Dirichlet forms and associated heat kernels on the Cantor set induced by random walks on trees

Jun Kigami Graduate School of Informatics Kyoto University Kyoto 606-8501, Japan e-mail:kigami@i.kyoto-u.ac.jp
Abstract Transient random walk on a tree induces a Dirichlet form on its Martin boundary, which is the Cantor set. The procedure of the inducement is analogous to that of the Douglas integral on S 1 associated with the Brownian motion on the unit disk. In this paper, those Dirichlet forms on the Cantor set induced by random walks on trees are investigated. Explicit expressions of the hitting distribution (harmonic measure) and the induced Dirichlet form on the Cantor set are given in terms of the eective resistances. An intrinsic metric on the Cantor set associated with the random walk is constructed. Under the volume doubling property of with respect to the intrinsic metric, asymptotic behaviors of the heat kernel, the jump kernel and moments of displacements of the process associated with the induced Dirichlet form are obtained. Furthermore, relation to the noncommutative Riemannian geometry is discussed.

Introduction

Transient random walks eventually go to innity, which is not just a single point but a collection of possible behaviors of random walks as the time tends to the innity. A rigorous way to describe this innity is the Martin boundary, where all the boundary values of harmonic functions lie. In certain cases, a transient random walk naturally induces a (Hunt) process (or equivalently a Dirichlet form) on its Martin boundary. In this paper, we are going to study such an induced (Hunt) process in the case of a random walk on a tree, whose Martin boundary is known to coincide with the Cantor set. A well-known example of such an induced Dirichlet form is the Douglas integral 2 2 (() ( ))( () ( )) D(, ) = (d) (d ) 4 0 sin2 ( ) 0 2 on the circle S 1 = {z |z = ei , R}, where (d) = d/2 is the uniform distribution on S 1 . In this case the process which induces the Douglas integral on its Martin boundary S 1 is (reected) Brownian motion on the unit disk 1

D = {z R, |z | < 1}, which is not even a random walk. We use, however, this example to illustrate the procedure of inducement of Dirichlet forms on Martin boundaries. The Dirichlet from associated with the Brownian motion on D is the (half of) Dirichlet integral ( 1 u v u v ) E (u, v ) = + dxdy 2 D x x y y
u 2 dened for u, v H 1 (D) = {u|u, u x , y L (D)}, where D is identied with {(x, y )|x, y R, x2 + y 2 < 1} and the derivatives are in the sense of distributions. Note that the Brownian motion starting from 0 D will hit the boundary S 1 uniformly due to the symmetry. In this case, the uniform distribution on S 1 is called the hitting distribution (or harmonic measure) of S 1 starting from 0. The key stone of the bridge between the Dirichlet integral on D and the Douglas integral on S 1 is the Poisson integral which gives harmonic functions on D from boundary values on S 1 . Let H be the operation of applying the Poisson integral. Namely, we dene H () by

(H ())(re ) =
0 i

1 r2 ( ) (d ). 1 2r cos ( ) + r2

Then Dom(D) = {| L2 (S 1 , ), H H 1 (D)} and D(, ) = E (H, H ) for any , Dom(D). The above example shows us the essence of obtaining Dirichlet forms on Martin boundaries from random walks. Suppose that we have a Markov process or a random walk on a space X , that its Martin boundary M is well-dened and that we have the following ingredients (a), (b) and (c): (a) the Dirichlet form E associated with the original process or the random walk (b) the hitting distribution of M starting from certain point in X (c) the map H which transforms functions on M into harmonic functions on X. Then the induced form EM on the Martin boundary M is given by EM (, ) = E (H, H ) for , FM , where FM = { | L2 (M, ), H Dom(E )}. The theory of the Martin boundary for a transient random walk originated with the classical works of Doob [4] and Hunt [6]. Since then, it has been developed by many authors and, as a result, all the necessary ingredients for constructing an induced form (EM , FM ) are already well-established. For example, one can nd them in [16]. In this paper, we are going to focus on the case of a transient random walk on a tree and study the induced form (EM , FM ) on its Martin boundary. The 2

original motivation of this work comes from the study of traces of the Brownian motion on the Sierpinski gasket. By removing the line segment on the bottom of the Sierpinski gasket, one can associate a random walk on a binary tree with the Brownian motion. We will present the details on this idea in a separate paper[10]. Due to Cartier[2], the Martin boundary of a transient random walk on a tree is known to be (homeomorphic to) the Cantor set = {1, 2}N = {i1 i2 . . . |ij {1, 2} for j N}. Hereafter we use to denote the Martin boundary in place of M . Let us x what we mean by a random walk and a tree at this point. Random walk: Let (V, E ) be a non-directed graph, i.e. V is the set of vertices and E = {(x, y )|x = y T, x and y are connected by an edge} is the set of edges. We assume that {y |(x, y ) E } is a nite set for any x V . For each (x, y ) E , we assign a conductance C (x, y ) > 0 which satises C (x, y ) = C (y, x). Dene p(x, y ) = C (x, z )/ (x,z)E C (x, y ). p(x, y ) gives the probability of transition from x to y in the unit time. Tree: A non-directed graph (T, E ) is a tree if and only if it is connected and there exists no non-trivial cyclic path. We always x a reference point T and assign conductances on (T, E ) as described above. In this framework, assuming the transience of the random walk, we are going to obtain explicit expressions of the hitting distribution starting from the reference point T and the induced form (E , F ) in terms of eective resistances, prove that (E , F ) is a regular Dirichlet form, and construct an intrinsic metric on the Martin boundary with respect to the random walk. Then, assuming the volume doubling property of with respect to the intrinsic metric, we are going to determine asymptotic behaviors of the heat kernel, the jump kernel and moments of displacements of the process generated by the regular Dirichlet form (E , F ). All the results in this paper are obtained in the generality of the above framework. In this introduction, however, we are going to present statements in the case of the binary tree for the sake of simplicity. Let T = m0 {1, 2}m , where {1, 2}0 = {} and {1, 2}m = {w1 . . . wm |w1 , . . . , wm {1, 2}} and let E = {(w, wi), (wi, w)|w T, i {1, 2}}. (T, E ) is called the (complete innite) binary tree. See Figure 1. For w T , we set rwi = C (w, wi)1 for i = 1, 2, Tw = {wv |v T } and w = {wi1 i2 . . . |ij {1, 2} for any j N}. Tw is a subtree of T = T and w is a subset of the Martin boundary = . The quadratic form Ew associated with the subtree Tw is given by 1 Ew (f, f ) = (f (v ) f (vi))2 . r vi i=1,2
v Tw

The eective resistance between w and w with respect to the subtree Tw is given by ( )1 Rw = min{Ew (f, f )|f (w) = 1, the support of f is a nite set} . Under these notations, assuming the transiency of the random walk, we have obtained the following results. 3

rw =

1 C (w ,w)

w = w i 11 12 21 22

rw1
w1

rw2
w2

The binary tree T

Local structure of T

Figure 1: Random walk on the binary tree T Explicit expression of the hitting distribution (Section 3): The hitting distribution is characterized by () = 1 and (wi ) = Rw (w ) rwi + Rwi

for any w T and any i {1, 2}. Explicit expression of (E , F ) (Section 4): For w T , let (u)w be the average of u on w with respect to , i.e. (u)w = w u( ) (d )/ (w ). Then ( )( ) ()w1 ()w2 ( )w1 ( )w2 E (, ) = rw1 + Rw1 + rw2 + Rw2
w T

(1.1)

for any , F = {|E (, ) < +}. In particular, (E , F ) is shown to be a regular Dirichlet form on L2 (, ). Wavelet base consisting of eigenfunctions (Section 4): Dene (w2 )w1 (w1 )w2 w = (w1 )2 (w2 ) + (w2 )2 (w1 ) for any w T , where A is the characteristic function of A . Then {1, w |w T } is a complete orthonormal system of L2 (, ) consisting of the eigenfunctions of the non-negative self-adjoint operator L associated with the Dirichlet form (E , F ). Furthermore, dene Dw = Rw (w ) for any w T . Then L w = (Dw )1 w for any w T . Such a wavelet base on the Cantor set consisting of eigenfunctions has been observed in Kozyrev[11] and PearsonBellissard[13]. In fact, the example in [13] is shown to be a special case of our framework in Section 12. Also it is noteworthy that Dw is expressed as the Gromov product of certain metric related to eective resistances. See (4.1). Construction of intrinsic metric (Section 5): For , with = , the conuence [, ] T is 1 . . . m , where = 1 2 . . . , = 1 2 . . . , 4

m = min{k |k = k } 1. Dene d(, ) = D[, ] . Then d(, ) is a metric on and it is thought of as the intrinsic metric with respect to the random walk. Moreover, we show that has the volume doubling property with respect to d, i.e. (B (, 2r)) c (B (, r)) for any and any r > 0, where c is independent of and r and B (, r) = { | , d(, ) < r}, if and only if there exist c > 0 and (0, 1) such that (wi ) (w ) and Dwi1 ...im c m Dw for any w T and any i, i1 , . . . , im {1, 2}. Asymptotic behavior of the heat kernel (Section 6): Assume that has the volume doubling property with respect to d. Dene p(t, , ) = e[]n1 t e[]n t []n ( ) ([]n )

n0

for any t > 0 and any , , where [ ]n = 1 . . . n for = 1 2 . . . . Then p(t, , ) is continuous on (0, +] and is the heat kernel (fundamental solution) of the heat equation u t = L u. Namely, (eL t u0 )( ) = p(t, , )u0 ( ) (d )

for any u0 L2 (, ). Moreover, p(t, , ) min { } t 1 , d(, ) ([, ] ) (B (, t)) (1.2)

for any t (0, 1] and any , , where the notation is dened at the end of introduction. This heat kernel estimate is a variant of that studied in Chen-Kumagai[3] and satises a typical asymptotic behavior of jump processes. From (1.2), we also have asymptotic behaviors of moments of displacement: if > 1, t E (d(, Xt ) ) t(| log t| + 1) if = 1, (1.3) t if 0 < < 1 for any and any t (0, 1], where E () is the expectation with respect to the Hunt process associated with the Dirichlet form (E , F ) on L2 (, ). In Section 14, (1.3) is shown to hold in general for jump processes which satisfy the heat kernel estimate (1.2). Generalization and inverse problem (Sections 9 and 10): We investigate the following class of quadratic forms Qs on L2 (, ) given by Q(, ) = aw ((),w1 (),w2 )(( ),w1 ( ),w2 ), (1.4)
w T

where aw > 0 for any w T , is a Borel regular probability measure on and (u),v = (v )1 v u( )(d ). We show an equivalent condition for Q being a regular Dirichlet form. Moreover, we consider when a quadratic 5

form Q given by (1.4) is induced by a transient random walk on T . Let w = aw (w )/((w1 )(w2 )) for w T . Dene two conditions (A) and (B) as follows: (A) w < wi for any (w, i) T {1, 2} and limm []m = + for any (B) w < wi for any (w, i) T {1, 2} and limm []m = + for -a.e. . Then we prove (A) Q is induced by a random walk on T (B). Furthermore, we show that neither (A) nor (B) is equivalent to Q being induced by a random walk on T in Section 11. Relation to noncommutative Riemannian geometry (Section 12): In [13], Pearson and Bellissard have constructed noncommutative Riemannian geometry on the Cantor set including Laplacians and Dirichlet forms from an ultra-metric. In fact, their Dirichlet forms belong to the class described by (1.4). As an application of the inverse problem, we prove that Dirichlet forms derived from self-similar ultra-metrics on the binary tree are induced by random walks and obtain heat kernel estimate (1.2) and moments of displacement (1.3). This extends the result of [13] where the authors have studied self-similar ultrametrics whose similarity ratios are equal. Behind all those results, the theory of resistance forms, which is briey reviewed in Section 13, plays an important role. For example, if a random walk on a graph (V, E ) is transient, we may add the innity I to the vertices V and consider a natural resistance form on V {I } associated with the random walk on V . See Section 2 for details. By using this fact, the Martin kernel is expressed by the resistance metric (eective resistance) associated with the resistance form on V {I }. This idea is essentially the key of obtaining the results of this paper. Finally we introduce several conventions in the notation in this paper. (1) Let f and g be real valued functions on a set A. We write f (x) g (x) on A if and only if there exist c1 , c2 > 0 such that c1 g (x) f (x) c2 g (x) for any x A. (2) Let X be a set. We dene (X ) = {u|u : X R}.

Weighted graphs as random walks

Although our main subject is a random walk on a tree, we introduce and study more general framework of weighted graphs in this section. The most important result is (2.3), where the Martin kernel is expressed by the resistance metric. This relation (2.3) is the foundation of all the theories in the following sections. Except that, most of the denitions and results are classical and can be found in [16] for example. First we dene weighted graphs and associated notions.

Denition 2.1. (1) A pair (V, C ) is called a weighted graph if and only if V is a countable set and C : V V [0, ) satises that C (x, y ) = C (y, x) for any x, y V and that C (x, x) = 0 for any x V . The points in V are called vertices of the graph (V, C ). Two vertices x, y V are said to be connected if and only if C (x, y ) > 0. The neighborhood Nx (V, C ) of x V is dened by Nx (V, C ) = {y |C (x, y ) > 0}. For x0 , x1 , . . . , xn V , p = (x0 , x1 , . . . , xn ) is called a path if and only if C (xi , xi+1 ) > 0 for any i = 0, . . . , n 1. We dene the length |p| of a path p = (x0 , x1 , . . . , xn ) by n. If x0 = x and xn = y , then a path (x0 , . . . , xn ) is called a path between x and y . A path p = (x0 , . . . , xn ) is called simple if and only if xi = xj for any i = j . (2) A weighted graph (V, C ) is called irreducible if and only if a path between x and y exists for any x, y V . (3) A weighted graph (V, C ) is called locally nite if and only if Nx (V, C ) is a nite set for any x V . In this paper, we always assume that V is an innite set and a weighted graph (V, C ) is irreducible and locally nite. A weighted graph gives a reversible Markov chain on V . Denition 2.2. Let C (x) = yV C (x, y ) and let p(x, y ) = C (x, y )/C (x). For n = 0, 1, 2, . . ., we dene p(n) (x, y ) for x, y V inductively by p(0) (x, y ) = xy , where xy is the Diracs delta, and p(n+1) (x, y ) = p(n) (x, z )p(z, y ).
z V

Dene G(x, y ) = n0 p(n) (x, y ), which may be innite. G(x, y ) is called the Green function of (V, C ). A weighted graph (V, C ) is said to be transient if and if G(x, y ) < + for any x, y V . We regard p(x, y ) as the transition probability from x to y in the unit time. Let ({Zn }n0 , {Qx }xV ) be the associated random walk (or the Markov chain) on V . Then p(n) (x, y ) = Qx (Zn = y ), i.e. p(n) (x, y ) is the probability of the transition from x to y at time n. Since p(x, y )C (x) = p(y, x)C (y ), this Markov chain is reversible. One can easily see that (V, C ) is transient if G(x0 , y0 ) < + for a single pair of points (x0 , y0 ). Next we dene the notions of Laplacian, harmonic functions and resistance form associated with a (locally nite irreducible) weighted graph (V, C ). Denition 2.3. (1) The Laplacian L : (V ) (V ) associated with (V, C ) is dened by (Lu)(x) = C (x, y )(u(y ) u(x))
y V

for any u (V ) and any x X . We say u (V ) is harmonic on V with respect to (V, C ) if (Lu)(x) = 0 for any x V . We use H(V, C ), H+ (V, C ) and H (V, C ) to denote the collection of harmonic functions, non-negative harmonic 7

functions, bounded harmonic functions on V with respect to (V, C ) respectively. (2) Dene F = {u|u : V R, C (x, y )(u(x) u(y ))2 < +} and 1 C (x, y )(u(x) u(y ))(v (x) v (y )) E (u, v ) = 2
x,y V x,y V

for any u, v F . For a reference point V , we dene E (u, v ) = E (u, v ) + u()v () for any u, v F . Dene C0 (V ) = {u|u (V ), #(supp(u)) < +}, where supp(u) is the support of u and #(A) is the number of elements in A. Proposition 2.4. (1) (E , F ) is a resistance form on V . (2) If u C0 (V ) and v F , then E (u, v ) = u(x)(Lv )(x).
x V

(2.1)

