SPECTRAL ASYMPTOTICS OF ONE-DIMENSIONAL GRAPH-DIRECTED SELF-SIMILAR MEASURES WITH OVERLAPS
SZE-MAN NGAI AND YUANYUAN XIE
Abstract. For the class of graph-directed self-similar measures onR, which could have over- laps but are essentially of finite type, we set up a framework for deriving a closed formula for the spectral dimension of these measures. For the class of finitely ramified graph-directed self- similar sets, the spectral dimension of the associated Laplace operators has been obtained by Hambly and Nyberg [7]. The main novelty of our result is that the graph-directed self-similar measures we consider do not need to satisfy the graph open set condition.
Contents
1. Introduction 1
2. Graph-directed iterated function systems and measures essentially of finite type 6
2.1. Graph-directed iterated function systems 6
2.2. The essentially finite type condition for the graph-directed self-similar measure 7
3. Eigenvalue counting function 8
3.1. Eigenvalue counting function onR 8
3.2. Unitarily equivalent operators 9
4. Renewal equation and Proof of Theorem 1.1 and Theorem 1.3 9
4.1. Renewal equation 9
4.2. Proof of Theorem 1.1 and Theorem 1.3 13
5. Strongly connected GIFSs onR 14
5.1. An example of strongly connected GIFS 15
5.2. Spectral dimension ofµin Example 5.1. 18
6. GIFSs that are not strongly connected 21
6.1. An example of a GIFS that is not strongly connected 21
6.2. Spectral dimension ofµin Example 6.1. 24
References 27
1. Introduction
Let U ⊆Rd be a bounded domain with smooth boundary, ∆ be the Dirichlet Laplacian on U, {λn} be the eigenvalues of −∆, and N(λ,−∆) be the number of eigenvalues that do not
Date: October 19, 2018.
2010Mathematics Subject Classification. Primary: 28A80, 35P20; Secondary: 35J05.
Key words and phrases. Fractal, spectral dimension, graph-directed self-similar measure, essentially of finite type.
The authors are supported in part by the National Natural Science Foundation of China, grant 11771136, and Construct Program of the Key Discipline in Hunan Province. The first author is also supported in part by the Hunan Province Hundred Talents Program and a Faculty Research Scholarly Pursuit Funding from Georgia Southern University.
1
exceedλ. Weyl [25] proved the following asymptotic formula for the Dirichlet Laplacian:
N(λ,−∆) = Bd
(2π)d|U|λd/2+o(λd/2) = 1
(4π)d/2Γ(d/2 + 1)|U|λd/2+o(λd/2), (1.1) where |U| denotes thed-dimensional volume of U and Bd is the volume of the unit ball in Rd. We point out that in [20, Equation (1.1)], the factor (2π)dis incorrectly typed as (4π)d/2.
There has been considerable interest in studying spectral dimension on various domains.
McKean and Ray [16] computed the spectral dimension of the Cantor measure. Fujita [5], Naimark and Solomyak [17] studied the spectral dimension of the self-similar measures satis- fying the open set condition (OSC) (see [9]). Kigami and Lapidus [12] obtained the spectral dimension of Laplacians on post-critically finite (p.c.f.) self-similar sets with a harmonic struc- ture. Croydon and Hambly [2, 6] studied the spectral dimension on the continuum random tree and random recursive affine nested fractals. For finitely ramified graph-directed self-similar set- s, Hambly and Nyberg [7] studied the spectral dimension of the associated Laplace operators.
Freiberg [4] investigated spectral asymptotics of generalized measure geometric Laplacians on Cantor like sets. Kajino [10, 11] studied asymptotics of the partition functions associated with self-similar sets. Alonso-Ruiz and Freiberg [1] obtained the spectral dimension of Laplacians on Hanoi attractors.
We say that an iterated function system (IFS) or a graph-directed iterated function system (GIFS), as well as any associated self-similar measure or graph-directed self-similar measure, have overlaps, if (OSC) or the graph open set condition (GOSC) (see Section 2.1) fails. In this case, it is much harder to compute the spectral dimension. For a class of IFSs on R with overlaps and satisfying second-order identities (see [23]), the first author [18] computed the spectral dimension of the corresponding measures. Tang and authors [20] defined measures that are essentially of finite type (EFT), a property describing the finiteness of basic measure types, and computed the spectral dimension of the Laplacian defined by a self-similar measure satisfying (EFT). The first author and Tang [19] computed the spectral dimension of a special class of graph-directed self-similar measures with overlaps. This paper studies the eigenvalue asymptotics of the Dirichlet Laplacians defined by graph-directed self-similar measure with overlaps in much greater generality.
Let Ω ⊆ Rd be a bounded open set, and µ be a positive finite Borel measure on Rd with supp(µ) ⊆Ω and µ(Ω)>0. We assume that the Poincar´e inequality (PI) for µ holds: There exists a constant C >0 such that
Z
Ω
|u|2dµ≤C Z
Ω
|∇u|2dx, for all u∈Cc∞(Ω). (1.2) (see, e.g., [8, 15, 17]) (PI) implies that each equivalence classu ∈H01(Ω) contains a unique (in the L2(Ω, µ) sense) memberu that belongs toL2(Ω, µ) and satisfies both conditions below:
(1) there exists a sequence {un} in Cc∞(Ω) such that un → u in H01(Ω) and un → u in L2(Ω, µ);
(2) u satisfies (1.2).
