This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
0 implies /xp(T) > 0. Hence by Theorem 1.3.8, we have the following corollary. p ™. For u G t(V+), we define P mu by 1, then A is called a Dirichlet eigenvalue (D-eigenvalue for short) of Aμ and any non-trivial (p G Ep (A) is called a Dirichlet eigenfunction (D-eigenfunction for short) of Aμ belonging to the Dirichlet eigenvalue A. Of course, eigenvalues and eigenfunctions of Aμ are expected to coincide with those of the non-negative self-adjoint operator H^ (or Hp, depending on the boundary condition) associated with the local regular Dirichlet form , u) M for any u G T§, (D3)
Corollary 1.4.9. If v (Kw)
= pw for any w G W*, then C is minimal.
Although the next theorem does not directly relate to self-similar measures, it tells us a useful fact: two Borel regular measures on a self-similar set are comparable if they are comparable on Kw for all w G W*. Theorem 1.4.10. Let C = (K, 5, {Fi}ies) be a self-similar structure. Let μ and v he Borel regular measures on (K,M^)) and (K,M(v)) respectively. Assume that v(K) < oo and v(I) = 0. // there exists c > 0 such that μ^yj) < cv(Kw) for any w G W*, then μ(A) < cis(A) for any A G M^μ) n M{y). In particular, //(I) = 0. Proof. Let U be an open subset of K. Set W(U) = {w G W* : Kw C U}. For w,v G W(U), we define w > v if and only if Hw D Ev. Then > is a partial order on W(U). If W~*~(U) is the collection of maximal elements in
28
Geometry of Self-Similar Sets
W(U) with respect to this order, then U = Uwew+(U)Kw and KWC\KV c X for w ^ v e W+(U). Therefore wew+(u)
wew+(U)
Now, by Proposition 1.4.2, for any A G A4(/x) D M(v), there exists a decreasing sequence of open sets {Ok}k>1 such that A C Ok for any k, μ(^k>lOk) = μ{A) and u(nk>iOk) = v(A). As ^(Ofc) < cv{Ok), we have /x(A) < ci/(A). D
1.5 Dimension of self-similar sets In this section, we will introduce the notion of the Hausdorff dimension of metric spaces and show how to calculate the Hausdorff dimension of self-similar sets. Definition 1.5.1. Let (X, d) be a metric space. For any bounded set A C X, we define H86(A) = i n f l ^ d i a m ^ ) 5 : A C U ^ i ^ d i a m ^ ) < 6}, i>1
where diam(i£) is the diameter of E defined in 1.2. Also, we define HS(A) = s limsup6ion 6(A). It is well-known that Hs is a complete Borel regular measure for any s s>0. See Rogers [157] and Falconer [32] for example. H is called the 5-dimensional Hausdorff measure of (X, d). Lemma 1.5.2. Let (X, d) be a metric space. For 0 < s
HUE) < S^HKE) for any E C X. Proof. If E C Ui>iEi and diam(Ei) < 6 for any i, then
Y^di&miEiY < ^diam^'MiamCEi) 5 < ^ " 5 ^ d i a m ( ^ ) s i>1
i>1
i>1
• By Lemma 1.5.2, we can obtain the following proposition. Proposition 1.5.3. For any E C X, sup{s : ns(E)
= oo} = inf{s : HS(E) = 0}.
(1.5.1)
1.5 Dimension of self-similar sets
29
S
Proof. By Lemma 1.5.2, if s < t, then H (E) < oo implies 7^(E) = 0 and S • also H\E) > 0 implies H (E) = oo. Now it is easy to see (1.5.1). Definition 1.5.4 (Hausdorff dimension). The quantity given by the equality (1.5.1) is called the Hausdorff dimension of E, which is denoted by dimuE. Remark. The Hausdorff measure and the Hausdorff dimension depend on a metric d. In this sense, if we need to specify which metric we are looking at, we will use the notation of dimn(^, d) instead of dimnE. The following lemma is often useful to calculate the Hausdorff dimension of a metric space. It is often called "Frostman's lemma". See, for example, Mattila [124]. It is also called the "mass distribution principle" in Falconer [33]. a Lemma 1.5.5. Let (K,d) be a compact metric space. IfH (K) < oo and there exist positive constants c, lo and a probability measure μ on Ksuch that
μ^iix))
< cl<*
for all x G K and any I G (0, IQ), then μ{A) < cHa(A) for any Borel set Ac K. In particular, 0 < Ha(K) < oo. Remark. According to the discussion of Moran [134], the converse of the above lemma is true: if 0 < 7ia(K) < oo, then there exists a probability measure μ on K such that, for some c > 0 , μ^xix))
for all x € K and / > 0. Moran proved this fact if K was a compact subset of Euclidean space. His argument, however, can be easily extended to this case. Proof. For U C K and x G 17, note that U C -Bdiam(i7)(x)A C \JiUi, then c^diam^r,
μ(A) < Y^M(Bdiam(tf,)te)) i
Hence, if
i
where x» G £/». Therefore μ{A) < cHf(A). Letting / —> 0, it follows that μ{A)
30
Geometry of Self-Similar Sets
Now let (K, { 1 , 2 , . . . , iV}, {i^}i
rVk for v = v\v2—
e W^,rWlW2,,,Wm_1 > a> rw}, Vk € W^.
Remark. A(r, a) is a partition of E. The following is our main theorem. This theorem was introduced in [88]. The ideas are, however, essentially the same as in Moran [134]. T h e o r e m 1.5.7. Assume that there exist r = ( r i , r 2 , . . . ,r/v) where 0 < 7*i < 1 and positive constants c\, c2, c* and M such that d i a m ^ ) < cxrw
(1.5.2)
#{w : w G A(r, a),d(x, Kw) < c2a) < M
(1.5.3)
for all w e W* and
31
1.5 Dimension of self-similar sets
for any x G K and any a G (0, c*), where d(x, Kw) = miy£Kw d(x, y). Then there exist constants C3, C4 > 0 such that for all A G B(K, d), czv(A) < na{A)
(1.5.4)
< c4v(A),
where v is a self-similar measure on K with weight (ria)i
and a is
e
f > « = l.
(1.5.5)
i=1
In particular, 0 < 7ia(K) < 00 and d i n i n g , d) = a. 1
Remark. Under the assumption (1.5.3), it is easy to see that #(7r~ (a:)) < M for any x G K. Hence, by Corollary 1.4.8, rwa
=
v(Kw) for any w G W*. Also u(I) = 0.
a
a
Proof We write A a = A(r,a). First we show that H (Kw) < (ci) v(Kw) for all w G W*. For w = W\W2 .. .wm G Ty*, we define Aa(w) = {v = v\V2 . •. Vk • wv G Att}^ where wv = W\W2 . . . wrnv\V2 . . . Vfc- Then we can see that A a (w) is a partition for sufficiently small a. Hence
rwa=
rwva.
^2
(1.5.6)
veAa(w)
By (1.5.2), it follows that diam(Kwv) < C\rwv < c\a for v G Aa(w). note that X^, = UveAa{w)Kwv Then
W?^^) < d
e
a
2
rwv
Also
a
= {Clrw) .
veAa(w)
Letting a —> 0, we obtain W Q ( ^ ) < (ci)Qr-wQ = ( d ) 0 * / ^ ) . a
a
Next we show that i/(KTO) < Mc2~ H (Kw). Let μ be the Bernoulli meaa sure on S with weight (ri )i
ir-l(BC2a{x))c
(J E™,
where Aa>a; = {t<;: w G A a , d(x, X^) < C2a}. Hence it follows that
v{BC2a{x))<
Y^ ^ € A a) X
M(S W ).
32
Geometry of Self-Similar Sets
Since μ ^ j = r£ < aa and #(A O , X ) < Af by (1.5.3), we have a
*(£<*«(*)) <
a
Mc2- (c2a) .
Lemma 1.5.5 implies that u(A) <
Mc2-ocna(A).
for any A G B(K, d). Hence there exist C3, C4 > 0 such that c 3 ^ C M < Ha(Kw)
<
c4v(Kw).
By Theorem 1.4.10, we can verify (1.5.4).
•
In the rest of this section, we show that the open set condition implies (1.5.2) and (1.5.3) of Theorem 1.5.7. Proposition 1.5.8. Suppose K is a subset ofRk, d is the Euclidean metric k k k ofR and Fi : R —> R is an ri-similitude for i = 1,2,... , N with respect to d. If the open set condition holds, i.e., there exists a bounded non-empty open set O C Rk such that N
\jFi(O)cO
and F^0) n Fd(0) = 0
fori^j,
i=1
then there exist constants c\,c2,M > 0 such that diam(Kw) < c\rw for all w G W* and #{w : w e A(r,a),d(x,Kw)
< c2a} < M
for all 0 < a < 1 and x G K. Proof. We can see that Kw C Ow for any w G W*, where O^ = F^(0). (By Exercise 1.2, it follows that O 2 -K\) Without loss of generality, we may assume that diam(O) < 1. Then, for all w G W*, dian^if^) < diam(0ly) < rw. Let m be thefc-dimensionalLebesgue measure and let < a}. Then U ^ A ^ O U ; C 52a(^)Aa,x = {w : w e A(r,a),d(x,Kw) Since the Ow are mutually disjoint, we have Y^weA m{Ow) < m(B2a(x)). Hence it follows that #(AaiX)rwkm(0) < 2kCak, where C = m(unit ball). Since rw > aR where R = min{ri,r2,... ,rN}, we see that #(Aa>iC) < k k 2 CR- m(0)-K D Corollary 1.5.9 (Moran's theorem). If K satisfies the open set condition, then dining, d) = a, where a is given by (1.5.5) with r^ = Lip(Ff).
1.6 Connectivity of self-similar sets
33
1.6 Connectivity of self-similar sets Let (K, 5, {Fi}i€s) be a self-similar structure. In this section we will give a simple condition for connectivity of K and also show that K is connected if and only if it is arcwise connected. As a reminder, the definition of connectivity is as follows. Definition 1.6.1. Let (X, d) be a metric space. (1) (X, d) is said to be connected if and only if any closed and open subset of X is X or the empty set. Also a subset A of X is said to be connected if and only if the metric space (A, <1\A) is connected. (2) A subset A of X is said to be arcwise connected if and only if there exists a path between x and y for any #, y G A: there exists a continuous map 7 : [0,1] —> A such that 7(0) = x and 7(1) = y. (3) (X, d) is said to be locally connected if and only if, for any x E X and any neighborhood U of x, there exists a connected neighborhood V of x with V C U. Of course, arcwise connectivity implies connectivity, but the converse is not true in general. Now we come to the main theorem of this section. Theorem 1.6.2. The following are equivalent. (1) For any i,j € S, there exists {ik}k=o,i,... ,n £ S such that IQ = i, in — j and Kik n Kik+1 ^ 0 for any k = 0 , 1 , . . . , n — 1. (2) K is arcwise connected. (3) K is connected. Proof. Obviously (2) =» (3). So let us show (3) => (1). Choose i € S and define A C S by A = {j E S : there exists {ik}k=o,i,... ,n Q S such that io = h *n= j and Kik D Kik+1 ± 0 for any k = 0 , 1 , . . . , n - 1}. If U = Uf e A #j and F = U^lf,-, then C/ H V = 0 and Cf U V = X. Also both J7 and V are closed sets because Ki is closed and A is a finite set. Hence U is an open and closed set. Hence U = K or U = 0. Obviously KiQU and hence [/ = X. Therefore V = 0 and hence A = 5. To prove (1) => (2), we need the following lemma. Lemma 1.6.3. For a map u : [0,1] —• K and for t G [0,1], we define D(u,t) = s\xp{\ims\xpd(u(tn),u(sn)) n—•<»
: lim tn = lim sn = t}. n—»-oo
n—>oo
7f / n : [0,1] —• K is uniformly convergent to f : [0,1] —* K as n —> 00 and lim<«_.~> D( f^.s) = 0. £/ien f is continuous at s.
34
Geometry of Self-Similar Sets
Proof of Lemma 1.6.3. Let d be a metric on K that is compatible with the original topology of K. litn-+s and sn —> s as n —> oo, then <*(/(*»), /(«»)) < d(/(*n), /m(tn)) + d(/m(*»), /„,(«„)) + d(f m(sn),
f(sn)).
Set r m = sup{d(/ m (£),/(£)) : 0 < t < 1}. Then the above inequality implies D(f,s) < 2rm + D(/ m ,s). Now letting ra —* oo, we can see that D(f, s) = 0. Hence / is continuous at s. D Now we return to the proof of (1) => (2). Define P = {/ : X 2 x [0,1] -> X : /(p,g,0) = p and /(p, g, 1) = g for any (p, g) € K2}. Also for f,g £ P, set dp(/,0) = sup{d(/(p,(7, t),(/(p, (/,*)) : (p,g,t) G X 2 x [0,1]}. Then (P, dp) is a complete metric space. By (1), for any (p, q) e K2, we can choose n(p,g)-, {ik(p,q)}o
1.6 Connectivity of self-similar sets
35
In the rest of this section, we study connectivity of K\Vo for a connected p. c. f. self-similar structure. Proposition 1.6.6. Let (K, 5, {Fi}ies) be a connected post critically finite self-similar structure. Let C be a connected component of K\VQ. Then C is arcwise connected. Moreover, for any x G C and any p G C, there exists a path between x and p in C U {p}. Proof Let x G C. Note that Km,x is arcwise connected. (See Proposition 1.6.4 and its proof.) Since C is open, Km^x is contained in C for sufficiently large m. Hence if O = {y : y G C, there exists a path between x and y in C}, then O is open. Therefore, O = C. So C is arcwise connected. Let p G C. If p G C, the statement is obvious because C is arcwise connected. So assume that p G Vo- Choose m so that KmyP n Vb = {p}. Let y G ifm,p n C. Since lfm,p is (arcwise) connected, there exists a path between y and p. This path is included in C U {p}. Also there exists a path between x and y which is included in C. Joining those two paths, we obtain the desired path between x and p. • Proposition 1.6.7. Let (K,S,{Fi}i^s) be a connected post critically finite self-similar structure. Let p be the fixed point of F{. If C is a connected component of K\{p}, then Cfl Vb ^ 0. In particular, the number of connected components of K\{p} is finite. Moreover, let {Cj}j=i^,...}m be the collection of all connected components of K\{p}. Then there exists a permutation of {1,2,... , m}, p, such that Fi(Ck) = Cp^ n Ki. Proof. Suppose that C/i,... ,17/ are connected components of K\{p}. Then we may choose n so that Uj is not contained in Fin(K) for all j = 1,2,... , /. By Proposition 1.3.5-(2), UJOF^CVQ) ^ 0 for any j = 1,2,... , I. Therefore I #(Vb)- This implies that the number of connected components of K\{p} is finite. Now, let Ci,... ,C m be the collection of all connected components of K\{p}. Note that Fi(Cj) is connected and Uf=1Fi(Cj) = Ki\{p}. Therefore, there exists p(j) such that Fi(Cj) C Cp(jy Lemma 1.3.14 implies that KijP = Ki. Hence Ck H K{ ^ 0 for any k and there exists j such that Cj n F^~1(Cfc) T^ 0. This implies that p is a permutation of {1,2,... , m}. As X A W = UJLiCKi n Ci), we see that i ^ n C p(j) = F^Cj) for any j . Next choose n so that Cj is not contained in Fin(K) for any j . Then, it follows that Cpn(k) n i^71 (Vb) ^ 0 for any k. Since Cpn(fc) n Fi n (F 0 ) = Fin(V0 n Ck), we obtain Cfc n Vb ^ 0 for any jfe. • Proposition 1.6.8. Let C = (K, 5, {Fi}ies) be a connected post critically
36
Geometry of Self-Similar Sets
finite self-similar structure. (1) For any p G Vo, let J(p) be the collection of all connected components of K\{p}. Then J(p) is a finite set. (2) The number of connected components of K\Vo is finite. (3) Forpe Vo, define J(p, Vo) = {C : C is a connected component of K\Vo,p G C}. Set m(p, Vo) = #(«/(p, Vo)) and m(p) = #(J{j>)). Ifp eV0 is a fixed point of Fw for some w G W*\Wo, then m(p, Vo) = m(p). Proof. (1) Suppose that J(p) is an infinite set. Let {Ci}i>\ be a collection of connected components of K\{p}. Assume that Cj ^ Cj if i ^ j . Claim 1. # { j : diam(Cj) > e} < oo for any e > 0 . Suppose that diam(Cjfc) > e for any k>0. Then for any k, there exists Xk G Cjh such that d(p,Xk) > e/2. Since K is compact, we may choose a subsequence {xkn}n>i so that Xkn —>• x as n —» oo for some x G K. Let C be the connected component of K\{p} with x G C. Since C is a neighborhood of x, £fcn belongs to C for sufficiently large n. This contradicts the fact that each Xkn belongs to a different Cj. Hence we have the claim. Now since C is p. c. f., we may write n~1(p) = (u;(l),... ,u;(ra)}. Then there exists k such that, for any j = 1,2,... , m, o;(j) = w(j)v(j) for some w(j) G Wfc and some v(j) G VF*. Let pj = 7r(v(j)) for jf = 1,2,... ,ra. Note that p^ G Vo and Fw^(pj) = p for all j . Also, by Proposition 1.6.7, the number of connected components of K\{pj} is finite. (We may change the self-similar structure to Cn = (K, W n , {Fw}wewn)i where n = \v(j)\. Then apply Proposition 1.6.7 to pj, which is a fixed point of Fv^y) Since Kkfp = \JjLiFw(j)(K), it follows that the number of connected components of Kk,p\{p} is finite. (See Proposition 1.3.6 for the definition of Kk,p.) By the way, since K^^ is a neighborhood of p, Claim 1 implies that K^^ contains infinitely many connected components of K\{p}. This contradiction shows that the number of connected components of K\{p} is finite. (2) Suppose that the number of connected components of K\Vo is infinite. Then there exists p G Vo such that J(p,Vo) is an infinite set. By the same discussion as in the proof of (1), it follows that, for any e > 0 , #{C|C G J(p,Vb), diain(C) < e} = oo. On the other hand, let e = ming€y0\{p} d(p, q)/2 and assume that C G J(p, Vo) and diam(C) < e. Then C is a connected component of K\Vo- By (1), {C\C G J(p,Vo), diam(C) < e} is a finite set. (3) Let k = m(p,Vo) and let U = (UCeJ(P,v0)C) u W - T n e n u i s a
1.6 Connectivity of self-similar sets neighborhood of p. Hence, if w(n) = w...w,
37
then Kw(n) C U for suffi-
n times
ciently large n. Therefore if k' is the number of connected components of Kw{n)\{p}i then k' > k. Since m(p) = k\ we see that m(p) >ra(p,Vb)On the other hand, a connected component of K\{p} contains at least one C G J(p, Vb). Hence m(p) < m(p, Vb). • The next proposition also concerns a connected p. c. f. self-similar structure. It gives a sufficient condition for K\VQ to be connected. be a connected post critically fiProposition 1.6.9. Let (K,S,{Fi}ies) nite self-similar structure. Assume that, for any p,q G Vb, there exists a homeomorphism g : K —* K such that g(Vo) = Vb and g(p) = q. Then K\Vo is connected. If a connected p. c. f. self-similar structure satisfies the assumption of the above proposition, we say that the self-similar structure is weakly symmetric. To prove the above proposition, we need the following lemmas. Lemma 1.6.10. Assume the conditions in Proposition 1.6.9. Let C be a connected component of K\VQ. Then #(C C\ Vo) > 2. Proof. Suppose that there exists a connected component of K\Vb satisfying #(C Pi Vo) = 1. Let po be the unique po G Vo with p0 G C. For any p G Vb, there exists a homeomorphism g : K —• K such that g(po) = PLetting Cp = ^(C), we see that Cp is a connected component of K\Vo and CpCiVo = {p}. Then it follows that Cp is a connected component of K\{p}. Now, since C is post critically finite, there exists p G Vo such that p is a fixed point of Fw for some w G W*\Wo. Let m = \w\. By exchanging the self-similar structure C with Cm = (K, Wmi {Fv}vewm), we can use Proposition 1.6.7 and obtain that Cp D Vb ^ 0. This contradicts the fact that CpnV0 = {p}. • Lemma 1.6.11. Assume the conditions in Proposition 1.6.9. Thenjor all pe Vo, m(p,Vo) =m(p). Proof. Since C is weakly symmetric, m(p) and m(p1 Vo) are independent of the choice of p G Vb- Hence, as in the proof of the last lemma, we may assume that Fw(p) = p for some w G W*\WQ. Then the statement is immediate by Proposition 1.6.8-(3). • Proof of Proposition 1.6.9. Let J be the collection of connected components of K\VQ. Define V = Vo U J and E = {(p, C):peV0,C G J(p, Vo)}.
38
Geometry of Self-Similar Sets
Let G = (V, E) be a undirected graph, where V is the set of vertices and E is the set of edges. Note that G is connected. First we show that this graph G contains no loop. Suppose that there exists a loop in G : there exist {pi}i=i,2y...,n C VQ and {Ci}i=it2,...,n C J such that (pi,d), {pi+i,Ci) G J51 for z = 1,2,... , n, wherep n +i = plm Then Ci and Ci-i-i are connected components of K\VQ whose closures contain p^. Let A = \Jj=1(Cj U {pj}). Then for any i, A\{pi} is connected and is contained in one connected component of K\{pi}. Hence, C% and C^+i are contained in the same connected component of K\{pi}. This contradicts Lemma 1.6.11. Since G contain no loop, G is a tree : for any x,y G V, there exists a unique sequence of edges from x to y. Hence G has an end point, i.e., there exists x G V such that #{?/ G V : (#,2/) € E or (y,x) E E} = 1. By Lemma 1.6.10, we see that x G Vfo. Hence m(x,Vo) = 1. Therefore, m (p, Vo) = ^i(x, Vo) = 1 for any p G Vb. On the other hand, if # J > 2, then there exists p G Vb such that m(p, Vo) > 2 because G is connected. Hence # J = 1. •
Notes and references The general references to this chapter are Falconer[32, 33] and Yamaguti et al.[lS4\. 1.1 The notion of self-similar sets could be traced back to the Cantor set, which is the first mathematical example of what are now called "fractals". In [134], Moran considered a class of sets which are extension of the Cantor set and calculated the Hausdorff dimensions of them. His class includes, for example, the Koch curve and the Sierpinski gasket. After Mandelbrot proposed the notion of fractals in [122, 123], Hutchinson[76] formulated the mathematical definition of self-similar sets, which is more restrictive than what we call "self-similar sets" in this book. See the remark after Theorem 1.1.4. Theorem 1.1.4 was essentially obtained in [76]. See also Hata[64] for an extension of Theorem 1.1.4. 1.2 The shift space E and the shift map a are important concepts in dynamical systems. For example, they play an essential role in the study of interval maps. See [133] for example. Theorem 1.2.3 was essentially obtained in [76]. 1.3 Partly motivated by Kameyama[80], the notion of the self-similar structure was introduced in [83]. 1.4 Hutchinson has given the definition of self-similar measures in [76].
Exercises
39
1.5 See Rogers[157] for details on the Hausdorff measures. See also [124] for results from the view point of geometric measure theory. 1.6 Theorem 1.6.2 was obtained in [64]. The results after Proposition 1.6.6 are new.
Exercises n
Exercise 1.1. Let / : R —* Rn be a similitude with Lipschitz constant r. Show that there exist a G Rn and U G O(n) such that f(x) = rUx + a for all x G R n . (Hint: if g(x) = (/(z) - / ( 0 ) ) / r , one can see that \g(x) — g(y)\ = \x-y\. This implies that the natural inner product of Rn is invariant under g. Also one should show that g is a linear map.) Exercise 1.2. Let (X,d) be a complete metric space and let fo : X —» X be a contraction for i = 1,2,... ,7V. For A C X, define F(A) = Ui if (resp. E(5) -• if7) is denoted by n (resp. TT'). Note that both TT and TT' are homeomorphisms. S e t / 3 = /iO7T/o7r-1.
(1) Show that fs is a contraction on if. (2) Let C - (if,{l,2,3},{/i,/ 2 ,/ 3 }). int(C£,K) = 0.
Show that int(C£) ^ 0 and
40
Geometry of Self-Similar Sets
Exercise 1.6. Prove that VC(A) = ^c for any partition A for the selfsimilar structures corresponding to Example 1.2.7, 1.2.8, 1.2.9 in the last section. Exercise 1.7. Let S = {1,2,3}. Set u ~ r if and only if {LJ,T} C {wl212, w3i} for some w e W*(S) or u> = r. (1) Let K = Y,(5)l~ with the quotient topology. Also define Fi : K —> K by Fi(x) = ^ ( ^ ( T T - ^ X ) ) ) for x 6 l Then prove that C = (K, 5, {i7!}^^) is a self-similar structure. (2) Let A = {1,21,22,23,3}. Prove that VC(A) is a proper subset of VcExercise 1.8. Let C = (K,S,{Fi}ies) be a self-similar structure and let A be a partition of E(5). Show that C is post critically finite if and only if £(A) is post critically finite. Exercise 1.9. Evaluate the Hausdorff dimensions of the self-similar sets introduced in Examples 1.2.6-1.2.9 under the Euclidean metric.
2 Analysis on Limits of Networks
In this chapter, we will discuss limits of discrete Laplacians (or equivalently Dirichlet forms) on a increasing sequence of finite sets. The results in this chapter will play a fundamental role in constructing a Laplacian (or equivalently a Dirichlet form) on certain self-similar sets in the next chapter, where we will approximate a self-similar set by an increasing sequence of finite sets and then construct a Laplacian on the self-similar set by taking a limit of the Laplacians on the finite sets. More precisely, we will define a Dirichlet form and a Laplacian on a finite set in 2.1. The key idea is that every Dirichlet form on a finite set can be associated with an electrical network consisting of resistors. Prom such a point of view, we will introduce the important notion of effective resistance. In 2.2, we will study a limit of a "compatible" sequence of Dirichlet forms on increasing finite sets. Roughly speaking, the word "compatible" means that the Dirichlet forms appearing in the sequence induce the same effective resistance on the union of the increasing finite sets. In 2.3 and 2.4, we will present further properties of limits of compatible sequences of Dirichlet forms.
2.1 Dirichlet forms and Laplacians on a finite set In this section, we present some fundamental notions of analysis on a finite set, namely, Dirichlet forms, Laplacians and effective resistance. Notation. For a set V, we define £(V) = {/ : / maps V into R}. If V is a finite set, £(V) is considered to be equipped with the standard inner product (•,•)defined by (u,v) = J2Pev u(p)v(p) f° r a n y v>, v e ^00First we give a definition of Dirichlet forms on a finite set V. In B.3, one 41
42
Analysis on Limits of Networks
can find a definition of Dirichlet forms for general locally compact metric spaces. Definition 2.1.1 (Dirichlet forms). Let V be a finite set. A symmetric bilinear form on £(V), 8 is called a Dirichlet form on V if it satisfies ( D F l ) £(u, u)>0
for any u G t(V),
(DF2) £(u, u) = 0 if and only if u is constant on V and (DF3) For any u G £(V), £(u,u) > £(u,u), where u is defined by
{
1
ifu(p)>l,
u(p)
ifO
0
if w(p) < 0.
We use VT(y) to denote the collection of Dirichlet forms on V. Also V!F{y) is the collection of all symmetric bilinear forms on £(V) with (DFl) and (DF2). Obviously VT{V) C VF(V). Condition (DF3) is called the Markov property. This definition is a special case of Definition B.3.2 when X is a finite set V and the measure μ is the discrete measure on V. Notation. Let V be a finite set. The characteristic function subset U C V is defined by v
.
XU(Q)
fi
=
Xu °f
a
if 9 e^,
., . otherwise.
If no confusion can occur, we write xu instead of \u- If U = {p} for a point p G V, we write x P instead of X{p}- H -^ : ^(^) ~^ ^(V) is a linear map, then we set Hpq = (Hxq){p) for p,q G V. For / G ^(V),
(Hf)(p) = Y,qeVHpgf(q).
Definition 2.1.2 (Laplacians). A symmetric linear operator H : £(V) —> ^(F) is called a Laplacian on F if it satisfies (LI) iJ is non-positive definite, (L2) Hu — 0 if and only if u is a constant on V, and (L3) flp, > 0 for all p ^ q G F. ^_ We use CA(V) to denote the collection of Laplacians on V. Also CA(V) is the collection of symmetric linear operators from £{V) to itself with (LI) and (L2).
