Uniqueness of Brownian motion on

Sierpinski carpets
Martin T. Barlow

Department of Mathematics, University of British Columbia
Vancouver B.C. Canada V6T 1Z2
Email: barlow@math.ubc.ca
Richard F. Bass

Department of Mathematics, University of Connecticut
Storrs CT 06269-3009 USA
Email: bass@math.uconn.edu
Takashi Kumagai

Department of Mathematics, Faculty of Science
Kyoto University, Kyoto 606-8502, Japan
Email: kumagai@math.kyoto-u.ac.jp
and
Alexander Teplyaev

Department of Mathematics, University of Connecticut

Storrs CT 06269-3009 USA
Email: teplyaev@math.uconn.edu
13 July 2009
Abstract
We prove that, up to scalar multiples, there exists only one local
regular Dirichlet form on a generalized Sierpinski carpet that is invari-
ant with respect to the local symmetries of the carpet. Consequently
for each such fractal the law of Brownian motion is uniquely determined
and the Laplacian is well dened.

Research partially supported by NSERC (Canada), and EPSRC (UK).

Research partially supported by NSF grant DMS-0601783.

Corresponding author

Research partially supported by the Grant-in-Aid for Scientic Research (B) 18340027.

Research partially supported by NSF grant DMS-0505622.

1
1 Introduction
The standard Sierpinski carpet F
SC
is the fractal that is formed by taking the unit
square, dividing it into 9 equal subsquares, removing the central square, dividing
each of the 8 remaining subsquares into 9 equal smaller pieces, and continuing. In
[3] two of the authors of this paper gave a construction of a Brownian motion on
F
SC
. This is a diusion (that is, a continuous strong Markov process) which takes
its values in F
SC
, and which is non-degenerate and invariant under all the local
isometries of F
SC
.
Subsequently, Kusuoka and Zhou in [30] gave a dierent construction of a
diusion on F
SC
, which yielded a process that, as well as having the invariance
properties of the Brownian motion constructed in [3], was also scale invariant. The
proofs in [3, 30] also work for fractals that are formed in a similar manner to the
standard Sierpinski carpet: we call these generalized Sierpinski carpets (GSCs).
In [5] the results of [3] were extended to GSCs embedded in R
d
for d 3. While
[3, 5] and [30] both obtained their diusions as limits of approximating processes,
the type of approximation was dierent: [3, 5] used a sequence of time changed
reecting Brownian motions, while [30] used a sequence of Markov chains.
Figure 1: The standard Sierpinski carpet
These papers left open the question of uniqueness of this Brownian motion
in fact it was not even clear whether or not the processes obtained in [3, 5] or [30]
were the same. This uniqueness question can also be expressed in analytic terms:
one can dene a Laplacian on a GSC as the innitesimal generator of a Brownian
motion, and one wants to know if there is only one such Laplacian. The main
result of this paper is that, up to scalar multiples of the time parameter, there
exists only one such Brownian motion; hence, up to scalar multiples, the Laplacian
is uniquely dened.
GSCs are examples of spaces with anomalous diusion. For Brownian motion
on R
d
one has E[X
t
X
0
[ = ct
1/2
. Anomalous diusion in a space F occurs
2
when instead one has E[X
t
X
0
[ = o(t
1/2
), or (in regular enough situations),
E[X
t
X
0
[ t
2/dw
, where d
w
(called the walk dimension) satises d
w
> 2. This
phenomena was rst observed by mathematical physicists working in the transport
properties of disordered media, such as (critical) percolation clusters see [1, 37].
Since these sets are subsets of the lattice Z
d
, they are not true fractals, but their
large scale structure still exhibits fractal properties, and the simple random walk
is expected to have anomalous diusion.
For critical percolation clusters (or, more precisely for the incipient innite
cluster) on trees and Z
2
, Kesten [23] proved that anomalous diusion occurs.
After this work, little progress was made on critical percolation clusters until the
recent papers [7, 8, 27].
As random sets are hard to study, it was natural to begin the study of anoma-
lous diusion in the more tractable context of regular deterministic fractals. The
simplest of these is the Sierpinski gasket. The papers [1, 37] studied discrete
random walks on graph approximations to the Sierpinski gasket, and soon af-
ter [19, 29, 11] constructed Brownian motions on the limiting set. The special
structure of the Sierpinski gasket makes the uniqueness problem quite simple, and
uniqueness of this Brownian motion was proved in [11]. These early papers used
a probabilistic approach, rst constructing the Brownian motion X on the space,
and then, having dened the Laplacian L
X
as the innitesimal generator of the
semigroup of X, used the process X to study L
X
. Soon after Kigami [24] and
Fukushima-Shima [18] introduced more analytical approaches, and in particular
[18] gave a very simple construction of X and L
X
using the theory of Dirichlet
forms.
Figure 2: The Sierpinski gasket (left), and a typical nested fractal, the Lindstrm
snowake (right)
It was natural to ask whether these results were special to the Sierpinski gas-
ket. Lindstrm [31] and Kigami [25] introduced wider families of fractals (called
nested fractals, and p.c.f. self-similar sets respectively), and gave constructions of
diusions on these spaces. Nested fractals are, like the Sierpinski carpet, highly
symmetric, and the uniqueness problem can be formulated in a similar fashion
to that for GSCs. Uniqueness for nested fractals was not treated in [31], and for
3
some years remained a signicant challenge, before being solved by Sabot [41].
(See also [33, 36] for shorter proofs). For p.c.f. self-similar sets, while some suf-
cient conditions for uniqueness are given in [41, 21], the general problem is still
open.
The study of these various families of fractals (nested fractals, p.c.f self-similar
sets, and GSCs) revealed a number of common themes, and showed that analysis
on these spaces diers from that in standard Euclidean space in several ways, all
ultimately connected with the fact that d
w
> 2:
The energy measure and the Hausdor measure are mutually singular,
The domain of the Laplacian is not an algebra,
If d(x, y) is the shortest path metric, then d(x, ) is not in the domain of the
Dirichlet form.
See [2, 26, 43] for further information and references.
The uniqueness proofs in [21, 33, 36, 41] all used in an essential way the fact
that nested fractals and p.c.f. self-similar sets are nitely ramied that is, they
can be disconnected by removing a nite number of points. For these sets there is
a natural denition of a set V
n
of boundary points at level n for the Sierpinski
gasket V
n
is the set of vertices of triangles of side 2
n
. If one just looks at the
process X at the times when it passes through the points in V
n
, one sees a nite
state Markov chain X
(n)
, which is called the trace of X on V
n
. If m > n then
V
n
V
m
and the trace of X
(m)
on V
n
is also X
(n)
. Using this, and the fact that
the limiting processes are known to be scale invariant, the uniqueness problem for
X can be reduced to the uniqueness of the xed point of a non-linear map on a
space of nite matrices.
While the boundaries of the squares (or cubes) have an analogous role to the
sets V
n
in the geometrical construction of a GSC, attempts to follow the same
strategy of proof encounter numerous diculties and have not been successful.
We use a dierent idea in this paper, and rather than studying the restriction of
the process X to boundaries, our argument treats the Dirichlet form of the process
on the whole space. (This also suggests a new approach to uniqueness on nitely
ramied fractals, which will be explored elsewhere.)
Let F be a GSC and the usual Hausdor measure on F. Let E be the set of
non-zero local regular conservative Dirichlet forms (c, T) on L
2
(F, ) which are
invariant with respect to all the local symmetries of F. (See Denition 2.15 for
a precise denition.) We remark that elements of E are not required to be scale
invariant see Denition 2.17. Our rst result is that E is non-empty.
Proposition 1.1 The Dirichlet forms associated with the processes constructed
in [3, 5] and [30] are in E.
Our main result is the following theorem, which is proved in Section 5.
Theorem 1.2 Let F R
d
be a GSC. Then, up to scalar multiples, E consists of
at most one element. Further, this one element of E satises scale invariance.
4
An immediate corollary of Proposition 1.1 and Theorem 1.2 is the following.
Corollary 1.3 The Dirichlet forms constructed in [3, 5] and [30] are (up to a
constant) the same.
(b) The Dirichlet forms constructed in [3, 5] satisfy scale invariance.
A Feller process is one where the semigroup T
t
maps continuous functions that
vanish at innity to continuous functions that vanish at innity, and lim
t0
T
t
f(x) =
f(x) for each x F if f is continuous and vanishes at innity. Our main theorem
can be stated in terms of processes as follows.
Corollary 1.4 If X is a continuous non-degenerate symmetric strong Markov
process which is a Feller process, whose state space is F, and whose Dirichlet form
is invariant with respect to the local symmetries of F, then the law of X under P
x
is uniquely dened, up to scalar multiples of the time parameter, for each x F.
Remark 1.5 Osada [35] constructed diusion processes on GSCs which are dier-
ent from the ones considered here. While his processes are invariant with respect
to some of the local isometries of the GSC, they are not invariant with respect to
the full set of local isometries.
In Section 2 we give precise denitions, introduce the notation we use, and
prove some preliminary lemmas. In Section 3 we prove Proposition 1.1. In Section
4 we develop the properties of Dirichlet forms c E, and in Section 5 we prove
Theorem 1.2.
The idea of our proof is the following. The main work is showing that if /, B
are any two Dirichlet forms in E, then they are comparable. (This means that /
and B have the same domain T, and that there exists a constant c = c(/, B) > 0
such that c/(f, f) B(f, f) c
1
/(f, f) for f T.) We then let be the
largest positive real such that ( = / B 0. If ( were also in E, then (
would be comparable to B, and so there would exist > 0 such that ( B 0,
contradicting the denition of . In fact we cannot be sure that ( is closed, so
instead we consider (

= (1 +)/B, which is easily seen to be in E. We then

need uniform estimates in to obtain a contradiction.
To show /, B E are comparable requires heat kernel estimates for an arbitrary
element of E. Using symmetry arguments as in [5], we show that the estimates
for corner moves and slides and the coupling argument of [5, Section 3] can be
modied so as to apply to any element c E. It follows that the elliptic Harnack
inequality holds for any such c. Resistance arguments, as in [4, 34], combined
with results in [20] then lead to the desired heat kernel bounds. (Note that the
results of [20] that we use are also available in [10].)
A key point here is that the constants in the Harnack inequality, and con-
sequently also the heat kernel bounds, only depend on the GSC F, and not on
the particular element of E. This means that we need to be careful about the
dependencies of the constants.
The symmetry arguments are harder than in [5, Section 3]. In [5] the approxi-
mating processes were time changed reecting Brownian motions, and the proofs
5
used the convenient fact that a reecting Brownian motion in a Lipschitz domain
in R
d
does not hit sets of dimension d 2. Since we do not have such approxi-
mations for the processes corresponding to an arbitrary element c E, we have
to work with the diusion X associated with c, and this process might hit sets
of dimension d 2. (See [5, Section 9] for examples of GSCs in dimension 3 for
which the process X hits not just lines but also points.)
We use C
i
to denote nite positive constants which depend only on the GSC,
but which may change between each appearance. Other nite positive constants
will be written as c
i
.
2 Preliminaries
2.1 Some general properties of Dirichlet forms
We begin with a general result on local Dirichlet forms. For denitions of local
and other terms related to Dirichlet forms, see [17]. Let F be a compact metric
space and m a Radon (i.e. nite) measure on F. For any Dirichlet form (c, T) on
L
2
(F, m) we dene
c
1
(u, u) = c(u, u) +|u|
2
2
. (2.1)
Functions in T are only dened up to quasi-everywhere equivalence (see [17] p.
67); we use a quasi-continuous modication of elements of T throughout the paper.
We write , ) for the inner product in L
2
(F, m) and , )
S
for the inner product
in a subset S F.
Theorem 2.1 Suppose that (/, T), (B, T) are local regular conservative irre-
ducible Dirichlet forms on L
2
(F, m) and that
/(u, u) B(u, u) for all u T. (2.2)
Let > 0, and c = (1 + )B /. Then (c, T) is a regular local conservative
irreducible Dirichlet form on L
2
(F, m).
Proof. It is clear that c is a non-negative symmetric form, and is local.
To show that c is closed, let u
n
be a Cauchy sequence with respect to c
1
.
Since c
1
(f, f) ( 1)B
1
(f, f), u
n
is a Cauchy sequence with respect to B
1
.
Since B is a Dirichlet form and so closed, there exists u T such that B
1
(u
n

u, u
n
u) 0. As / B we have /(u
n
u, u
n
u) 0 also, and so c
1
(u
n

u, u
n
u) 0, proving that (c, T) is closed.
Since / and B are conservative and F is compact, 1 T and c(1, h) = 0 for
all h T, which shows that c is conservative by [17, Theorem 1.6.3 and Lemma
1.6.5].
We now show that c is Markov. By [17, Theorem 1.4.1] it is enough to prove
that c( u, u) c(u, u) for u T, where we let u = 0 (u 1). Since / is
local and u
+
u

= 0, we have /(u
+
, u

) = 0 ([42, Proposition 1.4]). Similarly

B(u
+
, u

) = 0, giving c(u
+
, u

) = 0. Using this, we have

c(u, u) = c(u
+
, u
+
) 2c(u
+
, u

) +c(u

, u

) c(u
+
, u
+
) (2.3)
6
for u T. Now let v = 1 u. Then u = (1 v
+
)
+
, so
c(u, u) = c(v, v) c(v
+
, v
+
) = c(1 v
+
, 1 v
+
)
c((1 v
+
)
+
, (1 v
+
)
+
) = c( u, u),
and hence c is Markov.
As B is regular, it has a core ( T. Let u T. As ( is a core for B, there exist
u
n
( such that B
1
(u u
n
, u u
n
) 0. Since / B, /
1
(u
n
u, u
n
u) 0
also, and so c
1
(u
n
u, u
n
u) 0. Thus ( is dense in T in the c
1
norm (and it
is dense in C(F) in the supremum norm since it is a core for B), so c is regular.
Let A F be invariant for the semigroup corresponding to c. By [17, Theorem
1.6.1], this is equivalent to the following: 1
A
u T for all u T and
c(u, v) = c(1
A
u, 1
A
v) +c(1
FA
u, 1
FA
v) u, v T. (2.4)
Once we have 1
A
u T, since (1
A
u)(1
FA
u) = 0 we have /(1
A
u, 1
FA
u) = 0,
and we obtain (2.4) for / also. Using [17, Theorem 1.6.1] again, we see that
A is invariant for the semigroup corresponding to /. Since / is irreducible, we
conclude that either m(A) = 0 or m(X A) = 0 holds and hence that (c, T) is
irreducible.
Remark 2.2 This should be compared with the situation for Dirichlet forms on
nite sets, which is the context of the uniqueness results in [33, 41]. In that case
the Dirichlet forms are not local, and given /, B satisfying (2.2) there may exist