See Section 13 for the denition of resistance forms. Denition 2.5. (E , F ) dened in the last proposition is called the resistance form associated with the weighted graph (V, C ). The following theorem is one of the most essential results on the type problem of random walks. It has originated with Yamasaki in [17, 18]. See [15, Theorem 4.8] for details. Theorem 2.6. The following conditions are equivalent: (Tr1) (V, C ) is transient. (Tr2) 1 / (C0 (V ))E , where (C0 (V ))E is the closure of C0 (V ) with respect to E . (Tr3) (C0 (V ))E = F . (Tr4) sup{u(x)2 /E (u, u)|u C0 (V )} < + for any x V . By the above theorem, if (V, C ) is transient, we may introduce a point I , which is thought of as the point of innity, and construct a new resistance form on V {I }. Later in (2.3), the Martin kernel of (V, C ) is shown to be described by the resistance metric associated with this new resistance form on V {I }. Proposition 2.7. Assume that (V, C ) is transient. Dene F = (C0 (V ))E + R = {f + a|f (C0 (V ))E , a R}. Let I / V and dene u(I ) = a if u = f + a for f (C0 (V ))E and a R. Denote V = V {I }. Then (1) (E , F ) is a resistance form on V . (2) There exists g : V V [0, ) such that the following conditions (a), (b) and (c) are satised: (a) g (x, y ) = g (y, x) for any x, y V . 8

x x (b) For any x V , let g (y ) = g (x, y ). Then g is a unique element in x (C0 (V ))E which satises E (g , u) = u(x) for any u (C0 (V ))E . (c) For any x, y V ,

g (x, y ) =

R (x, I ) + R (y, I ) R (x, y ) , 2

(2.2)

where R (, ) is the resistance metric on V associated with (E , F ) given by } { (u(x) u(y ))2 u F , E (u, u) = 0 . R (x, y ) = max E (u, u) Note that the right-hand side of (2.2) is the Gromov product associated with the metric space (V {I }, R ). Remark. There is no weighted graph (V , C ) whose associated resistance form is (E , F ). Otherwise we have F , where is the characteristic function of the innity I . Then there exists u (C0 (V ))E such that = u + 1. Since u 1 on V and (T, C ) is transient, this contradicts to the fact that 1 / (C0 (V ))E . Remark. Let R(, ) be the resistance metric associated with the resistance form (E , F ) on V . Then R(, ) and R (, ) are called the limit resistance and the minimal resistance respectively in [14]. Also, they are called the free resistance and the wired resistance respectively in [7]. For the use of the terminology free and wired, see [12]. Proof. (1) We will verify the conditions (RFi) for i = 1, . . . , 5 in Denition 13.4. (RF1) and (RF2) are immediate. Let x be the characteristic function of a single point x V . Since x C0 (V ) for any x V , we have (RF3). For x = y V , since F F , the supremum in (RF4) with F = F is nite. If x V and y = I , then the supremum is also nite by the condition (Tr4) of Theorem 2.6. Finally, let u F . Then u = f + a for some f (C0 (V ))E and a R. Then f + a a (C0 (V ))E . Hence u(I ) = (f + a)(I ) = a = u(I ). This shows (RF5) for (E , F ). Thus (E , F ) is a resistance form on V . Applying Theorem 13.2 to the resistance form (E , F ) on V with B = {I }, we obtain the statements (2) and (3). Note that in this case, B is a single point. The remark after Theorem 13.2 shows that RB = R . The element I is considered as the point of innity. Denition 2.8. Assume that (V, C ) is transient. When it is necessary to specify (V, C ), we use I(V,C ) to denote the innity I associated with (E , F ) and write Rx (V, C ) instead of R (x, I ), which is the resistance metric between x and I(V,C ) with respect to (E , F ). If (V, C ) is not transient, then we dene Rx (V, C ) = . According to the use of terminology in Section 13, g (, ) is the I -Green function associated with (E , F ). In light of the following theorem, we call g (, ) the symmetrized Green function of the weighted graph (V, C ). 9

Theorem 2.9. Assume that (V, C ) is transient. Then, for any x, y V , g (x, y ) = G(x, y )/C (y ). To prove the above theorem, we need the next lemma.
(n) Lemma 2.10. Assume that (V, C ) is transient. Dene n (x, y ). y (x) = p n (1) y C0 (V ). (2) For any x, y V (n+1) (Ln (x, y ) p(n) (x, y )) = C (y )(p(n+1) (y, x) p(n) (y, x)). y )(x) = C (x)(p

(3)
m (n+m) E (n (y, x) p(n+m+1) (y, x)) x , y ) = C (y )(p

= C (x)(p(n+m) (x, y ) p(n+m+1) (x, y )). n y y (4) Dene Gy (x) = n0 y (x). Then G (C0 (V ))E and E (u, G ) = C (y )u(y ) for any u (C0 (V ))E . Proof. Since (V, C ) is locally nite, we verify (1). By using (2.1), routine calculations show (2) and (3). n1 (4) Let Gn,y (x) = i=0 i y . For n < m, by (3), E (Gm+1,y Gn,y , Gm+1,y Gn,y ) = C (y )
+m (n i=2n

p(i) (y, y )

2 m+1 i=n+m+1

) p(i) (y, y ) .

Since i=0 p(i) (y, y ) is nite, we see that E (Gm+1,y Gn,y , Gm+1,y Gn,y ) 0 n1 as n, m . Moreover, Gn,y (x) = i=0 p(i) (x, y ) G(x, y ) as n for x V . Hence {Gn,y }n1 is a Cauchy sequence with respect to E and its limit is Gy . Note that Gn,y C0 (V ). Hence Gy (C0 (V ))E . For any u C (V0 ), E (u, Gn,y ) =
n 1 i=0

E (u, i y) =
n 1

n 1

u(x)(Li y )(y )

i=0 xV

= C (y ) = C (y ) Letting n , we obtain

u(x)(p(i) (y, x) p(i+1) (y, x))

xV

i=1 xV

u(x)(p(0) (y, x) p(n) (y, x)).

E (u, Gy ) = C (y )

x V

u(x)p(0) (y, x) = C (y )u(y ).

10

RI = g (x, y ) rxI ryI Rx

Ry
y

rxy
Trace to {x, y, I }

After the Y transform

Figure 2: Calculation of g (x, y ) by means of Y transform Proof of Theorem 2.9. By Lemma 2.10-(4), it follows that Gy (C0 (V ))E and E (u, Gy /C (y )) = u(y ) for any u (C0 (V ))E . Proposition 2.7-(b) shows that G(x, y )/C (y ) = Gy (x)/C (y ) = g (x, y ). The next lemma is technically useful in the following sections. One can refer to Theorem 13.3 for the denition of a trace of a resistance form. Lemma 2.11. Assume that (V, C ) is transient. Set U = {x, y, I } for x, y X with x = y . Let (E|U , F |U ) be the trace of the resistance form (E , F ) on U . Suppose that E|U (u, u) = cxI (u(x) u(I ))2 + cyI (u(y ) u(I ))2 + cxy (u(x) u(y ))2 for any u : U R. Dene rxI = 1/cxI , ryI rxI ryI rxI +ryI +rxy 0 g (x, y ) = ryI rxI = 1/cyI and rxy = 1/cxy . Then if cxI cyI cxy > 0, if cxy = 0, if cxI = 0, if cyI = 0.

This lemma is proven by using the Y transform to E|U as illustrated in Figure 2. See [9, Lemma 2.1.15] for the Y transform. Proof. Assume cxy cyI cxI > 0. Let Rx = rxI rxy /R and Ry = ryI rxy /R and RI = rxI ryI /R, where R = rxI + ryI + rxy . Dene a resistance form E on four points {x, y, I, z } by E (u, u) =
p=x,y,I

1 (u(p) u(z ))2 . Rp

Then by the Y transform, (E|U , F |U ) is equal to the trace of resistance form E to U . Hence R (x, I ) = Rx + RI , R (y, I ) = Ry + RI and R (x, y ) = Rx + Ry . 11

Using (2.2), we immediately obtain g (x, y ) = RI . If any of cxy , cxI and cyI is 0, then direct calculation easily shows the proposition. In the rest of this section, we introduce the notion of the Martin boundary of a transient weighted graph and related results originally studied in 1960s. See [16, Section 24] for the references and details. Denition 2.12. Assume (V, C ) is transient. Dene Kx0 (x, y ) = G(x, y ) G(x0 , y )

for any x0 , x, y T . Kx0 (x, y ) is called the Martin kernel of (V, C ). Using (2.2), we have Kx0 (x, y ) = g (x, y ) R (x, I ) + R (y, I ) R (x, y ) = , g (x0 , y ) R (x0 , I ) + R (y, I ) R (x0 , y ) (2.3)

where R is the resistance metric with respect to (E , F ). Proposition 2.13. Assume (V, C ) is transient. Then there exists a unique minimal compactication V of V (up to homeomorphism) such that Kx0 is extended to a continuous function from V V to R. V is independent of the choice of x0 . Moreover, there exists a V \V -valued random variable Z such that ( ) Qx lim Zn = Z = 1
n

for any x X . Denition 2.14. Assume that (V, C ) is transient. V is called the Martin compactication of V . Dene M (V, C ) = V \V , which is called the Martin boundary of (V, C ). Dene a probability measure x on M (V, C ) by x (B ) = Qx (Z B ) for any Borel set B M (V, C ). x is called the hitting distribution on the Martin boundary M (V, C ) starting from x. Note that x s for x V are mutually absolutely continuous and satisfy x = yV p(x, y )y . The following theorem is the fundamental result on the Martin boundary and representation of harmonic functions on a weighted graph. Theorem 2.15. Assume that (V, C ) is transient. (1) Kx0 (, y ) H+ (V, C ) H (V, C ) for any x0 V and any y V . (2) For any h H+ (V, C ), there exists a Borel regular measure h on M (V, C ) such that h(x) = Kx0 (x, y ) h (dy ).
M (V,C )

(3) For any h H (V, C ), there exists f L (M (V, C ), x0 ) such that h(x) = Kx0 (x, y )f (y )x0 (dy ).
M (V,C )

12

Transient trees and their Martin boundaries

In this section, we introduce the notion of a tree, which is a special case of weighted graphs. The main goals are Theorems 3.8 and 3.13, where the hitting distribution on the Martin boundary and the Martin kernel are expressed in terms of resistance metrics of sub-trees. Denition 3.1. (1) A pair (T, C ) is called a tree if and only if it satises the following conditions (Tree1) and (Tree2): (Tree1) (T, C ) is an irreducible and locally nite weighted graph. (Tree2) For any x, y T , there exists a unique simple path between x and y . (2) Let (T, C1 ) and (T, C2 ) be trees. Then (T, C1 ) and (T, C2 ) is said to have a common graph structure if and only if C1 (x, y ) = 0 implies C2 (x, y ) = 0, and vise versa for any x, y T . We write (T, C1 ) (T, C2 ) if (T, C1 ) and (T, C2 )
G

have a common graph structure. Fix a countably innite set T . Then the relation is an equivalence relation on {(T, C )|(T, C ) is a tree}. We use G(T, C ) to denote the equivalence class of a tree (T, C ) with respect to . The equivalence class G(T, C ) determines the
G G

structure of T as a non-directed graph. Hereafter in this section, (T, C ) is always a tree. Denition 3.2. (1) For x, y T , the unique simple path between x and y is called the geodesic between x and y and denoted by xy . The path distance |x, y | between x and y is dened by the length of the geodesic between x and y . (2) For x T , dene x : T T by { xn1 if x = y and xy = (x0 , . . . , xn1 , xn ), x ( y ) = x if y = x. If x = y , x (y ) is called the predecessor of y with respect to x. Also we dene Sx (y ) = Ny (T, C )\{x (y )}. The points in Sx (y ) are called successors of y with respect to x. (3) An innite path (x0 , x1 , . . .) T N is called a geodesic ray originated from x V if and only if x0 = x and (x0 , x1 , . . . , xn ) = x0 xn for any n 1. Two geodesic rays (x0 , x1 , . . .) and (y0 , y1 , . . .) is equivalent if and only if (xm , xm+1 , . . .) = (yk , yk+1 , . . .) for some m and k . Dene (T, C ) as the collection of the equivalence classes of geodesic rays. We write T = T (T, C ). Easily, (T, C ) is identied with the collections of geodesic rays originated from a xed point x T . Lemma 3.3. Let x T . Dene x (T, C ) by the collection of the geodesic rays originated from x. Then (T, C ) is naturally identied with x (T, C ). x y (T, C )
x We need the notion of sub-tree Ty and associated collection of geodesic rays to describe ner structure of x (T, C ).

13

Denition 3.4. We write z xy if and only if xy = (x0 , . . . , xn ) and z = xi x for some i = 0, . . . , n. Dene Ty = {z |y xz } and
x x y (T, C ) = {(x0 , x1 , . . .)|(x0 , x1 , . . .) (T, C ), xm = y for some m 0}. x x Dene Ty = Ty x y (T, C ).

By Lemma 3.3, we always identify (T, C ) as x (T, C ). If no confusion x x may occur, we write , x and x y in place of (T, C ), (T, C ) and y (T, C ) respectively hereafter. Proposition 3.5. Dene O = {U |U T , for any x U ,
x there exists y T such that x y and Ty U .}

Then O gives a topology of T and (T , O) is compact. Moreover, the closure of T is T . The restriction of O to T is the discrete topology on T . (T , O) is called the end compactication of T . Remark. Note that the notions introduced in this section so far only depend on the equivalence class G(T, C ). In particular, if (T, C1 ) and (T, C2 ) are trees and G(T, C1 ) = G(T, C2 ), then (T, C1 ) is naturally identied with (T, C2 ) and the end compactications of (T, C1 ) and (T, C2 ) are the same. The following fundamental theorem on the Martin boundary of a tree is due to Cartier [2]. See [16] for details. Theorem 3.6. Assume (T, C ) is transient. Then the Martin compactication T of T coincides with the end compactication T . For any h H (T, C ), there exists a unique L (M (T, C ), x0 ) such that h(x) = Kx0 (x, y )(y )x0 (dy )
M (T,C )

for any x V . Moreover, h(xn ) ( ) as n x0 -almost every (T, C ), where (x0 , x1 , . . .) is the geodesic ray originated from x0 representing (T, C ). By the above theorem, we identify the Martin boundary M (T, C ) with hereafter. Let O be the relative topology of O on . Then (, O ) is compact. Recall that x is the hitting distribution on starting from x V dened by x (B ) = Qx (Z B ) for a Borel set B . We are going to give an expression x x of x (x y ) by means of resistance metrics of sub-trees (Ty , Cy ) dened below.
x x Denition 3.7. Let x, y T with x = y . Dene ry = 1/C (x (y ), y ). Let Cy x x x x x x x be the restriction of C onto Ty . We write Ry = Ry (Ty , Cy ) and y = ry + Ry .

14

x x x x x Remark. If (Ty , Cy ) is not transient, then Ry = x y = . Moreover, (Ty , Cy ) x x is not transient if and only if (Tz , Cz ) is not transient for all z Sx (y ).

Theorem 3.8. Assume that (T, C ) is transient. Then { Rx x (y ) x x x (x x x (y ) ) if (Tx (y ) , Cx (y ) ) is transient, x y x (y ) = 0 otherwise for any y T \{x}. By the inductive use of this theorem, if xy = (x0 , . . . , xn ), then x x x Rx0 Rx1 Rxn1 if (T x , C x ) is transitive, x x y y x x1 x2 x xn (y ) = 0 x x if (Ty , Cy ) is not transitive. Lemma 3.9. Assume (T, C ) is transient. Then
x x (x y ) = Rx (T, C )/y

(3.1)

for any y Nx (T, C ).

1 . Note that Rx (T, C ) = ( yNx (T,C ) 1/x y)

x x x Proof. If (Ty ) is not transient, then x (x , Cy y ) = Qx (Z y ) = 0. Assume x x that (Ty , Cy ) is transient. Let F (x, y ) = Qx (Zn = y for some n 0). By [16, (1.13)-(b)] and Theorem 2.9,

F (x, y ) = Using [16, (26.5)], we have

G(x, y ) g (x, y ) = G(y, y ) g (y, y )

(3.2)

F (x, y )(1 F (x, y )) (g (x, x) g (x, y ))g (x, y ) = . 1 F (x, y )F (y, x) g (x, x)g (y, y ) g (x, y )2

(3.3)

We consider the trace of (E , F ) to the three points {x, y, I }, where I = I(T,C ) as in Lemma 2.11. By using the same notations, g (x, x) = Rx (T, C ) = rxI (rxy + ryI ) , R ryI (rxy + rxI ) g (y, y ) = Ry (T, C ) = , R rxI ryI , g (x, y ) = R

where R = rxy + rxI + ryI . (3.3) and these imply

rxI Rx (T, C ) Rx (T, C ) = = . R rxy + ry x y

15

x x Proof of Theorem 3.8. Let Jy = {Zn Ty for suciently large n}. Then x x x x x (y ) = Qx (Jy ). Set Iz = {Zn Tz for any n 0}, where z = x (y ). By the Markov property, x x x x x x x x x x (x y ) = Qx (Jy ) = Qx (Jz )Qz (Jy Iz |Iz ) = x (z )Qz (Jy Iy |Iz ). x x x ) be the Markov chain associated with (T , C ) and Let ({Zn }n0 , {Qw }wTz z z let z is the associated hitting distribution. Then x x x x Qz (Jy Iz |Iz ) = Qz (Zn Ty for suciently large n) = z (x y ). x x x x Applying Lemma 3.9 to (Tz , Cz ), we obtain z (x y ) = Rz /y . Thus we have (3.1).

As an application of Theorem 3.8, we may identify the support of x , which is the Poisson boundary of (T, C ). Denition 3.10. Dene
x x x ) is transient} = {y |y T, (Ty , Cy T x x and let C be the restriction of C to T . x x (T , C ) is naturally identied with {(x0 , x1 , . . .)|(x0 , x1 , . . .) x , xm x x ) is regarded as a subset of . , C for any m 0}. In this sense, (T

x T

x x Corollary 3.11. The support of x is (T , C ).