2.1 Dirichlet forms and Laplacians on a finite set
43
Obviously CA(V) C CA(V). There is a natural correspondence between V!F(y) and CA(V). For a symmetric linear operator H : £(V) —> £(V), we can define a symmetric quadratic form £H(*, •) on £(V) by £H(U,V) = —(u,Hv) for u, v G £(V). If we write TT(H) = £H, it is easy to see that TT is a bijective mapping between symmetric linear operators and symmetric quadratic forms. This correspondence between Dirichlet forms and non-negative symmetric operators is a special case of the correspondence described in Theorem B.3.4. Proposition 2.1.3. n is a bijective mapping between CA(V) Moreover, ir(CA(V)) = VT{V).
andVTiV).
Proof. It is routine to show IT(CA(V)) = VF{V). To show n(CA(V)) = VTiy\ first note that £H{u,u) = ^EP,qev Hpq(u(p) - u(q))2. By this expression, it is easy to see that 7T(CA(V)) C VT{V). NOW suppose H e CA(V)\CA(Y). So there exist p ^ q e V with Hpq < 0. We can assume that Hpq = — 1 without loss of generality. Set u(p) = x, u(q) = y and u(a) = z for all a G V\{p, q}. Then we have £H(U, U) = a(x — z)2 + P(y — z)2 — (x — y)2. As £H is non-negative definite, a and β should be non-negative. If x = 1, z = 0 and y < 0, then £H{u, u) = a - 1 + 2y + (/3 - l)y 2 and ^H(^> ti) = a — 1. If |y| is small, we have £#(i6, w) < £i/(u, ix). Hence £H £ VTiy). This shows that n(H) G PJ^(F) if and only if H G CA(V). D Example 2.1.4. Let V be a set with three elements, say, pi,P2»P3- Set / - ( l + c) 1 e \ H= 1 -2 1 . Then5 f f («,u) = (;E-2/)2 + ( y - ^ 2 + V e 1 -(1 + e)/ e(a; — 2) 2 , where a; = u(pi),y = u(p2) and 2; = u(jpz). Letting X = x — y and y = y — z, we have £H(u, u) = X2 + Y2 + e(X + Y)2 = (1 + 2e)(X2 + y 2 ) - e(X - Y)2 So it is clear that if e > - | , then H € ^4(1^) and if c > 0, then H £ CA(V). If V is a finite set and H is a Laplacian on V, the pair (V, # ) is called a resistance network (an r-network, for short). In fact, we can relate an rnetwork to an actual electrical network as follows. For an r-network (V, H), we will attach a resistor of resistance rpq = Hpq~1 to the terminals p and q for p, q G V. Also the plus-side of a battery is connected to every terminal p while its minus-side is grounded so that we can put any electrical potential
44
Analysis on Limits of Networks
on each terminal. For a given electric potential v G £(V), the current ipq between p and q is given by ipq = Hpq(v(p) — v(q)). So the total current i(p) from a terminal p to the ground is obtained by i(p) — (Hv)(p). Let (V, if) be an r-network and let U be a proper subset of V. We next discuss the proper way of restricting if onto U from the analytical point of view. Lemma 2.1.5. Let V be a finite set and let U be a proper subset of V. For H G CA(V), we define Tv : £{U) -> £(U),JV : £(U) -» £(V\U) and Xu : £(V\U) - £(V\U) by H
-\Ju
Xv)>
where iJ\j is the transpose matrix of J\j. (When no confusion can occur, we use T, J and X instead ofTjj, Jjj and Xjj.) Then, X — Xu is negative definite and, for any u G £(V), £H(U,U) = £x(ui + X~1Juo,ui
+ X"1JUQ)
+£T-tJX-1j(uo,uo),
(2.1.1) where UQ = u\jj and u\ = u\y\u. Proof. For v € £(V\U), define v by v\jj = 0 and v\v\u = v. Then £x(y,v) = £H(V,V) > 0. By (L2), we see that if £x(v,v) = 0 then v should be a constant on V. This implies v = 0. Hence £x is positive definite. By the definition of £x, X is negative definite. Now (2.1.1) can be obtained by an easy calculation. • Theorem 2.1.6. Assume the same situation as in Lemma 2.1.5. For u e £(U), define h(u) e £(V) by h(u)\v = u and h(u)\V\u = -X~xJu. £H(V, V). Also Then h{u) is the unique element that attains mmve^v^v\u=u 1 define PV,u(H) =T- VX" J. Then, Pv,u : CA(V) -+ CA(U) and £pv,u{H){u,u) = £H(h(u),h(u)) =
min ve£(V),v\u=u
£H{v,v).
(2.1.2)
Moreover, if H G CA(V), then PV,u(H) G CA(U). Proof By (2.1.1), minve^V)}V\u=u £H(VJ V) is attained if and only iiv\v\u+ X~xJu = 0. Hence we have the first part of the theorem. Next we show that PV,u(H) = T- VX" 1 J G £A{V). By (2.1.1), we can verify (2.1.2). Hence, £pvu(H) is non-negative definite. By (2.1.2), £pv,u{H)(v">u) — 0 implies that h(u) is a constant on V and therefore u is a constant on U. Thus we can show that Py^u{H) G CA(U).
2.1 Dirichlet forms and Laplacians on a finite set
45
Finally, if H G CA(V), then £pv,u(H)(u,u) = £H(h(u),h(u)) > £H(h(u),h(u)). As h(u)\u = u, we obtain £H(h(u),h(u)) > £pV!U(H)(u,u). Hence £Pvu^H>} has the Markov property. By Proposition 2.1.3, PV^(H) G CA{U). • The linear operator Pγ^u (H) may be thought of as the proper restriction of H onto U from the viewpoint of electrical circuits. In fact, H and Py,u(H) give exactly the same effective resistance (which will be defined in Definition 2.1.9) on U. When no confusion can occur, we write [H]jj in place of Pv,u(H). Remark. In general, Pvu is not injective. For example, set V = {pi,P2^Ps} / - ( l + e) 1 e\ and U = {pi,p2}- If He = 1 - 1 0 for e > 0, then He €
V
CA(V) and [Ht]v = ("j
1
€
0
-e)
1
J ^.
Note that /i(w) is the unique solution of (Hv)\v\u = 0 a n ( l vlc/ = uTherefore, if we think of U as a boundary of V, /i(tt) is called the harmonic function with boundary value u G -£({/). For a Laplacian if on V, we obtain the following maximum principle for harmonic functions. Proposition 2.1.7 (Maximum principle). Let V be a finite set and let H e CA(V). Also let U be a subset ofV. For p <E V\U, set Up = {q eU : There exist pi,P2,... ,p m € V\U with pi = p such that HPiPi+1 > 0 for i = 1,2,... ,ra— 1 and ifPmQ > 0}. Then if (Hu)\v\u — 0? *ften, /or an?/ p G V\[/, min w(g) < u(p) < maxu(q).
qeUp
qeUp
Moreover, u(p) = maxg€[/p u(q) (or u(p) = mmqejjp u(q)) if and only if u is constant on Up. Proof. For p G V, set Np = {q : Hpq > 0}. Also define Wp = {q G V\17 : there exist pi,P2, • •,Pm • G V\J7 with p\=p and p m = # such that HPiPi+1 > 0 for i = 1,2,... ,ra— 1} and Vp = Wp U C/p. First assume u(p*) = maxqevpu(q) for p* G Wp. Then ^ C Vp and (ffu)(p#) = E gG iv p , Hp^q(u(q) - u(p*)) = 0. Since
46
Analysis on Limits of Networks
Hp*q > 0 and u(q) — u(p*) < 0 for any q G NPm, we have u(q) — u(p*)
for all q G Np#. Iterating this argument, we see that u is constant on Vp. Using the same argument, it follows that if there exists p* G Wp such that u (p*) — mmgeVp U(Q)I then u is constant on Vp. Hence, min u(q) — min u(q) < u(p) < m&xu(q) = maxii(^). The rest of the statement is now obvious.
•
The following corollary of the maximum principle is called the Harnack inequality. Corollary 2.1.8 (Harnack inequality). Let V be a finite set and let H G CA(V). Also let U be a subset of V'. Assume that A C V\U and that Vp = Vq for any p,q G A. Then there exists a positive constant c such that maxu(p) < c min u(p) peA
p£A
for any non-negative u G £(V) with (Hu)\y\u = 0. Proof Let V = Vp for some p G A. (Note that V is independent of the choice of p G A.) Set A = {u : (Hu)V\u = 0,minpev u(p) > 0,maxpGv" u(p) — !}• By Proposition 2.1.7, m.inpeAu(p) > 0 for u £ A. If Ao = {u\v : u G .4.}, then Ao is a compact subset of £(V). Therefore, c = inf{miiip^A u{p) '• u ^ A$} > 0. By the definition of Ao, it follows that c = inf{miupeA u(p) : u G A}. This immediately implies the Harnack inequality. • Next, we define effective resistances associated with a Laplacian or, equivalently, a Dirichlet form. Prom the viewpoint of electrical circuits, the effective resistance between two terminals is the actual resistance considering all the resistors in the circuit. Definition 2.1.9 (Effective resistance). Let V be a finite set and let H G CA(V). For p^qeV, we define RH(p,q) = (mm{£H(u,u) : u G l(V\u{p) = l,u(q) = O})"1.
(2.1.3)
Also we define RH(P,P) = 0 for all p G V. Rn(p,q) is called the effective resistance between p and q with respect to H. By Theorem 2.1.6, if U = {p, q}, then it follows that
2.1 Dirichlet forms and Laplacians on a finite set
47
Definition 2.1.10. Let V{ be a finite set and let H{ e llA(Vi) for i = 1,2. We write (Vi,#i) < (V2,H2) if and only if Vx C y 2 and Pv^Vx(H2) = Hx. The next proposition is obvious by the above definitions. Proposition 2.1.11. Let V* be a finite set and let Hi e CA(Vi) for i = 1,2. //(Vi,ffi) < (V2,H2), then RHlip,q) = RH (P,Q) for anyp,qe V^ 2
In fact, the converse of the above proposition is also true if both H\ and H2 satisfy (L3). This fact is a corollary of the following theorem, which says that a Laplacian is completely determined by its associated effective resistances. Theorem 2.1.12. Let V be a finite set. Suppose HuH2e CA(V). Then Hi = H2 if and only if RHl(p,q) = RH2(P,q) for anyp,qeV. Proof. We need to show the "if" part. We use an induction on #(V). When #(V) = 2, the theorem follows immediately by (2.1.4). Now suppose that the statement holds if # ( V ) < n. Let V = {pi,P2>--- >Pn}« We write hij = (H!)PiPj and Hi5 = (H2)PiPj. Also let V{ = V\{pi} and let
Di^Ti-tjtixir'ji for k = 1,2, where T^ : ^(^) -> ^(VJ), Jj : ^(^) -^ ^(fe» and Xj : ^({pj) -^ t({Pi}) a r e defined by tJl
ft=f^
)
Since (Vi,D{) < (V,Hk), Rpifaq) = RDi(p,q) for all p,q e V^. By the induction hypothesis, D\ = D\. Now define D% = D\ = D\ and dlkl = (D%)PkPl. Calculating directly and then using the fact that h^ — hki w e and Hik = Hku obtain tfsi — hki — hikhii/hii = Hki — HikHii/Hii. In particular, 2
dik = hkk - h ik/hu = Hkk - H?k/HH. 1
(2.1.5)
Exchanging k and i, we see that d^/d ^ = ha/hkk = Ha/Hkk- Therefore, there exists t > 0 such that Hα = thu for i = 1,2,... , N. Again, by (2.1.5), we have (hik) = hkkhii — dlkkhii
48
Analysis on Limits of Networks
and (Hik)2 = HkkHii - d\kHii = t2hkkhii - tdkkhu. As -hkk = J2i:i^k hik
and
~thkk = -Hkk = J2i:i^k H*k, we obtain
-hkk = 22 yhkkhii - d\khii = 22 yhkkhii - dlkkhii/t. iii^k
i-.i^k
As a function of t, the right-hand side of the above equation is monotonically increasing. Hence the above equality holds only for t = 1. Therefore we obtain H\ = H^ • • Corollary 2.1.13. Let Vi be a finite set and let Hi e CA(Vi) for i = 1,2. Then (Vi,#i) < (V2,H2) if and only if RHAP^Q) = R 0. Unfortunately, we don't know whether such an extension is true or not. One reason why effective resistance is important is that it is a metric on V if H e CA(V). This metric, called the effective resistance metric, will play a crucial role in the theory of Laplacians and Dirichlet forms on (post critically finite) self-similar sets. Theorem 2.1.14. Let V be a finite set and let H e CA(V). Then RH(-, •) is a metric on V. This metric RH is called the effective resistance metric on V associated with H. Remark. Not every metric on a finite set V corresponds to an effective resistance metric with respect to a Laplacian H e CA(V). See Exercise 2.1 and Exercise 2.2. We need the following well-known formula about electrical networks to show Theorem 2.1.14. Lemma 2.1.15 (A—Y transform). Let U = {pi,P2,P3} o,nd let V = {Po} U U. Set Rij = H~}p. for H G CA(U), HPiPj > 0. Define
R\2 "I" -^23 H" -R3I
R\2 H~ -R23 H~ R3I
where we assume that
R\2 "I" ^ 2 3 + -^31
A
2.1 Dirichlet forms and Laplacians on a finite set Pi
P2
Pi
Ri T>
K-1/^
PS
R2S
49
^ ^ \
Po
T>
\it3
P2
A-circuit
PS
(upside-down) Y-circuit
Fig. 2.1. A-Y transform
IfH'eCA{V)
is defined by H,
1
=(R-
PiPj
\o
if i = o, othewise,
for i < j , then [Hf]u = H. A direct calculation shows this formula. As we mentioned before, we can associate an actual electrical circuit with a Laplacian. In the above lemma, the circuit associated with H e CA(U) has three terminals {pi,P2,P3} and the terminals pi and pj are connected by a resistor of resistance Rij. Let us call this circuit a A-circuit, which reflects the triangular shape of the circuit. At the same time, the circuit associated with Hf e CA(V) consists of four terminals {po,Pi,P2,P3} and each terminal pi is connected only to po by a resistor of resistance Ri for i = 1,2,3. po is a focal point of the circuit. Let us call this circuit a Ycircuit because of its "upside-down Y" shape. The A-Y transform says that the A-circuit and the Y-circuit are equivalent to each other as electrical networks. See Figure 2.1. Proof of Theorem 2.1.14. Definition 2.1.9 and (2.1.4) imply that #jf(p,g) > 0 and that RH(P> Q) = 0 if and only if p = q. Next we must show the triangle inequality. We may assume that # ( ^ ) > 3. For U = {pi,P2,P3} C V, let H' = [H]u- By Proposition 2.1.11, we have RwipuPj) = RH(PUPJ)> First assume that H'PmPn > 0 for any m ^ n. Then the A-Y transform
50
Analysis on Limits of Networks
shows
1
where Rmn = (-H"pmPn}~ and {z,j, &} = {1,2,3}. Hence we easily see that RH(PUP2)
+ RH(P2,P3)
RH(PUPS)'
f
Next, suppose H piV3 = 0 for instance. Then RH'(PI,P2) = R23 and RH'(PI,P3)
= R12 + ^23, where Rij
= R12, 1
= {H^.)' .
verify the triangle inequality.
RH'(P2,P3) So we
can
•
RH('i •) is not a metric on V for general H e CA(V). In fact, if #(V) > 3, there exists H £ CA(V) such that RH{', ) is not a metric on V. (See Exercise 2.4 and Exercise 2.5.) As we will see, however, ^RH{-, •) is always a metric on F for all if e ^ 4 ( F ) . The following is an alternative expression for the effective resistance. Proposition 2.1.16. Let V be a finite set and let H e *CA(V). Then, for any p,q eV,
RHfaq) = m a x { K ^ 7 ^ ) | 2 : u G l(y),£H(u,u) ± 0)}.
(2.1.7)
Proof. Note that \u(p) — u(q)\2/£H(U,U) = \v(p) — v(q)\2/£H(V,V) if v = OLU^-β for any a ^ e R with a ^ 0 . For given u € £(V) with u(p) ^ u(q), there exist a and /? such that v(p) = 1 and v(q) = 0, where ^ = au + /?. Hence the right-hand side of (2.1.7) equals max{
*
: i; G ^(V),t;(p) = l, V (g) = 0}.
Now (2.1.3) immediately implies (2.1.7).
•
Applying (2.1.7), we can obtain an inequality between \u(p) — u(q)\, Rit(p,q) and £H(U,U). Corollary 2.1.17. Let V be a finite set and let H G CA(V). p,q€V and any u G £(V), \u(p)-u(q)\2
< RH(p,q)£H(u,u).
For any (2.1.8)
This estimate will play an important role when we discuss the limit of a sequence of r-networks in the following sections. As another application of Proposition 2.1.16, we can show that \/RH('I •) is a metric on V.
2.2 Sequence of discrete Laplacians
51
Theorem 2.1.18. Let V be a finite set and let H G CA(V). Set ## / 2 (p, q) = \JRH{PI Q)- Then R^ (•, •) is a metric on V. Proof. We only need to show the triangle inequality. By (2.1.7),
R]i2(p,q) = m a x { K g - ^ ) l
:u
e t(V),eH(u,u) ± 0}.
This immediately implies t h e triangle inequality for R^
(•,•)•
^
2.2 Sequence of discrete Laplacians In this section, we will discuss the limit of r-networks on an increasing sequence of finite sets that satisfy a certain compatible condition, namely: Definition 2.2.1. Let Vm be a finite set and let F m G lCA(V) for each m>0. {(Vm, #m)}m>o is called a compatible sequence if (V^n,i/m) < ( y m + i , i J m + i ) for all m > 0 . Let S = {(V r m ,i7 m )} m > 0 be a compatible sequence. Set V* = Um^oV^ and define JT(S) = {u : u e £(V*), lim eHm{u\VrnJu\vJ
< +oo},
(2.2.1)
m—>oo
£s(u,v) = Jim^SHm(u\vm,v\vJ,
(2.2.2)
for u, v £ J^(<S). Also, for p, q £V*, define the effective resistance associated with S by Rs(p,q) = RHm(P,q),
(2-2.3)
where m is chosen so that p,q G VmIn the next chapter, we will approximate a self-similar set by a sequence of increasing finite sets. Then we will construct Dirichlet forms and Laplacians on the self-similar set by taking a limit of a compatible sequence of r-networks. Throughout this section, S = {(V^n,ifm)}m>o is assumed to be a compatible sequence. Let us regard Vm as a boundary of F*. Then for any u G £{Vm), we consider a minimizing problem of £s(', *) under the fixed boundary value u as follows. Lemma 2.2.2. There exists a linear map hm : £(Vm) —> ^(S) such that, for any u G £(Vm), ^m(^)|vm = u and SHm(u,u)
= £s(hm(u),hm(u))
=
min
vEJr(S),v\vrn=u
£s(v,v).
(2.2.4)
52
Analysis on Limits of Networks
Moreover, if v £ J~(S) with v\ym = u attains the above minimum then v = hm(u). Proof As [fl"n]vm = Hm for n > m, we can apply Theorem 2.1.6 with V = Vn, U = Vm and H = Hn. Set hn,m = h where h is the linear map £(U) -> £(V) defined in Theorem 2.1.6. Then define hm(u)\Vn = ftn,m(w). For any n > ra, this definition is compatible and hm(u) G £(V*) is welldefined. By (2.1.2), we have £Hm K for all n > m. immediately.
w
) = £Hn
Therefore hm(u)
(hm(u)\vn,hm(u)\Vri) 6 ^ ( 5 ) . Also (2.1.2) implies (2.2.4) •
Let us fix m. Then hm(u) is also characterized as the unique solution of
{
(Hnvn)\Vn\Vrn v
\vm
=0
for all n> m,
= u,
where v G t(V*) and vn = v\vn. So hm(u) may be thought of as the harmonic function with boundary values u G Vm. If Hm G £^4(V^) for all m > 0, we can show the following maximum principle for harmonic functions. We will sometimes regard l(Vm) as a subset of T(S) by identifying £(Vm) with hm(£(Vm)) through the injective map hm. With this identification, one can write £um(% u) = £s(u, u) for any u G £(Vm) C ^"(5). Lemma 2.2.3 (Maximum principle). Assume ifm G £.4(1^) /or a// 7n > 0. If v £ £{y*) satisfies (Hnvn)\vn\vm ~ ^ for a^ n > m> where min v(flf) < v(p) < max v(a) w - v ^ ; - qeVrn K*J
qeVm
for any p G V*.
Proof This follows immediately by the maximum principle for harmonic functions on a finite set, Proposition 2.1.7. • Next we discuss the effective resistance lls( , *)• As in the case for finite sets, y/Rs( i') is a metric on V*. Proposition 2.2.4. If Kg (•, •) = y/Rs(m, •)> ^ e n -^5 ^5 a rne^c on V*. Moreover if Hm G £^4(1^) /or aZ/ m > 0 , t/ien i?,s is a metric on V*. Proof This is an easy corollary of Theorem 2.1.14 and Theorem 2.1.18 along with Proposition 2.1.11. •
2.2 Sequence of discrete Laplacians
53
The following lemma follows immediately from its counterpart, Proposition 2.1.16. Lemma 2.2.5. For any p, q G V*, Rsfaq) = (min{£s(u,u) : u G F(S),u(p) = l,u(q) = O})"1 = maxj
1 ^"^1 £s(u,u)
: u G F(S),£s(u,u)
> 0}. (2.2.5)
This lemma implies that \u(p) -u(q)\2 < Rs(p,q)£s(u,u) for any u G T(S)
and p,q e V*. By (2.2.6), f(S)
(2.2.6) C C{V*,R)!2).
For a
metric space (X, d), C(X, d) is the collection of real-valued functions on X that are uniformly continuous on (X, d) and bounded on every bounded subset of (X, d). Next we present important results on the limit of compatible sequences. In the following chapter, the results will be applied in constructing Dirichlet forms and Laplacians on a self-similar set. Theorem 2.2.6.
(1) f(5)cC(K,4 / 2 )-
(2) £s is a non-negative symmetric form onJ-(S). Moreover, £s(u,u) = 0 if and only if u is a constant on V*. (3) Define an equivalence relation ~ on F(S) by letting u ~ v if and only ifu — v is constant on V*. Then £s is a naturally defined positive definite symmetric form on F(S)/~ and (J-(S)/~,£s) is a Hilbert space. (4) Assume that iJ m G CA(V) for all m>0. If u is defined as in (DF3) for any u G F(S), then u G FiS) and £s{u,u) < £s(u,u). Proof. Every statement but (3) follows easily from the results and discussions in this and the previous section. We will use £ and T in place of £s and T{S) respectively. To show (3), first note that £(u,u) = £(v,v) if u rsj v. Hence £ is a well-defined positive definite symmetric form on F/~. Choose any p G V* and set Tv = {u : u G ,F,u(p) = 0}. Then (.77~,£) is naturally isomorphic to (!FP,£). Hence it suffices to show that (Tv, £) is a Hilbert space. Now let {vn}n>o be a Cauchy sequence in (.?>,£) and let v% = hm(vn\Vm). Then by Lemma 2.2.2 £(vF - < \ V ? - vD r
£(vk -vhvk-
vi).
Note that p G Vmfc> sufficiently large m. Hence £ is an inner product on Tv n £(Vm), where £(Vm) is identified with /im(^(Kn))« So there exists
54
Analysis on Limits of Networks
m
m+1
m
m
v G Tv Pi £(Vm) such that v™ -> v as n -> oo. As v | y m = v , there 171 exists v G ^(Vi) such that v\ym = v . On the other hand, let C = s u p n > 0 £(vn, vn). Then we have £(v™, vm) < sup n > m £ « \ O = C. Hence t; G > . Now, we fix e > 0. Then, we can choose n so that £(vn — Vk,vn — Vk) < e for all A: > n. Also, we can choose m so that 171 I ^ K - v,vn - v) - £(y™ - v ^™ - v^1 < 6.
Furthermore, we can choose k so that k > n and \£«
m - V™, < - V?) - £(V™ - V™,V™ - V )\ < 6.
As£(v™-v%l,v™-v%i) <£(vn-vk,vn-vk) < e, wehave £(vn-v,vn-v) < 3e. Thus we have completed the proof of (3). • Finally we show two examples. The first one is related to one of the most basic examples in probability theory. Example 2.2.7 (Simple random walk on Z). Let F m = {i e Z : \i\ < = 1 if \i-j\ = 1, ( # m ) ^ = m} and let Hm e CA(Vm) be defined by (H^ r 0 if \i — j\ > 1. Then S = {(V m ,i7 m )} m >i is a compatible sequence. We can easily see that V* = Z, Rs(i,j) = \i — j\ and £s(u, v) = Y^{u(i + 1) - u(i))(v(i + 1) - v(i)). Also we can see that (^,^"(5) PI L2(Z,//)) is a regular Dirichlet form on L 2 (Z, μ) for every Borel measure /i on Z that satisfies 0 < μ^i}) < oo for all i e Z . (See Definition B.3.2 for the definition of a regular Dirichlet form.) 2 Define a linear operator Aμ on L (Z,/x) by (Apu)(i) = M O '
1
^ + 1) + u{% - 1) - 2u{%)).
Then A^ is a non-positive self-adjoint operator on L 2 (Z, μ). Also Dom(AM) C J-(5)nL 2 (Z,/i)and £s(u,v) = - /
UA^μ
Jz for any u, v G Dom(AM). From this fact, AM is identified as the self-adjoint operator associated with the closed form (£s, F(S) DL2(Z, μ)) on L 2 (Z, /i). (See B.I for more information about closed forms and their associated selfadjoint operators.) Now if u(i) = 1 for all i G Z, then Au is the self-adjoint operator
2.3 Resistance form and resistance metric
55
associated with a simple random walk on Z in the following sense. Let u0 G Dom(AI/) and think about the following evolution equation with discrete time n = 0,1,2,...:
One can easily see that un = (I-\-Au/2)nuo for any n. For i G Z, if wo(^) = 1 and uo(k) = 0 for any k ^ z, then un(j) is the transition probability from i at time 0 to j at time n for the simple random walk on Z. The next example is an extreme case where Rs is the trivial metric on
Example 2.2.8 (Discrete topology). Let Vm = {1,2,... ,ra} and let tfm G £A(V m ) be defined by ( i f m ) ^ = 2/ra for i ^ j. Then 5 = {(y m , # m ) } m > 2 is a compatible sequence. We can easily verify that V* = N and Rs{i,j) = 1 for z ^ j. This metric ife induces the discrete topology on N. As Xi G F{S) for all i e N , (£5, ^ ( 5 ) ) is a regular Dirichlet form on L 2 (N, μ) for every Borel measure /i on N that satisfies 0 < /x({i}) < 00 for all i G N. In particular, if /i({z}) = 1 for all i G N, then L 2 (N,^) = ^ 2 (N). We see that ^ 2 (N) n ^ ( 5 ) = ^ 2 (N) and, for all u, v G ^ 2 (N),
£s(u,v) = 2
uvdμ. JN
2.3 Resistance form and resistance metric In the previous section, we constructed a quadratic form (£s,T(S)) and a metric Rs from a compatible sequence of r-networks <S = {(V^, #m)}m>oIn this section, we will give characterizations of the form (£5, .F(«S)) and the metric Rs and show that there is an one-to-one correspondence between such forms and metrics. First we give a characterization of quadratic forms.
Definition 2.3.1 (Resistance form). Let X be a set. A pair {£,F) is called a resistance form on X if it satisfies the following conditions (RFl) through (RF5). (RFl) J 7 is a linear subspace of £(X) containing constants and £ is a nonnegative symmetric quadratic form on T. £(u,u) = 0 if and only if u is constant on X. (RF2) Let rsj be an equivalent relation on T defined by u ~ v if and only if u — v is constant on X. Then (F/^, £) is a Hilbert space.
56
Analysis on Limits of Networks
(RF3) For any finite subset V C X and for any v G £(V), there exists u G T such that u\y = v. (RF4) For a n y p , g G X , lu{p)
snP{
Ui
- f:u^,S(u,u)>0}
C{U,U)
is finite. The above supremum is denoted by M(p,q). (RF5) If u G T, then u G T and £(iZ, u) < £(u,u), where u is defined in the same way as (DF3) in Definition 2.1.1. We use TZ!F{X) to denote the collection of resistance forms on X. Also we define UT(X) = {(£,?) : {£,T) satisfies the condtions (RFl) through (RF4)}. The condition (RF5) is called the Markov property. Let V be afinitejet. Then (£,£(V)) G nF(V) (or (£,£{V)) G nT(V)) if and only if £ G VT(V) (or £ G VT(V) respectively.) Also, immediately from Theorem 2.2.6, (£s,.F(<S)) belongs to KFiy*) for any compatible sequence S = {(Vm, i^m)}m>o- Moreover, if Hm G CA(Vm) for all m, then {£s,F(S)) is a resistance form on V*. Next we consider a characterization of metrics. Definition 2.3.2 (Resistance metric). Let X be a set. A function R : X x X -^ [0, +oo) is called a resistance metric on X if and only if, for any finite subset V C X, there exists Hγ G £*A(F) such that R\vxv = RHV, where RHV is the effective resistance with respect to Hγ. The collection of resistance metrics on X is denoted by IZM(X). Also we define UM{X) = {R : X x X -> M+ : for any finite subset V C X, there exists ffv € £A(V) with R\VxV
= RHv and fT^ = [HV2]Vl if VI C V2}.