0
> 0 such that (1 +)B / fails to be a Dirichlet form for (0,
0
).
For the remainder of this section we assume that (c, T) is a local regular
Dirichlet form on L
2
(F, m), that 1 T and c(1, 1) = 0. We write T
t
for the
semigroup associated with c, and X for the associated diusion.
Lemma 2.3 T
t
is recurrent and conservative.
Proof. T
t
is recurrent by [17, Theorem 1.6.3]. Hence by [17, Lemma 1.6.5] T
t
is
conservative.
Let D be a Borel subset of F. We write T
D
for the hitting time of D, and
D
for the exit time of D:
T
D
= T
X
D
= inft 0 : X
t
D,
D
=
X
D
= inft 0 : X
t
, D. (2.5)
Let T
t
be the semigroup of X killed on exiting D, and X be the killed process.
Set
q(x) = P
x
(
D
= ),
and
E
D
= x : q(x) = 0, Z
D
= x : q(x) = 1. (2.6)
Lemma 2.4 Let D be a Borel subset of F. Then m(D(E
D
Z
D
)) = 0. Further,
E
D
and Z
D
are invariant sets for the killed process X, and Z
D
is invariant for X.
7
Proof. If f 0,
T
t
(f1
ED
), 1
DED
q) = f1
ED
, T
t
(1
DED
q)) f1
ED
, T
t
q) = 0.
So T
t
(f1
ED
) = 0 on DE
D
and hence (see [17, Lemma 1.6.1(ii)]) E
D
is invariant
for X.
Let A = x : P
x
(
D
< ) > 0 = Z
c
D
. The set A is an invariant set of
the process X by [17, Lemma 4.6.4]. Using the fact that X = X, P
x
-a.s. for
x Z
D
and [17, Lemma 1.6.1(ii)], we see that A is an invariant set of the process
X as well. So we see that Z
D
is invariant both for X and X. In order to prove
m(D(E
D
Z
D
)) = 0, it suces to show that E
x
[
D
] < for a.e. x AD. Let
U
D
be the resolvent of the killed process X. Since A D is of nite measure, the
proof of Lemma 1.6.5 or Lemma 1.6.6 of [17] give U
D
1(x) < for a.e. x AD,
so we obtain E
x
[
D
] < .
Note that in the above proof we do not use the boundedness of D, but only
the fact that m(D) < .
Next, we give some general facts on harmonic and caloric functions. Let D be a
Borel subset in F and let h : F R. There are two possible denitions of h being
harmonic in D. The probabilistic one is that h is harmonic in D if h(X
t
D

) is
a uniformly integrable martingale under P
x
for q.e. x whenever D

is a relatively
open subset of D. The Dirichlet form denition is that h is harmonic with respect
to c in D if h T and c(h, u) = 0 whenever u T is continuous and the support
of u is contained in D.
The following is well known to experts. We will use it in the proofs of Lemma 4.9
and Lemma 4.24. (See [15] for the equivalence of the two notions of harmonicity
in a very general framework.) Recall that P
x
(
D
< ) = 1 for x E
D
.
Proposition 2.5 (a) Let (c, T) and D satisfy the above conditions, and let
h T be bounded. Then h is harmonic in a domain D in the probabilistic sense
if and only if it is harmonic in the Dirichlet form sense.
(b) If h is a bounded Borel measurable function in D and D

is a relatively open
subset of D, then h(X
t
D

) is a martingale under P
x
for q.e. x E
D
if and only
if h(x) = E
x
[h(X

)] for q.e. x E
D
.
Proof. (a) By [17, Theorem 5.2.2], we have the Fukushima decomposition
h(X
t
) h(X
0
) = M
[h]
t
+N
[h]
t
, where M
[h]
is a square integrable martingale addi-
tive functional of nite energy and N
[h]
is a continuous additive functional having
zero energy (see [17, Section 5.2]). We need to consider the Dirichlet form (c, T
D
)
where T
D
= f T : supp(f) D, and denote the corresponding semigroup as
P
D
t
.
If h is harmonic in the Dirichlet form sense, then by the discussion in [17,
p. 218] and [17, Theorem 5.4.1], we have P
x
(N
[h]
t
= 0, t <
D
) = 1 q.e. x F.
Thus, h is harmonic in the probabilistic sense. Here the notion of the spectrum
from [17, Sect. 2.3] and especially [17, Theorem 2.3.3] are used.
To show that being harmonic in the probabilistic sense implies being harmonic
in the Dirichlet form sense is the delicate part of this proposition. Since Z
D
is
8
P
D
t
-invariant (by Lemma 2.4) and h(X
t
) is a bounded martingale under P
x
for
x Z
D
, we have
P
D
t
(h1
ZD
)(x) = 1
ZD
(x)P
D
t
h(x) = 1
ZD
(x)E
x
[h(X
t
)] = h1
ZD
(x).
Thus by [17, Lemma 1.3.4], we have h1
ZD
T and c(h1
ZD
, v) = 0 for all v T.
Next, note that on Z
c
D
we have H
B
h = h, according to the denition of H
B
on
page 150 of [17] and Lemma 2.4, which implies H
B
_
h1
Z
c
D
_
= h1
Z
c
D
. Then from
[17, Theorem 4.6.5], applied with u = h1
Z
c
D
= h h1
ZD
T and B
c
= D, we
conclude that h1
Z
c
D
is harmonic in the Dirichlet form sense. Thus h = h1
Z
c
D
+h1
ZD
is harmonic in the Dirichlet form sense in D.
(b) If h(X
t
D

) is a martingale under P
x
for q.e. x E
D
, then E
x
[h(X
s
D

)] =
E
x
[h(X
t
D

)] for q.e. x E
D
and for all s, t 0, where we can take s 0
and t and interchange the limit and the expectation since h is bounded.
Conversely, if h(x) = E
x
[h(X

)] for q.e. x E
D
, then by the strong Markov
property, h(X
t
D

) = E
x
[h(X

)[T
t
D

] under P
x
for q.e. x E
D
, so h(X
t
D

)
is a martingale under P
x
for q.e. x E
D
.
We call a function u : R
+
F R caloric in D in the probabilistic sense if
u(t, x) = E
x
[f(X
tD
)] for some bounded Borel f : F R. It is natural to view
u(t, x) as the solution to the heat equation with boundary data dened by f(x)
outside of D and the initial data dened by f(x) inside of D. We call a function
u : R
+
F R caloric in D in the Dirichlet form sense if there is a function h
which is harmonic in D and a bounded Borel f
D
: F R which vanishes outside
of D such that u(t, x) = h(x)+T
t
f
D
. Note that T
t
is the semigroup of X killed on
exiting D, which can be either dened probabilistically as above or, equivalently,
in the Dirichlet form sense by Theorems 4.4.3 and A.2.10 in [17].
Proposition 2.6 Let (c, T) and D satisfy the above conditions, and let f T be
bounded and t 0. Then
E
x
[f(X
tD
)] = h(x) +T
t
f
D
q.e., where h(x) = E
x
[f(X
D
)] is the harmonic function that coincides with f on
D
c
, and f
D
(x) = f(x) h(x).
Proof. By Proposition 2.5, h is uniquely dened in the probabilistic and Dirichlet
form senses, and h(x) = E
x
[h(X
tD
)]. Note that f
D
(x) vanishes q.e. outside of
D. Then we have E
x
[f
D
(X
tD
)] = T
t
f
D
by Theorems 4.4.3 and A.2.10 in [17].
Note that the condition f T can be relaxed (see the proof of Lemma 4.9).
We show a general property of local Dirichlet forms which will be used in the
proof of Proposition 2.21. Note that it is not assumed that c admits a carre du
champ. Since c is regular, c(f, f) can be written in terms of a measure (f, f), the
energy measure of f, as follows. Let T
b
be the elements of T that are essentially
9
bounded. If f T
b
, then (f, f) is dened to be the unique smooth Borel measure
on F satisfying
_
F
gd(f, f) = 2c(f, fg) c(f
2
, g), g T
b
.
Lemma 2.7 If c is a local regular Dirichlet form with domain T, then for any
f T L

(F) we have (f, f)(A) = 0, where A = x F : f(x) = 0.

Proof. Let
f
be the measure on R which is the image of the measure (f, f) on
F under the function f : F R. By [13, Theorem 5.2.1, Theorem 5.2.3] and the
chain rule,
f
is absolutely continuous with respect to one-dimensional Lebesgue
measure on R. Hence (f, f)(A) =
f
(0) = 0.
Lemma 2.8 Given a m-symmetric Feller process on F, the corresponding Dirich-
let form (c, T) is regular.
Proof. First, we note the following: if H is dense in L
2
(F, m), then U
1
(H) is
dense in T, where U
1
is the 1-resolvent operator. This is because U
1
: L
2
T(L)
is an isometry where the norm of g T(L) is given by |g|
D(L)
:= |(I L)g|
2
,
and T(L) T is a continuous dense embedding (see, for example [17, Lemma
1.3.3(iii)]). Here L is the generator corresponding to c. Since C(F) is dense in L
2
and U
1
(C(F)) T C(F) as the process is Feller, we see that T C(F) is dense
in T in the c
1
-norm.
Next we need to show that u C(F) can be approximated with respect to the
supremum norm by functions in T C(F). This is easy, since T
t
u T for each
t, is continuous since we have a Feller process, and T
t
u u uniformly by [39,
Lemma III.6.7].
Remark 2.9 The proof above uses the fact that F is compact. However, it can be
easily generalized to a Feller process on a locally compact separable metric space
by a standard truncation argument for example by using [17, Lemma 1.4.2(i)].
2.2 Generalized Sierpinski carpets
Let d 2, F
0
= [0, 1]
d
, and let L
F
N, L
F
3, be xed. For n Z let Q
n
be the
collection of closed cubes of side L
n
F
with vertices in L
n
F
Z
d
. For A R
d
, set
Q
n
(A) = Q Q
n
: int(Q) A ,= .
For Q Q
n
, let
Q
be the orientation preserving ane map (i.e. similitude with
no rotation part) which maps F
0
onto Q. We now dene a decreasing sequence
(F
n
) of closed subsets of F
0
. Let 1 m
F
L
d
F
be an integer, and let F
1
be the
union of m
F
distinct elements of Q
1
(F
0
). We impose the following conditions on
F
1
.
(H1) (Symmetry) F
1
is preserved by all the isometries of the unit cube F
0
.
10
(H2) (Connectedness) Int(F
1
) is connected.
(H3) (Non-diagonality) Let m 1 and B F
0
be a cube of side length 2L
m
F
,
which is the union of 2
d
distinct elements of Q
m
. Then if int(F
1
B) is
non-empty, it is connected.
(H4) (Borders included) F
1
contains the line segment x : 0 x
1
1, x
2
= =
x
d
= 0.
We may think of F
1
as being derived from F
0
by removing the interiors of
L
d
F
m
F
cubes in Q
1
(F
0
). Given F
1
, F
2
is obtained by removing the same pattern
from each of the cubes in Q
1
(F
1
). Iterating, we obtain a sequence F
n
, where F
n
is the union of m
n
F
cubes in Q
n
(F
0
). Formally, we dene
F
n+1
=
_
QQn(Fn)

Q
(F
1
) =
_
QQ1(F1)

Q
(F
n
), n 1.
We call the set F =

n=0
F
n
a generalized Sierpinski carpet (GSC). The Hausdor
dimension of F is d
f
= d
f
(F) = log m
F
/ log L
F
. Later on we will also discuss the
unbounded GSC

F =

k=0
L
k
F
F, where rA = rx : x A.
Let

n
(dx) = (L
d
F
/m
F
)
n
1
Fn
(x)dx,
and let be the weak limit of the
n
; is a constant multiple of the Hausdor x
d
f
- measure on F. For x, y F we write d(x, y) for the length of the shortest path in
F connecting x and y. Using (H1)(H4) we have that d(x, y) is comparable with
the Euclidean distance [x y[.
Remark 2.10 1. There is an error in [5], where it was only assumed that (H3)
above holds when m = 1. However, that assumption is not strong enough to
imply the connectedness of the set J
k
in [5, Theorem 3.19]. To correct this error,
we replace the (H3) in [5] by the (H3) in the current paper.
2. The standard SC in dimension d is the GSC with L
F
= 3, m
F
= 3
d
1,
and with F
1
obtained from F
0
by removing the middle cube. We have allowed
m
F
= L
d
F
, so that our GSCs do include the trivial case F = [0, 1]
d
. The Menger
sponge (see the picture on [32], p. 145) is one example of a GSC, and has d = 3,
L
F
= 3, m
F
= 20.
Denition 2.11 Dene:
o
n
= o
n
(F) = Q F : Q Q
n
(F).
We will need to consider two dierent types of interior and boundary for subsets
of F which consist of unions of elements of o
n
. First, for any A F we write
int
F
(A) for the interior of A with respect to the metric space (F, d), and
F
(A) =
Aint
F
(A). Given any U R
d
we write U
o
for the interior of U in with respect
to the usual topology on R
d
, and U = U U
o
for the usual boundary of U. Let
A be a nite union of elements of o
n
, so that A =
k
i=1
S
i
, where S
i
= F Q
i
and
Q
i
Q
n
(F). Then we dene int
r
(A) = F ((
k
i=1
Q
i
)
o
), and
r
(A) = Aint
r
(A).
We have int
r
(A) = A(
k
i=1
Q
i
). (See Figure 3).
11
Figure 3: Illustration for Denition 2.11 in the case of the standard Sierpinski
carpet and n = 1. Let A be the shaded set. The thick dotted lines give int
F
A on
the left, and int
r
A on the right.
Denition 2.12 We dene the folding map
S
: F S for S o
n
(F) as follows.
Let
0
: [1, 1] R be dened by
0
(x) = [x[ for [x[ 1, and then extend the
domain of
0
to all of R by periodicity, so that
0
(x + 2n) =
0
(x) for all x R,
n Z. If y is the point of S closest to the origin, dene
S
(x) for x F to be the
point whose i
th
coordinate is y
i
+L
n
F

0
(L
n
F
(x
i
y
i
)).
It is straightforward to check the following
Lemma 2.13 (a)
S
is the identity on S and for each S

o
n
,
S
: S

S is
an isometry.
(b) If S
1
, S
2
o
n
then

S1

S2
=
S1
. (2.7)
(c) Let x, y F. If there exists S
1
o
n
such that
S1
(x) =
S1
(y), then
S
(x) =

S
(y) for every S o
n
.
(d) Let S o
n
and S

o
n+1
. If x, y F and
S
(x) =
S
(y) then
S
(x) =

S
(y).
Given S o
n
, f : S R and g : F R we dene the unfolding and restriction
operators by
U
S
f = f
S
, R
S
g = g[
S
.
Using (2.7), we have that if S
1
, S
2
o
n
then
U
S2
R
S2
U
S1
R
S1
= U
S1
R
S1
. (2.8)
Denition 2.14 We dene the length and mass scale factors of F to be L
F
and
m
F
respectively.
Let D
n
be the network of diagonal crosswires obtained by joining each vertex of
a cube Q Q
n
to the vertex at the center of the cube by a wire of unit resistance
see [4, 34]. Write R
D
n
for the resistance across two opposite faces of D
n
. Then it is
12
proved in [4, 34] that there exists
F
such that there exist constants C
i
, depending
only on the dimension d, such that
C
1

n
F
R
D
n
C
2

n
F
. (2.9)
We remark that
F
L
2
F
/m
F
see [5, Proposition 5.1].
2.3 F-invariant Dirichlet forms
Let (c, T) be a local regular Dirichlet form on L
2
(F, ). Let S o
n
. We set
c
S
(g, g) =
1
m
n
F
c(U
S
g, U
S
g). (2.10)
and dene the domain of c
S
to be T
S
= g : g maps S to R, U
S
g T. We write