We call u the L2(Ω, µ)-representative of u. Define a mapping I : H01(Ω) → L2(Ω, µ) by I(u) = u. It is easy to see that I is a bounded linear operator, but not necessarily injective.
Consider the subspaceN ofH01(Ω) defined as N :=n
u∈H01(Ω) :kI(u)kL2(Ω,µ) = 0o .
It follows from the continuity of I that N is a closed subspace of H01(Ω). Let N⊥ be the orthogonal complement of N in H01(Ω). Then I : N⊥ → L2(Ω, µ) is injective. With a slight abuse of notation, we will denoteu by u.
Consider a nonnegative bilinear form E(·,·) in L2(Ω, µ) given by E(u, v) :=
Z
Ω
∇u· ∇v dx (1.3)
with domain domE =N⊥. (PI) implies that (E,domE) is a closed quadratic form onL2(Ω, µ).
Hence these exists a nonnegative self-adjoint operator on L2(Ω, µ), which we denote by −∆µ and call the (Dirichlet) Laplacian with respect to µ, such that domE = dom(−∆µ)1/2 and E(u, v) =h(−∆µ)1/2u,(−∆µ)1/2viL2(Ω,µ)for allu, v∈domE. Letu∈domE. Thenu∈dom∆µ holds if and only if there existsf ∈L2(Ω, µ) such thatE(u, v) =hf, viL2(Ω,µ)for all v∈domE, where−∆µu=f. We remark that if d= 1, then (PI) holds for any such µ, and thus −∆µ is well defined.
We assume thatL2(Ω, µ) is infinite dimensional. It is known (see, e.g., [8]) that there exists an orthonormal basis{ϕn}∞n=1ofL2(Ω, µ) consisting of the eigenfunctions of−∆µ. The eigenvalues λn=λn(−∆µ) satisfy 0< λ1 ≤λ2 ≤ · · · and lim
n→∞λn=∞.
Let N(λ,−∆µ) be the number of eigenvalues of −∆µ (counting multiplicity) which do not exceedλ, i.e.,
N(λ,−∆µ) := #{n:λn≤λ}, (1.4)
where #A denotes the cardinality of a set A. Define the lower and upper spectral dimensions of−∆µ(or µ), respectively, as
ds(−∆µ) := lim
λ→∞
2 lnN(λ,−∆µ)
lnλ and ds(−∆µ) := lim
λ→∞
2 lnN(λ,−∆µ) lnλ .
Ifds(−∆µ) =ds(−∆µ), the common value, denoted ds(−∆µ) (or ds(µ)), is called the spectral dimension of −∆µ (or µ); it measures the asymptotic growth rate of the eigenvalue counting function as well as the magnitude of then-th eigenvalue.
(EFT) is introduced in [20]. Let µ = Pq
i=1µi be the graph-directed self-similar measure defined by the GIFS G = (V, E) on Rd, where V = {1, . . . , q} and E = {ei : i ∈ V}. We say thatµ satisfies (EFT) (see Definition 2.1) if there exist a family of bounded open subsets {Ωi}qi=1 with Ωi ⊆Rd, supp(µi)⊆Ωi, andµ(Ωi)>0, and a finite familyB:={B1,`:`∈Γ}of measure disjoint cells,B1,`⊆Ωi` for some i` ∈V, such that for any `∈Γ, there is a family of µ-partitions {Pk,`}k≥1 of B1,` satisfying the following conditions: (1)P1,` ={B1,`}, and there exists someB ∈ P12,` such that B 6= B1,`; (2) for any k ≥ 2, P1k+1,` contains all cells in P1k,`
that are µ-equivalent to some cell inB; (3) the sum of theµ-measures of those cells B ∈Pk,`
that are notµ-equivalent to any cell in Btends to 0 as k→∞. In this case, we call{Ωi}qi=1 an EFT-family, B a basic family of cells, and (B,P) := ({B1,`},{Pk,`}k≥1)`∈Γ a basic pair. We say that (B,P) isregular if each cell B ∈S
k≥1,`∈ΓPk,` is connected, and for any`∈Γ, there exist some similitude τ`, some Ωj`, and some constant w(`) > 0 such that τ`(Ωj`) ⊆ B1,` and µ≥w(`)µ◦τ`−1 onτ`(Ωj`).
Let µ=Pq
i=1µi be the graph-directed self-similar measure defined by the GIFSG= (V, E) on R. Assume that µ satisfies (EFT) with {Ωi}qi=1 being an EFT-family and assume that there exists a regular basic pair (B,P) := ({B1,`},{Pk,`}k≥1)`∈Γ. Then we can derive renewal equations for the eigenvalue counting functions, and express them in vector form as:
f =f ∗Mα+z, whereα∈R, and
f =f(α)(t) = [f`(α)(t)]`∈Γ, t∈R;
Mα = [µ(α)`0`]`,`0∈Γ is a finite matrix of Borel measures onR; z=z(α)(x) = [z(α)` (x)]`∈Γ is a vector of error functions.
