) = sup
x,yA
d(x, y).
Moreover, we set d

(x) = sup
yX
d(x, y) for any x X.
In the rest of this section, we assume that d and are distances on a set X.
The following notion semi-quasisymmetric is called weakly quasisymmetric
in [51] and can be traced back to [14] and [33]. See [51] for details.
Definition 11.1. is said to be semi-quasisymmetric with respect to d, or
(SQS)
d
for short, if and only if there exist (0, 1) and > 0 such that (x, z) <
(x, y) whenever d(x, z) < d(x, y).
In the above denition, we may assume < 1 without loss of generality.
Proposition 11.2. If is (SQS)
d
, then the identity map from (X, d) to (X, )
is continuous.
This fact has been obtained in [51].
Proof. Assume that d(x
n
, x) 0 as n + and (x
n
, x) a as n +,
where a > 0. Choose
1
(, 1). Then (x
n+m
, x) >
1
(x
n
, x) and d(x
n+m
, x) <
d(x
n
, x) for suciently large n and m. By (SQS)
d
, it follows that (x
n+m
, x) <
(x
n
, x). This contradiction implies the desired conclusion.
Proposition 11.3. Assume that (X, d) is uniformly perfect. Then is (SQS)
d
if and only if, for any > 0, there exists > 0 such that (x, z) < (x, y) whenever
d(x, z) < d(x, y).
11. SEMI-QUASISYMMETRIC METRICS 41
Proof. Assume that is (SQS)
d
. We will show that d(x, z) < (c)
n
d(x, y)
implies (x, z) <
n
(x, y) by induction, where c is the constant appearing in
Denition 7.1. The case n = 1 is obvious. Assume that it is true for n. Sup-
pose that d(x, z) < (c)
n+1
d(x, y). Since (X, d) is uniformly perfect, there exits
y

X such that c(c)


n
d(x, y) d(x, y

) < (c)
n
d(x, y). By induction assumption,
(x, y

) <
n
(x, y). Also since d(x, z) < d(x, y

). we have (x, z) < (x, y

).
Therefore (x, z) <
n+1
(x, y).
The converse is obvious.
Next we consider a geometric interpretation of semi-quasisymmetry. We say
that is semi-quasiconformal with respect to d, (SQC)
d
for short, if -balls are
equivalent to d-balls with a uniform distortion. (The precise denition is given in
Denition 11.4.) (SQS)
d
implies (SQC)
d
but not vise versa. To get an if and
only if assertion, we need a kind of uniform distortion condition regarding annuli
instead of balls, called annulus semi-quasiconformality. To give a precise statement,
we introduce the following notions.
Definition 11.4. (1) Dene d

(x, r) = sup
yB

(x,r)
d(x, y) for x X and
r > 0. d is said to be doubling with respect to if and only if there exist > 1
and c > 0 such that d

(x, r) cd

(x, r) < + for any r > 0 and any x X.


(2) is said to be semi-quasiconformal with respect to d, or (SQC)
d
for short, if
and only if d

(x, r) < + for any x X and any r > 0 and there exists (0, 1)
such that B
d
(x, d

(x, r)) B

(x, r) for any x X and r > 0.


(3) is said to be annulus semi-quasiconformal with respect to d, or (ASQC)
d
for
short, if and only if d

(x, r) < + for any x X and r > 0 and, for any (0, 1),
there exists (0, 1) such that d(x, y) d

(x, r) whenever (x, y) r.


(4) is called weak annulus semi-quasiconformal with respect to d, or (wASQC)
d
for short, if and only if d

(x, r) < + for any x X and r > 0 and there exist


(0, 1) and (0, 1) such that d(x, y) d

(x, r) whenever (x, y) r.


Remark. (1) If d is doubling with respect to , then
d

(x, ar) c
0
a

(x, r)
for any r > 0, a 1 and x X, where c
0
and are positive constants which are
independent of x, a and r. Hence the value of a itself is not essential. An easy
choice of a is two, and this is why we call this notion doubling.
(2) Note that B

(x, r) B
d
(x, d

(x, r)). Hence is (SQC)


d
if and only if, for any
x X and r > 0, there exist R
1
and R
2
such that B
d
(x, R
1
) B

(x, r) B
d
(x, R
2
)
and R
1
CR
2
, where C (0, 1) is independent of x and r. Therefore in this case,
a -ball is equivalent to a d-ball with a uniformly bounded distortion.
(3) Assume that d

(x, r) < + for any x and r. Then (ASQC)


d
is equivalent to
the following statement: for any (0, 1), there exists (0, 1) such that d(x, y)
d

(x, r) whenever r > (x, y) r. Also (ASQC)


d
implies that a -annulus
B

(x, r)\B

(x, r) is contained in a d-annulus B


d
(x, (1+)d

(x, r))\B
d
(x, d

(x, r))
for any > 0.
Theorem 11.5. Assume that both (X, d) and (X, ) are uniformly perfect and
that d

(x, r) < + for any x X and r > 0. Then the following four conditions
are equivalent.
(a) is (SQS)
d
.
42
(b) d is doubling with respect to and is (SQC)
d
.
(c) is (ASQC)
d
.
(d) is (wASQC)
d
.
Proof. (a) (b): First we show that d is doubling with respect to . Since
(X, ) is uniformly perfect, B

(x, r)\B

(x, cr) = unless X\B

(x, r) = , where c
is independent of x and r. By Proposition 11.3, we may assume < c
2
. Now by
(SQS)
d
, (x, z)/ (x, y) implies d(x, z) d(x, y).
Claim: Suppose r/

> (x, y). Then (x, z)/ > (x, y) for some z B

(x, r).
Proof of the claim: If X\B

(x, r) = , then there exists z X such that r >


(x, z) cr. Hence (x, z)/ cr/ > r/

> (x, y). In case X = B

(x, r),
let

(x) = sup
x

X
(x, x

). Then

(x)/ >

(x) (x, y). Hence there exists


z B

(x, r) = X such that (x, z)/ > (x, y). Thus we have shown the claim.
If (x, z)/ > (x, y), (SQS)
d
implies d(x, z) d(x, y). By the above claim, we
obtain that d

(x, z) d

(x, y/

). Hence d is doubling with respect to .


Next we show that is (SQC)
d
. Suppose that d(x, z) < d

(x, r). Then


there exists y B

(x, r) such that d(x, z) < d(x, y). Hence by (SQS)


d
, (x, z) <
(x, y) < r and hence z B

(x, r).
(b) (c): Let (0, 1). By (SQC)
d
, B

(x, r) B
d
(x, d

(x, r)). Since d is


doubling with respect to , d

(x, r) > c

(x, r), where c

is independent of x and
r. Therefore, B

(x, r) B
d
(x, c

(x, r)). This immediately implies (ASQC)


d
.
(c) (d): This is obvious.
(d) (a): Let (x, z) r. Then
n+1
r (x, z) <
n
r for some n 0. By
(ASQC)
d
, d

(x, r) d

(x,
n
r) d(x, z). Hence, d(x, z) < d

(x, r) implies
(x, z) < r. Now suppose d(x, z) < d(x, y). Since d(x, y) d

(x, (x, y)), we


have (x, y) < (x, y).
Next we present useful implications of (SQS)
d
, (SQC)
d
and (ASQC)
d
.
Definition 11.6. is said to decay uniformly with respect to d if and only if
(i) diam(X, d) < + and there exist r

> diam(X, d) and (a, ) (0, 1)


2
such
that
d
(x, r) a
d
(x, r) for any x X and r (0, r

]
or
(ii) diam(X, d) = + and there exists (a, ) (0, 1)
2
such that
d
(x, r)
a
d
(x, r) for any x X and r > 0.
Proposition 11.7. Assume that (X, d) is uniformly perfect and is (SQS)
d
.
Then decays uniformly with respect to d. More precisely, if diam(X, d) < +,
then, for any r

> 0, there exists (a, ) (0, 1)


2
such that
d
(x, r) a
d
(x, r) for
any x X and r (0, r

].
Remark. If is (SQS)
d
, then B
d
(x, d(x, y)) B

(x, (x, y)). Hence if r <


d

(x), we have
d
(x, r) < +. Note that d

(x) diam(X, d)/2.


Proof. Since is (SQS)
d
, there exist (0, 1) and (0, 1) such that
(x, z) < (x, y) whenever d(x, z) < d(x, y). Also there exists c (0, 1) such
that B
d
(x, r)\B
d
(x, cr) = unless X = B
d
(x, r). Suppose that diam(X, d) < +.
Choose n 1 so that c
n1
r

< diam(X, d)/2. Since d

(x) diam(X, d)/2, it


follows that X = B
d
(x, c
n1
r) for any r (0, r

]. Therefore, there exists y X


such that c
n
r d(x, y) < c
n1
r. If d(x, z) < c
n
r, we have d(x, z) < d(x, y).
Hence, (x, z) < (x, y). This shows that
d
(x, c
n
r) (x, y)
d
(x, r).
Similar argument suces as well in the case where diam(X, d) = +.
12. QUASISYMMETRIC METRICS 43
Proposition 11.8. Assume that d is doubling with respect to and is
(SQC)
d
. Let be a Borel regular measure on X. Then is (VD)

if is (VD)
d
.
Combining this proposition with Theorem 11.5, we see that the volume doubling
property is inherited from d to if is (SQS)
d
under uniform perfectness.
Proof. Since d is (VD)

,
B

(x, 2r) B
d
(x, 2d

(x, 2r)) B
d
(x, c

(x, r)),
where c

> 1. If is (VD)
d
, then
(B
d
(x, c

(x, r))) < c(B


d
(x, d

(x, r))).
Moreover, by (SQC)
d
, B
d
(x, d

(x, r)) B

(x, r). Thus, we have


(B

(x, 2r)) c(B


d
(x, d

(x, r)) c(B

(x, r)).

The following lemma is quite similar to Theorem 11.5 but is a little stronger
since it does not assume that (X, d) is uniformly perfect. We will take advantage
of this stronger statement later in the proofs of Proposition 11.10 and Lemma 13.9.
Lemma 11.9. Assume that (X, ) is uniformly perfect. If is (ASQC)
d
, then
d is doubling with respect to .
Proof. By the assumption, B

(x, r)\B

(x, cr) = unless B

(x, cr) = X for


some c (0, 1). Let = c
2
. By (ASQC)
d
, for some (0, 1), d(x, y) d

(x, r)
when (x, y) r. If B

(x, r) = B

(x, cr), then d

(x, r) = d

(x, cr) d

(x, r). If
B

(x, r)\B

(x, cr) = , then there exists y X such that r (x, y) cr. This
also implies that d

(x, cr) d

(x, r). Hence we have the doubling property of


d

.
Proposition 11.10. Let (X, ) be uniformly perfect. Assume that d

(x, r) <
+ for any x X and r > 0. If is (ASQC)
d
, then
d(x, y) d

(x, (x, y))


for any x, y X.
Proof. There exist (0, 1) and (0, 1) such that r (x, y) < r implies
d(x, y) d

(x, r). Choose > 1 so that < 1. For any y X, we have


(x, y) (x, y) < (x, y). Hence d(x, y) d

(x, (x, y)) d

(x, (x, y)).


By Lemma 11.9, d is doubling with respect to . Therefore, c
2
d

(x, (x, y))


d

(x, (x, y)) d(x, y), where c


2
only depends on .
12. Quasisymmetric metrics
In this section, we will introduce the notion of being quasisymmetric and relate
it to the notions obtained in the last section.
d and are distances on a set X through this section.
Definition 12.1. is said to be quasisymmetric, or QS for short, with respect
to d if and only if there exists a homeomorphism h from [0, +) to itself such that
h(0) = 0 and, for any t > 0, (x, z) < h(t)(x, y) whenever d(x, z) < td(x, y). We
write
QS
d if is quasisymmetric with respect to d.
44
The followings are basic properties of quasisymmetric distances.
Proposition 12.2. Assume that is quasisymmetric with respect to d. Then
(1) d is quasisymmetric with respect to .
(2) The identity map from (X, d) to (X, ) is a homeomorphism.
(3) (X, d) is uniformly perfect if and only if (X, ) is uniformly perfect.
(4) (X, d) is bounded if and only if (X, ) is bounded.
(5) Dene d

(x, r) = sup
yB

(x,r)
d(x, y) and
d
(x, r) = sup
yB
d
(x,r)
(x, y). Then
d

(x, r) and
d
(x, r) are nite for any x X and any r > 0.
Those statements, in particular (1) and (3), have been obtained in the original
paper [51].
Proof. (1) Note that (x, z) h(t)(x, y) implies d(x, z) td(x, y). Hence
if h(t)
1
(x, z) > (x, y), then 2t
1
d(x, z) > d(x, y). Set g(s) = 2/h
1
(1/t). Then
g(s) is a homeomorphism from [0, +) to itself and g(s)d(x, z) > d(x, y) whenever
t(x, z) > (x, y). Thus d is QS with respect to .
(2) If
QS
d, then is (SQS)
d
and d is (SQS)

. Now, Proposition 11.2 suces.


(3) There exists (0, 1) such that B
d
(x, r/)\B
d
(x, r) = if B
d
(x, r) = X by
the uniform perfectness. Choose t

(0, 1) so that h(t

) < 1. Suppose B

(x, r) =
X. There exists y X such that (x, y) > r. Let r = t

d(x, y). Since r <


d(x, y), B
d
(x, r) = . Hence there exists y
1
X such that t

d(x, y) d(x, y
1
) <
t

d(x, y). Since


QS
d, we have
1
(x, y) < (x, y
1
) <
2
(x, y), where 0 <

1
= h(2/(t

)) <
2
= h(t

) < 1. In the same way, we have y


2
which satises

1
(x, y
1
) < (x, y
2
) <
2
(x, y
1
). Inductively, we may construct {y
n
}
n1
such that

1
(x, y
n
) < (x, y
n+1
) <
2
(x, y
n
). Choose m so that (x, y
m+1
) < r (x, y
m
).
Then y
m
B

(x, r/
1
)\B

(x, r). Hence (X, ) is uniformly perfect.


(4), (5) Obvious.
By (1) of the above proposition,
QS
is an equivalence relation.
The following theorem relates the notion of begin semi-quasisymmetric with
being quasisymmetric. It has essentially been obtained in [51, Theorem 3.10],
where the notion of uniformly perfect is called homogeneously dense.
Theorem 12.3. Assume that both (X, d) and (X, ) are uniformly perfect.
Then is QS with respect to d if and only if is (SQS)
d
and d is (SQS)

.
Proof. If is QS with respect to d, then it is straightforward to see that
d and are semi-quasisymmetric with respect to each other. Conversely, as-
sume that d and are semi-quasisymmetric with respect to each other. Then
by Proposition 11.3, we may construct homeomorphisms h
1
: [0,
1
] [0,
1
] and
h
2
: [0,
2
] [0,
2
] which satisfy
(i) h
1
(0) = 0, h
2
(0) = 0,
(ii) (x, z) < h
1
()(x, y) whenever d(x, z) < d(x, y) for any (0,
1
] and
(iii) d(x, z) < h
2
()d(x, y) whenever (x, z) < (x, y) for any (0,
2
].
Dene
h
3
(t) =
_
2/(h
2
)
1
(1/t) for t [1/
2
, +),
2/
2
for t [0, 1/
2
].
Then (x, z) < h
3
()(x, y) whenever d(x, z) < d(x, y) for any (0, +). There
is no diculty to nd a homeomorphism h : [0, +) [0, +) with h(0) = 0
13. RELATIONS OF MEASURES AND METRICS 45
which satises that h(t) h
1
(t) for any t [0,
1
] and that h(t) h
3
(t) for any
t [
1
, +). Obviously d(x, z) < td(x, y) implies (x, z) < h(t)(x, y) for any
t > 0. Therefore, is QS with respect to d.
Combining this theorem with Theorem 11.5, we can produce several equivalent
conditions for quasisymmetry under uniform perfectness.
The next corollary is a modied version of Proposition 11.8.
Corollary 12.4. Assume that (X, d) is uniformly perfect and that
QS
d. Let
be a Borel regular measure on (X, d). Then is (VD)
d
if and only if it is (VD)

.
Proof. By Proposition 12.2-(3), (X, ) is uniformly perfect. Hence by Theo-
rem 12.3, is (SQS)
d
and d is (SQS)

. Theorem 11.5 shows that d is doubling with


respect to and is (SQC)
d
. By Proposition 11.8, if is (VD)
d
, then is (VD)

.
The converse follows by exchanging d and .
13. Relations of measures and metrics
To obtain a heat kernel estimate, one often show a certain kind of relations
concerning a measure and a distance. The typical example in the following relation:
(13.1) d(x, y)(B

(x, (x, y))) (x, y)

,
where d(x, y) is the resistance metric (may be written as R(x, y)), is a distance
used in the heat kernel estimate and is a positive exponent. The left hand side
corresponds to the mean exit time from a -ball. We generalize such a relation and
study it in the light of quasisymmetry in the present section.
Throughout this section, d and are distances on a set X which give the
same topology on X. is a Borel regular measure on (X, d). We assume that
0 < (B
d
(x, r)) < + and 0 < (B

(x, r)) < + for any x X and any r > 0.


Notation. We set V
d
(x, r) = (B
d
(x, r)) and V

(x, r) = (B

(x, r)).
Let H : (0, +)
2
(0, +) satisfy the following two conditions:
(H1) if 0 < s
1
s
2
and 0 < t
1
t
2
, then H(s
1
, t
1
) H(s
2
, t
2
),
(H2) for any (a, b) (0, +)
2
, dene
h(a, b) = sup
(s,t)(0,+)
2
H(as, bt)
H(s, t)
.
Then h(a, b) < + for any (a, b) (0, +)
2
and there exists c
0
> 0 such that
h(a, b) < 1 for any (a, b) (0, c
0
)
2
.
Also g : (0, +) (0, +) is a monotonically increasing function satisfying
g(t) 0 as t 0 and the doubling property, i.e. there exists c > 0 such that
g(2t) cg(t) for any t > 0.
Remark. The condition (H1) is the monotonicity of the function H. If (H2)
is satised, then
h(a, b)
n
H(s, t) H(a
n
s, b
n
t)
for any a, b, s, t and n. Since h(a, b) < 1 for (a, b) (0, c
0
)
2
, this inequality yields a
kind of homogeneity of the decay of H(s, t) as (s, t) (0, 0).
We will study several relations between conditions concerning d, , , H and
g.
46
Definition 13.1. (1) We say that the condition (DM1) holds if and only
if there exists : (0, 1] (0, +) such that is monotonically nondecreasing,
(t) 0 as t 0 and
()
g((x, y))
H(d(x, y), V

(x, (x, y)))



g((x, y)))
H(d(x, y), V

(x, (x, y)))


for any x, y X and any (0, 1].
(2) We say that the condition (DM2) holds if and only if
H(d(x, y), V

(x, (x, y))) g((x, y))


for any x, y X.
(3) We say that the condition (DM3) holds if and only if there exist r

> diam(X, )
such that
H(d

(x, r), V

(x, r)) g(r)


for any x X and any r (0, r

].
The relation (DM2) can be seen as a generalization of the above mentioned
relation (13.1), where H(s, t) = st and g(r) = r

. The relation (DM1) looks too


complicated but it is shown to be necessary if d
QS
and (DM2) is satised. See
Corollary 13.3.
Remark. If diam(X, ) = +, then we remove the statement there exists
r

> diam(X, ) such that and replace r (0, r

] by r > 0 in (3) of the above


denition.
In the next section, we are going to construct a distance on X which satises
all three conditions (DM1), (DM2) and (DM3) with g(r) = r

for suciently large


under a certain assumptions. See Theorem 14.1 for details.
The next theorem gives the basic relations. Much clearer description from
quasisymmetric point of view can be found in the corollary below.
Theorem 13.2. Assume that (X, ) is uniformly perfect, that lim
s0
h(s, 1) =
lim
t0
h(1, t) = 0 and that there exists c

> 0 such that (X) c

(x,

(x)) for
any x X, where

(x) = sup
yX
(x, y).
(1) is (VD)

if (DM1) and (DM2) hold.


(2) (DM1) and (DM2) hold if and only if (DM3) holds, d decays uniformly with
respect to and is (ASQC)
d
.
Remark. If diam(X, ) = +, then

(x) = + for any x X. In this case,


we dene B

(x, +) = X and V

(x, +) = (X). Hence letting c

= 1, we always
have (X) c

(x,

(x)).
On the other hand, if diam(X, ) < +, then diam(X, )/2

(x)
diam(X, ). In this case, X = B(x,

(x)).
Corollary 13.3. In addition to the assumptions in Theorem 13.2, suppose
that (X, d) is uniformly perfect. Then the following four conditions are equivalent:
(a) (DM1) and (DM2) hold.
(b)
QS
d and (DM2) holds.
(c)
QS
d and (DM3) holds.
(d) (DM3) holds, d decays uniformly with respect to and is (SQS)
d
.
Moreover, if any of the above conditions is satised, then is (VD)
d
and (VD)

.
13. RELATIONS OF MEASURES AND METRICS 47
The rest of this section is devoted to proving the above theorem and the corol-
lary.
Lemma 13.4. If (DM1) and (DM2) are satised, than, for any > 0, there
exists > 0 such that d(x, z) < d(x, y) whenever (x, z) < (x, y). In particular,
d is (SQS)

.
Proof. Assume that d(x, z) d(x, y) and that (x, z) (x, y). Let =
(x, z)/(x, y). Then by (DM2),
c
2
g(g(x, y)) H(d(x, z), V

(x, (x, z)))


H(d(x, y), V

(x, (x, y)))


h(1/, 1)
1
H(d(x, y), V

(x, (x, y))).


Hence
c
3

g((x, y))
H(d(x, y), V

(x, (x, y)))


,
where c
3
is a positive constant which depends only on . This combined with (DM1)
and (DM2) implies that 0 < c
4
(), where c
4
depends only on . Hence, there
exists > 0 such that (x, z) (x, y). Thus we have the contraposition of the
statement.
Lemma 13.5. Assume (DM1) and that lim
t0
h(1, t) = 0. Then, for any > 0,
there exists a > 0 such that V

(x, (x, y)) aV

(x, (x, y)) for any x, y X.