S
= [
S
.
Denition 2.15 Let (c, T) be a Dirichlet form on L
2
(F, ). We say that c is
an F-invariant Dirichlet form or that c is invariant with respect to all the local
symmetries of F if the following items (1)(3) hold:
(1) If S o
n
(F), then U
S
R
S
f T (i.e. R
S
f T
S
) for any f T.
(2) Let n 0 and S
1
, S
2
be any two elements of o
n
, and let be any isometry
of R
d
which maps S
1
onto S
2
. (We allow S
1
= S
2
.) If f T
S2
, then
f T
S1
and
c
S1
(f , f ) = c
S2
(f, f). (2.11)
(3) For all f T
c(f, f) =

SSn(F)
c
S
(R
S
f, R
S
f). (2.12)
We write E for the set of F-invariant, non-zero, local, regular, conservative Dirich-
let forms.
Remark 2.16 We cannot exclude at this point the possibility that the energy
measure of c E may charge the boundaries of cubes in o
n
. See Remark 5.3.
We will not need the following denition of scale invariance until we come to
the proof of Corollary 1.3 in Section 5.
Denition 2.17 Recall that
Q
, Q Q
1
(F
1
) are the similitudes which dene
F
1
. Let (c, T) be a Dirichlet form on L
2
(F, ) and suppose that
f
Q
T for all Q Q
1
(F
1
), f T. (2.13)
Then we can dene the replication of c by
c(f, f) =

QQ1(F1)
c(f
Q
, f
Q
). (2.14)
We say that (c, T) is scale invariant if (2.13) holds, and there exists > 0 such
that c = c.
13
Remark 2.18 We do not have any direct proof that if c E then (2.13) holds.
Ultimately, however, this will follow from Theorem 1.2.
Lemma 2.19 Let (/, T
1
), (B, T
2
) E with T
1
= T
2
and / B. Then ( =
(1 +)/B E for any > 0.
Proof. It is easy to see that Denition 2.15 holds. This and Theorem 2.1 proves
the lemma.
Proposition 2.20 If c E and S o
n
(F), then (c
S
, T
S
) is a local regular
Dirichlet form on L
2
(S,
S
).
Proof. (Local): If u, v are in T
S
with compact support and v is constant in a
neighborhood of the support of u, then U
S
u, U
S
v will be in T, and by the local
property of c, we have c(U
S
u, U
S
v) = 0. Then by (2.10) we have c
S
(u, v) = 0.
(Markov): Given that c
S
is local, we have the Markov property by the same proof
as that in Theorem 2.1.
(Conservative): Since 1 T, c
S
(1, 1) = 0 by (2.10).
(Regular): If h T then by (2.12) c
S
(R
S
h, R
S
h) c(h, h). Let f T
S
, so that
U
S
f T. As c is regular, given > 0 there exists a continuous g T such that
c
1
(U
S
f g, U
S
f g) < . Then R
S
U
S
f R
S
g = f R
S
g on S, so
c
S
1
(f R
S
g, f R
S
g) = c
S
1
(R
S
U
S
f R
S
g, R
S
U
S
f R
S
g)
c
1
(U
S
f g, U
S
f g) < .
As R
S
g is continuous, we see that T
S
C(S) is dense in T
S
in the c
S
1
norm. One
can similarly prove that T
S
C(S) is dense in C(S) in the supremum norm, so
the regularity of c
S
is proved.
(Closed): If f
m
is Cauchy with respect to c
S
1
, then U
S
f
m
will be Cauchy with
respect to c
1
. Hence U
S
f
m
converges with respect to c
1
, and it follows that
R
S
(U
S
f
m
) = f
m
converges with respect to c
S
1
.
Fix n and dene for functions f on F
f =
1
m
n
F

SSn(F)
U
S
R
S
f. (2.15)
Using (2.8) we have
2
= , and so is a projection operator. It is bounded
on C(F) and L
2
(F, ), and moreover by [40, Theorem 12.14] is an orthogonal
projection on L
2
(F, ). Denition 2.15(1) implies that : T T.
Proposition 2.21 Assume that c is a local regular Dirichlet form on F, T
t
is its
semigroup, and U
S
R
S
f T whenever S o
n
(F) and f T. Then the following
are equivalent:
14
(a) For all f T, we have c(f, f) =

SSn(F)
c
S
(R
S
f, R
S
f);
(b) for all f, g T
c(f, g) = c(f, g); (2.16)
(c) T
t
f = T
t
f a.e for any f L
2
(F, ) and t 0.
Remark 2.22 Note that this proposition and the following corollary do not use all
the symmetries that are assumed in Denition 2.15(2). Although these symmetries
are not needed here, they will be essential later in the paper.
Proof. To prove that (a) (b), note that (a) implies that
c(f, g) =

TSn(F)
c
T
(R
T
f, R
T
g) =
1
m
n
F

TSn(F)
c(U
T
R
T
f, U
T
R
T
g). (2.17)
Then using (2.15), (2.17) and (2.8),
c(f, g) =
1
m
n
F

SSn(F)
c(U
S
R
S
f, g)
=
1
m
2n
F

SSn(F)

TSn(F)
c(U
T
R
T
U
S
R
S
f, U
T
R
T
g)
=
1
m
2n
F

SSn(F)

TSn(F)
c(U
S
R
S
f, U
T
R
T
g).
Essentially the same calculation shows that c(f, g) is equal to the last line of the
above with the summations reversed.
Next we showthat (b) (c). If L is the generator corresponding to c, f T(L)
and g T then, writing f, g) for
_
F
fg d, we have
Lf, g) = Lf, g) = c(f, g) = c(f, g)
by (2.16) and the fact that is self-adjoint in the L
2
sense. By the denition of
the generator corresponding to a Dirichlet form, this is equivalent to
f T(L) and Lf = Lf.
By [40, Theorem 13.33], this implies that any bounded Borel function of L com-
mutes with . (Another good source on the spectral theory of unbounded self-
adjoint operators is [38, Section VIII.5].) In particular, the L
2
-semigroup T
t
of L
commutes with in the L
2
-sense. This implies (c).
In order to see that (c) (b), note that if f, g T,
c(f, g) = lim
t0
t
1
(I T
t
)f, g) = limt
1
(I T
t
)f, g)
= limt
1
(I T
t
)f, g) = limt
1
f, (I T
t
)g)
= c(f, g).
15
It remains to prove that (b) (a). This is the only implication that uses the
assumption that c is local. It suces to assume f and g are bounded.
First, note the obvious relation

SSn(F)
1
S
(x)
N
n
(x)
= 1 (2.18)
for any x F, where
N
n
(x) =

SSn(F)
1
S
(x) (2.19)
is the number of cubes o
n
whose interiors intersect F and which contain the point
x. We break the remainder of the proof into a number of steps.
Step 1: We show that if f = f, then (hf) = f(h). To show this, we
start with the relationship U
T
R
T
U
S
R
S
f = U
S
R
S
f. Summing over S o
n
(F)
and dividing by m
n
F
yields
U
T
R
T
f = U
T
R
T
(f) = f = f.
Since R
S
(f
1
f
2
) = R
S
(f
1
)R
S
(f
2
) and U
S
(g
1
g
2
) = U
S
(g
1
)U
S
(g
2
), we have
(hf) =
1
m
n
F

SSn
(U
S
R
S
f)(U
S
R
S
h) =
1
m
n
F

SSn
f(U
S
R
S
h) = f(h).
In particular, (f
2
) = ff = f
2
.
Step 2: We compute the adjoints of R
S
and U
S
. R
S
maps C(F), the contin-
uous functions on F, to C(S), the continuous functions on S. So R

S
maps nite
measures on S to nite measures on F. We have
_
f d(R

S
) =
_
R
S
f d =
_
1
S
(x)f(x) (dx),
and hence
R

S
(dx) = 1
S
(x) (dx). (2.20)
U
S
maps C(S) to C(F), so U

S
maps nite measures on F to nite measures
on S. If is a nite measure on F, then using (2.18)
_
S
f d(U

S
) =
_
F
U
S
f d =
_
F
f
S
(x) (dx) (2.21)
=
_
F
_

TSn
1
T
(x)
N
n
(x)
_
f
S
(x) (dx)
=

T
_
T
f
S
(x)
N
n
(x)
(dx).
Let
T,S
: T S be dened to be the restriction of
S
to T; this is one-to-one
and onto. If is a measure on T, dene its pull-back

T,S
to be the measure on
S given by
_
S
f d(

T,S
) =
_
T
(f
T,S
) d.
16
Write

T
(dx) =
1
T
(x)
N
n
(x)
(dx).
Then (2.21) translates to
_
S
f d(U

S
) =

T
_
T
f

T,S
(
T
)(dx),
and thus
U

S
=

TSn

T,S
(
T
). (2.22)
Step 3: We prove that if is a nite measure on F such that

= and
S o
n
, then
(F) = m
n
F
_
S
1
N
n
(x)
(dx). (2.23)
To see this, recall that

T,V
(
T
) is a measure on V , and then by (2.20) and (2.22)

=
1
m
n
F

V Sn
R

V
U

=
1
m
n
F

V Sn

TSn
_
1
V
(x)

T,V
(
T
)(dx)
=
1
m
n
F

T
_

T,V
(
T
)(dx).
On the other hand, using (2.18)
(dx) =

V
1
V
(x)
N
n
(x)
(dx) =

V
(dx).
Note that
V
and m
n
F

T,V
(
T
) are both supported on V , and the only way

can equal is if

V
= m
n
F

TSn

T,V
(
T
) (2.24)
for each V . Therefore
_
S
1
N
n
(x)
(dx) =
S
(F) = m
n
F

T
_
1
F
(x)

T,S
(
T
)(dx)
= m
n
F

T
_
1
F

T,S
(x)
T
(dx) = m
n
F

T
_

T
(dx)
= m
n
F

T
_
1
T
(x)
N
n
(x)
(dx) = m
n
F
_
(dx) = m
n
F
(F).
Multiplying both sides by m
n
F
gives (2.23).
17
Step 4: We show that if f = f, then

((f, f)) = (f, f). (2.25)

Using Step 1, we have for h C(F) T
_
F
h

((f, f))(dx) =
_
F
h(x) (f, f)(dx) = 2c(f, fh) c(f
2
, h)
= 2c(f, (fh)) c(f
2
, h) = 2c(f, fh) c(f
2
, h)
= 2c(f, fh) c(f
2
, h) =
_
F
h(f, f)(dx).
This is the step where we used (b).
Step 5: We now prove (a). Note that if g T L

(F) and A = x F :
g(x) = 0, then (g, g)(A) = 0 by Lemma 2.7. By applying this to the function
g = f U
S
R
S
f, which vanishes on S, and using the inequality

(f, f)(B)
1/2
(U
S
R
S
f, U
S
R
S
f)(B)
1/2

(g, g)(B)
1/2
(g, g)(S)
1/2
= 0, B S,
(see page 111 in [17]), we see that
1
S
(x)(f, f)(dx) = 1
S
(x)(U
S
R
S
f, U
S
R
S
f)(dx) (2.26)
for any f T and S o
n
(F).
Starting from U
T
R
T
U
S
R
S
f = U
S
R
S
f, summing over T o
n
and dividing by
m
n
F
shows that (U
S
R
S
f) = U
S
R
S
f. Applying Step 4 with f replaced by U
S
R
S
f,

((U
S
R
S
f, U
S
R
S
f))(dx) = (U
S
R
S
f, U
S
R
S
f)(dx).
Applying Step 3 with = (U
S
R
S
f, U
S
R
S
f), we see
c(U
S
R
S
f, U
S
R
S
f) = (U
S
R
S
f, U
S
R
S
f)(F)
= m
n
F
_
S
1
N
n
(x)
(U
S
R
S
f, U
S
R
S
f)(dx).
Dividing both sides by m
n
F
, using the denition of c
S
, and (2.26),
c
S
(R
S
f, R
S
f) =
_
S
1
N
n
(x)
(f, f)(dx). (2.27)
Summing over S o
n
and using (2.18) we obtain

S
c
S
(R
S
f, R
S
f) =
_
(f, f)(dx) = c(f, f),
which is (a).
18
Corollary 2.23 If c E, f T, S o
n
(F), and
S
(R
S
f, R
S
f) is the energy
measure of c
S
, then

S
(R
S
f, R
S
f)(dx) =
1
N
n
(x)
(f, f)(dx), x S.
We nish this section with properties of sets of capacity zero for F-invariant
Dirichlet forms. Let A F and S o
n
. We dene
(A) =
1
S
(
S
(A)). (2.28)
Thus (A) is the union of all the sets that can be obtained from A by local
reections. We can check that (A) does not depend on S, and that
(A) = x : (1
A
)(x) > 0.
Lemma 2.24 If c E then
Cap(A) Cap((A)) m
2n
F
Cap(A)
for all Borel sets A F.
Proof. The rst inequality holds because we always have A (A). To prove
the second inequality it is enough to assume that A is open since the denition of
the capacity uses an inmum over open covers of A, and transforms an open
cover of A into an open cover of (A). If u T and u 1 on A, then m
n
F
u 1
on (A). This implies the second inequality because c(u, u) c(u, u), using
that is an orthogonal projection with respect to c, that is, c(f, g) = c(f, g).
Corollary 2.25 If c E, then Cap(A) = 0 if and only if Cap((A)) = 0.
Moreover, if f is quasi-continuous, then f is quasi-continuous.
Proof. The rst fact follows from Lemma 2.24. Then the second fact holds
because preserves continuity of functions on -invariant sets.
3 The Barlow-Bass and Kusuoka-Zhou
Dirichlet forms
In this section we prove that the Dirichlet forms associated with the diusions
on F constructed in [3, 5, 30] are F-invariant; in particular this shows that E is
non-empty and proves Proposition 1.1. A reader who is only interested in the
uniqueness statement in Theorem 1.2 can skip this section.
19
3.1 The Barlow-Bass processes
The constructions in [3, 5] were probabilistic and almost no mention was made of
Dirichlet forms. Further, in [5] the diusion was constructed on the unbounded
fractal

F. So before we can assert that the Dirichlet forms are F-invariant, we
need to discuss the corresponding forms on F. Recall the way the processes in
[3, 5] were constructed was to let W
n
t
be normally reecting Brownian motion on
F
n
, and to let X
n
t
= W
n
ant
for a suitable sequence (a
n
). This sequence satised
c
1
(m
F

