Professional Documents
Culture Documents
HIGHER DIMENSIONS
Contents
1. Introduction and statement of main results 2
1.1. Introduction and background 2
1.2. Preliminaries 3
1.3. Main results 6
1.4. Applications 10
2. Form approach for Kreı̆n–Feller operators 12
2.1. Sobolev spaces and embeddings 12
2.2. Stein extension 13
FB 3 – Mathematik und Informatik, University of Bremen, Bibliothekstr. 1, 28359 Bremen,
Germany
E-mail addresses: mhk@uni-bremen.de, niemann1@uni-bremen.de.
2000 Mathematics Subject Classification. 35P20; 35J05; 28A80; 42B35; 45D05.
Key words and phrases. Kreı̆n–Feller operator; Laplace operator; spectral asymptotics; Lq -spectrum,
spectral partition function, Dirichlet forms, Minkowski dimension, coarse multifractal formalism,
Sobolev spaces, adaptive partition algorithm.
This research was supported by the DFG grant Ke 1440/3-1.
1
2 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
subtle connection between the multifractal concept of the Lq -spectrum and analytic
properties of the associated ‘fractal’ operators.
In this paper we extend ideas for the one-dimensional case developed in [KN22;
KN21b] to higher dimensions d ≥ 2 and in this way follow the line of investigation
outlined in [BS67; NS95; NS01; HLN06; NX21]. We will introduce the new
notion of partition functions, which is needed for higher dimensions and naturally
generalises Lq -spectra (Section 3). This new construction is closely related to certain
optimal embedding constants for Sobolev spaces (see Section 6.3) as elaborated by
Maz’ya and Preobrazenskii [Maz85; MP84] for d = 2 and Adams [Maz11, Section
1.4.1] for d > 2.
The spectral dimension of Kreı̆n–Feller operator for higher dimensions has
been first computed by Triebel [Tri97, Theorem 30.2] for Ahlfors–David regular
measures, by Naimark and Solomyak [Sol94; NS95] in the setting of self-similar
measures under the open set condition (OSC), and recently by Ngai and Xie [NX21]
for a class of graph-directed self-similar measures satisfying the graph open set
condition. As an application of our general results we extend these achievements to
self-conformal measures without any restriction on the separation conditions. In fact,
we prove that under the assumption (♠) the spectral dimension for self-conformal
measures can be identified as the unique intersection of the Lq -spectrum with the
line of slope 2 − d through the origin (Theorem 1.16).
1.2. Preliminaries. We now proceed to outline the theoretical preliminaries nec-
essary to determine the spectral properties of the Kreı̆n–Feller operator ∆νD/N for
a given finite non-zero Borel measure ν on the fixed d-dimensional unit cube
Q B di=1 Ii , d ≥ 2 with I j unit intervals, for j = 1, . . . , d, which are either half-
Q
open, open, or closed. For a bounded open set Ω ⊂ Rd with Lipschitz
boundary, let
us define the Sobolev spaces H (Ω) as the completion of Cb Ω with respect to the
1 ∞
kuk2L2 ≤ K kuk2H 1
ν
and in this case we can extend the operator to H 1 . Our first observation (Propo-
sition 2.14) is that ι is not continuous if the lower ∞-dimension of ν lies under a
certain threshold, namely,
maxQ∈DnN log ν (Q)
dim∞ (ν) B lim inf < d − 2,
n→∞ −n log 2
where DnN denotes a partition of Q by cubes of the form Q B di=1 Ii with (half-open,
Q
open, or closed) intervals Ii with endpoints in the dyadic grid, i. e. (k − 1) 2−n , k2−n
for some k ∈ Z, such that Dn+1 N is a refinement of DN for each n ∈ N, this means
n
that each element of Dn can be decomposed into 2d disjoint elements of Dn+1
N N .
is fulfilled, then a result of Maz’ya [Maz85] adapted to the dyadic grid (see
Lemma 6.7) ensures, that the embedding ι is compact and ∆νDN admits a count-
able 2
set ofeigenfunctions spanning Lν with a non-negative sequence of eigenvalues
λn ED/N tending to infinity. As mentioned above, the Hu–Lau–Ngai condition
n∈N
already appeared implicitly in [Tri97, Theorem 30.2 (Isotropic fractal drum)] in the
context of Ahlfors–David regular measures, for which we provide more details in
1.4 below). We note that for Q ∈ D with ν (Q) > 0 we have dim∞ (ν) ≤ dim∞ ν|Q
and hence the condition (♠) carries over to the restricted measure ν|Q . We define
the upper and lower exponent of divergence of the eigenvalue counting function
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 5
n o
N D/N (x) B sup n ∈ N : λn ∆νD/N ≤ x by
log N D/N (x) log N D/N (x)
sD/N B lim inf and sD/N B lim sup ,
x→∞ log(x) x→∞ log(x)
and refer to these numbers as the upper, resp. lower, spectral dimension of ED/N
(or of ∆D/N or just of ν). If the two values coincide we denote the common value by
sD/N , and call it the Dirichlet (respect. Neumann) spectral dimension.
Next, let us turn to the concept of partition functions, which in a certain extent
is borrowed from the thermodynamic formalism. For an arbitrary monotone set
function J : D → R≥0 we define the associated partition function, for q ∈ R≥0 ,
1 X
τJD/N (q) B lim sup τJ,n
D/N
(q) with τJ,n D/N
(q) B n
log J (Q)q (1.1)
n→∞ log 2 D/N
Q∈Dn
n o
with DnD B Q ∈ DnN : ∂Q ∩ Q = ∅ . The reason why the definition of DnD is ap-
propriate becomes apparent in constructing certain functions with compact support
contained in Q̊ for the proof of the lower bounds in the Dirichlet case (see proof
of Lemma 7.1). Note that we use the convention 00 = 0, that is for q = 0 we
neglect the summands with J (Q) = 0 in the definition of τJD/N . An important
quantity in the one-dimensional case and also in certain higher-dimensional cases
is the Lq -spectrum of ν, given by βνD/N B τνD/N . In this paper, we are particularly
interested in the set function
b
sup
Q∈D(Q) Q
ν e log Λ Qe , a = 0,
Jν,a,b (Q) B
e
eb ea
Q∈D(Q) ν Q Λ Q ,
sup e
a , 0,
with b ≥ 0 and a ∈ R. We write Jν,t (Q) B Jν,2/d−1,2/t (Q) and Jν (Q) B
Jν,2/d−1,1 (Q). We note that the general parameter a, b will also prove useful when
considering polyharmonic operators in higher dimensions (see e. g. [KN21a]).
Our most powerful auxiliary object is the (Dirichlet/Neumann) spectral partition
function with respect to ν given by the special choice J = Jν,a,b . As a consequence
of Lemma 3.7 we know that the spectral partition function does not depend on the
specific choice of the collection of dyadic cubes DnD/N . We consider the critical
exponent
X
κJ B inf q
< .
q ≥ 0 : J(Q) ∞
Q∈D
First, to obtain upper estimates of the spectral dimension, we construct optimal
partitions using an adaptive partition algorithm as presented in Section 4. As a
byproduct of Section 4, we are able to improve a result of [Bor71, Theorem 1]
(see Section 4.2). Let us define the set of J-partitions ΠJ to be the set of finite
o exists a partition P of Q by dyadic
collections P of dyadic ncubes such that there e
cubes from D with P = Q ∈ P : J(Q) > 0 . We define
e
and
log MJ (x) log MJ (x)
hJ B lim sup , hJ B lim inf
x→∞ log x x→∞ log x
will be called the upper, resp. lower, J-partition entropy. We will see in Section 4
and Section 5 that under some mild additional assumptions on J, hJ and κJ coincide.
1.3. Main results. The following abstract theorem provides an upper bound on the
upper spectral dimension in terms of the upper J-partition entropy.
Theorem 1.1. Suppose there exists a non-negative, monotone set function J on D,
which is uniformly vanishing, i. e. limn→∞ supQ∈Sk≥n DN J(Q) = 0, such that for all
k
b Q with Q u dΛ = 0, we have
R
Q ∈ D and all u ∈ C∞
Then s ≤ s ≤ hJ .
D N
For lower estimates of the spectral dimension we use certain disjoint families of
dyadic cubes and borrow ideas from the coarse multifractal analysis (see [Fal14;
Rie95]) which will be the topic of Section 5. In there we will also see how the
dyadic partition approach and the optimal partition approach are related by ideas
from large deviation theory. For all n ∈ N and α > 0, we define
D/N D/N D/N
n o
Nα,J (n) B card Mα,J (n) , Mα,J (n) B Q ∈ DnD/N : J (Q) ≥ 2−αn ,
and set
log+ Nα,J log+ Nα,J
D/N D/N
D/N (n) (n)
FJ (α) B lim sup , and F JD/N (α) B lim inf .
n log 2n n log 2n
We refer to the quantities
D/N
D/N FJ (α) F JD/N (α)
FJ B sup and F JD/N B sup
α>0 α α>0 α
as the upper, resp. lower, optimised (Dirichlet/Neumann) coarse multifractal dimen-
sion with respect to J. The lower estimate of the spectral dimension is based on
the following abstract observation which connects the optimised coarse multifractal
dimension and the spectral dimension.
Theorem 1.2. Assume there exists a non-negative monotone set function J on D
with dim∞ (J) > 0 (see Section 3) such that for every Q ∈ D with J (Q) > 0 there
exists a non-negative and non-zero function ψQ ∈ C∞ c with support contained in
˚ ˚
hQi3 (see Section 2.5 for the definition of hQi3 ) such that
2
2
ψQ
L2 ≥ J(Q)
∇ψQ
L2 Rd .
ν Λ( )
Then
N D
F JN ≤ sN and F J ≤ sN , F JD ≤ sD ≤ sN and F J ≤ sD ≤ sN .
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 7
then qJD/N is the unique zero of τJD/N and qJN = κJ . In general, κJ ≤ qJN .
Under the condition (♠) and for any t ∈ (2, 2 dim∞ (ν)/(d − 2)) the set function
Jν,t is uniformly vanishing and using [Maz85, Corollary, p. 54, 381–382], Theorem
1.1 is applicable for Jν,t (see Lemma 6.9 and Proposition 6.6). For the critical case
dim∞ (ν) = d − 2 there is the possibility of no continuous embedding, a continuous
but non-compact or a compact embedding. In Section 9.1 we give examples (for
d = 3) of absolutely continuous measures with dim∞ (ν) = d − 2 such that each
possibility is realised. For the case of compact embedding, Theorem 1.1 can be
employed to show that in our example sN = 3/2 (Example 9.2).
We will see that Theorem 1.2 is applicable for J = ν in the case d = 2 and J = Jν
for d > 2. The following list of results give the main achievements of this paper.