Remark. Note that h(1, t
n
) h(1, t)
n
. Hence lim
t0
h(1, t) = 0 if and only if
there exists t

(0, 1) such that h(1, t

) < 1.
Proof. For 1, we may choose a = 1. Suppose that (0, 1). Then by
(DM1) and the doubling property of g,
h
_
1,
V

(x, (x, y))


V

(x, (x, y))


_

H(d(x, y), V

(x, (x, y)))


H(d(x, y), V

(x, (x, y)))


()
g((x, y))
g((x, y))
c

> 0,
where c

depends only on . Since lim


t0
h(1, t) = 0, we have the desired conclusion.

Lemma 13.6. Assume that (X, ) is uniformly perfect and that lim
t0
h(1, t) =
0. If d is (SQS)

and (DM2) holds, then, for any suciently small (0, 1), there
exists a > 0 such that V

(x, (x, y)) aV

(x, (x, y)) for any x, y X.


Proof. Since d is (SQS)

, there exist (0, 1) and


0
(0, 1) such that
d(x, z) < d(x, y) whenever (x, z) <
0
(x, y). If
0
, then d(x, z) < d(x, y)
whenever (x, z) < (x, y). Also since (X, ) is uniformly perfect, there exists
c (0, 1) such that B

(x, r)\B

(x, cr) = if B

(x, r) = X. Let x and y X. Then


we may choose z B

(x, r)\B

(x, cr), where r = (x, y). Note that d(x, z) <


48
d(x, y) < d(x, y). By the doubling property of g and (DM2),
c

g((x, z)) c
2
g((x, z)/(c)) c
2
g(r) H(d(x, y), V

(x, r))
H(d(x, z), V

(x, r)) h
_
1,
V

(x, (x, z))


V

(x, (x, y))


_
1
H(d(x, z), V

(x, (x, z)))


c
1
h
_
1,
V

(x, (x, y))


V

(x, (x, y))


_
1
g((x, z)).
Therefore, it follows that
h
_
1,
V

(x, (x, y))


V

(x, (x, y))


_
c
3
.
Since lim
t0
h(1, t) = 0, we have V

(x, (x, y)) aV

(x, (x, y)). Letting = ,


we have the desired statement.
From Lemmas 13.5 and 13.6, we have the following lemma.
Lemma 13.7. Assume that (X, ) is uniformly perfect, that lim
t0
h(1, t) = 0,
and that there exists c

> 0 such that (X) c

(x,

(x)) for any x X. If


either (DM1) is satised or d is (SQS)

and (DM2) is satised, then is volume


doubling with respect to .
Proof. B

(x, r)\B

(x, cr) = unless B

(x, cr) = X by uniform perfectness.


Suppose r <

(x) = sup
yX
(x, y). Choose so that 0 < < c. Then
there exists y X such that cr (x, y) < r. By Lemmas 13.5 and 13.6, we
have V

(x, (x, y)) aV

(x, (x, y)) in either case. This immediately implies


V

(x, r) aV

(x, cr). Therefore, if r < c

(x), then V

(x,

r) aV

(x, r), where

= /c < 1. If diam(X, ) = +, then we have nished the proof. Otherwise,

(x) < + for any x X.


If r [c

(x),

(x)), there exists y X such that r (x, y)

(x).
Lemma 13.5 implies that V

(x, (x, y)) aV

(x, (x, y)) aV

(x, r). Since r/c

(x) (x, y), we have V

(x,

r) V

(x, (x, y)) aV

(x, r).
Finally, let r

(x). Then there exists y X such that

(x)/2 < (x, y)

(x) and V

(x,

(x))/2 V

(x, (x, y)). By Lemma 13.5, V

(x,

(x, y))
a

(x, (x, y)), where a

is independent fo x and y. Hence


a

2
V

(x, r) =
a

2
(X)
a

2
V

(x,

(x)) a

(x, (x, y))


V

(x,

(x, y)) V

(x,

r)

Proof of Theorem 13.2-(1). Combining Lemmas 13.4 and 13.7, we imme-


diately obtain Theorem 13.2-(1).
Lemma 13.4 implies the next fact.
Lemma 13.8. Assume that (X, ) is uniformly perfect and that is volume
doubling with respect to . If (DM1) and (DM2) are satised and lim
s0
h(s, 1) = 0,
then is (ASQC)
d
.
13. RELATIONS OF MEASURES AND METRICS 49
Proof. First we suppose that diam(X, ) < +. Lemma 13.4 implies that d
is (SQS)

. Let r

> diam(X, ). By Proposition 11.7-(1), there exist (0, 1) and


a (0, 1) such that d

(x, r) ad

(x, r) for any x X and any r (0, r

]. Let
r (0, r

]. (Note that d

(x,
n
r) a
n
d

(x, r). Hence can be arbitrarily small.)


Then
(13.2) d

(x, r) = sup{d(x, y)|y B

(x, r)\B

(x, r)}.
Since is (VD)

, there exists > 0 such that V

(x, r) V

(x, r). Now choose


x, y B

(x, r)\B

(x, r). Then (x, y) (x, z) (x, y)/. Therefore,


c
1
g((x, z)) H(d(x, z), V

(x, (x, z))) H(d(x, z), V

(x, (x, y)/)))


H(d(x, z), V

(x, (x, y))) c


2
h(1, )h
_
d(x, z)
d(x, y)
, 1
_
g((x, y)).
This along with the doubling property of g shows that
h
_
d(x, z)
d(x, y)
, 1
_
c
3
> 0,
where c
3
is independent of x, y and z. Since h(s, 1) 0 as s 0, there exists > 0
such that d(x, z) d(x, y). By (13.2), we see that d(x, z) d

(x, r). Hence


B

(x, r)\B

(x, r) B

(x, d

(x, r))\B

(x, d

(x, r)). Next we consider the case


where r > r

. Note that r

diam(X, d)

(x). Hence B

(x, r) = B

(x, r

) = X
and d

(x, r) = d

(x, r

) = d

(x). Also, (x, z) r

for any z X. Therefore


if r (x, y) < r, then r

(x, z) < r

and hence d(x, z) d

(x, r

) =
d

(x, r

). This completes the proof when diam(X, ) < +. Using similar argu-
ment, we immediately obtain the case where diam(X, ) = +.
Lemma 13.9. Assume that (X, ) is uniformly perfect, that is (VD)

and that
is (ASQC)
d
. If (DM2) holds, then, for any r

> diam(X, ),
H(d

(x, r), V

(x, r)) g(r)


for any x X and any r (0, r

]. In particular, (DM3) holds.


Remark. If diam(X, ) = +, then we remove , for any r

> diam(X, ),
and replace r (0, r

] by r > 0 in the statement of the above lemma.


Proof. There exists c (0, 1) such that B

(x, r)\B

(x, cr) = if X =
B

(x, cr) by the uniform perfectness. Recall that

(x) diam(X, )/2 for any


x X. Hence if c
n1
r

< diam(X, )/2, then c


n1
r <

(x) for any r (0, r

].
Therefore X = B

(x, c
n1
r). So, we have y X satisfying c
n
r (x, y) < c
n1
r.
By Proposition 11.10 and (DM2),
H(d

(x, (x, y)), V

(x, (x, y)) g((x, y)).


On the other hand, Lemma 11.9 implies that d is doubling with respect to . This
and the doubling property of show that
H(d

(x, (x, y)), V

(x, (x, y))) H(d

(x, r), V

(x, r)).
Moreover, by the doubling property of g, c
6
g(r) g((x, y)) g(r). Combining
the last three inequalities, we immediately obtain the desired statement.
Lemma 13.10. Assume that (X, ) is uniformly perfect and that is (ASQC)
d
.
Then (DM3) implies (DM2).
50
Proof. By Proposition 11.10,
d

(x, (x, y)) d(x, y).


Hence, letting r = (x, y), we obtain
H(d

(x, r), V

(x, r)) H(d(x, y), V

(x, r))
This immediately implies (DM2).
Lemma 13.11. Assume that (X, ) is uniformly perfect and lim
s0
h(s, 1) = 0.
If is (ASQC)
d
, d decays uniformly with respect to and (DM3) holds, then (DM1)
holds.
Proof. Since d decays uniformly with respect to , there exists c
0
> 0 and
> 0 such that d

(x, r) c
0

(x, r) for any x X and r (0, r

]. (If
diam(X, ) = +, we always replace (0, r

] by (0, +) in this proof.) Let r =


(x, y). By Proposition 11.10 and (DM3),
c
1
g(r) H(d

(x, r), V

(x, r)) H(c


0

(x, r), V

(x, r))
h(c
3

, 1)H(d(x, y), V

(x, r)),
where c
3
> 0 is independent of x and y. Moreover, by Lemma 13.10, we have
(DM2). Hence H(d(x, y), V

(x, r))/g(r) is uniformly bounded. So, there exists


c
4
> 0 such that
c
4
h(c
0

, 1)
H(d(x, y), V

(x, (x, y)))


g((x, y))
g((x, y))
H(d(x, y), V

(x, (x, y)))


for any x and y. Since lim
0
h(c
0

, 1) = 0, (DM1) follows by letting () =


c
4
h(c
0

, 1).
Proof of Theorem 13.2-(2). Assume (DM1) and (DM2). By Lemma 13.8,
is (ASQC)
d
. By Lemma 13.4, d is (SQS)

. This along with Proposition 11.7


implies that d decays uniformly with respect to . Now (DM3) follows by using
Lemma 13.9.
Lemmas 13.10 and 13.11 suce for the converse direction.
Proof of Corollary 13.3. (a) (b) Using Lemma 13.4 and 13.8 and
applying Theorem 11.5, we see that d and are semi-quasisymmetric with respect
to each other. Hence by Theorem 12.3,
QS
d.
(b) (c) Since
QS
d, Theorems 11.5 and 12.3 show that d is (SQS)

and that
is (ASQC)
d
. By Lemma 13.7, is (VD)

. Therefore Lemma 13.9 yields (DM3).


(c) (d) Since
QS
d, Theorems 11.5 and 12.3 show that d is (SQS)

and that
is (ASQC)
d
. Then by Proposition 11.7, d decays uniformly with respect to .
(d) (a) This immediately follows by Theorem 13.2.
14. Construction of quasisymmetric metrics
The main purpose of this section is to construct a distance which satisfy the
conditions (DM1) and (DM2) in Section 13 in the case where g(r) = r

.
In this section, (X, d) is a metric space and is a Borel regular measure on
(X, d) which is volume doubling with respect to d. We also assume that 0 <
14. CONSTRUCTION OF QUASISYMMETRIC METRICS 51
(B
d
(x, r)) < + for any x X and r > 0. Let H : (0, +)
2
(0, +) satisfy
(H1) and (H2) as in Section 13.
Theorem 14.1. Assume that (X, d) is uniformly perfect and that is (VD)
d
.
For suciently large > 0, there exists a distance on X such that
QS
d and
(DM3) holds with g(r) = r

.
Remark. If
QS
d and (X, d) is uniformly perfect, then Proposition 12.2 and
Corollary 12.4 imply that (X, ) is uniformly perfect and that is (VD)

.
Our distance satises the condition (c) of Corollary 13.3. If lim
s0
h(s, 1) =
lim
t0
h(1, t) = 0, then we have all the assumption of the corollary and hence obtain
the statements (a) through (d). In particular, d, and satisfy (DM1) and (DM2)
with g(r) = r

. In particular, letting H(s, t) = st, we establish the existence of a


distance which is quasisymmetric to the resistance metric and satises (13.1) if is
volume doubling with respect to the resistance metric. This fact plays an important
role in the next part.
Example 14.2. (1) If H(s, t) = t, (DM3) is
(B

(x, r)) r

.
Hence in this case, Theorem 14.1 implies the following well-known theorem: if
(X, d) is uniformly perfect and is (VD)
d
, then there exists a metric such that
the metric measure space (X, , ) is Ahlfors regular. ((X, , ) is called a metric
measure space if and only if (X, ) is a metric space and is a Borel-regular measure
on (X, ). A metric measure space (X, , ) is called Ahlfors -regular if and only
if (B

(x, r)) r

for any x X and any r diam(X, d). ) See Heinonen [32,


Chapter 14] and Semmes [48, Section 4.2] for details.
(2) Let F : (0, +) (0, +) be monotonically nondecreasing. Suppose that
there exist positive constants c
1
,
1
and
2
such that
F(xy) c
1
max{x

1
, x

2
}F(y)
for any x, y (0, +). Dene H(s, t) = F(s
p
t
q
). If p 0, q 0 and (p, q) = (0, 0),
then H satises (H1) and (H2). In fact,
H(as, bt) = F(a
p
b
q
s
p
t
q
) c
1
max{(a
p
b
q
)

1
, (a
p
b
q
)

2
}H(s, t).
Hence h(a, b) c
1
max{(a
p
b
q
)

1
, (a
p
b
q
)

2
}.
To prove Theorem 14.1, we need several preparations.
Notation. Dene v(x, y) = V
d
(x, d(x, y)) +V
d
(y, d(x, y)). Also dene
(x, y) =
_
H(d(x, y), v(x, y)) if x = y,
0 otherwise.
Note that (x, y) = (y, x) 0 and that (x, y) = 0 implies x = y.
Hereafter in this section, we always assume that (X, d) is uniformly perfect and
that is (VD)
d
. By the volume doubling property, we have the following lemma.
Lemma 14.3. For any x, y X,
v(x, y) V
d
(x, d(x, y)).
52
Lemma 14.4. Dene
f

1
,
2
(t) =
_
t

1
if t (0, 1),
t

2
if t 1.
Then there exist positive constants c
1
,
1
and
2
such that
V
d
(x, d(x, y)) c
1
f

1
,
2
()V
d
(x, d(x, y))
for any x, y X and any > 0.
Proof. If 1, this is immediate from the volume doubling property. Since
(X, d) is uniformly perfect, there exists c (0, 1) such that B
d
(x, r) = X implies
B
d
(x, r)\B
d
(x, cr) = . Let r = d(x, y). Choose z B
d
(x, r/2)\B
d
(x, cr/2). It
follows that V
d
(x, cr/4) + V
d
(z, cr/4) V
d
(x, r). Now by the volume doubling
property, V
d
(z, cr/4) aV
d
(x, cr/4), where a is independent of x, z and r. Hence
V
d
(x, cr/4) (1 +a)
1
V
d
(x, r). This shows the desired inequality when (0, 1).

Lemma 14.5. There exists a homeomorphism : [0, +) [0, +) such that


(0) = 0 and (x, z) < (t)(x, y) whenever d(x, z) < td(x, y).
Proof. Assume that d(x, z) < td(x, y). Write f = f

1
,
2
. Then by (H1), (H2)
and the above lemmas,
(x, z) = H(d(x, z), v(x, z)) H(td(x, y), Mc
1
f(t)V
d
(x, d(x, y)))
H(td(x, y), M
2
c
1
f(t)v(x, y)) h(1, M
2
c
1
)h(t, f(t))H(d(x, y), v(x, y)).
By the denition of h(a, b), it follows that h(t, f(t)) is monotonically nondecreasing.
Also if t < c
0
, then h(t, t

1
) < 1. Since h(t
n
, t
n
1
) h(t, t

1
)
n
for n 0, we see that
h(t, f(t)) 0 as t 0. Therefore, there exists a homeomorphism : [0, +)
[0, +) such that (0) = 0 and (t) h(1, M
2
c
1
)h(t, f(t)) for any t > 0.
Definition 14.6. f : X X [0, +) is called a quasidistance on X if and
only if f satises the following three conditions:
(QD1) f(x, y) 0 for any x, y X. f(x, y) = 0 if and only if x = y.
(QD2) f(x, y) = f(y, x) for any x, y X.
(QD3) There exists K > 0 such that f(x, y) K(f(x, z)+f(z, y)) for any x, y, z
X.
Lemma 14.7. (x, y) is a quasidistance.
Proof. Since d(x, y) d(x, z) +d(z, y), either d(x, y) d(x, z)/2 or d(x, y)
d(z, y). Assume that d(x, y) d(x, z)/2. Then Lemma 14.5 implies that (x, y)
(1/2)(x, z) (1/2)((x, z) +(z, y)).
Lemma 14.8. If f : X X [0, +) is a quasidistance on X, then there
exists
0
> 0 such that f

is equivalent to a distance for any (0,


0
], i.e.
f(x, y)

(x, y)
for any x, y X, where

is a distance on X.
See Heinonen [32, Proposition 14.5] for the proof of this lemma.
14. CONSTRUCTION OF QUASISYMMETRIC METRICS 53
Lemma 14.9. For suciently large > 0, there exists a distance on X such
that
QS
d and
(14.1) (x, y) (x, y)

for any x, y X.
Proof. By Lemmas 14.7 and 14.8, if is large enough, then there exists
a distance which satises (14.1). By Lemma 14.5, d(x, z) < td(x, y) implies
(x, z) < c(t)
1/
(x, y) for some c > 0. Hence
QS
d.
Since
QS
d, d and dene the same topology on X. Also since (X, d) is
uniformly perfect, so is (X, ). Then, Theorem 12.3 shows that d and are semi-
quasisymmetric with respect to each other. So we may enjoy the results in Theo-
rem 11.5 in the rest of discussions.
Lemma 14.10. For any x X and any r > 0,
(14.2) V

(x, r) V
d
(x, d

(x, r)).
Proof. Since is (SQC)
d
,
B
d
(x, cd

(x, r)) B

(x, r) B
d
(x, c

(x, r)).
This and the volume doubling property of imply (14.2).
Proof of Theorem 14.1. The rest is to show (DM3). Since (X, ) is uni-
formly perfect, there exists c (0, 1) such that B

(x, r)\B

(x, cr) = unless


B

(x, r) = X. We will consider the case when diam(X, ) < +. Let r

>
diam(X, ). Choose n 1 so that c
n
r < diam(X, )/2. Note that diam(X, )/2

(x). Hence if r (0, r

], then c
n
r <

(x). Therefore there exists y X such that


c
n+1
r (x, y) < c
n
r. By (ASQC)
d
, there exists > 0 such that d(x, z) d

(x, r)
for any r > 0 and any z B

(x, r)\B

(x, cr). This along with the doubling prop-


erty of d

(x, r) implies that d

(x, r) d(x, y) d

(x, c
n
r) c

(x, r). Then by


Lemma 14.3. the volume doubling property of , we have (14.1) and (14.2),
(c
n
r)

(x, y)

c
3
H(d(x, y), c
4
V
d
(x, d(x, y)))
c
3
H(c

(x, r), c
5
V
d
(x, d

(x, r))) c
6
H(d

(x, r), V

(x, r)).
Similarly,
(c
n+1
r)