(1.5)
Let
Mα(∞) :=h
µ(α)`0`(R)i
`,`0∈Γ. (1.6)
For each`∈Γ andα≥0, define F`(α) :=X
`0∈Γ
µ(α)`0`(R), D`:={α≥0 :F`(α)<∞}, αe`:= infD`. (1.7) If the error functions decay exponentially to 0 ast→ ∞, then ds(µ) is given by the uniqueα such that the spectral radius ofMα(∞) is equal to 1.
Theorem 1.1. Let µ=Pq
i=1µi be a graph-directed self-similar measure defined by a strongly connected GIFSG= (V, E) on R. Assume that µ satisfies (EFT) with {Ωi}qi=1 being an EFT- family and there exists a regular basic pair. Let Ω = Sq
i=1Ωi, ∆µ be the Dirichlet Laplacian defined by µ, and let Mα(∞), F`(α) and α˜` be defined as in (1.6) and (1.7). Assume that for each`∈Γ, lim
α→˜α+`
F`(α)>1.
(a) There exists a unique α >0 such that the spectral radius of Mα(∞) equals 1.
(b) If we assume, in addition, that for the uniqueαin (a), there existsσ >0such that for all
`∈Γ, z`(α)(t) =o(e−σt) ast→ ∞, thends(µ) = 2α; moreover, ifMα(∞) is irreducible, then there exist positive constants C1 and C2 such thatC1λα ≤N(λ,−∆µ)≤C2λα for all sufficiently large λ.
In section 5, we illustrate Theorem 1.1 by the strongly connected GIFS G = (V, E) with V ={1,2} and E ={ei : 1 ≤i≤5}, wheree1, e3 ∈ E1,1, e2 ∈E1,2, e4 ∈E2,2, e5 ∈E2,1. The six similitudes are defined by
Se1(x) =ρx, Se2(x) =rx+ρ(1−r), Se3(x) =rx+ (1−r),
Se4(x) =rx+ (1−r), Se5(x) =ρx, (1.8)
where
ρ+ 2r−ρr≤1, (1.9)
i.e.,Se2(1)≤Se3(0) (see Figure 1).
Corollary 1.2. Let µ = µ1+µ2 be a graph-directed self-similar measure defined by a GIFS G = (V, E) in (1.8) together with a probability matrix (pe)e∈E. Then there exists a unique positive real numberα satisfying
1−(pe4r)α
(1−(pe1ρ)α)(1−(pe3r)α)− (pe1e3 +pe2e5)ρrα
− pe2e4e5ρr2α
= 0. (1.10) Moreover,ds(µ) = 2α.
Theorem 1.3. Letµ=Pq
i=1µibe a graph-directed self-similar measure onRdefined by a GIFS G= (V, E) that is not strongly connected. Assume thatGhasγ strongly connected components.
Form = 1, . . . , γ, let 2αm be the spectral dimension of the graph-directed self-similar measure corresponding to the mth strongly connected component, and let
SCm:={i∈V :iis contained in the mth strongly connected component}. (1.11) Assume that µ satisfies (EFT) with {Ωi}qi=1 being an EFT-family and there exists a regular basic pair. LetΩ =Sq
i=1Ωi and∆µ be the Dirichlet Laplacian defined byµ. LetMα(∞), F`(α) andα˜` be defined as in (1.6)and (1.7). Assume that for m= 1, . . . , γ, each i∈SCm, and each
`∈Γi, lim
αm→˜α+`
F`(αm)>1.
(a) There exists a unique α= max{αm:m= 1, . . . , γ}>0 such that the spectral radius of Mα(∞) equals 1.
(b) If we assume, in addition, that for the unique α in (a), there exists σ > 0 such that for all ` ∈ Γ, z`(α)(t) = o(e−σt) as t → ∞, then we have ds(µ) = 2α; moreover, if Mα(∞) is irreducible, then there exist positive constants C1 and C2 such that C1λα ≤ N(λ,−∆µ)≤C2λα for all sufficiently large λ.
In section 6, we illustrate Theorem 1.3 by a GIFSG= (V, E) that is not strongly connected, withV ={1,2} andE ={ei : 1≤i≤5}, where e1, e2, e3∈E1,1, e4∈E2,2, e5 ∈E2,1. The five similitudes are defined by
Se1(x) =ρx, Se2(x) =rx+ρ(1−r), Se3(x) =rx+ (1−r),
Se4(x) =rx+ (1−r), Se5(x) =ρx, (1.12)
whereρ+ 2r−ρr ≤1, i.e., Se2(1)≤ Se3(0) (see Figure 5). For a probability matrix (pe)e∈E, we define
w(k) :=pe1
k
X
j=0
pje2pk−je3 , k≥0, (1.13) Corollary 1.4. Let µ=µ1+µ2 be a graph-directed self-similar measure defined by the GIFS G= (V, E) in (1.12) together with a probability matrix (pe)e∈E, and let w(k) be defined as in
(1.13). Then there exists a unique positive real number α satisfying (pe2r)α+ (pe3r)α+ 1−(pe2r)α
1−(pe3r)α
∞
X
k=0
(w(k)ρrk)α = 1. (1.14) Moreover,ds(µ) = 2α.