The proofs are postponed to Section 8. As an auxiliary quantity we need
!
N\D
dim∞ (ν) B lim inf − log max ν (Q) / log 2n
n→∞ Q∈DnN \DnD
D/N D/N
and we introduce the shorthand notation qD/N B qJD/N
ν
,F B F Jν , F D/N B
F JD/N
ν
, τD/N B τJD/N
ν
and h B hJν , h B hJν . In the following we write dim M (A) for
the upper Minkowski dimension of the bounded set A ⊂ Rd and—slightly abusing
notation—we also write dim M (ν) B dim M supp (ν) for the compactly support
Borel measure ν.
Theorem 1.3. Let ν be a Borel probability measure on Q such that dim∞ (ν) > d − 2.
(1) Under Neumann boundary conditions we have
N
F N ≤ sN ≤ h ≤ h = sN = qN = F , (1.2)
(2) Under Dirichlet boundary conditions and ν(Q̊) > 0 we have
D
F D ≤ sD , F = qD ≤ sD ≤ qN .
(3) In particular, if d = 2, then sN = 1, and under the assumption ν(Q̊) > 0, we
s = 1.
D
also have
N D
(4) If τ q = 0, or equivalently F = F , then the upper Dirichlet and
N D
or
• τD/N qD/N = lim inf n τJD/N ν ,n
qD/N and τD/N is differentiable in qD/N .
Remark 1.6. The above theorem and the notion of regularity give rise to the follow-
ing list of observations:
(1) If the Neumann spectral dimension exists, then it is given by purely measure-
geometric data encoded in the ν-partition entropy, namely we have h = h
and this value coincides with the spectral dimension.
(2) N-MF-regularity implies equality everywhere in the chain of inequalities
(1.2) and in particular the Neumann spectral dimension exists. If ν is D-
MF-regular, then we have equality everywhere in all chains of inequalities
above and in particular both Neumann and Dirichlet spectral dimensions
exist.
(3) To the best of our knowledge, all measures examined in the literature, for
which the spectral dimension is known, are PF-regular.
The following theorem shows that the spectral partition function is a valuable
auxiliary concept to determine the spectral behaviour for a given measure ν.
Theorem 1.7. Under the assumption dim∞ (ν) > d − 2 we have the following
regularity result:
(1) If ν is N-PF-regular, then it is N-MF-regular and the Neumann spectral
dimension sN exists.
(2) If ν is D-PF-regular and τN qD = 0, then both the Dirichlet und Neumann
spectral dimension exist and coincide, i. e. sD = sN .
This result is optimal in the sense that there is an example (derived from an
similar example for d = 1 in [KN21b]) of a measure ν which is not τ-regular and
for which sN > sN . It should be noted that PF-regularity is easy accessible if the
spectral partition function is essentially given by the Lq -spectrum.
In the following proposition we present lower bounds of the lower spectral
dimension in terms of the subdifferential of τD/N in q, defined as
n o
∂τD/N (q) B a ∈ R : ∀t ∈ R τD/N (t) ≥ a (t − q) + τD/N (q) .
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 9
Proposition 1.8. If for some 0 ≤ q ≤ qD/N we have τD/N (q) = limn τJD/N ν ,n
(q) and
−∂τ (q) = [a, b], then
D/N
aq + τD/N (q)
≤ sD/N .
b
Remark 1.9. In the case that τN (qN ) = lim τJNν ,n qN and τN is differentiable in qN ,
we infer qN ≤ sN and hence obtain a direct proof of the regularity statement, namely,
qN = sN = sN .
Also, if τD/N (1) = limn τJD/N
ν ,n
(1) = d − 2, we have the lower bound
+
τD/N (1) − d + 2
− ≤ sD/N ,
τD/N (1)
where f ± (x) denotes the left-sided, respectively right-sided, derivative of f : R≥0 →
R in x > 0.
Corollary 1.10. Let d = 2. Assume βνD/N is differentiable in 1. Then sD/N = 1.
1.3.2. General bounds in terms of fractal dimensions. We obtain general bounds for
sN in terms of the upper Minkowski dimension dim M (ν) and the possibly smaller
lower ∞-dimension dim∞ (ν) of ν (see also Figure 1.1 on page 10).
Corollary 1.11. Assume dim∞ (ν) > d − 2. Then for the Neumann upper spectral
dimension we have
Remark 1.12. Note that by choosing measures with dim M (ν) close to d − 2 we can
easily find examples where sN becomes arbitrarily large.
It is also worth mentioning that the analogous situation in the dimension d = 1 is
quite different (cf. [KN21b]), namely the lower bound becomes an upper bound,
dim M (ν) 1
sN/D ≤ ≤ .
dim M (ν) + 1 2
This inequality naturally links to the famous question by M. Kac [Kac66], ‘Can one
hear the shape of a drum?’ This question has been modified by various authors e. g.
in [Ber79; Ber80; BC86; Lap91], and closer to our context by Triebel in [Tri97].
In the plane, the spectral dimension does not encode any information about the
fractal-geometric nature of the underlying measure as we always have have sD/N = 1
for any bounded Borel measure with ν(Q̊) > 0. This has been observed in [Tri97]
for the special case of α-Ahlfors–David regular measures. For all other dimensions,
our results show that the upper spectral dimension sN is uniquely determined by
the spectral partition function τN , which in turn reflects many important fractal-
geometric properties of ν. For the case d > 2, this common ground provides
10 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
τN (q)
2 dim M (ν)
q
1 2 qN 3
interesting bounds on the upper Minkowski dimension of the support of ν and the
lower ∞-dimension of ν in terms of the upper spectral dimension as follows:
sN (d − 2)
dim M (ν) ≥ ≥ dim∞ (ν) .
sN − 1
So the answer to Kac’s question is ‘partially yes’. If additionally the Lq -spectrum
βνN is an affine function, we obtain βνN (q) = dim M (ν) + dim M (ν) (1 − q) and with
Corollary 1.11
sN (d − 2)
dim∞ (ν) = dim M (ν) = .
sN − 1
In this case, Kac’s question regarding dimensional quantities must be answered in
the affirmative.
1.4. Applications. Finally, we give some leading examples where the spectral
partition function is essentially given by the Lq -spectrum of ν (see Section 3.4.3
and Section 3.4.1) and in this case we are able to provide the following complete
picture.
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 11
vector (pi )i=1,...,` , ` ≥ 2 (see Section 3.4.3 for precise definitions). The spectral
partition function exists as a limit and is given by
τD/N (q) = βνN (q) + (d − 2) q,
where βνN denotes the Neumann-Lq -spectrum of ν (see Section 3.2 for definition).
Assuming dim∞ (ν) > d−2, then ν is D/N-PF-regular and the Dirichlet and Neumann
spectral dimension exist and equal sD = sN = qN . In particular, in the case d = 2,
we always have sD = sN = 1.
Remark 1.17. We remark that in the situation of Theorem 1.16 under OSC the
spectral dimension can be expressed in terms of an associated pressure function, i. e.
qN is the unique solution of P (q(ψ + (2 − d)ϕ)) = 0, where
. . . , `} → R :
ψ : {1, N
Remark 1.18. In general, it can be difficult to verify the condition dim∞ (ν) > d − 2,
but in the case d = 2 a sufficient condition is that the measure ν is invariant with
respect to an IFS given by a system of bi-Lipschitz contractions such that the
attractor is not a singleton ([HLN06, Lemma 5.1]). This carries over to self-similar
measures provided that the contractive similitudes do not share the same fixed point,
so that dim∞ (ν) > 0 and the spectral dimension is given by sD = sN = 1.
inequality (see [Rui12, Lemma 3, p. 500] and Lemma 2.1), reads, for some constant
c > 0, as follows
kukL2 ≤ c k∇ukL2 . (PWI)
Λ Λ
Since Q is bounded, the norm induced by the form h f, giH 1 is therefore equivalent
0
to the norm induced by h f, giH 1 on H01 . We will consider only those finite Borel
measures ν on the closure of some Lipschitz domain Ω ⊂ Rd for which the following
Poincaré inequality holds with c1 > 0.
kukL2 Ω ≤ c1 kukH 1 (Ω) for all u ∈ Cb∞ Ω . (PI)
ν
This then guarantees an embedding of the Sobolev spaces H 1 (Ω) and H01 (Ω) into
Lν2 . In fact, since Cb∞ Ω lies dense in H 1 (Ω) (this follows e. g. from the extension
1
domains), for every u ∈ H (Ω) there exists a sequence (un ) of
property of Lipschitz
elements of Cb Ω such that un → u with respect to the norm in H 1 (Ω). Now, (PI)
∞
implies that (un ) is also a Cauchy sequence in Lν2 , hence there exists u ∈ Lν2 Ω such
that un → u in Lν2 Ω . It is easy to see that this limit is independent of the particular
choice of (un ) and we therefore obtain in this way a bounded linear operator
ι B ιν B ιΩ,ν : H 1 (Ω) → Lν2 Ω , f 7→ f ,
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 13
with ι(u) = u for all u ∈ Cb∞ Ω . If ι is also injective, then we may regard H 1 (Ω)
as a subspace of Lν2 Ω . In case the map is not injective we consider the following
closed subspace of H 1 (Ω)
Nν B ker (ι) = f ∈ H 1 (Ω) : kι( f )kL2 Ω = 0
ν
n o
and have the natural embedding Nν⊥ B f ∈ H 1 (Ω) : ∀g ∈ Nν : h f, giH 1 (Ω) = 0 ,→
Lν2 Ω which is again given by ι. In particular, there exists a sequence (un ) ∈ Cb∞ Ω
such that un → u in H 1 (Ω) and ι (un ) = un → ι (u) in Lν2 Ω . The extended version
of the Poincaré inequality (PI) reads as
kι(u)kLν2 (Ω) ≤ c1 kukH 1 (Ω) for all u ∈ H 1 (Ω) .
This embedding carries over to
N̊ν B f ∈ H0 (Ω) : ∀g ∈ Nν ∩ H0 : h f, giH 1 (Ω) = 0
⊥ 1 1
(2.1)
0
From [NS01, Lemma 3, p. 500] we obtain that there exists CQ > 0 such that for all
u ∈ H 1 (Q)
Z ! Z 2
u dΛ ≤ k∇ukL2 (Q) + u dΛ .
2 2
CQ
Q Λ Q
Now, for u ∈ H 1 (Q) and T : Rd → Rd with T (x) = x0 + hx for some h ∈ (0, 1) and
x0 ∈ Q, such that T (Q) = Q , we have u ◦ T ∈ H 1 (Q) and
Z Z
CQ
u dΛ = CQ
2
u2 ◦ T dΛ
hd Q Q
Z 2
+ u ◦ T dΛ
T )k2L2 (Q)
≤ k∇ (u ◦
Λ Q
Z 2
= h2−d k∇uk2L2 (Q) + h−2d u dΛ .