Next we give an expression of the Martin kernel in terms of resistance metrics of sub-trees. Denition 3.12. Assume that (T, C ) is transient. For x = y T , dene { Rx y x x if (Ty , Cy ) is transient, x x y = y 1 otherwise. Theorem 3.13. Let x0 , x T with x0 = x and let x0 x = (x0 , x1 , . . . , xn ), where xn = x. For any y T , dene k (x0 , x, y ) as the unique k which satises x0 x y Tx Tx . Then, k k ( x x )1 x0 x x0 Kx0 (x, y ) = x x xk+1 x , 0 x1 n k1 where k = k (x0 , x, y ). The rest of this section is devoted to proving Theorem 3.13. Lemma 3.14. Assume that (T, C ) is transient. Let z T . If x Nz (T, C ) z and y Tx , then x (3.5) Kz (x, y ) = 1/z . (3.4)

16

rzx

x Rz

rxI

rxy

RxI

rxy

g (x, y )

ryI I

ryI y I

g (x, y )-1

g (x, y )-2

z rzw

rxw w rwI ryI

z rzw

z rzw w rwI ryI rwy g (z, y ) y I

rxy

g (z, y )-1

g (z, y )-2

g (z, y )-3

Figure 3: Calculation of g (x, y ) and g (z, y ) Proof. We have three cases, namely, x x z z Case 1: both (Tz , Cz ) and (Tx , Cx ) are transient. x x z z Case 2: (Tz , Cz ) is transient while (Tx , Cx ) is not. x x z z Case 3: (Tx , Cx ) is transient while (Tz , Cz ) is not. First we consider Case 1. Let cxI (u(x) u(I ))2 + cyI (u(y ) u(I ))2 + cxy (u(x) u(y )2
z z ). Assume that , Cz be the trace of the resistance form associated with (Tx cxI cyI cxy > 0. Set rxI = 1/cxI , ryI = 1/cyI and rxy = 1/rxy . Now our network corresponds to O at the upper left corner of Figure 3. We will follow the arrows in Figure 3 to calculate g (x, y ) and g (z, y ). To obtain g (x, y ), we x calculate the combined resistance RxI of rxz + Rz and rxI and get g (x, y )1 from O. Applying Y transform to g (x, y )-1, we have g (x, y )-2. Then Lemma 2.11 gives g (x, y ). After performing these processes,

g (x, y ) =

(ryI +

x rxy )(Rz

x rxI ryI (Rz + rxz ) . x + r )r + rxz + rxI ) + (Rz xz xI

To get g (z, y ), we rst apply the Y transform to three points network {z, x, I } in O and obtain g (z, y )-1. To set rwy = rxw + rxy gives g (z, y )-2. Finally applying the Y transform to three point network {I, w, y }, we get 17

g (x, y )-3. Then Lemma 2.11 gives g (z, y ). After performing these processes, g (z, y ) = (ryI
x rxI ryI Rz . x+r x + rxy )(Rz xz + rxI ) + (Rz + rxz )rxI

Since Kz (x, y ) = g (x, y )/g (z, y ), we have (3.5). If one of cxI , cyI and cxy = 0, then the corresponding edge in O of Figure 3 is disconnected. The calculation is considerably easier and one can conrm (3.5) as well. In Case 2, rxI and ryI are innite and the corresponding edges in O of Figure 3 are disconnected. Hence it is easy to obtain (3.5). x In Case 3, Rz is innite and the corresponding edge in O of Figure 3 is disconnected. Therefore g (x, z ) = g (x, y ) and this implies (3.5). Proof of Theorem 3.13. Let z = x0 . We use induction in |z, x|. Assume |z, x| = z z z x 1. Then y Tx or y Tx . If y Tx , then Lemma 3.14 shows (3.4). If y Tz , 1 z then Kz (x, y ) = (Kx (z, y )) = z and so (3.4) follows. Next assume that (3.4) holds if |z, x| m. Let |z, x| = m + 1 and let zx = (x0 , x1 , . . . , xm , xm+1 ). Then by Lemma 3.14, {( )1 x xm x if y Tx , m Kz (x, y ) = Kz (xm , y )Kxm (x, y ) = Kz (xm , y ) xm x x if y Txm .
xm z xm z Note that x = x and Tx = Tx . Hence using the hypothesis of the induction, we obtain (3.4). Thus we have completed the proof.

Induced form on the Martin boundary

In this section, we present a structure theorem (Theorem 4.4) of the Dirichlet form on the Cantor set induced by a transient random walk on a tree. Throughout this section, (T, C ) is a transient tree and T is a reference point. The trace on the Martin boundary (E , F ) is dened as follows. Denition 4.1. Fix a reference point T . Write T# = T \{}. Dene a linear map H : L1 ( , ) (T ) by H (f )(x) = K (x, y )f (y ) (dy )
x

for any x X . Moreover, dene F = {f |f L2 ( , ), H (f ) F} and E (f, g ) = E (H (f ), H (g )) for any f, g F .

18

Note that H (f ) H(T, C ) since K (, y ) H(T, C ) for any y . We are going to study the quadratic form (E , F ). Since the support of ) in place of (T, C ). Equivalently, we , C ), we will consider (T , C is (T x x assume that (Ty , Cy ) is transient for any x, y T hereafter. As a result, the Martin boundary is equal to the Poisson boundary and every edge in T must has at least one successor, i.e. Sx (y ) = for any x, y T . (Note that if Sx (y ) = , x x then (Ty , Cy ) is recurrent.) For ease of notation, we omit writing in notations. , x , Rx , Cx For example, we use , , Tx , Cx , Rx , x and x instead of , , Tx and x . The values Dx and x dened below play an essential role in this paper. For example, {x |x T } {0} will be identied with the collection of eigenvalues of the self-adjoint operator associated with (E , F ). Denition 4.2. Dene Dx = (x )Rx and x = 1/Dx for any x T . The map x x from T to (0, ) is called the eigenvalue map associated with (T, C ). By (2.2) and Theorem 5.2-(1), we will see that Dx = R (x, I ) + R (, I ) R (x, ) , 2 (4.1)

whose right-hand side coincides with the Gromov product of the metric R . Lemma 4.3. For any x T and any y S (x), Ry Dy = <1 Dx y and Dx Dy = ry (y ) In particular, Dx > Dy and y > x for any x T and any y S (x). Proof. By Theorem 3.8, Dy Rx Ry Ry Ry = = = <1 Dx y Rx y ry + Ry Again by Theorem 3.8, (y )y = (x )Rx . This immediately implies (4.3). Since every (Ty , Cy ) is assumed to be transient, it follows that (y ) > 0. The next theorem shows that E| has a simple expression by means of {x }xT and . Theorem 4.4. For any f L1 (, ), x E (H (f ), H (f )) = 2 (x ) Rx = 2
xT xT

(4.2)

(4.3)

( )2 (y ) (z ) (f )y (f )z (4.4)

y,z S (x)

where (f )y = (y )1 y f (x) (dx).

y,z S (x)

)2 1 ( (f )y (f )z y z

19

Remark. By Theorem 3.8, we see that Rx x (y ) (z ) = 2 (x ) 2y z for any x T and any y, z S (x). The statement of the above theorem includes E (H (f ), H (f )) < + if and only if the right-hand side of (4.4) is nite. Denition 4.5. (1) Dene |x| = |, x| for any x T . Let x = (x0 , . . . , xn ). Then dene [x]m = xm for any m = 0, 1, . . . , n = |x|. (2) Let = (x0 , x1 , x2 , . . .) . Dene [ ]n = xn for n 0. (3) For , with = , dene N (, ) = max{n|[ ]n = [ ]n } and [, ] = [ ]N (, ) . We use [, ]m to denote [[, ]]m . Using Theorem 4.4, we may realize aspects of the nature of the quadratic form (E , F ) as follows. Theorem 4.6. (1) (E , F ) is a regular Dirichlet form on L2 (, ). (2) Dene J : ( )\ [0, ) by ( ) N (, )1 [, ]m+1 [, ]m 1 + J (, ) = 2 ([, ]m+1 ) m=0 (4.5)

for any , with = , where = {(, )| }. Then { } F = u u L2 (, ), J (, )(u( ) u( ))2 (d ) (d ) < +

and, for any u, v F , E (u, v ) = J (, )(u( ) u( ))(v ( ) v ( )) (d ) (d ).

J (, ) is called the jump kernel of (E , F ). (3) Let L be the non-negative self-adjoint operator on L2 (, ) associated with (E , F ) on L2 (, ). Dene } { ay /y = 0 Ex = = ay y ,
y S (x) y S (x)

for x T . Then Ex is contained in the domain of L and L = x for any Ex . Moreover, let {x,i }i=1,...,#(S (x))1 be a orthonormal base of Ex with respect to L2 (, )-inner product. Then { , x,i |x T, i = 1, . . . , #(S (x)) 1} is a complete orthonormal system of L2 (, ). Remark. By Lemma 4.3, J (, ) > 0 for any (, ) ( )\.

20

We will prove Theorem 4.6 in Section 9 as a special case of generalized framework. Let us forget (T, C ) for the moment and think T as a non-direct graph. Namely, we x T and choose an equivalence class with respect to . Then we have : T T , , S (x), x and so on as we remarked after
G

Proposition 3.5. Let be a Borel regular probability measure on which satises (x ) > 0 for any x T and let : T [0, ). Dene a quadratic form (x) Q(u, v ) = (y )(z )((u)y, (u)z, )((v )y, (v )z, ), 2(x )
xT

where (f )x, = (x )1 x f (y )(dy ). Set D = {f |f L2 (, ), Q(f, f ) < +}. We study this quadratic form (Q, D) in Section 9. For example, (Q, D) will be shown to be a regular Dirichlet form on L2 (, ) if the counterpart of J (, ) is non-negative for any , with = . Moreover, let L be the self-adjoint operator on L2 (, ). Dene Ex, = associated non-negative { yS (x) ay y | yS (x) ay (y ) = 0}. Then we will see that L = (x) for any Ex . See Section 9 for details. The rest of this section is devoted to proving Theorem 4.4. Lemma 4.7. Let x = (x0 , . . . , xn ), where x0 = and xn = x. Then
x x (x ) = x x 0 n2 x Rx n 1 x0 x + rxn1 xn Rx + Rx n1

y,z S (x)

(4.6)

Proof. We use induction in |x|. First if |x| = 1, then by Theorem 3.8, (x ) = R . x

x 1 x x Since (R )1 = (rx + Rx )1 + (R ) , we have (x ) = R /(R + x ). This implies (4.6) in this case. Now it is enough to show that

My =

R(y) M(y) y

(4.7)

for any y T with |y | 2, where My is the right-hand side of (4.6). Let x = (y ) and z = (x). Then,
y x z My Ry Rx Rz + Rx + rzx = y z . y x+r x Mx Rz + rzx Rx + Ry R xy z

(4.8)

y x z y Note that Rz = Rz . Dene T0 = {x} T \(Tx Tx ) and let C0 be the restriction of C onto T0 . Set x = Rx (T0 , C0 ), i.e. x is the resistance between x and the x x innity I(T0 ,C0 ) . Set y = rxy + Ry and z = rzx + Rz . (See Figure 4. Three x ,C x ) and I(T x ,C x ) are identied as the innity I(T,C ) .) Then innities I(T0 ,C0 ) , I(Ty y z z y z it follows that Rx = x z /(z + x ) and Rx = y x /(y + x ). Applying these to (4.8), we have My /Mx = x /(y + x ). Since Rx = x y /(x + y ) and y = y , (4.7) follows.

21

I(T0 ,C0 )

x
x Rz
x ,C x ) I(Tz z

rzx z x

rxy y

x Ry
x ,C x ) I(Ty y

x z = Rz + rzx

x y = Ry + rxy

Figure 4: Calculation of My /Mx Theorem 4.8. For any (x, y ) T# T , K (x, y ) = x K ( (x), y ) + (1 x ) 1 b (y ). (x ) Tx

Proof. Let x = (x0 , . . . , xn1 , xn ) where x0 = and xn = x. Write z = xn1 (= (x)). If x / Tx , then Theorem 3.13 yields K (x, y ) = x K (z, y ). Suppose x Tx . Again by Theorem 3.13, ( x 1 ) x 1 K (x, y ) = (z ) K (z, y ) = x K (z, y ) + (z ) x K (z, y ) ( Rx + r ) z Rx zx z = x K (z, y ) + K (z, y ) x z +r Rz Rx zx z R x + Rx + rzx x x x = x K (z, y ) + (1 x ) z ( x1 x ) 1 n 2 x Rz Now Lemma 4.7 implies K (x, y ) = x K (z, y ) + (1 x )/ (x ). Corollary 4.9. For any f L1 (, ), { (f ) = f d H (f )(x) = x H (f )( (x)) + (1 x )(f )x

if x = , if x = .

(4.9)

Moreover, let f, u L1 (, ). If H (f )(x) = H (u)(x) for any x T , then f ( ) = u( ) for -almost every . Proof. (4.9) is direct from Theorem 4.8. Using (4.9) inductively, we see that (f )x = (u)x for any x T . Therefore, f ( ) = u( ) for -almost every .

22

Denition 4.10. Dene Wm = {x|x T, |x| = m} for m 0. For f L1 (, ), dene [f ]m = xWm (f )x x , E0 (f ) = 0 and, for m 1, Em (f ) =
xm k=1 Wk

)2 1( H (f )( (x)) H (f )(x) . rx

Corollary 4.9 implies the following lemma. Lemma 4.11. For any m 1 and any f L1 (, ), rx ( )2 Em (f ) = Em1 (f ) + H (f )( (x)) (f )x 2 (x ) xWm ry ( )2 = Em1 (f ) + H (f )(x) (f )y 2 (y )
xWm1 y S (x)

Denition 4.12. Let (Ex , Fx ) be the resistance form associated with (Tx , Cx ) for x T . Dene ( ) 1 Rk,x = inf {Ex (u, u)|u : Tx R, u(x) = 1, u(y ) = 0 if |y | |x| + k + 1}) for k 1 and R0,x = 0. Lemma 4.13. (1) limk Rk,x = Rx . (2) For any k 1 and any x T , 1 1 = Rk,x ry + Rk1,y
y S ( x )

Proof. (1) Note that (Tx , Cx ) is transient. Dene F,x = (C0 (Tx ))Ex +1, where Ex (u, v ) = E (u, v ) + u(x)v (x). Write Ix = I(Tx ,Cx ) . Recall that (Ex , F,x ) is a resistance form on Tx {Ix } and that ( )1 Rx = min{Ex (u, u)|u F,x , u(x) = 1, u(Ix ) = 0} . Hence Rx Rk,x . There exists F,x such that (x) = 1 and (Ix ) = 0 and Ex (, ) = 1/Rx . Since (Ix ) = 0, (C0 (Tx ))Ex . Hence there exists {n }n1 C0 (Tx ) such that n (x) = 1 and Ex (n , n ) 1/Rx as n . Choose kn so that supp(n ) {y |y Tx , |y | |x| + kn + 1}. Then Ex (n , n )1 Rkn ,x . Since kn as n and Rk,x Rk+1,x for any k , it follows that Rx limk Rk,x Rx . (2) This follows by applying the formula of combined resistance. Lemma 4.14. Let 1 m n.
g :g L1 (, ),[g ]m =[f ]m

min

En (g ) = Em (f ) +

xWm

Rnm,x ( )2 H (f )(x) (f )x (4.10) (Rx )2

for any f L1 (, ). 23

Proof. We use an inductive argument. If [f ]n = [g ]n , then En (f ) = En (g ) by (4.9). Hence (4.10) holds for n = m. Assume that we have (4.10) for m. Choose u L1 (, ) so that [u]m1 = [f ]m1 . By (4.9) and the induction hypothesis, min En (g ) = Em (u)+ Rnm,y ( )2 H (u)(y ) (u)y (Ry )2

g :g L1 (, ),[g ]m =[u]m

= Em1 (u) +

xWm1 y S (x)

ry + Rnm,y ( )2 H (u)(x) (u)y 2 (y ) ry + Rnm,y ( )2 H (f )(x) (u)y 2 (y )

xWm1 y S (x)

= Em1 (f ) +

xWm1 y S (x)

Now, we consider the minimum of the last expression subject to the constraint [u]m1 = [f ]m1 , i.e. Rx (u)y (f )x = y
y S (x)

for any x Wm1 . The method of Lagrange multipliers along with Lemma 4.13(2) shows ( ) min min E (g )
u:[u]m1 =[f ]m1

= Em1 (f ) + = Em1 (f ) +

g :[g ]m =[u]m

xWm1

)1 ( )2 1 1 ( H (f )(x) (f )x (Rx )2 ry + Rnm,y

xWm1

)2 Rnm+1,y ( H (f )(x) (f )x . 2 (Rx )

Lemma 4.15. For any f L1 (, ), (f )y (f )x (H (f )(x) (f )x ) (H (f )(y ) (f )y ) Rx Ry

= Proof. Note that (f )x =

Rx 2

y,z S (x)

)2 1 ( (f )y (f )z . (4.11) y z

y S (x) (Rx /y )(f )y .