This paper is organized as follows. In Section 2, we give a modified version of the definition of (EFT). In Section 3, we introduce some properties of the eigenvalue counting function. In Section 4, we derive renewal equations and prove Theorem 1.1 and Theorem 1.3. Section 5 illustrates Theorem 1.1 by a class of one-dimensional strongly connected GIFSs defined as in (1.8); we also prove Corollary 1.2. In Section 6, we study the one-dimensional GIFSs in (1.12), which are not strongly connected, and prove Corollary 1.4.
2. Graph-directed iterated function systems and measures essentially of finite type
2.1. Graph-directed iterated function systems. Agraph-directed iterated function system (GIFS) of contractive similitudes is an ordered pairG= (V, E) described as follows (see [14]).
V is a set of vertices labeled by{1, . . . , q} and E is a set of directed edges with each begining and ending at a vertex. It is possible for an edge to begin and end at the same vertex and we allow more than one edge between two vertices. To each edge e∈E, there corresponds a contractive similitudeSe(x) :Rd→Rd defined as
Se(x) =ρeRex+be,
whereρe∈(0,1) is the contraction ratio,Reis an orthogonal transformation, andbe∈Rd. Let Ei,j denote the set of all edges that begin at vertexiand end at vertexj. We calle=e1. . . ek
a path with length k, if the terminal vertex of each edgeei (1 ≤i≤ k−1) equals the initial vertex of the edgeei+1. It is well known that there exists a unique family of nonempty compact setsK1, . . . , Kq satisfying
Ki=
q
[
j=1
[
e∈Ei,j
Se(Kj), i= 1, . . . , q. (2.15) Define
K:=
q
[
i=1
Ki. (2.16)
We callK the graph self-similar set associated with G = (V, E). Assume that for each edge e∈E, there corresponds a transition probabilitype>0, and the weights of all edges leaving a given vertexisum to 1, namely,
X
j∈V
X
e∈Ei,j
pe= 1. (2.17)
Then for eachi∈V, there exists a unique Borel probability measuresµi such that µi=
q
X
j=1
X
e∈Ei,j
peµj◦Se−1. (2.18)
We note that supp(µi) =Ki for alli∈V. Finally, letµ:=Pq
i=1µi and call it agraph-directed self-similar measure. We say that G = (V, E) satisfies the graph open set condition (GOSC) (see [24]) if there exists a family {Oi}qi=1 ⊆Rd of nonempty bounded open sets such that for alli= 1, . . . , q,
[
e∈Ei,j
Se(Oj)⊆Oi and Se(Oj)∩Se0(Oj) =∅ for all distinct e, e0 ∈Ei,j.
It is obvious that Ki ⊆ Oi, i.e., supp(µi) ⊆ Oi. A GIFS, as well as any associated graph- directed self-similar measure, are said to be have overlaps if (GOSC) fails. Let {Ωi}qi=1 be a family of nonempty bounded open subsets of Rd. We say that {Ωi}qi=1 is invariant under the GIFS G = (V, E) if S
e∈Ei,jSe(Ωj) ⊆ Ωi for i = 1, . . . , q. We say G is connected if for each pair of vertices i, j∈V, there is a (non-directed) path between them. G is said to bestrongly connected if for each pair of vertices i, j ∈V, there is a directed path from ito j. A strongly connected component of G is a maximal subgraph H of G such that H is strongly connected.
Strongly connected components are pairwise disjoint and do not necessarily cover G. A single vertex may be a strongly connected component if it loops to itself. In this paper, we assume that each graph has at least one strongly connected component.
2.2. The essentially finite type condition for the graph-directed self-similar measure.
Let Ω⊆Rdbe a bounded open subset andµbe a positive finite Borel measure with supp(µ)⊆Ω and µ(Ω)>0. We call a µ-measurable subset U of Ω is a cell (in Ω) ifµ(U)>0. Clearly, Ω itself is a cell.
We say that two cellsU andV areµ-equivalent, denoted byU 'µ,τ,w V (or simplyU 'µV), if there exist some similitudeτ :U →V and some constant w >0 such thatτ(U) =V and
µ|V =wµ|U◦τ−1. (2.19)
It is easy to check that'µ is an equivalence relation.
LetU ⊆Ω be a cell. Two cellsV, W inU aremeasure disjoint with respect toµifµ(V∩W) = 0. We call a finite familyPof measure disjoint cells a µ-partition ofU ifV ⊆U for allV ∈P, and µ(U) = P
V∈P
µ(V). A sequence of µ-partitions {Pk}k≥1 is refining if for anyV ∈ Pk and any W ∈ Pk+1, either W ⊆ V or they are measure disjoint, i.e., each member of Pk+1 is a subset of some member ofPk.