(x, y)

c
7
H(d(x, y), c
8
V
d
(x, d(x, y)))
c
7
H(d

(x, r), V
d
(x, d

(x, r))) c
8
H(d

(x, r), V

(x, r)).
Thus we obtain (DM3) if diam(X, ) < +. The other case follows by almost the
same argument.
Part 3
Volume doubling measures and
heat kernel estimates
In this part, we will show results on heat kernel estimates, which answer the
two questions in the introduction, under the foundation laid by the previous two
parts. The rst question is how and when we can nd a metric which is suitable
for describing asymptotic behaviors of a heat kernel. As an answer, Theorem 15.11
shows that if the underlying measure is volume doubling with respect to the resis-
tance metric, then we can get a good (on-diagonal, at least) heat kernel estimate
by quasisymmetric modication of the resistance metric. The second question con-
cerns jumps. Namely, what kind of jumps can we allow to get a good heat kernel
estimate? Theorem 15.11 also gives an answer to this question, saying that the an-
nulus comparable condition (with the volume doubling property) is sucient and
necessary for a good heat kernel estimate.
15. Main results on heat kernel estimates
In this section, we present the main results on heat kernel estimates. There
will be three main theorems, 15.6, 15.10 and 15.11. The rst one gives a good (on-
diagonal and lower near diagonal) heat kernel estimate if (ACC) holds, the measure
is (VD)
R
and the distance is quasisymmetric with respect to the resistance metric
R. The second one provides geometrical and analytical equivalent conditions for
having a good heat kernel estimate. As an application, we will recover the two-
sided o-diagonal heat kernel estimates for the Brownian motions on homogeneous
random Sierpinski gaskets obtained by Barlow and Hambly in [7]. See Section 24
for details. Finally in the third theorem, (VD)
R
and (ACC) ensure the existence
of a distance d which is quasisymmetric to R and under which a good heat kernel
estimate holds.
Proofs of Propositions and Theorems in this section are given in latter sections
in this part.
Throughout this section, (E, F) is a regular resistance form on a set X and
R is the associated resistance metric on X. We assume that (X, R) is separable,
complete, uniformly perfect and locally compact. Let be a Borel regular measure
on (X, R) which satises 0 < (B
R
(x, r)) < + for any x X and any r > 0.
Under those assumptions, if D is the closure of F C
0
(X) with respect to the
E
1
-norm, then (E, D) is a regular Dirichlet form L
2
(X, ). Let ({X
t
}
t>0
, {P
x
}
xX
)
be the Hunt process associated with the regular Dirichlet form (E, D) on L
2
(X, ).
As we have shown in Section 7, if (X, R) is complete and is (VD)
R
, then
B
R
(x, r) is compact for any x X and any r > 0. Hence under (VD)
R
, Theo-
rem 10.4 implies the existence of a jointly continuous heat kernel (i.e. transition
density) p(t, x, y) associated with the Dirichlet form (E, D) on L
2
(X, ).
Definition 15.1. Let d be a distance on X giving the same topology as R.
Dene R
d
(x, r) = sup
yB
d
(x,r)
R(x, y), V
d
(x, r) = (B
d
(x, r)) and
h
d
(x, r) = R
d
(x, r)V
d
(x, r)
for any r > 0 and any x X.
Lemma 15.2. For each x X, R
d
(x, r) and V
d
(x, r) are monotonically non-
decreasing left-continuous function on (0, +). Moreover lim
r0
R
d
(x, r) = 0 and
lim
r0
V
d
(x, r) = ({x}).
By the above lemma, h
d
(x, r) is monotonically nondecreasing left-continuous
function on (0, +) and lim
r0
h
d
(x, r) = 0.
15. MAIN RESULTS ON HEAT KERNEL ESTIMATES 57
Definition 15.3. Let d be a distance on X which gives the same topology as
R. We say that the Einstein relation with respect d, (EIN)
d
for short, holds if and
only if
(EIN)
d
E
x
(
B
d
(x,r)
) h
d
(x, r),
for any x X and r > 0 with X = B
d
(x, r).
The name Einstein relation has been used by several authors . See [26] and
[50] for example.
We have two important equivalences between the resistance estimate, the an-
nulus comparable condition and the Einstein relation.
Proposition 15.4. Assume that d is a distance on X and d
QS
R. Then (RES)
is equivalent to
(15.1) R(x, B
d
(x, r)
c
) R
d
(x, r)
for any x X and any r > 0 with B
d
(x, r) = X.
Proposition 15.5. Assume that is (VD)
R
, that d is a distance on X and
that d
QS
R. Then (RES), (ACC) and (EIN)
d
are equivalent to one another.
The proofs of the above propositions are in Section 18.
Now we have the rst result on heat kernel estimate.
Theorem 15.6. Assume that (ACC) holds. Suppose that has volume doubling
property with respect to R. Then, there exists a jointly continuous heat kernel
p(t, x, y) associated with the Dirichlet form (E, D) on L
2
(X, ). Moreover, if a
distance d on X is quasisymmetric with respect to R, then (EIN)
d
holds and
(15.2)
c
1
V
d
(x, r)
p(h
d
(x, r), x, y)
and
(15.3) p(h
d
(x, r), x, x)
c
2
V
d
(x, r)
for any x X, any r > 0 and any y X with d(x, y) c
3
min{r, diam(X, d)}.
(15.2) will imply (15.5), which is the counterpart of the near diagonal lower
estimate (NDL)
,d
. In the case of diusions, if the distance is not geodesic, the
lower near diagonal estimate is known as a substitute for the o-diagonal lower
sub-Gaussian estimate.
Note that R
QS
R and R
d
(x, r) r if (X, R) is uniformly perfect. Hence,
h
d
(x, r) = rV
R
(x, r) and the above theorem shows
p(rV
R
(x, r), x, x)
1
V
R
(x, r)
.
This has essentially been obtained in [41].
To state the next theorem, we need several notions and results on monotonically
non-decreasing functions on (0, +) and their inverse.
Definition 15.7. Let f : (0, +) (0, +).
(1) f is said to be doubling if there exists c > 0 such that f(2t) cf(t) for any
t (0, +).
58
(2) f is said to decay uniformly if and only if there exists (, ) (0, 1)
2
such that
f(t) f(t) for any t (0, +).
(3) f is said to be a monotone function with full range if and only if f is mono-
tonically non-decreasing, lim
t0
f(t) = 0 and lim
t+
f(t) = +. For a monotone
function with full range on (0, +), we dene f
1
(y) = sup{x|f(x) y} and call
f
1
the right-continuous inverse of f.
Lemma 15.8. Let f : (0, +) (0, +) be a monotone function with full
range. (1) If f is doubling, then f
1
decays uniformly and f(f
1
(y)) y for any
y (0, +).
(2) If f decays uniformly, then f
1
is doubling and f
1
(f(x)) x for any x
(0, +).
This lemma is rather elementary and we omit its proof.
The following denition is a list of important relations or properties between a
heat kernel, a measure and a distance.
Definition 15.9. Let d be a distance of X giving the same topology as R and
let g : (0, +) (0, +) be a monotone function with full range.
(1) A heat kernel p(t, x, y) is said to satisfy on-diagonal heat kernel estimate of
order g with respect to d, (DHK)
g,d
for short, if and only if
p(t, x, x)
1
V
d
(x, g
1
(t))
for any x X and any t > 0, where g
1
is the right-continuous inverse of g.
(2) A heat kernel p(t, x, y) is said to have the doubling property. (KD) for short,
if and only if there exists c
1
> 0 such that
p(t, x, x) c
1
p(2t, x, x).
for any x X and any t > 0.
(3) We say that (DM1)
g,d
holds if and only if there exists : (0, 1] [0, +) such
that is monotonically nondecreasing, lim
t0
(t) = 0 and
g(d(x, y))
V
d
(x, d(x, y))

g(d(x, y))
V
d
(x, d(x, y))
()
for any x, y X and any (0, 1].
(4) We say that (DM2)
g,d
holds if and only if
R(x, y)V
d
(x, d(x, y)) g(d(x, y))
for any x, y X.
The conditions (DM1)
g,d
and (DM2)
g,d
corresponds to (DM1) and (DM2) with
H(s, t) = st respectively. (DM2)
g,d
is the counterpart of R() introduced dened in
[6] and introduced in the introduction. It relates the mean exit time with g(d(x, y))
through (EIN)
d
.
Remark. If diam(X, d) is bounded, it is enough for g to be only dened on
(0, diam(X, d)), for example, to describe (DM2)
g,d
. In such a case, the value of g for
[diam(X, d), +) does not make any essential dierences. One can freely extend
g : (0, diam(X, d)) (0, +) to g : (0, +) (0, +) so that g satises required
conditions as being doubling, decaying uniformly or being strictly monotone.
15. MAIN RESULTS ON HEAT KERNEL ESTIMATES 59
Here is our second theorem giving equivalent conditions for a good heat kernel.
Theorem 15.10. Assume that (ACC) holds. Let d be a distance on X giving
the same topology as R and let g : (0, +) (0, +) be a monotone function with
full range and doubling. Then the following statements (a), (b), (c) and (HK)
g,d
are equivalent.
(a) (X, d) is uniformly perfect, (DM1)
g,d
and (DM2)
g,d
hold.
(b) d
QS
R and (DM2)
g,d
holds.
(c) d
QS
R and, for any x X and any r diam(X, d),
(15.4) h
d
(x, r) g(r)
(HK)
g,d
d
QS
R, g decays uniformly, a jointly continuous heat kernel p(t, x, y)
associated with the Dirichlet form (E, D) on L
2
(X, ) exists and satises (KD) and
(DHK)
g,d
.
Moreover, if any of the above conditions holds, then there exist positive constants c
and c

such that
(15.5)
c

V
d
(x, g
1
(t))
p(t, x, y)
for any y B
d
(x, cg
1
(t)). Furthermore, assume that (r) = g(r)/r is a monotone
function with full range and decays uniformly. We have the following o-diagonal
estimates:
Case 1: If (E, F) has the local property, then
(15.6) p(t, x, y)
c
1
V
d
(x, g
1
(t))
exp
_
c
2
_
d(x, y)

1
(t/d(x, y))
_
_
for any x, y X and any t > 0, where c
1
, c
2
> 0 are independent of x, y and t.
Case 2: Assume that d(x, y) has the chain condition, i.e. for any x, y X and any
n N, there exist x
0
, . . . , x
n
such that x
0
= x, x
n
= y and d(x
i
, x
i+1
) Cd(x, y)/n
for any i = 0, . . . , n 1, where C > 0 is independent of x, y and n. Then,
(15.7)
c
4
V
d
(x, g
1
(t))
exp
_
c
5
_
d(x, y)

1
(t/d(x, y))
_
_
p(t, x, y)
for any x, y X and any t > 0, where c
3
, c
4
> 0 are independent of x, y and t.
Remark. Note that if d = R, then the above theorem says (DM2)
g,R
implies
(DM1)
g,R
. Moreover, in this case, (DM2)
g,R
shows the uniform volume doubling
property given by Kumagai in [41]. A Borel regular measure on a metric space
(X, d) is said to satisfy uniform volume doubling property if and only if there exists
a doubling function f such that
(B
d
(x, r)) f(r)
for any x X and any r > 0. In fact, he has shown the above theorem in this special
case including the o-diagonal estimates when (E, F) satises the local property.
The above theorem is useful to show a heat kernel estimate for a specic exam-
ple. In the next section, we will apply this theorem to (traces of) -stable processes
60
on R for (1, 2]. Also, in Section 24, we will apply (15.6) and (15.7) to homoge-
neous random Sierpinski gaskets and recover the o-diagonal heat kernel estimate
obtained by Barlow and Hambly in [7].
The next theorem assures the existence of a distance d which satises the
conditions in Theorem 15.10 for certain g if is (VD)
R
and (ACC) holds.
Theorem 15.11. Assume that (X, R) is uniformly perfect. Then the following
conditions (C1), (C2), ..., (C6) are equivalent.
(C1) (ACC) holds and is (VD)
R
.
(C2) is (VD)
R
and (EIN)
R
holds.
(C3) (ACC) holds and there exist a distance d on X and > 1 such that (HK)
g,d
with g(r) = r

is satised.
(C4) There exist a distance d on X and > 1 such that d
QS
R and
(15.8) E
x
(
B
d
(x,r)
) r

h
d
(x, r)
for any x X and any r > 0 with B
d
(x, r) = X.
(C5) is (VD)
R
. If d is a distance on X and d
QS
R, then (EIN)
d
holds.
(C6) is (VD)
R
. There exists a distance d on X such that d
QS
R and (EIN)
d
holds.
Moreover, if any of the above conditions holds, then we can choose the distance
d in (C3) and (C4) so that
(15.9) d(x, y)

R(x, y)(V
R
(x, R(x, y)) +V
R
(y, R(x, y)))
for any x, y X.
Both the volume doubling property and the Einstein relation are known to be
necessary to obtain a good heat kernel estimate. Hence the implication (C2) (C3)
shows that (ACC) is also necessary to get a reasonable two-sided diagonal heat
kernel estimate.
Remark. By Theorem 7.12, we may replace (ACC) by (RES) in (C1) and
(C3).
We have a simpler statement in the local case. Recall that (X, R) is assumed
to be uniformly perfect. Using Corollary 7.13, we have the next corollary.
Corollary 15.12. Assume that (E, F) has the local property. Then the fol-
lowing conditions (C1) and (C3) are equivalent:
(C1) is (VD)
R
.
(C3) There exist a distance d on X and > 1 such that (HK)
g,d
with g(r) = r

holds.
Moreover, if any of the above conditions is satised, then we have the coun-
terpart of the near diagonal lower estimate, (15.5), and o-diagonal sub-Gaussian
upper estimate (15.6).
Next we apply the above theorems to the Dirichlet form associated with a
trace of a resistance form (E, F) on X. Since (E, F) is assumed to be regular,
(E|
Y
, F|
Y
) is a regular resistance form by Theorem 8.4. Let Y be a closed subset of
X which is uniformly perfect. Assume that (E, F) satises (RES). By Theorem 8.6,
(E|
Y
, F|
Y
) satises (RES) as well. Recall that R|
Y
is the restriction of R to Y Y
and it coincides with the resistance metric associated with (E|
Y
, F|
Y
). Let be a
16. EXAMPLE: THE -STABLE PROCESS ON R 61
Borel regular measure on (Y, R|
Y
) which satisfy 0 < (B
R
(x, r) Y ) < +for any
x Y and r > 0. If is (VD)
R|
Y
, then (ACC) for (Y, R|
Y
) follows by Theorem 7.12.
Therefore, the counterpart of Theorems 15.6, 15.10 and 15.11 hold for (E|
Y
, F|
Y
)
with (ACC) granted. (Note that the conditions (a), (b), (c) and (HK)
g,d
imply the
volume doubling property.) In particular, we have the following result.
Theorem 15.13. Let be a Borel regular measure on (X, R) that satises
0 < (B
R
(x, r)) < + for any x X and any r > 0. Assume that (ACC)
holds for (E, F) and that there exists a distance d on X such that d
QS
R and
(HK)
g,d
with g(r) = r

is satised. Let Y be a non-empty closed subset of X and


let (E|
Y
, D|
Y
) be the regular Dirichlet form on L
2
(Y, ) induced by (E|
Y
, F|
Y
). If
(Y, R|
Y
) is uniformly perfect and there exist > 0 and a Borel regular measure
on (Y, R|
Y
) such that
(15.10) (B
d
(x, r)) r

(B
d
(x, r) Y )
for any x Y and any r > 0 with B
d
(x, r) = X, then it follows that > , that
there exists a jointly continuous heat kernel p
Y

(t, x, y) associated with the regular


Dirichlet form (E|
Y
, D|
Y
) on L
2
(Y, ) and that
(15.11) p
Y

(t, x, x)
1
(B
d
(x, t
1/()
) Y )
for any x X and any t > 0. In particular, if (B
d
(x, r)) r

for any x X and


any r > 0 with B
d
(x, r) = X, then
(15.12) p
Y

(t, x, x) t

for any x Y and any t > 0 with B


d
(x, t
1/()
) = X.
If (B
r
(x, d)) r

, then the Hausdor dimension of (X, d) is and (A)


H

(A) for any Borel set A, where H

is the -dimensional Hausdor measure of


(X, d). Hence, (X, d) is Ahlfors -regular set. In such a case, (15.10) implies that
(Y, d|
Y
) is Ahlfors ( )-regular set.
We will apply the above theorem for the traces of the standard resistance form
on the Sierpinski gasket in Example 20.11.
16. Example: the -stable process on R
In this section, we will apply the results in the last section to the resistance
form associated with the -stable process on R for (1, 2]. For = 2, the
-stable process is the Brownian motion on R. We denote the Euclidean distance
on R by d
E
.
Definition 16.1. (1) For (0, 2), dene
(16.1) F
()
=
_
u

u C(R),
_
R
2
(u(x) u(y))
2
|x y|
1+
dxdy < +
_
and
(16.2) E
()
(u, v) =
_
R
2
(u(x) u(y))(v(x) v(y))
|x y|
1+
dxdy
62
for any u, v F
()
. Moreover, dene D
()
= F
()
L
2
(R, dx).
(2) For = 2, dene
F
(2)
0
= {u|u C
1
(R),
_
R
(u

(x))
2
dx < +}
and
E
(2)
(u, v) =
_
R
u

(x)v

(x)dx
for any u, v F
(2)
0
.
For = 2, (E
(2)
, F
(2)
0
) does not satisfy (RF2). To make a resistance form, we
need to take a kind of closure of (E
(2)
, F
(2)
0
). The next proposition is elementary.
We omit its proof.
Proposition 16.2. If {u
n
}
n1
F
(2)
0
satises E
(2)
(u
n
u
m
, u
n
u
m
) 0 as
n, m + and u
n
(0) a as n + for some a R, then {u
n
}
n1
converges
compact uniformly to u C(R) as n +.
Definition 16.3. We use F
(2)
to denote the collection of all the limits u in
the sense of Proposition 16.2. Dene
E
(2)
(u, v) = lim
n+
E
(2)
(u
n
, v
n
)
for any u, v F
(2)
, where {u
n
}
n1
and {v
n
}
n1
are the sequences convergent to u
and v respectively in the sense of Proposition 16.2. Also set D
(2)
= F
(2)
L
2
(R, dx).
It is well-known that, for (0, 2], (E
()
, D
()
) is a regular Dirichlet form
on L
2
(R, dx) and the associated non-negative self-adjoint operator on L
2
(R, dx) is
()
/2
, where = d
2
/dx
2
is the Laplacian. The corresponding Hunt process is
called the -stable process on R. See [40, 15] for example. Note that (E
()
, F
()
)
has a natural scaling property. More precisely, for u F
()
, dene u
t
(x) = u(tx)
for any t > 0. Then,
E
()
(u
t
, u
t
) = t
1
E
()
(u, u)
for any t > 0. Combining this scaling property with [24, Theorem 8.1], we have
the following.
Proposition 16.4. For (1, 2], (E
()
, F
()
) is a regular resistance form on
R. The corresponding resistance metric R
()
(x, y) =

|xy|
1
for any x, y R,
where

is independent of x and y.
By this proposition, for (1, 2], if D
()

= L
2
(R, ) F
()
, then (E
()
, D
()

)
is a regular Dirichlet form on L
2
(R, ) for any Radon measure on R.
Theorem 16.5. (E
()
, F
(a)
) satises the annulus comparable condition (ACC)
for (1, 2].
Proof. By the scaling property with the invariance under parallel translations,
there exist c
1
, c
2
> 0 such that
R
()
(x, B(x, r)
c
) = c
1
r
R
()
(x, B(x, r)
c
B(x, 2r)) = c
2
r,
where B(x, r) = B
R
() (x, r). Now, it is obvious that (ACC) holds.
16. EXAMPLE: THE -STABLE PROCESS ON R 63
Due to this theorem, we can apply Theorems 15.6 and 15.11 to get an estimate
of the heat kernel associated with the Dirichlet form (E
()
, D
()

) on L
2
(R, ) if
has the volume doubling property with respect to the Euclidean distance. (Note
that R
()
is a power of the Euclidean distance.) As a special case, we have the
following proposition.
Proposition 16.6. Dene p
()

(t, x, y) as the heat kernel associated with the


Dirichlet form (E
()
, D
()
) on L
2
(R, x

dx) for > 1. For (1, 2],


p
()

(t, 0, 0) t

+1
+
for any t > 0.
Recall that ()
/2
is the associated self-adjoint operator for = 0. Hence
p
()
0
(t, x, y) = P

(t, |x y|), where P

(t, ) is the inverse Fourier transform of


e
ct|x|

for some c > 0. This immediately imply that p


()
0
(t, x, x) = a/t
1/
for
some a > 0.
Next we consider the trace of (E
()
, F
()
) on the Cantor set. Let K be the
Cantor set, i.e.
K =
_

m=1
i
m
3
m

i
1
, i
2
, . . . {0, 2}
_
.
The Hausdor dimension h of (K, d
E
) is log 2/ log 3. Let be the h-dimensional
normalized Hausdor measure. Dene
K
i
1
...i
m
=
_
m

k=1
i
k
3
k
+
1
3
m

n=1
j
n
3
n

j
1
, j
2
, . . . {0, 2}
_
for any i
1
, . . . , i
m
{0, 2}. Then (K
i
1
...i
m
) = 2
m
. Hence (B(x, r)) r
h
for
any r [0, 1] and any x K. It is easy to see that has the volume doubling
property with respect to d
E
. Also (K, d
E
|
K
) is uniformly perfect. Recall that
R
()
=

(d
E
)
1
. Also we have
(B
d
E
(x, r)) r
1h
(B
d
E
(x, r) K)
Since K is compact and any u F
()
is continuous, F
()
|
K
L
2
(K, ). Hence
(E
()
|
K
, F
()
|
K
) is a regular Dirichlet form on L
2
(K, ).
Using Theorem 15.13, we have the following result.
Theorem 16.7. Let (1, 2]. There exists a jointly continuous heat kernel
p
()
K
(t, x, y) on (0, +) K
2
associated with the Dirichlet form (E
()
|
K
, F
()
|
K
)
on L
2
(K, ). Moreover,
(16.3) p
()
K
(t, x, x) t

for any t (0, 1] and any x K, where =


log 2
(1) log 3+log 2
.
If = 2, the process associated with (E
(2)
|
K
, F
(2)
|
K
) on L
2
(K, ) is called
the generalized diusion on the Cantor set. Fujita has studied the heat kernel
associated with the generalized diusion on the Cantor set extensively in [20]. He
has obtained (16.3) for this case by a dierent method.
64
17. Basic tools in heat kernel estimates
The rest of this part is devoted to proving the theorems in Section 15. In this
section, we review the general methods of estimates of a heat kernel and make
necessary modications to them. The results in this section have been developed
by several authors, for example, [1], [38] and [22].
In this section, (X, d) is a metric space and (E, D) is a regular Dirichlet form
on L
2
(X, ), where is a Radon measure on X. (We do not assume that (E, D) is
derived from a resistance form.) We assume that there exists a jointly continuous
heat kernel (i.e. transition density) p(t, x, y) associated with this Dirichlet form.
First we introduce a result on diagonal-lower estimate of a heat kernel. The
Chapman-Kolmogorov equation implies the following fact.
Lemma 17.1. For any Borel set A X, any t > 0 and any x X,
P
x
(X
t
A)
2
(A)
p(2t, x, x).
The next lemma can be extracted from [22, Proof of Theorem 9.3].
Lemma 17.2. Let h : X (0, +) [0, +) satisfy the following conditions
(HKA), (HKB) and (HKC):
(HKA) For any x X, h(x, r) is a monotonically nondecreasing function of r and
lim
r0
h(x, r) = 0.
(HKB) There exists a
1
> 0 such that h(x, 2r) a
1
h(x, r) for any x X and any
r > 0.
(HKC) There exists a
2
> 0 such that h(x, r) a
2
h(y, r) for any x, y X with
d(x, y) r.
Assume that there exist positive constants c
1
, c
2
and r

(0, +){+} such


that
(17.1) c
1
h(x, r) E
x
(
B
d
(x,r)
) c
2
h(x, r)
for any x X and any r (0, r

]. Then,
(1) There exist (0, 1) and c > 0 such that
(17.2) E
x
(e

B
d
(x,r)
)
whenever h(x, r) c and r (0, r

/2].
(2) For any r (0, r

/2] and any t > 0,


(17.3) P
x
(
B
d
(x,r)
t) e
ct
h(x,r)
,
where and c are the same as in (1).
Combining the above two lemmas, we immediately obtain the following theo-
rem.
Theorem 17.3. Under the same assumptions of Lemma 17.1, there exist pos-
itive constants and such that

(B
d
(x, r))
p(h(x, r), x, x)
for any x X and any r (0, r

/2]. Moreover, if has the volume doubling


property with respect to d and h(x, r) h(x, r) for any x X and any r (0, r

],
17. BASIC TOOLS IN HEAT KERNEL ESTIMATES 65
where and belong to (0, 1) and are independent of x and r, then there exist