Λ Q
Z Z 2 Z 2
1
2 2
k∇uk2L2 (Q) +h −d
u dΛ ≤ k∇ukL2 (Q) +
2
u dΛ .
CQ u dΛ ≤ h
Q Λ Q Λ Λ(Q) Q
This gives
Z 2
min {1, CQ } 1
kukH 1 (Q) ≤ k∇ukL2 (Q) +
2 2
u dΛ .
2 Λ Λ(Q) Q
Lemma 2.2. Let T : Rd → Rd , x 7→ x0 + hx, with h ∈ (0, 1), x0 ∈ Q, such that the
cube Q B T (Q) belongs to D. Then we have:
(1) EQ : H 1 (Q) → H 1 Rd , u 7→ EQ (u ◦ T ) ◦ T −1 defines a Stein extension
with
EQ
≤ kEQ k /h2 ,
n o
(2)
EQ |NΛ (Q)
≤ kEQ k /DQ with NΛ (Q) B u ∈ H 1 (Q) : Q u dΛ = 0 .
R
EQ (u)
H 1 (Rd )
EQ (u ◦ T ) ◦ T −1
H 1 (Rd )
EQ |NΛ (Q)
= sup = sup
u∈NΛ (Q)? kukH 1 (Q) u∈NΛ (Q)? kukH 1 (Q)
R 2 R 2 1/2
∇ EQ (u ◦ T ) ◦ T −1 dΛ + EQ (u ◦ T ) ◦ T −1 dΛ
≤ sup 2 1/2
u∈NΛ (Q)?
R R
−1 2 dΛ + 1
Q
∇ (u ◦ T ) ◦ T Λ(Q) Q u dΛ
1/2
hd−2 (∇ (EQ (u ◦ T )))2 dΛ + hd E(u ◦ T )2 dΛ
R R
= sup 1/2
u∈NΛ (Q)?
R
hd−2 Q (∇ (u ◦ T ))2 dΛ
R 1/2
hd−2 (∇ (EQ (u ◦ T )))2 dΛ + E(u ◦ T )2 dΛ
R
≤ sup 1/2
u∈NΛ (Q)?
R
hd−2 Q (∇ (u ◦ T ))2 dΛ
kEQ (u ◦ T )kH 1 (Rd ) kEQ k
≤ sup ≤ ,
u∈NΛ (Q)? DQ ku ◦ T kH 1 (Q) DQ
Proof. Here we follow the arguments of [HLN06]. Let Cc (Q̊) denote the space of
continuous
functions from Q to R with compact support in Q̊. It is well known
that
Cc∞ Q̊ lies dense in Cc Q̊ with respect to the uniform norm, hence Cc∞ Q̊ lies
dense in Lν2 (Q̊). This carries over to the orthogonal projection onto ι N̊ν⊥ since
ι N̊ν is the zero space in Lν2 (Q̊). Similarly, it for the Neumann note that Cb (Q) lies
dense in Lν2 and by the Stone-Weierstrass theorem Cb∞ (Q) lies dense in Cb (Q) with
respect to the uniform norm. Hence, the claim follows.
Proposition 2.5. Assuming (PI), we have that dom ED/N equipped with the inner
product h f, giν + ED/N ( f, g) defines a Hilbert spaces, i. e. ED/N is a closed form
with respect to Lν2 .
Proof. Since both cases can be treated completely analogously, we only consider
the first case: We first observe that Nν⊥ is a closed linear subspace with respect
to h·, ·iH 1 , which by (PI) induces a norm that is equivalent to the norm induced
by h·, ·iH 1 + hι·, ι·iν . Therefore, Nν⊥ , h·, ·iH 1 + hι·, ι·iν is a Hilbert space space and
since EN + h·, ·iν is the push-forward of h·, ·iH 1 + hι·, ι·iν , the claim follows.
2.4. Kreı̆n–Feller operator. From
Proposition
2.5,Proposition 2.4 and [Kig01,
Theorem B.1.6.], we deduce that E , dom ED/N is a densely defined closed
D/N
D/N
form on Lν2 . Thus, for ED/N there exists non-negative
D/N 1/2
self-adjoint operator ∆ν
on Lν2 such that f ∈ dom ∆ν ⊂ dom ∆νD/N = dom(ED/N ), if and only
if f ∈ dom ED/N and there exists u ∈ Lν2 such that
ED ( f, g) = hu, giL2 (Q̊) , g ∈ dom ED
ν
and
EN ( f, g) = hu, giLν2 (Q) , g ∈ dom EN .
D/N D/N
In this we have u = ∆ν f . We call ∆ν Kreı̆n–Feller operator and f ∈
case,
dom ∆νD an (Dirichlet) eigenfunction with eigenvalue λ ∈ R, if
Z
E ( f, g) = λ
D
f g dν
Q̊
for all g ∈ dom ED . And we call f ∈ dom ∆νN an (Neumann) eigenfunction with
eigenvalue λ ∈ R, if Z
EN ( f, g) = λ f g dν
Q
for all g ∈ dom EN . In order to prove that the embedding dom ED/N , ED/N ,→
Lν2 is compact under the assumption dim∞ (ν) > d − 2, we need the following result
due to Maz’ya n [Maz85, Theorem 3, p. 387, o 2nd edition p. 579]: For d > 2 and
t > 2, the set u ∈ Cc∞ (Rd ) : kukH 1 (Rd ) ≤ 1 is precompact in Lνt , if and only if
lim sup %(1−d/2) ν (B(x, %))1/t = 0,
r↓0 x∈Rd ,%∈(0,r)
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 17
n o
and for d = 2, u ∈ Cc∞ (Rd ) : kukH 1 (Rd ) ≤ 1 is precompact in Lνt , if and only if
1/2
lim sup log(%) ν (B(x, %))1/t = 0.
r↓0 x∈Rd ,%∈(0,r)
Proposition
assumption dim∞ (ν) > d − 2 implies (PI) and the embedding
2.6. The
dom E D/N ,E D/N ,→ Lν2 is compact.
Proof. For t ∈ (2, 2 dim∞ (ν)/(d − 2)) and d > 2, the assumption (PI) implies
lim sup %(1−d/2) ν (B(x, %))1/t = 0,
r↓0 x∈Rd ,%∈(0,r)
and for d = 2,
1/2
lim sup log(%) ν (B(x, %))1/t = 0,
r↓0 x∈Rd ,%∈(0,r)
(see n[HLN06]).
Hence,
by [Maz85,
o Theorem 3 and Theorem
4, p. 583] we know
that u ∈ Cc R : kukH 1 (Rd ) ≤ 1 is precompact in Lν R . Thus, there exists c > 0
∞ d t d
such that for all u ∈ Cc∞ Rd , we have
kukLν2 (Rd ) ≤ kukLνt (Rd ) ≤ c · kukH 1 (Rd ) ,
which implies that the identity Cc∞ Rd → Lν2 Rd map permits a unique continuous
continuation ιRd : H 1 Rd → Lν2 Rd . To see that ιRd is compact, fix a bounded
sequence (un ) in H 1 Rd . We find another sequence (vn ) in Cc∞ Rd such that
kvn − un kH 1 (Rd ) → 0. The aforementioned precompactness leads to a subsequence
(vnk ) converging in Lν2 (Rd ) to an element v. Then ιRd (unk ) also converges to v in
Lν2 (Rd ), since
v − ιRd (unk )
L2 (Rd ) =
v − vnk + vnk − ιRd (unk )
L2 (Rd )
ν ν
≤
v − vnk
L2 (Rd ) +
ιRd (vnk − unk )
L2 (Rd )
ν
ν
≤
v − v
nk L2 (Rd ) + c ·
v − u
nk nk H 1 (Rd ) → 0.
ν
n o
This shows that ιRd u ∈ H 1 (Rd ) : kukH 1 (Rd ) ≤ k is precompact for every k > 0.
Since for the Stein extension EQ we have for all u ∈ H 1
kukLν2 ≤ c2
ι−1 u
H 1 .
The Dirichlet case follows by almost the same means without the use of the extension
operator (see also [HLN06]).
Corollary 2.7. Assume dim∞ (ν) > d − 2. Then the operator n D/N o ∆
D/N has compact
Remark 2.8. Let E be a closed form with domain dom (E) densely defined on Lν2 , in
particular dom (E) defines a Hilbert space with respect to ( f, g)E B h f, gi ν + E( f, g),
and assume that the inclusion from dom (E) , h·, ·iE into Lν2 is compact. Then the
Poincaré–Courant–Fischer–Weyl min-max principle is applicable, that is for the
i-th eigenvalue λi (E) of E, i ∈ N, we have (see also [Kig01, Theorem B.I.14] or
[Dav95; KL93])
λi (E) = inf sup R (ψ) : ψ ∈ G? : G <i dom (E) , h·, ·iE ,
n n o o
where we write G <i (H, h·, ·i), if G is a linear subspace of the Hilbert space H with
inner product h·, ·i and the vector space dimension of G is equal to i ∈ N; for the
Rayleigh–Ritz quotient is given by R (ψ) B E(ψ, ψ)/hψ, ψiν .
The following proposition will be crucial for the proof of the upper bound of the
spectral dimension as stated in Corollary 6.10.
Proposition 2.9. We have for all i ∈ N,
λi ED = inf sup RH 1 (ψ) : ψ ∈ G? : G <i
N̊ν⊥ , h·, ·iH 1
0 0
= inf sup RH 1 (ψ) : ψ ∈ G? : G <i H01 , h·, ·iH 1 ,
0 0
where the relevant Rayleigh–Ritz quotient is given by RH 1 (ψ) B hψ, ψiH 1 /hιψ, ιψiν .
0 0
The same result holds true for EN with N̊ν⊥ replaced by Nν⊥ and H01 by H 1 .
Proof. The first equality follows by the min-max principle and the fact that dom ED '
N̊ν⊥ . The part ‘≥’ for the second equality follows from the inclusion N̊ν⊥ ⊂ H01 . For
the reverse inequality we consider an i-dimensional subspace G = span( f1 , . . . , fi ) ⊂
H01 . There exists a unique decomposition f j = f j,1 + f j,2 with f j,1 ∈ N̊ν,⊥ and f j,2 ∈ N̊ν ,
j = 1, . . . , i. Suppose that f j,1 are not linearly independent, then there exists
j=1,··· ,i
a non-zero element g ∈ G ∩ Nν . To see this fix (λ1 , . . . , λn ) , (0, . . . , 0) with
λ1 f1,1 + · · · + λi fi,1 = 0. Then
λ1 f1,1 + f1,2 + · · · + λi fi,1 + fi,2 = λ1 f1,2 + · · · + λi fi,2 C g.
| {z } | {z }
∈G? ∈Nν
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 19
Note that span( f1,1 , . . . , fi,1 ) ⊂ N̊ν⊥ is also i-dimensional subspace in H01 . Hence, in
any case the reverse inequality follows.