F
/L
2
F
)
n
a
n
c
2
(m
F

F
/L
2
F
)
n
, (3.1)
where
F
is the resistance scale factor for F. It was then shown that the laws of
the X
n
were tight and that resolvent tightness held. Let U

n
be the -resolvent
operator for X
n
on F
n
. The two types of tightness were used to show there exist
subsequences n
j
such that U

nj
f converges uniformly on F if f is continuous on F
0
and that the P
x
law of X
nj
converges weakly for each x. Any such a subsequential
limit point was then called a Brownian motion on the GSC. The Dirichlet form
for W
n
is
_
Fn
[f[
2
d
n
and that for X
n
is
c
n
(f, f) = a
n
_
Fn
[f(x)[
2

n
(dx),
both on L
2
(F,
n
).
Fix any subsequence n
j
such that the laws of the X
nj
s converge, and the
resolvents converge. If X is the limit process and T
t
the semigroup for X, dene
c
BB
(f, f) = sup
t>0
1
t
f T
t
f, f)
with the domain T
BB
being those f L
2
(F, ) for which the supremum is nite.
We will need the fact that if U

n
is the -resolvent operator for X
n
and f is
bounded on F
0
, then U

n
f is equicontinuous on F. This is already known for the
Brownian motion constructed in [5] on the unbounded fractal

F, but now we need
it for the process on F with reection on the boundaries of F
0
. However the proof
is very similar to proofs in [3, 5], so we will be brief. Fix x
0
and suppose x, y are
in B(x
0
, r) F
n
. Then
U

n
f(x) = E
x
_

0
e
t
f(X
n
t
) dt
= E
x
_
S
n
r
0
e
t
f(X
n
t
) dt +E
x
(e
S
n
r
1)U

n
f(X
n
S
n
r
) +E
x
U

n
f(X
n
S
n
r
),
(3.2)
where S
n
r
is the time of rst exit from B(x
0
, r) F
n
. The rst term in (3.2) is
bounded by |f|

E
x
S
n
r
. The second term in (3.2) is bounded by
|U

n
f|

E
x
S
n
r
|f|

E
x
S
n
r
.
20
We have the same estimates in the case when x is replaced by y, so
[U

n
f(x) U

n
f(y)[ [E
x
U

n
f(X
n
S
n
r
) E
y
U

n
f(X
n
S
n
r
)[ +
n
(r),
where
n
(r) 0 as r 0 uniformly in n by [5, Proposition 5.5]. But
z E
z
U

n
f(X
n
S
n
r
) is harmonic in the ball of radius r/2 about x
0
. Using the
uniform elliptic Harnack inequality for X
n
t
and the corresponding uniform modu-
lus of continuity for harmonic functions ([5, Section 4]), taking r = [xy[
1/2
, and
using the estimate for
n
(r) gives the equicontinuity.
It is easy to derive from this that the limiting resolvent U

satises the property

that U

f is continuous on f whenever f is bounded.

Theorem 3.1 Each c
BB
is in E.
Proof. We suppose a suitable subsequence n
j
is xed, and we write c for the
corresponding Dirichlet form c
BB
. First of all, each X
n
is clearly conservative,
so T
n
t
1 = 1. Since we have T
nj
t
f T
t
f uniformly for each f continuous, then
T
t
1 = 1. This shows X is conservative, and c(1, 1) = sup
t
1 T
t
1, 1) = 0.
The regularity of c follows from Lemma 2.8 and the fact that the processes con-
structed in [5] are -symmetric Feller (see the above discussion, [5, Theorem 5.7]
and [3, Section 6]). Since the process is a diusion, the locality of c follows from
[17, Theorem 4.5.1].
The construction in [3, 5] gives a nondegenerate process, so c is non-zero. Fix
and let S o

(F). It is easy to see from the above discussion that U

S
R
S
f T for
any f T. Before establishing the remaining properties of F-invariance, we show
that

and T
t
commute, where

is dened in (2.15), but with o

n
(F) replaced
by o

(F). Let f, g)
n
denote
_
Fn
f(x)g(x)
n
(dx). The innitesimal generator for
X
n
is a constant times the Laplacian, and it is clear that this commutes with

.
Hence U

n
commutes with

, or

n
f, g)
n
= U

f, g)
n
. (3.3)
Suppose f and g are continuous and f is nonnegative. The left hand side is
U

n
f,

g)
n
, and if n converges to innity along the subsequence n
j
, this converges
to
U

f,

g) =

f, g).
The right hand side of (3.3) converges to U

f, g) since

f is continuous if f
is. Since X
t
has continuous paths, t T
t
f is continuous, and so by the uniqueness
of the Laplace transform,

T
t
f, g) = T
t

f, g). Linearity and a limit argument

allows us to extend this equality to all f L
2
(F). The implication (c) (a) in
Proposition 2.21 implies that c E.
3.2 The Kusuoka-Zhou Dirichlet form
Write c
KZ
for the Dirichlet form constructed in [30]. Note that this form is self-
similar.
21
Theorem 3.2 c
KZ
E.
Proof. One can see that c
KZ
satises Denition 2.15 because of the self-similarity.
The argument goes as follows. Initially we consider n = 1, and suppose f T =
T(c
KZ
). Then [30, Theorem 5.4] implies U
S
R
S
f T for any S o
1
(F). This
gives us Denition 2.15(1).
Let S o
1
(F) and S =
i
(F) where
i
is one of the contractions that dene
the self-similar structure on F, as in [30]. Then we have
f
i
= (U
S
R
S
f)
i
= (U
S
R
S
f)
j
for any i, j. Hence by [30, Theorem 6.9], we have
c
KZ
(U
S
R
S
f, U
S
R
S
f) =
F
m
1
F

j
c
KZ
((U
S
R
S
f)
j
, (U
S
R
S
f)
j
)
=
F
c
KZ
(f
i
, f
i
).
By [30, Theorem 6.9] this gives Denition 2.15(3), and moreover
c
S
(f, f) =
F
m
1
F
c
KZ
(f
i
, f
i
).
Denition 2.15(2) and the rest of the conditions for c
KZ
to be in E follow from
(1), (3) and the results of [30]. The case n > 1 can be dealt with by using the
self-similarity.
Proof of Proposition 1.1 This is immediate from Theorems 3.1 and 3.2.
4 Diusions associated with F-invariant Dirich-
let forms
In this section we extensively use notation and denitions introduced in Section 2,
especially Subsections 2.2 and 2.3. We x a Dirichlet form c E. Let X = X
(E)
be the associated diusion, T
t
= T
(E)
t
be the semigroup of X and P
x
= P
x,(E)
,
x F ^
0
, the associated probability laws. Here ^
0
is a properly exceptional set
for X. Ultimately (see Corollary 1.4) we will be able to dene P
x
for all x F, so
that ^
0
= .
4.1 Reected processes and the Markov property
Theorem 4.1 Let S o
n
(F) and Z =
S
(X). Then Z is a
S
-symmetric
Markov process with Dirichlet form (c
S
, T
S
), and semigroup T
Z
t
f = R
S
T
t
U
S
f.
Write

P
y
for the laws of Z; these are dened for y S ^
Z
2
, where ^
Z
2
is a
properly exceptional set for Z. There exists a properly exceptional set ^
2
for X
such that for any Borel set A F,

P
S(x)
(Z
t
A) = P
x
(X
t

1
S
(A)), x F ^
2
. (4.1)
22
Proof. Denote =
S
. We begin by proving that there exists a properly excep-
tional set ^
2
for X such that
P
x
(X
t

1
(A)) = T
t
1

1
(A)
(x) = T
t
1

1
(A)
(y) = P
y
(X
t

1
(A)) (4.2)
whenever A S is Borel, (x) = (y), and x, y F ^
2
. It is sucient to
prove (4.2) for a countable base (A
m
) of the Borel -eld on F. Let f
m
= 1
Am
.
Since T
t
1

1
(Am)
= T
t
U
S
f
m
, it is enough to prove that there exists a properly
exceptional set ^
2
such that for m N,
T
t
U
S
f
m
(x) = T
t
U
S
f
m
(y), if x, y F ^
2
and (x) = (y). (4.3)
By (2.8), (U
S
f) = U
S
f. Using Proposition 2.21,
T
t
U
S
f = T
t
U
S
f
m
= T
t
U
S
f,
for f L
2
, where the equality holds in the L
2
sense.
Recall that we always consider quasi-continuous modications of functions in
T. By Corollary 2.25, T
t
U
S
f
m
is quasi-continuous. Since [17, Lemma 2.1.4]
tells us that if two quasi-continuous functions coincide -a.e., then they coincide
q.e., we have that (T
t
U
S
f
m
) = T
t
U
S
f
m
q.e. The denition of implies that
(T
t
U
S
f
m
)(x) = (T
t
U
S
f
m
)(y) whenever (x) = (y), so there exists a properly
exceptional set ^
2,m
such that (4.3) holds. Taking ^
2
=
m
^
2,m
gives (4.2).
Using Theorem 10.13 of [16], Z is Markov and has semigroup T
Z
t
f = R
S
T
t
(U
S
f).
We take ^
Z
2
= (^
2
).
Using (4.3), U
S
R
S
T
t
U
S
f = T
t
U
S
f, and then
T
Z
t
f, g)
S
= R
S
T
t
U
S
f, g)
S
= m
n
F
U
S
R
S
T
t
U
S
f, U
S
g) = m
n
F
T
t
U
S
f, U
S
g).
This equals m
n
F
U
S
f, T
t
U
S
g), and reversing the above calculation, we deduce that
f, T
Z
t
g) = m
n
F
U
S
f, T
t
U
S
g), proving that Z is
S
-symmetric.
To identify the Dirichlet form of Z we note that
t
1
T
Z
t
f f, f)
S
= m
n
F
t
1
T
t
U
S
f U
S
f, U
S
f).
Taking the limit as t 0, and using [17, Lemma 1.3.4], it follows that Z has
Dirichlet form
c
Z
(f, f) = m
n
F
c(U
S
f, U
S
f) = c
S
(f, f).
Lemma 4.2 Let S, S

o
n
, Z =
S
(X), and be an isometry of S onto S

.
Then if x S ^,
P
x
((Z) ) = P
(x)
(Z ).
Proof. By Theorem 4.1 and Denition 2.15(2) Z and (Z) have the same Dirich-
let form. The result is then immediate from [17, Theorem 4.2.7], which states that
two Hunt processes are equivalent if they have the same Dirichlet forms, provided
we exclude an F-invariant set of capacity zero.
23
We say S, S

o
n
(F) are adjacent if S = QF, S

= Q

F for Q, Q

Q
n
(F),
and QQ

is a (d 1)-dimensional set. In this situation, let H be the hyperplane

separating S, S

. For any hyperplane H R

d
, let g
H
: R
d
R
d
be reection in
H. Recall the denition of
r
D, where D is a nite union of elements of o
n
.
Lemma 4.3 Let S
1
, S
2
o
n
(F) be adjacent, let D = S
1
S
2
, let B =
r
(S
1
S
2
),
and let H be the hyperplane separating S
1
and S
2
. Then there exists a properly
exceptional set ^ such that if x H D^, the processes (X
t
, 0 t T
B
) and
(g
H
(X
t
), 0 t T
B
) have the same law under P
x
.
Proof. Let f T with support in the interior of D. Then Denition 2.15(3)
and Proposition 2.20 imply that c(f, f) = c
S1
(R
S1
f, R
S1
f) + c
S2
(R
S2
f, R
S2
f).
Denition 2.15(2) implies that c(f, f) = c(f g
H
, f g
H
). Hence (g
H
(X
t
), 0
t T
B
) has the same Dirichlet form as (X
t
, 0 t T
B
), and so they have the
same law by [17, Theorem 4.2.7] if we exclude an F-invariant set of capacity zero.
4.2 Moves by Z and X
At this point we have proved that the Markov process X associated with the
Dirichlet form c E has strong symmetry properties. We now use these to obtain
various global properties of X. The key idea, as in [5], is to prove that certain
moves of the process in F have probabilities which can be bounded below by
constants depending only on the dimension d.
We need a considerable amount of extra technical notation, based on that in
[5], which will only be used in this subsection.
We begin by looking at the process Z =
S
(X) for some S o
n
, where n 0.
Since our initial arguments are scale invariant, we can simplify our notation by
taking n = 0 and S = F in the next denition.
Denition 4.4 Let 1 i, j d, with i ,= j, and set
H
i
(t) = x = (x
1
, . . . , x
d
) : x
i
= t, t R;
L
i
= H
i
(0) [0, 1/2]
d
;
M
ij
= x [0, 1]
d
: x
i
= 0,
1
2
x
j
1, and 0 x
k

1
2
for k ,= j.
Let

e
S = S (
d
i=1
H
i
(1)), D = S
e
S.
We now dene, for the process Z, the sets E
D
and Z
D
as in (2.6). The next
proposition says that the corners and slides of [5] hold for Z, provided that Z
0

E
D
.
Proposition 4.5 There exists a constant q
0
, depending only on the dimension d,
such that

P
x
(T
Z
Lj
<
Z
D
) q
0
, x L
i
E
D
, (4.4)

P
x
(T
Z
Mij
<
Z
D
) q
0
, x L
i
E
D
. (4.5)
24
A
0
A
1
A

1
s
v

Figure 4: Illustration for Denition 4.6 in the case of the standard Sierpinski carpet
and n = 1. The complement of D is shaded. The half-face A
1
corresponds to a
slide move, and the half-face A

1
corresponds to a corner move. In this case Q

is
the lower left cube in o
1
.
These inequalities hold for any n 0 provided we modify Denition 4.4 appropri-
ately.
Proof. Using Lemma 4.2 this follows by the same reection arguments as those
used in the proofs of Proposition 3.5 Lemma 3.10 of [5]. We remark that,
inspecting these proofs, we can take q
0
= 2
2d
2
.
We now x n 0. We call a set A R
d
a (level n) half-face if there exists
i 1, . . . , d, a = (a
1
, . . . , a
d
)
1
2
Z
d
with a
i
Z such that
A = x : x
i
= a
i
L
n
F
, a
j
L
n
F
x
j
(a
j
+ 1/2)L
n
F
for j ,= i.
(Note that a level n half-face need not be a subset of F.) For A as above set
(A) = i. Let /
(n)
be the collection of level n half-faces, and
/
(n)
F
= A /
(n)
: A F
n
.
We dene a graph structure on /
(n)
F
by taking A, B to be an edge if
dim(A B) = d 2, and A B Q for some Q Q
n
.
Let E(/
(n)
F
) be the set of edges in /
(n)
F
. As in [5, Lemma 3.12] we have that the
graph /
(n)
F
is connected. We call an edge A, B an i j corner if (A) = i,
25
(B) = j, and i ,= j and call A, B an i j slide if (A) = (B) = i, and the
line joining the centers of A and B is parallel to the x
j
axis. Any edge is either a
corner or a slide; note that the move (L
i
, L
j
) is an i j corner, while (L
i
, M
ij
) is
an i j slide.
For the next few results we need some further notation.
Denition 4.6 Let (A
0
, A
1
) be an edge in E(/
(n)
F
), and Q

be a cube in Q
n
(F)
such that A
0
A
1
Q

. Let v

be the unique vertex of Q

such that v

A
0
, and
let R be the union of the 2
d
cubes in Q
n
containing v

. Then there exist distinct

S
i
o
n
, 1 i m such that F R =
m
i=1
S
i
. Let D = F R
o
; thus
D = F R =
m
i=1
S
i
.
Let S

be any one of the S

i
, and set Z =
S
(X). Write
=
X
D
= inft 0 : X
t
, D = inft : Z
t

r
R. (4.6)
Let
E
D
= x D : P
x
( < ) = 1. (4.7)
We wish to obtain a lower bound for
inf
xA0ED
P
x
(T
X
A1
). (4.8)
By Proposition 4.5 we have
inf
yA0ED