Remark. Recall that [i^v2]vi = ^2,^1(^1^) by definition. Notice that by Corollary 2.1.13, the condition Hy1 = [Hy2]y1 is satisfied for a resistance metric R. If we could extend Theorem 2.1.12 to CA(V), which is quite likely, then we could remove the assumption HY1 = [Hy2]y1 from the definition of UM(X). Since RHV is a metric on V, a resistance metric R is a distance on X. Also, for R G IZM(X), y/R(-,-) is a distance on X. Let V be a finite set.JThen R G UM{V) (or i? G 72^4(V)) if and only if R = RH for some H G £A(V) (or iJ G £«A(V) respectively). Furthermore, it is natural to expect that Rs is a resistance metric for a compatible sequence S. More precisely, we have the following proposition.
2.3 Resistance form and resistance metric
57
Proposition 2.3.3. If S = {(V^,il m )} m >o is a compatible sequence, then Rs G TIM(V*). In particular, if Hm G CA(V) for all m, then Rs is a resistance metric. Proof Let V be a finite subset of V*. Then V C Vm for sufficiently large m. If iJy = [iirm]y, then RHV — -R«s|vxv- The rest of the conditions are obvious. • There is a natural one-to-one correspondence between resistance forms and resistance metrics. First we will construct a resistance metric from a resistance form. Theorem 2.3.4. // (£,.F) G TIT{X), then min{£(u,u) : u G J-,u(p) = l,u(q) = 0} exists for any p,q G X withjpjfi- q. If we define R(p, q)~x to be equal to the minimum value, then R G 71A4(X) and R(p,q) = m a x { | u ( p j ~ " ( g ) | 2 : u € T,£{u,u) > 0}. c(u, u)
(2.3.1)
Moreover, if (S.J7) G TIT{X), then R G TIM(X). To prove the above theorem, we need the following lemma. Lemma 2.3.5. // (£,T) G 11T{X) and V is a finite subset of X, then there exists a linear map hy : £(V) —> T such that hy{u)\y = u and £(hv(u),hv(u))=
min
E(v,v).
(2.3.2)
Furthermore, hy(u) is the unique element that attains thejibove minimum. Also set £v(u1,u2) = £(hv(ui),hv(u2)). Then £v G VTiy). Moreover, 7 v if (£, J ) is a resistance form, then £ G VTiV). Proof For p G V, let Tv = {u : u G T,u\v\{p} = 0}. Then by (RF3), Tv is not trivial. By (RF2), {Tv\S) is a Hilbert space. Define $ p : Tv -^ R by *p(w) = u(p). Then by (RF4) we have, for q G V\{p}, \$p(u)\2 < M(p,q)£(u,u) for all u G Tv. Hence <J>P is a continuous linear functional on (TV,E). Therefore there exists gp G Tv such that for all u G Tv\ £(gp,u) = $p(u) = u(p). Noting that £(gp,gp) = gp(p) > 0, we define ^p
=9p/9p(p)'
Now for any u G £(V), define hy(u) = J2peV u(p)ip^. If v G T with
58
Analysis on Limits of Networks
v\y — u, set v — v — hy(u). Then £ (v, v) = £ (v + hy(u), v + hy (u)) = S(v, v) + 2£(6, ftv(w)) + £(hy(u),
hv(u)).
As £(£, /iv(ii)) = ^Zpey u(p)v(p)/gp(p) = 0, we have £(v,v) = £(v,v) + £(hy(u),hy(u)) > £(hy (u), hy (u)). Equality holds only when v is constant on X and so v = 0 on X. It is easy to see that £v G VF(V). Also the Markov property (RF5) of [£,T) implies the Markov property (DF3) of £v. Thus we have completed the proof. • Proof of Theorem 2.34. By Lemma 2.3.5, min{£(u,u) : u G T, u(p) = l,u(q) = 0} exists for any p,q £ X with p ^ q. Now define Hy G CA(V) by £v = £HV . IfViCF 2 ,then 2
x
ve£(v2)Mv1=u
=
min
v'eJr,v'\v1=u
f^, i;;) = £H 1 (u, u).
Hence [i^y2]vi = Hyx. This fact also implies that RHV — R\vxv- Therefore we see that R G KM(X). If (f,^ 7 ) G HT{X)\ then # y G Pjr(y) and hence R G 7JA<(X). The same argument as in the proof of Proposition 2.1.16 implies (2.3.1). • Theorem 2.3.4 says that each (£, T) G 1Z,T(X) is associated with R G UM(X). So we can define a map F M X : 11T{X) -> 1ZM(X), which is called the "form to metric" map, by R = FMx{{£,F)). This form to metric map is, in fact, bijective: we can construct the inverse of FMx> Theorem 2.3.6. For R G IZM(X), there exists a unique (£.J7) G KJr(X) that satisfies (2.3.1). Moreover, if R G IZM(X), then [£,T) G 11T{X). Assuming the above theorem, we can define the "metric to form" map MFx : UM{X) -> nT(X). It is easy to see that MFX is the inverse of FMX. We will only present the proof of a special case of Theorem 2.3.6, namely Theorem 2.3.7. Theorem 2.3.6 can be proven by using routine and tedious arguments about limiting procedures similar to the special case. If R G nM(X), R1/2(-r) = VR(-r) is a metric on X. Assume that
2.3 Resistance form and resistance metric
59
the metric space (X, R1/2) is separable. Equivalently, there exists a family of finite subsets {Vm}m>o of X that satisfy Vm C Vm+i for m > 0 and K = Um>oVrm is dense in X. Set # m = HVrn. Then iJ m G CA(Vm) and [ffm+ilvm = #m by definition. Hence (Vm,Hm) < (y m + i,iJ m + i) and so S = {(Vm,iJm)}m>o is a compatible sequence. We know that (£s,f(S)) G T£F(V*). Also it is obvious that R = Rs on V*. Now as .F(«S) G C(V*,RJ ) , U E F(S) has a natural extension to a function in C(X, R1/2). We will think of F(S) as a subset of C(X, Rl/2) in this way. Then it is easy to see that (£5,.F(S)) satisfies (RFl) and (RF2). This (£s-,Jr(S)) is the candidate for (£,F) in Theorem 2.3.6. The problem is to show (RF3) and (2.3.1) for any p,q G X. (We already know that (2.3.1) holds for p, q G V* by Lemma 2.2.5.) We do this in the next theorem. Theorem 2.3.7. For R G UM(X), assume that (X,R1/2) is separable. Let {Vm}m>o be a family of finite subsets of X such that Vm C Kn+i for any m>0 and that V* = Um>oKn is dense in X. Set S = {(Vm, iJ m )} m > 0 where Hm = HVm. Then (£S,F(S)) G 11T{X) and R(p,q) = m a x j
1
^"^
1
: u G F(S),£s{u,u)
> 0}
(2.3.3)
for all p, q G X. Moreover, (£s,F(S)) is independent of the choice of {Vm}m>*. Also if Re 1ZM{X), then (£S,F(S)) G UT^X). Before proving the theorem, we need two lemmas. L e m m a 2.3.8. Let (£^) G TZJ-(X) and let {Vm} be a sequence of finite subsets of X such that Vm C Vm+i for m>0 and that V* = Dm>oVm is dense in (X.R1/2), where R - FM X ((£,.F)). If S = { ( F m , ^ m ) } m > 0 where Hm = HVrn, then ( ^ , ^ ( 5 ) ) = (£,T).
Remark. In this lemma, again we regard F(S) as a subset of C(X, R1/2) because R - Rs on V* and ^(5) C C(V*, R)/2). Proof. First we show that ^"(5) C T and £s(u,u) = £(u,u) for u G T(S). Let u G F(S). Set um = hvm(u\vm)i w n e r e ^ym is defined in Lemma 2.3.5. As £Hm(u\vmju\vm) ~ £(um->um), w e obtain £(um,um) < £(um+i,Um+i) < £s(u,u). Now without loss of generality we may assume that u(p) = 0 for some p eVo. (We can just replace u by u — u(p).) Note that (um — un)\vn = 0 for m > n. Then, recalling the definition of hy in the proof of Lemma 2.3.5, it follows that £(um - un,un) = ^2 (um(p) - un(p))u(p)I'gp(p) = 0 pevn
60
Analysis on Limits of Networks
for m> n. Hence £(um — ^n^m - un) = £(urn,urn) - £(un,un) —•0 as ra, n —> oo. Therefore {wm}m>o is a Cauchy sequence in {Tv, £). As (J^, £) is complete by (RF2), there exists u* G Tv such that £(u*—u m,w* — wm) —> 0 as m -^ oo. So £(u*,u*) = limrn^oo£(urn,urn) = £s(u,u). For g G V*, we have I^*(?) -Um(q)\2
R(p,q)£(u* -um,u*
-um).
Letting m —> oo, we obtain u\y^ = it*|v;. As it and w* are continuous on X with respect to i? 1 / 2 , it follows that u — u*. Thus we have shown that u e T and £(u,u) = £s(u,u). Secondly, for u G f , define um exactly the same as before. Then £{um, um) =
min
£(v, v) < £(u, u).
V$LJ~\V Y =U\V
Hence u G ^"(5). Now using the argument of the latter half of this proof, we see that £s('u>, u) = £(u,u). • Lemma 2.3.9. Let (£,T) G UT(Y) and let ( F , ^ 1 / 2 ) be the completion of(Y,R1^2), where R = F M y ( ( f , / ) ) . Then, for anyp,qe Y, R(p,q) = m a x { K p j " ^ ) | 2
: u G F,e(u,u)
> 0}.
Proof First we will show that R(p,q) = s u p { | t x ( ^ " t A ( g ) | 2 : u G .F,£(ti,u) > 0}. (2.3.4) £(u,u) We will denote the right-hand side of (2.3.4) by M(p, q). Choose {pn}, {qn} C Y so that Pn^ P and qn -^ q as n —> oo. Note that, i^,^)
=m
a x
{ M ^ M
: Me
jr,f (M,w) > 0}.
(2.3.5)
Hence, we have \u(pn) — u{qn)\2 < R(pn, qn)£(u,u) for any u G T. Letting n —± oo, we obtain \u(p)—u(q)\ 2 < R(p, q)£(u, u). Hence M(p, q) < R(p, q). Suppose M(p, q) < R(p, q). Then we can choose e so that for all u G T", |w(p) - ^(g)I < (yR(P>q) -
5e)y/£(u,u).
On the other hand, since R(pn,qn) —> R(p,q) as n —> oo, using (2.3.5), there exists {un} such that £[un, un) = 1 and \un(pn) - un(qn)\2 —> i?(p, ) as n —• oo. For sufficiently large n, we have \un(Pn) - un(qn)\ > (yJ~R(p,q) - e)
2.3 Resistance form and resistance metric
61
and R(pn,Pm),R(qn,qm) < e2 for all m > n. Furthermore, we can choose m so that m> n and un(pm) - un{p)\ < e and
\un{qm) - un(q)\ < e.
Now we have
\un(Pn) - Un(qn)\ < \un(pn) - Un(pm)\ + \un(pm) - Un{p)\ + \un(p) - un(q)\ + \un(q) - un(qm)\ + \un(qm) - un(qn)\ < \Un{Pn) - Un(Pm)\ + |^n(<7n) ~ Un(qm)\ + (yR(p,q) - 3c). Hence \un(pn) - un(pm)\ > e or |wn(^n) - un(qm)\ > e. This contradicts the fact that R(pniprn),R(qniqrn) < e2. Therefore we have shown (2.3.4). Now using the same argument as in the proof of Lemma 2.3.5, it follows that there exists \j) £ T such that ip{p) = 1, ip(q) = 0 and ip attains the supremum in (2.3.4). • Proof of Theorem 2.3.7. As we mentioned before, (£s, ^F(S)) (recall that we think of JF(S) as a subset of C ^ i ? 1 / 2 ) ) satisfies (RFl) and (RF2). To show (RF3), set V^ = Vm U V for a finite set V C X. Let H'm = Ev> and let S' = {(V^H'm)}. Then for any u G £(V), there exists v G F(S') such that v\v = u. As (Vm,Hm) < (V^H'J, £Hm{v\vmMvm) £ H^(v\v^,v\v^) < £s'{V)V). Hence ^nim^QoEH^v^Mvm) £s'(v,v). Therefore v G ^(5). This shows (RF3). Next, applying Lemma 2.3.9 to the case that Y = V*, we obtain (2.3.3) because X C F . This implies (RF4). Thus we have shown (£s,JF(<S)) G nF(X). Furthermore, (2.3.3) also implies R = FMx((£s,f ($)))• Let {Um} be a sequence of finite subsets of X that satisfies the same condition as {V^} and let Si be the compatible sequence associated with Um. Then, applying Lemma 2.3.8, we can see that (£s,.F(<S)) = (£s1,Jr(Si)). Hence (Ss^iS)) is independent of the choice of {V^}. Finally, if R G KM(X), then HVm G CA(Vm). Hence ( ^ , ^ ( 5 ) ) has the Markov property. • Using the discussions in this section, we can show another important fact about resistance forms and resistance metrics. If S = {(Fm,i]rm)}rn>o is a compatible sequence, then (£5,.F(S)) G T£F(V*) and Rs G 11M(V*). The space V* is merely a countable set. So if we constructed analytical objects like Laplacians or Dirichlet forms from (^5,^"(«S)), we would end up with analysis on a countable set. That is hardly what we want! One way of overcoming this difficulty is to consider the completion of
62
Analysis on Limits of Networks
V* with respect to the metric Rj . Let (Qs^Rg ) be the completion of (F*,i?£ ). Then (Qs^Rj ) might be an interesting uncountably infinite set. As we mentioned before, F(S) can be naturally regarded as a ). Hence (£s,3~(S)) can be considered as a quadratic subset of C(QS,RJ form on (Qs, RJ ). There is, however, a delicate question about this completion procedure. Is the extended R$ in 1ZAi(Qs)7 Equivalently, do we have (£s,.F(<S)) G 7^^-*(X)? This is not a trivial problem. In fact, this is not true in general. See Exercise 2.7 for a counterexample. Fortunately, if we assume the Markov property, i.e., Hm G CA(Vm) for all m > 0, then it follows that RlJ2 G nM(tts) and (£S,F(S)) G ilF^s) by virtue of the next theorem.
Theorem 2.3.10. Let (S.J7) G HT(X). If (X,R) is the completion of (X, R), where R = FMX((£, J 7 )), then (£, T) G TLT{X) and R G UM(X). Proof (RFl), (RF2) and (RF5) follow immediately. Also (RF4) is an obvious consequence of Lemma 2.3.9. Instead of (RF3), we will show the following (RF3*). (RF3*) For each finite subset V C X and for each p G V, there exists u G T such that u\y — Xp, where Xp i s the characteristic function of the one point set {p}. We use an induction on #(V) to prove (RF3*). If #(V) = 2, say V = {p, q}, then by Lemma 2.3.9, there exists u G T such that u(p) ^ u(q). If we set / = (u- u(q))/(u(p) - u(q)), we have f\v = xPNext suppose (RF3*) holds for #(V) < n. Let V = {pi,P2,-- ,Pn}Then by the induction hypothesis, there exists u G T such that u(pi) = 1 and u(pi) = 0 for i > 3. Case 1. If u(p2) < u(pi), then for some a,/? E M, v = aw + /? satisfies v{p1) = 1 and v(pj) < 0 for j > 2. Define I; as in (DF3) of Definition 2.1.1. Then by the Markov property (RF5), we have v G T. Obviously v\y — xPl • Case 2. If ufa) = w(pi), then choose / G T that satisfies /(pi) > /(P2) and |/(pi)| < 1/2 for alH = 1,2,... , n. For some a, /? G R, v = a(w+/)-h^ has the same properties as v in Case 1. Case 3. If u(pi) < u(p2), then using the same argument as in Case 1, we can find v G T that satisfies v\y = x P2 . Thus if / = u — u{p2)v, then f\v = XpiThus we have shown that (RF3*) holds for #(V) = n.
D
63
2.4 Dirichlet forms and Laplacians on limits of networks 2.4 Dirichlet forms and Laplacians on limits of networks
In the last section, we studied relations among a compatible sequence of r-networks, a resistance form and a resistance metric. In this section, we will take a first step to establish an "analysis" on limits of networks. In particular, we are interested in constructing a counterpart of the Laplacian defined as a differential operator in classical calculus. By the results in the last section, it is reasonable to start from a compatible sequence of r-networks S = {(V m ,# m )}. (We are not concerned with how to obtain a compatible sequence of r-networks in this section.) Then we obtain a resistance form (£,«7r) and a resistance metric R by taking a limit of S. Naturally, the resistance form (£, T) and the resistance metric it! are important elements in our "analysis". However, those are not enough. We need to introduce integration, that is, to introduce a measure on the space. The following is a general result concerning a resistance form, a resistance metric and a measure. Theorem 2.4.1. Let R G UM(X) and suppose that (X.R1^2) is separable. Set (£, J7) — MFx(R). Also let μ be a α-finite Borel measure on (X,^ 1 / 2 ). Define £\(u, v) = £(u,v) + / u{x)v{x)μ(dx) Jx for u, v G L2(X, μ) n T. Then (L2(X, μ) fl T, £\) is a Hilbert space. Moreover, if μ(X) < oo and Jx R(p,p*)μ(dp) < oo for some p* G X, then the 2 2 2 identity map from L (X, μ)C\J• with £\-norm to L (X, μ) with L -norm is a compact operator. Proof. Let {un}n>o be a Cauchy sequence in (L2(X, μ) D T,£\) and let vn = un — un(p) for p G X. Then by (RF2), there exists v G Tv such that £(vn — v, vn — v) —> 0 as n —> oo. By (RF4), we have \vn(q) - v(q)\2 < R(p,q)£(vn -v,vn-
v).
(2.4.1)
Since μ is cr-finite, there exists {if m } m >o such that Km C X is bounded, 0 < //(ifm) < oo and Um>oifm = X. By (2.4.1), we see that vn —> v as h —> oo in L 2 (i^ m ,^|x m )- Also {wn|Km}n>o is a Cauchy sequence in L2(Km^\Krn). As un(p) = (un - vn)\Krn, there exists c G R such that u n(p) - ^ c a s n - > oo. If we let u = v + c, then £ (u — un, u — un) —> 0 as n —> oo. Also, Unlxm -* ^\'Km as n ^ oo in L2(JFCm,//|i<:m). On the other hand, {^n}n>o is a Cauchy sequence in L2(X^) and so there exists w* G L2 (X, μ) such that un —• u* as n —> oo in L2 (X, μ). As
64
Analysis on Limits of Networks u
n
2
2
u*\Km = \Km i L (Krn^\Krn), u = u* in L (X, μ). Hence « n -> u as 2 7 n - > o o in (L (X, μ) n J , £\). 2 Now suppose μ(X) < oo. Let W be a bounded subset of (L (X, μ) D T,£i). If C = s u p u G W £i(u, u), then 2
|u(p)-u((Z)|
(2.4.2)
for all u G £/ and all p, g G X. Let V be a countable dense subset of 1S bounded for any p G V by (2.4.2). Hence, X. Note that {^(p)}new by standard diagonal construction, we can find v G ^(V) and {vn} C ZY satisfying vn(p) —>• v(p) as n —> oc for all p G V. Using (2.4.2), we see that v satisfies (2.4.2) for p , g G 7 . Therefore v extends naturally to a function 1 2 and it satisfies (2.4.2) for all p,q G X as well. For any v G C(X,R / ) p e l , choose {pn} C V so that pn —•> p as n —> ex). Then M P ) - V(P)| < \Vk(p) -Vk(Pn)\ + |Vfc(Pn) - V(pn)\ + |v(pn) ~ v(p)| + |Vfc(pn) - V(pn)\. < 2y/CR(p,pn) Hence vn(p) —> v(p) a s n - ^ o o for all p G X. By (RF4), K^/CCP) - vi(p)) - (vk(p*) - ^ ( p * ) ) | 2 < 5(vfc - ^z,^fc -
vi)R(p,p*).
As f ( u n , t ; n ) < C, the above inequality implies \vk(p) - vi(p)\ < ^/4CR(p,p*) -f |^(p*) - vt(p*)\. Letting / —> oo, we have |v/c(p) — v(p)\2 4Ci?(p,p*) 4-1 for large /c. Since 2 $xR(p,p*)lJL{dp) < oo, it follows that Vk —> v in L (X, μ) as A: —>• oo by Lebesgue's dominated convergence theorem. • Now we have collected enough facts to use some abstract theory from functional analysis. In fact, by Theorem 2.4.1, we can apply the welldeveloped theory of closed forms and self-adjoint operators, which is introduced in Appendix B.I. Theorem 2.4.2. Let R G 11M(X) and suppose that (X.R1/2) is separable. Set (£,F) = MFx(R)' Also let μ be a α-finite Borel measure on (X,R}/ 2 ). Also assume that L2(X^)nJ: is dense in L 2 (X,μ) with respect 2 to the L -norm. Then there exists a non-negative self-adjoint operator H 2 such that Bom(H^2) = T and £{u,v) = (H^u.H^v) for on L (K^) all u,v G T. Moreover, if ^(X) < oo and jx R(p,p*)μ(dp) < oo for some p* G X, then H has compact resolvent. Proof. Set n = L 2 ( X ^ ) , Q(-,-) = £ and Dom(Q) = T. Then Theorem 2.4.1 along with Theorem B.1.6 immediately implies the required results. •
2.4 Dirichlet forms and Laplacians on limits of networks
65
Assume that R G 7ZM(X). Also, in addition to the assumptions of Theorem 2.4.2, let us assume that X is a locally compact metric space. Then (£, T n L 2 (X, μ)) is a Dirichlet form on L2(X^). Moreover, if 2 2 JFnL (X, μ^coix] is dense in C 0 (X), then (£, JF n L (X, //)) is a regular Dirichlet form. (See B.2 for the definition of (regular) Dirichlet forms and Co(X).) In fact, if (S^J7) comes from a regular harmonic structure, which is defined in 3.1, we can verify all the conditions above and get a Dirichlet form and a Laplacian immediately from the theorems in this section. See the next chapter for details. From an abstract point of view, the self-adjoint operator —H should be our Laplacian. However, this abstract construction is too general to study detailed information about our Laplacian. For example, it is quite difficult to get concrete expressions for harmonic functions and Green's function from this abstract definition. So, we also need to construct a Laplacian on a self-similar set in a classical way, specifically as a direct limit of discrete Laplacians. See Chapter 3, in particular 3.7. Example 2.4.3. Let K be any closed subset of R. We can always find an increasing sequence of finite sets {T4n}m>o that satisfy Vm C Vm+i and Um>oKi = K- If Vm = {Pm,i}7=i a n d Pm,i < Pm,t+i for all i, then we define Hm G CA{Vm) by 1
,TT N
j \Pm,i - Prnjl'
(Mm)pm,iPm,j -
i
n
[0
if |i - j | = 1, ^U
•
otherwise,
for i ^ j . Then {(Vm>#m)}m>o is a compatible sequence. Also if i^ is the effective resistance defined on Um>oV^, then R coincides with the restriction of the Euclidean metric. Let (S,^) be the corresponding resistance form and let /x be a α-finite Borel regular measure on K. First note that f\x belongs to T for any piecewise linear function / on R with supp(/) compact. (supp(/) denotes the support of /.) By this fact, it follows that T"Pi I?{K,\L) is dense in I?{K^μ) with respect to the L2-norm. Set T\ = T n I?{K,\i). Then {Z,T\) is a local regular Dirichlet form on This example contains many interesting cases. The most obvious one is 1 the case where K = R. In this case, T\ coincides with i? (R), which is the completion of {u G C 1 (R) : / v!(x)2dx < oo,supp(^) is compact.}
66
Analysis on Limits of Networks 1
with respect to the iJ -norm || • ||i defined by f
2
u {x) )dx. Also £(u,v) — fRuf(x)v'(x)dx. If μ is the Lebesgue measure on R, then the non-negative self-adjoint operator H coincides with the standard — A = 2 2 -d /dx . One of the other interesting cases is the Cantor set. Let K be the Cantor set defined in Example 1.2.6. Let //bea self-similar measure on K. (See 1.4 for the definition of self-similar measures.) Then (£,T) is a local regular 2 Dirichlet form on L (if,/x). By Theorem 2.4.2, we obtain a non-negative self-adjoint operator H. Set Aμ = —H. Then A^ may be thought of as a Laplacian on the Cantor set K. In [39, 40], Fujita studied the spectrum of Aμ and the asymptotic behavior of the associated (generalized) diffusion process on the Cantor set.
Notes and references Most of the contents in this chapter are taken from [87]. New aspects are the introduction of 7XF(-), &A(-), UM(-) and T£F(-). Indeed, in [87], condition (DF3) was forgotten in the definition of the Dirichlet forms. (The author thanks V. Metz for having pointed this out.) This motivated the author to study what would happen without (DF3). The relation between electrical networks and Dirichlet forms (or Laplacians) on graphs is a classical subject. See, for example, Doyle & Snell [31]. 2.1 The operator Pjj,v appearing in Theorem 2.1.6 is known as the trace of Dirichlet forms. In 2.1, we only treat the special case of Dirichlet forms on finite sets. See [43, §6.2] for the definition of the trace of Dirichlet forms for general cases. This operation has been known in various areas. For example, it is called the shorted operator in electrical network theory and is an example of a Shur complement. See Metz [127] and Barlow [6, Remark 4.26] for details.
Exercises Exercise 2.1. Show that every metric on V coincides with an effective resistance metric associated with a Laplacian H £ CA(V) if #(V) = 3.
Exercises
67
Exercise 2.2. Let V = {pi,P2,P3,P4} and let d be a metric on V defined by
{
x if (i,i) = (1,2), 0 if» = i, for some x with 0 < x < 2. Show that there exists a Laplacian H £ CA(V) such that Rfj = d if and only if x < 3/2. Exercise 2.3. Verify that the A-Y transform remains true even if H G Exercise CA(V). 2.4. Show that if #(V) = 3, H G CA(V) if and only if RH(-, •) is a metric on V. Exercise 2.5. Let V = {pi,P2,P3,P4}- For i ^ j , set PxVo
\-e
if (ij) = (1,4),
where e > 0 . Show that if e is sufficiently small, H G CA(V) and it^j(•, •) becomes a metric on V. Exercise 2.6. Let V = {^1,^2,^3}- Define
(
—(1 + m)
l + 2m
—m
1 + 2m -2(1 + 2m) 1 + 2m -m l + 2m - ( 1 + m), Show that RHm(pi,Pj) converges a s m - > o o . Also show that there exists no H G CA(V) such that R = RH, where R(pi,Pj) = limm_+oo RHrn(pi,pj). Exercise 2.7. Let X = {a, 6} U {pm}m>i- Define #(a,&) = 2,iJ(a,p m ) = R(b,Pm) = (l + m)/(l + 2m) and R{pj,pk) = \k - j\/(l + 2fc)(l + 2j). (1) Show that i? G UM{X). (2) Let (X, R1/2) be the completion of (X, J?1/2). Show that R <£ UM(X). (Hint: see Exercise 2.6.)