Lemma 2.10. There exists c > 0 such that we have for all i ∈ N
λi EN ≤ cλi ED .
and H01 (Q) ⊂ H 1 (Q). Hence, the claim follows from Proposition 2.9.
Lemma 2.11. Let ν1 , ν2 be bounded Borel measures on Q . Suppose (PI) holds for
ν1 and
ν2 (E) ≤ ν1 (E), for all E ∈ B(Q).
Then we have ιν1 (u) = ιν2 (u) holds ν2 -almost surely for every u ∈ H 1 (Q). In
particular, if both ∆νD/N
1
and ∆νD/N
2
have compact resolvent, then
sνD/N
2
≤ sνD/N
1
and sνD/N
2
≤ sνD/N
1
.
Proof. We have for all u ∈ Cb∞ Q
kukLν2 ≤ c1 kukH 1 .
(Q) ≤ kukLν2 (Q)
2 1
N
Now, for each u ∈ H 1 there exists (un )n∈N ∈ Cb∞ Q such that un → ιν1 (u) in
Lν21 (Q) and un → u in H 1 (Q). Moreover, by assumption, we have Lν21 (Q) ⊂ Lν22 (Q)
and
un − ιν1 (u)
L2 ≤
un − ιν1 (u)
L2 → 0,
ν2 ν1
which yields ιν1 (u) = ιν2 (u) holds ν2 -almost surely. Now, the last claim follows from
the min-max principle stated in Proposition 2.9 and ιν1 = ιν2 ν2 -almost surely.
20 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
Lemma 2.12. For m > 1 and r > 0, let Q be a cube with side length mr and Q0 ⊂ Q
a centred and parallel sub-cube with side length r. Then there exists ϕQ,m ∈ Cc∞ (Rd )
which satisfies the following properties:
Proof. Let ψ : x 7→ 1B̊1 (0) c2 exp 1/ kxk2 − 1 be the normalised Friedrichs’ mol-
Q0
Q00
Q + B1 (0)
00
Z Z
ϕQ,m (x) = 1Q00 (y)ψ (x − y) dΛ(y) = 1Q00 (y)ψ (x − y) dΛ(y)
Rd x−Bε (0)
Z Z
= ψ (x − y) dΛ(y) = ψ (y) dΛ(y) = 1
x−Bε (0) Bε (0)
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 21
and consequently supp ϕQ,m ⊂ Q00 + Bε (0) ⊂ Q̊. Further, for x ∈ Q00 + Bε (0), we
have
Z
1 00 (y) (∂/∂xi ) ψ (x − y) dΛ(y)
(∂/∂xi ) ϕQ,m (x) =
Q00 Q
Z
2 |x − y |
≤ 1Q00 (y)
i
i 2 ψ (x − y) dΛ(y)
Q 00
ε2
x−y
2 − 1
ε
Z
1 2
≤ 1Q00 (y)
ψ (x − y) dΛ(y)
ε Q00 ∩Bε (x)
x−y
2 − 1 2
ε
Z
1 2
≤
ε Bε (0)+x
x−y
2 2 ψ (x − y) dΛ(y)
ε
−1
Z
1 2
= 2 ψ (y) dΛ(y)
ε Bε (0)
y
2
ε
− 1
Z
1 2
= ψ(y) dΛ(y).
ε B1 (0) kyk2 − 12
−2
Observing B (0) kyk2 − 1 ψ(y) dΛ(y) < ∞, the claim follows.
R
1
For s > 0 let hQi s denote the cube centred and parallel with respect to Q such that
Λ(Q) = s−d Λ hQi s , s > 0 (i. e. hQi s = T (Q)+(1− D T (x) = sx, x ∈ R Eand
d
D s)xE0 with
x0 ∈ R is the centre of Q). Note that we have hQi1/s = s Q + (1 − s )x0 =
d −1 −1
s s
Q.
Lemma 2.13. Let Q be a with side length mr > 0, m > 1. Then there exists a
constant C > 0 depending on m > 1 and d such that for ϕQ,m as defined in Lemma
2.12 we have
2 1−2/d
Λ hQi1/m
R
∇ϕQ,m dΛ
2 ≤ C(m − 1)−2 m1−2/d .
ν hQi1/m
R
ϕQ,m dν
Proof. Using Lemma 2.12, we find
2
Λ (Q) / (m − 1)2 r12
R
∇ϕQ,m dΛ 2 dC 2 Λ (Q)1−2/d
≤ dC ≤
ϕ2Q,m dν
R
(m − 1)2 m ν hQi1/m
ν hQi1/m
1−2/d
Λ hQi1/m
= dC 2 (m − 1)−2 m1−2/d .
ν hQi1/m
The following proposition applies only in the case d > 2.
Proposition 2.14. If dim∞ (ν) < d − 2 and d > 2, then the identity operator
C∞b Q , h·, ·i H
2
1 → Lν
22 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
is not continuous.
implies that there exists a sequence of cubes (Qn ) ∈ DN with strictly decreasing
diameters such that ν (Qn ) ≥ Λ (Qn )a/d , n ∈ N, for some a ∈ (dim∞ (ν) , d − 2).
Now we have for un B Λ hQn i2 1/d−1/2 ϕhQn i2 ,2 with C > 0 given in Lemma 2.12
Z Z !
2 2
kun k2H 1 = Λ hQn i2 2/d−1 + ϕ
∇ϕ
hQn i2 ,2 dΛ hQn i2 ,2 dΛ
Q Q
Z !
2/d−1 CΛ hQn i2 2
≤ Λ hQn i2 + ϕhQn i2 ,2 dΛ
Λ(Qn )2/d Q
≤ Λ hQn i2 2/d−1 4CΛ hQn i2 −2/d+1 + Λ hQn i2 ≤ 4(C + 1).
Even though, the particular choice of the set of dyadic cubes is not unique, we will
see that this does not affect our results (see Lemma 3.7, Lemma 3.9 and Lemma
3.12 ).
In this chapter we introduce the new notion of partition functions with respect
to a non-negative, monotone set function J defined on the dyadic cubes D. One
further natural assumption is that J is locally non-vanishing, that is, if J(Q) > 0 for
Q ∈ D, then there exists Q0 ( Q, Q0 ∈ D with J(Q0 ) > 0. Note that this assumption
is satisfied for each specific choice for J that we consider.
Of particular interests are the cases when J is chosen to be Jν as defined in
the introduction or J = ν, where in the latter case we obtain the well-known
Lq -spectrum, for which we also write βνDN .
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 23
3.1. The partition function. Recall the definition in (1.1) of the partition function
with respect to J given by
P
log q
D/N J(C)
C∈D
τJD/N (q) B lim sup τJ,n (q), with τJ,n
D/N
(q) B n
n→∞ log(2n )
n o
together with the critical values qJD/N B inf q ≥ 0 : τJD/N (q) < 0 and
X
κJ B inf q
< .
q ≥ 0 : J(Q) ∞
Q∈D
We start with some general observations for which we need the following objects:
\[n o maxQ∈DnN log J (Q)
supp (J) B Q : Q ∈ DnN , J (Q) > 0 and dim∞ (J) B lim inf .
k∈N n≥k
n→∞ − log 2n
We call dim∞ (J) the ∞-dimension of J which generalises the lower ∞-dimension
for ν defined in (♠). Obviously, the following holds.
Lemma 3.1. If dim∞ (J) > 0, then J is uniformly vanishing, i. e. we have
limn→∞ maxC∈DnN J(C) = 0.
Remark 3.2. Note that thanks to the monotonicity of J the assumption limn→∞ maxC∈DnN J(C) =
0 is equivalent to limn→∞ supC∈Sk≥n DN J(C) = 0.
k
Now, assume qJN < ∞. It follows there exists q > 0 such that τJN (q) < 0. Conse-
quently, we obtain from (3.1) − dim∞ (J)q ≤ τJN (q) < 0, which gives dim∞ (J) > 0.
Reversely, suppose dim∞ (ν) > 0. In the case dim∞ (J) = ∞, using (3.1), we have
qJN = 0 due to τJN (q) = −∞ for q > 0. Now, let us consider the case 0 < dim∞ (J) <
∞. Then it follows from (3.1) that τJN (q) < 0 for all q > dim M (J) / dim∞ (J) which
proves the implication.
24 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
Now, assume qJN < ∞. Then we have τJN (q) < 0 for all q > qJN , and therefore, for
every ε > 0 with τJN (q) < −ε < 0 and n large enough, we obtain
X
J(Q)q ≤ 2−nε ,
Q∈DnN
n o
implying Q∈D J(Q)q < ∞. This shows inf q ≥ 0 : Q∈D J(Q)q < ∞ ≤ qJN . For
P P
the reversed inequality we note that if qJN = 0, then the claimed equality is clear.
If, on the other hand, qJN > 0, then we necessarily have dim∞ (J) < ∞. Since,
τJN is decreasing, convex and proper, it follows that qJN is a zero of τJN and for all
0 < q < qJN
0 < τJN (q).
This implies that for every 0 < δ < τJN (q), there is subsequence (nk ) such that
2nk δ ≤
X
J(Q)q
Q∈DnNk
and therefore, X X X
∞= J(Q)q ≤ J(Q)q
k∈N Q∈DnN Q∈D
k
n o
and consequently qJN ≤ inf q ≥ 0 : q <∞.
P
Q∈D J(Q)
Remark 3.4. Note that in the case dim∞ (J) ≤ 0, we deduce from Lemma 3.3 τJN (q)
is non-negative for q ≥ 0, hence qJN = ∞. However, it is possible that κJ < ∞.
Indeed, in Example 9.2 we give an example of a measure ν, where κJν gives the
precise upper bound for the spectral dimension, while κJν < qJNν = ∞.
Lemma 3.5. If dim∞ (J) ∈ (0, ∞), then τJN is convex and strictly decreasing on R≥0 .
In particular, if qJN > 0, then qJN is the only zero of τJN .
Proof. First, note that Lemma 3.3 implies τJN (q) ∈ R for all q ≥ 0 and limq τJN (q) =
−∞. Since dim∞ (J) > 0 it follows Lemma 3.1 that for n large J(Q) < 1, Q ∈ DnN .