Let B := {B1,`}`∈Γ be a finite family of measure disjoint cells in Ω, and for each `∈Γ, let {Pk,`}k≥1 be a family of refiningµ-partitions of B1,` withP1,`:={B1,`}. We divide eachPk,`, k≥ 2, into two (possibly empty) subcollections, P1k,` and P2k,`, with respect to B, defined as follows:
P1k,`:=
B ∈Pk,`:B'µB1,i for somei∈Γ , P2k,`:=Pk,`\P1k,`=
B ∈Pk,`:B /∈P1k,` . (2.20) Definition 2.1. We say that a graph-directed self-similar measure µ = Pq
i=1µi on Rd is essentially of finite type (EFT) if there exist a family of bounded open subsets {Ωi}qi=1 with Ωi ⊆Rd, supp(µi)⊆Ωi and µ(Ωi)>0, and a finite family B:={B1,`}`∈Γ of measure disjoint
cells,B1,`⊆Ωi` for some i`= 1, . . . , q, such that for any`∈Γ, there is a family of µ-partitions {Pk,`}k≥1 of B1,` satisfying the following conditions:
(1) P1,`={B1,`}, and there exists some B ∈P12,` such that B6=B1,`;
(2) if for some k≥2, there exists some B∈P1k,`, then B∈P1k+1,` and henceB ∈P1m,` for allm≥k;
(3) lim
k→∞
P
B∈P2k,`
µ(B) = 0.
HereP1k,`andP2k,`(k≥2) are defined as in (2.20). In this case, we call{Ωi}qi=1anEFT-family, B a basic family of cells, and(B,P) := ({B1,`},{Pk,`}k≥1)`∈Γ a basic pair.
For k≥2 and`∈Γ, let Pk,`={Bk,`,i, i= 1,2, . . .}.
Definition 2.2. Assume that a graph-directed self-similar measureµ=Pq
i=1µisatisfies (EFT) with {Ωi}qi=1 being an EFT-family and(B,P) := ({B1,`},{Pk,`}k≥1)`∈Γ being a basic pair. We say that (B,P) is regular if each cell B ∈S
k≥1,`∈ΓPk,` is connected, and for any`∈Γ, there exist some similitude τ`, some Ωj` and some constant w(`) > 0 such that τ`(Ωj`) ⊆ B1,` and µ≥w(`)µ◦τ`−1 onτ`(Ωj`). In this case, we call B a regular basic family of cells.
3. Eigenvalue counting function
3.1. Eigenvalue counting function on R. ( [20, Section 4.1] ) In this subsection, we only consider the one-dimensional eigenvalue counting function. Let (E,domE) be defined as in (1.3) with Ω = (a, b) and let −∆µ be the associated Dirichlet Laplacian on L2((a, b), µ). Let P ={ai}n+1i=0 be a partition of [a, b] satisfying
a0 :=a < a1 <· · ·< an+1 =:b fori∈ {0, . . . , n+ 1}.
Define F := F(P) = {u ∈ domE : u(ai) = 0 for alli = 0. . . , n+ 1}. Then F is a closed subspace of domE. Define a relation ∼E on domE, induced by F, by u ∼E v if and only if u−v∈ F. Then∼E is an equivalence relation on domE. Define the quotient space
domE/F :={[u]E :u∈domE},
where [u]E is the equivalence class ofu. Define addition and scalar multiplication on domE/F as usual. For each i= 1, . . . , n, letfi be a function in domE that satisfies
fi(aj) =δij, i, j= 1, . . . ,2n,
whereδij is the Kronecker delta. Such anfi clearly exists. It is easy to prove that domE/F = span{[fi]E :i= 1, . . . , n} and dim(domE/F) =n.
Let −∆Fµ|
(a,b) be the Laplacian defined by the Dirichlet form (1.3) with domE =F, and let N(λ,−∆Fµ|
(a,b)) := #{n:λn(−∆Fµ|
(a,b))≤λ} be the associated eigenvalue counting function. If
F =N⊥, whereN is defined as in Section 1, then N(λ,−∆Fµ|
(a,b)) reduces toN(λ,−∆µ|
(a,b)).
It follows from the variational formula that N(λ,−∆Fµ|
(a,b))≤N(λ,−∆µ|
(a,b))≤N(λ,−∆Fµ|
(a,b)) + #P−2. (3.21)
If supp(µ) = [a, b], then N(λ,−∆Fµ|
(a,b)) = Pn
i=0N(λ,−∆µ|
(ai,ai+1)). Next, we state a similar formula. A proof can be found in [20, Proposition 4.1].
Proposition 3.1. Let µbe a continuous positive finite Borel measure on [a, b] withsupp(µ)⊆ [a, b]. Suppose there exists a nonempty subset Λ ⊆ {0,1, . . . , n} such that µ(ai, ai+1) > 0 for anyi∈Λ and µ(aj, aj+1) = 0 for anyj /∈Λ. Then
N(λ,−∆Fµ|
(a,b)) =X
i∈Λ
N(λ,−∆µ|
(ai,ai+1)).
3.2. Unitarily equivalent operators. In this subsection, we state a slightly modified version of [18, Propositions 2.2 and 2.3] below.
Proposition 3.2. ( [18, Proposition 2.2])LetS :R→Rbe a similitude, with Lipschitz constant r, such that S(a, b) = (c, d). Let µ be a continuous positive finite Borel measure on [a, b] with supp(µ)⊆[a, b]. Then
(a) −∆µ◦S−1|
(c,d) and r−1· −∆µ|(a,b)
are unitarily equivalent.
(b) If, in addition,µ|(c,d) =wµ◦S−1 on (c, d) for some constantw >0, then −∆ν|
(c,d) and (rw)−1· −∆µ|(a,b)
are unitarily equivalent.