Construction of Laplacians on P. C. F. Self-Similar Structures
In this chapter, we will construct the analysis associated with Laplacians on connected post critically finite self-similar structures. In this chapter, C = (K,S,{Fi}ies) is a post critically finite (p. c. f. for short) self-similar structure and K is assumed to be connected. (Also in this chapter, we always set £ = {1,2,... , N}.) Recall that a condition for K being connected was given in 1.6. The key idea of constructing a Laplacian (or a Dirichlet form) on K is finding a "self-similar" compatible sequence of r-networks on {V^n}m>o, where Vm = Vm(C) was defined in Lemma 1.3.11. Note that {Vm}m>$ is a monotone increasing sequence of finite sets. We will formulate such a self-similar compatible sequence in 3.1. Once we get such a sequence, we can use the general theory in the last chapter and construct a resistance form (£,F) and a resistance metric R on V*, where V* = Um>oV^. If the closure of V* with respect to the metric R were always identified with K, then we could apply Theorem 2.4.2 and see that (£,J-) is a regular local Dirichlet form on L2(K, μ) for any self-similar measure μ on K. Consequently, we could immediately obtain a Laplacian associated with 2 the Dirichlet form (£,.F) on L (if,//). Unfortunately, as we will see in 3.3, in general, the closure of V* with respect to R is merely a proper subset of K in certain cases. In spite of this difficulty, we will give a condition on the probability measure μ which is sufficient for (£, J-) to be a regular 2 local Dirichlet form on L (K^) in 3.4. As was mentioned in 2.4, there is an abstract way of constructing the Laplacian from a Dirichlet form (£,.?*) on L2(K^). (See B.I for details.) However, we will develop our analysis on a p. c. f. self-similar set K in a 2 2 classical and explicit way similar to that of the ordinary Laplacian d /dx on the unit interval. In 3.5, the Green's function will be given in a constructive manner. In the following sections, we will study some counterparts of 68
3.1 Harmonic structures
69
classical analysis on Euclidean spaces, for example, Green's operator in 3.6 and Gauss-Green's formula in Theorem 3.7.8. Finally, we will define a Laplacian as a scaling limit of discrete Laplacians on Vm in 3.7. Throughout this chapter, d is a metric on K which gives the original topology of K as a compact metric space. Also we write C(K) — C(K, d). Since (K, d) is compact, C(K) is the collection of all real-valued continuous functions on K. 3.1 Harmonic structures In this section, we start constructing Dirichlet forms and Laplacians on K. As was mentioned above, the basic idea is to find a "self-similar" compatible sequence of r-networks on {Vm}m>$ and to take a limit. (Recall that Vm C Vm+i by Lemma 1.3.11.) For any initial D G £A(Vo), we can construct a sequence of self-similar Laplacians i/ m G CA(Vm) as follows. Definition 3.1.1. If D G CA(V0) and r = (ri,r 2 ,... ,rN), where for i G 5, we define £^ G VTiy^) by
for w, v G £(ym),
w h e r e rw = rWl . . . rWrn for w = wiw2
Hm G CA(Ym) is characterized by S^
...wme
n>0
Wm. Also
= £Hm.
It is easy to see that N
u, v) - y
1 - £ ( m ) ( u o Fi, v o Fi)
(3.1.1)
for u,v e £(Vm). Also ffm = J2wewm ^RwDRw, where i ^ : e(ym) -> £(Vb) is defined by i ^ / = / o F^ for w G Wm. We write 5 m = S^ hereafter. Considering (3.1.1), we may regard (ym,Hm) as a self-similar sequence of r-networks. If it is also a compatible sequence, then it is possible to construct a Laplacian on K using the theory in the previous chapter. Definition 3.1.2 (Harmonic structures). (D,r) is called a harmonic structure if and only if {(K n ,iJ m )} m >o is a compatible sequence of rnetworks. Also, a harmonic structure (-D,r) is said to be regular if 0 < n < 1 for all i G S.
70
Construction of Laplacians on P. C. F. Self-Similar Structures
Once we get a harmonic structure, we can use the general framework in Chapter 2 (in particular, Theorem 2.2.6, Theorem 2.4.1 and Theorem 2.4.2) to construct a quadratic form {£, J7) on V* = Um>oV^ and an associated non-negative self-adjoint operator H on L2(17, μ), where ft is the completion of V* under the resistance metric associated with (5, J7) and ^ is a given cr-finite Borel regular measure. This H should be our Laplacian. It seems easy but there remains a "slight" problem: the topology on V* given by the resistance metric may be different from the original topology of K. In such a case, ft does not coincide with K. In fact, we will see in the next section that ft = K if and only if the harmonic structure is regular. Another important problem is whether there exists any harmonic structure on a p. c. f. self-similar structure. By virtue of the self-similar construction of Hm, we can simplify the condition for harmonic structures as follows. Proposition 3.1.3. (D, r) is a harmonic structure if and only if(V0, D) < Proof Assume that (ym-i,Hm-i) < (Vm,Hm). Then, for any u G we\mve£m-i(uoFi,uoFi) = min{£m(voFi,voF») : v G £(Vm+i), v\Vm u}. Hence, by (3.1.1), £m{u,u) = min{f m+ i(v,^) : v G i(Vm+1),v\Vm u). Therefore (V^,# m ) < ( V ^ + i ^ + i ) . So, by induction, if (V0,D) (Vijjffi), then (Vm,Hm) < (Vm+i,Hm+i) for any m > 0 . The converse obvious.
= = < is •
For given r = (n, r 2 ,... , rN), define Ur : CA(V0) -+ CA(V0) by
where Hi G CA(Vi) is given by Definition 3.1.1. 1Zr is called a renormalization operator on £A(Vo). By the above proposition, D is a harmonic structure if and only if D is a fixed point of 1Zr. Also, it is easy to see that 1Z\r(aD) — (\)~la1Zr(D) for any a, A > 0. Hence, if D is an eigenvector of 7£r, i.e.,7^(D) = AD, then D is a fixed point of TZ\r. So, the existence problem of harmonic structures is reduced to a fixed point problem (or eigenvalue problem) for the non-linear homogeneous map lZr. In general, this problem is not easy and we do not fully understand the situation yet. For example, it is not known whether there exists at least one harmonic structure on a p. c. f. self-similar set. The only general result on existence of a harmonic structure is the theory of nested fractals by Lindstr0m [116]. Nested fractals are highly symmetric self-similar structures. (See 3.8 for
3.1 Harmonic structures
71
the definition.) We will present a slightly extended version of Lindstr0m's result on the existence of a harmonic structure on nested fractals in 3.8. Example 3.1.4 (Interval). Set Fx(x) = x/2 and F2(x) = x/2 + 1/2. Then C = ([0,1], {1,2}, {i^, F2}) is a p. c. f. self-similar structure. We see that Vm = {i/2rn}i=o 1 2m. Let us define D e CA(V0) by
Then (D, r) is a harmonic structure on C if r = (7*1, r 2 ) satisfies r\ -f r 2 = 1 and 0 < r^ < 1 for i = 1, 2. Also, it is easy to see that those are all the harmonic structures on C Example 3.1.5 (Sierpinski gasket). Recall Examples 1.2.8 and 1.3.15. The Sierpinski gasket is a p. c. f. self-similar set with Vb = {pi,P2>P3}- Define D e CA(Vo) by D= Also set r = (3/5,3/5,3/5). Then we see that (D, r) is a harmonic structure on the Sierpinski gasket. (D, r) is called the standard harmonic structure on the Sierpinski gasket. There are other harmonic structures on the Sierpinski gasket if we weaken the symmetry. See Exercise 3.1. Example 3.1.6 (Hata's tree-like set). Let C be the self-similar structure associated with Hata's tree-like set appearing in Examples 1.2.9 and 1.3.16. Then Vb = {c, 0,1} as in the previous example. Define D € CA(Vo) by D= and define r = (r, 1 — r 2 ) for r G (0,1). If rh = 1, then (Z), r) is a regular harmonic structure on C. So far we have presented examples of regular harmonic structure. Of course, there are many examples of non-regular harmonic structures. Example 3.1.7. Set Fx(z) = z/2,F2(z) = z/2 + 1/2 and F3(z) = y/^ +1/2. Let K be the self-similar set with respect to {Fi,F 2 ,F 3 } and let C = (K, {1,2,3}, {Fi, F 2 , F3}). Then £ is a p. c. f. self-similar structure. In
72
Construction of Laplacians on P. C. F. Self-Similar
Structures
fact, Cc = { 1 2 , 2 i , 3 i } , P £ = {i,2} and Vo = {0,1}. If D = and r = (r, 1 — r, s) for r € (0,1) and s > 0 , then (D, r) is a harmonic structure on £. Obviously, (D,r) is not regular for s > 1. See Example 3.2.3 and Exercise 3.2 for a more natural example of nonregular harmonic structures. Proposition 3.1.8. For w G W*, let w denote the periodic sequence in S defined by w = www . . . . Let (D, r) be a harmonic structure and let w G P /or w G VF*. XTien r^ < 1. /n particular, there exists i £ S such that
n<1. Remark. If (if, 5, {jFi}ies) is a p. c. f. self-similar structure, then the post critical set P consists of eventually periodic points: for any uj e P , there exist w G W* and m > 0 such that o~mu) = w. Corollary 3.1.9. Let (D,r) be a harmonic structure. If r± = • ••= rjy, then Ti < 1 for any i G S. In particular, (D, r) is a regular harmonic structure. To prove Proposition 3.1.8, we need the following lemma. Lemma 3.1.10. Let V be a finite set and let H G CA(V). Suppose U C V and p G U. If there exists q* G V\U such that Hvq^ ^ 0 , then —h pp < -Hpp, where (hki)k,ieu = [H]uProof. H can be expressed as (
), where T : £(U) -> i(U),J
:
\J XJ £(U) -> e(V\U) and X : £(V\U) -+ i(V\U). Then [i7]t/ = T - VX" 1 J. Now let ^p = - X " 1 Jxjf, where Xp(x) = 1 if a; = p and x^(#) = 0 if x ^ p on [/. It follows that
qev\u As Hpq^ T^ 0, the maximum principle (Proposition 2.1.7) implies that
Vvte*) > 0. Therefore Eqev\u
Hpq^P{q) > 0.
D
Proof of Proposition 3.1.8. First we assume that #CPi(Vb) n Vo) < 1 for all i G 5.
As # i = E i l i £*/WW*i, we have (H^
(3.1.2)
= E^o^Vo^fa)^ ^ ^ « -
Set
p = 7r(ii;), where it; = WiW2 . . . w m G W* and w e V. By Lemma 1.3.14, TT-^TT^)) = {w}. Hence, {(^,i) : ^ G Vb,*i(tf) = P} = {(7T(CT^),K;I)}.
3.2 Harmonic functions
73
Therefore, (H\)pp = ^^Dqq, where q = 7r(aw). So, letting pi = 7r(al~1w) for i = 1,2,... ,ra+ 1 , then (Hi)PiPi = ^DPi+lPi+1, where we set wm+i = w\. Now by (3.1.2), we can apply Lemma 3.1.10 and obtain — DPi+lPi+1 < -{Hi)Pi+lPi+1. So we have II™ i ~ ( # i W i < r^~1112=1 -(Hi)Pl+lPi+1. Hence 7\y < 1. If (3.1.2) is not satisfied, we replace the original self-similar structure
C = (K,S,{Fi}ieS)
by Cm = (K, Wm, {Fw}weWJ.
Then, by Propo-
sition 1.3.12, Vc — T^cm' Also, it is easy to see that (Z},r m ), where r m = (rw)weWm, is a harmonic structure on £m. For sufficiently large m, Cm satisfies (3.1.2) and hence we can apply the above argument to the harmonic structure (JD, r m ) . Therefore ( r ^ ) m < 1. Thus we obtain that
rw < 1.
•
3.2 Harmonic functions Let (D, r) be a harmonic structure on a connected p. c. f. self-similar structure C = (K, S, {F,} i e 5 ), where S = {1,2,... , N}. Then {(Vm, ^ m )}m>o is a compatible sequence of r-networks. So we can construct (£,T) as in (2.2.1) and (2.2.2). By Theorem 2.2.6, (E,F) G nF(V*), where K = Um>oKn- In this section, we consider harmonic functions associated with (8,^). The arguments in the last chapter, in particular Lemma 2.2.2, imply the following result. Proposition 3.2.1. For any p G £(Vo), there exists a unique u G T such that u\y0 = p and £(u,u) = min{£(v,v) : v G .T7, v|y0 — p}. Furthermore, u is the unique solution of j(Hrnv)\Vm\Vo
=0
for all m>1,
<
{
v\v0
[6.Z.I)
= p.
The function u obtained in the above theorem is called a harmonic function with boundary value p. Corollary 3.2.2. Let u be a harmonic function. and any m>0, (Hmu)(p) = (Du)(p).
Then, for any p G Vo
Proof. By Lemma 2.2.2, £(u, u) — £Hrn (u, u) for any m > 0 . Hence, if v is a harmonic function, then £(u, v) — £um (uiv) f° r a n y m> 0- Now let v\y0 = Xp- Then £(u,v) = £Hrn(u,v) = - 52q£Vm v(q)(Hmu)(q) = (Hmu)(p). • Example 3.2.3. Let {pi,P2,P3}
De
the vertices of an equilateral triangle
74
Construction of Laplacians on P. C. F. Self-Similar Structures
in C and let
^ +)
^ +)
( +)
( +
Let 5 - {1,2,... ,7}. For z G S, let #(*) = A(z - p{) + ph where ft = ft = ft = ft ft = & = ft = 1 - 2/?, /?7 = 1 - W for 1/3 < /? < 1/2. If if is the self-similar set with respect to {Fi}ies, then the self-similar structure (if, S, {F^ j ^ 5 ) is independent of the value of P and is post critically finite. In fact 3
C= [J {Ik, (k + 3)k} k=1
U
(J
{Jfe/,raifc,raZ,/fc},
k
P = {i,2,3} and pk = n(k) for k = 1,2,3. Define qk = n(7k) = n((k + 3)fe) for jfe = 1,2,3. Also let
and r = (1,1,1,1,1,1,^) for £ > 0. Considering the symmetry of (D,r) and if, we see that 1Zr(D) = XD for some A. (Note that 1Zr preserves the symmetry of D.) So (D, Ar) is a harmonic structure. Now let us calculate the value of A by using Corollary 3.2.2. Let u be the harmonic function that satisfies u(p2) = ^(^3) = 0 and u(pi) = 1. (Note that the harmonic function u is independent of the value of A because it is the solution of (3.2.1).) Taking the symmetry into account, we deduce that u(qie) = u(q15),u(q2e) = u(q^),u(q2) = u(q^),u(q2^) = u(q34). By (3.2.1),
t+3 By Corollary 3.2.2, (Dw)(pi) = - 2 = A - ^ ^ i e J - l ) = (ffiu)(pi). Therefore, we obtain A = (3t + 7)/(7t + 15). Note that 0 < A < 1 for any t > 0. However, At > 1 if t > VE. So the harmonic structure (J9, Ar) is regular if and only if 0 < t < y/E. Note that (K,S,{Fi}ies) is an affine nested fractal. (See 3.8 for the definition of an affine nested fractal.)
3.2 Harmonic functions
75
Fig. 3.1. Pi,qi and ^ Let R be the resistance metric on V* associated with (S.J7). Then, by (2.2.6), u G C(K,-#). Note that V* is a countable subset of K and the topology of ("14, i?) may be different from that of V* with the relative topology from the original metric on K. We will see, however, that a harmonic function has a unique extension to a continuous function on K. Now recall that d is a metric on K which is compatible with the original topology of K. Theorem 3.2.4. Letu be a harmonic function. Then there exists a unique u e C(K) such that u\y^ = u\v+ Remark. As is shown in 3.3, the closure of (V*,R) equals K with the original topology if and only if (£), r) is regular harmonic structure. In such a case, the above theorem is obvious. Proof. Let u b e a harmonic function with boundary value p. Set H\ = T
J \ where T : i(V0) -> e(V0),J : £(V0) -> J X, ^(VA^o) -> ^(^A^o). Then it follows that
(uoFi)\Vo=Ri(u\Vl)
((V^VQ)
and X :
=R
As Fw is a bijective mapping between Vb and Fw(VQ) for to £ If,, we will
76
Construction of Laplacians on P. C. F. Self-Similar
Structures
identify £(V0) and £(FW(VO)) through Fw. Define a linear map A{ : £(V0) e(Fi(Vo)) = i(Vo) by
Then u\Fw(Vo)
= AWmAWm_1
. . . AWlp
(3.2.3)
for w = W\W2 •..Wm G W* and (^)pQ > 0 for any p, g G Vb and ^
:
=
:
Ww
.
(3.2.4)
First we prove the theorem assuming (3.1.2). Claim 1. Set v(f) = maxp,g€Vb |/(p) - f(q)\ for / G ^(F o ). Then v(AJ) < Proof of Claim 1: If fi"i^|vi\vb = 0 an< ^ fl'lvb — /> then Aif — g\Fi(y0)Applying the maximum principle (Proposition 2.1.7) and taking (3.1.2) into account, we can see that max geFi(Vb) 9(Q) ~ mmgGFi(Vb) 9(o) < v(f)Hence v(Aif)
/ e e(Yo).
Proof of Claim 2: Define Q : ^(F o) -> ^(^o) by
then v(/) = v(Q/) and v ( ^ / ) = v(A{Qf) ^
^
for any / G £(V0). Hence
: / G €(Vb),t;(/) ^ 0} = s u p { ^ ^ : / G £(Vb),t;(/) ^ 0}
f : / G <(%), J]) /(g) = 0,t;(/) = 1}. As {/ G ^(Vb) : Y^qzvo /(^) ~ ^' 'L'(^) = -^} ^s c o m P a c t ? t n e above supremum is less than 1. Now by Claim 2 and the maximum principle, vw(u) = sup{|/(p) - /((/)| :p,qeKwnV.}<
cmv(p)
for any it; G W*, where c = m a x ^ s Q . Hence, if {pi}i>i is a Cauchy sequence with respect to a metric on K which is compatible with the original
3.2 Harmonic functions
77
topology of K, then {u(pi)}i>i is convergent as i —• oo. Using this limit, we can extend u t o a continuous function u on K. Next if (3.1.2) is not satisfied, we can exchange the harmonic structure as in the proof of Proposition 3.1.8. Then again we can use the result under a new self-similar structure £ m and harmonic structure (D,r m ). Note that harmonic functions remain the same after we replace the harmonic structure. • Hereafter, we identify u with its extension u and think of a harmonic function as a continuous function on K. Immediately, by Lemma 2.2.3, we have the maximum principle for harmonic functions. Theorem 3.2.5 (Maximum principle: weak version). Let u be harmonic. Then, for any w G W*, min
u(p) < u(x) <
max
u(p)
for any x e Kw. This maximum principle is a little weak because we say nothing about the case when u(x) attains the maximum. We will give a stronger version of the maximum principle in Theorem 3.2.14. Example 3.2.6 (Sierpinski gasket). Let us calculate the stochastic matrices {Ai}ies for the standard harmonic structure on the Sierpinski gasket given in Example 3.1.5. Recall that Vb = {pi,P2,P3}- See Figure 1.2. It follows that Vi = {Pi,qi}ies, where S = {1,2,3}. Now let f(p1) = a, f(jp2) — b and f(ps) = c and solve the linear equation (Hif)(qi) = 0 for i e S. Then we get f(q1) = (2b + 2c + a)/5, f(q2) = (2c + 2a + 6)/5 and f(q3) = (2a + 26 + c)/5. By this result,
It is easy to see that the eigenvalues of A{ are 1,3/5,1/5. Note that the second eigenvalue 3/5 is equal to r^. In fact, this is not a coincidence. (Recall that r = (3/5,3/5,3/5).) In A.I, we will give a general result on the second eigenvalue of Ai. See Exercise 3.3 for more examples. The stochastic matrices {Ai}ies determine the harmonic functions through (3.2.3). The behavior of a harmonic function around a point TT(UJ) for u> £ ^(S) is given by the asymptotic behavior of (3.2.3) for m —> oo.
78
Construction of Laplacians on P. C. F. Self-Similar
Structures
This is the problem of random iterations of matrices and, in general, it is very difficult. Even in the above example, we do not know how to calculate the behavior of Au>rnAUJrn_1 . . . AUl as m —> oo unless the sequence u) is (eventually) periodic. Kusuoka used {Ai}ies to construct Dirichlet forms on finitely ramified self-similar sets in Kusuoka [107] and obtain some result about almost sure behavior of the random iteration of {Ai}iesAn important property of harmonic functions is the Harnack inequality, which follows from the discrete version, Corollary 2.1.8. Proposition 3.2.7 (Harnack inequality). If X is a compact subset of K that is contained in a connected component of K\VQ, then there exists a constant c>0 such that ma,xX£x u(x) < cminxex u(x) for any nonnegative harmonic function u on K. Proof Set Xm = U w e ^ m : ^ n x / 0 ^ - Then we can choose m so that Xm n Vo = 0. Now set V = Vm, U = Vo, H = Hm and A = Xm n Vm. The Harnack inequality (Corollary 2.1.8) implies that there exists c > 0 such that maxpGA u(jp) < cminpeA u(p) for any non-negative harmonic function u on K. By using the maximum principle, it follows that max x €x u(x) < maXpG,4 u(p) and min p €^ u(p) < min a ; e x^(^)- Hence we have shown the required inequality. • For p G Vo, let i^v be the harmonic function that satisfies ipp\v0 — X^0• It is easy to see that {ipp}Pev0 is a partition of unity on K: J2peVo ipP(x) = 1 for any x G K. Theorem 3.2.8. Let p G Vo and let x G K\V0. Then ipp{x) > 0 if and only if there exists C G J(p, Vo) which contains x. Recall that J(p,Vo) is the collection of all connected components of K\Vo whose closure contains p. Definition 3.2.9. (1) For p,q G Vm, {pi}f=i is called an iJ m -path between p and q if Pi G Vm\Vo for any i = 1,2,... ,n, (Hrn)PPl > 0, (Hm)Pnq > 0 and (Hm)PiPi+1 > 0 for any i = 1,2,... , n - 1. (2) For p, q G V^\Vo, we write p ~ q if and only if there exists an i7 m -path between p and q.
771
The notion ~ will not appear later in this section. It plays, however, an m
important role in 3.5, where we will show that p ~ q if and only if p and q m
belong to the same connected component of K\Vo by using the results in this section.
3.2 Harmonic functions
79
Note that (Hm)pq = T,W:P,qeFw(v0)(r™)~ lD(Fw)-Hp)(Fw)-Hq)Since D is a Laplacian on Vb, we may immediately verify the following lemma. Lemma 3.2.10. Let w G Wm and letp,q G FW(VO). (1) / / Fw(Vo) n Vo = 0, then there exists an H^-path between p and q contained in Fw(Vo). (2) If Fw(Vo) nVo = {q} and p ^ q, then there exists an H^-path between p and q contained in Fw(Vo). Proof of Theorem 3.2.8. First assume that x G C G J(p,V0). Then, by Proposition 1.6.6, there exists a path 7 : [0,1] —> C U {p} between x and p. Choose m so that sup diam(if m f 7 ( t ) )
i€[0,l]
where dist(A,JB) = mixeA,yeB d(x,y). Let { p i , . . . ,p n } = ^ m ri7([0,l]), where pi = p(U) and £1 < £2 < • •<•tn = 1. Note that p n = p. Then there exist w(0),... , i y ( n - 1 ) G W m such that x,p1 G ^ ( o ) and p»,Pi+i € ^ ( t ) for all i = 1,2,... ,n — 1. Note that Kw^ C\ Vo C {p} for any z. Let j = min{i : K^ ( i ) n Vo = {p}}. Suppose i^p{x) = 0. Then, by Theorem 3.2.5, there exists q G FW^(VQ) such that ipp(q) = 0. Let g = p$. Then, by Lemma 3.2.10, there exists an # m - p a t h between pi and pi+\ for any i = 0 , 1 , . . . , j — 1. Also, by Lemma 3.2.10, there exists an i7 m -path between pj and p. Joining those i7 m -paths, we may construct an if m -path between q and p. Let V — V m and let U = Vo. Then the maximum principle, Proposition 2.1.7, implies that i/;p(q) > 0. Therefore, if>p(x) > 0. Next let C be the connected component of K\Vo. Assume that p £ C. We show that ipp(q) = 0 for any q G C C\ V*. Let q G C C\ Vm- Define Um = U2,EVbnc(^™>2/ n Vrn)' Since ipp is continuous and i/>p(y) = 0 for any p G Vb n C, it follows that maxzGj7m i/)p(z) —> 0 as m —> CXD. Also g ^ C/m and p £ Um for sufficiently large m. Now apply the maximum principle, Proposition 2.1.7 with V = F m , t/ = Um U Vb and if = ffm. Since p £ C, any i7 m -path between p and g must intersect Um. Hence Uq C f/m, where f/q is defined in Proposition 2.1.7. So 0 < ipp(q) < max z6 [/ m i^p(z). Letting m —»• 00, we see that ^p(q) = 0. Since CflF* is dense in C, it follows that ^p\c = 0. • By the above theorem, we get a relation between the topological properties of K and the positivity of the elements of D, where (D,r) is the harmonic structure.
80
Construction of Laplacians on P. C. F. Self-Similar
Theorem 3.2.11. Let (D,r) be a harmonic structure. and only if J(p, Vo) D J(q, Vo) ^ 0.
Structures Then Dpq > 0 if
The condition that J(p, Vo) n J(q, Vo) ^ 0 is equivalent to the one that there exists a connected component of K\Vo whose closure contains both p and q. Proof: Parti. Let (£,T) be the resistance form associated with (D,r). Then it follows that Dpq = -£o(ipp\v0,ipq\v0) = -£m(^P\vm^q\vm)' Hence Dpq = (Hm^p)(q)
= (Hm^q)(p).
(3.2.5)
Assume that J(p,Vo) D J(q,Vo) — 0. Since Kmjq C Ucej(g,vo)Cf f° r sur E~ ciently large ra, Theorem 3.2.8 implies that ipp\Kmjq = 0. Hence, by (3.2.5), Dpq = 0. ' • To prove the rest of this theorem, we need the following lemmas. L e m m a 3.2.12. Let p and q belong to Vo. If Dpq > 0, then there exists a path 7 : [0,1] —> K between p and q that satisfies 7((0,1)) C K\VoProof. By the above proof of Theorem 3.2.11, if Dpq > 0, then J(p, Vo) H J(g 5 Vo) T^ 0- Let C G J(p, Vo)n J(q, Vo). Then, by Proposition 1.6.6, there exists a path 7 : [0,1] —> C U {p, g} with 7(0) = p and 7(1) = q. D Lemma 3.2.13. Let p and q belong to Vm. Let {pi}i=i,2,... ,n be an Hmpath between p and q. Then there exists a connected component of K\Vo, C, such that pi € C for all i = 1,2,... ,n, p e C and q e C. Proof. For each i G {1, 2 , . . . , n — 1}, there exists w G Wm such that Pi,Pi+i £ Kw and ^(Fu; )-1 (pi),(Ft/; )-1 (pi+ i) > 0- By Lemma 3.2.12, it follows that there exists a path 7^ between pi and Pi+i that satisfies 7i((0,1)) C i;f\T4n. Hence there exists a connected component of K\Vo which contains both pi and Pi+i. Inductively, we see that the connected component, say C, contains all pi. By the same argument as pi and p^+i, we may find a path 7 between p and pi that satisfies 7((0,1)) C K\Vm. So it follows that 7((0,1)) C C. Hence p G C. Exactly the same argument implies that qeC as well. • Part II of the proof of Theorem 3.2.11. Assume that J(p, Vo) H J(q, Vo) ^ 0. Let C G J(p,Vo) n J(q,Vo). Then, by Proposition 1.6.6, there exists a path 7 : [0,1] —> C U {p,^} with 7(0) = p and 7(1) = q. Now by the same argument as in the proof of Theorem 3.2.8, we may construct an iJ m -path between p and q from the continuous path 7. Let {pi}f:=1 be the iJ m -path between p and q. Then, by Lemma 3.2.13, pi belongs to C for
3.2 Harmonic functions
81
alH = 1,2,... ,ra. Since (Hm)Priq > 0 and ipp(pn) > 0, (3.2.5) implies that Dpq = (Hmil>p)(q) > (Hm)Pnq(^p(pn) - i/>p(q)) > 0. • By the above discussion, we also obtain a stronger version of the maximum principle.