Also by (4.9), if y S (x), then

24

H (f )(y ) (f )y = y (H (f )(x) (f )y ). By those relations, (f )y (f )x (H (f )(x) (f )x ) (H (f )(y ) (f )y ) Rx Ry y S (x) ( ) (f )y 1 = (f )y (H (f )(x) (f )x ) (H (f )(x) (f )y ) y y y S (x) y S (x) ( )2 1 1 = Rx (f )y + ((f )y )2 y y y S ( x ) y S (x) ) (1 Rx 1 2 (( f ) ) (f )y (f )z = R y x y (y )2 y z y S (x) y,z S (x),y =z ) 1( 1 1 =Rx (f )y (f )z ((f )y )2 Rx y z y z
y S (x)

Rx = 2

z S (x),z =y

y,z S (x),y =z

y,z S (x)

)2 1 ( (f )y (f )z . y z

Lemma 4.16. Let f L1 (, ). For any m 0, (f )x (H (f )(x) (f )x ) = Rx

xWm

Rx 2

y,z S (x)

)2 1 ( (f )y (f )z . y z (4.12)

For m = 0, the right-hand side of (4.12) is considered as 0. Proof. Since H (f )() = (f ) , we have (4.12) for m = 0. Suppose that (4.12) holds for m = 0, 1, . . . , n 1. Then Lemma 4.15 suces to show (4.12) for m = n. Lemma 4.17. Let f L1 (, ). Then Em (f )+
xWm

)2 1 ( H (f )(x) (f )x = Rx

1 xm j =0 Wj

Rx 2

y,z S (x)

)2 1 ( (f )y (f )z . y z

Proof. Dene Vm = (m j =1 Wj ) (xWm {Ix }), where Ix = I(Tx ,Cx ) . For u (Vm ), dene Em (u, u) =
xm j =1 Wm

1 )2 1( (u(x) u(Ix ))2 . u(x) u( (x)) + rx Rx

Then (Em , (Vm )) is a resistance form on Vm . Let Lm be the associated Laplacian, i.e. non-positive denite symmetric operator from (Vm ) to (Vm ) which satises E (u, v ) = xVm u(x)(Lm v )(x). Dene F : Vm R by F (x) = 25

H (f )(x) for x m j =0 Wj and F (Ix ) = (f )x for x Wm . By (4.9), for any x Wm , (Lm F )(x) = 1 1 (F (Ix ) F (x)) + (F ( (x)) F (x)) Rx rx 1 1 = ((f )x H (f )(x)) + (H (f )( (x)) H (f )(x)) = 0. Rx rx

In the same manner, it follows that (Lm F )(x) = 0 for any x m j =0 Wm . Therefore, Em (F, F ) =
xWm

F (Ix )(Lm F )(Ix ) =

xWm

(f )x (H (f )(x) (f )x ), Rx

which equals to the left-hand side of (4.12). On the other hand, Em (F, F ) = Em (f ) +
xWm

)2 1 ( H (f )(x) (f )x Rx

by denition. Combining these facts with Lemma 4.16, we immediately deduce the desired equation. Proof of Theorem 4.4. By Lemma 4.14, En (f ) Em (f ) + Rnm,x ( )2 H (f )(x) (f )x (Rx )2

xWm

for any n m 1. Applying Lemma 4.13-(1) and Lemma 4.17, we have E (H (f ), H (f ))

Rx 2

y,z S (x)

)2 1 ( (f )y (f )z Em (f ) y z

for any m 1. Since Em (f ) E (H (f ), H (f )) as m , we immediately verify the theorem.

Intrinsic metric and volume doubling property

In the last section, we have introduced {Dx }xT and shown that Dxn is monotonically decreasing if = (x0 , x1 , . . .) . In this section, we construct an ultra-metric d on where the diameter of x equals to Dx for any x T and provide a simple condition which is equivalent to the volume doubling property of with respect to this ultra-metric d. The ultra-metric d will turn out to be suitable for describing asymptotic behaviors of the Hunt process associated with the regular Dirichlet form (E , F ) on L2 (, ) in the next section.

26

x x In this section, (T, C ) is a transient tree. We assume that (Ty , Cy ) is transient for any x, y T . We x a reference point X and use the same notation as in the previous sections. To avoid nonessential complications, we further assume that #(S (x)) 2 for any x T . Even without this assumption, the statements in this and following sections, except Corollary 6.9, hold with minor modication. This assumption implies the following proposition.

Proposition 5.1. (, O ) has no isolated point and is a Cantor set, i.e. compact, perfect and totally disconnected. To describe the next theorem, we use the notations from Denition 2.8. Recall that F = (C0 (T ))E + R and that g is the symmetrized Green function of (T, C ). Theorem 5.2. (1) g (, x) = Dx for any x T . (2) xWn (x )Dx 0 as n . (3) D[]n 0 as n for -almost every . Notation. We write Dn, = D[]n . We have an example of (T, C ) where limn D[]n > 0 for some in Section 11. Proof. (1) Dene (x) = g (, x). Then is a {, I }-harmonic function with boundary value () = R and (I ) = 0. Let (Em , Fm ) be the trace of (E , F ) on Tm {I }, where Tm = m n=0 Wm and let Lm be the associated discrete Laplacian on Wm {I }. Note that |Tm {I } is also a {, I }-harmonic function with respect to (Em , Fm ). Hence, for x Wm , Lm (x) = Since () = 0, we obtain (x)/ ( (x)) = Dx /D(x) (5.1) ( (x)) (x) () (x) + =0 rx Rx

by (4.2). As () = R = D , (5.1) implies (x) = Dx on Wm for any m 0 inductively. (2) Let (Em , (Tm )) be the resistance form on a nite set Tm associated with (Tm , C |Tm ) and let Hm be the associated (discrete) Laplacian. Then by (4.3) Em (, ) = (x)(Hm )(x) Dy D D(x) Dx Dx ry rx xTm y S () = R Dx (x ). = R
x T m x T m

(5.2)

27

Note that Em (, ) E (, ) = R as m . Hence by (5.2), we have desired result. (3) Dene n : R by n = xWn Dx x . By (4.2), n is positive and monotonically decreasing as n . By (2), n ( ) (d ) = Dx (x ) 0
xTm

as n . This immediately implies that n ( ) 0 as n for -almost every . Since n ( ) = Dn, , we have completed our proof. Now we dene an ultra-metric on the Cantor set by means of Dx . Denition 5.3. Dene { D[, ] d(, ) = 0

if = , if = .

Proposition 5.4. (1) d(, ) is a metric on . Moreover it is an ultra-metric, i.e. max{d(, ), d(, )} d(, ). (5.3) for any , , . (2) max d(, ) = R for any . (3) Dene B (, r) = { |d(, ) < r} for any and r > 0. B (, r) = []n if and only if D[]n < r D[]n1 . (4) The identity from (, d) to (, O ) is continuous. Moreover, (, d) is homeomorphic to (, O ) if and only if D[]n 0 as n for any . Proof. (1) It is enough to show (5.3), which implies the triangle inequality. Other properties of metric are immediate. If T[,] T[, ] , then multiple applications of (4.2) show that d(, ) < d(, ). Hence we have (5.3). Otherwise, d(, ) = d(, ) and (5.3) holds. (2) For any , we may choose so that [, ] = . Then d(, ) = D = R . (3) Fix . Then d(, ) {Dm, |m 0}. Moreover, d(, ) < Dn1, d(, ) Dn, []n . Therefore, B (, r) = []n if and only if Dn, < r Dn1, . (4) Note that {x }xT is a fundamental system of neighborhoods of O . For any x T# , choose x . If r = Dx , then B (, r) = x . Hence x is open with respect to (, d) for any x T . This implies the continuity of the identity from (, d) to (, O ). Assume that Dn, 0 as n for any . Let U be an open set with respect to (, d). For any U , B (, r) U for some r > 0. By the assumption, there exists n such that Dn, < r and hence []n U . Therefore, U is an open set with respect to O . This shows that (, d) and (, O ) are homeomorphic. Finally assume Dn, D > 0 for some . If 0 < r < D, then B (, r) = { }. Therefore { } is an open set with respect to (, d). On the 28

other hand, { } / O by Proposition 5.1. Hence (, d) is not homeomorphic to (, O ). The following theorem gives necessary and sucient condition for to have the volume doubling property with respect to the metric d. Theorem 5.5. has the volume doubling property with respect to d, i.e., there exists c > 0 such that (B (x, 2r)) c (B (x, r)) for any x and any r > 0, if and only if the following two conditions (EL) and (D) hold: (EL) : There exists c1 (0, 1) such that c1 (x )/ ((x) ) for any x T# . (D): For any n 0 and any , D[]n+m D[]n , where m 1 and (0, 1) are independent of n and . By the condition (D), D[]n 0 for any if has the volume doubling property with respect to d. Also the number of neighboring vertices is shown to be uniformly bounded under the volume doubling property as follows. Proposition 5.6. If (EL) hold, then supxT,yS (x) (y )/ (x ) < 1 and supxT #(S (x)) < +. Proof. If #(S (x)) > 1, then c1 zS (x),z=y (z )/ (x ) = 1 (y )/ (x ). Hence (y )/ (x ) 1 c1 . Moreover, c1 #(S (x)) yS (x) (y )/ (x ) = 1. This shows #(S (x)) 1/c1 . The following alternative denition of the volume doubling property is sometimes useful. Proposition 5.7. has the volume doubling property with respect to d if and only if there exist c > 0 and (0, 1) such that (B (, r)) c (B (, r ) for any and any r (0, R ]. Proof of Theorem 5.5. Assume the volume doubling property. Then, (B (, Dn, + )) c (B (, Dn, /2 + /2)) Choose so that Dn, + < Dn1, and Dn, /2 + /2 Dn, . Proposition 5.4(3) shows ([]n ) c (B (, Dn, )) = c ([]n+1 ). If x T and any y S (x), then x = [ ]n and y = [ ]n+1 for some . Hence 1/c (y )/ (x ). Thus we have (EL) . By Proposition 5.6, there exists (0, 1) such that (y )/ (x ) for any x T and any y S (x). Choose m 1 so that m < 1/c. For any , we have (B (, Dn+m, )) = ([]n+m+1 ) m ([]n+1 ) = m (B (, Dn, ). Using the volume doubling property, we obtain 1 (B (, 2Dn+m, )) (B (, Dn+m, )) m (B (, Dn, ). c 29

Hence Dn+m, Dn, /2. Thus we have (D). Conversely, we assume (EL) and (D). If Dn, < r D0, = R , then we may choose k n such that Dk, < r Dk1, . Since Dk+m, Dk, < r, we have ([]k+m ) (B (, r)). On the other hand, by (EL) , ([]k+m ) (c1 )m ([]k ) = (c1 )m (B (, r)). Thus (B (, r) c (B (, r )) (5.4)

for any r > Dn, , where c = (c1 )(m+1) . By Proposition 5.7, we have the volume doubling property of with respect to d.

Asymptotic behaviors of the process

In this section, we study the asymptotic behavior of the heat kernel (transition density), the jump kernel J (, ) and moments of displacement associated with the Dirichlet form (E , F ) on L2 (, ) under the assumption that is volume doubling with respect to d. As in the last section, (T, C ) is a transient tree, x x ) is assumed to be transient for any x, y T . We x a reference point , Cy (Ty . Moreover, we continue to assume that #(S (x)) 2 for any x T . All the statements except Corollary 6.9, however, will hold without this assumption. Making use of Theorem 4.6, we have a formal expression of a heat kernel associated with the Dirichlet form (E , F ) on L2 (, ) as follows: p(t, , ) = 1 + By Lemma 6.1 below,
#(S (x))1

x T

#(S (x))1

ex t

j =1

x,j ( )x,j ( ).

(6.1)

j =1

x,j ( )x,j ( ) =

y S (x)

1 1 y ( )y ( ) ( )x ( ). (y ) (x ) x

Combining this with (6.1), we obtain ( ) 1 1 1 + e[]n t ( ) ( ) [ ] [ ] n +1 n n=0 p(t, , ) = N (, ) ( [, ] t ) 1 n1 e[, ]n t e ([, ]n ) n=0

if = , (6.2) if = ,

where we dene []1 = 0 and write [, ]n = [[, ]]n . If we allow as a value, p(t, , ) is well-dened on (0, ) 2 through (6.2). The value ( may occur ) on the diagonal. Also, by (6.2), p(t, , ) is continuous on (0, ) ( )\ , where is the diagonal.

30

Lemma 6.1. Let V = {1, . . . , n} and let : V (0, +). Dene an inner product (, ) of (V ) by (u, v ) = kV (k )u(k )v (k ) for any u, v (V ). If (1 , . . . , n1 ) is a orthonormal base of L = {u|u (V ), (u, V ) = 0} with respect to (, ) , then
n 1 i=1

i (k )i (m) =

n j (k )j (m) j =1

(j )

j =1 (j )

1 km n , (k ) j =1 (j )

where km is the Kronecker delta. Proof. We use induction in n. To distinguish between dierent ns, we write Vn and Ln, instead of V and L . If n = 2, then ( ) 1 (2) 1 = (1)(2)((1) + (2)) (1) 2 This implies 1 (k )1 (m) = j =1 j (k )j (m)/(j ) ((1) + (2))1 . Suppose the lemma is true for n. Let {1 , . . . , n1 } be a orthonormal base of Ln, . We extend the domain of i to {1, . . . , n, n +1} by setting i (n +1) = 0 for any i = 1, . . . , n 1. Then (1 , . . . , n1 ) is a orthonormal system of Ln+1, as well. Moreover, by the induction hypothesis,
n 1 j =1

j (k )j (m) =

n j (k )j (m) j =1

(j )

1 n

j (k )

n j =1

j ( m)

(6.3)

j =1 j =1

Dene n Ln+1, by { 1 (n + 1) Pn n (k ) = n n+1 j =1 (j ) (n+1) j =1 (j ) j =1 (j )

if k {1, . . . , n}, if k = n + 1.

Then (1 , . . . , n ) is a orthonormal base of Ln+1, . Adding n (k )n (m) to (6.3), we obtain the statement of the lemma for n + 1. In fact, the formal heat kernel p(t, , ) is shown to be a transition density of a Hunt process associated with the regular Dirichlet form (E| , F| ) under a suitable assumption. Proposition 6.2. Assume that limn D[]n = 0 for any , . Dene pt, ( ) = p(t, , ) for any , and any t > 0. Then pt, d = 1 and pt, ps, d = p(t + s, , ) (6.4)

for any , with = and any t, s > 0. In particular, dene (pt u)( ) = pt, ud for any Borel measurable bounded function u : R. Then {pt }t>0 is a Markovian transition function in the sense of [5, Section 1.4]. 31

Proof. If = , then p(t, , ) = e[]n1 t e[]n t []n ( ) ([]n ) (6.5)

n0

Note that this is an innite sum of non-negative functions. Using (6.5), we obtain (6.4) by a routine but careful calculation. The fact that {pt }t>0 is a Markovian transition function is immediate from (6.4). By this proposition, if limn D[]n = 0 for any , then (pt u)( ) = (Tt u)( ) for -a.e. , where {Tt }t>0 is the strongly continuous semigroup on L2 (, ) associated with the Dirichlet form (E , F ) on L2 (, ). Moreover, we have the following theorem. Theorem 6.3. If limn D[]n = 0 for any , then there exists a Hunt process ({Xt }t>0 , {P } ) on whose transition density is p(t, , ), i.e. E (f (Xt )) = p(t, , )f ( ) (d )

for any and any Borel measurable bounded function f : R, where E () is the expectation with respect to P . Note that the Hunt process ({Xt }t>0 , {P } ) is naturally associated with the Dirichlet form (E , F ) on L2 (, ). The essence of the proof of Theorem 6.3 is to show that the transition function {pt }t>0 is a Feller transition function. Namely we are going to prove that pt (C ()) C () and ||pt u u|| 0 as t 0 for any u C (), where C () is the collection of continuous functions on . Lemma 6.4. Let Cm = { xWm ax x |ax R for any x Wm } and let C = m0 Cm . Then ||pt u u|| 0 as t 0 for any u C . Proof. For any u Cm , by using an inductive argument, it follows that u=c+
m k=1 xWk #(S (x))1

i=1

bx,i x,i .

(See Lemma 9.2 for details.) This implies pt u = c +

i=1

ex t bx,i x,i .

Hence there exists c > 0 such that ||pt u u|| c(1 e t ).