> 0
and c

(0, 1) such that

(B
d
(x, r))
p(h(x, r), x, x)
for any x X and r (0, c

].
Next we give a result on o-diagonal upper estimate.
Hereafter, h(x, r) is assumed to be independent on x X. We write h(r) =
h(x, r).
The following line of reasoning has essentially been developed in the series of
papers by Barlow and Bass [2, 3, 4]. It has presented in [1] in a concise and
organized manner. Here we follow a sophisticated version in [22]. Generalizing the
discussion in [22, Proof of Theorem 9.1], we have the following lemma.
Lemma 17.4. Let (E, D) be local. Also let h : (0, +) (0, +) be a mono-
tone function with full range, continuous, strictly increasing and doubling.
(1) If there exist (0, 1) and c > 0 such that (17.2) holds for any r (0, r

] and
any x X with h(r) c, then, for any q > 0,
(17.4) E
x
(e

B
d
(x,r)
) c
1
exp
_
c
2
r
h
1
(c/)
_
for any > 0 and any r (0, qr

], where c
1
=
2 max{1,q}
and c
2
= log .
(2) Moreover, assume that (r) = h(r)/r is a monotone function with full range
and strictly increasing. If (17.4) holds for any > 0 and any r (0, R], then, for
any (0, 1), any t > 0 and any r (0, R],
(17.5) P
x
(
B
d
(x,r)
t) c
1
exp
_

c
3
r

1
(c
4
t/r)
_
,
where c
3
= c
2
(1 ) and c
4
= c/(c
2
).
Remark. The local property of a Dirichlet form is equivalent to that the
associated Hunt process is a diusion. See [21, Theorem 4.5.1] for details.
Proof. (1) First assume that r/h
1
(c/) 2 and r

/h
1
(c/) 2. Then
there exists n N such that
(17.6)
r
h
1
(c/)
n
r
2h
1
(c/)

r
r

.
If n is the maximum natural number satisfying (17.6), then n r/h
1
(c/) 1.
Since h(r/n) c and r/n (0, r

], argument in [22, Proof of Theorem 9.1] works


and implies
(17.7) E
x
(e

B
d
(x,r)
)
n

exp
_

c
2
r
h
1
(c/)
_
.
If r/h
1
(x/) 2, then
(17.8) E
x
(e

B
d
(x,r)
) 1
1

2
exp
_

c
2
r
h
1
(c/)
_
.
Finally if r

/h
1
(c/) 2, then r/h
1
(c/) qr

/h
1
(x/) 2q for any r
(0, qr

]. Hence
(17.9) E
x
(e

B
d
(x,r)
) 1
1

2q
exp
_

c
2
r
h
1
(c/)
_
.
66
Combining (17.7), (17.8) and (17.9), we obtain the desired inequality.
(2) By [22, Proof of Theorem 9.1],
(17.10) P
x
(
B
d
(x,r)
t) e
t
E
x
(e

B
d
(x,r)
) c
1
exp
_
t
c
2
r
h
1
(c/)
_
for any t > 0, any > 0 and any r (0, R]. Let =
c
2
r
t
1
(ct/(c
2
r))
. Then
t =
c
2
r
h
1
(c/)
=
c
2
r

1
(ct/(c
2
r))
. Hence we have (17.5).
Theorem 17.5. Let (E, D) be local. Also let h : (0, +) (0, +) be a
monotone function with full range, continuous, strictly increasing, doubling and
decays uniformly. Assume that (r) = h(r)/r is a monotone function with full
range and strictly increasing and that is (VD)
d
. If there exist c
1
, c
3
, c
4
> 0 such
that (17.5) holds for any t > 0 and any r (0, R] and
p(t, x, x)
c
5
(B
d
(x, h
1
(t)))
for any t (0, t

] and any x X, then there exist c


6
and c
7
such that
(17.11) p(t, x, y)
c
6
(B
d
(x, h
1
(t)))
exp
_
c
7
d(x, y)

1
(2c
4
t/d(x, y))
_
for any t (0, t

] and any x, y X with d(x, y) R.


The next two lemmas are technically the keys in proving the above theorem.
The rst one is well-known. See [22, Lemma 11.1].
Lemma 17.6. Assume that is (VD)
d
. There exist c
0
> 0 and > 0 such
that
(B
d
(x, r
1
)) c
0
(r
1
/r
2
)

(B
d
(x, r
2
))
for any r
1
r
2
> 0 and
(B
d
(x, r)) c
0
_
1 +
d(x, y)
r
_

(B
d
(y, r))
for any x, y X and r > 0. In particular, there exists M > 0 such that
(B
d
(x, r)) M(B
d
(y, r))
if d(x, y) r.
Lemma 17.7. Let be a monotone function with full range, strictly increasing
and continuous. Set h(r) = r(r). For any > 0, any > 0, any s > 0 and any
r > 0,
(17.12) 1 +
r
h
1
(s)
max{
1
, 1 +} exp
_
r

1
(s/r)
_
.
Proof. Set x = r/h
1
(s). If 0 x , then (17.12) holds. Assume that
x . Then

1
_
s
r
_
=
1
_

_
r
x
__

r
x
.
This implies
exp
_
r

1
(s/r)
_
exp x 1 +x.
Hence we have (17.12).
17. BASIC TOOLS IN HEAT KERNEL ESTIMATES 67
Proof of Theorem 17.5. We can prove this theorem by modifying the dis-
cussion in [22, Section 12.3]. Note that the counterpart of [22, (12.20)], which is
one of the key ingredients of the discussion in [22], is obtained by Lemma 17.7.
Next we give an o-diagonal lower estimate. For our theorem, the local property
of the Dirichlet form is not required but the estimate may not be best possible
without the local property, i.e. if the Hunt process associated with the Dirichlet
form has jumps. One can nd the original form on this theorem in [1].
Theorem 17.8. Let : (0, +) (0, +) be a monotone function with full
range, strictly increasing and continuous. Set h(r) = r(r). Assume that is
(VD)
d
and that d(x, y) satises the chain condition dened in Case 2 of Theo-
rem 15.10. Also assume that there exist c
1
and c
2
such that
(17.13)
c
1
V
d
(x, h
1
(t))
p(t, x, y)
for any t (0, t

] and any x, y X with d(x, y) c


2
h
1
(t). Then
(17.14)
c
3
V
d
(x, h
1
(t))
exp
_
c
4
d(x, y)

1
(c
5
t/d(x, y))
_
p(t, x, y)
for any t (0, t

] and any x, y X.
Lemma 17.9. Let C, D, T (0, +). Then D Ch
1
(T) if and only if
D/C
1
(TC/D). Also D Ch
1
(T) if and only if D/C
1
(TC/D).
The ideas of the following proof is essentially found in [1]. We modify a version
in [38].
Proof of Theorem 17.8. If d(x, y) c
2
h
1
(t), then (17.13) implies (17.14).
So we may assume that d(x, y) ch
1
(t), where c = min{c
2
/(6C), 1/(2C)}, with-
out loss of generality. By Lemma 17.9, we have
d(x, y)
c
1
(ct/d(x, y))
1.
Therefore, there exists n N such that
(17.15)
d(x, y)
2c
1
(2ct/d(x, y))
n
d(x, y)
c
1
(ct/d(x, y))
.
Note that (17.15) is equivalent to
(17.16) ch
1
_
t
n
_

d(x, y)
n
2ch
1
_
t
n
_
.
Now we use the classical chaining argument. (See [1] for example.) Note that
p(t, x, y) =
_
X
n1
p
_
t
n
, x, z
1
_
p
_
t
n
, z
1
, z
2
_
p
_
t
n
, z
n1
, y
_
(dz
1
) (dz
n1
).
By the chain condition, we may choose a sequence {x
i
}
n
i=0
such that x
0
= x, x
n
= y
and d(x
i
, x
i+1
) D/n for any i = 0, 1, . . . , n 1, where D = Cd(x, y). De-
ne B
i
= B
d
(x
i
, D/n) for i = 1, . . . , n 1. If z
i
B
i
and z
i+1
B
i+1
, then
d(z
i
, z
i+1
) 3D/n. By (17.16), 3D/n c
2
h
1
(t/n) and D/n h
1
(t/n), (17.13)
and Lemma 17.6 yield
p(t, z
i
, z
i+1
)
c
1
V
d
(z
i
, h
1
(t/n))

c
1
MV
d
(x
i
, h
1
(t/n))
,
68
Hence
p(t, x, y)

_
B
1
...B
n1
p
_
t
n
, x, z
1
_
p
_
t
n
, z
1
, z
2
_
p
_
t
n
, z
n1
, y
_
(dz
1
) (dz
n1
)
(c
1
/M)
n
1
V
d
(x, h
1
(t))
n1

i=1
V
d
(x
i
, D/n)
V
d
(x
i
, h
1
(t/n))
(c
1
/M)
n
1
V
d
(x, h
1
(t))
n1

i=1
V
d
(x
i
, D/n)
V
d
(x
i
, h
1
(t/n))
By Lemma 17.6 and (17.16),
V
d
(x
i
, D/n)
V
d
(x
i
, h
1
(t/n))
(c
0
)
1
_
D
nh
1
(t/n)
_

c
0
1
(cC)

.
Therefore there exists L > 1 such that
p(t, x, y)
L
n
V
d
(x, h
1
(t))
.
Now the desired estimate follows immediately from (17.15).
18. Proof of Theorem 15.6
We assume the same prerequisites on a resistance form (E, F) and the associated
resistance metric R as in Section 15.
Lemma 18.1. Let A be an open set containing x X. Assume that A = X
and that A is compact. Then, for any (0, 1),
(1 )R(x, A
c
)V
R
(x, R(x, A
c
)) E
x
(
A
) R(x, A
c
)(A)
Proof. Set B = A
c
. Note that E
x
(
A
) =
_
A
g
B
(x, y)(dy) by Corollary 10.11.
Since g
B
(x, y) g
B
(x, x) = R(x, B), the upper estimate is obvious. If y
B
R
(x, R(x, B)), then (GF4) implies that g
x
B
(y) (1 )g
x
B
(x). Therefore,
E
x
(
A
)
_
B
R
(x,R(x,B))
g
x
B
(y)(dy) (1 )R(x, B)V
R
(x, R(x, B)).

Proposition 18.2. Assume that d


QS
R.
(1) There exists > 0 such that B
d
(x, d
R
(x, r)) B
R
(x, r) B
d
(x, d
R
(x, r))
and B
R
(x, R
d
(x, r)) B
d
(x, r) for any x X and any r > 0, where d
R
(x, r) =
sup
yB
R
(x,r)
d(x, y).
(2) There exists c > 0 such that R
d
(x, 2r) cR
d
(x, r) for any x X and any
r > 0.
(3) If diam(X, d) < +, then, for any r

> 0, there exist (0, 1) and (0, 1)


such that R
d
(x, r) R
d
(x, r) for any x X and any r (0, r

]. If diam(X, d) =
+, then we have the same statement with r

= +.
(4) If is (VD)
R
, then it is (VD)
d
.
18. PROOF OF THEOREM 15.6 69
Proof. If d
QS
R, then by Theorem 12.3, d is (SQS)
R
and R is (SQS)
d
. Since
(X, R) is assumed to be uniformly perfect, Proposition 12.2-(3) implies that (X, d)
is uniformly perfect. Hence we may apply Theorem 11.5. Note that the statement
(a) of Theorem 11.5 holds.
(1) By the statement (b) of Theorem 11.5, d is (SQC)
R
and R is (SQC)
d
.
(2) By the statement (b) of Theorem 11.5, R is doubling with respect to d.
(3) Proposition 11.7 suces to deduce the desired result.
Proof of Proposition 15.4. Dene

R(x, r) = R(x, B
R
(x, r)
c
).
Assume (RES). By Proposition 18.2-(1),
(18.1) B
R
(x, R
d
(x, r)) B
d
(x, r) B
R
(x, R
d
(x, r)).
Hence by (RES),

R(x, r)

R(x, R
d
(x, r)) cR
d
(x, r).
If B
R
(x, R
d
(x, r)) = X, then (RES) also shows

R(x, r)

R(x, R
d
(x, r)) cR
d
(x, r).
In case X = B
R
(x, R
d
(x, r)), we have diam(X, R)/2 R
d
(x, r). Hence

R(x, r) diam(X, R) 2R
d
(x, r).
Conversely assume that

R(x, r) R
d
(x, r) for any x X and any r > 0 with
B
d
(x, r) = X. By (18.1),
(18.2) c
1
R
d
(x, r)

R(x, R
d
(x, r)) and

R(x, R
d
(x, r)) c
2
R
d
(x, r).
On the other hand, by Proposition 18.2-(1) and (2), there exists (0, 1) such
that
(r) R
d
(x, d
R
(x, r)) r (r)
for any x X and r > 0, where (r) = R
d
(x, d
R
(x, r)). Hence by (18.2),
r (r)

R(x, (r))

R(x, r/)

R(x, r)

R(x, (r)) c
2
(r) c
2
r/.
This suces for (RES).
Proof of Proposition 15.5. Assume (RES). Let (0, 1). By Proposi-
tion 15.4 and the volume doubling property of , we obtain
(B
R
(x, R(x, B
d
(x, r)
c
)) (B
R
(x, c
3
R
d
(x, r)))
c

(B
R
(x, R
d
(x, r))) c

(B
d
(x, r)).
By Lemmas 18.1 and 15.4 , c

(1 )c
3
h
d
(x, r) E
x
(
B
d
(x,r)
) c
4
h
d
(x, r).
Conversely, assume (EIN)
d
. By Lemma18.1,
(18.3) c
1
R
d
(x, r) R(x, B
d
(x, r)
c
).
Also Lemma 18.1 and the volume doubling property of yield
(18.4) c
2
R(x, B
d
(x, r)
c
)(B
R
(x, R(x, B
d
(x, r)
c
)) R
d
(x, r)V
d
(x, r).
By (18.3), it follows that
V
d
(x, r) V
R
(x, R
d
(x, r)) c
3
V
R
(x, c
1
R
d
(x, r)) V
R
(x, B
d
(x, r)
c
).
70
This and (18.4) show that c
4
R(x, B
d
(x, r)
c
) R
d
(x, r). Thus we obtain (15.1).
Now Proposition 15.4 implies (RES).
Proof of Theorem 15.6. By Lemma 7.9, B
R
(x, r) is totally bounded for
any x X and any r > 0. Hence B
R
(x, r) is compact. By Theorem 10.4, there
exists a jointly continuous heat kernel p(t, x, y). Since d
QS
R, B
d
(x, r) is compact
for any x X and any r > 0. Using (10.4) with A = B
d
(x, r) and letting t =
h
d
(x, r), we have
p(h
d
(x, r), x, x)
2 +

2
V
d
(x, r)
.
The rest is to show (EIN)
d
and the lower estimate of the heat kernel. Since is
(VD)
R
, (X, R) has the doubling property. Hence by Theorem 7.12, (ACC) im-
plies (RES). Proposition 15.5 shows (EIN)
d
. Next we show that h
d
(x, r) satises
the conditions (HKA), (HKB) and (HKC) in Lemma 17.2. (HKA) is immediate.
(HKB) follows from Proposition 18.2-(2) and (4). Note that diam(B
d
(x, r), R)
R
d
(x, r) diam(B
d
(x, r), R)/2. If d(x, y) r, then B
d
(y, r) B
d
(x, 2r). Hence,
R(y, r) diam(B
d
(y, r), R) diam(B
d
(x, 2r)) 2R(x, 2r).
By Proposition 18.2-(2), we have (HKC). Now assume that diam(X, R) = +.
Then we have (EIN)
d
. By Proposition 18.2-(3), Theorem 17.3 shows that, for some
c > 0,
(18.5)
c
V
d
(x, r)
p(h
d
(x, r), x, x)
for any x X and any r > 0. Next we consider the case where diam(X, R) < +.
If B
d
(x, r) = X, then r diam(X, d)/2. Therefore, assumptions of Lemma 17.2
hold with r

= diam(X, d)/3. Hence by Theorem 17.3, (18.5) is satised for any


x X and any r (0, diam(X, d)], where is independent of x. Next we show
(18.6)
1
(X)
inf
xX,rdiam(X,d)
p(h
d
(x, r), x, x)
and
(18.7) sup
xX,rdiam(X,d)
1
V
d
(x, r)
< +.
Letting A = X in Lemma 17.1, we have (X)
1
p(t, x, x) for any x X and
any t > 0. This yields (18.6). Let r
1
= diam(X, d). Since X B
R
(x, r) for
some r > 0, it follows that X is compact. Hence we may choose N > 0 so that,
for any x, y X, there exists {x
i
}
i=1,...,N+1
X such that x
1
= x, x
N+1
= y and
d(x
i
, x
i+1
) r
1
for any i = 1, . . . , N1. Since has the volume doubling property
with respect to d, there exists a
1
> 0 such that V
d
(y
1
, r
1
) a
1
V
d
(y
2
, r
1
) for any
y
1
, y
2
X with d(x, y) r
1
. Hence, for any x, y X, V
d
(x, r
1
) (a
1
)
N
V
d
(y, r
1
).
This shows (18.7). Thus we have obtained (18.6) and (18.7). Therefore, there exists
C > 0 such that
C
V
d
(x, r)
p(h
d
(x, r), x, x)
for any x X and any r diam(X, d). Hence changing c, we have (18.5) for any
x X and any r > 0 in this case as well.
Now, by Proposition 18.2-(3), there exists (0, 1) such that R
d
(x, r)
(c/4)R
d
(x, r) for any x X and any r diam(X, d), where c is the constant
19. PROOF OF THEOREMS 15.10, 15.11 AND 15.13 71
appearing in (18.5). Since R
d
(x, r) = diam(X, d) for any r diam(X, d), we see
that R(x, y) (c/4)R
d
(x, r) if d(x, y) min{r, diam(X, d)}. Let T = h
d
(x, r).
Then, this and (18.5) imply
|p(T, x, x) p(T, x, y)|
2
R(x, y)E(p
T,x
, p
T,x
)
R(x, y)p(T, x, x)
T

cR
d
(x, r)p(T, x, x)
4R
d
(x, r)V
d
(x, r)
=
1
4
c
V
d
(x, r)
p(T, x, x)
1
4
p(T, x, x)
2
.
Hence,
p(h
d
(x, r), x, y)
p(h
d
(x, r), x, x)
2

1
2
c
V
d
(x, r)
.
Thus we have shown Theorem 15.6.
19. Proof of Theorems 15.10, 15.11 and 15.13
The proofs of Theorems 15.10 and 15.11 depend on the results in Sections 13
and 14. We use those results by letting H(s, t) = st. Note that all the assumptions
on H in Sections 13 and 14 are satised for this particular H.
Proof of Theorem 15.10. The equivalence between (a), (b) and (c) is im-
mediate form the corresponding part of Corollary 13.3. Next assume that (a), (b)
and (c) hold. By Corollary 13.3, has the volume doubling property with respect
to d and R. Also by Lemma 14.4 and Proposition 18.2-(2), there exists , (0, 1)
such that h
d
(x, r) h
d
(x, r) for any x X and r > 0. Hence by (15.4), g
decays uniformly. Lemma 15.8 implies that g
1
is doubling and decays uniformly.
Now apply Theorem 15.6. There exists a jointly continuous heat kernel p(t, x, y).
Furthermore, combining (15.2), (15.3) and (15.4) along with the volume doubling
property and the above mentioned properties of g and g
1
, we obtain
1
V
d
(x, g
1
(t))
p(t, x, x)
for any t cg(diam(X, d)) and any x X, where c is independent of x and r. The
same arguments as in the proof of Theorem 15.6, in particular, (18.6) and (18.7)
show (DHK)
g,d
for t cg(diam(X, d)). Now (KD) is straightforward by the volume
doubling property. Thus we have obtained (HK)
g,d
.
Conversely, assume (HK)
g,d
. (KD) and (DHK)
g,d
imply that has the volume
doubling property with respect to d. Since d
QS
R, we have the volume doubling
property of with respect to R. Also (X, d) is uniformly perfect. By Theorem 15.6,
we have (15.2) and (15.3). Comparing those with (DHK)
g,d
, we see that
(19.1) V
d
(x, g
1
(h
d
(x, r))) V
d
(x, r)
for any x X and any r > 0. Set r

= diam(X, d). Note that B


d
(x, r) = X
for any r < r

/2. By Lemma 14.4, for any > 1, there exists (0, 1)


such that V
d
(x, r)
1
V
d
(x, r) for any r < r

/2. This along with (19.1)


shows that r g
1
(h
d
(x, r)) for any r < r

/2. This and Lemma 15.8 im-


ply (15.4) for r < r

/2. Let us think about r [r

/2, r

]. If r

< +,
then (X, d) is compact and so is (X, R). Therefore, R
d
(x, r) diam(X, R) and
72
V
d
(x, r) (X) < +. Let r [r

/2, r

]. By the volume doubling prop-


erty, V
d
(x, r) V
d
(x, r

/2) cV
d
(x, r

) = c(X). Also, by Proposition 7.6-


(2), we have R
d
(x, r) R
d
(x, r

/2) cV
d
(x, r

) c

diam(X, R)/2. Hence


cc

diam(X, R)(X)/2 h
d
(x, r) diam(X, R)(X) for any r [r

/2, r

]. Also
g(r

/2) g(r) g(r

) for any r [r

/2, r

]. Therefore, adjusting constants, we


obtain (15.4) for r (0, r

]. Thus the condition (c) has been veried.