Hence, it follows that τJN is decreasing and as pointwise limit superior of convex
functions again convex. Now, we show that τJN is strictly decreasing. Assume there
exist 0 ≤ q1 < q2 such that τJN (q1 ) = τJN (q2 ). Since, τJN is decreasing function we
obtain τJN (q1 ) = τJN (q) for all q ∈ [q1 , q2 ]. Let q00 ∈ (q1 , q2 ). The convexity of τJN
implies for all q0 > q00
τJN (q00 ) − τJN (q1 ) τJN (q0 ) − τJN (q1 )
0= ≤ ≤ 0,
q00 − q1 q0 − q1
which implies τJN (q) = τJN (q1 ) for all q > q1 which contradicts limq τJN (q) = −∞.
For the second claim note that, since τJN is convex, it follows that τJN is continuous
on R>0 . Hence, we obtain τJN (qJN ) = 0. Finally, the uniqueness follows from the fact
that τJN is a finite strictly decreasing function.
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 25
Definition 3.6. We say that a non-negative, monotone set function J defined on all
possible dyadic sub-cubes of Q is locally almost subadditive if for any two sets of
e D of Q by dyadic cubes there exists a constant C > 0 such that
dyadic partitions D,
N
for every Q ∈ Dn we have
X
J Q0 .
J (Q) ≤ C
e N :Q∩Q0 ,∅
Q 0 ∈D n+2
Lemma 3.7. For a non-negative, monotone and locally almost subadditive set
function J defined on all possible dyadic sub-cubes of Q, we have that the definition
of τJD/N does not depend on the particular choice of the dyadic partition.
Proof. We only give a proof of the assertion in (3), namely τJN (0) = dim M (J).
First, observe that, if Q ∈ DnN , Q ∩ supp(J) , ∅, then there exists Q0 ∈ DnN with
Q0 ∩ Q , ∅ and J(Q0 ) > 0. This can be seen as follows: For x ∈ Q ∩ supp(J)
there exists a subsequence (nk ) such that x ∈ Qnk , Qnk ∈ Dnk and ν(Qnk ) > 0. For
k ∈ N such that nk ≥ n there exists exactly one Q0 ∈ DnN with Qnk ⊂ Q0 . Now,
x ∈ Qnk ⊂ Q0 , implying Q0 ∩ Q , ∅ and since J nis monotone, we have ν(Q o ) > 0.
0
Lemma 3.9. The definition of βνD/N and dim∞ (ν) does not depend on the particular
choice of the dyadic partition.
Fact 3.10. We make the following elementary observations:
(1) βνN (0) = dim M (ν) .
(2) dim∞ (ν) ≤ d.
(3) βνN (1) = 0 and if ν Q̊ > 0 , then also βνD (1) = 0.
(4) For the Dirichlet Lq -spectrum we have βνD = βν|D .
Q̊
(5) For q ≥ 0, we have −qd ≤ βνN (q) and if ν Q̊ > 0, also −qd ≤ βνD (q).
(6) If supp(ν) ⊂ Q̊, then we have βνD = βνN .
(7) If ν is absolutely continuous with density h ∈ LΛ t for some t > d/2, then
(8) The condition dim∞ (ν) > d −2 requires that the upper Minkowski dimension
dim M (ν) and the Hausdorff dimension dimH (ν) must also lie in (d − 2, d].
This in particular rules out the possibility of atomic parts of ν, if d ≥ 2; for
d = 1, atomic examples have been studied extensively in [KN22].
Remark 3.11. It is worth noting that the situation is much simpler in the one-
dimensional case, which follows from the fact that the boundary contains only two
points. Suppose ν is a non-zero Borel probability measure on (0, 1). Then for all
q ∈ [0, 1], βνD (q) = βνN (q).
3.3. The spectral partition function and connections to the Lq -spectrum. This
section is devoted to the special case J = Jν,a,b , where for b ≥ 0 and a ∈ R,
b
supQ∈D(Q) ν Qe log Λ Q e , a = 0,
Jν,a,b (Q) B
e
eb Λ Q e a,
sup ν Q a , 0,
e Q∈D(Q)
Hence, we have that J is non-negative and monotone uniformly vanishing and for
every Q ∈ DnN we have supQ0 ∈D(Q) ν(Q0 )b Λ(Q0 )a = ν (Qmax )b Λ (Qmax )a for some
28 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
N with m ≥ n. Hence,
Qmax ∈ D(Q) ∩ Dm
X b
Jν,a,b (Q) = ν (Qmax )b Λ (Qmax )a ≤ 2adm ν Q0
e N :Q0 ∩Qmax ,∅
Q 0 ∈D m+2
≤ 2−2ad 6d ν Q0 b Λ(Q0 )a
max
e N :Q0 ∩Qmax ,∅
Q 0 ∈D m+2
≤ 2−2ad 6d Jν,a,b Q0
max
e N :Q0 ∩Qmax ,∅
Q 0 ∈D m+2
≤ 2−2ad 6d Jν,a,b Q0
max
e N :Q0 ∩Q,∅
Q 0 ∈D n+2
X
≤ 2−2ad 6d Jν,a,b Q0 .
e N :Q0 ∩Q,∅
Q 0 ∈D n+2
We now elaborate some connections between the Lq -spectrum and the spectral
partition function.
Proposition 3.13. Let a < 0 and assume b dim∞ (ν) + ad > 0, then for all q ≥ 0, we
have
βνD/N (bq) − adq ≤ τJD/N
ν.a,b
(q) ≤ βνD/N (q (b + ad/ dim∞ (ν))) andτJD/N
ν,a,b
(0) = βνD/N (0) .
Proof. Let q > 0. For dim∞ (ν) > ε > 0, we have for n large enough and all
C ∈ DnD/N
ν(C) ≤ 2−εn
or, alternatively, ν(C)d/ ≤ Λ (C). Since, ν(C) becomes uniformly small for all
C ∈ DnD/N and n large, for every 0 < δ < b, we obtain
log (1/ν(C)) q ≤ ν(C)−qδ .
This leads to
X X
log(2)d ν(C)bq ≤ Jν,0,b (C)q
C∈DnD/N C∈DnD/N
d X
sup log(1/ν(Q) q ν(Q)bq .
≤
ε D/N Q∈D(C)
C∈Dn
d X
≤ ν(C)q(b−δ)
ε D/N
C∈Dn
Hence,
βνD/N (qb) ≤ τJD/N
ν,0,b
(q) ≤ βνD/N (q(b − δ))
and for δ & 0, the continuity of βνD/N gives βνD/N (qb) = τJD/N
ν,0,b
(q). In the same way we
obtain under assumption that βνD/N exists as limit that βνD/N (bq) = lim inf n→∞ τJD/N
ν,0,b ,n
(q).
Corollary 3.16. If d = 2 and dim∞ (ν) > 0, then τJNν (1) = βνN (1) = 0, or equiva-
lently, qJNν = 1. If additionally ν Q̊ > 0, then τJDν (1) = βνD (1) = 0, or equivalently,
qJDν = 1.
By virtue of Proposition 3.13 and Proposition 3.15 we arrive at the following list
of facts.
Fact 3.17. Assuming b dim∞ (ν) + ad > 0, the following list of properties of the
spectral partition function applies:
(1) supp Jν,a,b = supp (ν).
(2) dim∞ (Jν,a,b ) = b dim∞ (ν) + ad > 0.
(3) We have that qJNν,a,b is the unique zero of τa,b
N and
dim∞ (ν)
qJNν,a,b ≤ .
b dim∞ (ν) + ad
(4) We have dim∞ (ν) ≤ dim M (ν) and
d dim M (ν)
(2/d−1) ≤ qJν .
N N
≤ ≤ qνΛ
2 dim M (ν) − d + 2
If additionally, dim∞ (ν) = dim M (ν), then
dim M (ν)
(2/d−1) = qJν = .
N N
qνΛ
dim M (ν) − d + 2
30 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
(5) If ν is absolutely continuous with density h ∈ LΛ r for some r > d/2, then
τJDν (q) = τJNν (q) = βνN (q) + (d − 2)q, for all q ∈ [0, r].
(6) For the Dirichlet spectral partition function we have τJDν,a,b = τJDν| ,a,b .
Q̊
(7) For c > 0, we have τJD/N cν,a,b
= τJD/N
ν,a,b
and we can assume without loss of
generality that ν is a probability measure.
3.4. Special cases. In this section we show that for some particular cases (abso-
lutely continuous measures, Ahlfors–David regular measures, and self-conformal
measures) the spectral partitions function is completely determined by the Lq -
spectrum assuming dim∞ (ν) > d − 2. Furthermore, for these classes of measures
we investigate under which conditions the Dirichlet and the Neumann Lq -spectra
coincide. Later, we will use the results to calculate the spectral dimension for these
classes of measures.
q
≤ 2−n(2q−d) k f kLq .
Λ
This proves the claim for d/2 ≤ q ≤ r. For the remaining case, we use the convexity
of τJD/N
ν
, the lower bound obtained above and the fact that τJD/N
ν
(0) ≤ d, to obtain for
q ∈ [0, r],
d − 2q ≥ τJD/N
ν
(q) ≥ βνD/N (q) − (2 − d)q = d − 2q.
This implies
n o
log card Q ∈ DnN : ν (Q) > 0
dim M (ν) = lim
n→∞ log 2n
n o
log card Q ∈ DnD : ν (Q) > 0
= lim = α = dim∞ (ν) .
n→∞ log 2n
Indeed, since ν Q̊ > 0 we find an element E ∈ D with E ⊂ Q̊ and ν (E) = ε > 0.
Then, on the one hand, for large n ∈ N we have
n o
n o log card Q0 ∈ DnD : ν (Q0 ) > 0
card Q0 ∈ DnD : ν Q0 > 0 C2−nα ≥ ε =⇒ lim inf ≥ α.
n log 2n
On the other hand,
n o
n o log card Q0 ∈ DnN : ν (Q0 ) > 0
card Q0 ∈ DnN : ν Q0 > 0 C −1 2−nα 3−d ≤ 1 =⇒ lim sup ≤ α.
n log 2n
Next we prove that for all q ≥ 0
τJD/N
ν
(q) = βνD/N (q) + (2 − d) q = (α + 2 − d) q − α
exists as a limit. Indeed,
X X n o
Jν (Q)q ≤ C 2−n(α+2−d)q ≤ C card Q ∈ DnN : ν (Q) > 0 2−n(α+2−d)q
Q∈DnN Q∈DnN
and
X X
Jν (Q)q ≥ ν (Q)q 2(d−2)nq
D
Q∈Dn+1 D
Q∈Dn+1
X D E q
≥ 2(d−2)nq 4−dq−d ν Q̊
2
Q∈DnD ,ν(Q)>0
n o
≥ C −1 4−dq−d card Q ∈ DnD : ν (Q) > 0 2−n(α+2−d)q ,
giving the claim.