32

Proof of Theorem 6.3. Let , and let N = N (, ). Then by (6.5), |p(t, , ) p(t, , )| ( ) e[]n1 t e[]n t e[]n1 t e[xi]n t = []n ( ) + []n ( ) ([]n ) ([]n )
n>N

Since limn D[]n = 0 for any , we have |pt, pt, |d = 2e[,] t

Hence if u is a bounded Borel measurable function on , then |(pt u)( ) (pt u)( )| 2e[,] t ||u|| . Again by the fact that limn D[]n = 0 for any , we see pt u C (). In particular pt (C ()) C (). Let u C () and x > 0. Then there exist m 0 and um Cm such that ||u um || < /3. By Lemma 6.4, ||pt u u|| ||pt u pt um || + ||pt um um || + ||u um || < for suciently small t > 0. Hence ||pt u u|| 0 as t 0. Thus {pt }t>0 is a Feller transition function. Then by [5, Theorem A.2.2], (see also [1, Theorem I.9.4]), we obtain the desired statement. Notation. We write (m, ) = ([]m ) for any and any m 0. Without any further assumptions, p(t, , ) satises the following estimates. Proposition 6.5. (1) For any and any t > 0, p(t, , ) (2) If 0 < t d(, ), then p(t, , ) t . d(, ) ([, ] ) (6.7) 1 1 e (B (, t)) (6.6)

Proof. (1) If Dn, < t Dn1, for some n 1, then t/Dm, 1 for m = 0, 1, . . . , n 1. Hence p(t, , ) 1 +
n 1 m=0

) 1 1 1 e1 . (m + 1, ) (m, ) e (n, )

By Proposition 5.4-(3), it follows that (n, ) = (B (, t)). Hence we have (6.6) for t (0, R ]. For t R , B (, t) = and hence p(t, , ) 1 e1 = e1 / (B (, t)). 33

(2) Write N = N (, ), n = []n , Dn = Dn, and n = ([]n ). Then d(, ) = DN . By letting f (t) = p(t, , ), (6.2) implies f (t) = f (t) = f (t) =
N en1 t en t , n n=0 N n en t n1 en1 t , n n=0 n (n1 )2 en1 t (n )2 en t , n n=0

where 1 = 0. Since n1 t n t DN /Dn 1, we see that f (t) 0 and f (t) 0 for any t [0, DN ]. Hence f (DN )t f (t) f (0)t for any t [0, DN ]. (6.8) along with f (0) =
N 1 n=0

(6.8)

(1 1 ) N 1 1 + = n n+1 N DN N d(, ) ([, ] )

shows (6.7) for any t (0, d(, )]. The volume doubling property of leads us to two-sided estimates of p(t, , ) and J (, ). Note that under the volume doubling property of , we have limn D[]n = 0 for any by Theorem 5.5. Theorem 6.6. Suppose has the volume doubling property with respect to d. (1) p(t, , ) is continuous on (0, ) . Dene t if 0 < t d(, ), d(, ) ([, ] ) q (t, , ) = 1 if t > d(, ). (B (, t)) Then p(t, , ) q (t, , ) on (0, ) . (2) For any (, ) ( )\, J (, ) 1 . d(, ) ([, ] ) (6.10) (6.9)

The proof of this theorem is given at the end of this section. The heat kernel estimate (6.9) can be thought of as a generalized version of the counterpart in [3], where is supposed to satisfy the uniform volume doubling property: i.e. (B (x, r)) f (r) for any r > 0 and f (r) has the doubling property. In this paper, however, we do not require the uniform volume doubling property. The following proposition gives an alternative expression for q (t, , ). 34

Proposition 6.7. For any t > 0 and any , , { q (t, , ) = min } t 1 , . d(, ) ([, ] ) (B (, t))

The above proposition is immediate from the following lemma. Lemma 6.8. t (B (, t)) d(, ) ([, ] ) if and only if t d(, ). Proof. Assume 0 < t D = R . Then Dn, < t Dn1, for some n 1. By Proposition 5.4-(3), B (, t) = []n . If t d(, ), then N (, ) n 1 and hence t ([]n ) d(, ) ([, ] ). Otherwise, t > d(, ) implies N (, ) n. Therefore, t ([]n ) > d(, ) ([, ] ). Next assume t > D . Then B (, t) = and t > d(, ). Hence t ([]n ) > d(, ) ([, ] ). Next we have estimate of moments of the displacement. In the next corollary, the assumption that #(S (x)) 2 for any x T is essential. Corollary 6.9. Suppose that has the volume to d. Then t E (d(, Xt ) ) t(| log t| + 1) t for any and any t (0, 1]. Note that we have an extra log-term when = 1. This is due to the slow polynomial decay of the o-diagonal part of p(t, , ) with respect to the space variable . See the discussion after Theorem 14.1 for details. Corollary 6.9 is a special case of Theorem 14.1, which shows that the above behavior of moments of displacement occurs whenever p(t, , ) enjoys certain type of heat kernel estimate which is typical in certain class of jump processes. Proof. Let us verify all the assumptions of Theorem 14.1 with X = , = and (r) = r. Theorem 6.3 suces to show the assumption (1) in Section 14. By Proposition 5.6, there exist c1 , c2 , c3 (0, 1) and R > 0 such that c1 (B (, r)) (B (, c3 r) c2 (B (, r)) for any and any r (0, R]. Hence we have (14.1). Theorem 6.6 implies (14.2). Therefore, we may apply Theorem 14.1 and obtain the above corollary. The rest of this section is devoted to showing Theorem 6.6. Proof of Theorem 6.6-(1). Combining (6.1) and Lemma 6.1, we have p(t, , ) = 1+
n0

doubling property with respect if > 1, if = 1 if 0 < < 1.

(6.11)

e[]n t

1 ([]n+1 )

[]n+1 ( )

) 1 []n ( ) . (6.12) ([]n )

Note that every term in the sum is continuous on (0, ) . 35

On the other hand, by Theorem 5.5, the volume doubling property implies that there exist (0, 1) and c > 0 such that ([]n ) n and Dn, cn

for any and any n 0. Hence the absolute value of each term in (6.12) 1 n is no greater than c n ec T on [T, ) . Since the innite sum of these terms is convergent, the innite sum in (6.12) is uniformly convergent on [T, ) by the Weierstrass M-test. Therefore, p(t, , ) is continuous on (0, ) . Proof of Theorem 6.6-(2). We use the notations in the proof of Proposition 6.5. By Theorem 5.5, there exist (0, 1) and M N such that m n n+m Dn+M Dn for any n and m. Choosing suitable (0, 1) and c > 0, we also have Dn+m cm Dn (D) (EL) (D)

for any n and m. Note that , c and M are independent of . The upper estimate for 0 < t d(, ) is immediate by Proposition 6.5-(2). Lower estimate for 0 < t d(, ): Assume that M N + 1. By (EL) , f (DN )
N n=N M +1 N n=N M +1

n en DN n1 en1 DN n N n n en DN n1 en1 DN N (6.13)

) DN M 1 ( 1 DN D N M . e N DN e DN M

By (D), we see DN /DN M < 1. Hence f (DN ) M 1 (e1 e ) C . N DN N DN (6.14)

Next we consider the case where N + 1 M . Since Dm+1, < Dm, for any and any m 0, there exists (0, 1) such that Dm+1, Dm, for any and any m M . Note that N 1 M and hence DN DN 1 . Therefore f (DN ) N e1 N 1 eN 1 DN N ) e1 e N 1 (1 DN DD e N 1 . N DN e DN 1 N DN 36

Using this and (6.14), we have the desired lower estimate for 0 < t d(, ) from (6.8). Upper estimate for d(, ) < t: By (6.2), p(t, , ) p(t, , ). Hence it is enough to show that p(t, , ) c/ (B (, t)). Suppose Dm < t Dm+1 . Then 1+
m 1 ( n=0

1 n+1

m 1 ( 1 1 ) n t 1) 1 1 e 1+ = = (6.15) n n+1 n m (B (, t)) n=0

Let n m. Then by (D), t/Dn Dm /Dn c1 mn . Also by (EL) , 1/n+1 1/n 1/n+1 (n+1m) /m . These imply
( 1 ) n t k c1 k C 1 e e . ( B ( , t)) n +1 n m n =m k=0

(6.16)

Since C is independent of and m, (6.15) and (6.16) yield the upper estimate for d(, ) < t D0 . If t > D0 , then (B (, t)) = 1. Since p(t, , ) = 1 +
( 1 n 1 1 ) n t e 1+ (n+1) ec , n+1 n n=0 n=0

we have the upper estimate in this case. Lower estimate for d(, ) < t: If D0 < t, then B (, t) = and p(t, , ) 1 1 e1 (1 e0 t ) 1 e1 = . 0 (B (, t))

Assume d(, ) < t D0 . Then Dm < t Dm1 for some m {1, 2, . . . , N }. Suppose M N . Then we may choose k {0, . . . , m 1} so that m k + M N . By (EL) , it follows that m M n for any n = k, k + 1, . . . , k + M . Hence p(t, , )
k +M n=k+1

en1 t en t ek t ek+M t eb ea M M , n m m

where a = t/Dk+M and b = t/Dk . Using (D), we see that b 1 a and b a. Since min{eb ea |b 1 a, b a} > 0, it follows that p(t, , ) C / (B (, t)), where C is independent of and t. Finally assume N + 1 M . As in the proof of the lower estimate for 0 < t d(, ), we have Dm Dm1 , where (0, 1) is independent of and t. Hence em1 t em t . p(t, , ) m Since t/Dm1 1 t/Dm and t/Dm1 t/Dm , the similar argument as above shows p(t, , ) C / (B (, t)).

37

Proof of Theorem 6.6-(3). We continue to use the notations in the proof of Proposition 6.5. By (4.5), J (, ) =
N 1 1( m+1 m ) 0 + , 2 m+1 m=0

where N = |x|. Note that (EL) , (D) and (D) in the proof of Theorem 6.6-(2) hold under the volume doubling property of . First we show the lower estimate. Assume that N M + 1, where M is the constant appearing in (D). By (EL) , N i M N for 0 i M . Hence, by (D), J (, ) M 2
N 1 m=N M

m+1 m M N N M N 2 N

M (1 )N M (1 ) . = 2N 2d(, ) ([, ] ) In case N M + 1, we may apply the same discussion as in the lower estimate for d(, ) < t in the proof of Theorem 6.6-(2) and obtain the lower estimate in this case. Next by (D), it follows that N i ci N . This implies
N 1 m c(1 + + . . . + N )N c J (, ) . 2 m=0 N 2N 2(1 )d(, )([, ] )

Thus we have shown the upper estimate.

Completion with respect to resistance metric

As in the last section, (T, C ) is a transient tree, (E , F ) and R are the associated resistance form and resistance metric on T respectively. Let T be the completion of T with respect to the resistance metric. Then by [9, Theorem 2.3.10], (E , F ) can be thought of as a resistance form on T . Write TR = T \T . Then by Theorem 13.3, we have a resistance form (E|TR , F|TR ) on TR which is the trace of the resistance from (E , F ) on TR . The natural question is if TR = and (E , F ) = (E|TR , F|TR ) or not. In particular, if this is true, then (E , F ) is a resistance form on . In the followings, we establish a sucient condition in Theorem 7.1 and a necessary condition in Theorem 7.2 for being TR = and (E|TR , F|TR ) = (E , F ). Using these theorems, we will show that both TR = and TR = can occur by examples in the next section. Theorem 7.1. Let rx = C (x, (x))1 . If m1 (maxxWm rx ) < +, where Wm = {x|x T, |x| = m}, then there exists a homeomorphism between T and T which is the identity on T . Moreover, (E , F ) coincides with (E| , F| ) which is the trace of (E , F ) on , where is identied with T \T through the homeomorphism. In particular, (E , F ) is a resistance form on . 38

Proof of Theorem 7.1. Dene rm = maxxWm rx . Let . For m > n, R([ ]n , [ ]m ) =

r[]i

m i=n+1

ri

Therefore, {[ ]n }n0 is an R-Cauchy sequence. Denote the R-limit of {[ ]n }n0 by ( ) T . Since R([ ]n , [ ]n ) r[]N (, )+1 for n > N (, ), : T is injective. Similar argument shows that () T \T . Let {xn }n0 be an R-Cauchy sequence and let x be its R-limit in T . Assume x / T . Then there exists a subsequence {yn }n0 of {xn }n0 such that |yn | as n . For any m 0, there exists zm Wm such that yn Tzm for suciently large n. Note that = (z0 , z1 , . . .) is a geodesic ray and that R(, z ) i=m+1 ri for any z Tzm . Those facts imply x = . Hence is bijective. Now, extend : T T by (x) = x for x T . It is routine to see that is continuous. Thus is a homeomorphism between T and T and |T is identity. Hereafter, we identify with T \T . Now, let f F . Then H (f ) F . Note that H (f ) is extended to a continuous function on T = T . Hence Theorem 3.6 shows that H (f )(x) = H (u)(x) for any x T , where u = H (f )| . Making use of Corollary 4.9, we have (u)x = (f )x for any x T . This implies u = f for -almost everywhere. Hence f F| = {h| |h F}. Let f F| . Let h(f ) be the -harmonic function with boundary value f , i.e. h(f ) F attains the following minimum: min{E (v, v )|v F , v | = f }. Since h(f ) is harmonic on T and is bounded, Theorem 3.6 shows that h(f )(x) = H (u)(x) for some u L (, ) and h(f )([ ]n ) u(x) for -almost every . Note that h(f ) is continuous on T and h(f )| = f . It follows that u = f and hence h(f ) = H (f ). Therefore, f F and hence F| = F . Finally E| (f, f ) = E (h(f ), h(f )) = E (H (f ), H (f )) = E (f, f ) for any f F = F| . Thus (E| , F| ) is identied with (E , F ). Next we state a necessary condition. Theorem 7.2. If (E , F ) is a resistance form on and the associated resistance metric gives the same topology as O , then R[]n 0 as n for any . Proof of Theorem 7.2. Using [8, Theorem 9.4], we see that p(t, , ) is jointly continuous and hence it is bounded. (Since p(t, , ) is continuous except the diagonal , the heat kernel given in [8, Theorem 9.4] is equal to p(t, , ).) By [8, Corollary 9.6], tp(t, , ) 0 as t 0 for any . Using Proposition 6.5-(1), we have t/ (B (, t)) 0 as t 0. If Dn, < t Dn1, , then B (, t) = []n . Hence t/ (B (, t)) Dn, / ([]n ) = R[]n . Since n as t 0, we have R[]n 0 as n .

39

Example: binary tree

In this section, we are going to illustrate the results obtained in the previous sections by examples which are (innite complete) binary trees. Our examples are divided into two classes. The rst class is the collection of self-similar trees, where the volume doubling property is automatic under the assumption of transience. The other class is homogeneous trees, through which we will explore various phenomena when the volume doubling property fails. Denition 8.1. Let Wm = {1, 2}m for m 0, where W0 = {}. Dene W = m0 Wm . We denote (w1 , . . . , wm ) as w1 . . . wm . For w1 . . . wm W \W0 , dene (w1 . . . wm ) = w1 . . . wm1 and S (w1 . . . wm ) = {w1 . . . wm 1, w1 . . . wm 2}. Assume that C : W W [0, ) satises C (w, v ) = C (v, w) and C (w, v ) > 0 if and only if (w) = v or (v ) = w. (W , C ) is called the (innite complete) binary tree. For binary trees, we always choose as the reference point. Then, the notions (w), S (w) and Wm are consistent with those dened in Sections 3 and 4. For any (W , C ), the collection of innite geodesic rays originated from , , is identied with the Cantor set {1, 2}N . As a standard metric on , we introduce d (, ). Denition 8.2. Dene d (, ) = 2N (, ) for any = and d (, ) = 0 if = . It is easy to see that d (, ) is a distance on . First we consider a kind of self-similar binary tree (W , CS ). Denition 8.3. Let r1 , r2 > 0. For w W , dene CS (w, wi) = (rw ri )1 , where rw = rw1 rwm for any w = w1 . . . wm W . Theorem 8.4. (W , CS ) is transient if and only if r1 r2 /(r1 + r2 ) < 1. In particular, if r1 = r2 = r, then (W , CS ) is transient if and only if 0 < r < 2. Proof. Let ({Xn }n0 , {Pw }wW ) be the random walk associated with (W , CS ). Then P (|Xn+1 | = |Xn | + 1||Xn | 1) = (r1 + r2 )/(r1 + r2 + r1 r2 ) = p1 and P (|Xn+1 | = |Xn | 1||Xn | 1) = r1 r2 /(r1 + r2 + r1 r2 ) = p2 . Therefore, we may associate a random walk ({Zn }n0 , {Qk }kN ) on N = {0, 1, . . .} such that Qk (Zn+1 = Zn + 1|Zn 1) = p1 and Qk (Zn+1 = Zn 1|Zn 1) = p2 if Zn 1 and Qk (Zn+1 = 1|Zn = 0) = 1. This random walk is transient if and only if p1 > p2 r1 r2 /(r1 + r2 ) < 1. By applying the results in the previous sections, we obtain the following statements. Lemma 8.5. Assume that (W , CS ) is transient. )1 (r + r 1 2 1 . (1) R = r1 r2 (2) Rw = rw R for any w W . 40

(3) Let 1 = r2 /(r1 + r2 ) and 2 = r1 /(r1 + r2 ). Then (w ) = w1 wm for any w = w1 . . . wm W . (4) Dw = (w )Rw = (r1 r2 /(r1 + r2 ))|w| R for any w W . (5) has the volume doubling property with respect to d(, ), where d(, ) has been given in Denition 5.3. Proof. (1) Considering the self-similarity, R satises 1 1 1 + = . r1 (R + 1) r2 (R + 1) R (2) (3) (4) (5) This is obvious from the self-similarity. Use Theorem 3.8. Combine (2) and (3). One can easily verify the conditions (EL) and (D) in Theorem 5.5.