Note that unitarily equivalent operators have the same set of eigenvalues.
Proposition 3.3. ( [18, Proposition 2.3])Let µ,ν be continuous positive finite Borel measures on [a, b] and assume that there exists some constant w >0 such that µ ≤wν on [a, b]. Then for anyn≥1, λn(−∆µ)≥w−1·λn(−∆ν).
The following result follows by combining Propositions 3.2 and 3.3. The proof can be found in [20].
Proposition 3.4. Let µ be a continuous positive finite Borel measure on R and assume that there exist a similitudeS with Lipschitz constant r, and a constantw >0 such thatS([a, b]) = [c, d]and µ≥wµ◦S−1 on[c, d]. Then N(wrλ,−∆µ|
(a,b))≤N(λ,−∆µ|
(c,d)).
4. Renewal equation and Proof of Theorem 1.1 and Theorem 1.3 4.1. Renewal equation. Let µ = Pq
i=1µi be a graph-directed self-similar measure defined by G = (V, E) on R. In the rest of this section, assume that µ satisfies (EFT) with {Ωi}qi=1 being an EFT-family, with (B,P) := ({B1,`},{Pk,`}k≥1)`∈Γ being a regular basic pair. The regularity of (B,P) implies that each cell B ∈ S
k≥1,`∈ΓPk,` is an interval. This allows us to apply Propositions 3.1–3.4. Let
Γi :={`∈Γ :B1,`⊆Ωi} and Bi :={B1,`:`∈Γi} for i∈V. (4.22)
Then Γ = Sq
i=1Γi and B = Sq
i=1Bi. Note that Γi and Bi maybe empty. The following Proposition has been modified from [20, Proposition 4.5] to suit our purpose. The proof is similar.
Proposition 4.1. Let µ=Pq
i=1µi be a graph-directed self-similar measure defined by a GIFS G = (V, E) on R. Assume that µ satisfies (EFT) with{Ωi}qi=1 being an EFT-family and with B := {B1,` :` ∈Γ} being a regular basic family of cells. Let Ω = Sq
i=1Ωi, and Γi, Bi defined as in (4.22). Then for i∈V and any `∈Γi, there exists some constantc` >0 such that
N(λ,−∆µi|B
1,`)≤N(λ,−∆µ|Ω)≤N c`λ,−∆µi|B
1,`
. (4.23)
Proposition 4.1 implies that the asymptotic behavior of N(λ,−∆µ) is controlled by that of N(λ,−∆µ
i|B
1,`) for i∈V and `∈Γi.
Step 1. Derivation of functional equations. For `∈Γi and k≥2, letP1k,` and P2k,` be defined as in (2.20) with respect to B, where i∈ V. Without loss of generality, we may assume that Γi can be partitioned into two (possibly empty) sub-collections, Γ0i and Γ∗i, defined as follows.
An index ` ∈ Γi belongs to Γ0i if there exists some integer k satisfying P2k,` = ∅. Let κ` ≥ 2 (depending on `) denote the smallest of suchk. Define Γ∗i := Γi\Γ0i and letκ` :=∞ for`∈Γ∗i. Let Γ0 =Sq
i=1Γ0i and Γ∗=Sq
i=1Γ∗i. Then Γ = Γ0∪Γ∗.
For i∈ V, fix any ` ∈Γi. The definition of (EFT) implies that for any 2≤ k ≤ κ`, there exist two finite disjoint G1k,`, G2k,`⊆Nsuch that
P1k,`=
k
[
m=2
Bm,`,p:p∈G1m,` and P2k,`=
Bk,`,p:p∈G2k,` .
Condition (1) of (EFT) implies that G12,` 6= ∅. If `∈ Γ∗, condition (3) of (EFT) implies that
k→∞lim P
p∈G2k,`
µ(Bk,`,p) = 0.
Proposition 4.2. Assume that µsatisfies (EFT). Let `∈Γi, and J`:=
j∈V :Se(Ωj)⊆B1,` for e∈Ei,j , (4.24) where i∈V. Let2≤k≤κ`. IfG1k,`6=∅, then for each p∈G1k,`, there exist someξ(k, `, p)>0 and c(k, `, p)∈Γj,j ∈J`, such that
N(λ,−∆µi|
Bk,`,p) =N(ξ(k, `, p)λ,−∆µj|B
1,c(k,`,p)). (4.25)
Proof. For any p ∈ G1k,`, by the definition of P1k,`, there exist some similitude Se(k,`,p) with Lipschitz constant re(k,`,p), as well as constants w(k, `, p) > 0 and c(k, `, p) ∈ Γj such that µi|Bk,`,p =w(k, `, p)µj|B
1,c(k,`,p)◦Se(k,`,p)−1 , wherej∈J`. Combining this with Proposition 3.2(b),
we obtain (4.25) with ξ(k, `, p) :=w(k, `, p)re(k,`,p).