3.4.3. Conformal iterated function systems. Let U ⊂ Rd be an open set. We say a
C 1 -map S : U → Rd is conformal if for every x ∈ U the matrix S 0 (x), giving the
total derivative of S in x, satisfies |S 0 (x) · y| = kS 0 (x)k |y| for all y ∈ Rd .
Let us assume that Q is closed. A family of mappings {S i : Q → Q}i=1,...,` , ` ≥ 2
is a conformal iterated function system if each S i extends to an injective conformal
map S i : U → U on an open set U and the following conditions are satisfied:
n
o
sup
S 0 (x)
: x ∈ U < 1
i
and there exists a constant D ≥ 1 such that for all n ∈ N and u ∈ {1, . . . , `}n
−1
S u0 (x)
D ≤ ≤D
kS u0 (y)k
32 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
Let (pi )i∈I be the associated positive probability vector and define pu = |u|
Q
i=1 pui .
Then there is a unique Borel probability measure ν with support K such that
`
X
ν(A) = pi ν S −1 (A)
i=1
for A ∈ B Rd . We refer to ν as the self-conformal measure.
We need the following result from [PS00, Theorem 1.1]:
For a self-conformal measure ν, the Lq -spectrum βνN exists as a
limit on R>0 .
Proposition 3.19. For d > 2 let ν denote a self-conformal measure on the closed
cube Q such that (♠) holds. Then for q ≥ 0,
0
If Q0 ∈ Dn+m
N ∩ D (Q), m ∈ N and u ∈ I Q we have diam S u−1 (Q0 ) ≤ L2−m
for some L > 0 and hence it is contained in at most 3d cubes from Dm−kN with
−1 0
k B log(L)/ log(2) (this gives diam S u (Q ) ≤ 2 −m+k ). Also, by definition of
Q 0
I and Wn , we have
[ [ [
S u (Q) ⊂ Q00 ⊂ Q03 B Q00 .
0
u∈I Q Q00 ∈DnN ,Q00 ∩Q0 ,∅ Q00 ∈DnN ,Q00 ∩Q,∅
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 33
Then we have
X
ν Q0 Λ Q0 a/d = 2−a(n+m) pu ν S u−1 Q0
u∈Wn
X
=2 −an
pu 2−am ν S u−1 Q0
0
u∈I Q
X X
≤ 2−an pu 2−ak 2−a(m−k) ν(C)
u∈I Q0 N ,S −1 (Q0 )∩C,∅
C∈Dm−k u
n o X
≤ 2−ak 3d max ν (C) Λ (C)a/d 2−an pu
N
C∈Dm−k 0
u∈I Q
n o [
= 2−ak 3d max ν (C) Λ (C)a/d 2−an ν S u (Q)
N
C∈Dm−k 0
u∈I Q
≤ 2−ak 3d Kν Q03 2−an .
Since in the above inequality Q0 ∈ D (Q) was arbitrary, we deduce for q > 0,
X X q
Jν (Q)q ≤ 2(d−2)kq 3dq K q 2−naq ν Q03
Q∈DnN Q∈DnN
q
X X
≤ 2(d−2)kq 3dq K q 2−naq ν(Q )
0
Q∈DnN Q0 ∈Dn ,Q0 ∩Q,∅
X
≤ 2(d−2)kq 3dq K q 2−naq 3dq max ν(Q0 )q
Q0 ∈DnN ,Q0 ∩Q,∅
Q∈DnN
X
≤ 2(d−2)kq 3dq K q 2−naq 3dq+d ν(Q)q .
Q∈DnN
This gives βνN (q) − adq ≥ τJNν (q). Furthermore, observe that βJNν (0) = dim M (ν) =
βνN (0). To complete the proof, observe that
X X
ν (Q)q Λ (Q)a ≤ Jν (Q)q .
Q∈DnN Q∈DnN
Finally, [PS00, Theorem 1.1] gives βνN (q)−adq ≤ lim inf n→∞ τJNν ,n (q) for q > 0.
Proposition 3.20. Let ν denote a self-conformal measure on Q̊ and dim∞ (ν) > d−2
. Then
βνN (q) = βνD (q) = lim inf βν,n
D
(q) = lim inf βν,n
N
(q)
n→∞ n→∞
for all q > 0.
Proof. We use the same notation as in the proof of Proposition 3.19. By our
assumption there exists n ∈ N such that S u (Q) ⊂ Q̊ for some u ∈ Wn . Indeed
assume for all n ∈ N and u ∈ Wn , we have
S u (Q) ∩ ∂Q , ∅.
34 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
Further, using supu∈Wn diam (S u (Q)) ≤ 2−n → 0 for n → ∞ and K ⊂ u∈Wn S u (Q),
S
we deduce that K ⊂ ∂Q. This gives ν(∂Q) > 0 contradicting our assumption.
Let us√assume that the distance of S u (Q) to the boundary of Q is at least
2 −n−m 0 +2 d for some m0 ∈ N. Then all cubes Q ∈ Dn+m N intersecting S u (Q)
lie in Dn+m for all m > m0 . Therefore, using the self-similarity and [PS00, Lemma
D
2.2 \{}& 2.4] (with constant C1 from there) we have for q > 0
q
X X X
ν (Q) =
q pv ν S v Q
−1
D
Q∈Dn+m D
Q∈Dn+m v∈Wn
q
X q q
X
≥ pu ν S u−1 Q ≥ C1−1 pu ν (Q)q .
N
Q∈Dn+m N
Q∈Dm
This gives βνN (q) ≤ βνD (q) and lim inf n→∞ βν,n
N (q) ≤ lim inf
n→∞ βν,n (q) for q > 0.
D
The reverse inequalities are obvious. Hence, the claim follows from [PS00, Theorem
1.1].
Corollary 3.21. Let ν denote a self-conformal measure on Q̊ and dim∞ (ν) > d − 2.
Then we have
βνN (q) + (d − 2)q = τJNν (q) = lim inf τJNν ,n (q) = τJDν (q) = lim inf τJDν ,n (q).
n→∞ n→∞
Proof. The case d = 2 follows immediately from Proposition 3.15 and Proposi-
tion 3.20. For d > 2, we obtain from Proposition 3.20 and Proposition 3.19 the
following chain of inequalities
βνN (q) + (d − 2)q = lim inf βν,n
D
(q) + (d − 2)q
n→∞
≤ lim inf τJDν ,n (q) ≤ lim inf τJNν ,n (q)
n→∞ n→∞
= τJNν (q) = βνN (q) + (d − 2)q.
Proof. The fourth inequality is always true. We only have to consider the case κJ <
∞. The first statement follows from the monotonicity of J, limn→∞ supi≥n,C∈DN J(C) =
i
0 and the definition of D. Further, Lemma 3.3 gives κJ ≤ qJN (where equality holds
if dim∞ (J) > 0, otherwise qJN = ∞). Let 0 < t < J(Q). Setting
Rt B {Q ∈ D : J(Q) ≥ t} ,
we note that for Q ∈ Pt there is exactly one Q0 ∈ Rt ∩DNlog Λ(Q) /d−1 with Q ⊂ Q0 and
| 2 |
for each Q0 ∈ Rt ∩ DNlog Λ(Q) /d−1 there are at most 2d elements of Pt ∩ DNlog Λ(Q) /d
| 2 | | 2 |
such that they are subsets of Q0 . Hence,
card Pt ≤ 2d card Rt .
For q > κJ we obtain
∞
X X ∞
X X
tq card Pt = tq 1 ≤ 2d tq
n=1 Q∈Pt ∩DnN n=0 Q∈Rt ∩DnN
∞
X X ∞ X
X
≤ 2d q
J(Q) ≤ 2 d
J(Q)q < ∞.
n=0 Q∈Rt ∩DnN n=0 Q∈DnN
This implies
card Pt
lim sup ≤ q.
t↓0 − log(t)
Now, q tending to κJ proves the third inequality. The second inequality follows
immediately form the observation that MJ (x) ≤ card P1/x . The first inequality
follows from the following observation. For α > 0, n ∈ N and P ∈ ΠJ such that
maxC∈P J (C) < 2−nα , we have
n o
N
Nα,J (n) = card Q ∈ DnN : J(Q) ≥ 2−αn ≤ card (P) ,
where we used the fact that for each Q with J(Q) ≥ 2−αn there exists at least one
Q0 ⊂ D (Q) ∩ P and the this assignment is injective. The last claim follows from
the fact that for every α > 0 and q = qJD , we have
(n) ≤ 2αnq
X
D
Nα,J J(Q)q .
Q∈DnD
Remark 4.2. Alternatively, the algorithm stated in Proposition 4.1 can be also
formulated as follows (here we follow [HKY00]):
Let 0 < t < J(Q). We say Q ∈ D is bad, if J(Q) ≥ t, otherwise we call Q
good. We generate a partition of Q of elements of D into good intervals which will
be denoted by Gt . Since by the choice of t, we see that Q is bad. Hence, we put
B0 B {Q}. Now, we divide each element of B0 into 2d cubes of D of equal size and
check whether they are good, in which case we move these cubes to Gt , or they are
bad, in which case they are put into B1 .We repeat this procedure until the set of bad
cubes is empty. The process terminates, which is ensured by the assumption that J
36 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
where qnk is the unique zero of τnNk . Further, suppose lim inf qnk > 0. Then we have
hJ ≤ lim inf qnk .
k→∞
Proof. First of all note that we have Q∈DnN J(Q)qnk = 1. Further, since J is
P
k
uniformly decreasing, we choose k large enough such that J(Q) < 1 for all Q ∈ DnNk .
This ensures that τnNk has a unique zero. Hence, we obtain
K 1/qnk
max J(Q) ≤ .
2nk τnk (0)/qnk
N
Q∈DnNk
This implies
log 2nk τnk (0)
N
In particular, we are interested in the special choice J J,a (Q) B J(Q)Λ(Q)a , a > 0,
Q ∈ D, where J is a non-negative, finite, locally non-vanishing, subadditive function
on D, that is, if Q ∈ D is decomposed into a finite number of disjoint cubes Q j of
j
D, then J(Q j ) ≤ J(Q). We are interested now in the growth properties of γJJ,a ,n .
P
Upper estimates for γJJ,a ,n have been first obtained in [BS67; Bor71]. Here, we
proceed as follows: First we present an adaptive partition algorithm (in the sense
of [DeV87]) going back to Birman/Solomyak [BS67; Bor71] to obtain well-known
upper bounds on γJJ,a ,n . After that we employ the estimates in Proposition 4.1 to
partially improve and extend the results in [BS67; Bor71].