The above lemma shows that d(, ) = D[, ] = (r1 r2 /(r1 + r2 ))N (, ) R . Hence we have d(, ) = d (, ) R for any , , where (r + r ) 1 2 = log / log 2. r1 r2 Combining those with Theorem 6.6, we have the following heat kernel estimate. Theorem 8.6. Assume that (W , CS ) is transient. Let p(t, , ) be the associated heat kernel. Dene t if 0 < t d (, ) , d (, ) ([, ] ) q (t, , ) = 1 if t > d (, ) . (B (, t1/ )) Then p(t, , ) q (t, , ) (8.1) on (0, ) , where B (, r) = { |d (, ) < r}. In particular, if r1 = r2 , then = 1 log r/ log 2 and we may replace q in (8.1) by t if 0 < t d (, ) , q (t, , ) = d (, )+1 t1/ if t > d (, ) . Theorem 8.7. (E , F ) is a resistance form on and the associated resistance metric gives the same topology as O if and only if 0 < r1 < 1 and 0 < r2 < 1. Proof. Assume that 0 < r1 < 1 and 0 < r2 < 1. Let r = max{r1 , r2 }. Then 0 < r < 1 and maxxWm rx rm . By Theorem 7.1, (E , F ) is a resistance form on . Conversely say r1 1. Let = 1111 . . .. Then R[]n = (r1 )n R . Now Theorem 7.2 yields the desired conclusion. 41

Next we study a class of homogeneous binary trees. Denition 8.8. Let r(i) > 0 for any i 0. Dene CH (w, wi) = r(|w|)1 for any w W and any i {1, 2}. Theorem 8.9. (W , CH ) is transient if and only if n0 2(n+1) r(n) < +. Proof. Let us reduce Wm to an single point m N {0}. Then the resulting weighted {0}, C ), where C (m, m + 1) = 2m+1 /r(m). Hence R = graph is ((N n+1) R0 = n0 r(n)/2 . Since the transience is equivalent to the condition that R0 < +, we have the claim of the theorem. By the homogeneity of the tree, we can easily see that (x ) = (y ) for any x, y Wm . Also the same method as in the proof of the above theorem gives Rw . As a consequence, we have the followings. Lemma 8.10. Assume that (W , CH ) is transient. Then for any w W , (w ) = 2|w| , Rw = r(|w| + n) 2n+1 and Dw = r(n) . 2n+1 (8.2)

n0

n|w|

Hereafter, we always assume that (W , CH ) is transient. By (8.2), Dw only depends on |w|. For n 0, we dene Dn = Dw and n = 1/Dn for w Wn . Note that {Dn }n0 is strictly decreasing and limn Dn = 0. Conversely, given a strictly decreasing sequence {Dn }n0 with limn Dn = 0, we may construct an associated {r(n)}n0 by letting r(n) = 2n+1 (Dn Dn+1 ). By (6.2), we have p(t, , ) = 1 + and for = ,
N (, ) n=0

2n en t

(8.3)

p(t, , ) =

2n (en1 t en t ) (8.4) 2 e
n n t

n=0 N (, )1

=1+

N (, ) N (, ) t

n=0

From these, if p(T, , ) < + for some T > 0, then p(t, , ) is continuous on [T, ) . Choosing an appropriate decreasing sequence {Dn }n0 , we may obtain a variety of heat kernels with interesting behaviors. For instance, we have an example where pt, () = p(t, , ) is getting more and more regular as t log 2 as follows. Example 8.11. Let Dn = (n + 1)1 for n 0. Then 2N +1 e(N +1)t 1 if t = log 2, (1 et ) p(t, , ) = 2et 1 (N + 1)(1 et ) if t = log 2, 42

(8.5)

where N = N (, ). Dene pt, by pt, ( ) = p(t, , ). Then by (8.2) and (8.5), for q 1, pt, Lq (, ) if and only if (1 1/q ) log 2 < t. (We regard 1/q as 0 if q = .) In other words, if 0 < t log 2, then pt, Lq (, ) for 1 q < log 2/(log 2 t) and p(t, , ) is nite and continuous on (log 2, ) . To describe the diagonal part p(t, , ), we introduce the eigenvalue counting function N () of the non-negative denite self-adjoint operator associated with the Dirichlet from (E , F ) on L2 (, ). The eigenvalue counting function N () is dened as the number of eigenvalues of L which are no greater than . By Theorem 4.6, n is an eigenvalue with multiplicity 2n and 0 is an eigenvalue with multiplicity 1. Hence, N () = 1 + 2n = 2F () ,
n:n

where F is dened by F () = n if and only if n1 < n . (Recall that 1 = 0.) (8.3) yields the next proposition. Proposition 8.12. For the homogeneous (W , CH ), (s) st p(t, , ) = t e N (s)ds = es N ds t 0 0 Furthermore, if f : [0, ) [0, ) is a monotonically non-decreasing function and f (n ) = n + 1 for any n 1, then for any t > 0, 1 s f (s/t) e 2 ds p(t, , ) es 2f (s/t) ds. 2 0 0 By using the above proposition, an asymptotic behavior of p(t, , ) as t 0 may be determined even if the volume doubling property fails as in the next example. Example 8.13. Let Dn = (n + 1)2 for n 0. Then n = (n + 1)2 . In this case, does not have the volume doubling property with respect to d. Proposition 8.12 implies that p(t, , ) es 2 s/t ds.
0

on (0, ) . Now, for any c > 0, ( c2 ) 2 c c s/ts e ds = 1 + exp ey dy. c 4t t 0 2 t ( (log 2)2 ) 1 p(t, , ) 1 + exp 4t t on (0, ) . Since (B (, t))1 2 1/t , the degree of divergence of p(t, , ) as t 0 is actually much higher than what is expected by the formula which holds with the volume doubling property. 43 Hence

Generalization and jump kernel

In this section, as a generalization of (E , F ), we study a class of Dirichlet forms and/or closed forms on the space of innite geodesic rays of a tree as a non-directed graph. Let T be a countably innite set. To give an a priori structure of a tree as a non-directed graph, we x an equivalence class with respect to . This is equivalent to choose T and : T T such that () = and, for any x T \{}, (n) (x) = for some n 1, where (n) = . . . . All the notions
n-times G

associated with the structure of a tree as a non-directed graph can be derived from and , for example, T# = T \{}, S (x) = {y | (y ) = x}, Tx = {y |y T, (n) (y ) = x for some n 0}, = {(xi )i0 |x0 = , (xi+1 ) = xi for any i 0}, x = {(xi )i0 |(xi )i0 , xm = x for some m 0}. To avoid unnecessary technical complexity, we assume that 2 #(S (x)) < + for all x T . The space is equipped with the canonical topology O which generated by the basis of open sets {x |x T }. Note that (, O ) is compact. Denition 9.1. (1) Let : T [0, ) and let MP (), where MP () = {| is a Borel regular probability measure on (x ) > 0 for any x T }. Then for = (, ), we dene { D = u u L2 (, ), (x) 2(x )
y,z S (x)

} ( (y )(z ) (u),y (u),z )2 < + ,

x T

where (u),x Q (u, v ) =

= (x )1 x ud, and (x) 2(x )

( (y )(z ) (u),y (u),z )((v ),y (v ),z )

xT

for any u, v D . (2) Dene Ex, { } = f f= ay y , f (y )(dy ) = 0 .

Comparing with (4.4), it is apparent that Q is a generalization of E . We are going to show that (Q , D ) has properties which are analogous to (E , F ). 44

Lemma 9.2. Let 0 ( ) = 1 for any . For any x T , choose an L2 (, )orthonormal base (x,1 , . . . , x,M (x) ) of Ex, , where M (x) = #(S (x)) 1. Then {0 , x,n |x T, 1 n M (x)} is a complete orthonormal system of L2 (, ). M (x) Proof. Let Cm = {a0 0 + xTm n=1 ax,n x,n |a0 , ax,n R} and let C = k m0 Cm . Then C = { i=1 k xk |k = 0, 1, . . . , x1 , . . . , xk T, 1 , . . . , k R}. Hence C is dense in C () with respect to the supremum norm. This implies that C is dense in L2 (, ). Since {0 , x,n |x T, 1 n M (x)} is orthonormal, it is a complete orthonormal system. Theorem 9.3. Let : T [0, ) and let MP (). Let = (, ). Then
(x) { M ax,n n,x , D = u u L2 (, ), u = a0 0 + xT n=1 (x) M xT n=1

} (x)(ax,n )2 < + . (9.1) ax,n x,n and v = b0 0 +

Moreover, if u, v D , u = a0 0 + M (x) x T n=1 bx,n x,n , then Q (u, v ) =

x T

M (x)
n=1

x T

M (x)

(x)

ax,n bx,n .

(9.2)

n=1

In particular, (Q , D ) is a closed quadratic form on L2 (, ) and L u = x u for any x T and any u Ex, , where L is the nonnegative self-adjoint operator on L2 (, ) associated with (Q , D ). We can also obtain a (formal) expression on the integral kernel p(t, x, y ) of the semigroup etL by the same formula as (6.2). Proof. Let Tm = m n=0 Wn and let Qm (u, v ) =
xTm

(x) 2(x )

y,z S (x)

(y )(z )((u)y, (u)z, )((v )y, (v )z, )

for any u, v L2 (, ). If u = a0 0 + M (x) x T n=1 bx,n x,n , then Qm (u, v ) =

M (x)
n=1

ax,n x,n and v = b0 0 +

M (x)

(x)

ax,n bx,n .

n=1

Since u D if and only if Qm (u, u) is convergent as m , we have (9.1). The rest follows immediately. 45

To obtain an alternative expression of Q by using an integral kernel, we dene a transformation and . Denition 9.4. For any x T , dene [x]n = xn for n = 1, 2, . . . , M , where M = |x| and (x0 , x1 , . . . , xM 1 , xM ) is the geodesic between and x. (Hence x0 = and xM = x.) Let be a Borel regular probability measure on . Dene a linear map : (T ) (T ) and : (T ) (T ) by ) ( |x|1 ([x]m+1 ) ([x]m ) 1 ( ())(x) = ([x]0 ) + 2 ([x]m+1 ) m=0 for any (T ) and any x T , and
|x|1

(9.3)

( (J ))(x) = 2J (x)(x ) + 2 for any J (T ) and any x T .

( ) J ([x]m ) ([x]m ) ([x]m+1 )

(9.4)

m=0

Simple calculation shows that is the inverse of . Lemma 9.5. = = Identity on (T ). Denition 9.6. For J : T [0, ), dene LJ : ( )\ [0, ) by LJ (, ) = J ([, ]) for any , with = , { } DJ, = u u L2 (, ), LJ (, )(u( ) u( ))2 (d )(d ) < +
2

and, for any u, v DJ, , QJ, (u, v ) = LJ (, )(u( ) u( ))(v ( ) v ( ))(d )(d ).
2

Theorem 9.7. Let : T [0, ) and let MP (). Write = (, ). Then (Q , D ) is a regular Dirichlet from on L2 (, ) if and only if ( ())(x) 0 for any x T . Moreover, assume that ( ())(x) 0 for any x T . Then D = D (), and Q (u, v ) = Q (), (u, v ) for any u, v D . Remark. Dene + (T ) = {u|u : T [0, )}. Then ()(x) 0 for any x T if and only if (+ (T )). Before proving Theorem 9.7, we state an immediate corollary which follows by (9.4) and Lemma 9.5. Corollary 9.8. Let J : T [0, ) and let Mp (). Then (QJ, , DJ, ) is a regular Dirichlet from on L2 (, ). Moreover, (QJ, , DJ, ) = (Q , D ), where = ( (J ), ). Next two lemmas are needed to prove Theorem 9.7. 46

Lemma 9.9. Let J : T [0, ), let MP () and let x T . Then for any Ex, and any u L2 (, ), LJ (, )(( ) ( ))(u( ) u( )) is -integrable on and LJ (, )(( ) ( ))(u( ) u( ))(d )(d ) = x (, u) , (9.5)

Proof. Dene Yx = y,zS (x),y=z y z . Then LJ = J (x)Yx . Let x T Ku,v (, ) = LJ (, )(u( ) u( ))(v ( ) v ( ) and let = yS (x) ay y . Since ( )(d ) = 0, it follows that y S (x) ay (y ) = 0. We devide {(, )|( ) = ( )} into three regions Yx , x (\x ) and (\x ) x . For the rst part, since LJ (, ) = J (x) on Yx , K,u (, ) is integrable on Yx and K,u (, ) Yx ( ) = (ay az ) (z ) u( )(d ) (y ) u( )(d )
y,z S (x) y z

where x = ( (J ))(x) and (, u) =

( )u( )(d ).

= 2(x )J (x)(, u) . For the second region, let Ux,m = [x]m \[x]m+1 . Then \x = m=0 Ux,m . Note that LJ (, ) = J ([x]m ) on x Ux,m . Hence K,u (, ) is integrable on each x Ux,m . Now K,u (, )(d )(d ) = J ([x]m )( )u( )(d )( )
x Ux,m x Ux,m |x|1

= J ([x]m )(([x]m ) ([x]m+1 )(, u) . Hence

K,u (, )(d )(d ) =

J ([x]m )(([x]m ) ([x]m+1 ))(, u) .

m=0

The third part is the same as the second part. As a whole, we obtain (9.5). Lemma 9.10. Let J : T [0, ) and let MP (). Then (QJ, , DJ, ) is a regular Dirichlet form on L2 (, ). Moreover,
(x) M (x) { } M ax,n x,n (a0 )2 + (1 + x ) DJ, = a0 0 + (ax,n )2 < + , xT n=1 xT n=1

(9.6) where x = ( (J ))(x), and QJ, (u, v ) =

ax,n bx,n

(9.7)

n=1

47

for any u = a0 0 + DJ, .

xT

M (x)
n=1

ax,n x,n , v = b0 0 +

x T

M (x)
n=1

bx,n x,n

Proof. By [5, Example 1.2.4], it follows that (QJ, , DJ, ) is a Dirichlet form. Note that {0 , x,n |x T, n = 1, . . . , M (x)} is a complete orthonormal system of L2 (, ). Moreover, Lemma 9.9 implies that x,n DJ, and QJ, (x,n , u) = x (x,n , u) for any u DJ, . These facts yield (9.6) and (9.7). Let C be the same as in the proof of Lemma 9.2. Then C is dense in C () with respect to the supremum norm and in L2 (, ) as well. By (9.6) and (9.7), it follows that the (QJ, )1 -closure of C is DJ, . Hence C is a core and (QJ, , DJ, ) is regular. Proof of Theorem 9.7. Let J (x) = ( ())(x) for any x T . First assume that J (x) 0 for any x T . Then, by Theorem 9.3 and Lemma 9.10, we have all the desired statements. Conversely, assume (Q , D ) is a regular Dirichlet form on L2 (, ). In particular, it has the Markov property. If u = yWm+1 ay y , then Q (u, u) = (x) (y )(z )((u)y, (u)z, )2 (x ) x T m y,z S (x) = J (x) (u( ) u( ))2 (d )(d )
x T m Yx

(9.8)

For any x T , choose y = z S (x). Dene u, = y + z . Then by (9.8), Q (u, , u, ) = +

( ) 2J ([y ]n ) ([y]n ) ([y]n+1 ) 2 (9.9)

( ) 2J ([z ]n ) ([z]n ) ([z]n+1 ) 2 4J (x)(y )(z )

Since Q is non-negative, (9.9) can be summarized as Q (u, , u, ) = A2 + B + C 2 , where A, B and C are constants, A, C > 0 and B = 4J (x)(y )(z ). Applying the unit contraction to u1, for < 0, we have Q (u1, , u1,b ) Q (u1,0 , u1,0 ). This implies that B + C 2 0 for any < 0. Therefore, B 0 and so J (x) 0. Recall that if (x )xT is the eigenvalue map associated with a transient tree (T, C ), then []n is monotonically increasing for any . Next we consider such an class of : T (0, ).

48

Lemma 9.11. Dene (T ) = {u|u : T (0, ), u(x) > u( (x)) for any x T# }. If MP (), then ( (T )) = (T ) and ( (T )) = (T ). Proof. Let (T ) and let J = (). Then (9.3) and (9.4) imply (x) ( (x)) = 2((x) )(J (x) J ( (x))) (9.10)

for any x T# and () = 2( )J (). This equality suces for the proof. The above lemma and Theorem 9.3 imply the following corollary. Corollary 9.12. Let MP (). If (T ), then (Q , D ) is a regular Dirichlet form on L2 (, ), where = (, ). Moreover, D = DJ, and Q = QJ, , where J = (). Finally we give a proof of Theorem 4.6. Proof of Theorem 4.6. Lemma 4.3 shows that (x )xT (T ). Hence by the above corollary, we have the statements (1) and (2) of the theorem. The statement (3) is immediate from (9.2).

10

Inverse problem

As in the last section, T is assumed to have a structure of a tree as a nondirected graph by specifying T and : T T . In other words, we have xed an equivalence class G with respect to . Let T R(G) be the collection of transient trees whose structure as a nondirect graph are G, i.e. T R(G) = {(T, C )|(T, C ) is a tree, G(T, C ) = G
x x and (Ty , Cy ) is transient for any x, y T }. G

Then by the results of the previous sections, we have a map from T R(G) to (T ) MP (T ) dened by (T, C ) ((x )xT , ), where (x )xT is the eigenvalue map of (T, C ) and is the hitting distribution starting from associated with (T, C ). Let us call this map , which is actually injective because (x )xT and give the values of (rx )xT# by (4.3). Note that Dx = 1/x . In this section, we are interested in the inverse of : T R(G) (T ) MP (). In particular, we try to understand the image of . Dene
, (T ) = {| (T ), lim ([ ]n ) = + for -a.e. } n

and
(T ) = {| (T ), lim ([ ]n ) = + for all }. n

49

Then by Theorem 5.2, (T R(G))

) , (T ) {} .