(15.5) follows from its counterpart (15.2).
The rest is o-diagonal estimates. Note that both g and are doubling and
decay uniformly. Then by Lemma19.1 below, we may replace g and by h and
which are continuous and strictly increasing. For the upper o-diagonal estimate,
since d
QS
R, Theorem 15.6 implies (17.1) with r

= diam(X, d)/2. Then by


Lemmas 17.2 and 17.4, we obtain (17.5) for any t > 0 and any r (0, diam(X, d)).
Applying Theorem 17.5, replacing h and by g and , and using the doubling
properties, we obtain (15.6) for any x, y X and any t > 0. Finally, since we have
(15.5), Theorem 17.8 shows an o-diagonal lower estimate, which easily implies
(15.7) by similar argument as in the case of upper o-diagonal estimate.
Lemma 19.1. Suppose that g : (0, +) (0, +) is a monotone function with
full range and doubling. Then there exists h : (0, +) (0, +) such that h is
continuous and strictly monotonically increasing on (0, +) and g(r) h(r) for
any r (0, +). Moreover, if g decays uniformly, then g
1
(t) h
1
(t) for any
t (0, +).
Proof. Assume that g(2r) cg(r) for any r. Set (r) = 1 + (1 + e
r
)
1
.
Note that is strictly monotonically increasing and 1 < (r) < 2 for any r. Let
G(r) = (r)g(r). Then H is strictly monotonically increasing. There exists a
continuous function F : (0, +) (0, +) such that F(G(r)) = r for any r > 0.
Dene f(x) = (x)F(x). Then f is strictly monotonically increasing and continuous
and so is the inverse of f, which is denoted by h. Since f(G(r)) = (G(r))F(G(r)) =
(G(r))r, we have (r)g(r) = h((G(r))r). This implies h(r)/2 g(r) ch(r).
Now assume that g decays uniformly. Then so does h. By Lemma 15.8, h
1
is
doubling. Since h(r)/2 g(r) ch(r), we have h
1
(t/c) g
1
(t) h
1
(2t) for
any t (0, +). Hence the doubling property of h
1
shows h
1
(t) g
1
(t) for
any t (0, +).
Proof of Theorem 15.11. By Theorem 15.6, (C1) implies (C5). Note that
R
QS
R. Since (X, R) is uniformly perfect, it follows that R
R
(x, r) r. Hence
(C5) implies (C2). Obviously (C2) implies (C6). By Proposition 15.5, (C6) implies
(C1).
(C1) (C3) and (C4): Assume (C1). Then by Theorem 14.1, there exists
a metric d which satises the condition (c) of Theorem 15.10 with some > 1.
Therefore, we have (C3) by Theorem 15.10. Also, (C4) follows by Theorem 15.6.
(C3) (C1): Assume (C3). Then, (DHK)
g,d
and (KD) imply the volume
doubling property of with respect to d. Since d
QS
R, has the volume doubling
property with respect to R as well. Hence we have (C1).
(C4) (C1): Assume (C4). Then (2r)

R
d
(x, 2r)V
d
(x, 2r). By Proposi-
tion 18.2-(2), R
d
(x, 2r) R
d
(x, r). Hence is (VD)
d
. Since d
QS
R, is (VD)
R
.
Also we have (EIN)
d
. Hence Proposition 15.5 shows (ACC). Thus (C1) is veried.
19. PROOF OF THEOREMS 15.10, 15.11 AND 15.13 73
Finally, (15.9) follows by the process of construction of d in Section 14, in
particular, by (14.1).
Proof of Theorem 15.13. Let g(r) = r

and let h(r) = r

. By The-
orem 15.10, (HK)
g,d
shows (DM2)
g,d
and (DM1)
g,d
. Using (15.10), we obtain
(DM2)
h,d|
Y
and (DM1)
h,d|
Y
, where we replace by . Since we have (ACC) for
(E|
Y
, F|
Y
), Theorem 15.10 implies the counterpart of (HK)
h,d|
Y
. Thus we have
(15.11).
Part 4
Random Sierpinski gaskets
The main purpose of this part is to apply theorems in the last part to resistance
forms on random Sierpinski gaskets. The notion of random (recursive) self-similar
set has introduced in [44], where basic properties like the Hausdor dimension have
been studied. Analysis on random Sierpinski gaskets has been developed in a series
of papers by Hambly[27, 28, 29]. He has dened Brownian motion on a random
Sierpinski gasket associated with a natural resistance form and studied asymptotic
behaviors of associated heat kernel and eigenvalue counting function. He has found
possible uctuations in those assymptotics, which have later conrmed in [30].
In this part, we will rst establish a sucient and necessary condition for a
measure to be volume doubling with respect to the resistance metric associated
with the natural resistance form in Theorem 23.2. This result is a generalization of
the counterpart in [39] on self-similar sets. Using this result, we show that a certain
class of random self-similar measure always has the volume doubling property with
respect to the resistance metric, so that we may apply theorems on heat kernel
estimates in the last part. Note that Hambly has used the Hausdor measure
associated with the resistance metric, which is not a random self-similar measure in
general. In fact, in Section 25, we show that the Hausdor measure is not volume
doubling with respect to the resistance metric almost surely. On the contrary,
in the homogeneous case, the Hausdor measure is a random self-similar measure
and is shown to satisfy the volume doubling condition in Section 24. Applying
Theorem 15.10, we will recover the two-sided o-diagonal heat kernel estimate in
[7]. See Theorem 24.7.
Throughout this part, we x p
1
=

1, p
2
=

3/2

1/2 and p
3
=

3/2

1/2 and set V


0
= {p
1
, p
2
, p
3
}. Note that p
1
+ p
2
+ p
3
= 0 and that V
0
is the set of vertices of a regular triangle. Let T be the convex hull of V
0
. We will
always identify R
2
with C if no confusion may occur.
Dene f
i
: C C by f
i
(z) = (zp
i
)/2+p
i
for i = 1, 2, 3. The self-similar set K
associated with {f
1
, f
2
, f
3
} is the Sierpinski gasket, i.e. K is the unique nonempty
compact set which satises K = f
1
(K) f
2
(K) f
3
(K). Recall the denition of
self-similar sets in Example 7.2. To distinguish K from other generalized Sierpinski
gasket, we call K the original Sierpinski gasket.
20. Generalized Sierpinski gasket
In this section, as a basic component of random Sierpinski gasket, we dene a
family of self-similar sets in R
2
which can be considered as a modication of the
original Sierpinski gasket. Then making use of the theory in [36], we briey review
the construction of resistance forms on those sets. Also, in Example 20.11, we apply
Theorem 15.13 to subsets of the original Sierpinski gasket and obtain heat kernel
estimates for the traces onto those subsets.
The following is a standard set of denitions and notations regarding self-similar
sets.
Definition 20.1. Let S be a nite set.
(1) We dene W
m
(S) = S
m
= {w
1
w
2
w
m
|w
j
S for j = 1, . . . , m} for m 1
and W
0
(S) = {}. Also W

(S) =
m0
W
m
(S). For any w W

(S), the length of


w, |w|, is dened to be m where w W
m
(S). For any w = w
1
w
2
w
m
W
m
(S),
20. GENERALIZED SIERPINSKI GASKET 77
dene [w]
0
= and
[w]
n
=
_
w
1
w
2
w
n
if 1 n < m,
w if n m.
(2) (S) is dened by (S) = S
N
= {
1

2
. . . |
j
S for any j N}. For any
=
1

2
. . . (S), dene []
0
= and []
n
=
1

2

n
for any n 1. For
w W

(S), dene
w
(S) = {| (S), []
|w|
= w} and dene
w
: (S)
(S) by
w
() = w.
A generalized Sierpinski gasket is dened as a self-similar set which preserves
some of the good properties possessed by the original Sierpinski gasket.
Definition 20.2. Let K be a non-empty compact subset of R
2
and let S =
{1, . . . , N} for some integer N 3. Also let F
i
(x) =
i
A
i
x + q
i
for any i S,
where
i
(0, 1), A
i
O(2), where O(2) is the 2-dimensional orthogonal matrices,
and q
i
R
2
. Recall the denitions of p
1
, p
2
, p
3
, V
0
and T at the beginning of this
part. (K, S, {F
i
}
iS
) is called a generalized Sierpinski gasket, GSG for short, if and
only if the following four conditions are satised:
(GSG1) K =
iS
F
i
(K),
(GSG2) F
i
(p
i
) = p
i
for i = 1, 2, 3,
(GSG3) F
i
(T) T for any i S and F
i
(T) F
j
(T) F
i
(V
0
) F
j
(V
0
) for any
i, j S with i = j,
(GSG4) For any i, j {1, 2, 3}, there exist i
1
, . . . , i
m
such that i
1
= i, i
m
= j and
F
i
k
(V
0
) F
i
k+1
(V
0
) = for all k = 1, . . . , m1.
Write L = (K, S, {F
i
}
iS
). For any w = w
1
w
2
w
m
W

(S)\W
0
(S), we dene
F
w
= F
w
1
. . . F
w
m
and K
w
= F
w
(K). Also V
m
(L) =
wW
m
(S)
F
w
(V
0
).
By (GSG1), (GSG2) and (GSG3), the results in [36, Sections 1.2 and 1.3] show
that a GSG L = (K, S, {F
i
}
iS
) is a post critically nite self-similar structure whose
post critical set is V
0
. Also by (GSG4), K is connected.
Using the results in [36, Section 1.4], we have the following proposition.
Proposition 20.3. Let L = (K, S, {F
i
}
iS
) be a GSG and let (
s
)
sS
(0, 1)
S
satisfy

sS

s
= 1. Then there exists a Borel regular measure on K such that
(K
w
) =
w
1

w
2

w
m
for any w = w
1
w
2
w
m
W

(S).
Definition 20.4. The Borel regular measure in Proposition 20.3 is called
the self-similar measure on K with weight (
s
)
sS
.
Next we give a brief survey on the construction of a (self-similar) resistance
form on a GSG. See [36] for details.
Proposition 20.5. Let L = (K, S, {F
i
}
iS
) be a GSG and let (D, r) LA(V )
(0, +)
S
. Then there exists a Laplacian L
m
LA(V
m
(L)) such that, for any
u, v (V
m
(L)),
E
L
m
(u, v) =

wW
m
(S)
1
r
w
E
D
(u F
w
, v F
w
),
where r = (r
i
)
iS
and r
w
= r
w
1
r
w
m
for w = w
1
w
2
w
m
W
m
(S).
Definition 20.6. Let L = (K, S, {F
i
}
iS
) be a GSG and let (D, r) LA(V )
(0, +)
S
. Dene L
m
(D, r) LA(V
m
(L)) as the Laplacian L
m
obtained in Proposi-
tion 20.5. (D, r) is called a regular harmonic structure on L if and only if r (0, 1)
S
and {(V
m
(L), L
m
(D, r)}
m0
is a compatible sequence.
78
By [36, Proposition 3.1.3], {(V
m
(L), L
m
(D, r)}
m0
is a compatible sequence if
and only if
E
D
(u, u) = inf{E
L
1
(v, v)|v (V
1
(L)), v|
V
0
= u}
for any u (V
0
). Hence we only have to verify nite number of equations to show
that (D, r) is a regular harmonic structure.
Now let (D, r) be a regular harmonic structure. By Theorem 3.13, the com-
patible sequence S = {(V
m
(L), L
m
(D, r)}
m0
produces a resistance form (E, F) on
X, where X is the completion of (V

, R
S
). By the results in [36, Sections 3.2 and
3.3], we may identify X = K and obtain the following proposition.
Proposition 20.7. Let L = (K, S, {F
i
}
iS
) be a GSG and let (D, r) be a
regular harmonic structure on L. There exists a resistance form (E, F) on K which
satises the following properties (a), (b) and (c):
(a) The associated resistance metric R gives the same topology as the restriction
of the Euclidean metric does,
(b) L
(E,F),V
m
= L
m
(D, r) for any m 0,
(c) For any u F, u F
s
F for any s S and
(20.1) E(u, u) =

sS
1
r
s
E(uF
s
, uF
s
).
In particular, (K, R) is compact and (E, F) is a regular resistance form.
The equation (20.1) is called the self-similarity of the resistance form (E, F).
Recall that the chain condition of a distance is required to get a (lower) o-
diagonal estimate of a heat kernel. In [35], we have obtained a condition for the
existence of a shortest path metric which possesses the chain condition. The next
denitions and the following theorems are essentially included in [35].
Definition 20.8. Let L = (K, S, {F
i
}
iS
) be a GSG.
(1) For p, q V
n
(L), (x
1
, . . . , x
m
) is called an n-path between p and q if x
1
=
p, x
m
= q and for any i = 1, . . . , m1, there exists w W
n
(S) such that x
i
, x
i+1

F
w
(V
0
).
(2) L is said to admit symmetric self-similar geodesics if and only if there exists
(0, 1) such that

1
= min{m1|(x
1
, . . . , x
m
) is a 1-path between p and q}
for any p, q V
0
with p = q. is called the symmetric geodesic ratio of L.
Definition 20.9. Let (X, d) be a metric space. For x, y X, a continuous
curve g : [0, d(x, y)] X is called a geodesic between x and y if and only if
d(g(s), g(t)) = |s t| for any s, t [0, 1]. If there exists a geodesic between x and
y for any x, y X, then d is called a geodesic metric on X.
Obviously, a geodesic metric satises the chain condition. The following theo-
rem shows the existence of geodesic metric.
Theorem 20.10. Let L = (K, S, {F
i
}
iS
) be a GSG. Assume that L admits
symmetric self-similar geodesics. Then there exists a geodesic metric d on K which
gives the same topology as the Euclidean metric does on K and
d(F
i
(x), F
i
(y)) = d(x, y)
20. GENERALIZED SIERPINSKI GASKET 79
for any x, y K and any i S, where is the symmetric geodesic ratio of L.
Moreover, for any p, q V
n
(S),
d(p, q) =
n
min{m1|(x
1
, . . . , x
m
) is an n-path between p and q}.
Proof. We can verify all the conditions in [35, Theorem 4.3] and obtain this
theorem.
We present two examples which will used as a typical component of random
Sierpinski gaskets in the following sections.
Example 20.11 (the original Sierpinski gasket). Let f
i
(z) = (z p
i
)/2 + p
i
and let K be the original Sierpinski gasket. Set S = {1, 2, 3}. Then (K, S, {f
i
}
iS
)
is a generalized Sierpinski gasket. We write L
SG
= (K, S, {f
i
}
iS
). Dene
D
h
=
_
_
2 1 1
1 (1 +h) h
1 h (1 +h)
_
_
for h > 0. By [36, Exercise 3.1], there exists a unique (r
1
, r
2
, r
3
) such that
(D
h
, (r
1
, r
2
, r
3
)) is a harmonic structure for each h > 0. Also the unique (r
1
, r
2
, r
3
)
satises r
2
= r
3
and (D
h
, r
1
, r
2
, r
3
)) is regular. We write r
i
= r
SG
i
for i = 1, 2, 3.
Note that r
SG
i
depends on h in fact.
Hereafter in this example, we set h = 1. Then r
SG
1
= r
SG
2
= r
SG
3
= 3/5. Set r =
(3/5, 3/5, 3/5). Let be the self-similar measure on K with weight (1/3, 1/3, 1/3).
Let (E, F) be the regular resistance form on K associated with (D
1
, r) and let R
be the associated resistance metric on K. Then by Barlow-Perkins [9], it has been
known that the heat kernel p(t, x, y) associated with the Dirichlet form (E, F) on
L
2
(K, ) satises
(20.2) c
1
t

d
S
2
exp
_
c
2
_
|x y|
d
w
t
_ 1
d
w
1
_
p(t, x, y)
c
3
t

d
S
2
exp
_
c
4
_
|x y|
d
w
t
_ 1
d
w
1
_
for any t (0, 1] and any x, y K, where d
s
= log 9/ log 5 and d
w
= log 5/ log 2.
The exponents d
s
and d
w
are called the spectral dimension and the walk dimension
of the Sierpinski gasket respectively. In this case, L
SG
admits symmetric self-
similar geodesics with the geodesic ratio 1/2. The resulting geodesic metric on K
is equivalent to the Euclidean metric.
Next we consider the traces of (E, F) on Ahlfors regular subsets of K. It is
known that
R(x, y) d
E
(x, y)
(log 5log 3)/ log 2
for any x, y K, where d
E
(x, y) = |x y|. Hence d
E

QS
R. Also,
(20.3) (B
d
E
(x, r)) r
d
H
for any x K and r (0, 1], where d
H
= log 3/ log 2 is the Hausdor dimension
of (K, d
E
). Let Y be a closed Ahlfors -regular subset of K, i.e. there exists
a Borel regular measure on Y such that (B
d
(x, r) Y ) r

for any x Y
and any r (0, 1]. Then by (20.2) and (20.3), we may verify all the assumptions
80
Figure 2. the Sierpinski spiral
of Theorem 15.13. Thus there exists a jointly continuous heat kernel p
Y

(t, x, y)
associated with the regular Dirichlet form (E|
Y
, F|
Y
) on L
2
(Y, ) and
p
Y

(t, x, x) t

for any x K and any t (0, 1], where =


log 2
log 5log 3+ log 2
. In particular, if Y is
equal to the line segment p
2
p
3
, then = 1 and = log 2/ log (10/3).
Example 20.12 (the Sierpinski spiral). For i = 1, 2, 3, dene h
i
(z) = (z
p
i
)/3 + p
i
for any z C. Also dene h
4
(z) = z/

3. The unique non-empty


compact subset K of C satisfying K =
i=1,2,3,4
h
i
(K) is called the Sierpinski spiral,
the S-spiral for short. See Figure 2. Let S = {1, 2, 3, 4}. Then (K, S, {h
i
}
iS
) is
a generalized Sierpinski gasket. We use L
SP
to denote this generalize Sierpinski
gasket associated with the S-spiral. Let D
h
be the same as in Example 20.11 for h >
0. Dene r
SP
1
= (h)/(h+1), r
SP
2
= (1h)/(h+1), r
SP
3
= (1)/2 and r
SP
4
= .
Then (D
h
, (r
SP
i
)
iS
) is a regular harmonic structure for (0, min{h, 1/h}). Let
(E, F) be the regular resistance form on K associated with (D
h
, (r
SP
i
)
iS
) and let
R be the resistance distance induced by (E, F). Note that K is a dendrite, i.e. for
any two points x, y K, there is a unique simple path between x and y. It follows
that R is a geodesic metric. The Hausdor dimension d
H
of (K, R) is given by the
unique d which satises
4

i=1
(r
SP
i
)
d
= 1.
By Theorem 23.8, any self-similar measure on K has the volume doubling property
with respect to R. (Note that a generalize Sierpinski gasket itself is a special
random Sierpinski gasket. Also for the S-spiral, all the adjoining pair are trivial,
i,e. ((j
1
, i
1
), (j
2
, i
2
)) is an adjoining pair if and only if j
1
= j
2
and i
1
= i
2
. See
Denition 23.7 for the denition of adjoining pair.) In particular, letting be the
self-similar measure with weight ((r
SP
i
)
d
H
)
iS
, we have
R(x, y)V
R
(x, R(x, y)) R(x, y)
d
H
+1
.
21. RANDOM SIERPINSKI GASKET 81
By Theorem 15.10, the heat kernel p(t, x, y) associated with the Dirichlet form
(E, F) on L
2
(K, ) satises
c
1
t

d
H
d
H
+1
exp
_
c
2
_
R(x, y)
d
H
+1
t
_ 1
d
H
_
p(t, x, y)
c
3
t

d
H
d
H
+1
exp
_
c
4
_
R(x, y)
d
H
+1
t
_ 1
d
H
_
for any t (0, 1] and any x, y K. Note that the S-spiral admits symmetric
self-similar geodesics with the ratio 1/3 and this geodesic metric coincides with the
resistance metric R when h = 1 and = 1/3.
21. Random Sierpinski gasket
In this section, we will give basic denitions and notations for random (recur-
sive) Sierpinski gaskets. Essentially the denition is the same as in [44, 27, 28, 29].
However, we will not introduce the randomness until Section 25.
For j = 1, . . . , M, let L
j
= (K(j), S
j
, {F
j
i
}
iS
j
) be a generalized Sierpinski gas-
ket, where S
j
= {1, . . . , N
j
}. Set N = max
j=1,...,M
N
j
and dene S = {1, . . . , N}.
Those generalized Sierpinski gaskets L
1
, . . . , L
M
are the basic components of our
random Sierpinski gasket.
Definition 21.1. Let W

(S) and let : W

{1, . . . , M}. (W

, )
is called a random Sierpinski gasket generated by {L
1
, . . . , L
M
} if and only if the
following properties are satised:
(RSG) W

and, for m 1, w = w
1
w
2
w
m
W
m
(S) belongs to W

if and
only if [w]
m1
W

and w
m
S
([w]
m1
)
.
Strictly speaking, to call (W

, ) a random Sierpinski gasket, one need to


introduce a randomness in the choice of (w) for every w, i.e. a probability measure
on the collections of (W

, ). We will do so in the nal section, Section 25. Until


then, we study each (W

, ) respectively without randomness.