In following we use the terminology as in [DKS20]. Let Ξ0 be a finite partition
of Q of dyadic cubes from D. We say a partition Ξ0 of Q is an elementary extension
of Ξ0 , if it can be obtained from by uniformly splitting some of its cubes into 2d
equal sized disjoint cubes lying in D with half side length. We call a partition Ξ
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 37
Let (Pk )k∈N denote a sequence of dyadic partitions obtained recursively as follows:
set P0 B Ξ0 and, for k ∈ N, construct an elementary extension Pk of Pk−1 by
subdividing all cubes Q ∈ Pk−1 , for which
J J,a (Q) ≥ 2−daGa (Pk−1 ),
into 2d equal sized cubes. Then, for k ∈ N, we have
Ga (Pk ) B max J J,a (Q) ≤ Cεmin(1,a) (Nk − N0 )−(1+a) J(Q)
Q∈Pk
1
1+a 1 1 X a 1
2−daGa (Pk−1 ≤ min J J,a (Q) 1+a ≤ Λ(Q) 1+a J(Q) 1+a .
Q∈S ki ti,k i
Q∈S k
1
X X
1+a 1 ≤ 1 ε min(1,a)
2−daGa (Pk−1 ) ≤ Λ(Q) J(Q) 1+a .
ti,k i
i
ti,k
Q∈S k Q∈S k
This is equivalent to
ad min(1,a) 1
ti,k ≤ 2 1+a ε 1+a (Ga (Pk−1 ))− 1+a .
ad min(1,a) 1
Since tk = t1,k + t2,k , we have tk ≤ 21+ 1+a ε 1+a (Ga (Pk−1 ))− 1+a . By the definition of
the dyadic subdivision P j , j ∈ N,
Ga (P j ) ≤ 2−ad Ga (P j−1 ). (4.1)
Now, applying (4.1) recursively, for all integers j ≤ k we obtain Ga (Pk−1 ) ≤
2−ad(k− j)Ga (P j−1 ). Since for all j ∈ N we have N j − N j−1 = (2d − 1)t j . Hence, for
38 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
all k ∈ N , we deduce
k
X
Nk − N0 = (2 − 1)
d
tj
l=1
k
ad min(1,a) X 1
≤ 21+ 1+a ε 1+a (2d − 1) Ga (P j−1 )− 1+a
j=1
k
ad min(1,a) 1
X ad
≤ 21+ 1+a ε 1+a (2d − 1)Ga (Pk−1 )− 1+a 2− 1+a (k− j)
j=1
ad −1
ad min(1,a) 1
≤ 1 − 2− 1+a 21+ 1+a ε 1+a (2d − 1)Ga (Pk−1 )− 1+a
ad −1
ad min(1,a) ad 1
≤ 1 − 2− 1+a 21+ 1+a ε 1+a (2d − 1)2− 1+a Ga (Pk )− 1+a .
This proves our first claim. For second claim note that Nk−1 ≤ Nk ≤ 2d Nk−1 . Hence,
for n ∈ N with Nk−1 ≤ n ≤ Nk and using n2−d ≤ Nk−1 , we obtain
γJJ,a ,n ≤ γJJ,a ,Nk−1
≤ Cεmin(1,a) (Nk−1 − N0 )−(1+a)
−(1+a)
≤ Cεmin(1,a) 2−d n − N0 .
Definition 4.5. We call J singular function with respect to Λ, if for every ε > 0
there exist two partitions Ξ00 ⊂ Ξ0 ⊂ D of Q such that
X X
Λ(Q) ≤ ε and J(Q) ≤ ε.
0 0
Q∈Ξ0 \Ξ0 Q∈Ξ0
Using Proposition 4.1, we are able to extend the class of set functions considered
in [BS67, Theorem 2.1.] (i. e. we allow set functions J for which J is only assumed
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 39
to be non-negative, monotone and dim∞ (J) > 0). We obtain the following estimate
for the upper exponent of divergence of γJ,n given by
log γJ,n
αJ B lim sup .
n→∞ log(n)
Proposition 4.8. Let J be a non-negative monotone function on D such that
dim∞ (J) > 0. Then
1 1 dim∞ (J)
− = αJ ≤ − N ≤ − .
hJ qJ dim M (J)
In particular, for J = J J,a , we have
1 1 dim M (J) + ad
− = αJJ,a ≤ − ≤− ≤ −(1 + a).
hJJ,a qJNJ,a dim M (J)
Remark 4.9. If τJNJ,a (q) < d(1 − q(1 + a)) for some q ∈ (0, 1), then this estimate
improves the corresponding results of [Bor71; BS67, Theorem 2.1.], where only
αJJ,a ≤ −(1 + a) has been shown.
Proof. For all ε > 0, we have for n large
1/(hJ +ε)
MJ n ≤ n,
this gives minP∈ΠJ ,card(P)≤n maxQ∈P J(Q) ≤ n−1/(hJ +ε) . Thus, in tandem with Propo-
sition 4.1 and Fact 3.8, we see that αJ ≤ −1/hJ ≤ −1/qJN ≤ − dim∞ (J)/dim M (J).
To prove the equality, it is left to show αJ ≥ −1/hJ . First, assume αJ > −∞, then
for all ε > 0 with αJ + ε < 0, for n large, we have
min max J(Q) ≤ nαJ +ε .
P∈ΠJ ,card(P)≤n Q∈P
This implies qJNJ,a ≤ τNJ (0)/(τNJ (0) + ad) ≤ 1/(1 + a). From Proposition 4.1 we
deduce
1 1 dim M (J) + ad
− ≤− N ≤− ≤ −(1 + a).
hJ qJJ,a dim M (J)
40 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
D/N D/N
Recall the definition of Nα,J and F J , F JD/N from the introduction. Since the
following arguments do not depend on the particular choice DD/N , we omit the
dependence on D/N and write τ B τJD/N and τn B τJ,n D/N
.
Lemma 5.3. For α ∈ (0, dim∞ (J)) and n large, we have
D/N
Nα,J (n) = 0.
In particular,
D/N D/N
D/N log Nα,J (n) log Nα,J (n)
FJ = sup lim sup , F JD/N = sup lim inf .
α≥dim∞ (J) k→∞ log 2n α α≥dim∞ (J) k→∞ log 2n α
Proof. For fixed α > 0 with 0 < α < dim∞ (J) we have, by the definition of
dim∞ (J), we have for n large maxQ∈DnN J(Q) ≤ 2−αn . Hence, for all Q ∈ DnN we
have J(Q) ≤ 2−nα . For every 0 < α0 < α, it follows that NαD/N
0 ,J (n) = 0. This proves
the claim.
We need the following elementary observation from large deviation theory which
seems not to be standard in the relevant literature.
Lemma 5.4. Suppose (Xn )n∈N are real-valued random variables on some proba-
bility spaces (Ωn , An , µn ) such that the rate function
R c (t) B lim supn→∞ cn (t) is a
proper convex function with cn (t) B a−1 n log exp tX n dµn , t ∈ R, an → ∞ and
such that 0 belongs to the interior of the domain of finiteness {t ∈ R : c (t) < ∞}. Let
I = (a, d) be an open interval containing the subdifferential ∂c (0) = [b, c] of c in 0.
Then there exists r > 0 such that for all n sufficiently large,
µn a−1
n Xn < I ≤ 2 exp (−ran ) .
Proof. We assume that ∂c (0) = [b, c] and I = (a, d) with with a < b ≤ c < d. First
note that the assumptions ensure that −∞ < b ≤ c < ∞. We have by Chebychev
inequality for all q > 0,
Z
µn a−1n X n ≥ d = µn (qX n ≥ qan d) ≤ exp (−qan d) exp (qXn ) dµn
implying
n log µn an Xn ≥ d ≤ inf c (q) − qd = inf c (q) − qd,
lim sup a−1 −1
q>0 q∈R
where the equality follows from the assumption d < ∂c (0) and c (q)−qd ≥ (c−d)q ≥
0 for all q ≤ 0. Similarly, we find
n log µn an Xn ≤ a ≤ inf c (q) − qa = inf c (q) − qa.
lim sup a−1 −1
q<0 q∈R
We are left to show that both upper bounds are negative. We show the first case
by contradiction – the other case follows in exactly the same way. Assuming
inf q∈R c (q) − qd = 0 implies for all q ∈ R that c (q) − qd ≥ 0, or after rearranging,
c (q) − c (0) ≥ dq. This means, according to the definition of the sub-differential,
that d ∈ ∂c (0), contradicting our assumptions.
42 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
Proposition 5.5. For a subsequence (nk ) define the convex function on R≥0 by
B B lim sup τnD/N
k
and for some q ≥ 0 we assume B(q) = lim τnD/Nk
(q) and set
[a , b ] B −∂B (q). Then we have a ≥ dim∞ (J) and
0 0 0
D/N
a0 q + B(q) log Nα,J (nk )
≤ sup lim inf
b0
α>b0 k→∞ α log 2 k
n
D/N D/N
log Nα,J (nk ) log Nα,J (nk )
≤ sup lim inf = sup lim inf .
α≥dim∞ (J) k→∞ α log 2nk α>0 k→∞ α log 2nk
Moreover, if B(q) = τD/N (q), then [a, b] = −∂τD/N (q) ⊃ −∂B (q). Further, if
additionally 0 ≤ q ≤ qJD/N , then
which gives a0 ≥ dim∞ (J) > 0. Let q ≥ 0. Now, for all k ∈ N and s < a0 ≤ b0 < t,
we have
D/N
n o
Nt,J (nk ) ≥ card C ∈ DnD/N
k
: 2−snk
> J (C) > 2−tnk
| {z }
BLns,tk
J (C)q 2 snk q ≥ 2 snk q+nk τnk (q) 1Lns,t (C)J (C)q 2−nk τnk (q)
X X
≥
k
C∈Lns,tk C∈DnD/N
k
snk q+nk τnk (q)
1 s,t { (C)J (C) 2 k .
τ
X
=2 q −n (q)
1 − k n
L nk
D/N
C∈Dnk
We use the lower large deviation principle for the process Xk (C) B log J (C) with
probability measure on DnD/N
k
given by µk ({C}) B J (C)q 2−nk τnk (q) . We find for the
free energy function
1
c (x) B lim sup n
log Eµk exp xXk
nk log 2 k
X
1 τ
= lim sup J (C) /2 k
x+q n (q)
n
log k n
nk log 2 k D/N
C∈Dnk
with −∂c (0) = [a0 , b0 ] ⊂ (s, t) and hence there exists a constant r > 0 depending on
s, t and q such that for k large by Lemma 5.4
!
Xk
1 s,t { (C)J (C) /2 τ
X
q nk n k
(q)
= µk < (−t, −s) ≤ 2 exp (−rnk ) .
D/N
L nk log (2nk )
C∈Dnk
Therefore,
D/N
log Nt,J (nk )
lim inf ≥ sq + B(q)
k→∞ log 2nk
for all s < a0 and t > b0 and hence
D/N
log Nt,J (nk ) a0 q + B(q) a0 q + B(q)
sup lim inf ≥ sup = .
t>b0 k→∞ t log 2nk t>b0 t b0
The fact that −∂τ (q) ⊃ −∂B (q) if τ(q) = B(q) follows immediately from lim supk τnk ≤
τ.