P
y
(T
Z
A1
) q
0
. (4.9)
Z hits A
1
if and only if X hits (A
1
), and one wishes to use symmetry to prove
that, if x A
0
E
D
then for some q
1
> 0
P
x
(T
X
A1
) q
1

P
x
(T
Z
A1
) q
1
q
0
. (4.10)
This was proved in [5] in the context of reecting Brownian motion on F
n+k
, but
the proof used the fact that sets of dimension d 2 were polar for this process.
Here we need to handle the possibility that there may be times t such that X
t
is
in more than two of the S
i
. We therefore need to consider the way that X leaves
points y which are in several S
i
.
Denition 4.7 Let y E
D
be in exactly k of the S
i
, where 1 k m. Let
S

1
, . . . , S

k
be the elements of o
n
containing y. (We do not necessarily have that
S
1
is one of the S

j
.) Let D(y) = int
r
(
k
i=1
S

i
); so that D(y) =
k
i=1
S

i
. Let D
1
,
D
2
be open sets in F such that y D
2
D
2
D
1
D
1
D(y). Assume further
that (D
i
) D(y) = D
i
for i = 1, 2, and note that we always have (D
i
) D
i
.
For f T dene

D1
f = k
1
m
n
F
1
D1
f; (4.11)
the normalization factor is chosen so that
D1
1
D1
= 1
D1
.
26
sy s
v

D(y)
D1
D2
Figure 5: Illustration for Denition 4.7 in the case of the standard Sierpinski
carpet and n = 1. The complement of D is shaded, and the dotted lines outline
D(y) D
1
D
2
.
As before we dene T
D1
T as the closure of the set of functions f
T : supp(f) D
1
. We denote by c
D1
the associated Dirichlet form and by
T
D1
t
the associated semigroup, which are the Dirichlet form and the semigroup
of the process X killed on exiting D
1
, by Theorems 4.4.3 and A.2.10 in [17]. For
convenience, we state the next lemma in the situation of Denition 4.7, although
it holds under somewhat more general conditions.
Lemma 4.8 Let D
1
, D
2
be as in Denition 4.7.
(a) Let f T
D1
. Then
D1
f T
D1
. Moreover, for all f, g T
D1
we have
c
D1
(
D1
f, g) = c
D1
(f,
D1
g)
and T
D1
t

D1
f =
D1
T
D1
t
f.
(b) If h T
D1
is harmonic (in the Dirichlet form sense) in D
2
then
D1
h is
harmonic (in the Dirichlet form sense) in D
2
.
(c) If u is caloric in D
2
, in the sense of Proposition 2.6, then
D1
u is also caloric
in D
2
.
Proof. (a) By Denition 2.15, f T. Let be a function in T which has
support in D(y) and is 1 on D
1
; such a function exists because c is regular and
Markov. Then f T, and f = km
n
F

D1
f. The rest of the proof follows
from Proposition 2.21(b,c) because c(
D1
f, g) = k
1
m
n
F
c(f, g).
27
(b) Let g T with supp(g) D
2
. Then
c(
D1
h, g) = k
1
m
n
F
c(h, g) = k
1
m
n
F
c(h, g) = c(h,
D1
g) = 0. (4.12)
The nal equality holds because h is harmonic on D
2
and
D1
g has support in
D
2
. Relation (4.12) implies that
D1
h is harmonic in D
2
by Proposition 2.5.
(c) We denote by T
t
the semigroup of the process X
t
, which is X
t
killed at exiting
D
2
. The same reasoning as in (a) implies that T
t

D1
=
D1
T
t
. Hence (c) follows
from (a), (b) and Proposition 2.6.
Recall from (2.19) the denition of the cube counting function N
n
(z). Dene
the related weight function
r
S
(z) = 1
S
(z)N
n
(z)
1
for each S o
n
(F). If no confusion can arise, we will denote r
i
(z) = r
S

i
(z).
Let (T
Z
t
) be the ltration generated by Z. Since T
Z
0
contains all P
x
null sets,
under the law P
x
we have that X
0
= x is T
Z
0
measurable.
Lemma 4.9 Let y E
D
, D
1
, D
2
be as in Denition 4.7. Write V =
X
D2
.
(a) If U
F
(D
2
) satises (U) D(y) = U, then
E
y
(r
i
(X
V
)1
(XV U)
) = k
1

P
S(y)
(Z
V

S
(U)), for i = 1, . . . , k = N
n
(y).
(4.13)
(b) For any bounded Borel function f : D
1
R and all 0 t ,
E
y
(f(X
tV
)[T
Z
tV
) =
_

D1
f
_
(Z
tV
). (4.14)
In particular
E
y
(r
i
(X
tV
)[T
Z
tV
) = k
1
. (4.15)
Proof. Note that, by the symmetry of D
2
, V is a (T
Z
t
) stopping time.
(a) Let f T
D1
be bounded, and h be the function with support in D
1
which
equals f in D
1
D
2
, and is harmonic (in the Dirichlet form sense) inside D
2
. Then
since
S

i
(y) = y for 1 i k,

D1
h(y) = k
1
k

i=1
h(
S

i
(y)) = h(y).
Since
D1
h is harmonic (in the Dirichlet form sense) in D
2
and since y E
D
, we
have, using Proposition 2.5, that
h(y) =
D1
h(y) = E
y
(
D1
h)(X
V
) = k
1
E
y
k

i=1
h(
S

i
(X
V
)).
Since f = h on
F
(D
2
),
E
y
(f(X
V
)) = h(y) = k
1
E
y
k

i=1
f(
S

i
(X
V
)).
28
Write
x
for the unit measure at x, and dene measures
i
(, dx) by

1
(dx) =
XV
(dx),
2
(dx) = k
1
k

i=1

i
(XV )
(dx) = k
1
k

i=1

i
(ZV )
(dx).
Then we have
E
y
_
f(x)
1
(dx) = E
y
_
f(x)
2
(dx)
for f T
D1
, and hence for all bounded Borel f dened on
F
(D
2
). Taking
f = r
i
(x)1
U
(x) then gives (4.13).
(b) We can take the cube S

in Denition 4.6 to be S

1
. If g is dened on S

then
U
S
g is the unique extension of g to D(y) such that
D1
U
S
g = U
S
g on D(y). Thus
any function on S is the restriction of a function which is invariant with respect
to
D1
. We will repeatedly use the fact that if
D1
g = g then g(X
t
) = g(Z
t
), and
so also g(X
tV
) = g(Z
tV
).
We break the proof into several steps.
Step 1. Let T
D2
t
denote the semigroup of X stopped on exiting D
2
, that is
T
D2
t
f(x) = E
x
f(X
tV
).
If f T
D1
is bounded, then Proposition 2.6 and Lemma 4.8 imply that q.e. in D
2
T
D2
t

D1
f =
D1
T
D2
t
f. (4.16)
Note that by Proposition 2.6 and [17, Theorem 4.4.3(ii)], the notion q.e. in D
2
coincides for the semigroups T, T
D2
and T, where T is dened in Lemma 4.8.
Step 2. If f, g T
D1
are bounded and
D1
g = g, then we have
D1
(gf) = g
D1
f.
Hence
T
D2
t
(g
D1
f) = T
D2
t

D1
(gf) =
D1
T
D2
t
(gf). (4.17)
Step 3. Let be a Borel probability measure on D
2
. Set

= (
D1
)

. Suppose
that (^
2
) = 0, where ^
2
is dened in Theorem 4.1. If f, g are as in the preceding
paragraph, then we have
E

g(Z
tV
)f(X
tV
) =
_
D2
T
D2
t
_
gf
_
(x)
_

D1
_

(dx)
=
_
D2

D1
_
T
D2
t
(gf)
_
(x)(dx)
=
_
D2
T
D2
t
_
g
D1
f
_
(x)(dx)
= E

g(Z
tV
)
D1
f(X
tV
)
= E

g(Z
tV
)
D1
f(Z
tV
), (4.18)
where we use the denition of adjoint, (4.17) to interchange T
D2
and
D1
, and
that g(X
tV
) = g(Z
tV
).
29
Step 4. We prove by induction that if (^
2
) = 0, m 0, 0 < t
1
< < t
m
< t,
g
1
, . . . , g
m
are bounded Borel functions satisfying
D1
g
i
= g
i
, and f is bounded
and Borel, then
E

(
m
i=1
g
i
(Z
tiV
))f(X
tV
) = E

(
m
i=1
g
i
(Z
tiV
))
D1
f(Z
tV
). (4.19)
The case m = 0 is (4.18). Suppose (4.19) holds for m1. Then set
h(x) = E
x
(
m
i=2
g
i
(Z
(tit1)V
))f(X
(tt1)V
). (4.20)
Write

x
= (
x
)

. By (4.19) for m1, provided x is such that

x
(^
2
) = 0,

D1
h(x) = E

x
(
m
i=2
g
i
(Z
(tit1)V
))f(X
(tt1)V
) (4.21)
= E
x
(
m
i=2
g
i
(Z
(tit1)V
))
D1
f(Z
(tt1)V
). (4.22)
So, using the Markov property, (4.18) and (4.21)
E

(
m
i=1
g
i
(Z
tiV
))f(X
tV
) = E

g
1
(Z
t1V
)h(X
t1V
)
= E

g
1
(Z
t1V
)
D1
h(X
t1V
)
= E

g
1
(Z
t1V
)E
Xt
1
V
_
(
m
i=2
g
i
(Z
(tit1)V
))
D1
f(Z
(tt1)V
)
_
= E

(
m
i=1
g
i
(Z
tiV
))
D1
f(Z
tV
),
which proves (4.19). Therefore since (

x
)

x
,
E

x
(
m
i=1
g
i
(Z
tiV
))f(X
tV
) = E

x
(
m
i=1
g
i
(Z
tiV
))
D1
f(Z
tV
),
and so
E

x
(f(X
tV
)[T
Z
tV
) =
_

D1
f
_
(Z
tV
).
To obtain (4.14), observe that

y
=
y
. Equation (4.15) follows since
D1
r
i
(x) =
k
1
for all x D
1
.
Corollary 4.10 Let f : D(y) R be bounded Borel, and t 0. Then
E
y
(f(X
t
)[T
Z
t
) =
_

D(y)
f
_
(Z
t
). (4.23)
Proof. This follows from Lemma 4.9 by letting the regions D
i
in Denition 4.7
increase to D(y).
Let (A
0
, A
1
), Z be as in Denition 4.6. We now look at X conditional on T
Z
.
Write W
i
(t) =
Si
(Z
t
) S
i
. For any t, we have that X
t
is at one of the points
W
i
(t ). Let
J
i
(t) = j : W
j
(t ) = W
i
(t ),
M
i
(t) =
m

j=1
1
(Wj(t)=Wi(t))
= #J
i
(t),
p
i
(t) = P
x
(X
t
= W
i
(t )[T
Z
t
)M
i
(t)
1
= E
x
(r
i
(X
t
)[T
Z
t
).
30
Thus the conditional distribution of X
t
given T
Z
t
is
k

i=1
p
i
(t)
Wi(t)
. (4.24)
Note that by the denitions given above, we have M
i
(t) = N
n
(W
i
(t)) for 0 t < ,
which is the number of elements of o
n
that contain W
i
(t).
To describe the intuitive picture, we call the W
i
particles. Each W
i
(t) is a
single point, and for each t we consider the collection of points W
i
(t), 1 i m.
This is a nite set, but the number of distinct points depends on t. In fact, we
have W
i
(t), 1 i m = X
t
D. For each given t, X
t
is equal to some of the
W
i
(t). If X
t
is in the r-interior of an element of o
n
, then all the W
i
(t) are distinct,
and so there are m of them. In this case there is a single i such that X
t
= W
i
(t).
If Z
t
is in a lower dimensional face, then there can be fewer than m distinct points
W
i
(t), because some of them coincide and we can have X
t
= W
i
(t) = W
j
(t) for
i ,= j. We call such a situation a collision. There may be many kinds of collisions
because there may be many dierent lower dimensional faces that can be hit.
Lemma 4.11 The processes p
i
(t) satisfy the following:
(a) If T is any (T
Z
t
) stopping time satisfying T on T < then there exists
() > 0 such that
p
i
(T +h) = p
i
(T) for 0 h < .
(b) Let T be any (T
Z
t
) stopping time satisfying T on T < . Then for each
i = 1, . . . k,
p
i
(T) = lim
sT
M
i
(T)
1

jJi(T)
p
j
(s).
Proof. (a) Let D(y) be as dened as in Denition 4.7, and D

=
S
(D(X
T
)).
Let
T
0
= infs 0 : Z
s
, D

, T
1
= infs T : Z
s
, D

;
note that T
1
> T a.s. Let s > 0,
0
be a bounded T
Z
T
measurable r.v., and

1
=

m
j=1
f
j
(Z
(T+tj)T1
), where f
j
are bounded and measurable, and 0 t
1
<
< t
m
s. Write

1
=

m
j=1
f
j
(Z
(tj)T0
). To prove that p
i
((T +s)T
1
) = p
i
(T)
it is enough to prove that
E
x

1
r
i
(X
(T+s)T1
) = E
x

1
p
i
(T). (4.25)
However,
E
x

1
r
i
(X
(T+s)T1
) = E
x
_

0
E(
1
r
i
(X
(T+s)T1
)[T
X
T
)
_
= E
x
_

0
E
XT
(

1
r
i
(X
sT0
))
_
= E
x
_

j
p
j
(T)E
Wj (T)
(

1
r
i
(X
sT0
))
_
. (4.26)
31
If W
j
(T) , S
i
then
E
Wj (T)
(

1
r
i
(X
sT0
)) = 0.
Otherwise, by (4.15) we have
E
Wj (T)
(

1
r
i
(X
sT0
)) = M
i
(T)
1

E
ZT

1
. (4.27)
So,

j
p
j
(T)E
Wj (T)
(

1
r
i
(X
sT0
)) =

j
p
j
(T)1
(jJi(T))
M
i
(T)
1

E
ZT

1
= p
i
(T)