Note that (W

, ) is not a geometrical object. The set K(W

, ) R
2
dened
in Proposition 21.3-(2) is the real geometrical object considered as the random
self-similar set generated by (W

, ).
Definition 21.2. Let (W

, ) be a random Sierpinski gasket generated by


{L
1
, . . . , L
M
}. Dene W
m
= W

W
m
(S).
(1) Dene F

= I, where I is the identity map from R


2
to itself. For any m 1
and w = w
1
w
2
w
m
W
m
, dene F
w
: R
2
R
2
by
F
w
= F
([w]
0
)
w
1
F
([w]
1
)
w
2
. . . F
([w]
m1
)
w
m
.
(2) (W

, ) = {w
1
w
2
. . . |w
1
w
2
. . . (S), w
1
. . . w
m
W
m
for any m 1}.
(3) Dene T
m
(W

, ) =
wW
m
F
w
(T) and V
m
(W

, ) =
wW
m
F
w
(V
0
) for any
m 0.
The following results are basic properties of random Sierpinski gaskets which
are analogous to self-similar sets.
Proposition 21.3. Let (W

, ) be a random Sierpinski gasket generated by


{L
1
, . . . , L
M
}.
(1)
m0
T
m
(W

, ) equals to the closure of


m0
V
m
(W

, ) with respect to the


82
Euclidean metric.
(2) Dene K(W

, ) =
m0
T
m
and K
w
(W

, ) = K(W

, ) F
w
(T) for any
w W

. Then, K
w
(W

, ) K
v
(W

, ) = F
w
(V
0
) F
v
(V
0
) for any w, v W

with

w
(S)
v
(S) = .
(3) Let =
1

2
. . . (W

, ). Then
K
[]
m
(W

, ) K
[]
m+1
(W

, )
for any m 0 and
m1
K
[]
m
(W

, ) is a single point. If we denote this single


point by
W

,
(), then the map
W

,
: (W

, ) K(W

, ) is continuous and
onto. For any k = 1, 2, 3, (
W

,
)
1
(p
k
) = {(k)

}, where (k)

= kkk . . . (S).
(4) For any x K(W

, ), set n(x) = #(
1
W

,
(x)). Then 1 n(x) 5.
Moreover n(x) 2 if and only if there exist w W

, i
1
, . . . , i
n(x)
S
(w)
with
i
m
= i
n
for any m = n and k
1
, . . . , k
n(x)
{1, 2, 3} such that

1
W

,
(x) = {wi
m
(k
m
)

|m = 1, . . . , n(x)}.
Next we try to describe the self-similarity of random Sierpinski gasket.
Definition 21.4. Let (W

, ) be a random Sierpinski gasket generated by


{L
1
, . . . , L
M
}.
(1) For any w W

, dene W
w

= {v|wv W

} and
w
: W
w

{1, . . . , M} by

w
(v) = (wv) for any v W
w

.
(2) A subset W

is called a partition of W

if and only if (W

, )

w
(S) and
w(1)
(S)
w(2)
(S) = for any w(1), w(2) with w(1) = w(2).
The following theorem gives the self-similarity of random Sierpinski gasket.
(21.1) is the counterpart of the ordinary self-similarity K =
N
i=1
F
i
(K).
Proposition 21.5. Let (W

, ) be a random Sierpinski gasket generated by


{L
1
, . . . , L
M
}. For any w W

, (W
w
,
w
) is a random Sierpinski gasket generated
by {L
1
, . . . , L
M
}, K
w
(W

, ) = F
w
(K(W
w

,
w
)) and F
w

W
w

,
w =
W

w
.
Moreover, if is a partition of W

, then
(21.1) K(W

, ) =
_
w
K
w
(W

,
w
) =
_
w
F
w
(K(W
w

,
w
)).
The following proposition describes the topological structure of a random Sier-
pinski gasket.
Proposition 21.6. Let K
m,x
(W

, ) =
wW
m
,xK
w
(W

,)
K
w
(W

, ). Then
K
m,x
(W

, ) is a neighborhood of x and sup


xK(W

,)
diam(K
m,x
, d
E
) 0 as m
+, where d
E
is the Euclidean distance.
Proof. Set A
m,x
=
wW
m
,x/ K
w
(W

,)
K
w
(W

, ). Then A
m,x
is compact
and x / A
m,x
. Hence = min
yA
m,x
|x y| > 0. Write K
m,x
= K
m,x
(W

, ).
For any s (0, ), B
d
E
(x, s) K(W

, ) K
m,x
. Hence K
m,x
is a neighbor-
hood of x. Let L be the maximum of the Lipschitz constants of F
j
i
for j
{1, . . . , M} and i S
j
. Then diam(K
w
(W

, ), d
E
) L
m
diam(T, d
E
). Thus
sup
xK(W

,)
diam(K
w
(W

, )) L
m
diam(T, d
E
) 0 as m +.
Figure 3 shows two random Sierpinski gaskets generated by {L
SG
, L
SP
}.
22. RESISTANCE FORMS ON RANDOM SIERPINSKI GASKETS 83
Figure 3. Random Sierpinski gaskets
22. Resistance forms on Random Sierpinski gaskets
The main purpose of this section is to introduce the construction of a (random
self-similar) resistance form on a random Sierpinski gasket. We follow the method
of construction given in [28]. Furthermore, we are going to study the resistance
metric associated with the constructed resistance form.
In this section, we x a random Sierpinski gasket (W

, ) generated by {L
j
}
M
j=1
,
where L
j
= (K(j), S
j
, {F
j
i
}
iS
j
) and S
j
= {1, . . . , N
j
}. We write T
m
, V
m
, K, K
w
and in place of T
m
(W

, ), V
m
(W

, ) and so on.
Let (D, r
(j)
) be a regular harmonic structure for each j {1, . . . , M}. Set
r
(j)
= (r
(j)
i
)
iS
j
. (Note that D is independent of j.) Dene r = max{r
(j)
i
|j
{1, . . . , M}, i S
j
} and r = min{r
(j)
i
|j {1, . . . , M}, i S
j
}.
We rst construct a series of resistance forms and/or Laplacians on {V
m
}
m0
as in the case of a generalized Sierpinski gasket.
Definition 22.1. Let (W

, ) be a random Sierpinski gasket generated by


{L
1
, . . . , L
M
}. For any w W
m
, dene r
w
= r
([w]
0
)
w
1
r
([w]
1
)
w
2
r
([w]
m1
)
w
m
. (We set
r

= 1.) Dene a symmetric bilinear form E


m
on (V
m
) by
E
m
(u, v) =

wW
m
1
r
w
E
D
(uF
w
, vF
w
)
for any u, v (V
m
). We use L
m
to denote the symmetric linear operator from
(V
m
) to itself satisfying E
m
(u, v) = (u, L
m
v)
V
m
for any u, v (V
m
).
Since each (D, r
(j)
) is a harmonic structure, we have the following fact imme-
diately.
Proposition 22.2. E
m
is a resistance form on V
m
for any m 1 and L
m
is
a Laplacian on V
m
. Moreover, {(V
m
, L
m
)}
m0
is a compatible sequence.
Let S = {(V
m
, L
m
)}
m0
be the compatible sequence obtained in Proposi-
tion 22.2. Then we have a resistance form (E
S
, F
S
) on V

by Theorem 3.13. Let


R
S
be the associated resistance metric on V

. Note that if x, y V
m
, then R
S
(x, y)
84
is equal to the resistance metric with respect to the resistance form (E
m
, (V
m
)) on
V
m
. We use this fact in the followings. Let (X, R) be the completion of (V

, R
S
)
as in Theorem 3.13. We are going to identify X with K and show that R gives the
same topology as the restriction of the Euclidean metric on K does. Hereafter we
use E, F and R to denote E
S
, F
S
and R
S
if no confusion may occur.
The following denition is an analogue of the notion of scales in [39].
Definition 22.3. For s (0, 1) dene

s
= {w|w W

\W
0
, r
[w]
|w|1
> s r
w
}
and
1
= {}. For any x X and any s (0, 1],

s,x
= {w|w
s
, x K
w
}, K
s
(x) =
w
s,x
K
w
,

1
s,x
= {w|w
s
, K
w
K
s
(x) = } and U
s
(x) =
w
1
s,x
K
w
.
Also Q
s
(x) =
w
s
\
1
s,x
K
w
, C
s
(x) = U
s
(x) Q
s
(X).
We think of K
w
s for w
s
a ball of radius s with respect to the resistance
metric. Also, U
s
(x) is regarded as a s-neighborhood of x. Such a viewpoint will be
justied in Corollary 22.8. First we show that {U
s
(x)}
s>0
is a fundamental system
of neighborhoods with respect to the Euclidean metric.
Lemma 22.4. Let d
E
be the restriction of Euclidean metric on K. In this
lemma, we use the topology of K induced by d
E
.
(1) K
s
(x) and U
s
(x) are compact.
(2) U
s
(x) is a neighborhood of x with respect to d
E
. Moreover,
lim
s0
sup
xK
diam(U
s
(x), d
E
) = 0.
(3) C
s
(x)
w
1
s,x
F
w
(V
0
) and C
s
(x) is the topological boundary of U
s
(x).
Proof. (1) and (3) are immediate. About (2), for w = w
1
w
2
w
m

s
,
since r
w
1
w
2
w
m1
> s r
w
, it follows that r
m1
s r
m
. Hence
(22.1)
log s
log r
m
log s
log r
+ 1
Let m(s) be the integral part of log s/ log r and let m(s) be the integral part of
log s/ log r + 2. Then (22.1) implies U
s
(x) K
m(s),x
. By Proposition 21.6, U
s
(x)
is a neighborhood of x. Also by (22.1),
diam(U
s
(x), d
E
) 4 sup
w
s
diam(K
w
, d
E
) 4 sup
xK
diam(K
m(s),x
).
Now Proposition 21.6 yields the desired result.
In the next lemma, we show that the diameter of K
w
for w
s
is roughly s.
Lemma 22.5. (1) There exists c
0
> 0 such that sup
x,yK
w
V

R(x, y) c
0
r
w
for any w W

.
(2) There exists c
1
> 0 such that R(x, y) c
1
s for any s (0, 1], any x V

and
any y U
s
(x) V

.
Proof. (1) First we enumerate two basic facts.
Fact 1: r
w
(r)
|w|
.
Fact 2: Dene R

= max{R(x, y)|x, y V
0
}. Then R(x, y) r
w
R

for any
x, y F
w
(V
0
).
22. RESISTANCE FORMS ON RANDOM SIERPINSKI GASKETS 85
Assume that x F
w
(V
0
) and y F
w
(V
1
(L
(w)
)) for some w W

. Note that
F
w
(V
1
(L
(w)
)) =
iS
(w)
F
wi
(V
0
). Recall that we set N = max
j=1,...,M
N
j
, where
N
j
= #(S
j
). Since #(V
1
(L
(w)
)) 3N, we may nd m 3N, i
1
, . . . , i
m
S
(w)
and x
0
, x
1
, . . . , x
m
F
w
(V
1
(L
(w)
)) satisfying x
0
= x, x
m
= y and x
k1
, x
k

F
wi
k
(V
0
) for any k = 1, . . . , m. By the above facts,
R(x, y)
m

k=1
R(x
k1
, x
k
) 3NR

rr
w
.
Now, let x F
w
(V
0
) and let y K
w
V

. Then y F
wv
(V
0
) for some wv W

.
Choose y
i
F
[wv]
|w|+i
for i = 1, . . . , |v| 1. Set y
0
= x and y
|v|
= y. By the above
argument,
R(x, y)
|v|1

i=0
R(y
i
, y
i+1
)
|v|1

i=0
3NR

rr
[wv]
|w|+i

i=0
3NR

(r)
i+1
r
w
=
3NR

rr
w
1 r
.
This shows that sup
x,yK
w
V

R(x, y) 6NR

r(1 r)
1
r
w
.
(2) Let y U
s
(x)V

. There exist w(1), w(2)


1
s,x
and z F
w(1)
(V
0
)F
w(2)
(V
0
)
such that x K
w(1)
V

and y K
w(2)
V

. By (1), R(x, y) R(x, z) +R(z, y)


c
0
(r
w(1)
+r
w(2)
) 2c
0
s.
Next lemma is the heart of the series of discussions. It shows that U
s
(x) contains
a resistance ball of radius cs, where c is independent of s.
Lemma 22.6. There exists c
2
> 0 such that R(x, y) c
2
s for any s (0, 1],
any x V

and any y Q
s
(x) V

.
Proof. Set K

= K
s
(x) V

and Q

= Q
s
(x) V

.
Claim 1 Let z (K
s
(x) Q
s
(x))
c
V

. For any a, b, c R, there exists u F


such that u|
K

= a, u|
Q

= b and u(z) = c.
Proof of Claim 1 Set m

= max
w
1
s,x
|w|. We may choose m m

so that
z F
w
(V
0
), K
w
Q
s
(x) = and K
w
K
s
(x) = for some w W
m
. Consid-
ering the resistance form (E
m
, (V
m
)), we nd u (V
m
) such that u|
K
s
(x)V
m
=
a, u|
Q
s
(x)V
m
= b and u(z) = c. Since (E, F) is the limit of the compatible sequence
(V
m
, L
m
), the harmonic extension of u, h
V
m
( u), possesses the desired properties.
Claim 2 Let F
s,x
= {u|u F, u|
K

and u|
Q

are constants}. Then (E, F


s,x
) is a
resistance form on (V

\(K
s
(x) Q
s
(X))) {K

} {Q

}.
Proof of Claim 2 By Claim 1, we see that (K

)
F
= K

. Theorem 4.3 implies


that (E, F
K

) is a resistance form on (V

\K
s
(x)) {K

}. Again by Claim 1,
(Q

)
F
K

= Q

. Using Theorem 4.3, we verify Claim 2.


Claim 3 Let R

(, ) be the resistance metric associated with (E, F


s,x
). Then
R

(K

, Q

) c
2
s for any x V

and any s (0, 1], where c


2
is independent
of x and s.
Proof of Claim 3 Let

V =
w
1
s,x
\
s,x
F
w
(V
0
). Dene V = (

V \(K
s
(x) Q
s
(x)))
{K
0
} {Q
0
}. Note that V is naturally regarded as a subset of (V

\(K
s
(x)
86
Q
s
(X))) {K

} {Q

}. Also, :

V V is dened by
(x) =
_

_
x if x / K
s
(x) Q
s
(x),
K
0
if x K
s
(x),
Q
0
if x Q
s
(X).
Let
E
V
(u, v) =

w
1
s,x
\
s,x
1
r
w
E
D
(u , v )
for any u, v (V ). Then (E
V
, (V )) is a resistance form on V . If R
V
(, ) is the
resistance metric associated with (E
V
, (V )), then R
V
(K
0
, Q
0
) = R

(K

, Q

). Let
us consider R
V
(K
0
, Q
0
). Any path of resistors between K
0
and Q
0
corresponds to
(r
w
)
1
E
D
(u, v) for some w
1
s,x
\
s,x
. Let w
1
s,x
\
s,x
. If F
w
(V
0
)K
s
(x)
or F
w
(V
0
) Q
s
(x) is empty, then this part does not contribute to the eective
resistance between K
0
and Q
0
. So assume that both p
w
= K
s
(x) F
w
(V
0
) and
q
w
= Q
s
(x) F
w
(V
0
) are non-empty. Let r(w) be the eective resistance between
p
w
and q
w
with respect to the resistance form derived from the resistance form
(r
w
)
1
E
D
(, ) on F
w
(V
0
). Since the choice of p
w
and q
w
in F
w
(V
0
) is nite, it
follows that
1
r
w
r(w)
2
r
w
, where
1
and
2
are independent of x, s and
w. Since r
w
rs, we have
3
s r(w)
2
s, where
3
=
1
r. Now R
V
(K
0
, Q
0
)
is the resistance of the parallel circuit with the resistors of resistances r(w). Since
#(
s,x
1
) is uniformly bounded with respect to x and s, in fact 45 is a sucient
upper bound, we have

4
s R
V
(K
0
, Q
0
)
5
s,
where
4
and
5
are independent of x and s. This completes the proof of Claim 3.
Since R

(K

, Q

) R(x, y) for any y Q


s
(x), Claim 3 suces for the proof
of this lemma.
Combining all the lemmas, we nally obtain our main theorem of this section.
Theorem 22.7. The resistance metric R and the Euclidean metric d
E
give the
same topology on V

. Moreover, the identity map on V

is extended to a homeo-
morphism between the completions of (V

, R) and (V

, d
E
).
Proof. If {x
n
}
n1
V

and x V

, then the following three conditions (A),


(B) and (C) are equivalent.
(A) lim
n+
R(x
n
, x) = 0
(B) For any s > 0, there exists N > 0 such that x
n
U
s
(x) for any n N.
(C) lim
n+
|x
n
x| 0.
In fact, by Lemmas 22.5-(2) and 22.6, (A) is equivalent to (B). Lemma 22.4-(2)
shows that (B) is equivalent to (C).
Hence, the identity map between (V

, R) and (V

, d
E
) is homeomorphism. Next
assume that {x
n
}
n1
is a d
E
-Cauchy sequence. Let x K be the limit of {x
n
}
n1
with respect to d
E
. Since U
s
(x) is a neighborhood of x with respect to d
E
by
Lemma 22.4-(2), x
n
U
s
(x) for suciently large n. Lemma 22.5-(2) shows that
{x
n
}
n1
is an R-Cauchy sequence. Conversely assume that {x
n
}
n1
is not a d
E
-
Cauchy sequence. There exist > 0 and subsequences {x
n
i
} and {x
m
i
} such that
|x
n
i
x
m
i
| for any i 1. By Lemma 22.5-(2), we may choose s (0, 1]
so that diam(U
s
(x), d
E
) < . This shows that x
n
i
/ U
s
(x
m
i
). By Lemma 22.6,
it follows that R(x
n
i
, x
m
i
) c
2
s. Hence {x
n
}
n1
is not an R-Cauchy sequence.
23. VOLUME DOUBLING PROPERTY 87
Thus we have shown that the completions of (V

, R) and (V

, d
E
) are naturally
homeomorphic.
By this theorem, we are going to identify the completion of (V

, R) with K. In
other words, the resistance metric R is naturally extended to K. Using [36, Theo-
rem 2.3.10], we think of (E, F) as a resistance form on K and R as the associated
resistance metric from now on. Note that (K, R) is compact and hence (E, F) is
regular.
By the identication described above, Lemmas 22.4, 22.5 and 22.6 imply that
U
s
(x) is comparable with the resistance ball of radius s.
Corollary 22.8. There exist
1
,
2
> 0 such that
B
R
(x,
1
s) U
s
(x) B
R
(x,
2
s)
for any x X and any s (0, 1].
Since (E, F) is a resistance form on K and (K, R) is compact, we immediately
obtain the following result.
Corollary 22.9. (E, F) is a local regular resistance form on K. Moreover,
let be a Borel regular measure on (K, R) with (K) < +. Then (E, F) is a
local regular Dirichlet form on L
2
(K, ).
Definition 22.10. (E, F) constructed in this section is called the resistance
form on K associated with ((D, r
(j)
))
j=1,...,M
. (E, F) is also called the Dirichlet
form associated with ((D, r
(j)
))
j=1,...,M
.
23. Volume doubling property
In this section, we will give a criterion for the volume doubling property of
a measure with respect to the resistance metric in Theorem 23.2. For random
self-similar measures, we will obtain a simpler condition in Theorem 23.8.
As in the last section, (W

, ) is a random Sierpinski gasket generated by


{L
1
, . . . , L
M
}, (D, r
(j)
) is a regular harmonic structure on L
j
for any j and (E, F)
is the resistance form on K associated with {(D, r
(j)
)}
j=1,...,M
. We continue to use
the same notations as in the previous section.
The rst theorem is immediate from Theorem 9.4 and Corollary 22.9.
Theorem 23.1. Let be a nite Borel regular measure on K. Then (E, F) is
a local regular Dirichlet form on L
2
(K, ).
The following theorem gives a necessary and sucient condition for the volume
doubling property with respect to the resistance metric. It is a generalization of
[39, Theorem 1.3.5]. The conditions (EL) and (GE) correspond to (ELm) and (GE)
in [39] respectively.
Theorem 23.2. Let be a nite Borel regular measure on K. has the
volume doubling property with respect to the resistance distance R if and only if the
following two conditions (GE) and (EL) are satised:
(GE) There exists c
1
> 0 such that (K
w
) c
1
(K
v
) for any w, v
s
with
K
w
K
v
= and any s (0, 1].
(EL) There exists c
2
> 0 such that (K
wi
) c
2
(K
w
) for any w W

and any
i S
(w)
.
88
We need two lemmas to prove this theorem.
Lemma 23.3. Let be a nite Borel regular measure on K. has the volume
doubling property with respect to the resistance metric R if and only if there exist
(0, 1) and c > 0 such that (U
s
(x)) c(U
s
(x)) for any x X and any
s (0, 1].
Proof. By Corollary 22.8, B
R
(x,
1
s) U
s
(x) B
R
(x,
2
s). Assume that
(U
s
(x)) c(U
s
(x)). Choose n so that
n

2
<
1
. Then
(B
R
(x,
1
s)) (U
s
(x)) c
n
(U

n
s
(x)) (B
R
(x,
n

2
s)).
Hence has the volume doubling property with respect to R.
Conversely, assume that (B
R
(x, s)) c

(B
R
(x, s)) for some c

> 0 and
(0, 1). Choose n so that
n

2
<
1
. Then
(U
s
(x)) (B
R
(x,
2
s)) (c

)
n
(B
R
(x,
n

2
s)) (c

)
n
(U

2
(
1
)
1
s
(x)).
Letting =
n

2
(
1
)
1
, we have the desired statement.
Lemma 23.4. Let s (0, 1] and let w
s
. If r
2
, then there exists x K
w
such that U
s
(x) K
w
.
Proof. Set w = w
1
w
2
w
m
, where m = |w|. Choose k and l so that k, l
{1, 2, 3}, k = w
m
and l = k. Note that K
wkl
F
w
(V
0
) = . Since r
w
1
w
2
w
m1
> s,
it follows that r
wk
> r
2
s. If r
2
, then r
wklv

s
for some v W

(S).
Set w

= wklv. Choose x K
w

\F
w

(V
0
). By Proposition 21.3-(2) and (4),

s,x
= {w

} and [w

]
m
= w for any w


1
s,x
. Hence U
s
(x) K
w
.
Proof of Theorem 23.2. Assume (GE) and (EL). Fix (0, 1). Let w

s,x
and let wv
s,x
. For any w


1
s,x
, there exists w


s,x
such that
K
w
K
w
= and K
w
K
w
= . Hence by (GE), (K
w
) (c
1
)
2
(K
w
). Since
#(
1
s,x
) 45,
(23.1) (U
s
(x)) 45(c
1
)
2
(K
w
).
(In the above argument we only need that #(
1
s,x
) is uniformly bounded. Since
V
0
is a regular triangle and every K(W