Theorem 3.2.14 (Maximum principle: strong version). Let x G K\Vo and let C be the connected component of K\Vo with x G C. Then, for any harmonic function u, min_u(p) pevonc
< u(x) <
ma,x_u(p). pevonc
Moreover, if any of the equalities holds, then u is constant on C. Proof First assume that x G Vm for some m. Then the same arguments as in the proof of Theorem 3.2.8 imply that, for any p G C fl Vb, there exists an i/ m -path between x and p. Conversely, by Lemma 3.2.13, if there exists an i/ m -path between x and p G Vb, then p G C DVo- Hence, let V = Vm, U = Vo and H = Hm and apply Proposition 2.1.7. Then Ux = C n Vb. Therefore, the desired results are immediate from the conclusion of Proposition 2.1.7. Next if x £ V*, then choose w G W* so that x G Kw C C. Then, by the weak version of the maximum principle (Theorem 3.2.5),
min u(q) < u(x) < max u(q). F(V)
eF(V)
Also, if u(x) = max 9eFu; (y 0 ) u(q), then u(x) = u(q) for some q G FW(VO). Since FW(VQ) C C fl Vm, where m = \w\, the first part of this proof implies the desired results for this case as well. • Next we define the notion of piecewise harmonic functions. The following is an immediate corollary of Theorem 3.2.4. Corollary 3.2.15. For p G £(Vm), there exists a unique continuous function u on K such that ^|y m = p and £(u\v*,u\v+) = min{£'(v,^) : v G u in the above corollary is called an m-harmonic function with boundary value p. Another characterization of ra-harmonic functions is that u is an ra-harmonic function if and only if uoFw is a harmonic function for any w G Wm. For p G Vm, define 1$)™ to be the ra-harmonic function with boundary value xjfm. Then any ra-harmonic function u is a linear combination of {il>™}. In fact, u = £ p e Vm
82
Construction of Laplacians on P. C. F. Self-Similar •If
Structures
We Write U
m = PrnU, t h e n £ m ( l i | Vm , u\ Vrn) =
In the rest of this section, we will give an expansion of u G £(V*) in a piecewise harmonic basis {V^lpGV*, where ipp = ip™ if p G V m \ y m _ i . Lemma 3.2.16. Let u be an m-harmonic function. Then £(u,f) — 0 for
feFiff\vm=0. Proof. For n > m, we have (Hnu)(p) = 0 if p € Vn\Vm and /(p) = 0 if peVm. Hence £n(«, / ) = - £ p e V n /(p)(ffn«)(p) = 0. • L e m m a 3.2.17. For u G J-, £(u — um,u — um) —>• 0 as m —>• oo? where Proof By Lemma 3.2.16, £(u — w m , i^m) = 0. Hence, £(iz — ixm, w — ixm) = S(u, u) - £{um, Um) = £{u, u) - £m{u, u). n Definition 3.2.18. Let u G £(V*). For p G Vrm\Vrm_i, define a p (i/) = Um(p) - um-i(p)' {pLp{u) = u(p) if p G Vo-) Also define a w (w) = ix m+ i o Fw - Um o F w for ^ G W m . Note that ap(u) — u(p) — um-i(p) definition,
for p G Vm\Vm-i-
By the above
where the infinite sum in the latter equality is pointwise convergent. (In fact, it is a finite sum for each p G V*.) This is the expansion of a function on V* in the piecewise basis {'0p}pev*Also, by Definition 3.2.18, aw(u) is a 1-harmonic function and aw{u)\yQ = 0. Using ap(u), aw{u) is given by a
Fw(p)(u)ipp.
(3.2.6)
Proposition 3.2.19. For u G £{V*), u G T if and only if £o(u,u) -f Y^ Moreover, ifuEJ7,
^2
—£i(a>w(u),a w(u))
< oc.
then the sum in (3.2.7) is equal to £(u,u).
Proof By Lemma 3.2.16, m-l
£m(u,u) = ^2 £k+i(uk+i -uk,uk+i k=0
-uk)
+Eo(u,u).
(3.2.7)
83
3.3 Topology given by effective resistance By the self-similarity of £k+i, we obtain
—£i(a w (u),a w (u)). weWk
Combining the above two equalities, we easily verify the statement of the proposition. •
3.3 Topology given by effective resistance As in the last section, we assume that (D,r) is a harmonic structure on C = (K, S, {Fi}ieS), where S = {1,2,... , N}. Let {(Vm, # m )}m>o be the compatible sequence induced by (D, r). Then, there exist (£, J7) G 1ZJ-(V*) and R G KM(V*) associated with { ( F m , # m ) } m > o . Moreover, if (il,R) is 7 the completion of (V*,#), the results in 2.3 and 2.4 imply that (S.J ) G In this section, we study the difference between Q and K, in other words, the difference between two topologies on V*. One is given by the resistance metric R and the other is given by the relative topology from the original metric space (K,d). Roughly speaking, it turns out that Q can be always identified with a subset of K and that Q, = K if and only if (Z), r) is regular. First, we observe an important property of (E^J7), that is, the selfsimilarity of (S^J7). 7
P r o p o s i t i o n 3 . 3 . 1 . For any u G f and i^S^uoFi^J . c< yU) VJ
==
Moreover,
Ls [U O JΓi) V O ±*i)
- y •—1 ^**
for any u,v G T. The above proposition is easily verified by (3.1.1). Moreover, if A is a partition of T, = 5 N , then an inductive argument on #(A) shows that —£(uoFw,voFw).
(3.3.1)
Now we start to study the difference between O and K. Proposition 3.3.2. There exists a continuous infective map 0 : Q, —> K which is the identity on V*. Lemma 3.3.3. If FW(V*) D FV(V*) = 0 for w,v G W*, then in£{R(p,q) :
84
Construction of Laplacians on P. C. F. Self-Similar
Structures
Proof. Choose m > 0 so that w, v G Wm. Define p G £(Vm) by {
0
otherwise.
Then p\Fv(y*)nVm ~ 0- Hence, if u is the ra-harmonic function with bound ary value p, then u\Fw{y^ = 1 and W|FW(V.) = 0- BY ( 2 - 2 -6), if P € and g G F , ( 1 4 ) , then
D Recall that d is a metric on K which is compatible with the original topology of if. Note that (if, d) is a compact metric space. Proof of Proposition 3.3.2. Step 1: Construction of 6 Let {pi} be a Cauchy sequence in (V*,R). By Lemma 3.3.3, defining 6m by min
(
inf
,Fw(v*)nFv(v*)=® \
then we have dm > 0. Choose n(m) so that R(pk,Pi) <$m for /c, / > n(m), then there exists w G W m such that pk G U vG ^ m :F u; (v,)nF.(y,)^0^( V *) f o r k > n(m). As max 1(;€ ^ m diam(K-u;,d) —>> 0 as m —• oo, {p^} becomes a Cauchy sequence in (if, d). (diam(-, d) means the diameter with respect to a metric d.) So let p be the limit point of {pi} in (V*, i^). Then we define 0(p) by the limit point of {pi} in (K,d). It is routine to show that 6 is a well-defined continuous map from ft to K and 0|y; is the identity. Step 2: # is injective. Assume that 0(p) = 6{q) for p,q G ft. If #(pi,p) - • 0 and iJ(^,g) -> 0 as i —> oo, then the discussion in Step 1 shows that d(pi,0(p)) —> 0 and d(qi,0(q)) —>• 0 as i —> oo. Let i? be an ?n-harmonic function on V*. Since v can be extended as a continuous function on both (if, d) and (ft, i2), we see that v(p) = limi_>oo v(Pi) = v(5(p)) = v(^(g)) = lim^oo vfe) = v(q). Now, for u G J", by (2.2.6)
where po G Vo- (Recall that um is an m-harmonic function defined by um = Y^xeVm u(x)^m-) Hence, using Lemma 3.2.17, we have limm-+oo Um(p) = u(p). In the same manner, we can also show that lim™-,^ Um{q) = ^((z)Since, wm is a m-harmonic function, it follows that um(p) = um(q)' Hence u(p) = u{q). Therefore, using (2.3.1), we can conclude that p — q. •
3.3 Topology given by effective resistance
85
By virtue of Proposition 3.3.2, we will identify ft with O(ft) and think of ft as a subset of K. Now we ask when ft is equal to K. The answer is given by the following theorem. Theorem 3.3.4. The following are equivalent. (1) ft = K. (2) (ft, R) is compact. (3) (ft, R) is bounded. (4) For any u G T, suppGfi \u(p)\ < oo. (5) (D, r) is regular. Moreover, if (D, r) is regular, then R is a metric on K which is compatible with the original topology. To prove the above theorem, we need the following lemmas. Lemma 3.3.5. For any w G W* and for any p,q G ft, rwR(p,q) > R(Fw(p),Fw(q)).
(3.3.2)
Proof. By Proposition 3.3.1, £(u,u) =
^
rw>~x£(u o Fwr,u o FW') > rw~xE(u o Fw,u o Fw)
for w G Wm. Therefore, we have 'W
\uoFw(p)-uoFw(q)\2 £(uo Fw,uo Fw)
\u(Fw(p))-u(Fw(q))\2 £(u,u)
>
Hence, by (2.3.1), we obtain (3.3.2).
•
Lemma 3.3.6. Let uo = u)\U)2 ... G S. Assume that limsup min 8(arnuj,uf) > 0,
(3.3.3)
where 6 is any of the metrics Sr on E defined in Theorem 1.2.2. If lim inf r m—>oo
m
^> 0
then 7T(UJ) £ ft. Proof Set if = J2peVl\Vo ipp and [0
otherwise,
for w G W*. By (3.3.3), there exist r G T\V and {m n } n >i C N such
86
Construction of Laplacians on P. C. F. Self-Similar
t h a t mn
<
ran+i
for any n > 1 and a^co
•(jOmn and define u = X ^ n > 1 n~1(Pwn>
Structures
—• r as n —• oo. Set wn = Then
M^) =-I 0 Note that liminfn^oo r w n > 0. Hence ^o(w,w)+^
^
otherwise.
—fi(a W ;(u),a 1( ;(^))
= ^o(^)^J + 2 ^ ~2 n>1
ti\fp,ip) < OO. ^
By Proposition 3.2.19, we see that £{u,u) < oo and u G T. Now we assume (3.1.2). Then, by Lemma 1.6.10, ip(x) > 0 for any x G K\VQ. Since r ^ P , we can choose m > 0 so that C = mhixeKT1T2...Trn (p(x) > 0. Without loss of generality, we may assume that o~mriuj G E riT2 ... Tm . So, if w(n) = wnT\T2 . . . Tm, then w(n) = o^i^ . . . a; m n + m and min
99wn (x) =
min
^(^) = C.
Hence inf{w(x) : x G Kw^ D V*} > X]fc=i ^ / ^ "^ °° a s n ~^ °°- Suppose that 7T(U) G fi. Then there exists {pi}i>i C V* such that pi —>• TT(LO) as i -^ oo in (fi, ,R). Since 0 is continuous, it follows that pi —> 7T(O;) as 2 —> oo in (K,d) as well. Therefore, p^ G i ^ ( n ) for sufficiently large i. Hence u (Pi) ~* °° a s ^ ~^ °°- O n the other hand, w is a continuous function on (£2,i?). Hence {w(pi)}i>i converges to a finite limit as i —> oo. Thus we conclude that TT(O;) ^ ft. If (3.1.2) is not satisfied, again we may repeat the same arguments as in the proofs of Proposition 3.1.8 and Theorem 3.2.4. • Lemma 3.3.7. If (D,r) is not regular, then Q, ^ K. u G T such that suppGV^ \u(p)\ — °°-
Also there exists
Proof. If (D,r) is not regular, there exists r^ > 1. Let u = k. Then Proposition 3.1.8 implies that UJ ^ V. Hence uo satisfies the conditions in Lemma 3.3.6. So TT(UJ) £ Q and Vt^ K. Also, by the proof of Lemma 3.3.6, we can find u G T that is unbounded on V*. • Proof of Theorem 3.3.4- Assume that (D, r) is regular. Then Lemma 3.3.5 implies that {Fi}ies is a system of contractions on a complete metric space (fl, i^). Hence, by Theorem 1.1.4, there exists a non-empty compact subset K of (fi,i?) such that K = UiesFi(K). As 6 is continuous, K is also a
87
3.3 Topology given by effective resistance
compact subset of (K,d). For x G K, Dwew*Fw(x) is a subset of K. It is easy to see that Uwew*Fw(%) is dense in (K,d). Hence K = K. Therefore ft = K, (ft, J?) is compact, and 9 is a homeomorphism between (ft, it!) and (K,d). Obviously (2) implies (3). Also as \u(p) — u(q)\2 < £(u,u)R(p,q) for u G F, we can see that (3) implies (4). The above discussion with Lemma 3.3.7 completes the proof of Theorem 3.3.4. • Next, we consider how large ft is in K if (D,r) is not regular. For instance, we will calculate /x(ft) for a self-similar measure μ on K. Theorem 3.3.8. Let μ be a self-similar measure on K with weight (μ^gs. (2) IfYliesW^l,
then μ^il) = 0.
Remark. A. Teplyaev showed that if Y[ieS(ri)^
— 1, then μ(Q) = 0.
This theorem is somewhat intriguing, in particular, in the second case where μ(Q) = 0. As is shown in the next section, even in such a case, if riμi < 1 for all i G S, then (S^J7) is a local regular Dirichlet form on L2{K^). So u G T is originally a continuous function on (V*,R) and it can be extended to a continuous function on (fi, R). This is not surprising. But, despite the fact that 17 is a null set in K, u can be extended to an L2class function on a far larger space K. This fact might suggest that some of the analysis on a large (higher dimensional) space could be determined by information on a small subset of the space. Lemma 3.3.9. Let u> = n(u) G ft.
UJIUJ2
Proof. Let r e V and set p m = using Lemma 3.3.5, we see that
... G E. // -K(UJIUJ2
Z^OVL^...^™
< °°; then
•. .tc;mr). Then p m G F m , and,
max R(x,y). Hence {pm} is a Cauchy sequence in (ft, i?). Let p be its limit. On the other hand, d(pm,
TT(U>)) —> 0 as m —» oo. Hence #(p) = TT(U;) G
Proof of Theorem 3.3.8. Set nm(i,Lj) = # { j : ^ w G E, i G S and m>1. Define A/"={^G£:
lim n ™( 2 ^) m—>oo
777,
= M
ft.
= z, 1 < j < m} for
. for all 2 G 5}.
D
88
Construction of Laplacians on P. C. F. Self-Similar Structures
Then the strong law of large number implies that μ(.^/') = 1. Note that # ^ logr<. Hence, for w&N, l i m Wu uj m—>-oo
(jj m I
/ m
=
ies
Now, if nieS(ri)^ < l> t n e n X^m>or^i^2..-ojm < oo for Lu e JV. Therefore, Lemma 3.3.9 implies TT(UJ) G ft for any LJ G N. Hence μ^tl) — 1. Next, suppose I l i e s f a ) ^ > ^' ^ e t ^ m = {^1^2 • • -Mm : v G V}. Then we can choose m so that #(Vm) < #(W m ). Set
and ^4 = l i m i n f ^ ^ o o ^ . If min r G p8(a n uj,r) —> 0 as n —>• CXD, then, for sufficiently large k, ujk^k+i • • • ^fc+m-i ^ 'Pm- Hence c<; G A. Now, as Bk+mi for / = 1,2,... are independent and μ^k) = μ(Bk+ml) < 1, we see that μ{(^j kBj) = 0. Hence μ{A) = 0. So μ{N D Ac) = 1. Assume that LO G A/" fl A c . Then limsup m _, 0 0 min rG -p ^(cr m o;,r) > 0. Also, liminf m ^ o o r U ; i a ; 2 . < . a ; m > 0 because lim m _ > o o |r u ; i a ; 2 > .. C t , m | 1 / r n = n*Gs( r i) M i > !• Hence, by Lemma 3.3.6, TT(UJ) £ Q. As μ^f] Ac) = 1, it follows that ^(n) = 0. •
3.4 Dirichlet forms on p. c. f. self-similar sets In this section, we continue with the assumptions of the previous section: (D,r) is a harmonic structure on (K,S,{Fi}ies)Then we have 7 (£:^F) G 1U-(Q), where (S^J ) is the natural limit of the compatible sequence {(K n ,ff m )} m >o. If the harmonic structure (D,r) is regular, it was shown in Theorem 3.3.4 that K = Q. Then, results from Chapter 2, in par7 ticular, Theorems 2.2.6, 2.3.10 and 2.4.1, immediately imply that (S^J ) is 2 a local regular Dirichlet form on L (X, μ) for any Borel regular probability measure μ on K. See B.3 for the definition of Dirichlet forms. On the other hand, by Theorem 3.3.4, fi is a proper subset of K if (D, r) is not regular. In this case, functions in T may not be extendable to continuous functions on K. (In fact, T turns out to contain an unbounded function by Theorem 3.3.4.) In this section, however, we will show that T 2 7 can be embedded in L (K^) for a certain measure μ and that (S^J ) is a 2 local regular Dirichlet form on L (K^). Let M(K) be the collection of all Borel regular probability measures on
89
3.4 Dirichlet forms on p. c.f. self-similar sets K. Also we define M{K) = Jμ G M{K) : μ(V;) = 0 and p(O) > 0 for any non-empty open set O C K.}
Of course, all self-similar measures belong to M(K). The following sufficient condition for μ to be in M(K) is sometimes useful. Lemma 3.4.1. Let μ G M(K). If there exists 7 > 0 such that /yμ(Kw) < ) for any (w,i) eW*xS, then μ G M(K). A measure satisfying the condition in the above theorem is called a 7elliptic measure. Proof. If we choose m large enough, then for any w G W m, there exists w1 G Wm such that Kw n Kw> = 0. So, changing the self-similar structure to Cm = (K,Wm,{Fw}w€Wm) if necessary, without C = (K,S,{Fi}ieS) loss of generality we may assume that for any i G 5, there exists j G S such that Ki D Kj = 0. Then 1 = μ^lt) > μ^i) + μ^j) > 27. Therefore 7 < 1/2. Let 6 = 1 - 7. Then 0 < 8 < 1. Now, for any it; G W* and any
> μ^i)
+ μ^j)
>
where we choose j G S so that KiO Kj = 0 . This implies that m m 6μ^w). Hence 7 < μ ^ w ) < ^ for any w G T ^ . This inequality implies that μ(p) = 0 for any p e K and that μ(O) > 0 for any non-empty open subset O C K. • u ( I n 3-2> p m ^ As in 3.2, set P m : T -> C(tf) by Pmu = T,Pevm (p)^is denoted by it m . We also use this notation in this section.) Recall that Um is an m-harmonic function and it can be regarded as a continuous function on K. Also, define Hi$ = {u : w is a 1-harmonic function and it|v0 = 0}.
L e m m a 3 . 4 . 2 . Then there exists c > 0 such that (UI^μ any u G 7^i,o and for any μ G M(K),
where (u, v)μ = fK
< c£(u,u)
for
uvdμ.
Proof. Note that both (•, -)M a n d £ are inner-products on a finite dimensional vector space Hito- As u = ]CpeVi\Vb U(P)^P
^or
a n
^ ^ ^ ^i,o» we
90
Construction of Laplacians on P. C. F. Self-Similar Structures
have (u, u)p =
Y^
apqu(p)u(q) <-
(u(p)2 + u(q)2)
^
V
where apq = §K'^p1^)qdμ. (Since 0 < i/jp(x) < 1 for any x G K, we have u( 2 0 < apq < 1.) Also there exists c > 0 such that #(Vi\Vb) EPeVi\vb &) c£(^, ix) for any w G Wi,o- This proves the lemma. • Lemma 3.4.3. Let μ G M{K). Sm>o Rm(n) < °°> then {Pmu}
Define Rm(f^) — m a x ^ ^ ^ fwl^{Kw). If 2 converges in L {K^) as m —> oc for any
For ease of notation, we write flm instead of Rm{^) if no confusion can occur. w
Proof Jov w G W*, define ^ ^ by μ (A) = μ{Fw(A))/μ{Kw). Then w /x G Af ( # ) . Note that / K / o F^μ™ = ^ — j ]K^ fdμ. Now, by Proposition 3.3.1 and Lemma 3.4.2, it follows that, for u G J-, -Um) £((um+1 weWm
- Um) O Fw, (Um+1 - Um) O Fw)
w
>c-^7(nm+1-Um)^. tiyn JK
Hence cRm£(um+i - % , V i i. So
- um) > \\um+i - um\\% where ||u||£ =
n-l
- Wfc) > | | w n
fc
- U( m | Iμ u
u
for n > m. By Lemma 3.2.16, we have X]fc=m^(^ +i ~~ ki k+\ 5(ifcn — w m , ii n — Txm). Hence
— u k) =
n-l
c( ^
iifc)5K - «m, ^n - ^m) > I K ~ t ^ | | ^ .
(3.4.1)
91
3.4 Dirichlet forms on p. c.f self-similar sets This shows that {w m } m >o is a Cauchy sequence in L2(K^).
•
Define Lμ(u) for u G T as the limit of {ii m } m >o in L2(K, μ) as m —• oo. Then Lμ'. T ^ L2(K, μ) is a linear map. ^μ will be shown to be an injective map. If swpwew* J2m>o Rm(nw)
L e m m a 3.4.4. Let μ G M(K). Iμ'. T ^ L2(K, μ) is injective.
< oo, then
Proof. By (3.4.1), we have R
k)£(U
C( J2
-Um,U-
Um) > \^(u)
- U^.
(3.4.2)
k>m
Note that c is independent of μ. First we will show the following claim. Claim: There exists c\ > 0 such that, for any μ € M(K) and for any u G T with Lμ (u) — 0, > cx max|^(p)| 2 .
max{l, y^ Rm^)}£(u,u) ^
(3.4.3)
peVo
m>0
Proof of Claim. Let S = max{l, ^ in (3.4.2). Then we have cS£(u,u) > cS£(uo,uo) + /
.Rm(/x)}. If Lμ(u) — 0, set m — 0
m > 0
|^o|2^ 2
/
JKK
^/i > cE(uo,uo) + (tto) 2 ,
(3.4.4)
where u$ = JK U^μ = ^2peVo u(p) JK IJJPC^μ. Let Tio be the collection of all harmonic functions. If ap > 0 and ^2peVo OLV = 1 for a = (a p ) pG vb? w e define 7Va(?;) - -y/c5(i;, v) + ( E p G v 0 a p v < » ) 2 f o r v G Ho- T h e n ^ α is a norm on Ho. Since the parameter space a is compact, there exists c 0 such that Na(v) > c^ maxpGvb l v (p)| f° r a n Y v € Wo and for any a. Hence, it is easy to see that (3.4.4) implies the claim. • Now we prove Lemma 3.4.4 by using the claim. If Lμ(u) = 0, then Lμ™ (u o Fw) = 0 for w G W*. Hence, applying (3.4.3) to μw and u o Fw, c3£(uo Fw,uo 1
Fw) >
max
\u(p)\2,
£F(V) R
where c 3 = c i " max{l,sup w e M / ^ J2m>o m(^w)}Note that c 3 is independent of w G W*. Since £^(ti, u) > rw~1£(u o i ^ , w o F w ) , we have max F(V)
\u(p)\ 2
92
Construction of Laplacians on P. C. F. Self-Similar Structures
So if p = TT(U) G Vm, then Wm
£(u,u) > \u{p)\2.
Using Proposition 3.1.8, we see that r(JJlUJ2,^UJrn —> 0 as m —> oo. Hence tx(p) = 0 for any p eV*. Therefore u = 0. D
Now we may identify ^ " a s a subset of L2(K, μ) through Lμ.
Lemma 3.4.5. Let μ G M(K). Assume sup^GVF+ E™>o ^m(/^w) < oo. // £*(u,v) = £(u,v) + JKUVJ^μ, then Lμ is a compact operator from (T,£*) Proof. By (3.4.2),
k>m
Hence | | ^ — Pm\\(jr1s*)-+L2(K^) -^ 0 as m -^ oo. Since Pm is a finite rank operator (Definition B.1.9), Proposition B.I. 10 and Theorem B.I. 11 imply that Lμ is a compact operator. • Now, we are ready to show that (£,F) is a local regular Dirichlet form on L2(K^). See Definitions B.3.1 and B.3.2 for the definitions concerning Dirichlet forms. Theorem 3.4.6. Assume s u p ^ €^ ^ ^ ^ ( / i 1 " ) < oo. Then {£,J~) is a local regular Dirichlet form on L2(K, μ). The corresponding non-negative self-adjoint operator H^ on L2(K^) has compact resolvent. Remark. If μ is a self-similar measure with weight fai)ies, then the condition s u p ^ t ^ J2m>o R"m{lJ>w) < cxD is equivalent to r ^ < 1 for all i € S. Recall that H^ is characterized as a non-negative self-adjoint operator on L2(K^) that satisfies £ = QHN and T = Dom(i7jv1//2). Existence of such a self-adjoint operator is ensured by Theorem B.1.6. In other words, for u e T, u e Dom(iJjv) if and only if there exists / G T such that £(u, v) = (/, v)μ for any v G T. In this case, / = H^u. The subscript "JV" in HN represents the first letter of "Neumann" because — ifAT corresponds to the Laplacian with Neumann boundary condition. We call —HN the Neumann Laplacian. See 3.7 for details. As we mentioned at the beginning of this section, the above theorem is an easy corollary of theorems in Chapter 2 if (D, r) is regular or, equivalently, K = fJ. In such a case, the condition sup t t ; G ^ Sm>o ^0^™) < °° is automatically satisfied by any μ G M{K) because r^ < 1 for any i G S.
3.4 Dirichlet forms on p. c.f self-similar sets 2
Proof Closedness: Note that L (K^)
C ^(K^).
93
Set
= 0}. JK
Then £ is an inner-product on T and (T, £) is complete. For u^T, define a 7 u = u — JK ^μ. If {i£n}n>o is Cauchy sequence in (.T , £*), then {un}n>o is a Cauchy sequence in (.T7, £). Hence f(ii n — v, un — v) —> 0 as n —> oo 2 2 2 for some v G f . Using the fact that / K ?/ d/i = JK u dμ+(JK ^μ) , we see that {JK undμ}n>Q is a Cauchy sequence. Let c be the limit of JK U^μ as n —> oo and let tx = v + c. Then it is easy to see that {un}n>0 is convergent to u in (.?*,£*) as n —• oo. Hence (^",5*) is complete. Regularity: Set W* = {u : u is an m-harmonic function for some m > 0 } . Then H* C C(K). The maximum principle shows that {P m ii} m >o converges to u uniformly on K for u G C(K). Hence H* is dense in C(K). Also, W* is dense in (.F, £*). Therefore W* is a core for (£^T). Markov property: For u E J7, define u in the same way as in (DF3) of Definition 2.1.1. As (fi,?) is a resistance form, u £ T and f(/u, u) < £(u,u). We need to show that ^Mf2 = T^u. This is obvious for any u G J 7 n C(K) because {P m ^} m >o converges to u uniformly on K as m —> oo. Therefore, (5,FC\C(K)) has the Markov property. Since f n C ( K ) D W*, it follows that (£,.F) is the minimal closed extension of (£,,Fn C(K)). Therefore, by [43, Theorem 3.1.1], (5, J 7 ) has the Markov property. Local property: Assume ^ ( u ) = 0 on a open set O C K. For any p G fiflO, there exists w G VF* such that p G X^ C O. Then ^ ( w o F^) = 0. As w μ satisfies the condition of Lemma 3.4.4, it follows that u o Fw — 0 in T. Hence u = 0 on Q D O. Now if supp(^(ii)) D supp(^(v)) = 0 for it, v G J 7 , then supp('u) D supp(v) = 0. Since Sm(u,v) = 0 for large m, we obtain f(w,v) = 0 . Finally, by Theorem B.I. 13, Lemma 3.4.5 shows that H has compact resolvent. • As an easy corollary of Theorem 3.4.6, we also have the Dirichlet form which corresponds to Dirichlet boundary conditions. Corollary 3.4.7. Define To = {u : u G T,u\yQ = 0}. Suppose that 8\ipweW^ ^ m > 0 Rrn^w) < oo. Then (£,J-Q) is a local regular Dirichlet form on L2(K, μ). Moreover the corresponding non-negative definite selfadjoint operator H& on L2(K, μ) has compact resolvent. The operator is called the Dirichlet Laplacian. Remark. Strictly speaking, as T§ is not dense in C(K) with respect to the
94
Construction of Laplacians on P. C. F. Self-Similar Structures 2
supremum norm, (S^o) is not a regular Dirichlet form on L (if,/i). It is, however, a regular Dirichlet form on L2(AT\Vb,//). As with Hjsr, we see that HD is characterized by £\j?QX:pQ — QHD and 1 2 T§ — Dom(Hz) / ). Also u £ Dom(i7/)) if and only if there exists / £ J-Q such that £(u, v) = (/, v)μ for any v £ T§. In this case, / = Proposition 3.4.8. HD is invertible. Moreover, GD — (HD)~ 2 pact operator on L (K^) characterized by £(Gof,v) = (f,v)μ 2 v £ FQ and for any f £ L (K, μ). X
is a comfor any
GD will be called the Green's operator. See 3.6 for details. We will give two proofs of this proposition below. For those who are not familiar with the subjects in B.I, the second proof may be more accessible. Proof. Note that (To, £) is a Hilbert space. First we will give a proof using some general theory. Since £ is positive definite on To, 0 is not an eigenvalue 1 of Hj). As HD has compact resolvent, GD — HD~ is a compact operator. As £(u, v) = (HDU, v)μ for any u £ DOTII(HD) and for any v £ T§, we have
£{GDf,v) = {f,v)n f° r any v £ T^ and for any / £ L2{K^). Next we give another proof, which is less abstract. By (3.4.2), we have 2 \(u, f)μ\ < c£(u,u) > (iz,iz)/x for any u £ TQ. Hence, for any / £ L (K^), u \\ \\μ\\f\\μ yjc£{u,u) I I/I Iμ for u £ T§. This means that u —• (u,/) M is a bounded functional on a Hilbert space (J~o,£). Therefore, there exists h £ J-Q such that £(u, h) = (u, / ) M for any u £ T§. It is easy to see that h is uniquely determined by / and the correspondence / —> h is linear. Hence we write h = Gz>/. Since (GDf, j % = £(GDf, GDf)
> c(GDf, GDf)^
(3.4.5)
we have CHGD/IU < \\f\\μ^ So GD is a bounded operator from L2(K^) 1 to itself. It is obvious that GD = HD~ - Moreover, we also see that Iμ > \/£{GDf'> GDf) by (3.4.5). Hence GD is a bounded operator from 2 2 L (K, μ) to (Vo, £). So, if U is a bounded subset oiL (K, μ), then GD(U) is > 2 a bounded subset of {To, £). Now, by Lemma 3.4.5, Lμ : (^b? £) — L (K, μ) is a compact operator. Hence GD{U) = tμ{GD{U)) is a relatively compact 2 2 -» L2(K^) is a compact subset of L (K, μ). Therefore GD • L (K^) operator. •
3.5 Green's function We continue to assume that (D, r) is a harmonic structure on a connected p. c. f. self-similar structure (K,S,{Fi}ieS), where S = {1,2,... ,JV}. In
95
3.5 Green's function
this section, we will construct the Green's function for (£, J7) derived from the harmonic structure (D,r). The Green's function will be the integral kernel of Green's operator in the next section. Roughly speaking, the Green's operator is the inverse of the Dirichiet Laplacian, GΒ = ( i J ^ ) " 1 , given in Proposition 3.4.8. More precisely, (GDu)(x) = /
JKK
2
for u G L (K, /i), where g is the Green's function which we will define in this section. The above relation is justified in 3.7, in particular, Lemma 3.7.10 and Theorem 3.7.14. Recall that Hm : e(Vm) -> t{Vm) is given by Hm = Ewewm ^RwDRw, where Rw : £(Vm) —»> £(Vo) is defined by Rwu = u o Fw. Also, Hm can be expressed as ( \
m
™ ) , where Tm : £(Vo) -> £(V0); Jm
: £(V0)
and Xm : ^(KA^o) i(Vm\V0). First we will study a discrete version of Green's function —X^1. For this purpose, we need the following lemma. Lemma 3.5.1. Let X — (Xij) be a symmetric n x n real matrix. Assume that X satisfies the following four conditions. (1) Xα < 0 for any i = 1,2,... , n. (2) X^OifijLj. (3) 5Zr=i Xij < 0 for any j = 1,2,... , n and this sum is negative for some 3(4) If i ^ j , then there exist m>1 and io, • • • , im € {l> 2 , . . . , n} such = that iQ — i, im j and X{kik+1 > 0 for all k = 0 , . . . , m — 1. T/ien X is invertible. Moreover, if G — (G{j) = (—X)" 1 , £/ien Gti > Gij > 0
/or anyij
G {1,2,... , n } .