E
ZT

1
. (4.28)
Here we used the fact that p
j
(T) = p
i
(T) if j J
i
(T). Combining (4.26) and
(4.28) we obtain (4.25).
(b) Note that

jJi(T)
r
j
(x) is constant in a neighborhood of X
T
. Hence
lim
sT

jJi(T)
r
j
(X
s
) =

jJi(T)
r
j
(X
T
),
and therefore
lim
sT

jJi(T)
p
j
(s) =

jJi(T)
p
j
(T) = M
i
(T)p
i
(T),
where the nal equality holds since p
i
(T) = p
j
(T) if W
i
(T) = W
j
(T).
Proposition 4.12 Let (A
0
, A
1
), Z be as in Denition 4.6. There exists a constant
q
1
> 0, depending only on d, such that if x A
0
E
D
and T
0
is a nite (T
Z
t
)
stopping time, then
P
x
(X
T0
S[T
Z
T0
) q
1
. (4.29)
Hence
P
x
(T
X
A1
) q
0
q
1
. (4.30)
Proof. In this proof we restrict t to [0, ]. Lemma 4.11 implies that each process
p
i
() is a pure jump process, that is it is constant except at the jump times. (The
lemma does not exclude the possibility that these jump times might accumulate.)
Let
K(t) = i : p
i
(t) > 0,
k(t) = [K(t)[,
p
min
(t) = minp
i
(t) : i K(t) = minp
i
(t) : p
i
(t) > 0.
Note that Lemma 4.11 implies that if p
i
(t) > 0 then we have p
i
(s) > 0 for all
s > t. Thus K and k are non-decreasing processes. Choose I(t) to be the smallest
i such that p
I(t)
(t) = p
min
(t).
32
To prove (4.29) it is sucient to prove that
p
min
(t) 2
dk(t)
2
d2
d
, 0 t . (4.31)
This clearly holds for t = 0, since k(0) 1 and p
i
(0) = r
i
(X
0
), which is for each i
either zero or at least 2
d
.
Now let
T = inft : p
min
(t) < 2
dk(t)
.
Since p
i
(T + h) = p
i
(T) and k(T + h) = k(T) for all suciently small h > 0, we
must have
p
min
(T) < 2
dk(T)
, on T < . (4.32)
Since Z is a diusion, T is a predictable stopping time so there exists an increasing
sequence of stopping times T
n
with T
n
< T for all n, and T = lim
n
T
n
. By the
denition of T, (4.31) holds for each T
n
. Let A = : k(T
n
) < k(T) for all n.
On A we have, writing I = I(T), and using Lemma 4.11(b) and the fact that
k(T
n
) k(T) 1 for all n,
p
min
(T) = p
I
(T) = M
I
(T)
1

jJI (T)
p
j
(T)
= lim
n
M
I
(T)
1

jJI (T)
p
j
(T
n
) 2
d
lim
n
p
min
(T
n
)
2
d
lim
n
2
dk(Tn)
2
d
2
d(k(T)1)
= 2
dk(T)
.
On A
c
we have
p
min
(T) = lim
n
M
I
(T)
1

jJI (T)
p
j
(T
n
)
lim
n
p
min
(T
n
)
lim
n
2
dk(Tn)
= 2
dk(T)
.
So in both case we deduce that p
min
(T) 2
dk(T)
, contradicting (4.32). It follows
that P(T < ) = 0, and so (4.31) holds.
This gives (4.29), and using Proposition 4.5 we then obtain (4.30).
4.3 Properties of X
Remark 4.13 is a doubling measure, so for each Borel subset H of F, almost
every point of H is a point of density for H; see [44, Corollary IX.1.3].
Let I be a face of F
0
and let F

= F I.
Proposition 4.14 There exists a set ^ of capacity 0 such that if x / ^, then
P
x
(
F
< ) = 1.
33
Proof. Let A be the set of x such that when the process starts at x, it never
leaves x. Our rst step is to show F A has positive measure. If not, for almost
every x, T
t
f(x) = f(x), so
1
t
f T
t
f, f) = 0.
Taking the supremum over t > 0, we have c(f, f) = 0. This is true for every
f L
2
, which contradicts c being non-zero.
Recall the denition of E
S
in (2.6). If (E
S
S) = 0 for every S o
n
(F) and
n 1 then (F A) = 0. Therefore there must exist n and S o
n
(F) such that
(E
S
S) > 0. Let > 0. By Remark 4.13 we can nd k 1 so that there exists
S

o
n+k
(F) such that
(E
S
S

)
(S

)
> 1 .
Let S

o
n+k
be adjacent to S

and contained in S, and let g be the map that

reects S

across S

. Dene
J
i
(S

) = T : T o
n+k+i
, T int
r
(S

),
and dene J
i
(S

) analogously. We can choose i large enough so that

(E
S
J
i
(S

)) > (1 2)(S

). (4.33)
Let x E
S
J
i
(S

). Since x E
S
, the process started from x will leave S

with probability one. We can nd a nite sequence of moves (that is, corners or
slides) at level n +k +i so that X started at x will exit S

by hitting S

. By
Proposition 4.12 the probability of X following this sequence of moves is strictly
positive, so we have
P
x
(X(
S
) S

) > 0.
Starting from x E
S
, the process can never leave E
S
, so X will leave S

through B = E
S
S

with positive probability. By symmetry, X

t
started
from g(x) will leave S

in B with positive probability. So by the strong Markov

property, starting from g(x), the process will leave S with positive probability. We
conclude g(x) E
S
as well. Thus g(E
S
J
i
(S

)) E
S
J
i
(S

), and so by (4.33)
we have
(E
S
J
i
(S

)) > (1 2)(S

).
Iterating this argument, we have that for every S
j
o
n+k
(F) with S
j
S,
(E
S
S
j
) (E
S
J
i
(S
j
)) (1 2)(S
j
).
Summing over the S
i
s, we obtain
(E
S
S) (1 2)(S).
Since was arbitrary, then (E
S
S) = (S). In other words, starting from almost
every point of S, the process will leave S.
34
By symmetry, this is also true for every element of o
n
(F) isomorphic to S.
Then using corners and slides (Proposition 4.12), starting at almost any x F,
there is positive probability of exiting F

. We conclude that E
F
has full measure.
The function 1
E
F

is invariant so T
t
1
E
F

= 1, a.e. By [17, Lemma 2.1.4],

T
t
(11
E
F

) = 0, q.e. Let ^ be the set of x where T

t
1
E
F

(x) ,= 1 for some rational

t. If x / ^, then P
x
(X
t
E
F
) = 1 if t is rational. By the Markov property,
x E
F
.
Lemma 4.15 Let U F be open and non-empty. Then P
x
(T
U
< ) = 1, q.e.
Proof. This follows by Propositions 4.12 and 4.14.
4.4 Coupling
Lemma 4.16 Let (, T, P) be a probability space. Let X and Z be random vari-
ables taking values in separable metric spaces E
1
and E
2
, respectively, each fur-
nished with the Borel -eld. Then there exists F : E
2
[0, 1] E
1
that is jointly
measurable such that if U is a random variable whose distribution is uniform on
[0, 1] which is independent of Z and

X = F(Z, U), then (X, Z) and (

X, Z) have
the same law.
Proof. First let us suppose E
1
= E
2
= [0, 1]. We will extend to the general case
later. Let Q denote the rationals. For each r [0, 1] Q, P(X r [ Z) is a
(Z)-measurable random variable, hence there exists a Borel measurable function
h
r
such that P(X r [ Z) = h
r
(Z), a.s. For r < s let A
rs
= z : h
r
(z) > h
s
(z).
If C =
r<s; r,sQ
A
rs
, then P(Z C) = 0. For z / C, h
r
(z) is nondecreasing in
r for r rational. For x [0, 1], dene g
x
(z) to be equal to x if z C and equal
to inf
s>x,sx; sQ
h
s
(z) otherwise. For each z, let f
x
(z) be the right continuous
inverse to g
x
(z). Finally let F(z, x) = f
x
(z).
We need to check that (X, Z) and (

X, Z) have the same distributions. We have
P(X x, Z z) = E[P(X x [ Z); Z z] = lim
s>x,sQ,sx
E[P(X s [ Z); Z z]
= limE[h
s
(Z); Z z] = E[g
x
(Z); Z z].
On the other hand,
P(

X x, Z r) = E[P(F(Z, U) x [ Z); Z z] = E[P(f
U
(Z) x [ Z); Z z]
= E[P(U g
x
(Z) [ Z); Z z] = E[g
x
(Z); Z z].
For general E
1
, E
2
, let
i
be bimeasurable one-to-one maps from E
i
to [0, 1],
i = 1, 2. Apply the above to X =
1
(X) and Z =
2
(Z) to obtain a function F.
Then F(z, u) =
1
1
F(
2
(z), u) will be the required function.
We say that x, y F are m-associated, and write x
m
y, if
S
(x) =
S
(y)
for some (and hence all) S o
m
. Note that by Lemma 2.13 if x
m
y then also
35
x
m+1
y. One can verify that this is the same as the denition of x
m
y given in
[5].
The coupling result we want is:
Proposition 4.17 (Cf. [5, Theorem 3.14].) Let x
1
, x
2
F with x
1

n
x
2
, where
x
1
S
1
o
n
(F), x
2
S
2
o
n
(F), and let =
S1
[
S2
. Then there exists a
probability space (, T, P) carrying processes X
k
, k = 1, 2 and Z with the following
properties.
(a) Each X
k
is an c-diusion started at x
k
.
(b) Z =
S2
(X
2
) =
S1
(X
1
).
(c) X
1
and X
2
are conditionally independent given Z.
Proof. Let Y be the diusion corresponding to the Dirichlet form c and let Y
1
, Y
2
be processes such that Y
i
is equal in law to Y started at x
i
. Let Z
1
=
S1
(Y
1
)
and Z
2
=
S2
(Y
2
). Since the Dirichlet form for
Si
(Y ) is c
Si
and Z
1
, Z
2
have the
same starting point, then Z
1
and Z
2
are equal in law. Use Lemma 4.16 to nd
functions F
1
and F
2
such that (F
i
(Z
i
, U), Z
i
) is equal in law to (Y
i
, Z
i
), i = 1, 2,
if U is an independent uniform random variable on [0, 1].
Now take a probability space supporting a process Z with the same law as Z
i
and two independent random variables U
1
, U
2
independent of Z which are uniform
on [0, 1]. Let X
i
= F
i
(Z, U
i
), i = 1, 2. We proceed to show that the X
i
satisfy
(a)-(c).
X
i
is equal in law to F
i
(Z
i
, U
i
), which is equal in law to Y
i
, i = 1, 2, which
establishes (a). Similarly (X
i
, Z) is equal in law to (F(Z
i
, U
i
), Z
i
), which is equal
in law to (Y
i
, Z
i
). Since Z
1
=
S1
(Y
1
) and Z
2
=
S2
(Y
2
), it follows from the
equality in law that Z =
S1
(Y
1
) and Z =
S2
(Y
2
). This establishes (b).
As X
i
= F
i
(Z, U
i
) for i = 1, 2, and Z, U
1
, and U
2
are independent, (c) is
immediate.
Given a pair of c-diusions X
1
(t) and X
2
(t) we dene the coupling time
T
C
(X
1
, X
2
) = inft 0 : X
1
(t) = X
2
(t). (4.34)
Given Propositions 4.12 and 4.17 we can now use the same arguments as in [5]
to couple copies of X started at points x, y F, provided that x
m
y for some
m 1.
Theorem 4.18 Let r > 0, > 0 and r

= r/L
2
F
. There exist constants q
3
and ,
depending only on the GSC F, such that the following hold:
(a) Suppose x
1
, x
2
F with [[x
1
x
2
[[

< r

and x
1

m
x
2
for some m 1. There
exist c-diusions X
i
(t), i = 1, 2, with X
i
(0) = x
i
, such that, writing

i
= inft 0 : X
i
(t) , B(x
1
, r),
we have
P
_
T
C
(X
1
, X
2
) <
1

2
_
> q
3
. (4.35)
(b) If in addition [[x
1
x
2
[[

< r and x
1

m
x
2
for some m 1 then
P
_
T
C
(X
1
, X
2
) <
1

2
_
> 1 . (4.36)
36
Proof. Given Propositions 4.12 and 4.17, this follows by the same arguments as
in [5], p. 694701.
4.5 Elliptic Harnack inequality
As mentioned in Section 2.1, there are two denitions of harmonic that we can
give. We adopt the probabilistic one here. Recall that a function h is harmonic
in a relatively open subset D of F if h(X
t

D
) is a martingale under P
x
for q.e. x
whenever D

is a relatively open subset of D.

X satises the elliptic Harnack inequality if there exists a constant c
1
such that
the following holds: for any ball B(x, R), whenever u is a non-negative harmonic
function on B(x, R) then there is a quasi-continuous modication u of u that
satises
sup
B(x,R/2)
u c
1
inf
B(x,R/2)
u.
We abbreviate elliptic Harnack inequality by EHI.
Lemma 4.19 Let c be in E, r (0, 1), and h be bounded and harmonic in B =
B(x
0
, r). Then there exists > 0 such that
[h(x) h(y)[ C
_
[x y[
r
_

(sup
B
[h[), x, y B(x
0
, r/2), x
m
y. (4.37)
Proof. As in [5, Proposition 4.1] this follows from the coupling in Theorem 4.18
by standard arguments.
Proposition 4.20 Let c be in E and h be bounded and harmonic in B(x
0
, r).
Then there exists a set ^ of c-capacity 0 such that
[h(x) h(y)[ C
_
[x y[
r
_

(sup
B
[h[), x, y B(x
0
, r/2) ^. (4.38)
Proof. Write B = B(x
0
, r), B

= B(x
0
, r/2). By Lusins theorem, there exist
open sets G
n
such that (G
n
) 0, and h restricted to G
c
n
B

is continuous.
We will rst show that h restricted to any G
c
n
satises (4.37) except when one
or both of x, y is in ^
n
, a set of measure 0. If G =
n
G
n
, then h on G
c
is
H older continuous outside of ^
n
, which is a set of measure 0. Thus h is H older
continuous on all of B

outside of a set E of measure 0.