, ) T, we may deduce that the uniform


bound is no greater than 45.) Now, since wv
s
and w
s
,
s < r
w
r
v

sr
v

sr
|v|1
,
where v

= [v]
|v|1
. Letting m

be the integral part of


log
log r
+2, we have |v| m

.
Note that m

only depends on . By (EL), (K


wv
) (c
2
)
m

(K
w
). Hence (23.1)
shows that
(U
s
) (c
2
)
m

(K
w
) (c
2
)
m

(c
1
)
2
1
45
(U
s
(x)).
By Lemma 23.3, has the volume doubling property with respect to R.
Next assume that (GE) do not hold. For any C > 0, there exist s (0, 1] and
w, v
s
with K
w
K
v
= such that (K
v
) C(K
w
). Let (0, r
2
]. By
Lemma 23.4, U
s
(x) K
w
for some x K
w
. Since v
1
s,x
,
(U
s
(x)) (1 +C)(K
w
) (1 +C)(U
s
(x)).
Lemma 23.3 shows that does not have the volume doubling property with respect
to R.
23. VOLUME DOUBLING PROPERTY 89
Finally, if (EL) do not hold, then for any > 0 there exist w W

and i S
(w)
such that (K
wi
) (K
w
). Set s = r
w
. Let (0, r
3
]. Then s r
3
s r
2
r
wi
.
By Lemma 23.4, there exists x K
wi
such that U
s
(x) K
wi
. Now,
(U
s
(x)) (K
wi
) (K
w
) (U
s
(x)).
Using Lemma 23.3, we see that does not have the volume doubling property with
respect to R.
Next we introduce the notion of random self-similar measures, which is a natu-
ral generalization of self-similar measures on ordinary self-similar sets. See Deni-
tion 20.4 for the denition of self-similar measures on generalized Sierpinski gaskets.
Proposition 23.5. Let
(j)
= (
(j)
i
)
iS
j
(0, 1)
S
j
satisfy

iS
j

(j)
i
= 1
for each j = 1, . . . , M. Dene
w
=
([w]
0
)
w
1

([w]
1
)
w
2

([w]
m1
)
w
m
for any w =
w
1
w
2
w
m
W

. Then there exists a unique Borel regular probability measure


on K such that (K
w
) =
w
for any w W

. Moreover, satises the condition


(EL) in Theorem 23.2
Note that the Hausdor measure associated with the resistance metric, which
has been studied in [28, 29, 30] is not a random self-similar measure in general
except for the homogeneous case.
Definition 23.6. The Borel regular probability measure in Proposition 23.5
is called the random self-similar measure on (W

, ) generated by (
(1)
, . . . ,
(M)
).
In the next denition, we introduce a notion describing relations of neighboring
K
w

s for w
s
in order to apply Theorem 23.2.
Definition 23.7. A pair ((j
1
, i
1
), (j
2
, i
2
)) {(j, i)|j = 1, . . . , M, i {1, 2, 3}}
2
is called an adjoining pair for (W

, ) if and only if there exist w, v W

such that
wi
1
, vi
2

s
for some s (0, 1], w = v, j
1
= (w), j
2
= (v) and (w(i
1
)

) =
(v(i
2
)

).
Theorem 23.8. Let be a random self-similar measure on (W

, ) generated
by (
(1)
, . . . ,
(M)
). has the volume doubling property with respect to the resistance
metric R if
(23.2)
log
(j
1
)
i
1
log r
(j
1
)
i
1
=
log
(j
2
)
i
2
log r
(j
2
)
i
2
for any adjoining pair ((j
1
, i
1
), (j
2
, i
2
)) for (W

, ).
Before proving this theorem, we give an example where the condition of the
above theorem is realized.
Example 23.9. Let L
1
= L
SG
and let L
2
= L
SP
, where L
SG
and L
SP
are
the original Sierpinski gasket and the Sierpinski spiral respectively introduced in
Section 20. Set S
1
= {1, 2, 3} and S
2
= {1, 2, 3, 4}. Dene H = {(h, )|0 < h,
(0, min{h, 1/h})}. Fix (h, ) H and set r
(1)
i
= r
SG
i
for i S
1
and r
(2)
i
= r
SP
i
for i S
2
. (Recall that r
SG
i
only depends on h and r
SP
i
depend on h and . See
Examples 20.11 and 20.3.) Denote r
(j)
= (r
(j)
i
)
jS
j
for j = 1, 2. Dene

by the
unique satisfying

iS
1
(r
(1)
i
)

= 1. Note that

depends only on h. When


90
h = 1, then r
(1)
i
= 3/5 for any i S
1
and hence

= log 3/(log 5 log 3). Let

(1)
i
= (r
(1)
i
)

for i S
1
. Dene
H
0
= {(h, )|(h, ) H,

i=1,2,3
(r
(2)
i
)

< 1}.
If h = 1, r
(2)
i
= (1 )/2 for any i {1, 2, 3}. This implies (1, ) H
0
for any
(0, 1). Hence H
0
is a non-empty open subset of R
2
. Let
(2)
i
= (r
(2)
i
)

for any
i {1, 2, 3} and let
(2)
4
= 1

3
i=1

(2)
i
. Applying Theorem 23.8, we have the
following proposition:
Proposition Assume that (h, ) H
0
. Let (W

, ) be any random Sierpinski gas-


ket generated by generated by {L
1
, L
2
}. Let

be the random self-similar measure


generated by ((
(j)
i
)
iS
j
)
j=1,2
and let R be the resistance metric on K = K(W

, )
associated with ((D
h
, r
(j)
))
j=1,2
. Then

has the volume doubling property with


respect to R.
The rest of this section is devoted to the proof of Theorem 23.8.
Proof of Theorem 23.8. Let (1) and (2) belong to (W

, ). Write
(i) = (i)
1
(i)
2
. . . for i = 1, 2. Assume that (1) = wi
1
(k)

, (2) = wi
2
(l)


(W

, ), where w W

\W
0
, i
1
= i
2
S
(w)
, k, l {1, 2, 3} and ((1)) =
((2)). Set r
i,n
= r
(([(i)]
n1
))
(i)
n
and
i,n
=
(([(i)]
n1
))
(i)
n
for i = 1, 2 and n 1.
Dene {m
n
}
n0
and {M
n
}
n0
inductively by
_

_
m
0
= M
0
= |w|
m
n+1
= inf{m|m > m
n
, r
[(1)]
m
= r
[(2)]
m

for some m

}
M
n+1
= inf{m|m > M
n
, r
[(1)]
m

= r
[(2)]
m
for some m

}
(If inf{m|m > m
n
, r
[(1)]
m
= r
[(2)]
m

for some m

} = , then we dene m
N
=
M
N
= + for all N n +1.) Also dene s
n
= r
[(1)]
m
n
for n 0. (If m
n
= +,
then dene s
n
= 0.) Note that s
n
= r
[(1)]
m
n
= r
[(2)]
M
n
.
Claim 1 Let n 1. Then there exists
n
such that
1,m
= (r
1,m
)

n
for any
m = m
n
+ 1, . . . m
n+1
and
2,m
= (r
2,m
)

n
for any m = M
n
+ 1, . . . , M
n+1
.
Proof of Claim 1 For suciently small > 0, [w(1)]
m
n
+1
, [w(2)]
M
n
+1

s
n
+
.
Hence (([w(1)]
m
n
), k), (([w(2)]
M
n
), l) is an adjoint pair. By (23.2),
(23.3)
log
1,m
n
+1
log r
1,m
n
+1
=
log
2,M
n
+1
log r
2,M
n
+1
Set
n
= log
1,m
n
+1
/log r
1,m
n
+1
. Let m
n
+ 1 m < m
n+1
. Then there ex-
ists m

[M
n
, M
n+1
1] such that r
[w(2)]
m

< r
[w(1)]
m
< r
[w(2)]
m

+1
. Set s

=
r
[w(1)]
m
. Then we see that [w(1)]
m
, [w(2)]
m

+1

s

and [w(1)]
m+1
, [w(2)]
m

+1

+
for suciently small > 0. Hence ((([w(1)]
m1
), k), (([w(2)]
m
), l)) and
((([w(1)]
m
, k), (([w(2)]
m
, l)) are adjoint pairs. Using (23.2), we see that
(23.4)
log
1,m
log r
1,m
=
log
2,m

+1
log r
2,m

+1
=
log
1,m+1
log r
1,m+1
.
By similar argument,
(23.5)
log
2,m
log r
2,m
=
log
2,m+1
log r
2,m+1
23. VOLUME DOUBLING PROPERTY 91
for any m = M
n
+1, . . . , M
n+1
1. The equations (23.3), (23.4) and (23.5) imme-
diately imply the claim. (End of Proof of Claim 1)
Claim 2 Set s

= min{r
wi
1
, r
wi
2
}. Dene m

= min{m|s

> r
[(1)]
m
s
1
} and
M

= min{m

|s

> r
[(2)]
m

s
1
}. There exists
0
> 0 such that
1,m
= (r
1,m
)

0
and
2,m
= (r
2,m
)

0
for any m = m

, . . . , m
1
and any m

= M

, . . . , M
1
.
Proof of Claim 2 If m
1
= m
0
+ 1, then s

= s
1
. Hence we have Claim 2.
Similarly, if M
1
= M
0
+ 1, then we have Claim 2. Thus we may assume that
m
1
m
0
+ 2 and M
1
M
0
+ 2. Then [(1)]
m
1
, [(2)]
M
1

s
1
, and so it follows
that (([(1)]
m
1
1
), k), (([(2)]
M
1
1
), l)) is an adjoining pair. By (23.2),
(23.6)
log
1,m
1
log r
1,m
1
=
log
2,M
1
log r
2,M
1
Let m {m

, . . . , m
1
1}. Then there exists m

[M
0
+ 1, M
1
1] such that
r
[w(2)]
m

< r
[w(1)]
m
< r
[w(2)]
m

+1
. Using similar argument as in the proof of Claim
1, we obtain counterparts of (23.4) and (23.5). These equalities along with (23.6)
yield the claim. (End of Proof of Claim 2)
Claim 3 Dene L = min{n|n N, r
n
< r}. If [(1)]
m
, [(2)]
m

s
for some
s [s
1
, s
0
), then
(23.7) ()
L
(r)

0
(L+1)

[(1)]
m

[(2)]
m

()
L
(r)

0
(L+1)

[(1)]
m
,
where = min{
(j)
i
|j {1, . . . , M}, i {1, 2, 3}}.
Proof of Claim 3 Assume that r
wi
1
= s

. Note that m

= m
0
+ 2. First we
consider the case where s [s

, s
0
). It follows that m = m
0
+ 1 and wi
1

s
.
Since r
[(2)]
m

1
s

= r
wi
1
, we have r
2,M
0
+1
r
2,m

1
r
1,m
0
+1
. This shows
that r
m

M
0
1
r. Therefore, m

M
0
L. Now
w

wi
1

w
and
w

[(2)]
m


w
()
m

M
0

w
()
L
. This immediately imply Claim 3 in this case.
Next suppose s [s
1
, s

). By Claim 2,
1,m
0
+2

1,m
= (r
1,m
0
+2
r
1,m
)

0
and

2,M


2,m
= (r
2,M

r
2,m
)

0
. On the other hand, r
[(1)]
m1
> s r
[(1)]
m
.
Hence
s
r
wi
1
r
1,m
0
+2
r
1,m

rs
r
wi
1
.
This implies
(23.8)

w
s

0
(r
w
)

0
1
r

[(1)]
m


w
s

0
(r
w
)

0
r

0
.
Similarly, we have
s
r
[(2)]
M

1
r
2,M

r
2,m

rs
r
[(2)]
M

1
and hence
(23.9)

w
s

0
(r
w
)

0
1
r

0
L

[(2)]
m



w
s

0
(r
w
)

L
r

0
Combining (23.8) and (23.9), we have the claim. If r
wi
2
= s

, then an analogous
discussion yields the claim as well. (End of Proof of Claim 3)
Claim 4 Dene

= max{log
(j)
i
/ log r
(j)
i
|j {1, . . . , M}, i {1, 2, 3}}. If
[(1)]
m
, [(2)]
m

s
for some s (0, s
0
), then
(23.10) ()
L
(r)

(L+2)

[(1)]
m

[(2)]
m

()
L
(r)

(L+2)

[(1)]
m
92
Proof of Claim 4 By Claim 1,
1,m
n
+1

2,m
n+1
= (r
1,m
n
+1
r
1,m
n+1
)

n
=
(r
2,M
n
+1
r
2,M
n+1
)

n
=
2,M
n
+1

2,M
n+1
. for any n 1. Suppose s
[s
p+1
, s
p
) for some p 1. Then

[(1)]
p

[(2)]
p
=

1,m
0
+1

1,m
1

2,M
0
+1

2,M
1


1,m
p
+1

1,m

2,M
p
+1

2,m

.
By Claim 3, we have an estimate of the rst part of the right-hand side of the
above equality. For the second part,
1,m
p
+1

1,m
= (r
1,m
p
+1
r
1,m
)

p
. On
the other hand, s/s
p
r
1,m
p
+1
r
1,m
rs/s
p
. Hence we have
_
s
s
p
_

p

1,m
p
+1

1,m
(r)

p
_
s
s
p
_

p
.
Similarly,
_
s
s
p
_

p

2,M
p
+1

2,m
(r)

p
_
s
s
p
_

p
.
Hence (
1,m
p
+1

1,m
)/(
2,M
p
+1

2,m
) (r)

p
(r)

. Combining this
with Claim 3, we obtain Clam 4. (End of Proof of Claim 4)
Finally, we prove the theorem. If w(1), w(2)
s
for some s (0, 1], w(1) =
w(2) and K
w(1)
K
w(2)
= , then w(1) = [(1)]
m
and w(2) = [(2)]
m
for
some (1) = wi
1
(k)

, (2) = wi
2
(l)

(W

, ), where w W

\W
0
, i
1
=
i
2
S
(w)
, k, l {1, 2, 3} and ((1)) = ((2)). By Claim 4, (K
w(2)
)
()
L
(r)

(L+2)
(K
w(1)
). Hence we have (GE). Proposition 23.5 shows that
satises (EL). Using Theorem 23.2, we see that has the volume doubling prop-
erty with respect to R.
24. Homogeneous case
In this section, we treat a special class of random Sierpinski gasket called ho-
mogeneous random Sierpinski gaskets. In this case, the Hausdor measure is a
random self-similar measure and it is always volume doubling with respect to the
resistance metric. The associated diusion process has been extensively studied in
[27, 29, 7]. Most of the results in this section are the reproduction of their works
from our view point.
As in the previous sections, L
j
= (K(j), S
j
, {F
i
j
}
iS
j
) is a generalized Sierpinski
gasket for j = 1, . . . , M and S
j
= {1, . . . , N
j
}.
Definition 24.1. Let (W

, ) be a random Sierpinski gasket generated by


{L
1
, . . . , L
M
}.
(1) (W

, ) is called homogeneous if and only if (w) = (v) for any w, v W


m
and for any m 0.
(2) Let (W

, ) be homogeneous. For m 1, we dene


m
= (w) for w
W
m1
. Set
(j)
i
= (N
j
)
1
for any j = 1, . . . , M and any i S
j
. The random self-
similar measure on (W

, ) generated by {(
(1)
i
)
iS
1
, . . . , (
(M)
i
)
iS
M
} is called
the canonical measure on (W

, ).
This canonical measure coincides with the measure used in [27, 28, 29]. We
will show in Theorem 24.5 that the canonical measure is equivalent to the Hausdor
measure associated with the resistance metric.
Throughout this section, (W

, ) is a homogeneous random Sierpinski gasket.


Let (D, r
(j)
) be a regular harmonic structure on L
j
for each j = 1, . . . , M. We will
also require homogeneity for the resistance scaling ratio r
(j)
. Namely, the following
24. HOMOGENEOUS CASE 93
Figure 4. Homogeneous random Sierpinski gaskets
condition (HG):
(HG) r
(j)
i
1
= r
(j)
i
2
for any j and any i
1
, i
2
S
j
.
is assumed hereafter in this section. Under (HG), we write r
(j)
i
= r
(j)
.
Proposition 24.2. Assume (HG).
(1) Dene r(m) = r
(
1
)
r
(
m
)
. Then
s
= W
m
for s (r(m1), r(m)].
(2) Let be the canonical measure on (W

, ). Then (K
w
) = #(W
m
)
1
=
(N

1
N

m
)
1
for any w W
m
.
Note that j
1
= j
2
for any adjoining pair ((j
1
, i
1
), (j
2
, i
2
)) in the case of a
homogeneous random Sierpinski gasket. Hence by Theorem 23.8, we immediately
obtain the following result.
Theorem 24.3. Assume (HG). The canonical measure has the volume dou-
bling property with respect to the resistance metric R on K = K(W

, ) associated
with ((D, r
(j)
))
j=1,...,M
.
We can describe more detailed structure of the canonical measure in terms
of the resistance metric.
Definition 24.4. Dene (s) = #(
s
)
1
for any s (0, 1]. For s 1, we
dene (s) = (1). For any > 0 and any A K, we dene
H

(A) = inf
_

i1
(diam(U
i
, R))|A
i1
U
i
, diam(U
i
, R) for any i 1
_
and H

(A) = lim
0
H

(A). H

is called the -Hausdor measure of (K, R).


It is known that H

is a Borel regular measure. See [46]. The next theorem


shows that is equivalent to the -Hausdor measure.
Theorem 24.5. Assume (HG). The canonical measure is equivalent to the -
dimensional Hausdor measure H

of (K, R). More precisely, there exist c


1
, c
2
> 0
such that
(24.1) c
1
(A) H

(A) c
2
(A)
94
for any Borel set A K and
(24.2) c
1
(r) (B
R
(x, r)) c
2
(r)
for any x X and any r (0, 1].
Proof. Note that has the doubling property, i.e. (2s) c(s) for any
s, where c is independent of s. Now, if w
s
, then (K
w
) = (s). Since
1 #(
1
s,x
) 45,
(24.3) (s) (U
s
(x)) 45(s)
By Corollary 22.8, (B
R
(x,
1
s)) (U
s
(x)) (B
R
(x,
2
s)). This along with
(24.3) and the doubling property of implies (24.2).
Next we show (24.1). By Lemma 22.5-(1),
3
r
w
diam(K
w
, R)
4
r
w
for
any w W

, where
3
and
4
are independent of w. Since (r
w
) = (K
w
), we
have
6
(K
w
) (diam(K
w
, R))
7
(K
w
). Let A be a compact subset of
K. Dene
s
(A) = {w|w
s
, K
w
A = } and K
s
(A) =
w
s
(A)
K
w
. Then

n1
K
1/n
(A) = A. Note that max
w
s
diam(K
w
, R)
8
s. Hence,
H

8
s
(A)

w
s
(A)
(diam(K
w
, R))

w
s
(A)

7
(K
w
)
7
(K
s
(A)).
Letting s 0, we obtain H

(A)
7
(A) for any compact set A. Since both H

and are Borel regular, H

(A)
7
(A) for any Borel set A. Finally, let A be a
Borel set. Suppose A
i1
U
i
. Choose x
i
U
i
. Then by (24.2)
(A)

i1
(U
i
)

i1
(B
R
(x, diam(U
i
, R))) c
2

i1
(diam(U
i
, R))
This shows (A) c
2
H

(A). Thus we have (24.1).