Proof Define a n n x n diagonal matrix A by Aα = —Xα- Also define an n x n symmetric matrix B by Bij = X^ if i ^ j and i ^ = 0 for any z = 1,2,... ,n. Then - X ^ " 1 = / - BA~X, where / is the n x n unit matrix. Let C = BA~X. Then C^ = Bij/Ajj > 0. By the assumption (3), we see that 0 < 5^=1 ^ j — ^ ^ o r a n y ^ = 1? 2,... , n and that this sum is strictly less than 1 for some j , say, / G {1,2,... , n}. Set \x\ = YZ=1 \xi\ for x = (^)i=i,2,...,n € Mn. Then | • | is a norm of
96
Construction of Laplacians on P. C. F. Self-Similar Structures
R n . It follows that
\cx\ <
it
and that \Cx\ < \x\ if xx ^ 0. For j G {1,2,... ,n}, define e5 G RN by (ej)i = 6ij for all i E {1, 2,... , n}, where 6 is Kronecker's delta. By the assumption (4), we see that there exist rrij and z0, i i , . . . , i m j G {1,2,... , n} such that zo = /, imj = j and Cikik+1 > 0 for any k = 0,1,... , m^ — 1. This implies that (C^e^i > 0. Hence it follows that ( C ^ + ^ 1 < IC^e^l < \ej\ — 1. Let s = maxjG{l52,... ,n} m j + 1- If x E Mn and x ^ 0 , then |C s x| < E ^ i k j l i ^ e ? ! < N- Hence it follows that ||C 5 || < 1, where ||C S || is the operator norm of Cs with respect to the norm | •| on R n . Hence -XA'1 = / - C is invertible and (/ - C)" 1 = / + C + C 2 + • • •. This implies that X is invertible and that m>l(io,...,t m )€5m (t,i)
where Sm(iJ)
= {(*o,... ,*m) :io = «, *m = j and {ik}k=o,...,m C { 1 , 2 , . . . , n } } .
By condition (4), we see that G^ > 0 for any i, j G { 1 , 2 , . . . , n } . Now define Mi = maxfc=i?2,...,nGifc. Suppose that G^ = M^ and that i / j . Then
By condition (3), it follows that ^2k:k^j —X kj/Xjj < 1. Hence G^ = G^ for any k with X ^ > 0. This implies that Gik = Mi if there exists m>1 and io,.. •, i m such that zo = k, im = j and X^il+1 > 0 for any / = 0,... , m — 1. By the condition (4), it follows that G^ = Mi for any i = 1,2,... , n. Hence G^ > G^ for any i,j. • Lemma 3.5.2. X m is negative definite. If Gm = —X^ 1, then (G m ) pp > (Gm)pq 0 for anyp.q G F m \F 0 , ^ f e e ^ G m = ((Gm)Pg r, forp,q G V^\Vfc, (Gm)pq > 0 if and only if p ~ q. m
Recall Definition 3.2.9 for the definition of ~. m
1}
Proof. Let (V m \V 0 )/- = {V4 ,... , vi }, where each V^} C Vm\V0 is an 771
k)
equivalence class. Then Xm is decomposed as X$ : ^(vdi}) -> ^(Vdi}) for i = 1,2,... ,/c. Applying Lemma 3.5.1 to each X^ , we obtain the first
3.5 Green's function
97
half of the lemma. Also, by Lemma 3.5.1, it follows that (Gm)pq > 0 if and • only if p ~ q. m
Set X = Xu Xpq = (X^pq, G = Gi and Gpq = (Gi)OT for any p,q € Definition 3.5.3. Define *(x,y) = J2P,qev1\v0GP
{xy)=
[0
for
*,V
otherwise,
x
for x,yeK.
Write ^ (y) = tf(s,y) and ^ ( y ) = *w(a:,2/).
^ ^ is a non-negative continuous function on K x K. \J/£, is an ra + 1harmonic function if w G T^m-
Lemma 3.5.4. For any u e T, i/a; e Kw, otherwise. Proof. For u
Therefore, if x G /fw, ly G Wm and z = F~1{x), then
= rw~x((u o Fw)i(z) - (uo Fw)o(z)) = r w " 1 (u m + i(a:) -
um(x)). D
Set m-l
and set g^(y) = gm(z, y)- Then g^ is an m-harmonic function and g^(y) = 0 if y G Vo- By Lemma 3.5.4, we see that £{g*m,u)=um{x)-uo{x) for any u 6 f . Therefore, gm(x,y) =
(3.5.1) l n
n
Y^PtqeVm\Vo{Gm)pqi p {x)ipq {y)-
98
Construction of Laplacians on P. C. F. Self-Similar Structures
Now note that ^w(x,y) > 0. Hence {gm(x,y)}m>Q is a non-decreasing sequence. Therefore the following definition makes sense, allowing o o a s a possible value of the limit below. P r o p o s i t i o n 3 . 5 . 5 . For x,y G K,
g(x,y)=
define
lim gm(x,y) = m—>oo
Then g is continuous on {(x,y) € K x K : x ^ y}. Moreover, (1) Let x, y G K\Vo. Then g(x, y) > 0 if and only if x and y belong to the same connected component of K\Vo. (2) // (D,r) is regular, then g is a continuous function on K x K. g is called the Green's function associated with the harmonic structure (AΓ).
We need the following lemma to prove this proposition. Lemma 3.5.6. For any p, q G Vm\Vo, the following three conditions are equivalent. (1) p and q belong to the same connected component of K\Vo, (2) p ~ q, (3) (Gm)Pq > 0. Proof By Lemma 3.5.2, we immediately see that (2) and (3) are equivalent. Also, Lemma 3.2.13 implies that (2) => (1). Now the remaining part is to show that (1) implies (3). Let C be the connected component of K\Vo which contains both p and q. Then since C is arcwise connected by Proposition 1.6.6, there exists a continuous path 7 such that 7(0) = p, 7(1) = q and 7([0,1]) C C. Let {pi,p 2 ,..- ,Pk} = Vn H 7([0,1]), where Pi = p(U) and 0 = ti < £2 < • * • < tk = 1- Note that p\ = p and pk = q. Then there exist w(0),... ,it;(/c — 1) G Wn such that Pi,Pi+i G ^ ( i ) for all i = 1,2,... , k — 1. For sufficiently large n, it follows that Kw^ C C for i = 1,2,... , k — 1. By Lemma 3.2.10-(l), there exists an i7n-path between pi and pi+i for any i = 1,2,... , k — 1. Combining those iJn-paths, we have an i/ n -path between p and q. Hence (Gn)pq > 0. Note that (Gm)pq = grn(p,q) = ^ ( p , ? ) = {Gn)pq H Tu < Ti sad p,q G Vm.
Therefore
we see that (Gm)pq > 0. This completes the proof. Proof of Proposition 3.5.5. Set Ktn,p = UW€Wm,p€^w-Kti; f° r P £ K.
• If
x y^ y, then ifm}a5 fl ifm,3/ is empty for sufficiently large m. Hence the sum in the above definition is really a finite sum if x ^ y. Since tyw is continuous, we can verify that g is continuous on {(x, y) € K x K : x ^ y}.
3.5 Green's function
99
(1) Choose m so that Km,x n Vo = Kmty D Vo = 0. Set Vm,x = Km^ n V^n and V^y = i^m,y ^ Vo- Since K m , x and lfm)1/ are connected, there exist connected components C\ and C2 of K\Vo such that K m ? x G C\ and #m,y € C2 . Note that if ^™(a) > 0, then p € V^>a for a = x,y. First assume that C\ —C^. We may choose p e V^ >a. and g' G F m , y that satisfy ^ ( x ) > 0 and ip™(y) > 0. Then Lemma 3.5.6 implies (Gm)pq > 0. Therefore,
9m(x,y) > (GrJn^ixWfiy)
> 0.
Next suppose that C\ ^ C2- If ip™(x) > 0 and ip™(y) > 0, then p e V m,x and q G Vmty. So p and ^ belong to different connected components of K\Vo. Hence, by Lemma 3.5.6, (Gm)pq = 0. Therefore gm(x,y) — 0 for sufficiently large m. This immediately implies that g(x, y) — 0. (2) As (jD,r) is regular, r = maxiesri < 1. There exists C > 0 such that su C rm Px, 2 / GKS U ;Giy m r ^^( : r '2/) ' - S i n c e Y<wewTnr™^™ i s continuous, r P = T,m>o(T,wewm w^w) is continuous. D If (D,r) is not regular, it could happen that g(x,x) = 00. In fact, we will show that g(x, x) < 00 if and only if x G ft. Theorem 3.5.7. The following three conditions are equivalent. (1) g(x,x) < 00 (2) gx G F, where gx(y) = g(x,y)
(3)
xen.
Moreover, if any of the above conditions is satisfied, then £(gx,u) u{x) — UQ(X) for any u G T. Lemma 3.5.8. There exist ci, c 2 > 0 such that, for any u G E,
m>0 m>0
where u> = '
Proof Let x = ir(u>). Then g(x,x) = Note that there exist ci, C2 > 0 such that, for any u G i(Vi\Vo),
pGVi\Vo
—
100
Construction of Laplacians on P. C. F. Self-Similar
Structures
Note that *Wla;a...Wm(ar,x) = EP,qev1\v0 Gpqipp(n(amu;))
< *Ul»2..Mm(x,x)
< c2
D Proof of Theorem 3.5.7. (3) => (2): Set To = {u G J 7 : u|vb = 0}Then (.Fo,£) is a Hilbert space. If x G ft and p G Vo, then |?x(:z)|2 < £(u,u)R(x,p) for li G ^o- This shows that u —> it(x) is a continuous functional To — > R- Hence there exists h e To such that £(/i, w) = ix(x) for any u G ^b- Now £(hmju) = £{h,um) = um(x) = 8(g^,u) for any u G ToHence hm = g^. So h = gx and gx G To(2) =^> (1): This is obvious by the following equality.
£(gx,gx)=
lim £{gxn,,g^,)= lim gm(x,x)
m—>^oo
=g(x,x).
m—>^cxD
(1) =^> (3): l£g(x,x) < oo, then, by Lemma 3.5.8, J2m>o r^^i-..o;m^(^m)2 < oo, where 7r(a;) = x and # m = 7r(crmc<;). Hereafter we will use \I>m and and r a;ia;2...a;m respectively. r(m) to denote ^u^-ujm First, if lim m ^oo (p(xm) = 0, then m i n r 6 p 6(amu,r) —> 0 as m —• oo, where ^ is a metric on E. Then, by Proposition 3.1.8, we see that Sm>o r ( m ) < °°- So, Lemma 3.3.9 implies that x G ( l . Next assume that limsup m _^ oo (p(xm) > 0. Choose {nii}i>o so that rrii < mi+i and (p(xmi) > limsup m _ >oo ip(xm)/2. We may assume that (j mi u; -+ T £ T>\P as z —^ oo. Set ^ = 7T(T). Then we can choose k so that iffc,g = UW£wk:qeKwKw does not intersect VQ. Also we may assume that xmi G i^ffc,^ for all i > 1. Choose q* G i^fc,9 H V^ and define pi = Claim 1: There exists c\ > 0 such that
^m{puPi) < c1^m(x,x)
and ^m(pux)
<
cl^rn{x,x)
for any i and for any m G [0, m^ — 1] U [m^ +fc,oo). Proof of Claim 1: If m > ra» + fc, then ^m(pi,y) = 0 for any y E K. So the claim is obvious in this case. Assume that m < m^ — 1. If ii(y) = \I/ m (p, FWlW2..iW (y)), -u is a non-negative harmonic function for any p £ K. Set X = i£fc,g. Then the Harnack inequality (Proposition 3.2.7) implies that there exists c > 0 such that u(yi) < cu(y2) for any 2/1,2/2 € X. Hence, we see that ^Tn(p,x1) < c^ m (p,x 2 ) for any x^ x2 G i^i^ 2 ...^m(^M)Therefore
*m(Pi,Pi)
c*m(pi,x)
3.5 Green's function
101
This immediately implies the claim. Claim 2: Let Ni = Ui>i{mi,... ,ra*+fc- 1}. Then XlmeNi r ( m ) < °°Proof of Claim 2: Since liminf^oo ip(xmi) > 0 and Y2m>or(rn)(P(x™>) 2 < oo, it follows that X^>i r{mi) < oo. So y ^ r(m) < V^(l + i? + R2 H mGNi
h Rk~1)r(mi) < oo,
i>1
where i? = maxsG,s r s . Claim 3: lim*—«> g(puPi) = lini^oo g(pi,x) = ^(x, x). Proof of Claim 3: Define ^ci^ m (o:,^)
otherwise.
Then Claim 2 along with the fact that g(x,x) < oo implies X^m>o a™ < °°Also, by Claim 1, it is easy to see that tym(pi,pi) < am and \£m(pi, x) < a m for all m. As ^ m (x,x) - lim i _ oo * m (pi,pi) = lim^oo * m (p»,x), the dominated convergence theorem proves the claim. Claim 4: There exists C3 > 0 such that R(z, y) < 2£(gz - g», g* - g») + c3 m^c \^p{z) - i>p(y)\2. for any z,y eft. Proof of Claim 4: Since u(z) — u(y) = £(gz — gy,u) + (uo(x) — uo(y)) for any u G T, we have |«(^) - u(y)\2 < 2£(gz - g», u)2 + 2\uo(z) - uo(y)\2 <2£(gz - gy,gz - gy)£{u,u) + 2 ( £ \u(p) - «(p,) - gy,gz - gy)£(u,u)
p€Vo
where p* G Vb a n d M = #(Vo). Also recall t h a t £(UQ,UO) = fo( ^D(^|vo» M lvb)j where D G CA(Vo). So there exists c > 0 such t h a t
(max\u(p) -u(p*)\)2 < c£(uo,uo) < c£(u,u) peVo
for any u G T. Hence we have Hz) U{ 2
-
f
< 2£{g> -gy,gz-
gy )
This immediately implies the claim. Claim 5: {pi}i>i is a Cauchy sequence in (
102
Construction of Laplacians on P. C. F. Self-Similar Structures
Proof of Claim 5: Note that {pi}i>\ is a Cauchy sequence in (K, d), where d is a metric on K that is compatible with the original topology of K. Hence maxpGvb \ipp(Pi) ~ i/>p(Pj)\ —> 0 as z, j —• oo. Also, by Claim 3, as i —> oo, i
Pi
- 9^9
X
- 9 ) = g(Pi,Pi) - 2g(Pi, x) + g(x, x) -> 0.
As £{u — v,u — v) < (yj£(u — h,u — h) + y/£(h — v, h — ^)) 2 , it follows that £(gPi — gPj ,gPi — gPj) -^ 0 as i,j —» oo. Therefore, by Claim 4, we have R(pi,pj) —> 0 as i, j —• oo. So {pi}i>i is a Cauchy sequence in (fi, ii). Now as lim^^oo d(x,pi) — 0, the limit point of {pi} in (0, i?) is identified with x. (Recall that Q, is identified as 9(Q) C K. See Proposition 3.3.2.) Hence x e ft. • Remark. The idea of using the Harnack inequality is originally due to Dr. A. Teplyaev.
3.6 Green's operator As in the previous sections, (D, r) is a harmonic structure on a self-similar structure (K, 5, {F;}iGs), where S = {1,2,... , N}. In the last section, we defined the Green's function g. In this section we define the (extended) Green's operator Gμ, which is an integral operator whose kernel is the Green's function. Let μeM(K). r
Theorem 3.6.1. Define R^μ) = maxwewm wV>(Kw) ' for 1 < t < oo. (Ift = oo, Rln = maxwewmrw') Assume ^ m > 0 #m(AO < °°- If s is the constant {including oo) that satisfies 1/t + 1/s = 1, then for f G LS(K^),
(GJ)(x) = f g(x,y)f(yMdy)
(3.6.1)
JK
is well-defined for all x G K and Gμ f G C(K) D T§. Moreover, Gμ : S L {K^) —• C(K) is a compact operator. Also £{u, Gμ f) = [ (u- uo)fdμ
(3.6.2)
JK
for any u G T. The operator Gμ is called the (extended) Green's operator. Remark. If /i is a self-similar measure with weight (/zi,//2,... ,^iv), then the condition J2m>o ^mil2) < oo is equivalent to r^/x^)1/* < 1 for all i G S.
103
3.6 Green's operator
Remark. If (-D,r) is regular, then g is a continuous function on K x K. Therefore the above theorem is almost obvious in this case. In particular, we see that E m > o i ? m < oo. Hence Gμ : Ll(K^) -> C(K). Comparing the above theorem with Proposition 3.4.8, we see that Gμ = 1 w GD = Hr>~ if s\xpweW^ Y^m>0 Rm(v ) < oo. More precisely, (3.6.2) imS plies that Gμf = GDf for / G L (K^) n L2(K^). Hence, if t > 2, then the (extended) Green's operator Gμ is really a natural extension of the Green's operator GD and GD(L2(K^)) C C(K) n T^. Hence Dom(if D ) C C(K) n JF0. Also, if 1 < t < 2, then G D | L S ( K , M ) = GM and GD(LS(K, μ)) C C(K) n J^To prove Theorem 3.6.1, we need several lemmas. Lemma 3.6.2. / / E m > 0 C M < oo, then gx G L\K^) for any x eK. Moreover, x —> gx is a continuous map from K to Ll{K, μ). For u G La(K,μ),
^{"(Iμ)1?0'.
we set | | u | | M j a = ( / K
Proof. For X = -K{UJ\UL)2 . . . ), ^ ( y ) = E m > 0 rwiW2-wm^a;ia;2...Wm (^5 !/)• Hence y ^ ruj1uj2 m>0
11<7^1 Iμ t
Wmll^^ cj
w
Hμ*
m>0
Hence 0*
m>0
eL\K^). Note that x
y
\\g -9 \\^t<
\\9Xm-9Vm\U +
\\9 m-:^lUt+ X
11^-^11
Since | | < ^ - ^ | U t < 2-^k>m+l Rk\lu) and gm is continuous on iC x Bf, we x y see that \\g — ^ || M ,t —> 0 if d(x, y) •
Lemma 3.6.3.
(?/)
"
Proo/. Note that £(ip™,u) = £m(il>™, u) = -(Hmu)(p) Since £(gp,u) = u(p) — uo(p), it follows that q
£( Y^ (Hm)pqg ,u)
for any u e T.
= (Hmu)(p) - (Hmuo)(p) = £(-i>™,u)
peVm r
Since UQ is harmonic, Lemma 3.2.16 implies £{$ ? — il>p, UQ) = 0 for p G VQ.
104
Construction of Laplacians on P. C. F. Self-Similar Structures
Also (Hrnuo)(p) — 0 for p 6 V^\Vb. Hence we have
7
for any u e J . This implies the desired equality.
•
Lemma 3.6.4.
Proo/. Note that (H7n(Gμf))(p)
=
The result follows by Lemma 3.6.3.
•
Proof of Theorem 3.6.1. By Holder's inequality,
Hence Lemma 3.6.2 implies that Gμf G C(K). Let {/ n } n >i be a bounded sequence in LS(K^). Then {G M / n }n>i is uniformly bounded and equicontinuous. So, by Ascoli-Arzela's theorem, {Gμfn}n>l contains a subsequence that is convergent in C{K). Hence Gμ is a compact operator from LS(K^) to C(K). By Lemma 3.6.4, = [ (um- v * ^ .
(3.6.3)
JK
peJm Set u = Gμ?. Then SmiG^Gnf)
= [ (Gμf)rnfdμ
< II/IU,! SUp \(Gμf)(x)\.
(Note that (G M /) 0 = 0.) Hence S(Gμf, G M /) < oo and G M / G To. For any w e f , letting m -> oo in (3.6.3), we obtain (3.6.2). • Next we study the n-th power of the Green's operator Gμ for n G N. The integral kernel #(n) for G^; is given by the following inductive formula. Set g^\x,y) = g(x,y) and
[
K
for n G Z.
105
3.6 Green's operator
Theorem 3.6.5. Assume that s u p ^ ^ E m ^ ^ i ^ 1 " ) < o c - Then, / o r any n, there exist cn > 0 and \I/™ (x,y) swc/i t/iat ^w is continuous on KxK,
ff
(3.6.4)
and o < tfW (*, y)
(3.6.5)
CnXKw (x)xKw (y)
for any w G VF*. Moreover, if μ is a self-similar measure on K, then there exists a non-negative continuous function ^^ on K x K such that
[0
»*»«*.. otherwise.
(
3.6.6)
Prom the above theorem, it is easy to see that g^ is continuous on {(x, y) : x,y € K,x ^ y} because the sum in (3.6.4) is, in fact, a finite sum if x T^ y. Note that g^ = g is independent of μ but this is no longer true if n > 2. Corollary 3.6.6. Assume that sup^ eVF ^ E m > o ^ ( ^ ) < °°- V there exists t>1 such that Y^m>0 fl^(^) < oo7 then, for any n > (1 — 1/t)" 1 , #(n) zs continuous on KxK and G™ Z5 extended to a compact operator from LX{K, μ) to C{K). In particular, suppose that μ is a self-similar measure with weight (μ^izs which satisfies μ^i < 1 for any i G S. Then g^ is a continuous function on K x K for any n > max^s log A** / log μ^i. Remark. If ^ is a self-similar measure with weight fai)ies, the quantity maxie^log/ii/log^r^ coincides with ds&(v, μ)/2, where ds is the unique positive number that satisfies Y2ies(Viri)ds^2 = 1, ^ is a self-similar meads 2 sure with weight (^iri) / )ies and <5(i/, μ) is the distortion between v and μ defined in Proposition 4.5.2. In 4.1, ds is called the spectral exponent and it determines the asymptotic behavior of the eigenvalue counting function of the Laplacian. See Theorem 4.1.5 for details. Proof of Corollary 3.6.6. Set
Then hm is continuous on K x K and \hm (x,y)\ < c n (i?^(^)) n , where a = (1 - l / n ) " 1 . As 1 > 1/n -h 1/t, it follows that t' < t and hence E m > 0 < W < °°- S O E m > 0 ^ n ( < ( M ) ) n < C n ( E m > o < ( ^ ) ) n < 00.
106
Construction of Laplacians on P. C. F. Self-Similar Structures
Therefore ^™>o ^™ *s uniformly convergent on K x K. Hence g^ continuous on K x K. The rest of the statement is straightforward.
is •
Proof of Theorem 3.6.5. The proof is by induction on n. For n = 1, we have g(x,y) = Y,weWti rw^w(x,y) and 0 < ^w(x,y) < ciXKw(x)xKw(y), where C\ = maxX)1/Gx ty(x,y). Next assume that the claim of the theorem is true for n. Set ^ ™β = { \ for a,PeW*. Note that
Hence we have
0 < Vffifay) Now, let hw,m(x,y)
< clC^(KanK0)xKa(x)xK0(y).
= E 7 e ^ m rwrwynμ(KW7)n-x¥^)(x,y).
0 < &w,i,\\x,y) < c1cnμ(KW7)xKm(x)xKu,^(y)• \hw>m(x,y)\ < clCnrwn+^(Kw)n
max
(3.6.7) By (3.6.7),
Therefore, ^^{KJT
Since £ m > 0 ( / U / ^ ) ) n < Mn, where M = supw€W, E m > o ^ ( ^ ) . w e see that J2m>o hw,m is uniformly convergent on KxK. Hence, if hw(x, y) = Sm>o ^w,m(x,y), hw is continuous on K x K. Also 0 < hw{x,y) < ClcnMnrwn+^{Kw)nXKw{x)xKw{y)<
(3.6.8)
Next define fw,m(x,y) = Y^ieWrn rwlrwnμ(Kw)n~1^^,w (x,y). Similar arguments to those given above for /i W j m will show that fw = ^2m>1 fw,m is continuous on K x K and that 0 < fw(x,y)
< dc n Mr^ n + V(^) n Xi^Wx^(2/)-
So define ^L n + 1 ) by r ^ 1 μ^v)"V^(x,y) by (3.6.8) and (3.6.9), it follows that ^ + that 0 < ^+l\x,y) < cn+lXKw(x)xKw(y), Now by definition,
1 )
(3.6.9)
= hw(x,y)+fw(x,y). Then, is continuous on K x K and where cn+l = c n ci(M n
107
S.I Laplacians
Finally assume that /i is a self-similar measure. Note that (3.6.6) is true for n = 1. So suppose (3.6.6) is true for n. Then
"
X l
'
e J r
-
(3.6.10)
otherwise. Recall that r^+^iK^V^ix^y) = hw(x,y) + / w (a;,y). By (3.6.10), routine calculations show that (3.6.6) is true for n + 1. • Obviously, g^n\x,y) is non-negative by definition. Moreover, Theorem 3.6.5 implies that g^ (x, y) > 0 if x and y belong to the same connected component of K\Vo. More precisely, we have the following result. Corollary 3.6.7. Let m
£ r for any ra, n G N. If x and y belong to the same connected K\Vo,
then there exists m>0
component
such that g™ (x,y) > 0 for any
of
n>1.
Proof By Proposition 3.5.5-(l), there exists ra > 0 such that gin(x,y) = grn{x,y) > 0 and ^ } ( x , x ) > 0. Suppose that g{m\x,y) > 0. By (3.6.4),
gm(x,z)gW(z,yMdz) < g£+1\x,y).
(3.6.11)
Note that gm and gin are continuous on K x K. Since gm(x,x) > 0 and { { +l g m\x,y) > 0, (3.6.11) implies that g m \x,y) > 0. Now the corollary follows by induction. •
3.7 Laplacians As in the previous sections, (£>, r) is a harmonic structure on a self-similar structure (K, 5, { F j i G 5 ) , where S = {1,2,... , N}. In this section, we define the Laplacian AM associated with (D, r) and μ, w where μ G Af(jRT). If sup l u e V ^ m Xlm>o ^m{l^ ) < oo, we already introduced in 3.4 non-negative definite self-adjoint operators HN and Up associated with Dirichlet forms (S.J7) and (£,.Fo)> respectively, on I?(K^μ). (See Theorem 3.4.6 and Corollary 3.4.7.) We might think of -HN and -HD as Laplacians for the corresponding boundary conditions. However, in this section, we will give a more direct way of defining the Laplacian. Our Laplacian, Aμ, corresponds to the classical second derivative and its domain Dom(AM) corresponds to the C2-class of functions. See Example 3.7.2.