So x n and let H = G
c
n
. Let x, y be points of density for H; recall Remark
4.13. Let S
x
and S
y
be appropriate isometries of an element of o
k
such that x S
x
,
y S
y
, and (S
x
H)/(S
x
)
2
3
and the same for S
y
. Let be the isometry
taking S
x
to S
y
. Then the measure of (S
x
H) must be at least two thirds the
measure of S
y
and we already know the measure of S
y
H is at least two thirds
that of S
y
. Hence the measure of (S
y
H) ((S
x
H)) is at least one third the
measure of S
y
. So there must exist points x
k
S
x
H and y
k
= (x
k
) S
y
H
that are m-associated for some m. The inequality (4.37) holds for each pair x
k
, y
k
.
37
We do this for each k suciently large and get sequences x
k
H tending to x and
y
k
H tending to y. Since h restricted to H is continuous, (4.37) holds for our
given x and y.
We therefore know that h is continuous a.e. on B

. We now need to show the

continuity q.e., without modifying the function h. Let x, y be two points in B

for
which h(X
tB
) is a martingale under P
x
and P
y
. The set of points ^ where this
fails has c-capacity zero. Let R = [x y[ < r and let > 0. Since (E) = 0, then
by [17, Lemma 4.1.1], for each t, T
t
1
E
(x) = T
t
(x, E) = 0 for m-a.e. x. T
t
1
E
is in
the domain of c, so by [17, Lemma 2.1.4], T
t
1
E
= 0, q.e. Enlarge ^ to include
the null sets where T
t
1
E
,= 0 for some t rational. Hence if x, y / ^, then with
probability one with respect to both P
x
and P
y
, we have X
t
/ E for t rational.
Choose balls B
x
, B
y
with radii in [R/4, R/3] and centered at x and y, resp., such
that P
x
(X
Bx
^) = P
y
(X
By
^) = 0. By the continuity of paths, we can
choose t rational and small enough that P
x
(sup
st
[X
s
X
0
[ > R/4) < and the
same with x replaced by y. Then
[h(x) h(y)[ = [E
x
h(X
tBx
) E
y
h(X
tBy
)[
[E
x
[h(X
tBx
); t <
Bx
] E
y
[h(X
tBy
); t <
By
][ + 2|h|

C
_
R
r
_

|h|

+ 4|h|

.
The last inequality above holds because we have P
x
(X
t
^) = 0 and similarly
for P
y
, points in B
x
are at most 2R from points in B
y
, and X
tBx
and X
tBy
are not in E almost surely. Since is arbitrary, this shows that except for x, y in
a set of capacity 0, we have (4.37).
Lemma 4.21 Let c E. Then there exist constants > 0, C
i
, depending only
on F, such that if 0 < r < 1, x
0
F, y, z B(x
0
, C
1
r) then for all 0 < < C
1
,
P
y
(T
B(z,r)
<
B(x0,r)
) >

. (4.39)
Proof. This follows by using corner and slide moves, as in [5, Corollary 3.24].
Proposition 4.22 EHI holds for c, with constants depending only on F.
Proof. Given Proposition 4.20 and Lemma 4.21 this follows by the same argument
as [5, Theorem 4.3].
Corollary 4.23 (a) c is irreducible.
(b) If c(f, f) = 0 then f is a.e. constant.
Proof. (a) If A is an invariant set, then T
t
1
A
= 1
A
, or 1
A
is harmonic on F. By
EHI, either 1
A
is never 0 except for a set of capacity 0 or else it is 0, q.e. Hence
(A) is either 0 or 1. So c is irreducible.
(b) The equivalence of (a) and (b) in this setting is well known to experts. Suppose
38
that f is a function such that c(f, f) = 0, and that f is not a.e. constant. Then
using the contraction property and scaling we can assume that 0 f 1 and there
exist 0 < a < b < 1 such that the sets A = x : f(x) < a and B = x : f(x) > b
both have positive measure. Let g = b (a f); then c(g, g) = 0 also. By Lemma
1.3.4 of [17], for any t > 0,
c
(t)
(g, g) = t
1
g T
t
g, g) = 0.
So g, T
t
g) = g, g). By the semigroup property, T
2
t
= T
2t
, and hence T
t
g, T
t
g) =
g, T
2t
g) = g, g), from which it follows that g T
t
g, g T
t
g) = 0. This implies
that g(x) = E
x
g(X
t
) a.e. Hence the sets A and B are invariant for (T
t
), which
contradicts the irreducibility of c.
Given a Dirichlet form (c, T) on F we dene the eective resistance between
subsets A
1
and A
2
of F by:
R
e
(A
1
, A
2
)
1
= infc(f, f) : f T, f [
A1
= 0, f [
A2
= 1. (4.40)
Let
A(t) = x F : x
1
= t, t [0, 1]. (4.41)
For c E we set
[[c[[ = R
e
(A(0), A(1))
1
. (4.42)
Let E
1
= c E : [[c[[ = 1.
Lemma 4.24 If c E then [[c[[ > 0.
Proof. Write H for the set of functions u on F such that u = i on A(i), i = 0, 1.
First, observe that T H is not empty. This is because, by the regularity of c,
there is a continuous function u T such that u 0 on the face A(0) and u 1
on the opposite face A(1). Then the Markov property for Dirichlet forms says
0 (u 1) T H.
Second, observe that by Proposition 4.14 and the symmetry, T
A(0)
< a.s.,
which implies that (c, T
A(0)
) is a transient Dirichlet form (see Lemma 1.6.5 and
Theorem 1.6.2 in [17]). Here as usual we denote T
A(0)
= f T : f[
A(0)
= 0.
Hence T
A(0)
is a Hilbert space with the norm c. Let u T H and h be
its orthogonal projection onto the orthogonal complement of T
A(0)A(1)
in this
Hilbert space. It is easy to see that c(h, h) = [[c[[.
If we suppose that [[c[[ = 0, then h = 0 by Corollary 4.23. By our def-
inition, h is harmonic in the complement of A(0) A(1) in the Dirichlet form
sense, and so by Proposition 2.5 h is harmonic in the probabilistic sense and
h(x) = P
x
(X
T
A(0)A(1)
A(1)). Thus, by the symmetries of F, the fact that h = 0
contradicts the fact that T
A(1)
< by Proposition 4.14.
An alternative proof of this lemma starts with dening h probabilistically and
uses [14, Corollary 1.7] to show h T
A(0)
.
39
4.6 Resistance estimates
Let now c E
1
. Let S o
n
and let
n
=
n
(c) be the conductance across S.
That is, if S = Q F for Q Q
n
(F) and Q = a
i
x
i
b
i
, i = 1, . . . , d, then

n
= infc
S
(u, u) : u T
S
, u [
{x1=a1}
= 0, u [
{x1=b1}
= 1.
Note that
n
does not depend on S, and that
0
= 1. Write v
n
= v
E
n
for the
minimizing function. We remark that from the results in [4, 34] we have
C
1

n
F

n
(c
BB
) C
2

n
F
.
Proposition 4.25 Let c E
1
. Then for n, m 0

n+m
(c) C
1

m
(c)
n
F
. (4.43)
Proof. We begin with the case m = 0. As in [4] we compare the energy of v
0
with
that of a function constructed from v
n
and the minimizing function on a network
where each cube side L
n
F
is replaced by a diagonal crosswire.
Write D
n
for the network of diagonal crosswires, as in [4, 34], obtained by
joining each vertex of a cube Q Q
n
to a vertex at the center of the cube by a
wire of unit resistance. Let R
D
n
be the resistance across two opposite faces of F in
this network, and let f
n
be the minimizing potential function.
Fix a cube Q Q
n
and let S = Q F. Let x
i
, i = 1, . . . 2
d
, be its vertices,
and for each i let A
ij
, j = 1, . . . d, be the faces containing x
i
. Let A

ij
be the face
opposite to A
ij
. Let w
ij
be the function, congruent to v
n
, which is 1 on A
ij
and
zero on A

ij
. Set
u
i
= minw
i1
, . . . w
id
.
Note that u
i
(x
i
) = 1, and u
i
= 0 on
j
A

ij
. Then
c(u
i
, u
i
)

j
c(w
ij
, w
ij
) = d
n
.
Write a
i
= f(x
i
), and a = 2
d

i
a
i
. Then the energy of f
n
in S is
c
S
D
(f
n
, f
n
) =

i
(a
i
a)
2
.
Now dene a function g
S
: S R by
g
S
(y) = a +

i
(a
i
a)u
i
(y).
Then
c
S
(g
S
, g
S
) Cc(u
1
, u
1
)

i
(a
i
a)
2
C
n
c
S
D
(f
n
, f
n
).
We can check from the denition of g
S
that if two cubes Q
1
, Q
2
have a common
face A and S
i
= Q
i
F, then g
S1
= g
S2
on A. Now dene g : F R by
40
taking g(x) = g
S
(x) for x S. Summing over Q Q
n
(F) we deduce that
c(g, g) C
n
(R
D
n
)
1
. However, the function g is zero on one face of F, and 1 on
the opposite face. Therefore
1 =
0
= c(v
0
, v
0
) c(g, g) C
n
(R
D
n
)
1
C
n

n
F
,
which gives (4.43) in the case m = 0.
The proof when m 1 is the same, except we work in a cube S o
m
and use
subcubes of side L
nm
F
.
Lemma 4.26 We have
C
1

n

n+1
C
2

n
. (4.44)
Proof. The left-hand inequality is immediate from (4.43). To prove the right-hand
one, let rst n = 0. By Propositions 4.12 and 4.14, we deduce that v
0
C
3
> 0
on A(L
1
F
); recall the denition in (4.41). Let w = (v
0
C
3
)/C
3
. Choose a cube
Q Q
1
(F
1
) between the hyperplanes A
1
(0) and A
1
(L
1
F
); A
1
(t) is dened in
(4.41). Then

1
= c
F1
(v
1
, v
1
) c
F1
(w, w) c(w, w)
= C
2
3
c(v
0
C
3
, v
0
C
3
) C
2
3
c(v
0
, v
0
) = C
4

0
.
Again the case n 0 is similar, except we work in a cube S o
n
.
Note that (4.43) and (4.44) only give a one-sided comparison between
n
(c)
and
n
(c
BB
); however this will turn out to be sucient.
Set
= log m
F
/ log L
F
,
0
= log(m
F

F
)/ log L
F
.
By [5, Corollary 5.3] we have
0
2, and so
F
m
F
L
2
F
. Let
H
0
(r) = r
0
.
We now dene a time scale function H for c. First note that by (4.43) we
have, for n 0, k 0.

n
m
n
F

n+k
m
n+k
F
C
k
F
m
k
F
. (4.45)
Since
F
m
F
L
2
F
> 1 there exists k 1 such that

n
m
n
F
<
n+k
m
n+k
F
, n 0. (4.46)
Fix this k, let
H(L
nk
F
) =
1
nk
m
nk
F
, n 0, (4.47)
and dene H by linear interpolation on each interval (L
(n+1)k
F
, L
nk
F
). Set also
H(0) = 0. We now summarize some properties of H.
41
Lemma 4.27 There exist constants C
i
and

, depending only on F such that the

following hold.
(a) H is strictly increasing and continuous on [0, 1].
(b) For any n, m 0
H(L
nkmk
F
) C
1
H(L
nk
F
)H
0
(L
mk
F
). (4.48)
(c) For n 0
H(L
(n+1)k
F
) H(L
nk
F
) C
2
H(L
(n+1)k
F
). (4.49)
(d)
C
3
(t/s)
0

H(t)
H(s)
C
4
(t/s)

for 0 < s t 1. (4.50)

In particular H satises the fast time growth condition of [20] and [10, Assump-
tion 1.2].
(e) H satises time doubling:
H(2r) C
5
H(r) for 0 r 1/2. (4.51)
(f ) For r [0, 1],
H(r) C
6
H
0
(r).
Proof. (a), (b) and (c) are immediate from the denitions of H and H
0
, (4.43)
and (4.44). For (d), using (4.48) we have
H(L
kn
F
)
H(L
knkm
F
)
C
7
H(L
kn
F
)
H(L
kn
F
)H
0
(L
km
F
)
= C
7
L
km0
F
= C
7
_
L
kn
F
L
knkm
F
_
0
,
and interpolating using (c) gives the lower bound in (4.50). For the upper bound,
using (4.44),
H(L
kn
F
)
H(L
knkm
F
)
C
km
8
= L
km

F
=
_
L
kn
F
L
knkm
F
_

, (4.52)
where

= log C
8
/ log L
F
, and again using (c) gives (4.50). (e) is immediate from
(d). Taking n = 0 in (4.48) and using (c) gives (f).
We say c satises the condition RES(H, c
1
, c
2
) if for all x
0
F, r (0, L
1
F
),
c
1
H(r)
r

R
e
(B(x
0
, r), B(x
0
, 2r)
c
) c
2
H(r)
r

. (RES(H, c
1
, c
2
))
Proposition 4.28 There exist constants C
1
, C
2
, depending only on F, such that
c satises RES(H, C
1
, C
2
).
Proof. Let k be the smallest integer so that L
k
F

1
2
d
1/2
R. Note that if
Q Q
k
and x, y Q, then d(x, y) d
1/2
L
k
F

1
2
R. Write B
0
= B(x
0
, R) and
B
1
= B(x
0
, 2R)
c
.
42
We begin with the upper bound. Let S
0
be a cube in Q
k
containing x
0
: then
S
0
F B. We can nd a chain of cubes S
0
, S
1
, . . . S
n
such that S
n
B
1
and S
i
is adjacent to S
i+1
for i = 0, . . . , n 1. Let f be the harmonic function
in F (S
0
B
1
) which is 1 on S
0
and 0 on B
1
. Let A
0
= S
0
S
1
, and A
1
be the opposite face of S
1
to A
0
. Then using the lower bounds for slides and
corner moves, we have that there exists C
1
(0, 1) such that f C
1
on A
1
. So
g = (f C
1
)
+
/(1 C
1
) satises c
S1
(g, g)
k
. Hence
R
e
(S
0
, B
1
)
1
= c(f, f) c
S1
(f, f) (1 C
1
)
2

k
,
and by the monotonicity of resistance
R
e
(B
0
, B
1
) R
e
(S
0
, B
1
) C
2

1
k
,
which gives the upper bound in (RES(H, c
1
, c
2
)).
Now let n = k +1 and let S Q
n
. Recall from Proposition 4.25 the denition
of the functions v
n
, w
ij
and u
i
. By the symmetry of v
n
we have that w
ij

1
2
on
the half of S which is closer to A
ij
, and therefore u
i
(x)
1
2
if [[xx
i
[[

1
2
L
n
F
.
Now let y L
n
F
Z
d
F, and let V (y) be the union of the 2
d
cubes in Q
n
containing y. By looking at functions congruent to 2u
i
1 in each of the cubes in
V (y), we can construct a function g
i
such that g
i
= 0 on F V (y), g
i
(z) = 1 for
z F with [[z y[[

1
2
L
n
F
, and c(g
i
, g
i
) C
n
. We now choose y
1
, . . . y
m
so
that B
0

i
V (y
i
): clearly we can take m C
5
. Then if h = 1 (

i
g
i
), we have
h = 1 on B
0
and h = 0 on B
1
. Thus
R
e
(B
0
, B
1
)
1
c(h, h) c
_

g
i
,

g
i
_
C
6

n
,
proving the lower bound.
4.7 Heat kernel estimates
We write h for the inverse of H, and V (x, r) = (B(x, r)). We say that p
t
(x, y)
satises HK(H;
1
,
2
, c
0
) if for x, y F, 0 < t 1,
p
t
(x, y) c
1
0
V (x, h(t))
1
exp(c
0
(H(d(x, y))/t)
1
),
p
t
(x, y) c
0
V (x, h(t))
1
exp(c
1
0
(H(d(x, y))/t)
2
).
The following equivalence is proved in [20]. (See also [10, Theorem 1.3, (a)
(c)] for a detailed proof of (a) (b), which is adjusted to our current setting.)
Theorem 4.29 Let H : [0, 1] [0, ) be a strictly increasing function with
H(1) (0, ) that satises (4.51) and (4.50). Then the following are equiva-
lent:
(a) (c, T) satises (V D), (EHI) and (RES(H, c
1
, c
2
)) for some c
1
, c
2
> 0.
(b) (c, T) satises HK(H;
1
,
2
, c
0
) for some ,
1
,
2
, c
0
> 0.
Further the constants in each implication are eective.
43
By saying that the constants are eective we mean that if, for example (a)
holds, then the constants
i
, c
0
in (b) depend only on the constants c
i
in (a), and
the constants in (VD), (EHI) and (4.51) and (4.50).
Theorem 4.30 X has a transition density p
t
(x, y) which satises HK(H;
1
,
2
, C),
where
1
= 1/(
0
1),
2
= 1/(

1), and the constant C depends only on F.