By (24.3), we have the uniform volume doubling property, which has been
dened in [41]. By the above theorem, we have
R(x, y)V
R
(x, R(x, y)) R(x, y)(R(x, y)),
Hence Theorem 15.10 implies the following theorem. (In fact, since we have the
uniform volume doubling property, [41, Theorem 3.1] suces to have (24.4) and
(24.5). Recall the remark after Theorem 15.10.)
Theorem 24.6. Let (E, F) be the regular local Dirichlet form on L
2
(K, ) as-
sociated with ((D, r
(j)
))
j=1,...,M
. Assume (HG). There exists a jointly continuous
heat kernel p(t, x, y) associated with the Dirichlet form (E, F) on L
2
(K, ). Dene
g(r) = r(r). (Note that (g
1
(t)) t/g
1
(t).) Then
(24.4) p(t, x, x)
g
1
(t)
t
for any t > 0 and any x K and
(24.5) p(t, x, y)
c
1
g
1
(t)
t
exp
_
c
2
_
R(x, y)

1
(t/R(x, y))
_
_
for any t > 0 and any x, y K.
24. HOMOGENEOUS CASE 95
By (24.4), the uctuation from a power law in the on-diagonal behavior of heat
kernels given in [27, 29] is now understood as the uctuation of (r) versus r

,
where is the Hausdor dimension of (K, R).
Unfortunately, the resistance metric is not (equivalent to a power of) a geodesic
metric in general, and hence (24.5) may not be best possible. To construct a
geodesic metric, we dene the notion of n-paths for homogeneous random Sierpinski
gaskets in the similar way as in Denition 20.8.
Theorem 24.7. Assume that L
j
admits symmetric self-similar geodesics with
the ratio
j
. Set (m) =

1
. . .

j
. Then there exists a geodesic metric d on K
which satises
d(p, q) = (n) min{m1|(x
1
, . . . , x
m
) is an n-path between p and q}
for any p, q V
n
. Moreover, assume (HG) and that r
(j)
/N
j
<
j
for any j =
1, . . . , M. Set (m) = #(W
m
)
1
and T
m
= (m)r(m) for any m 0. Dene

m
= log T
m
/ log (m) and
h(s) =
_
s

m
if T
m
s < T
m1
,
s
2
if t 1.
Then
(24.6)
c
1
V
d
(x, h
1
(t))
exp
_
c
2
_
d(x, y)

1
(t/d(x, y))
_
_
p(t, x, y)

c
3
V
d
(x, h
1
(t))
exp
_
c
4
_
d(x, y)

1
(t/d(x, y))
_
_
,
where (s) = h(s)/s.
The above two-sided o-diagonal estimate is essentially same as that obtained
by Barlow and Hambly in [7]. More precisely, they have shown (24.7) and (24.8)
given below.
We will prove this theorem at the end of this section.
Remark. (24.6) has equivalent expressions. Set
m
= log (m)/ log (m).
Then (24.6) is equivalent to
(24.7)
c
5
t

m
/
m
exp
_
c
6
_
d(x, y)

n
t
_ 1

n
1
_
p(t, x, y)

c
7
t

m
/
m
exp
_
c
8
_
d(x, y)

n
t
_ 1

n
1
_
.
if T
m
t < T
m1
and T
n
/(n) t/d(x, y) < T
n1
/(n 1).
Also (24.6) is equivalent to
(24.8)
c
9
(m)
exp
_
c
10
T
m
T
n
_
p(t, x, y)
c
11
(m)
exp
_
c
12
T
m
T
n
_
.
if T
m
t < T
m1
and T
n
/(n) t/d(x, y) < T
n1
/(n 1).
96
Example 24.8. As in Example 23.9, let L
1
= L
SG
and let L
2
= L
SP
. We
consider homogeneous random Sierpinski gasket generated by {L
1
, L
2
}. See Fig-
ure 4. Fix h = 1 and set r
(1)
= (3/5, 3/5, 3/5) and r
(2)
= (1/3, 1/3, 1/3, 1/3).
Then (D
1
, r
(j)
) is a regular harmonic structure on L
j
for j = 1, 2. Note that
((D
1
, r
(j)
))
j=1,2
satises the assumption (HG). In this case,
(1)
i
= 1/3 for i S
1
and
(2)
i
= 1/4 for i S
2
. Also in this case, by Examples 20.11 and 20.12, both
L
1
and L
2
admit symmetric self-similar geodesics with the geodesic ratio
1
= 1/2
and
2
= 1/3 respectively. Since r
(1)
/N
1
= 1/5 <
1
and r
(2)
/N
2
= 1/12 <
2
, we
have (24.6), (24.7) and (24.8).
The rest of this section is devoted to proving Theorem 24.7. The existence of a
geodesic distance d is shown by similar argument in the proof of [35, Theorem 4.3].
Using the same discussion, we obtain the following lemma as well.
Lemma 24.9. Dene (m) = max
wW
m
diam(K
w
, d). Then (m) (m).
Lemma 24.10. For x, y K, dene M(x, y) = inf{m|y / U
r(m)
(x)}. Then
(1) For some m

N,
1 inf{n|(x
0
, . . . , x
n
) is an M(x, y)-path and there exist
w(1), w(2) W
M(x,y)
such that x, x
0
K
w(1)
and x
n
, y K
w(2)
} m

for any x, y K,
(2) R(x, y) r(M(x, y)),
(3) d(x, y) (M(x, y))
Proof. Note that
r(k)
= W
k
for any k. Since y U
r(m1)
(x)\U
r(m)
(x) for
m = M(x, y), we have (1). Combining (1) and Corollary 22.8, we obtain (2). (3)
follows from Lemma 24.9 and (1).
Lemma 24.11. d
QS
R.
Proof. By Lemma 24.10-(2), for any n 1, there exists
n
> 0 such that
R(x, z)
n
R(x, y) implies M(x, z) M(x, y)+n. Fix (0, 1). By Lemma 24.10-
(2), for suciently large n, d(x, z) d(x, y) whenever M(x, z) M(x, y) + n.
Hence d is (SQS)
R
. The similar discussion shows that R is (SQS)
d
. Hence we have
d
QS
R by Theorem 12.3.
Lemma 24.12. V
d
(x, d(x, y)) (M(x, y)).
Proof. Since is (VD)
R
and d
QS
R, is (VD)
d
. Lemma 24.10-(3) implies
V
d
(x, d(x, y)) V
d
(x, (M(x, y)). Set m = M(x, y). Note that B
d
(x, (m))
U
r(m)
(x). Hence
V
d
(x, (m)) (U
r(m)
(x)) #(
1
r(m),x
)(m) C(m),
where C is independent of x and m. On the other hand, if w W
m
and x K
w
,
then K
w
B
d
(x, (m)). Combining this with Lemma 24.9, we obtain (m) =
(K
w
) V
d
(x, (m)) cV
d
(x, (m)). Thus, we have the desired result.
Proof of Theorem 24.7. By Lemmas 24.10 and 24.12, we obtain
R(x, y)V
d
(x, d(x, y)) T
M(x,y)
h(d(x, y)).
25. INTRODUCING RANDOMNESS 97
Hence (DM2)
h,g
follows. At the same time, h is a monotone function with full
range and doubling. Now we have the condition (b) of Theorem 15.10. Moreover,
since r
(j)
/N
j
<
j
for any j, is a monotone function with full range and decays
uniformly. Hence Theorem 15.10 implies (24.6).
25. Introducing randomness
Finally in this section, we introduce randomness in the random Sierpinski gas-
kets. As mentioned before, the Hausdor measure associated with the resistance
metric is almost surely not (equivalent to) a random self-similar measure.
As in the previous section, we x a family of generalized Sierpinski gaskets
{L
1
, . . . , L
M
}. Let L
j
= (K(j), S
j
, {F
j
i
}
iS
j
), where S
j
= {1, . . . , N
j
}. Set N =
max
j=1,...,M
N
j
and S = {1, . . . , N} as before.
Definition 25.1. Let =
{(W

, )|(W

, ) is a random Sierpinski gasket generated by {L


1
, . . . , L
M
}}.
Dene
w,j
= {(W

, )|(W

, ) , w W

, (i) = j} for i W

(S) and j
{1, . . . , M}. Let B
m
be the -algebra generated by {
w,j
|w
m1
n=0
W
n
(S), j
{1, . . . , M}} and dene B =
m1
B
m
.
For = (W

, ) , we write W

() = W

, () = , K

= K(W

, ),
W

m
= W
m
(W

, ) and so on.
According to [28, 29], we have the following fact.
Proposition 25.2. Let (
j
)
j=1,...,M
(0, 1)
M
satisfy

M
j=1

j
= 1. Then
there exists a probability measure P on (, B) such that {
w,j
|w W

(S), j
{1, . . . , M}} is independent and P((w) = j|w W

) =
j
for any w W

(S) and
any j {1, . . . , M}.
We x such a probability measure P on (, B) as in Proposition 25.2.
Now let (D, r
(j)
) be a regular harmonic structure on L
j
for j = 1, . . . , M. We
use (E

, F

) to denote the resistance form on K

associated with ((D, r


(j)
))
j=1,...,M
for . In [28], Hambly has introduced a probability measure on K

which
is natural from the view point of the resistance metric in the following way.
Definition 25.3. Let = (W

, ) . Choose x
w
K

w
for w W

. For
n 1, dene

n
=

wW

m
(r
w
)
1

vW

m
(r
v
)
1

x
w
,
where
x
is the Diracs point mass. Let =

be one of the accumulating points


of {
n
} in the weak sense.
Note that since K

is compact, {
n
} has accumulating points. This measure

is known to be equivalent to the proper dimensional Hausdor measure and it


is not a random self-similar measure for P-a. s. . See [28, 29] for details.
In [28, 30], Hambly and Kumagai have shown some uctuations in the asymptotic
behavior of heat kernels associated with the Dirichlet form (E

, F

) on L
2
(K

)
for P-a. s. . In particular, by [30, Theorem 5.5], we have the following
theorem.
Theorem 25.4.

is not (VD)
R
for P-a. s. .
98
As in the homogeneous case, a uctuation of the diagonal behavior of heat
kernels from a power law has been shown in [30] as well. By the above theorem,
however, the uctuation in this case may be caused by the lack of volume doubling
property. (Recall that the volume doubling property always holds in the homoge-
neous case.) Hence those two uctuations in homogeneous and non-homogeneous
cases are completely dierent in nature.
Proof. Using [30, Theorem 5.5], we see that (GE) do not hold for P-a. s. .
Hence by Theorem 23.2,

is not (VD)
R
for P-a. s. .
Bibliography
[1] M. T. Barlow, Diusion on fractals, Lecture notes Math. vol. 1690, Springer, 1998.
[2] M. T. Barlow and R. F. Bass, Transition densities for Brownian motion on the Sierpinski
carpet, Probab. Theory Related Fields 91 (1992), 307330.
[3] , Coupling and Harnack inequalities for Sierpinski carpets, Bull. Amer. Math. Soc.
(N. S.) 29 (1993), 208212.
[4] , Brownian motion and harmonic analysis on Sierpinski carpets, Canad. J. Math. 51
(1999), 673744.
[5] M. T. Barlow, R. F. Bass, Z.-Q. Chen, and M. Kassmann, Non-local Dirichlet forms and
symmetric jump process, Trans. Amer. Math. Soc. 361 (2009), 19631999.
[6] M. T. Barlow, T. Coulhon, and T. Kumagai, Characterization of sub-Gaussian heat kernel
estimates on strongly recurrent graphs, Comm. Pure Appl. Math. 58 (2005), 16421677.
[7] M. T. Barlow and B. M. Hambly, Transition density estimates for Brownian motion on scale
irregular Sierpinski gaskets, Ann. Inst. H. Poincare 33 (1997), 531557.
[8] M. T. Barlow, A. A. Jarai, T. Kumagai, and G. Slade, Random walk in the incipient innite
cluster for oriented percolation in high dimensions, Comm. Math. Phys. 278 (2008), 385431.
[9] M. T. Barlow and E. A. Perkins, Brownian motion on the Sierpinski gasket, Probab. Theory
Related Fields 79 (1988), 542624.
[10] M.T. Barlow, R. F. Bass, and T. Kumagai, Stability of parabolic Harnack inequalities on
metric measure spaces, J. Math. Soc. Japan 58 (2006), 485519.
[11] R. F. Bass, Probabilistic Techniques in Analysis, Probability and its Applications, Springer-
Verlag, 1995.
[12] K. Bogdan, A. Stos, and P. Sztonyk, Harnack inequality for stable processes on d-sets, Studia
Math. 158 (2003), 163198.
[13] M. Brelot, On Topologies and Boundaries in Potential Theory, Lecture Note in Math., vol.
175, Springer-Verlag, 1971.
[14] A. Buerling and L. Ahlfors, The boundary correspondence under quasiconformal mappings,
Acta Math. 96 (1956), 125142.
[15] Z.-Q. Chen and T. Kumagai, Heat kernel estimates for stable-like processes on d-sets, Sto-
chastic Process Appl. 108 (2003), 2762.
[16] , Heat kernel estimates for jump processes of mixed types on metric measure spaces,
Probab. Theory Related Fields 140 (2008), 277317.
[17] D. A. Croydon, Heat kernel uctuations for a resistance form with non-uniform volume
growth, Proc. London Math. Soc. (3) 94 (2007), 672694.
[18] E. B. Davies, Heat Kernels and Spectral Theory, Cambridge Tracts in Math. vol 92, Cam-
bridge University Press, 1989.
[19] P. G. Doyle and J. L. Snell, Random Walks and Electrical Networks, Math. Assoc. Amer.,
Washington, 1984.
[20] T. Fujita, Some asymptotics estimates of transition probability densities for generalized dif-
fusion processes with self-similar measures, Publ. Res. Inst. Math. Sci. 26 (1990), 819840.
[21] M. Fukushima, Y. Oshima, and M. Takeda, Dirichlet Forms and Symmetric Markov Pro-
cesses, de Gruyter Studies in Math. vol. 19, de Gruyter, Berlin, 1994.
[22] A. Grigoryan, Heat kernel upper bounds on fractal spaces, preprint 2004.
[23] , The heat equation on noncompact Riemannian manifolds. (in Russian), Mat. Sb.
182 (1991), 5587, English translation in Math. USSR-Sb. 72(1992), 4777.
[24] , Heat kernels and function theory on metric measure spaces, Cont. Math. 338 (2003),
143172.
99
100 BIBLIOGRAPHY
[25] A. Grigoryan and A. Telcs, Sub-Gaussian estimates of heat kernels on innite graphs, Duke
Math. J. 109 (2001), 451 510.
[26] , Harnack inequalities and sub-Gaussian estimates for random walks, Math. Ann. 324
(2002), 521556.
[27] B. M. Hambly, Brownian motion on a homogeneous random fractal, Probab. Theory Related
Fields 94 (1992), 138.
[28] , Brownian motion on a random recursive Sierpinski gasket, Ann. Probab. 25 (1997),
10591102.
[29] , Heat kernels and spectral asymptotics for some random Sierpinski gaskets, Frac-
tal Geometry and Stochastics II (C. Bandt et al., eds.), Progress in Probability, vol. 46,
Birkhauser, 2000, pp. 239267.
[30] B. M. Hambly and T. Kumagai, Fluctuation of the transition density of Brownian motion
on random recursive Sierpinski gaskets, Stochastic Process Appl. 92 (2001), 6185.
[31] W. Hebisch and L. Salo-Coste, On the relation between elliptic and parabolic Harnack in-
equalities, Ann. Inst. Fourier 51 (2001), 14271481.
[32] J. Heinonen, Lectures on Analysis on Metric Spaces, Springer, 2001.
[33] J. A. Kelingos, Boundary correspindence under quasiconformal mappings, Michigan Math.
J. 13 (1966), 235249.
[34] J. Kigami, Harmonic calculus on limits of networks and its application to dendrites, J.
Functional Analysis 128 (1995), 4886.
[35] , Hausdor dimensions of self-similar sets and shortest path metrics, J. Math. Soc.
Japan 47 (1995), 381404.
[36] , Analysis on Fractals, Cambridge Tracts in Math. vol. 143, Cambridge University
Press, 2001.
[37] , Harmonic analysis for resistance forms, J. Functional Analysis 204 (2003), 399444.
[38] , Local Nash inequality and inhomogeneity of heat kernels, Proc. London Math. Soc.
(3) 89 (2004), 525544.
[39] , Volume doubling measures and heat kernel estimates on self-similar sets, Memoirs
of the American Mathematical Society 199 (2009), no. 932.
[40] T. Kumagai, Some remarks for stable-like jump processes on fractals, Fractals in Graz 2001
(P. Grabner and W. Woess, eds.), Trends in Math., Birkhauser, 2002, pp. 185196.
[41] , Heat kernel estimates and parabolic Harnack inequalities on graphs and resistance
forms, Publ. Res. Inst. Math. Sci. 40 (2004), 793818.
[42] P. Li and S.-T. Yau, On the parabolic kernel of the Schrodinger operator, Acta Math. 156
(1986), 153 201.
[43] R. Lyons and Y. Peres, Probability on Trees and Networks, Book in preparation, current
version available at http://mypage.iu.edu/rdlyons/.
[44] R. D. Mauldin and S. C. Williams, Random recursive constructions: asymptotic geometric
and topological preperties, Trans. Amer. Math. Soc. 295 (1986), 325346.
[45] V. Metz, Shorted operators: an application in potential theory, Linear Algebra Appl. 264
(1997), 439455.
[46] C. A. Rogers, Hausdor Measures, Cambridge Math. Library, Cambridge University Press,
1998, First published in 1970, Reissued with a foreword by K. Falconer in 1998.
[47] L. Salo-Coste, A note on Poincare, Sobolev, and Harnack inequalities, Internat. Math. Res.
Notices (1992), 2738.
[48] S. Semmes, Some Novel Types of Fractal Geometry, Oxford Math. Monographs, Oxford
University Press, 2001.
[49] P. M. Soardi, Potential Theory on Innite Networks, Lecture Note in Math., vol. 1590,
Springer-Verlag, 1994.
[50] A. Telcs, The Einstein relation for random walks on graphs, J. Stat. Phys 122 (2006), 617
645.
[51] P. Tukia and J. Vaisala, Quasisymmetric embeddings of metric spaces, Ann. Acad. Sci. Fenn.
Ser. A I Math. 5 (1980), 97114.
[52] W. Woess, Random walks on innite graphs and groups a surveey on selected topics, Bull.
London Math. Soc. 26 (1994), 160.
[53] , Denumerable Markov Chains, European Math. Soc., 2009.
[54] K. Yosida, Functional Analysis, sixth ed., Classics in Math., Springer, 1995, originally pub-
lished in 1980 as Grundlehren der mathematischen Wissenschaften band 123.
ASSUMPTIONS, CONDITIONS AND PROPERTIES IN PARENTHESES 101
Assumptions, Conditions and Properties in Parentheses
(ACC), 23
(ASQC)
d
, 41
(C1), 60
(C1), 60
(C2), 60
(C3), 60
(C3), 60
(C4), 60
(C5), 60
(C6), 60
(DHK)
g,d
, 58
(DM1), 46
(DM1)
g,d
, 58
(DM2), 46
(DM2)
g,d
, 58
(DM3), 46
(EIN)
d
, 57
(EL), 87
(GE), 87
(GF1), 14
(GF2), 14
(GF3), 14
(GF4), 14
(GSG1), 77
(GSG2), 77
(GSG3), 77
(GSG4), 77
(H1), 45
(H2), 45
(HG), 93
(HK)
g,d
, 59
(HKA), 64
(HKB), 64
(HKC), 64
(KD), 58
(L1), 12
(L2), 12
(L3), 12
(LYD)
,d
, 4
(LYU)
,d
, 3
(NDL)
,d
, 3
(QD1), 52
(QD2), 52
(QD3), 52
(R1), 21
(R2), 21
(R4), 21
(RES), 25
(RF1), 10
(RF2), 10
(RF3), 10
(RF3-1), 11
(RF3-2), 11
(RF4), 10
(RF5), 10
(RSG), 81
(SQC)
d
, 41
(SQS)
d
, 40
(TD1), 32
(TD2), 32
(TD3), 32
(TD4), 32
(WG1), 11
(WG2), 11
(WG3), 11
(wASQC)
d
, 41
R(), 5
102 Indexes
List of Notations
B, 97
B
Y
R
(x, r), 27
B
F
, 8
B
m
, 97
C
0
(X), 21
Cap(), 29
C
F
, 9
C
s
(x), 84
d

(), 40
d
E
(, ), 61
d(x, r), 41
D, 28
D
U
, 31
E
S
(, )
for a compatible sequence S, 13
E
U
(, )
for a subset U, 31
E
1
(, ), 28
E
L
(, )
for a Laplacian L on a nite set,
12
E |
Y
(, ), 26
F(B), 8
F
+
+
, 20
F
B
, 15
F|
Y
, 25
F
S
for a compatible sequence S, 13
g
x
1
, 30
g
B
(, ), 14
g
x
B
, 14
h
Y
, 26
h
d
(x, r), 56
H
Y
, 26
K(W

, ), 82
K
m,x
(W

, ), 82
K
s
(x), 84
K
w
(W

, ), 82
(), 8
LA(V ), 12
L
(E,F),V
, 13
L
SG
, 79
L
SP
, 80
M, 30
M
U
, 31
N(B, r), 18
n(x), 82
N
d
(B, r), 18
O
F
, 20
O
F
(O, ), 20
O
F
(), 20
O
R
, 20
p
t
, 31
p
U
(t, x, y), 32
Q
s
(x), 84
R
+
+
, 19
R
d
(x, r), 56
R(U, V ), 18
R(x, A), 32
R|
Y
, 26
supp(), 21
T
m
(W

, ), 81
U
s
(x), 84
V
d
(x, r), 1
V
m
(W

, ), 81
W

(S), 76
W
w

, 82
W
m
, 81
W
m
(S), 76
X
B
, 15

V
U
, 8

w
, 82

s
, 84

s,x
, 84
LIST OF NOTATIONS 103

1
s,x
, 84

U
, 31
, 97

w,j
, 97
, 82

x
B
, 15

(), 46
(S), 77

w
, 77

w
, 77

w
(S), 77

U
, 31

x
1
, 30
(, )
V
, 8
( )

, 82
[ ]
n
, 77
, 6

QS
, 43
|| ||
,K
, 21
|| ||

, 21
#(), 18
104 Indexes
Index
adjoining pair, 89
Ahlfors regular, 51
-stable process, 5
annulus comparable condition, 23
annulus semi-quasiconformal, 41
canonical measure, 92
capacity, 29
C
F
-topology, 17
chain condition, 59
closed ball, 40
compatible sequence, 13
contraction, 23
decay uniformly
among metrics, 42
function, 58
diameter, 40
Dirichlet form
associated with ((D, r
(j)
))
j=1,...,M
,
87
induced by a resistance form, 28
doubling
among metrics, 41
doubling property
of a function, 45, 57
of a heat kernel, 58
of a metric space, 24
doubling space, 24
Einstein relation, 4, 57
exit time, 31
ne topology, 19
generalized Sierpinski gasket, 77
geodesic, 78
geodesic metric, 78
Green function, 14
GSG, 77
harmonic function, 13, 26
harmonic structure
regular, 77
heat kernel, 32
Laplacian, 12
length
of a word, 76
Li-Yau type on-diagonal estimate, 4
local property
resistance form, 23
Markov property, 2
metric measure space, 51
monotone function with full range, 58
near diagonal lower estimate, 3
on-diagonal hear kernel estimate
of order g, 58
original Sierpinski gasket, 76
partition, 82
path, 78
QS, 43
quasi continuous, 30
quasidistance, 52
quasisymmetric, 43
random self-similar measure, 89
random self-similar set
homogeneous, 92
random Sierpinski gasket, 81
R(), 5
regular
resistance form, 21
resistance estimate, 25
resistance form, 10
associated with ((D, r
(j)
))
j=1,...,M
,
87
associated with a Laplacian, 12
shorted, 15
resistance metric, 2, 11
associated with a trace, 27
shorted, 15
R-topology, 17
S-spiral, 80
self-similar measure, 77
self-similar set, 23
semi-quasiconformal, 41
semi-quasisymmetric, 40
INDEX 105
Sierpinski gasket, 76
Sierpinski spiral, 80
spectral dimension, 79
stable under the unit contraction, 9
standard inner product, 8
sub-Gaussian estimate
Li-Yau type, 1
sub-Gaussian upper estimate, 60
Li-Yau type, 3
symmetric geodesic ratio, 78
symmetric self-similar geodesics, 78
trace
of a resistance form, 26
transition density, 32
transition semigroup, 31
uniform volume doubling property, 59
uniformly perfect, 22
unit contraction, 9
volume doubling property, 24
walk dimension, 79
weak annulus semi-quasiconformal, 41
weighted graph, 11
-Hausdor measure, 93