108
Construction of Laplacians on P. C. F. Self-Similar Structures
Definition 3.7.1. Let μ G M(K). Define Vμ = {u G C(K) : there exists / G C(K) such that - f(p)\ = 0}, lim max ^]p(Hmu)(p) ™-+oo peV\Vo
where μ-m^p — JK ^β™ Aμ. Then for u G Vμ, we write / = Aμu, where / is the function appearing in the definition of Vμ. Aμ is called the Laplacian associated with (D, r) and μ. Obviously, Aμ : Vμ —• C{K) is a linear map and Vμ — Dom(AM). If u is a harmonic function on K, then it immediately follows that u £ Vμ and Aμ!! = 0. Example 3.7.2 (Interval). Recall Example 3.1.4. Set r = (1/2,1/2). Then (D,r) is a harmonic structure. For this harmonic structure, if p = m i/2 £ Vm, then
{
ix(2-^)-u(0) u(l - 2~m) - u(l)
if p = if p = 1 .
(3.7.1) Obviously, a harmonic function is a linear function. Now let μ be the selfsimilar measure with weight (1/2,1/2). Then μ is, in fact, a restriction of Lebesgue measure on R to [0,1]. For p G V^i\Vo, since ip™ is a piecewise linear function, we see that μrn^p = 2~m. Hence Routine calculus arguments show that Vμ = C2([0,1]) and Aμ?! = d2u/dx2 for u e Vμ. Example 3.7.3 (Sierpinski gasket). Let (D,r) be the harmonic structure on the Sierpinski gasket given in Example 3.1.5. Define a linear operator Lm : £(Vm) -+ e(Vm) by
where, for p G Vm, Vmjp = {q € Vm: q ^ p, there exists w G W^ such that p, g G -F^ (Vb)}.
109
S.I Laplacians Then Hm = (5/3) m L m . Note that
\{P
#{Vm,p) = l
^
e
Now let μ be the self-similar measure on the Sierpinski gasket with weight (1/3,1/3,1/3). Then it follows that JK ^ μ = 1/3 for any p G Vo- Using self-similarity and symmetry, we obtain,
IK
P
[1/3-+1
ifpe^oV
Hence μ^p(Hmu){p)
^5m(Lmu)(p)
=
for any p G Vm\Vo. The associated Laplacian Aμ is called the standard Laplacian on the Sierpinski gasket. Lemma 3.7.4. For v G C(K) and u G Vμ, V
-> / vAμUdμ
v(p)(Hmu){p)
as m —• oo.
Proof. If fm(x) = ^2peVrn\vQv(p)^m)p(Hmu)(p)ip^l(x), then fm —»- vA^.w as m —> oo uniformly on any compact subset of K\Vo. Also {/m}m>o is uniformly bounded. Hence, fK / m d ^ —> / K vAμudμ as m —• oo. • Lemma 3.7.5. For / G Pμ and pG^o? lim -(H
Proo/. Recall that Hm = (T™
*^ m ), where T m :
i{Vo) - ^(^m\Vo) and X m : £(Vm\V0) Tm - tJmX^Jm, hence we have
where / 0 = /|y 0 and /i = /Iv^vb-
T n i s
-> ^ m \ V b ) -
Note that D =
implies
(^m/)|Vo = ^rn^m
(Hmf)\vm\V0-
110
Construction of Laplacians on P. C. F. Self-Similar Structures
Since ( V ^ " 1 ) ^ = -i/>p(q) for psV0
and q G Vm\V0, we have
Letting ra —> oo, Lemma 3.7.4 implies the lemma.
•
Definition 3.7.6. For / G Vμ and p G Vo, we define
p is called the Neumann derivative of / at p G VoFor the case of Example 3.7.2, (3.7.1) implies that (d/) 0 = ~/'(0) and (df)i = f ( l ) . Lemma 3.7.7. For v G C(K) and u G Vμ, lim fm(i;,ifc)= y^v(p)(dw)p- / vAμUdμ.
m—>^oo
'
J
JITS K
pevo
Proof. Note that
^2
Y^ v(p)(Hmu)(p). pevm\Vo
Apply Lemma 3.7.4 to complete the proof.
•
Using the above lemma, we can immediately obtain the following theorem. Theorem 3.7.8 (Gauss-Green's formula). Vμ C T. For any v G T and u G Vμ, S(v,u) = ] T v(p)(du)p - J vAμudμ.
(3.7.2)
If both u and v belong to Pμ, then p
- u(p)(dv)p) = / (vAμU - uAμv)dμ. J K
(3.7.2) is a counterpart of the Gauss-Green's formula in ordinary calculus on R n and Riemannian manifolds. For example, let U be a bounded domain in R 2 and assume that dU is a smooth curve. For u,v € C2(f7), set
, dv du
dv du x
S.I Laplacians
111
Then the Gauss-Green's formula says £(v,u) = / v—ds — I vAudxdy. Jdu un Ju w h p r p L\U A?/ — @^ -t-Iwnere — dx2
Next we will study relations between the Laplacian Aμ and the abstract non-negative definite self-adjoint operators HN and HoTheorem 3.7.9. Assume that μ G M(K) and s u p ^ G ^ S m > o oo. // Z>D,M = {u G Vμ : u\v0 = 0}, then VD^ = Dom(HD) n Vμ, Ho\vD p — ~^μ\'DD,μ and Ho is the Friedrichs extension of—Aμ onV-p^. Also define VN^ = {u e Vμ : (du)\p = Ofor all p G Vo}. Then VN^ — Dom(H]s^)^Vμ, HN\VN^ = ~^μ\vN,μ and HN is the Friedrichs extension of —Aμ on Visf^. This theorem shows that — H^ is the Neumann Laplacian associated with (S, J-) and μ and that —Ho is the Dirichlet Laplacian associated with {£,J~) and μ. We write Vo and £>TV instead of Vo^ and X>AT^ respectively if no confusion can occur. Lemma 3.7.10. For any f G C(K), G M / G VD and AM(G?M/) = - / . Proo/. By Lemma 3.6.4, for p G Fm\Vb,
where e m = max p G v m (sup 2 / e j f t : m \f{y) — f(p)\)- As / is uniformly continuous on K, em —> 0 as m —> oo. • Proof of Theorem 3.7.9, Part I. Recall that, for u G To, u G Dom(# D ) and Hou = f if and only if £(v,u) = (v, / ) M for any i; G ^o- By (3.7.2), if u G DDJ £(V,U) = (v, —Aμu)μ for any i; G ^>• Hence D ^ C Dom(Ho) l and F D | P D = - A M | p D . Since Gμ\c^K) = Hp \C(K), li u e VD and -AμU = /, then Gμ? = u. Hence P D C Gμ(C(K)). On the other hand, by Lemma 3.7.10, we have VD D G^CiJK)). Hence VD = Gμ(C(K)). set Hou = f for u G Dom(if/p). Then, there exists {fn}n>o C such that / n ^> / as n —> oo in L2(K^). If wn = Gμfn, then u n G ). Recall that G^) = Ho~x- By (3.4.5) (and the discussions following it), < c(ft, ft)M for any ft G Z/2(i^, //), where f* = 5 + (•, -)M. As () |c(K), wehavef*(u n -Tx,w n -w) < c(fn-f,fn-f)μ -> 0 as n —>• oo. Hence the completion oiVo with respect to £* contains Dom(if^). Now obviously, the completion of Dom(Ho) with respect to £* equals Jo
112
Construction of Laplacians on P. C. F. Self-Similar Structures
because the Friedrichs extension of Hp is HD and T§ Therefore the completion of Vp with respect to £* equals T§. This means that HD is the Friedrichs extension of — A M | p D . • To prove the rest of Theorem 3.7.9, we need several preliminaries. For / G L\K^)Z define Pf = f -JKfdμ. Set IP(K^μ) = P(If(K^)\ f = P(J=), VN = P(VN)znd C(K) = P(C(K))L Note that (J^£) is 2 a Hilbert space, hence {£,F) is a closed form on L (K, μ). Let H^ be the non-negative self-adjoint operator associated with (£, OF) on L2(K, μ). Then, for u G T and / G L 2 (if,/i), w e pom(HN) and #7v^ = / if and only if £(v,u) = (v,/) for any v e J7. So it is easy to see that Dom(HN) = Dom(iJjv) H ^ and HN = ^iv| D o m ( j f f i v ) n ^- Note that HN is injective because £ is positive definite on T. The following lemma is obvious from Lemma 3.6.4. Lemma 3.7.11. For f G L°°{K^), peV0.
(d(G M /)) p - -JK'^βpfdμ
for any
Lemma 3.7.12. For any f G L2(K, μ), there exists a unique u G f such that £(v, u) = (vo, / ) M /or an?/ v e T, where v0 = J2pev0 v{v)^p' Moreover, define Q : L2(K, μ) —> L2(AT,/i) 6?/ Q/ = u. T/ien Q/ is harmonic and (d(Qf))p = $K^vWv f°r anype Vo. Also Q : L2(K^) -> L 2 (X,^) is a compact operator. Proof Let ap = / x ^ p / d ^ for any p G V^. Then (v0, / ) M = Z)p€Vb Since / G L2(K^)y
J2peVo ap = 0. Choose any # G fi. Then
Therefore, v —> (VQT^μ is a continuous functional on (iF,£). Hence there exists u G T such that £(v, u) = (i>o, /) for any v G J^. Let it = Q/. Then, for p G Vm\Vo, -(Hm(Qf))(p) = eU>?,Qf) = S(Py>?),Qf) = ( ( P « ) ) O , / ) M = 0, because (P(^^))o is a constant. Hence Qf is harmonic. Since A^Q/) = 0, (3.7.2) implies £(«,Q/) = E p€vb v(p)(d(Qf))p. Therefore (d(Q/)) p = ap. Since Qf is harmonic, (3.7.3)
113
S.I Laplacians
Let Ho = {Pu : u is a harmonic funtion on K}. As dimW0 < oo, there exists c > 0 such that c{u,u)μ < £o(u,u) for any u G Ho- Hence, by (3.7.3), we have c||Q/|| M < ||/||M.^Therefore Q : Z 2 ( i < » -> L2(K^) is a bounded operator. Also, as dim'Wo < °°5 Q is a finite rank operator and hence it is a compact operator by Proposition B.I. 10. • Proof of Theorem 3.7.9, Part 2. For / G Z 2 (if,/i), define GNf = P(GDf + Qf). Then for any iz G J 7 , £(u, GAT/) = (u - u0, /) M + (u 0 , /) M = (u, /) M Hence Gjv/ G Dom(iJjv) and -ffjv(Gjv/) = /. As G^ and Q are compact operators, we see that GN ' L2(K^) —> L2(K^) is a compact operator. Since i7/v is injective, it follows that H^ is invertible and GN = H^1. Also we obtain £(GNf,GNf)
= {GfffJ^
IIGAT/IUII/IU
< cWfWl
(3.7.4)
Next for, u G T>N-, (3.7.2) implies that £(v, u) = (v, —Aμu)μ for any v G J^. Also, by (3.7.2), 0 - £(l,u) = - fK A^μ. This implies A ^ G C(i^). Therefore, we have -u G Dom(i?^) and Hjyu = —Aμu. Thus we obtain that Vjsf C Dom(ifiv) and that ^T/vlp^ = —AM|p . On the other hand, for / G C(K), GNf = P(G£f + Qf). Hence, it follows from Lemma 3.7.11 and Lemma 3.7.12 that Gjv/ € £>w Therefore GN(C(K))=VN. Now for ti G Dom(iifjv), let / = Hjyu. Then there exists {fn}n>i C C(K) such that / n —* / as n -^ oo in L2(K^). If wn = GNfm then, by (3.7.4), £(un—u,un—u) < c\\fn — /I| 2 . This implies that itn —* iz as n —> oo in (J7, £*). Since un G Pjv? we have shown that the completion of V^ with respect to £* contains Dom(if/v). Since i7^ is itself a non-negative selfadjoint operator, the Friedrichs extension of H^ on Dom(i/^) is HN and T — Dom(Hj ). Hence, the completion of Dom(ifjv) with respect to £* equals to T. Therefore the completion of V^ with respect to £* is j ^ . Finally, note that T = T 0 {constants} and Piv = T^N © {constants}. Also we obtained H^ = HN\Dom(HN\n
114
Construction of Laplacians on P. C. F. Self-Similar Structures
Corollary 3.7.13. For b = D and N, if f G Dom(Hb) and Hbf G C(K), then f G Dom(Hb) D Vμ = Vb^. To end this section, we study the Dirichlet problem for Poisson's equation. Theorem 3.7.14. For any f G C(K) and p G £{Vo), there exists a unique ii G Vμ such that (3.7.5) u\Vo
= p.
Moreover, u is given by u = J2Pev0 Pip)^
~ Gμ!•
Proof. If u = Y2pev0 P(P)^P ~~ G>/> then, by Lemma 3.7.10, u satisfies (3.7.5). Assume that both u\ and u^)" 1 , it follows that t> = 0. Therefore U\ = U2' n Corollary 3.7.15. The collection of harmonic functions on K coincides with {u : u G Pμ, Aμ^ = 0}. Combining Lemma 3.6.4, Corollary 3.7.13 and Theorem 3.7.14> we obtain the following theorem. Theorem 3.7.16. Let
) - /
9(x,
JK
pevo (4) u G C(K) and, for any m>1
and any p G Vm\Vo,
(Hmu)(p) = f
JKK
(5) u G T and, for any v G Vμ^, £{v,u) = — JK
3.8 Nested fractals
115
3.8 Nested fractals In 3.1, we introduced the notion of harmonic structures. In the following sections, we then constructed Dirichlet forms, Green's functions and Laplacians under the assumption of the existence of a harmonic structure. In this section, we will show that there exists a harmonic structure on a class of highly symmetric p. c. f. self-similar structures based on the idea of Lindstr0m [116], where the notion of nested fractals was introduced. Recall 3.1, where we observed that the existence problem for harmonic structures is equivalent to the eigenvalue problem (or fixed point problem) of the renormalization operator lZr. In [116], Lindstr0m showed the existence of a fixed point of a probabilistic counterpart of the renormalization map lZr for nested fractals. In this section, we will extend the notion of nested fractal and introduce strongly symmetric p. c. f. self-similar structures. In Theorem 3.8.10, we will show existence of harmonic structures for strongly symmetric p. c. f. self-similar structures. Throughout this section, £ = (K, 5, {Fi}ies) be a connected p. c. f. selfsimilar structure, where S = {1, 2,... , N}. First we introduce the notion of symmetry of a p. c. f. self-similar structure. Definition 3.8.1. A homeomorphism g : K —» K is called a symmetry of C if and only if, for any m > 0 , there exists an injective map g^ : Wm —• Wm such that g(Fw(V0)) = Fg(m){w)(V0) for any w G Wm. In this section we also assume that K cRd and that i^ is the restriction of a similitude with Lip (i^) = Q on K for any i G S. Recall that a map / : Rd -> Rd is called a similitude with Lip(/) = c if there exist U G 0(d) and a eRd such that f(x) = cUx + a for any x G R. See Definition 1.1.2 and the following remark. Let #(Vb) = M and let Vo = {Pi}i=i,... ,M- Then, without loss of generality, we may assume that X^=i Pi = 0Definition 3.8.2. For x,y G Rd, x ^ y, let Hxy = {zeRd:\x-z\
=
\y-z\}.
(HXy is the hyperplane bisecting the line segment [#,y].) Also let gxy : Rd -> Rd be reflection in Hxy. Remark. Let g be an affine transformation from Rd to itself. Suppose that g\x is a symmetry of C. Then g(Vo) = Vo- Since Y^L^Vi — 0, it follows that #(0) = 0. Hence g is a linear transformation.
116
Construction of Laplacians on P. C. F. Self-Similar
Structures
Definition 3.8.3. (1) Let ra* = # { | x - y\ : x,y € V0,x ^ y}. Define l0 = min{|a; - y\ : x,y G Vb,x ^ y}. Also we define {Zt}t=o,...,m«.-i inductively by U+\ = min{|x — y\ : x, y G Vb, \x — y\ > k}. (2) Let Xi G Vm for i = 1,2,... , n. Then (xi)i=i,2,... , n is called an m-walk (between #i and x n ) if and only if there exist u> 1 ,... , u>72"1 G W m such that #i, Xi+\ G i ^ t (Vo) for alH = 1,2,... , n — 1. (3) A 0-walk (xi)i=i j2,... ,n is called a strict 0-walk (between x\ and x n ) if and only if \xi — Xi+\ \ — lo for any z = 1,2,... , n — 1. (4) Qs = {g : ^ G 0(d), # | K is a symmetry of £ } . ^ s is a group with respect to the natural composition. Recall the remark after Definition 3.8.2. Then g G Qs is equivalent to the condition that g : Wd —• W* is an affine transformation which preserves the Euclidean metric and g\x is a symmetry of C Definition 3.8.4. (1) We say that C = (K, 5, {Fi}ieS) is strongly symmetric if it satisfies the following four conditions: (551) For any x ^ y G Vb, there exists a strict 0-walk between x and y. (552) If £,2/, z G Vb and \x — y\ = \x — z\, then there exists g G Qs such
that g(x) = x and g(y) = z. (553) For any i = 0 , . . . , ra* — 2, there exist x, y and z £ Vo such that |x — 2/| = /;, |x — 2| = Zi-i-i and ^ ^ I A : is a symmetry of £. (554) Vb is ^-transitive, that is, for any x ^ y G Vb, there exists g € Qs such that #(:E) = 7/. (2) We say that C = (K, 5, {Fi}iEs) is an affine nested fractal if gxy\K is a symmetry of C for any x ^ y G VoIt is easy to see that if C is strongly symmetric, then it is weakly symmetric, which has been defined in 1.6. The following proposition is obvious by (SS4). Proposition 3.8.5. Let Vb = {Pi}i=i,2,...,M- V^ then |pi| = . . . = \pM\.
^s strongly symmetric,
It follows that an affine nested fractal is strongly symmetric. In fact, the conditions (SS2), (SS3) and (SS4) are obvious. We can verify (SSI) as well by virtue of the following lemma. L e m m a 3.8.6. Let C = (K,S,{Fi}ies) be an affine nested fractal. there exists a strict 0-walk between x and y for any x ^ y EVQ. Proof For p eVo, set Bp = {q G Vb : there exists a strict 0-walk between p and q}.
Then
3.8 Nested fractals
111
Define A — {{x,y) : x,y G VQ,BX ^ By}. (Note that there exists no strict 0-walk between x and y if (x, y) G A ) Assume that A ^ 0. Set a = min{|x — y\ : (x, y) G A} and choose (x, y) G A satisfying \x — y\ = a. (Note that a > Zo-) Let z e Bx with |z — x| = fo- Then (z,y) G A and therefore |y — z\ > \y — x\ = a. Since \z — x\ = lo < \x — y\ = a, it follows that x and z belong to the same side of the hyperplane Hyz. Hence, if e = (y — z)/\y — z\, then (x,e) < 0. Let 7 = —(x,e) and let XQ = gyz(x). Then \y — xo\ = lo and x$ G By. Note that xo = x + 2je. Using Proposition 3.8.5, we obtain (y -x,e)
= - ( x , e) H- (y, e) = 7 + (2/ - (y + *)/2, e) = 7 4- |y - 2;|/2.
Therefore, |y - x| = a > (y - x, e) = 7 + \y - z\/2 > 7 4- a / 2 .
(3.8.1)
This implies that \y — x\ > 27 = |x — xo| > a — \y — x|. Therefore equalities hold in (3.8.1). Thus we obtain that \y — x\ = (y — x, e) and that \y - z\ = \y - x|. So e = (y - z)/\y - z\ = (y - x)/\y - x\ = (y - x)/\y - z\. Hence x = z. This contradiction immediately implies that A = 0. • So we obtain the following proposition. Proposition 3.8.7. If C = (K, S, {i^}ies) is an affine nested fractal, then it is strongly symmetric. The converse of the above proposition is not true. See Example 3.8.13 below for an example. Originally Lindstr0m introduced the notion of nested fractals in [116]. Afterwards, the notion of affine nested fractals was introduced as an extension of nested fractals in [37]. The definition of (affine) nested fractals in [116] and [37] contained the following conditions: Connectivity: For any i,j G 5, there exist z i , . . . ,im such that i\ — z, im = j and Fik(V0) n Fik+1(V0) ^ 0 for any k = 1 , . . . ,ra - 1. Nesting: For any m and any w, wr G Wm with w ^ wf, Kwr\Kw,
=
Fw(Vo)nFw,(Vo).
Since we assume that K is connected, the connectivity condition is satisfied by Corollary 1.6.5. Also the nesting condition is immediately verified by Proposition 1.3.5-(2). Since the boundary VQ of an affine nested fractal is highly symmetric, the geometry of VQ seems quite restricted. In fact, it is conjectured in [6,
118
Construction of Laplacians on P. C. F. Self-Similar
Structures
§5] that Vo is a regular planar polygon, a d-dimensional tetrahedron or a d-dimensional simplex. Hereafter in this section, we always assume that C = (X, 5, {Fi}ies) is strongly symmetric. We also impose the following assumption on C: #(Fi(Vo) n Fj(Vo))
1 whenever i ^ j .
(3.8.2)
This may be regarded as a technical assumption in order to avoid nonessential difficulties. The forthcoming results in this section may hold without (3.8.2). In [37], it is claimed that (3.8.2) holds for any affine nested fractals. However, this does not seem quite clear at this moment. See also [6, §5] for some comments. Definition 3.8.8. (1) on Vb, Cs(Vo), by
We define the collection of symmetric Laplacians
£s(Vo) = {DG CA(Vo) : Dxy = Dx,y, if \x-y\
= \x' - yf\}.
(2) Let r — ( r i , r 2 , . . . , r ^ ) £ (0,00)^. We say that r is ^-invariant if Ti — Tj whenever there exists g G Gs such that g(Fi(Vo)) = FJ(VQ). (3) For a symmetry g, D G CA(Vo) is said to be ^-invariant if Dxy = Dg(x)g(y) f° r any x, y G Vb. Also, if D G CA(Vo) is ^-invariant for any g G Gs, then D is said to be ^-invariant. Obviously, if r = ( 1 , . . . ,1) then r is ^-invariant. P r o p o s i t i o n 3.8.9. The collection of all Gs-invariant Laplacians equals Proof. Since Gs C O(d), D is ^-invariant for any D G Cs(Vo). Conversely, suppose that D G CA(Vo) is ^-invariant. Let x, y,x',y' G Vb and assume that \x — y\ = \x' — y'\. By (SS4), there exists g\ G Gs such that gi(x) = x'. Let y" = gi(y'). Then \x' - y'\ = \x' - y"\. By (SS2), there exists g2 G Gs such that ^2(^0 — x' and g2{y") = 2/'- Set g — g^°9\- Then p(a:) = x' and = 2/;. Hence Dxy = Dx*y>. So D G £ s (Vb). D Theorem 3.8.10. Le£ £ = (iiT, 5, { F j ^ s ) 6e strongly symmetric. If T = {riir2T • • ,TN) ^ (0,00)^ is Gs-invariant, then there exist D G £ s (Vb) and A > 0 such that (D,Ar) is a harmonic structure on C Moreover, the resistance form {£,T) constructed from the harmonic structure (D,Xr) is Gs-invariant, i.e., u o g G T and E{u,u) — £ (u o g, u o g) for any g G Gs and u G T'. Remark. In 3.1, we have seen that (J9, Ar) is a harmonic structure if and only if Ur{D) = XD (or equivalently H\r{D) = D), where Hr : CA(V0) -+
119
3.8 Nested fractals
CA(Vo) is the renormalization map on CA(Vo) defined by lZr(D) = [i?i]vbIt is easy to show that such a A is uniquely determined, (i.e., if there exist f D, D' e CA(Vo) and A, A' > 0 such that nr(D) = XD and 1Zr(D ) = \'D', then A = A'.) Note that Ur{Cs{Vo)) C C8(V0), which will be shown in Lemma 3.8.22. Using Sabot's arguments in [160], we can also show that the fixed point of 7l\v\cs{v0) 1S unique if (K, 5, {Fi}ies) is an affine nested fractal. See also Metz [129] for another proof of the uniqueness of the fixed point of Uxr\cs(Vo)' Remark. The harmonic structure (D, Ar) obtained in the above theorem is not always regular. For example, in Example 3.2.3, the self-similar structure (K, S, {Fi}ies) is an affine nested fractal and the harmonic structure (D, Ar) is (?5-invariant. In this example, (D, Ar) is not regular if t > y/b. Example 3.8.11 (Pentakun). Let S = {1,2,... ,5}. For k e 5, let pk = e2fc7Tv/=rr/5 e c a n d d e f i n e a contraction Fi : C -^ C by
where we naturally identify C with R 2 . The pentakunf is the self-similar set with respect to {Fi}ies- See Figure 3.2. The self-similar structure that corresponds to the pentakun is post critically finite. In fact
k=1
V — {k}kes and pk = 7r(h) for k e 5, where [i]5 e S is defined by [i]5 = i mod 5. It follows that Vo = {pk}kes is the collection of vertices of a regular pentagon. The pentakun is an affine nested fractal. In fact, Qs coincides with {g e O(2) : g(Vo) = F o }, which is generated by the reflections {gxy : x,y G Vo,x ^ y}. D e CA(Vo) belongs to CS(VO) if and only if
PiPj
Jα
if |i — jl = ± 1
mod 5
~ \b
i f | i - j | = ±2
mod 5,
Therefore,
where a, b > 0 with a + b > 0. In this case, r — (vk)k=i121... ,5 is ^-invariant if and only if n = ••• = r 5 . Let r = ( 1 , . . . ,1). By Theorem 3.8.10, there exist A > 0 and D £ Cs(Vo) such that 1Zr(D) = AD, or equivalently t In the same way, we can also define hexakun, heptakun, octakun and so on. 'kun' is 'Mr' in Japanese.
120
Construction of Laplacians on P. C. F. Self-Similar
Structures
Fig. 3.2. Pentakun (D, Ar) is a harmonic structure. In this case, it is also easy to check uniqueness of such A and D directly. (Use a symbolic calculus program, such as Mathematica or Maple V.) Example 3.8.12 (Snowflake). Let pk = e 2fc7r ^ /irT / 6 G C for k = 1 , . . . , 6 and let p7 = 0. We naturally identify C with R 2 . Define F{ : C -> C by Fk(z) = (z — Pk)/3+Pk for k — 1,2,... ,7. The snowflake is the self-similar set with respect to the contractions {Fk}k=i,2,...,7- See Figure 3.3. The self-similar structure associated with the snowflake is post critically finite. In fact, pk = IT(h) for k = 1,2,... ,7,
C=
(J
, k[k + 2] 6 , k[k + 3] 6 , k[k + 4 ] J ,
fc=l,2,...,6
fc=i,2,...,6, where [i]6 G { 1 , 2 , . . . ,6} is V = {fc}fc=i,2 6 and Fo = defined by [Z]Q =i mod 6. The snowflake is an affine nested fractal. In fact, Qs coincides with {g € 0(2) : g(V0) = Vo}, which is generated by the reflections {gxy : x,y G Vo,x ^ y}. In this case, we see that # { | x — y\ : x,y G Vo,x ^ y} = 3 and hence Cs(Vo) is a 3dimensional manifold. Also r = (rk)k=i,2,...j is ^-invariant if and only if n = • •. = r 6 . Let r = ( 1 , . . . , 1, i) for t>0. Then, by Theorem 3.8.10, for any t > 0 , there exist A > 0 and D G CS(VO) such that 7lr(D) = A£>, or equivalently, (D, Ar) is a harmonic structure.
3.8 Nested fractals
121
Fig. 3.3. Snowflake The next example is strongly symmetric but it is not an affine nested fractal. Example 3.8.13. Let (K, 5, {Fi}ies) be a connected p. c. f. self-similar set, where K C R3 = {(x,y,z) : x,y,z G R} and Fi : R3 -• R3. We assume that Vo is the set of vertices of a cube, that is, Vo = {((-1)*, ( - I F , (-1)*) : i,j, k € {0,1}}. This case is not included in the class of affine nested fractals, because 9pq(Vo) ^Voifp = ( - 1 , - 1 , - 1 ) and q = (1,1,1). Let Q = {g e O(3) : (y ) = V }. Then it follows that #(£) = 48. We define three subgroups g 0 o 1 2 ^°,G and Q of Q. First, let Q° be the group generated by the three reflections in the xy, yz and xz-planes. Then #(£°) = 8. Next, let Q 1 be the group generated by the four rotations by 2TT/3 around the lines x = y = z,x 2
= y = —z, x — —y = —z and x = — y = z. T h e n #{G X) 1
2
= 12.