Proof. This is immediate from Theorem 4.29, and Propositions 4.22 and 4.28.
Let
J
r
(f) = r

_
F
_
B(x,r)
[f(x) f(y)[
2
d(x)d(y),
N
r
H
(f) = H(r)
1
J
r
(f),
N
H
(f) = sup
0<r1
N
r
H
(f),
W
H
= f L
2
(F, ) : N
H
(f) < . (4.53)
We now use Theorem 4.1 of [28], which we rewrite slightly for our context. (See
also Theorem 1.4 of [10], which is adjusted to our current setting.) Let r
j
= L
kj
,
where k is as in the denition of H.
Theorem 4.31 Suppose p
t
satises HK(H,
1
,
2
, C
0
), and H satises (4.51) and
(4.50). Then
C
1
c(f, f) limsup
j
N
rj
H
(f) N
H
(f) C
2
c(f, f) for all f W
H
, (4.54)
where the constants C
i
depend only on the constants in (4.51) and (4.50), and in
HK(H;
1
,
2
, C
0
). Further,
T = W
H
. (4.55)
Theorem 4.32 Let (c, T) E
1
.
(a) There exist constants C
1
, C
2
> 0 such that for all r [0, 1],
C
1
H
0
(r) H(r) C
2
H
0
(r). (4.56)
(b) W
H
= W
H0
, and there exist constants C
3
, C
4
such that
C
3
N
H0
(f) c(f, f) C
4
N
H0
(f) for all f W
H
. (4.57)
(c) T = W
H0
.
Proof. (a) We have H(r) C
2
H
0
(r) by Lemma 4.27, and so
N
H
(f) C
1
2
N
H0
(f). (4.58)
Recall that (c
BB
, T
BB
) is (one of) the Dirichlet forms constructed in [5]. By
(4.58) and (4.55) we have T T
BB
. In particular, the function v
E
0
T
BB
(see
Subsection 4.6).
44
Now let
A = limsup
k
H(r
k
)
H
0
(r
k
)
;
we have A C
2
.
Let f T. Then by Theorem 4.31
c
BB
(f, f) C
3
limsup
j
H
0
(r
j
)
1
J
rj
(f)
= C
3
limsup
j
H(r
j
)
H
0
(r
j
)
H(r
j
)
1
J
rj
(f)
C
3
limsup
j
AN
rj
H
(f) C
4
Ac(f, f).
Taking f = v
E
0
,
1 c
BB
(v
E
0
, v
E
0
) C
4
Ac(v
E
0
, v
E
0
) = C
4
A. (4.59)
Thus A C
5
= C
1
4
. By Lemma 4.27(c) we have, for n, m 0,
H(r
n+m
)
H
0
(r
n+m
)
C
6
H(r
n
)
H
0
(r
n
)
.
So, for any n
H(r
n
)
H
0
(r
n
)
C
1
6
A C
5
/C
6
,
and (a) follows.
(b) and (c) are then immediate by Theorem 4.31.
Remark 4.33 (4.56) now implies that p
t
(x, y) satises HK(H
0
,
1
,
1
, C) with

1
= 1/(
0
1).
5 Uniqueness
Denition 5.1 Let W = W
H0
be as dened in (4.53). Let /, B E. We say
/ B if
B(u, u) /(u, u) 0 for all u W.
For /, B E dene
sup(B[/) = sup
_
B(f, f)
/(f, f)
: f W
_
,
inf(B[/) = inf
_
B(f, f)
/(f, f)
: f W
_
,
h(/, B) = log
_
sup(B[/)
inf(B[/)
_
;
h is Hilberts projective metric and we have h(/, B) = h(/, B) for any (0, ).
Note that h(/, B) = 0 if and only if / is a nonzero constant multiple of B.
45
Theorem 5.2 There exists a constant C
F
, depending only on the GSC F, such
that if /, B E then
h(/, B) C
F
.
Proof. Let /

= //[[/[[, B

= B/[[B[[. Then h(/, B) = h(/

, B

). By Theorem
4.32 there exist C
i
depending only on F such that (4.57) holds for both /

and
B

. Therefore
B

(f, f)
/

(f, f)

C
2
C
1
, for f W,
and so sup(B

[/

) C
2
/C
1
. Similarly, inf(B

[/

) C
1
/C
2
, so h(/

, B

)
2 log(C
2
/C
1
).
Proof of Theorem 1.2 By Proposition 1.1 we have that E is non-empty.
Let /, B E, and = inf(B[/). Let > 0 and ( = (1 + )B /. By
Theorem 2.1, ( is a local regular Dirichlet form on L
2
(F, ) and ( E. Since
((f, f)
/(f, f)
= (1 +)
B(f, f)
/(f, f)
, f W,
we obtain
sup(([/) = (1 +) sup(B[/) ,
and
inf(([/) = (1 +) inf(B[/) = .
Hence for any > 0,
e
h(A,C)
=
(1 +) sup(B[/)

_
e
h(A,B)
1
_
.
If h(/, B) > 0, this is not bounded as 0, contradicting Theorem 5.2. We must
therefore have h(/, B) = 0, which proves our theorem.
Proof of Corollary 1.4 Note that Theorem 1.2 implies that the P
x
law of X
is uniquely dened, up to scalar multiples of the time parameter, for all x / ^,
where ^ is a set of capacity 0. If f is continuous and X is a Feller process, the
map x E
x
f(X
t
) is uniquely dened for all x by the continuity of T
t
f. By a limit
argument it is uniquely dened if f is bounded and measurable, and then by the
Markov property, we see that the nite dimensional distributions of X under P
x
are uniquely determined. Since X has continuous paths, the law of X under P
x
is
determined. (Recall that the the processes constructed in [5] are Feller processes.)
Remark 5.3 In addition to (H1)-(H4), assume that the (d1)-dimensional fractal
F x
1
= 0 also satises the conditions corresponding to (H1)-(H4). (This
assumption is used in [22, Section 5.3].). Then one can show (f, f)(F F
0
) = 0
for all f T where (f, f) is the energy measure for c E and f T. Indeed, by
the uniqueness we know that c is self-similar, so the results in [22] can be applied.
46
For h given in [22, Proposition 3.8], we have (h, h)(F [0, 1]
d
) = 0 by taking
i in the last inequality of [22, Proposition 3.8]. For general f T, take an
approximating sequence g
m
T as in the proof of Theorem 2.5 of [22]. Using
the inequality
[(g
m
, g
m
)(A)
1/2
(f, f)(A)
1/2
[ (g
m
f, g
m
f)(A)
1/2
2c(g
m
f, g
m
f)
1/2
,
(see page 111 in [17]), we conclude that (f, f)(F [0, 1]
d
) = 0. Using the self-
similarity, we can also prove that the energy measure does not charge the image
of F [0, 1]
d
by any of the contraction maps.
Remark 5.4 One question left over from [3, 5] is whether the sequence of approx-
imating reecting Brownian motions used to construct the Barlow-Bass processes
converges. Let

X
n
t
= X
n
cnt
, where X
n
is dened in Subsection 3.1 and c
n
is a
normalizing constant. We choose c
n
so that the expected time for

X
n
started at
0 to reach one of the faces not containing 0 is one. There will exist subsequences
n
j
such that there is resolvent convergence for

X
nj
and also weak conver-
gence, starting at every point in F. Any of the subsequential limit points will
have a Dirichlet form that is a constant multiple of one of the c
BB
. By virtue
of the normalization and our uniqueness result, all the limit points are the same,
and therefore the whole sequence

X
n
converges, both in the sense of resolvent
convergence and in the sense of weak convergence for each starting point.
Acknowledgment. The authors thank Z.-Q. Chen, M. Fukushima, M. Hino,
V. Metz, and M. Takeda for valuable discussions, and D. Croydon for correcting
some typos.
References
[1] Alexander, S. and Orbach, R.: Density of states on fractals: fractons. J.
Physique (Paris) Lett. 43 (1982), 625-631.
[2] Barlow, M.T.: Diusions on fractals. Lectures on probability theory and statis-
tics (Saint-Flour, 1995), 1121, Lecture Notes in Math. 1690, Springer, Berlin,
1998.
[3] Barlow, M.T., Bass, R.F.: The construction of Brownian motion on the Sier-
pinski carpet. Ann. Inst. H. Poincare 25 (1989) 225257.
[4] Barlow, M.T., Bass, R.F.: On the resistance of the Sierpinski carpet. Proc. R.
Soc. London A. 431 (1990) 345360.
[5] Barlow, M.T., Bass, R.F.: Brownian motion and harmonic analysis on Sier-
pinski carpets. Canad. J. Math. 54 (1999), 673744.
47
[6] Barlow, M.T., Bass, R.F., Kumagai, T.: Stability of parabolic Harnack inequal-
ities on metric measure spaces. J. Math. Soc. Japan (2) 58 (2006), 485519.
[7] Barlow, M.T., Kumagai, T.: Random walk on the incipient innite cluster on
trees. Illinois J. Math., 50 (2006), 3365.
[8] Barlow, M.T., J arai, A.A., Kumagai, T., Slade, G.: Random walk on the
incipient innite cluster for oriented percolation in high dimensions. Commun.
Math. Phys., 278 (2008), 385431.
[9] Barlow, M.T., Bass, R.F., Kumagai, T.: Note on the equivalence of parabolic
Harnack inequalities and heat kernel estimates.
http://www.math.uconn.edu/bass/papers/phidfapp.pdf
[10] Barlow, M.T., Bass, R.F., Kumagai, T., Teplyaev, A.: Supplementary notes
for Uniqueness of Brownian motion on Sierpinski carpets.
http://www.math.uconn.edu/bass/papers/scuapp.pdf
[11] Barlow, M.T., Perkins, E.A.: Brownian Motion on the Sierpinski Gasket.
Probab. Theory Rel. Fields 79 (1988), 543623.
[12] Ben-Avraham, D., Havlin, S.: Diusion and reactions in fractals and disor-
dered systems. Cambridge University Press, Cambridge, 2000.
[13] Bouleau, N., Hirsch, F.: Dirichlet forms and analysis on Wiener space. de
Gruyter Studies in Mathematics, 14. Walter de Gruyter and Co., Berlin, 1991.
[14] Chen, Z.-Q.: On reected Dirichlet spaces. Probab. Theory Rel. Fields 94
(1992), 135162.
[15] Chen, Z.-Q.: On notions of harmonicity. Proc. Amer. Math. Soc., to appear.
[16] Dynkin, E.B.: Markov Processes - I. Springer, Berlin, 1965.
[17] Fukushima, M., Oshima, Y., Takeda, M.: Dirichlet Forms and Symmetric
Markov Processes. de Gruyter, Berlin, 1994.
[18] Fukushima, M., Shima, T.: On a spectral analysis for the Sierpi nski gasket.
Potential Anal. 1 (1992), 135.
[19] Goldstein, S.: Random walks and diusions on fractals. Percolation theory
and ergodic theory of innite particle systems (Minneapolis, Minn., 19841985),
121129, IMA Vol. Math. Appl. 8, Springer, New York, 1987.
[20] Grigoryan, A., Telcs, A.: Two-sided estimates of heat kernels in metric mea-
sure spaces. In preparation.
[21] Hambly, B.M., Metz, V., Teplyaev, A.: Admissible renements of energy on
nitely ramied fractals. J. London Math. Soc. 74 (2006), 93112.
[22] Hino, M., Kumagai, T.: A trace theorem for Dirichlet forms on fractals. J.
Func. Anal. 238 (2006), 578611.
48
[23] Kesten, H.: Subdiusive behavior of random walk on a random cluster. Ann.
Inst. H. Poincare Probab. Statist., 22 (1986), 425487.
[24] Kigami, J.: A harmonic calculus on the Sierpinski space. Japan J. Appl.
Math. 6 (1989), 259290.
[25] Kigami, J.: A harmonic calculus for p.c.f. self-similar sets. Trans. Amer. Math.
Soc. 335 (1993), 721755.
[26] Kigami, J.: Analysis on fractals. Cambridge Univ. Press, Cambridge, 2001.
[27] Kozma, G., Nachmias, A.: The Alexander-Orbach conjecture holds in high
dimensions. Preprint 2008.
[28] Kumagai, T., Sturm, K.-T.: Construction of diusion processes on fractals,
d-sets, and general metric measure spaces. J. Math. Kyoto Univ. 45 (2005), no.
2, 307327.
[29] Kusuoka, S.: A diusion process on a fractal. Probabilistic methods in math-
ematical physics (Katata/Kyoto, 1985), 251274, Academic Press, Boston, MA,
1987.
[30] Kusuoka, S., Zhou, X.Y.: Dirichlet forms on fractals: Poincare constant and
resistance. Probab. Theory Related Fields 93 (1992), no. 2, 169196.
[31] Lindstrm, T.: Brownian motion on nested fractals. Mem. Amer. Math. Soc.
420 (1990), 1128.
[32] Mandelbrot, B.: The Fractal Geometry of Nature. W.H. Freeman, San Fran-
cisco, 1982.
[33] Metz, V.: Renormalization contracts on nested fractals. J. Reine Angew.
Math. 480 (1996), 161175.
[34] McGillivray, I.: Resistance in higher-dimensional Sierpi nski carpets. Potential
Anal. 16 (2002), no. 3, 289303.
[35] Osada, H.: Singular time changes of diusions on Sierpinski carpets. Stochas-
tic Process. Appl. 116 (2006), 675689.
[36] Peirone, R.: Convergence and uniqueness problems for Dirichlet forms on
fractals. Boll. Unione Mat. Ital. Sez. B Artic. Ric. Mat. (8) 3 (2000), 431460.
[37] Rammal, R., Toulouse G.: Random walks on fractal structures and percola-
tion clusters. J. Physique Lettres 44 (1983) L13L22.
[38] Reed, M., Simon, B.: Methods of modern mathematical physics. I. Functional
analysis. Academic Press, 1980.
[39] Rogers, L.C.G., Williams, D.: Diusions, Markov Processes, and Martingales.
Volume one: Foundations, 2nd ed. Wiley, 1994.
49
[40] Rudin, W.: Functional analysis. McGraw-Hill, 1991.
[41] Sabot, C.: Existence and uniqueness of diusions on nitely ramied self-
similar fractals. Ann. Sci.

Ecole Norm. Sup. (4) 30 (1997), 605673.
[42] Schmuland, B.: On the local property for positivity preserving coercive forms.
Dirichlet forms and stochastic processes (Beijing, 1993), 345354, de Gruyter,
Berlin, 1995.
[43] Strichartz, R.S.: Dierential Equations on Fractals: a Tutorial. Princeton
University Press, Princeton, NJ, 2006.
[44] Torchinsky, A.: Real-Variable Methods in Harmonic Analysis. Academic
Press, Orlando FL, 1986.
50