(10.1)

On the other hand, the next theorem implies that (T ) MP () (T R(G)). (10.2)

In the next section, we have examples which show that the equality does not holds any of (10.1) and (10.2). Theorem 10.1. For any (, ) (T ) MP (), there exists (T, C ) T R(G) such that the hitting distribution starting from is and the eigenvalue map is ((x))xT , where (rx )xT# is given by rx = 1 ( 1 1 ) (x ) ( (x)) (x) (10.3)

for any x T# for any y T . In particular, (10.2) holds. Remark. By (4.3), the value given by (10.3) is the only possible (rx )xT# under which is the eigenvalue map and is the hitting distribution starting from . The rest of this section is devoted to a proof of Theorem 10.1. First we prove the following key lemma. Lemma 10.2. Let (T, C ) be a tree. Set rx = C ( (x), x)1 for any x T# . Let (E , F ) be the corresponding resistance form on T . Assume that u : T R and that there exists a Borel regular probability measure on such that u( (x)) u(x) = (x ) rx for any x T# . Then (1) u F if and only if xWm u(x)(x ) is convergent as n . (2) Assume u F . Then for any v F , E (u, v ) = v () lim v (x)(x ).
m xWm

(10.4)

(10.5)

In particular E (u, v ) = v () for any v (C0 (T ))E and (rx )xT# is transient. (4) Assume u F . Let g (x, y ) be the symmetrized Green function of (T, C ). Also let be the hitting distribution starting from . If u() = E (u, u) and xWm u(x) (x ) 0 as n , then u(x) = g (, x) for any x T . Moreover, = . Proof. Let L be the (discrete) Laplacian associated with (E , F ), i.e. (Lf )(x) = f (y ) f (x) f ( (x)) f (x) + . rx ry
y S (x)

50

for any f and any x T . Then by (10.4), for any x T# , (Lu)(x) = u(y ) u(x) u( (x)) u(x) + rx ry y S ( x ) = (x ) (y ) = 0, u(y ) u() = 1. ry
y S(x)

(10.6)

(Lu)() =

y S ()

Dene Em (f, g ) = xTm (f ( (x))f (x))(g ( (x))g (x))/rx . Note that Em (f, f ) E (f, f ) as m for any f : T R and Em (f, g ) E (f, g ) as m for any f, g F . By (10.6), Em (u, v ) =
xTm1

v (x)(Lu)(x) v (x)(x ).

xWm

v (x)

u( (x)) u(x) rx (10.7)

= v () (1) If
xWm

xWm

u(x)(x ) is convergent as m , then (10.7) shows

lim Em (u, u) = u() lim u(x)(x ).

Hence u F . Conversely, if u F , then Em (u, u) is convergent as m . Considering (10.7), we see that xWm u(x)(x ) is convergent as m . (2) Taking m in (10.7), we obtain (10.5). If v C0 (T ), this immediately imply that E (u, v ) = v (). Let v (C0 (T ))E . Then there exists {vn }n1 C0 (T ) such that E (v vn , v vn ) 0 and vn () v () as n . Since E (u, vn ) = vn (), it follows that E (u, v ) = v () by taking the limit as n . Note that E (u, 1) = 0 = 1. This implies that 1 / (C0 (T ))E . By Theorem 2.6, the corresponding Markov chain (T, C ), i.e. {rx }xT# , is transient. (3) Write (x) = g (, x) for any x T . Then Em (u, ) = u() (x) () ( (x)) p(x) u(x) rx rx xWm xS () = u() (x ) u(x) (x )
xS ()

= u()

xWm

u(x) (x ).

If xWm u(x) (x ) 0 as n , then we obtain E (u, ) = u(). Also we have E (u, ) = () as (C0 (T ))E . Assume E (u, u) = u(). Then E (u , u ) = 0 and hence u is a constant. Since u() = (), it follows that u = . By (4.3), Theorem 5.2-(1) and (10.3), (x ) = g (, (x)) g (, x) u( (x)) u(x) = = (x ) rx rx 51

xWm

for any x T# . Therefore = . For (, ) (T ) MP (), we dene (rx )xT# by (10.3) and the corresponding tree (T, C ). Let (E , F ) be the associated resistance form on T .
1 Lemma 10.3. Dene : T R by (x) = (x) . Assume , (T ). (1) xWm (x)(x ) 0 as m . (2) F and E (, ) = (). For any v (C0 (T ))E , E (, v ) = v (). (3) (rx )xT# is transient. Proof. (1) Dene m : R by m = xWm (x)x . Since (T ), m is monotonically decreasing. Moreover, m ( ) = ([ ]m ) 0 as m for -almost every . Hence xWm (x)(x ) = m ( )( ) 0 as m . (2), (3), (4) Let u = . Then by (10.3), u satises (10.4). Using (1) and applying Lemma 10.2, we immediately obtain (2), (3) and (4).

Finally we give a proof of Theorem 10.1. Proof of Theorem 10.1. Assume (T ). Let m = xWm (x)x . Then, ( ) = ([ ] ) 0 as m . Since is monotonically non-increasing, m m m ( x ) ( ) 0 as m . Combining this fact with ( ) ( ) = x xWm m Lemma 10.3, we can verify all the assumptions in Lemma 10.2 with u = . Therefore, (x) = g (, x) for any x T , where g (x, y ) be the symmetrized Green function of (T, C ). By Theorem 5.2-(1), (g (, x)1 )xT is the eigenvalue map of {rx }xT# . Note that we have dened (x) = (x)1 . Hence ((x))xT is the eigenvalue map of (rx )xT# and is the hitting distribution starting from .

11

Examples: the inverse problem

Let T = W , where W is the (innite complete) binary tree dened in Section 8. Let G represents the structure of the binary tree as a non-directed graph. We dene : by (i1 i2 . . .) = i2 i3 . . .. (1) First we describe an example ((1) , ) where (1) / , ) (T ) but ( (T R(G)). Lemma 11.1. Dene xm = 22 . . . 2 Wm for m 0, where x0 = , and ym = xm1 1 for m 1. Then = (m1 ym ) {222 . . .}. Denition 11.2. (1) Dene as the Bernoulli measure on with weight (1/2, 1/2), that is, is a Borel regular probability measure and (x ) = 2|x| . (2) We dene (1 (x))xT by 1 (xm ) = 1 + 2m for any m 0 and 1 (ym w) = 2(|w|+1) 1 (xm1 ) for any m 1 and w T . Set (1) (x) = 1 (x)1 for any x T. By the above lemma, 1 ([ ]n ) 0 as n if and only if = 222 . . .. Hence we have the following proposition. 52

 Proposition 11.3. (1) , (T )\ (T ).

By (4.3) and (10.3), it follows that the only candidate for (rx )xT# associated with ((1) (x), ) is { 1 if x = xm for m 1, (1) rx = m1 2 +1 if x = ym w for m 1 and w T . Let (T, C (1) ) be the tree corresponding to (rx )xT# . Lemma 10.3 shows that (T, C (1) ) T R(G). Theorem 11.4. ((T, C (1) )) = ((1) , ). Proof. By the denition of (1) , { (1) ([ ]n )}n0 does not converge to 0 as n if and only if = 222 . . .. Let be the hitting distribution starting from (1) of (rx )xT# . Then by (4.3), we have (xm ) = (Dxm1 Dxm )/rxm = Dxm1 Dxm . Since {Dxm }m0 is convergent as m , it follows that (xm ) 0 as m . This shows that ({222 . . .}) = 0. Hence (1) ([ ]n ) 0 (1) (1) (x)x . as n for -almost every . Dene m = xWm Then m ( ) = ([ ]m ) 0 as m for -almost every . Hence (1) (1) (x) (x ) = m ( ) (d ) 0 as m . Lemma 10.2 implies xWm that (1) = ( (1) )1 is equal to the hitting distribution of (rx )xT# and that = . , (T ), Next example ((2) , ) does not belong to (T R(G)) but satises that (2) where is the same as in Denition 11.2.
(1) (1)

Denition 11.5. (1) Dene V2m W2m by V2m = {11, 22}m for m 0. Also dene V2m1 = {wi|w V2m2 , i {1, 2}} and U2m = {wij |w V2m2 , i, j {1, 2}, i = j } for m 1. (2) Let { 1 + 2n if x Vn for n 0, (2) (x) = 2(|w|+1) (1 + 2(2m1) ) if x = yw for y U2m and w T . Dene (2) (x) = ( (2) (x))1 . Note that = {11, 22}N (m1 (yU2m y )). Since ({11, 22}N ) = 0, we have the next proposition.
Proposition 11.6. (2) , (T )\ (T ).

As in the case of (1) , the only candidate for (rx )xT# is given by (10.3). We (2) use (rx )xT# to denote this candidate whose values are { 1 if x Vn for some n 1, (2) rx = 1 + 22m1 if x = yw for y V2m and w T . We use (T, C (2) ) to denote the corresponding tree. 53

Theorem 11.7. Let ((T, C (2) )) = ((x )xT , ). Then (x )xT (T ) and ({11, 22}N ) > 0, where {11, 22}N = {i1 i2 . . . |i2m1 = i2m for any m 1}. In particular, ((2) (x))xT = (x )xT and is not absolutely continuous with respect to . In particular, ((2) , ) / (T R(G)). Constructing ((T, C (2) )), we prove Theorem 11.7 in the rest of this section. Lemma 11.8. Dene {m }m0 by ( ( 1 )2m+1 )1 ( ( 1 )2m1 )1 m = (1)m1 1 1 2 2 2m+1 and let F (z ) = m0 m z . Then ( (z ) ) 6z 2 F (2z ) F (z ) + 2 F F (z ) = 2 F (z ) 2 4z + 1

(11.1)

(11.2)

for any |z | < 1/2 and F (t) is strictly monotonically increasing on [0, 1/2]. Proof. Let F (z ) = m0 m z 2m+1 . Then (11.2) yields (
m1

Xm +

) ) ( 32 2 12 m z 2m+1 + Xm + 3 m+1 z 2m+1 Xm Xm m0 m z 2m+1 =6

where Xm = 22m+3 . This implies 1 = 80 /7 and (22m+3 16)m + (22m+3 1)m+1 = 0. for m 1. In fact, by (11.1), these equalities do hold if m = m for any m 0. Hence we have (11.2). Note that 0 = 2, 1 = 16/7 and that 0 < 2n < 2n1 for n 1. Hence for t 0, (4n 1)2n1 t4n2 (4n + 1)2n t4n + F ( t) = 2 + ( ) 2+ (4n 1)2n1 (4n + 1)|2n |t2 t4n2 .
n1 n1 n1

Since (4n 1) > (4n + 1) 1/4 for n 1, it follows that F (t) > 2 for t [0, 1/2] and hence F (t) is strictly monotonically increasing on [0, 1/2]. Using (11.2), we have the following lemma. Lemma 11.9. Dene tm = 2m / 2. Then for m 2, ( ) 3 1 2m1 F (tm1 ) F (tm ) + 2(F (tm+1 ) F (tm )) = ( 2 )22 1 m1
2

+1

F ( tm )

(11.3)

54

Routine calculations using (11.3) show the next lemma. Lemma 11.10. Let = 3(5F (t1 ) 4F (t2 ))1 . Dene ( (x))xT by F (t1 ) + 1 if x = , 2 F (t ) if x V2m1 for m 1, m (x) = (2F (tm+1 ) + F (tm ))/3 if x V2m for m 1, (|w|+1) 2 F (tm ) if x = yw for y U2m and w T . Then (L(2) )() = 1 and (L(2) )(x) = 0
(2) (rx )xT# .

(11.4)

for any x T# , where L(2) is the Laplacian associated with

Proof of Theorem 11.7. By Lemma 11.8, if (x) = (x)1 , then (T ). (2) Dene () = 1 and (x ) = ( ( (x)) (x))/rx for any x T# . Then (11.4) shows that (x ) = (x1 ) + (x2 ). Hence is identied as a Borel regular probability measure on and MP (). Applying Lemma 10.2 with u = and = , we have that ( , ) = ((T, C (2) )). Obviously = (2) . If ({11, 22}N ) = 0, then (2) ([ ]n ) 0 as n for -almost every . Then Lemma 10.2 implies that (2) is the eigenvalue map of (r(2) )xT# . This contradiction shows that ({11, 22}N ) > 0.

12

Relation to noncommutative Riemannian geometry

In [13], Pearson and Bellissard have constructed a framework for noncommutative Riemannian geometry on the Cantor set. Starting from an ultra-metric, they have obtained a probability measure , Dirichlet forms depending on a parameter s and associated Laplacians. From our point of view, they have dened Dirichlet forms by giving J : T and as in Section 9. Let us give a quick introduction to their construction of a measure and a Dirichlet form. Unfortunately, any avor of noncommutative geometry, which was the spirit of the original paper [13], does not remain in our brief review. Let x a structure of a tree T as a non-directed graph as in Section 9. Namely, we x T and : T T . Let : T (0, ) satisfy that (x) > (y ) for any x T and any y S (x) and that limn ([ ]n ) = 0 for any . We dene an ultra-metric d (, ) on by d (, ) = ([, ]). It is easy to see that d is a metric on which gives the same topology as the original one. Dene the zeta function (z ) associated with the ultra-metric d by (z ) = (x)z
xT

and assume that there exists s0 > 0 such that (z ) has a singularity at z = s0 and is holomorphic on {z |z C, Re(z ) > s0 }. 55

Theorem 12.1 (Pearson and Bellissard). With some regularity assumptions on , there exists a unique Borel regular probability measure on which satises s y T (y ) (x ) = lim x s s s 0 y T (y ) for any x T . Denition 12.2. For s R, dene Js : T [0, ) by Js (x) = (x)s2 , (w,v )Yx (w )(v )

where Yx = y,zS (x),y=z y z . Also dene s : T R by s = (Js ). The Dirichlet form (Qs , Dom(Qs )) on L2 (, ) dened in [13] coincides with (QJs , , DJs , ). Note that Js (x) > 0 and hence s (x) > 0 for any s R and any x T . Combining Theorem 9.3 and Corollary 9.8, we obtain the following fact which was originally obtained in [13]. Theorem 12.3. Let Ls be the non-negative denite self-adjoint operator associated with the Dirichlet form (QJs , , DJs , ) on L2 (, ). Then the spectrum of Ls is pure point. By Theorem 10.1, if s (T ), then (s , ) = ((T, C )) for some transient tree (T, C ). Unfortunately, it is dicult to tell whether s (T ) or not in general. From now on, we discuss this problem in the case of a self-similar on the (complete innite) binary tree W which has been introduced in Section 8. We let T = W for the rest of this section. Denition 12.4. Let 1 , 2 (0, 1). Dene : T (0, 1) by (w1 . . . wm ) = w1 wm for w1 . . . wm T . The case when 1 = 2 has been studied in [13]. Proposition 12.5. (1) The zeta function (z ) associated with the ultra-metric d is given by (z ) =
( m=0

(1 )z + (2 )z

)m

1 ( ) 1 (1 )z + (2 )z

and the singularity s0 is the unique real number which satises (1 )s0 + (2 )s0 = 1. (2) Let i = (i )s0 for i = 1, 2. Then the measure given in Theorem 12.1 is the self-similar measure with weight (1 , 2 ), i.e. (w1 ...wm ) = w1 wm for any w = w1 . . . wm T .

56

By the above proposition, it follows that (w ) = (w)s0 . Hence Js (w) = (w)s22s0 21 2 (12.1)

for any w T . Lemma 9.11 implies that s (T ) if and only if s < 2 + 2s0 . By (12.1), s (w) =
m 1 ) 1 ( (w1 . . . wm )s2s0 + (1wn+1 )(w1 . . . wn )s2s0 (12.2) 1 2 n=0

for any w = w1 . . . wm W . (12.2) has been obtained in [13] for the case when 1 = 2 . Proposition 12.6. s (T ) if and only if s 2+ s0 . Moreover, if s 2+ s0 , then { (w)s2s0 if s < 2 + s0 , s (w ) |w | if s = 2 + s0 for any w T . Corollary 12.7. There exists a transient tree (T, Cs ) such that ((T, Cs )) = (s , ) if and only if s 2 + s0 . Let d(s) be the metric associated with (T, Cs ) dened in Denition 5.3. Since d (, ) = s ([, ])1 , Proposition 12.6 yields { d (, )2+s0 s if s < 2 + s0 , (s) d (, ) (12.3) N (, )1 if s = 2 + s0
( s)

for any , . By (12.3), has the volume doubling property with respect to d(s) if and only if s < 2 + s0 . Applying the results in Section 6, we have the following asymptotic estimates of the heat kernel and moments of displacement. Theorem 12.8. Assume s < 2+ s0 . There exists a jointly continuous transition density p(t, , ) on (0, ) for the Hunt process associated with the Dirichlet form (QJs , , DJs , ) on L2 (, ). Dene t if d (, )2+s0 s > t, d (, )2+2s0 s q1 (t, , ) = 1 if d (, )2+s0 s t. 1 (Bd (, t /(2+s0 s) )) Then, p(t, , ) q1 (t, , ) for any (t, , ) (0, ) . Moreover, t if > 1, ( ) (2+s0 s) E d (, Xt ) t(| log t| + 1) if = 1, (12.4) t if 0 < < 1 for any and any t (0, 1]. 57

In the case when 1 = 2 , Pearson and Bellissard have obtained an averaged version of (12.4) with an exact expression of the leading term in [13]. See Theorem 14.1 and the following discussion for an observation on the appearance of the | log t| term in the case of = 1.