Also let g be the group generated by G° and Q . Then #(£ ) = 24. Note that g2 is a proper subgroup of Q For example, gxy £ Q2 if x = (1,1,1) and y = (1,-1,-1). Also the rotation by TT/2 around the x-axis does not belong to Q2. Now if Qs includes Q2, then (K, 5, {Fi}ies) is strongly symmetric. Let us present a concrete example. Let {pi}i=i,2,... ,8 be the vertices of a cube. Also let PQ = 0. For any i = 1,2,... ,9, define Fi(a) = (a—pi)/3-\-pi for any a G R3. If K is the self-similar set with respect to {Fi} i=i)2,...,9, then (K,S,{Fi}i(zs), where S = {1,2,... ,9}, is a connected p.c.f.self-
122
Construction of Laplacians on P. C. F. Self-Similar
Structures
similar structure satisfying Vo = {pi}i=i,2,... ,8- It is easy to see that Qs 2 contains Q . In fact, Qs = Q. Moreover, r = (r^)j=i?2,... ,9 is ^-invariant if and only if r 1 = • • • = rg. In the rest of this section, we will prove Theorem 3.8.10. The idea of the proof is essentially due to Lindstr0m [116]. L e m m a 3.8.14. Let C*(VQ) be the collection of symmetric linear operators D from £(Vo) to itself which satisfy the following three conditions: (VI) DXy >§ ifx^y and ^2qeVo Dpq = 0 for any p G Vo. (V2) If \x-y\ = l0, then Dxy > 0. (V3) // \x - y\ > \x' - y'\ > 0, then Dxy < Dx>y>. Proof. Let D G £*(Vo). Let £D be the quadratic form on £(Vo) defined by £D(U,V) = —(u,Dv). Then obviously £& satisfies (DFl) because £D(U,U) = J2p qeVo Dpq(u(p) ~U(Q))2/^(Recall Definitions 2.1.1 and 2.1.2.) Assume that Du = 0 for u G £(V0). Then u(p) = u(q) if Dpq > 0. Now by (SSI), for any x, y G Vo, there exists a strict 0-walk between x and > 0 for any i = 1,2,... , n - l . 2/, say, (fft)i=i,2,...,n. Then, by (V2), Dx.Xi+1 Therefore u{x) = u(y). This shows (DF2). Obviously D satisfies (L3) and hence £D satisfies (DF3). By Proposition 2.1.3, it follows that D G £A(V0). f Moreover, by (V3), Dxy = Dx>y> if \x-y\ = \x -y'\. Hence D G CS(VO). D For ( ^ Γ ) G CA(VO) X (0,OO)^, we defined H± G CA{Vi) in Definition 3.1.1. Recall that the renormalization operator lZr is defined by nr{D) = [H^. D e f i n i t i o n 3.8.15. For any x,y G V\ a n d n > 0 , define Pn(x,y)
= { x = (xo,xu...
, x n , x n + i ) :x0 = x,xn+i
=y,
x is a 1-walk between x and y and xi G Vi\Vb for any i = 1, 2,... , n.} Moreover, let D G £A(Vo). Pn(x,y) by
Define /i(x) for x = ( x o , x i , . . . ,xn,xn+i)
, / x _ hXQXlhXlX2
•-•
hXri_lXnhXnXri+1
r*'X\X\'*'X2X2 ' ' ' '*'XriXn
where =
Pq
f(i?i)p 9
if p^ q,
l - ( F i ) p p if p = q.
In particular, if n = 0, then Po(x,y) = {(x,y)}
and h((x,y)) = hxy.
G
3.8 Nested fractals
123
The following lemma holds for any connected p. c. f. self-similar structure and any (L>,r) G CA(V0) x (0,00)*. Lemma 3.8.16. For any x 7^ y G Vo, ^—> n>0
^—> x.ePn(x,y)
Proof. We decompose Hi into the following form: Hl
=
(j
x
where T : t(V0) - *(Vb), J : *(Vb) - €(Vi\V 0 ) and X : (Vi\F0) -> £(Vi\Vo). Then, by Theorem 2.1.6, it follows that [H1]v0=T-tJX~1J. Note that hpp = —(Hi)pp > 0 and hpp > hpq for any p,q G Vi\Vb. Define A : ^(Vfc) -+ ^(Vb) by A p p - hpp = -Xpp for any p e V^Vo and Apq = 0 if p / g. Also define L : ^V^Vb) -> ^(F^Vb) by L pq = / i p q / / i ^ for any Pi Q € ViVK)- Then X = - 2 ( 7 - L/2)A, where / is the identity map. Therefore, ( - X ) " 1 - A " 1 ^ Lβ)"1^. Recall the proof of Lemma 3.5.1. Let C be the counterpart of the C defined there. Define an equivalence relation ~ on Vi\Vb by p ~ q if and only if there exists x G Un>oPn(x, y) such that ft(x) > 0. Let (Vb\Vi)/~ = {[/i,... , C/m}, where C/"i C Vb\Vi is an equivalence class. Then X and C are decomposed as J*Q : £(Ui) —• ^(C/"i) and C* : ^(f/i) —> £{Ui), respectively, for i = 1, 2 , . . . , m. Then, by the argument in the proof of Lemma 3.5.1, it follows that | | ( C t ) n i | | < 1 for some m G N. This implies that | | C n | | < 1 for some n. Note that | | C | | < 1. Since L = (C + J), we obtain | | ( L / 2 ) n | | < 1. Hence,
(-X)-1 = A-\I - L/2)"V2 = A-1 J2 2-(fc+1)Lfe, where L° = /. Hence, for p, g G Vi\Vb, )pq
This immediately implies the desired expression of TZr(D)xy.
D
Remark. Let (K, S, {Fi}ie$) be strongly symmetric and let D G £*(Vb). Then we can show that X itself satisfies the assumptions of Lemma 3.5.1. (This fact is not used afterwards.)
124
Construction of Laplacians on P. C. F. Self-Similar
Structures
L e m m a 3.8.17. Let D G £*(Vo) and let x,y,z G Fs(Vo) for some s G S. If \x -y\ <\x - z\, then hxy > hxz. Proof. First suppose that x — y. Then hxy
— hxx — —\Hi)xx
—
/ \H\)px ^ \Hi)xz pev1\{x}
—
hxz.
Next suppose that x ^ y. Then, by (3.8.2), there exist XQ, yo,zo £ Vo such that Fs(x0) = x,Fs(y0) = y,Fs(z0) = z,hxy = (ra)'1DXoyo and hxz = (7's)~1-Dxo«o" Since DXoyo > DXoZo, it follows that hxy > hxz. D Lemma 3.8.18. Let w G Wm and let x G K\Vm. Then x G ^ \ ^ ( V b ) if and only if J D Vm — ^ ( V o ) , where J is the connected component of K\Vm containing x. Proof. Suppose that x G Kw\Fw(Vo) and let J be the connected component of K\Vm containing x. Then, by Proposition 1.6.9, it follows that J — Kw\Fw(Vo). Hence, 7 n Vm = FW(VO). Conversely, if x £ Kw\Fw(Vo), there exists w' G Wm such that w ^ w' and x G Kw,\Fw,(Vo). Then In Vm = FW>(VO). Hence FW(VO) = FJ{V0). This is impossible because #(Fw(Vo) fl Fwt{Vo)) < 1 by the assumption (3.8.2). • Proposition 3.8.19. Let g be a symmetry of C. If Fwf(V0) = g(Fw(V0)) for w,w' G Wm, then Kw> — g(Kw). Proof. Let x G Kw\Fw(Vo) and let J be the connected component of K\V m containing x. Then g( J) is the connected component of K^m containing g(x). Since g{J) C\ Vm = Fw'(Vo), Lemma 3.8.18 implies that g{x) G Kw>. Hence g(Kw) C Kw>. Note that g~l is also a symmetry and g~1(Fwf (Vo)) = Fw(Vo). Using entirely the same arguments, we see that g(Kw) 3 Kw>. • Originally Fi was a map from K to itself. Prom now on, we think of Fi as a similitude on Rd. Proposition 3.8.20. For any w G W* and any g G Gs, set Ug,w = FW>~1
ogoFw,
where w' G W* is the unique word that satisfies Fw'(Vo) = g(Fw(Vo)). Ug,w egs. Note that Ug$ = g.
Then
3.8 Nested fractals
125
Proof. By Proposition 3.8.19, we see that g(Kw) = Kw>. Therefore, it follows that, for any v € W*, there exists v' G W* such that g(Fwv(Vo)) = Fw'v'(Vo). This implies that U9iW(Fv(Vo)) = Fv'(Vo). Hence U9jW is a symmetry. Note that U9iW is a linear transformation. Also, from the definition, U9iw preserves distance. Hence U9iW G O(d). D Let A(-, •) G VT{ym)' We say that A(-, •) is ^-invariant if and only if A{u ° g,v o g) = A(u, v) for any u , u G i(Vm) and any g € GsCorollary 3.8.21. Let D G C8(V0). Ifr is gs-invariant, then £<m)(-, •) is Qs-invariant, where S^ G VFiVm) is defined in Definition 3.1.1. Proof. We use induction on m. First £(°' = £& is <5S-invariant. Now assume that £( m ) is ^-invariant. Let g G Qs. By (3.1.1), iV
£
(m+1)
(wo^,vo^) = y
—f(m) {uogoFuvogoFi)
(3.8.3)
for ix,i; G £(Vm+i). Now let g(Ki) = Kj. Then r^ = r^ because r is ^-invariant. Also the induction hypothesis along with Proposition 3.8.20 implies
£ ( m ) (uogoFi,vogoFi)=
£{m) (u o Fj o Ug4,v o F, o U9ii)
Again, by (3.1.1), it follows that ^ m + 1 ) ( « o 5 , U o 5 ) = S^+^^u).
•
Lemma 3.8.22. Ifr is Qs-invariant, then fcr(£s(Vo)) C £ e (Vb). Proof. Let D G £ S (F O ). By Corollary 3.8.21, Hi is ^-invariant. Hence (Hi)pq = (Hi)g{p)g(q) for any p,q G Vi and any g e Gs. Now, as we have seen in the proof of Proposition 3.8.9, if x,xf,y,y' G Vo and |x — y\ = |x' — 2/'|, then there exists # G Qs such that ^(x) — x' and ^(y) = y'. Define #(x) = (^(xi))i=o,i,..., n +i for x = (xi)i=o,i,..., n +i € Pn(x,y). Then ^ is a bijective mapping between Pn(x,y) and P n (x / ,y / ). Also it follows that ft(x) = h(g(x)) because i?i is ^-invariant. Hence, by Lemma 3.8.16, y,. D Lemma 3.8.23. Let p^ q G Vo and assume that gpq G Qs. If x,y G i for some w e W*, \x — q\ < \x—p\ and \y — q\ > \y—p\, then gpq(Kw) = Kw. Proof. Let J = K-u;\Ft;(;(Vro). Then, as K\Vo is connected, J is a connected component of K\Vm. If gpq{Kw) = Kw>, then ^ ( J ) = KW\FW>(VQ) and 9pq(J) is also a connected component of K\Vm. By Proposition 1.6.6, there exists a continuous path between x and y contained in J U {x,y}. Since
126
Construction of Laplacians on P. C. F. Self-Similar
Structures
x and y belong to different sides of the hyperplane Hpq, this continuous path between x and y intersects Hxy. Hence JnHpq ^ 0. By the fact that gpq(JnHpq) = JnHpq, we see that Jngpq(J) ^ 0. Hence J = gpq(J) and therefore w = w1. • Lemma 3.8.24. Let a{ G R for i = 1, 2,... , n. Then,
J:JC{l,2,...,n}
jeJ
i=1
Lemma 3.8.25. If r is Qs-invariant, then TZr(C*(Vo)) C £*(Vb). Proo/. Let D G £*(Fo)- Since £> G £ e (Vb), Lemma 3.8.22 implies that Tip(D) G £s(Vfc). Therefore we need to show (V3). Now by (SS3), there exist x, y, z G VQ such that \x — y\ — Z&, |a: — J2T| = Z^+i and ^ G ^ s . Let V = {p G Rd : \p - y\ < \p - z\} and define T : Rd -+ Rd by
1
(9yz(p) otherwise. Note that x G V and T(x) = x. Also for x = (#i)i=o,...,n+i €Pn(x,z) U P n (x, z), define T(x) = (T(xi))i=0,... ,n+i- Then, by Lemma 3.8.23, T(x) G Pn(a;,y)Now let x = (^)^=o,... ,n+i ^ ImT and let us consider T~1(x). Define /(x) — {i G {1,... , n + 1} : there exists s € S such that Xi-ugyz(xi)
e ^(Vb) and ^ ^ i7 y2 }.
For J C /(x), we define r(x, J) = (2/i)i=0|... ,n+i by ^ = ( 5 (i) = # { j £ J : j
127
Notes and references
where α ^ x ) = hXj_igyz{xj)/hXj_lXj. Since \xj-! - x5\ < \xj-X - gyz(xj)\, Lemma 3.8.17 implies that c*j(x) < 1. So, using Lemma 3.8.24, we obtain
E
xGlmT
E
J:JC/(x) #(J) is even.
M
xGlmT
J:JC/(x) #(J) is odd.
= E Mx) E (xelmT
J:JC/(x)
We also see that /ix?/ > /i x z . (If / i ^ > 0, then x,z G Fs(Vo) for some 5 G T^). Lemma 3.8.23 implies y G Fs(Vb). By Lemma 3.8.17, we obtain hXy > hxz.) Hence, by Lemma 3.8.16, we verify that 7lr(D)xy > lZr(D)xz. This immediately shows (V3). • Proof of Theorem 3.8.10. Choose x,y G Vo so that \x — y\ — /0-
c*(Vo) = {D G cjy*V
D
*y =}}-
Define
K(Dl=
Let
nr(D)/nr(D)xy.
Then we see that IZ : £*(Vb) —> £*(Vo) and that IZ is continuous. Now C*(Vo) is homeomorphic to a compact ball in E 7 7 1 " 1 , where m = #{\p — q\ : p,q €Vo,p y^ q}. By Brouwer's fixed point theorem, it follows that IZ has a fixed point D G £*(y 0 ). Let A = 1Zr(D)xy. Then 7£r(£>) = A£>. Hence (L), Ar) is a harmonic structure. By Corollary 3.8.21, f ^ ) (= VTiVm) induced by the harmonic structure (Z), AΓ) is C/s-invariant for any k>0. Letting k —• oo, we verify that (f, J7) is ^-invariant. D
Notes and references The use of Dirichlet forms to study analysis on fractals was first developed in Fukushima & Shima [44], where they studied the asymptotic behavior of the eigenvalues of the standard Laplacian on the Sierpinski gasket. Then, in [107] and [83], such an approach using Dirichlet forms was extended to more general classes of (finitely ramified) self-similar sets. See also Fukushima [41], Kusuoka [108], Kumagai [100] and Metz [125, 128]. Although the main source of this chapter is [83], we add many results for the case of non-regular harmonic structures. 3.1 The notion of harmonic structures was introduced in [83]. The counterpart of the renormalization map lZr on £>.F(Vo) had already been intro-
128 Construction of Laplacians on P. C. F. Self-Similar Structures duced in Hattori, Hattori & Watanabe [70], Kusuoka [107] and Lindstr0m [116]. In particular, Kusuoka constructed a self-similar Dirichlet form from a fixed point of such a renormalization map in [107]. Also, Lindstr0m constructed a diffusion process (called Brownian motion) on a nested fractal from decimation invariant random walks in [116]. The fixed point problem of lZr has been studied by many authors, for example, Lindstr0m [116], Metz [126, 127, 130, 129] and Sabot [160]. All of those works focus on (affine) nested fractals. The existence of the fixed point was shown by Lindstr0m in [116]. Moreover, Sabot proved the uniqueness of the fixed point in [160]. See Theorem 3.8.10 and the remark after it. Also, Metz obtained a kind of stability of the fixed point in [129]. 3.2 Theorems 3.2.8, 3.2.11 and 3.2.14 have not been explicitly stated elsewhere before. 3.3 Self-similarity of Dirichlet forms, Proposition 3.3.1, was first observed in Fukushima [41] for nested fractals. Theorem 3.3.4 is an extension of [85, Theorem 3.1]. Theorem 3.3.8 is a new result. 3.4 The key ideas of the proof of Lemmas 3.4.3, 3.4.4 and 3.4.5 are due to Kumagai [100]. Those lemmas play an essential role in proving Theorem 3.4.6 when the harmonic structure is not regular. 3.5 The explicit definition of the Green's function (as we see in Proposition 3.5.5) was given in [83]. Theorem 3.5.7 is a new result. In [94], one can find a figure of the graph of the Green's function for the standard harmonic structure on the Sierpinski gasket. 3.7 The explicit definition of the Laplacian in Definition 3.7.1 was given in [83]. Theorem 3.7.9 has been known to specialists for sometime. In [44], Pukushima & Shima showed this fact for the standard Laplacian on the Sierpinski gasket with Dirichlet boundary condition. Exercises Exercise 3.1. Let C be the harmonic structure associated with the Sierpinski gasket. (See Example 3.1.5.) Set D= and r = (s,st,st), where /i, s and t are positive real numbers. Show that if we fix h > 0, there exist unique s and t such that (D, r) becomes a harmonic structure on the Sierpinski gasket. Also prove that (D, r) is a regular harmonic structure.
Exercises
129
Hint: let R = 1/h. Then calculate the effective resistances for D and Hi by using the A-Y transform (Lemma 2.1.15). Then apply Corollary 2.1.13. You will find that the condition for (D, r) being a harmonic structure is
Exercise 3.2 (modified Sierpinski gasket). Let {pi}i=i,2,3 be the vertices of a regular triangle in the complex plane C. Set p4 = (p2 +ps)/2, Ps — (Ps + Pi)l^ a n d PG = {Pi H-p2)/2. Choose real numbers a and P so that 2a + ft = 1 and a > /? > 0. We define F;(z) = a(z - pi) + Pz for i = 1,2,3 and Fi(z) = β^z—pi) -\-pi for i = 4,5,6. Let X be the self-similar set with respect to {Fi}ies, where S = {1,2,... , 6}. (1) Prove that the self-similar structure C associated with K is post critically finite with Vo = {pi,P2,P3}> (2) Define D € CA{V0) by
Let r = (r, r, r, re, rs, rs) where r, 5 > 0. Show that if we fix s, then there exists an unique r such that (D,r) is a harmonic structure on C Is this harmonic structure regular? Hint: use the A-Y transform and calculate effective resistances as in Exercise 3.1. Exercise 3.3. Calculate A\ for the harmonic structures on Hata's tree-like set (Example 3.1.6) and the modified Sierpinski gasket (Exercise 3.2).
Exercise 3.4 (Neumann problem for Poisson's equation). Prove that, for / G C(K) and rj G ^(Vo), there exists u G Vμ such that
\{du)p
= Tj(p) for all peV^
if and only if JK ^μ = T,peVo rj(p)Exercise 3.5. Let V C Vo and assume that V i=- 0. Define Tγ = {u G T : (1) Prove that (Tγ, £) is a Hilbert space. Also show that (£, JV) is a regular Dirichlet form on L2(K, μ). (2) Show that, for any / G L2(K^), there exists ix G ^V such that £(v,u) = (y^ Dμ for any v G JV- Denote tx by Gy/. Then prove that
130 Construction of Laplacians on P. C. F. Self-Similar Structures 2 1 Gγ is a bounded operator from I? [K^μ) to L (lf, μ) and Gγ = Hγ , where Hγ is the non-negative definite self-adjoint operator associated with the Dirichlet form (£,JV) on L2(K^). (3) Set Vγ = {u e Vμ : u\v = 0, (du)p = 0 for any p e V0\V}. Then prove that Vγ C Dom(ifvr) and Hv\vv = — A M |x> v . Also show that Hγ is the Friedrichs extension of —/S.μ\x>v^
4 Eigenvalues and Eigenfunctions of Laplacians
In this chapter, we will study eigenvalues and eigenfunctions for the Laplacian AM associated with (D, r) and μ. In particular, we will be interested in the asymptotic behavior of the eigenvalue counting function and present a Weyl-type result (Theorem 4.1.5) in 4.1. It turns out that the nature of eigenvalues and eigenfunctions of Aμ is quite different from that of Laplacians on a bounded domain of R n . For example, we will find localized eigenfunctions in certain cases. More precisely, in 4.3, we will define the notion of pre-localized eigenfunctions, which are the eigenfunction of Aμ satisfying both Neumann and Dirichlet boundary conditions. It is known that such an eigenfunction does not exists 1 for the ordinary Laplacian on a bounded domain of W . Proposition 4.3.3 shows that if there exists a pre-localized eigenfunction, then, for any open set O C K, there exists a pre-localized eigenfunction whose support is contained in O. One important consequence of the existence of pre-localized eigenfunctions is the discontinuity of the integrated density of states. See Theorem 4.3.4 and the remark after it. We will give a sufficient condition for the existence of pre-localized eigenfunctions in 4.4. In particular, we will see that there exists a pre-localized eigenfunction for the Laplacian on an affine nested fractal associated with the harmonic structure appearing in Theorem 3.8.10. See Corollary 4.4.11. Throughout this chapter, C — (K, 5, {Fi}i£s) is a connected p. c. f. selfsimilar structure with S = {1, 2,... , N} and (D, r) is a harmonic structure on £, where r = (ri)ies. Also μ e M(K) and s u p ^ G ^ I2m>oR™(vw) < oo. Note that if // is a self-similar measure with weight (/i^)^^, then the 7 above condition on μ is equivalent to μ^i < 1 for any i G S. We use {£, J ) to denote the resistance form associated with (Z), r). Also d is a metric on 131
132
Eigenvalues and Eigenfunctions of Laplacians
K that is compatible with the original topology of K. (Note that (K, d) is compact.)
4.1 Eigenvalues and eigenfunctions In this section, we first define eigenvalues and eigenfunctions of the Laplacians Aμ. Then we will define the eigenvalue counting function and present a theorem which gives the asymptotic behavior of the eigenvalue counting function.
Definition 4.1.1. Let EN{\) = {ipeVN
: A^
= -A?}. IidimEN(\)
>
1, then A is called a Neumann eigenvalue (N-eigenvalue for short) of Aμ and any non-trivial ip G E^ (A) is called a Neumann eigenfunction (Neigenfunction for short) of Aμ belonging to the Neumann eigenvalue A. Also define ED(X) = {
(£, f) (or (£,f0)) on
L2(K^).
P r o p o s i t i o n 4.1.2. Suppose that X^m>o-^m(/i) < ° ° f Then the following three conditions are equivalent. (Nl) if G Dom(iifjv) o/ad Hjycp = \tp, (N2) (f G T and S(cp, u) = X((p, ^) M for any u G T, (N3) ipeEN(\). Also, the following three conditions are equivalent. (Dl) (f G Dom(#£)) and HDV = X(p, (D2) (f G TQ and £(
o r s o m e
t > 0 .
r Remark. Let /z be a self-similar measure. Then 5 ] m > 0 ^ M < °° f° some t > 0 if and only if r ^ < 1 for all i G S. See the remark after Corollary 3.6.6.
Proof. By Corollary 3.6.6, # ( n ) is continuous on K x K for n > (1 -
Recall that G^c^d) (GDnu)(x)
l/t)~l.
= GD\c(K,d)- Hence, = (G^u)(x) = f gW(x,y)u(yMdy) JK
(4.1.1)
4-1 Eigenvalues and eigenfunctions
133
U 2 for u G C{K,d). Since GD is a bounded operator from L (K, μ) to itself and g^ is continuous, it follows that (4.1.1) is true for all u G L2(K, μ). This implies that GDnu G C(K, d) for any u G L2(K, μ). Now the equivalence between (Dl) and (D2) is obvious from the characterization of a self-adjoint operator associated with a closed form in Lemma B.1.4. Also, by Theorem 3.7.9, (D3) implies (Dl). Now suppose (f G Dom(Hz)) and Hrup = X(p. Then, as XnGDny> = <£> we see that if G C(K,d). Hence XGDV = XGμ(p = (p. So applying Lemma 3.7.10, we see that (p G Vp. Therefore (p G ED(X). Next we consider the Neumann case. It follows immediately that (Nl) and (N2) are equivalent by Lemma B.1.4. Assume that ip G Dom(if/v) and H]y(p = \(f. Let (po = Y^pev0 ^p(p)^p' Then we see that (P—
By Theorem 3.4.6 and Corollary 3.4.7, the self-adjoint operators and HD have compact resolvents. Therefore, N-eigenvalues (and also Deigenvalues) are non-negative, of finite multiplicity and the only accumulation point is oo. In other words, for b = D and N, there exist {A^}^>i and ip\ G Eh(\\) such that
2
and {<£i}i>i is a complete orthonormal system for L (K, μ). So we can define eigenvalue counting functions as follows. Notation. Hereafter, the symbol b always represents a boundary condition: b is either N or D, where N represents the Neumann boundary condition and D represents the Dirichlet boundary condition. Definition 4.1.3 (Eigenvalue counting function). Define p&(x,/x) for b = D and N by pb(x^) = ^ A < x ^ i m ^ ( ^ ) * Pb{x,v) is called the beigenvalue counting function. Obviously, p\,(x^) = max{A: : Xbk < x} and pb(x^) —> oo as x —» oo. For ordinary Laplacians on bounded domains in R n , we have Weyl's famous theorem on the asymptotic behavior of eigenvalue counting functions. Theorem 4.1.4 (WeyPs theorem). Let Q, be a bounded domain in Rn. Let Xi be the i-th eigenvalue of the Dirichlet eigenvalue problem of —A on
134
Eigenvalues and Eigenfunctions
of Laplacians
Q, that is, ( Au = -Xu where A = Y^=1 d2 /dx2.
Set p(x) — # { i : A^ < x). Then as x —> oo,
n — (97r\~ lOl Jb srn/2 _|i_ OyX n(^rn/^\J , n(rr\J — P\Jb \^" J R&n\^L\ n
w/iere | • \n is n-dimensional
Lebesgue measure
and Bn — \{x : \x\ < l } | n .
Remark. Weyl [182, 183] proved the above result under some extra conditions on the domain D. It is now known that the above result is true for any bounded domain. See Lapidus [111]. We are interested in an analogue of Weyl's theorem for our Laplacian Aμ, or more specifically, the asymptotic behavior of the eigenvalue counting functions p&(x,/i) as x —> oo. If /i is a self-similar measure, we have the following result. Theorem 4.1.5. Let μ be a self-similar measure with weight (^i=1,2,... ,NAssume that μ^i < 1 for all i G {1,2,... , iV}. Let ds be the unique real number d that satisfies N
2=1 /
where 7^ = y r^7- Then 0 < liminf ph(x^)/xds/2
< limsuppb(x^)/xds/2
< 00
for b = D and N. ds is called the spectral exponent of (£,T^μ). Moreover (1) Non-lattice case: If Y^i=1 Zlog7i is a dense subgroup ofR, then the limit limx_^oo pb(x^)/xds/2 exists and is independent of the boundary conditions. (2) Lattice case: IfY^=1 Zlog7; is a discrete subgroup ofR, let T > 0 be its generator. Then, as x —> 00,
where G is a right-continuous, T-periodic function satisfying 0 < inf G(x) < supG(x) < 00 and o(l) is a term which vanishes as x —> 00. Moreover, the periodic function G is independent of the boundary conditions. Remark. In the proof, we will get more concrete expressions for the limit for the lattice case and the periodic function G for the non-lattice case. See (4.1.4) and (4.1.5).
135
4-1 Eigenvalues and eigenfunctions
Remark. If we use the full version of the renewal theorem (Theorem B.4.3), ds 2 in the we get more detailed information about the error term o(l)x ^ lattice case. For example, if the 7^ are all equal, then pb(x^)
= G(logx/2)xds^2 + O(1)
(4.1.3)
a s x - ^ o o , where O(l) is a term which is bounded as x —» 00. Remark. According to [41], G((logx)/2)xds^2 is the integrated density of states of K associated with the Laplacian A^. Example 4.1.6 (Sierpinski gasket). For the standard Laplacian on the Sierpinski gasket in Example 3.7.3, r* = 3/5 and μ^ = 1/3 for all i = 1,2,3. Let ds be the spectral exponent. Then, by (4.1.2), we have ds = log 9/log 5. In this case, ds is also called the spectral dimension of the Sierpinski gasket. Note that ds differ