For all i∈V, each `∈Γi, and 1≤n≤κ`, we define a partitionPn,` of B1,` as follows:
Pn,` :=
x:x is an end-point of some interval inPn,` ,
and let Fn,` := F(Pn,`). Note that for all i ∈ V, any ` ∈ Γi and 2 ≤ n ≤ κ`, we have
#Pn,`≤2#Pn,`. It follows from Proposition 3.1 that for i∈V,`∈Γi, and 2≤n≤κ`, N(λ,−∆Fµn,`
i|B
1,`
) =
n
X
k=2
X
p∈G1k,`
N(λ,−∆µi|
Bk,`,p) + X
p∈G2n,`
N(λ,−∆µi|
Bn,`,p).
Combining this with (3.21) and Proposition 4.2, for any`∈Γ0i, N(λ,−∆µ
i|B
1,`) =
κ`
X
k=2
X
p∈G1k,`
N(ξ(k, `, p)λ,−∆µ
j|B
1,c(k,`,p)) +(κ`, `), (4.26) where 0≤(κ`, `)≤2#Pκ`,`−2. Similarly, for `∈Γ∗i and n≥2, we have
N(λ,−∆µi|B
1,`) =
n
X
k=2
X
p∈G1k,`
N(ξ(k, `, p)λ,−∆µj|B
1,c(k,`,p))
+ X
p∈G2n,`
N(λ,−∆µi|
Bn,`,p) +(n, `),
(4.27)
where 0≤(n, `)≤2#Pn,`−2.
Step 2. Derivation of the vector-valued equation.
Case (1). G is strongly connected. For all i∈V, each `∈Γi, and α >0, define f`(t) =f`(α)(t) :=e−αtN(et,−∆µi|
B1,`), t∈R. (4.28)
Let λ=et. Then e−αtN(βλ,−∆µi|B
1,`) =βαf`(t+ lnβ) for anyβ >0. Now, multiply both sides of (4.26) and (4.27) bye−αt. Then for`∈Γ0i, we have
f`(t) =
κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)αfc(k,`,p) t+ lnξ(k, `, p)
+z(α)` (t), (4.29)
wherez`(α)(t) :=e−αt(κ`, `). For `∈Γ∗i andn≥2, we obtain f`(t) =
∞
X
k=2
X
p∈G1k,`
ξ(k, `, p)αfc(k,`,p) t+ lnξ(k, `, p)
+z`(α)(t)
−
∞
X
k=n+1
X
p∈G1k,`
ξ(k, `, p)αfc(k,`,p) t+ lnξ(k, `, p) ,
(4.30)
where
z`(α)(t) :=e−αt
X
p∈G2n,`
N(λ,−∆µi|
Bn,`,p) +(n, `)
. (4.31)
Since λ1(−∆µi|B
1,`) >0 for any i∈V and any `∈Γi, there exists t0 ∈R such that f`(t) = 0 for anyt < t0 and any `∈Γi. For each t ∈R, all i∈V and all `∈Γ∗i, let nt := nt(`) be the positive integer such that
t+ max
lnξ(k, `, p) :p∈G1k,` < t0 for allk > nt. (4.32)
Letn=nt in (4.30). Then f`(t) =
∞
X
k=2
X
p∈G1k,`
ξ(k, `, p)αfc(k,`,p) t+ lnξ(k, `, p)
+z(α)` (t) for`∈Γ∗i, (4.33)
wherez`(α)(t) is obtained from that in (4.31) by replacingnwith nt. Fori∈V,`∈Γi, letµ(α)`0`
be the discrete measure such that µ(α)`0` −lnξ(k, `, p)
:=ξ(k, `, p)α for 2≤k≤κ`, p∈G1k,`, `0=c(k, `, p)∈Γj and j∈J`. (4.34) Then (see (1.7))
µ(α)`0`(R) =
κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)α and
F`(α) =X
`0∈Γ κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)α.
Case (2). G is not strongly connected. If G = (V, E) is not strongly connected, then there exists somei, j ∈V satisfying Ei,j =∅. That is,S
`∈ΓiJ` =∅. Assume thatG hasγ strongly connected components. For m= 1, . . . , γ, letSCm be defined as in (1.11).
For m= 1, . . . , γ, each i∈SCm and each `∈Γi, define f`(t) =f`(αm)(t) :=e−αmtN(et,−∆µ
i|B
1,`), αm>0, t∈R. (4.35) Let λ= et. Then e−αmtN(βλ,−∆µi|B
1,`) = βαmf`(t+ lnβ) for any β > 0. Now, multiply both sides of (4.26) by e−αmt. Then for`∈Γ0i, we have
f`(t) =
κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)αmfc(k,`,p) t+ lnξ(k, `, p)
+z(α` m)(t), (4.36)
where z`(αm)(t) :=e−αmt(κ`, `). Similarly, for`∈Γ∗i and n≥2, we obtain f`(t) =
∞
X
k=2
X
p∈G1k,`
ξ(k, `, p)αmfc(k,`,p) t+ lnξ(k, `, p)
+z`(αm)(t)
−
∞
X
k=n+1
X
p∈G1k,`
ξ(k, `, p)αmfc(k,`,p) t+ lnξ(k, `, p) ,
(4.37)
where
z`(αm)(t) :=e−αmt X
p∈G2n,`
N(λ,−∆µi|
Bn,`,p) +(n, `)
. (4.38)
Since λ1(−∆µi|B
1,`) > 0 for all i∈ V and all ` ∈ Γi, there exists t0 ∈ R such that f`(t) = 0 for any t < t0 and any `∈Γi. For each t∈ R, all i∈V and all `∈Γ∗i, let nt := nt(`) be the positive integer such that
t+ max
lnξ(k, `, p) :p∈G1k,` < t0 for allk > nt. (4.39)
Letn=nt in (4.37). Then for m= 1, . . . , γ, eachi∈SCm and `∈Γ∗i, we have
f`(t) =
∞
X
k=2
X
p∈G1k,`
ξ(k, `, p)αmfc(k,`,p) t+ lnξ(k, `, p)
+z`(αm)(t), (4.40)
where z(α` m)(t) is obtained from that in (4.38) by replacing n with nt. For m = 1, . . . , γ, i∈SCm,`∈Γi, let µ(α`0`m) be the discrete measure such that
µ(α`0`m) −lnξ(k, `, p)
:=ξ(k, `, p)αm for 2≤k≤κ`, p∈G1k,`, `0 =c(k, `, p)∈Γj and j∈J`. (4.41) Then (see (1.7))
µ(α`0`m)(R) =
κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)αm
and
F`(αm) = X
`0∈Γ κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)αm.