13

Appendix A: transience of resistance forms

In this appendix, we rst introduce several basic notions on resistance forms which are needed in this paper. Then, we will study the transience of general resistance forms as an extension of the case of weighted graphs in Section 2. Denition 13.1 (Resistance form). Let X be a set. A pair (E , F ) is called a resistance form on X if it satises the following conditions (RF1) through (RF5). (RF1) F is a linear subspace of (X ) containing constants and E is a nonnegative symmetric quadratic form on F . E (u, u) = 0 if and only if u is constant on X . (RF2) Let be an equivalent relation on F dened by u v if and only if u v is constant on X . Then (F /, E ) is a Hilbert space. (RF3) For any x, y X , if x = y , then there exists u F such that u(x) = u(y ). (RF4) For any p, q X , { } |u(p) u(q )|2 sup u F , E (u, u) > 0 E (u, u) is nite. The above supremum is denoted by R(E ,F ) (p, q ). (RF5) u F and E ( u, u ) E (u, u) for any u F , where if u(p) 1, 1 u(p) = u(p) if 0 < u(p) < 1, 0 if u(p) 0. Let (E , F ) be a resistance form on a set X . We call R(E ,F ) (, ) the resistance metric on X associated with (E , F ). (It is known that R(E ,F ) (, ) is a metric on X . See [9] for example.) Hereafter in this section, (E , F ) is always a resistance from on X and R(, ) is the associated resistance metric. We now introduce two important notions, the Green function and trace. Let B be a nonempty subset of X . The following result enables us to dene the B -Green function gB (, ). See [8, Section 4] for details and the proof. Theorem 13.2 (Green function). Let F (B ) = {u|u F , u|B 0}. Assume that uF (B ) u1 (0) = B . x F (B ) and (1) There exists a unique gB : X X [0, ) satisfying that gB x x E (gB , u) = u(x) for any u F (B ) and any x X , where gB (y ) = gB (x, y ). Moreover, gB (x, y ) = gB (y, x) and gB (x, y ) > 0 if and only if x, y X \B . 58

(2) Let b B and dene u(B ) = u(b) for u F (B ) + R. Then (E , F (B ) + R) can be regarded as a resistance form on XB = (X \B ) {B } Let RB (, ) be the associated resistance metric, then gB (x, y ) = for any x, y X . The function gB (, ) is called the B -Green function of the resistance form (E , F ) on X . Remark. If B is a single point, then the assumption of Theorem 13.2 is satised by the denition of the resistance form and the next theorem shows that F (B )+ R = F . Hence RB = R in this case. The next theorem is from [8, Section 7]. Theorem 13.3 (Trace). Set F|B = {u|B : u F}. (1) There exists a linear map hB : F|B F such that hB (f )|B = f and E (hB (f ), hB (f )) = min{E (u, u)|u F , u|B = f } for any f F|B . Moreover, hB (f ) is the unique element in {u|u F , u|B = f } that attains the above minimum and F = F (B ) Im(hB ), where means that E (u, v ) = 0 for any u F (B ) and any v Im(hB ). (2) Dene E|B (f, f ) = E (hB (f ), hB (f )) for f F|B . Then (E|B , F|B ) is a resistance form on B and the associated resistance metric on B is equal to the restriction of R onto B . (E|B , F|B ) is called the trace of (E , F ) on B and hB (f ) is called the B harmonic function with the boundary value f . To extend the notion of transience for general resistance forms, we need to study the collection of resistance forms (E , F ) on X which are restrictions of (E , F ) i.e. F F and E (u, v ) = E (u, v ) for any u, v F . Denition 13.4. Let A be a subset of F . (1) A is said to have the Markov property if and only if u A for any u A. (2) A is said to separate points in X if and only if, for any x, y X with x = y , there exists u A such that u(x) = 1 and u(y ) = 0. (3) Dene N and A(E ,F ) by N = {F |F F , (E , F ) is a resistance form on X } and A(E ,F ) =
F N ,AF

RB (x, B ) + RB (y, B ) RB (x, y ) 2

F .

The above denition of separation of points is slightly stronger than the usual one, which only requires the existence of function u with u(x) = u(y ).

59

Proposition 13.5. Let p X . Dene Ep (u, v ) = E (u, v ) + u(p)v (p) for any u, v F . Then Ep is an inner product on F and (F , Ep ) is a Hilbert space. The topology induced by Ep is independent of the choice of p X . Let 1 be the characteristic function of X . For a linear subspace A of F , we use A + 1 to denote the direct sum of A and the space of constants on X . Proposition 13.6. Let A be a linear subspace of F . If A separates points on X , then A(E ,F ) N . Moreover, if A + 1 has the Markov property, then A(E ,F ) = (A)Ep + 1, where (A)Ep is the closure of A with respect to the inner product Ep . Proof. It is immediate to verify that (E , A(E ,F ) ) is a resistance form on X . Assume that A + 1 has the Markov property. Then we can easily show that (A)Ep + 1 is a resistance form on X . Since every F N contains (A)Ep + 1, it follows that A(E ,F ) = (A)Ep + 1. Denition 13.7. Let A be a subset of F . For any U X , dene AU = {u|u A, u(x) 1 for any x U } and { + if AU = , CapA (U) = inf uAU E (u, u) if AU = . The following result and the method of its proof are analogous to those of Theorem 2.6. Theorem 13.8. Let A be a subspace of F . Assume that 1 / A. Then the following conditions are equivalent: (T1) 1 / (A)Ep . (T2) For any x X , there exists c > 0 such that u(x)2 cE (u, u) for any u A. (T3) CapA (K) > 0 for any non-empty compact subset K X . (T4) CapA (K) > 0 for some non-empty compact subset K X . Proof. (T1) (T3): Assume that CapA (K) = 0 for a non-empty compact set K . Then AK = . Choose x K . There exists un A such that un (x) = 1 for any n and E (un , un ) 0 as n . Dene vn = un un (p). Then {vn } is a Cauchy sequence with respect to Ep . Let v be the limit of {vn }. Since vn (x) = un (x) un (p) = 1 un (p) v (x) as n , {un (p)} converges to some R as n . Hence un = vn + un (p) converges to v + with respect to Ep . Now E (v + , v + ) = limn E (un , un ) = 0. Therefore v + is a constant. Note that v (x) + = 1, we see that {un } converges to 1 as n . Hence 1 (A)Ep . (T3) (T4): This is obvious. (T4) (T1): Assume 1 (A)Ep . Then there exists {un }n1 A such that Ep (1 un , 1 un ) 0 as n . Let x K . Then un (x) 1 as n . Let M = diam((K, R)). Then |un (x) un (y )|2 M E (un , un ) 60

for any y K . Hence we may choose N > 0 such that 2un AK for any n N . Since E (2un , 2un ) 0 as n , we have CapA (K) = 0. (T1) (T2): If (T2) does not hold, then there exists {un } A which has exactly the same properties as in the implication of (T1) (T3). The rest of the argument is entirely the same. (T2) (T4): This is obvious. Denition 13.9. Let A be a subspace of F . We say that (E , F ) is A-transient if and only if 1 / (A)Ep and denote (A)Ep + 1 by F (A). In particular, we say (E , F ) is transient if it is F C0 (X )-transient, where C0 (X ) is the collection of continuous functions with compact supports with respect to the resistance metric. We write F = F (F C0 (X )). In Example13.12, we have non-trivial example of A other than F C0 (X ). If (E , F ) is the resistance form associated with a weighted graph (V, C ), then C0 (V ) F and hence (V, C ) is transient if and only if (E , F ) is transient. If (E , F ) is A-transient, it is natural to expect that (E , F (A)) is a resistance form on X {IA }, where IA should correspond to the innity. To realize such a statement, we need to assume an additional property of A which is an extension of the Markov property. Denition 13.10. Let A be a subspace of F . A is said to have the extended Markov property if and only if u + a a A for any u A and any a R. It is easy to see that F C0 (X ) has the extended Markov property. Note that if A has the extended Markov property, then both A and A + 1 have the Markov property. Theorem 13.11. Let A be a linear subspace of F which separates points in X and has the extended Markov property. Assume that (E , F ) is A-transient. Let IA / X and extend u F (A) to u : X {IA } R by setting u(IA ) = a, where u = f + a for f (A)Ep and a R. Then (E , F (A)) is a resistance form on X {IA }. Proof. Write I = IA . Since A + 1 has the Markov property, (E , F (A)) is a resistance form on X by Proposition 13.6. We show that (E , F (A)) is a resistance form on X {I }. (RF1), (RF2) and (RF3) are immediate. Since u u(I ) (A)Ep for any u F (A), (T2) of Theorem 13.8 implies that { sup } |u(x) u(I )|2 u F (A), E (u, u) > 0 < +. E (u, u)

This yields (RF4). Let u F (A) and set v = u u(I ). Then v (A)Ep and u u(I ) = v + u(I ) u(I ) (A)Ep . Hence (u) = u(I ) and we have (RF5). Example 13.12. Let (T, C ) be the (complete innite) binary tree dened in Section 8 and let (E , F ) be the associated resistance form on T . Dene A = {u|u F , #{w|w = w1 . . . wm T, w1 = 1, u(w) = 0} < +}. Then A satises 61

the extended Markov property and separates points in T . Let C = CS , where CS is the self-similar weight dened in Denition 8.3. Then by the similar arguments as in the proof of Theorem 8.4, it follows that (E , F ) is A-transient if and only if r1 r2 /(r1 + r1 ) < 1. Also we have F = F (C0 (V )) F (A) F and none of the equalities hold.

14

Appendix B: moments of displacement

In this section, we will consider general Hunt processes on a compact metric space (X, d). Let be a Borel regular nite measure on (X, d). Let (E , F ) be a regular Dirichlet form on L2 (X, ). We use ({Xt }t>0 , {Px }xX ) to denote the associated Hunt process on X . Assumptions (1) There exists a jointly continuous transition density p(t, x, y ) on (0, ) X X such that p(t, x, y )f (y )(dy ) Ex (f (Xt )) =
X

for any x X , any t > 0 and any bounded -measurable function f : X R. (2) There exist c1 , c2 , c3 (1, ) and R > 0 such that c1 (Bd (x, r)) (Bd (x, c3 r)) c2 (Bd (x, r)) (14.1)

for any x X and r (0, R]. (3) There exists a strictly increasing function : [0, ) [0, ) with (0) = 0 such that the following transition density estimate holds, i.e. if we dene q (t, x, y ) by t ) ( ) if t (d(x, y )), ( d(x, y ) Bd (x, d(x, y )) q (t, x, y ) = 1 ) if 0 < (d(x, y )) t, ( Bd (x, 1 (t)) then p(t, x, y ) q (t, x, y ) for any x, y X and any t (0, 1]. (4) There exist c4 , c5 , c6 (1, ) such that c4 (r) (c5 r) c6 (r) for any r 0. Remark. (14.2) may be thought of as one of typical heat kernel estimates for jump processes. In [3], Chen and Kumagai have studied a Hunt process associated with a Dirichlet form (E , F ) dened by E (u, u) = (u(x) u(y ))2 J (x, y )(dx)(dy ),
X

(14.2)

(14.3)

62

where J (x, y ) is a given jump kernel. Assuming that (Bd (x, r)) V (r) for some strictly monotone function V : [0, ) [0, ), they have shown that (14.2) holds if J (x, y )1 (d(x, y ))V (d(x, y )). Theorem 14.1. Suppose that the above four assumptions hold. Then if > 1, t Ex ((d(x, Xt )) ) t(| log t| + 1) if = 1, t if 0 < < 1, for any t (0, 1] and any x X . The | log t| + 1 term in the case of = 1 is due to the (relatively) slow decay of o-diagonal part t/((d(x, y ))(Bd (x, d(x, y ))) of the heat kernel estimate with respect to the space variable x. Note that the decay of the o-diagonal part is exponential if a heat kernel enjoys (sub-)Gaussian asymptotic behavior which is typical in many cases of diusion processes, for example, the Brownian motion on Euclidean domain and the Sierpinski gasket. In such a case, there appears no | log t| + 1 term in moments of displacement. Remark. Suppose that (X, d) is not bounded. If Assumption (1) holds, (14.1) holds with R = , (14.2) holds for any x, y X and any t > 0 and (14.4) hold, then an analogous argument as the proof of Theorem 14.1 shows { + if 1, Ex ((d(x, Xt )) ) = t if 0 < < 1. We use V (x, r) = (Bd (x, r)). Lemma 14.2. Suppose that the above four assumptions hold. Then there exists , , c (1, ) and T > 0 such that V (x, 1 (s)) V (x, 1 (s)) cV (x, 1 (s)) for any x X and any s (0, T ]. Proof of Theorem 14.1. Note that Ex ((d(x, Xt )) ) =
X

(14.4)

p(t, x, y )(d(x, y )) (dy ).

Since p(t, x, y ) is jointly continuous, it is enough to show (14.4) for t (0, T ]. We divide the domain X of the above integral into X1 = {y |(d(x, y )) > t} and X2 = {y |(d(x, y )) t}. Dene N = max{n|n Z, n t < T }. Set Yn = {y |n t (d(x, y )) < n+1 t} for n < N and YN = {y |N t (d(x, y ))}. By (14.2), t(d(x, y )) 1 p(t, x, y )(d(x, y )) (dy ) (dy ) X1 X1 V (x, d(x, y )) (14.5) N t(d(x, y )) 1 = (dy ) V (x, d(x, y )) n=0 Yn 63

Now, for n = 0, 1, . . . , N 1, (n t) 1 (V (x, 1 (n+1 t)) V (x, 1 (n t))) t V (x, 1 (n+1 t))
Yn

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

(n+1 t) 1 (V (x, 1 (n+1 t)) V (x, 1 (n t))) . (14.6) t V (x, 1 (n t)) By Lemma 14.2, we have
Yn

(d(x, y )) 1 (dy ) t(n t) 1 V (x, d(x, y ))

(14.7) that (14.7) is

for n = 0, 1, . . . , N 1. Using the similar arguments, we conrm valid for n = N . Hence by (14.5), (14.6) and (14.7), N t p(t, x, y )(d(x, y )) (dy ) t(n t) 1 t(| log t| + 1) X1 n=0 t Next let Zn = {y |n t > (d(x, y )) (n+1) t} for n 0. Then p(t, x, y )(d(x, y )) (dy )
X2 X2

if > 1, if = 1, if 0 < < 1. (14.8)

(d(x, y )) (d(x, y )) ( dy ) = (dy ) V (x, 1 (t)) V (x, 1 (t)) n=0 Zn

(n t) (V (x, 1 (n t)) V (x, 1 ((n+1) t)) V (x, 1 (t)) n=0

(14.9)

(n t) ct .

n=0

Combining (14.8) and (14.9), we obtain (14.4).

References
[1] R. M. Blumenthal and R.K. Getoor, Markov Processes and Potential Theory, Academic Press, 1968. [2] P. Cartier, Fonctions harmoniques sur un arbre, Sympos. Math., vol. 9, pp. 203270, Academic Press, 1972. [3] Z.-Q. Chen and T. Kumagai, Heat kernel estimates for jump processes of mixed types on metric measure spaces, Probab. Theory Related Fields 140 (2008), 277317.

64

[4] J. L. Doob, Discrete potential theory and boundaries, J. Math. Mech. 8 (1959), 433458. [5] M. Fukushima, Y. Oshima, and M. Takeda, Dirichlet Forms and Symmetric Markov Processes, de Gruyter Studies in Math. vol. 19, de Gruyter, Berlin, 1994. [6] G. A. Hunt, Markov chains and Martin boundaries, Illinois J. Math. 4 (1960), 313340. [7] P. T. Jorgensen and E. P. J. Pearse, A Hilbert space approach to effective resistance metric, to appear in Complex Anal. Oper. Theory, arXiv:0906.2535. [8] J. Kigami, Resistance forms, quasisymmetric maps and heat kernel estimates, preprint. [9] , Analysis on Fractals, Cambridge Tracts in Math. vol. 143, Cambridge University Press, 2001.

[10] J. Kigami and K. Takahashi, Trace of the standard resistance form on the Sierpinski gasket and the structure of harmonic functions, in preparation. [11] S. V. Kozyrev, Wavelet analysis as a p-adic spectral analysis, Izv. Ross. Akad. Nauk Ser. Mat. 66 (2002), 149158. [12] R. Lyons and Y. Peres, Probability on Trees and Networks, Book in preparation, current version available at http://mypage.iu.edu/rdlyons/. [13] J. Pearson and J. Bellissard, Noncommutative Riemannian geometry and diusion on ultrametric Cantor sets, J. Noncommutative Geometry 3 (2009), 447480. [14] P. M. Soardi, Potential Theory on Innite Networks, Lecture Note in Math., vol. 1590, Springer-Verlag, 1994. [15] W. Woess, Random walks on innite graphs and groups a surveey on selected topics, Bull. London Math. Soc. 26 (1994), 160. [16] , Random Walks on Innite Graphs and Groups, Cambridge Tracts in Math., vol. 138, Cambridge University Press, 2000.

[17] M. Yamasaki, Parabolic and hyperbolic innite networks, Hiroshima Math. J. 7 (1977), 135146. [18] , Discrete potentials on an innite network, Mem. Fac. Sci. Shimane Univ. 13 (1979), 3144.

65