We summarize the above derivations in the following theorem.
Theorem 4.3. Let µ= Pq
i=1µi be a graph-directed self-similar measure on R. Assume that µ satisfies (EFT) with{Ωi}qi=1 being an EFT-family and there exists a regular basic pair. Let Ω =Sq
i=1Ωi and ∆µ is defined on Ω. Let f,Mα andz be defined as in (1.5). Thenf satisfies the vector-valued renewal equation f =f∗Mα+z.
4.2. Proof of Theorem 1.1 and Theorem 1.3.
Proof of Theorem 1.1. (a) Since eachF`(α) is a strictly decreasing positive continuous function ofα, lim
α→∞F`(α) = 0, and lim
α→αe+`
F`(α)>1, there exists a uniqueαsuch that the spectral radius ofMα(∞) is 1.
(b) Let α be the unique number in part (a). Letm:= [m(α)`0`] = [R∞
0 x dµ(α)`0`] be the moment matrix. Following the proof of [18, Theorem 1.1(b)], we need to show that some moment condition holds, and it suffices to show that 0 < P
`0∈Γ
m(α)`0` < ∞. It is easy to check that for
`∈Γ, P
`0∈Γ
m(α)`0` takes the following values:
X
`0∈Γ κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)α
ln(ξ(k, `, p)) .
It follows from lim
α→αe+`
F`(α)>1 that there exists >0 such that 0< F`(α−)<∞. Thus
0<X
`0∈Γ κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)α
ln(ξ(k, `, p))
=X
`0∈Γ κ`
X
k=2
X
p∈G1k,`
ξ(k, `, p)α−ξ(k, `, p)
ln(ξ(k, `, p))
<∞.
The last inequality follows from the fact lim
t→0+tlnt= 0. By (4.34), we have P
`0∈Γ
µ(α)`0`(0) = 0<
P
`0∈Γ
µ(α)`0`(∞), i.e., each column ofMα is nondegenerate at 0. From Theorem 4.3,f =f∗Mα+z, where, by assumption,z is directly Riemann integrable onR.
We first consider the case Mα(∞) is irreducible. It follows from the above observations and [18, Theorem 4.1] that there exist positive constants C1 and C2 such that 0 < C1 ≤
lim
t→∞
f`(t) ≤ lim
t→∞f`(t) ≤ C2 < ∞ for all ` ∈ Γ. The definition in (4.28) implies that C1 ≤ λ−αN(λ,−∆µi|
B1,`) ≤ C2, which, together with (4.23), yields C1λα ≤ N(λ,−∆µ|
Ω) ≤ C2λα. Combining this, part (a), and the definition of ds(µ), we get ds(µ) = 2α.
It remains to consider the case Mα(∞) is reducible. As in the proof of [18, Theorem 1.1(b), Case 2], we have
t→∞lim f`(β)(t) = 0 for all `∈Γi, i∈ {1, . . . , q} and allβ < α.
Moreover, there exists some `0 ∈Γi such that lim
t→∞
f`(α)
0 (t) >0. The definition of f`(t) implies that N(λ,−∆µ
i|B
1,`) = o(λβ) for ` ∈ Γi and β < α and lim
λ→∞
λ−αN(λ,−∆µ
i|B
1,`0
) > 0. Thus N(λ,−∆µ) = o(λβ) for any β < α and lim
λ→∞
λ−αN(λ,−∆µ) > 0. Hence ds(−∆µ) ≤ 2α and ds(−∆µ)≥2α, which completes the proof.
Proof of Theorem 1.3. (a) We observe that each F`(αm) is a strictly decreasing positive con- tinuous function ofαm, lim
αm→∞F`(αm) = 0, and lim
αm→αe+`
F`(αm)>1. Thus there exists a unique α= max{αm:m= 1, . . . , γ}such that the spectral radius of Mαm(∞) is 1.
(b) The proof is similar to that of Theorem 1.1(b).
5. Strongly connected GIFSs on R
In this section, we compute the spectral dimension of some graph-directed self-similar mea- sures defined by strongly connected GIFSsG= (V, E), which have overlaps.