The following corollary shows that our result in Proposition 5.5 covers the
corresponding statement for the one-dimensional case in [KN22, Prop. 4.17].
Corollary 5.6. Let J(Q) B ν(Q)Λγ (Q), Q ∈ D with γ > 0 and d ≥ 1. Then
τJD/N (q) = βνD/N (q) − γdq, q ≥ 0 and dim∞ (J) = dim∞ (ν) + dγ > 0. Suppose there
exists a subsequence (nk ) and q ∈ [0, 1] such that τJD/N (q) = limk τJ,n
D/N
k
(q). Then for
D/N
B B lim sup τJ,nk
, we have −∂B (q) B [a0 , b0 ] ⊂ ∂τJD/N (q) B [a, b] and
Proof. Due to Proposition 4.1, we can restrict our attention to the case qJD/N > 0.
First, assume τ (q) = lim inf n τn (q) for q ∈ qJD/N − ε, qJD/N , for some ε > 0 and
set [a, b] = −∂τ qJD/N . Then by the convexity of τ we find for every ∈ 0, qJD/N
an element q ∈ qJD/N − , qJD/N such that τ is differentiable in q with − (τ)0 (q) ∈
(b−, b] since the points where τ is differentiable on (0, ∞) lie dense in (0, ∞) which
follows from the fact that τ is a decreasing function and the left-hand continuity of
left-hand derivative of the convex function τ. Then we have by Proposition 5.5
log+ Nα,J
D/N D/N
(n) log Nα,J (n)
sup lim inf ≥ sup lim inf
α≥τ+ n→∞ α log 2n
α>−(τ)0 (q) n→∞ α log 2 n
J
6. Upper bounds
6.1. Upper bound for spectral dimension. This section establishes a relation
between embedding constants on sub-cubes and the spectral dimension.
Proposition 6.1. Suppose there exist a non-negative, uniformly vanishing,
mono-
∞
tone set function J on D such that for all Q ∈ D and all u ∈ Cb Q with
u dΛ = 0, we have
R
Q
kuk2L2 (Q) ≤ J(Q) k∇uk2L2 (Q) .
ν Λ
Then we have
sD ≤ sN ≤ hJ and sD ≤ sN ≤ hJ .
Proof. For a partition Ξ of Q into elements of D, let us define the following closed
linear subspace of H 1
( Z )
FΞ = u ∈ H : 1
u dΛ = 0, Q ∈ Ξ .
Q
Next we show that Cb∞ (Q) ∩ FΞ lies dense in FΞ with respect to H 1 . Since Q
has the extension property we readily see that Cb∞ (Q) lies dense in H 1 . Hence for
every u ∈ FΞ , there exists a sequence un in Cb∞ (Q) such that un → u in H 1 . The
Cauchy-Schwarz inequality gives for all Q ∈ Ξ
Z Z Z
un dΛ = un − u dΛ ≤ (un − u) dΛ → 0.
2
Q Q
obtain Z
kEQ k
ι(u)2 dν ≤ max J(Q) k∇uk2L2 (Q) .
DQ Q∈Ξ Λ
Define for i ∈ N
λiν,FΞ B inf sup RH 1 (ψ) : ψ ∈ G? : G <i FΞ , h·, ·i H 1
n n o o
n o
RH 1 (ψ) B hψ, ψiH 1 /hιψ, ιψiν and N N (y, FΞ ) B card i ∈ N : λiν,FΞ ≤ y ,y > 0.
Hence, maxQ∈Ξ J(Q) < DQ /(kEQ k x), implies
λ1ν,FΞ > x.
In view of the min-max principle as stated in Proposition 2.9, we deduce
N N (x) ≤ N N (x, FΞ ) + card(Ξ) = card(Ξ),
implying N N (x) ≤ MJ (kEQ k x/DQ ) and hence sN ≤ hJ and sN ≤ hJ .
Remark 6.2. Note that in the one dimensional case the assumption of Proposition 6.1
is always valid. Indeed, R exists C > 0 such that for all intervals I contained in
there
[0, 1] and u ∈ C∞
b I with I
u dΛ = 0, we have
kuk2L2 (I) ≤ Cν(I)Λ(I) k∇uk2L2 (I) = CJν,1,1 (I) k∇uk2L2 (I) ,
ν Λ Λ
(see for instance the proof of [BS67, Theorem 3.3.]). With this observation our
general results reproduce the upper bounds for spectral dimension in d = 1 in terms
of the fixed point of the Lq -spectrum ([KN22]).
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 47
Remark 6.3. The ideas underlying in Proposition 6.1 correspond to some extent to
those developed in[NS95; Sol94],[NS01, Chapter 5], that is, reducing the problem
of estimating the spectral dimension to an auxiliary counting problem. To illustrate
the parallel, we present an alternative proof of the upper estimate of the eigenvalue
counting function for self-similar measures under OSC ([Sol94, Theorem 1]). As
in the setting in [Sol94] we let ν denote a self-similar measure under OSC with
contractive similitudes S 1 , . . . , S m and corresponding contraction ratios
hi ∈ (0, 1)
and probability weights pi ∈ (0, 1), for i = 1, · · · , m . We assume ν Q̊ > 0 and
dim∞ (ν) > d − 2, which is in this case equivalent to maxi pi h2−d
i < 1. For simplicity
we assume the feasible set is given by Q̊, i. e. S j (Q̊) ⊂n Q̊. Instead
of D
o we will
consider a symbolic partition by the cylinder sets D e B T ω Q̊ : ω ∈ I ∗ with I B
{1, . . . , m}. Then J will be replaced by e J:D e → R≥0 with e J T ω Q̊ B pω h2−d ω ,
ω ∈ I ∗ . Now, observe that for 0 < t < mini=1,...,m pi h2−d
i , we have
n o
et B ω ∈ I ∗ : pω h2−d
P ω < t ≤ pω− hω− ,
2−d
P (2−d) δ
where δ is the unique solution of m i=1 pi hi = 1. Then there exists K > 0 such
that for all u ∈ H 1 with T (Q̊) u dΛ = 0, ω ∈ P
R
et
ω
Z Z Z
ι(u) dν ≤ K max J T ω Q̊
2 e |∇u| dΛ < tK
2
|∇u|2 dΛ
ω∈P
et Q Q
(see [NS01, p. 502]). Then a simple computation gives the two-sided estimate
t−δ
t−δ ≤ card P
et ≤ .
mini=1,...,m pi h2−d
i
hence the results of [NS01; NS95, Theorem 1.] follow from this simple counting
argument without any use of renewal theory. The drawback of the ideas [NS95;
Sol94; NS01, Chapter 5] is that they rely heavily on the specific structure of the
self-similar measures, whereas our approach via dyadic cubes avoids the use of
specific properties of the underlying measure.
6.2. Upper bounds for the Jν,a,b -partition entropy. This section is devoted to
the study of the Jν,a,b -partition entropy, which is ultimately associated with the
spectral dimension for a certain choice of parameters a, b. Let us introduce the
following notation: Ma,b (x) B MJν,a,b (x). x > 0 as well as
Proof. First, note that Proposition 3.15 implies dim∞ (Jν,0,b ) = b dim∞ (ν) > 0.
Therefore, we deduce that Jν,a,b is uniformly vanishing by Lemma 3.1. An applica-
tion of Proposition 4.1 with J =Jν,0,b gives h0,b ≤ qJNν,0,b . Furthermore, we have by
Proposition 6.4 that βνN (qb) = τJNν,0,b (q). Thus, we obtain by Lemma 3.3
X
q
< = qJNν,0,b = 1/b.
inf q ≥ 0 : J (C) ∞
ν,0,b
C∈D
The rest of this section deals with the case a , 0. Recalling the definition of
qJN , N ≤ qN
we find qνΛ a Jν,a,1
with equality for the case a > 0. We need the following
elementary lemma.
Lemma 6.5. For a, b ∈ R with a < b, let ( fn : [a, b] → R)nN be a sequence of
decreasing functions converging pointwise to a function f . We assume that fn has a
unique zero in xn , for all n ∈ N and f has a unique zero in x. Then x = limn→∞ xn .
Proposition 6.6. If b dim∞ (ν) + ad > 0 and a , 0, then
qJNν,a/b,1 dim∞ (ν)
ha,b ≤ ≤ .
b b dim∞ (ν) + ad
In the case a > 0, we have
n o dim M (ν) 1
ha,b ≤ qJNν,a,b = inf q > 0 : βνN (bq) < adq ≤ ≤ .
bdim M (ν) + ad b+a
In particular, if dim∞ (ν) > d − 2, then for all t ∈ (0, 2 dim∞ (ν) /(d − 2)) we have
t dim∞ (ν)
h2/d−1,2/t ≤ qJNν,t(2/d−1)/2,1 ≤ .
2 2 dim∞ (ν)/t + 2 − d
Moreover, limt↓2 qJNν,t(2/d−1)/2,1 = qJNν .
Proof. Since b dim∞ (ν) + ad > 0, we obtain from Fact 3.8 that dim∞ (Jν,a/b,1 ) =
∞ (ν)
dim∞ (ν) + ad/b > 0. This is equivalent to qJNν,a/b,1 ≤ dimdim
∞ (ν)+ad/b
due to Fact 3.8.
Using the definition of Ma,b (x), Proposition 4.1 applied to J = Jν,a/b,1 , we obtain
log Ma,b (x)
log Ma/b,1 x1/b qJNν,a/b,1
lim sup = lim sup ≤ .
x→∞ log(x) x→∞ b log(x1/b ) b
The estimate of qJNν,a,b for the case a > 0 follows from βν (bq) ≤ dim M (ν) (1 − qb)
for all 0 ≤ q ≤ 1/b. Now, let dim∞ (ν) > d − 2 and t ∈ (0, 2 dim∞ (ν) /(d − 2)).
Thus we obtain dim∞ Jν,t(2/d−1)/2,1 = dim∞ (ν) + dt(2/d − 1)/2 > 0. Hence, the
third claim follows from the first part. The rest of the proof is devoted to prove
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 49
limt↓2 qJNν,t(2/d−1)/2,1 = qJNν,(2/d−1),1 . First, observe that, using 0 < t(d − 2)/2 < s <
dim∞ (ν), we have for n large
ν(C) ≤ 2−sn , C ∈ DnN .
Set a B 2/d − 1 and consider for fixed q ≥ 0,
P
D/N ν(QC ) (Λ(QC ) )
log q qa t/2
C∈D
t 7→ τJD/N (q) = lim sup n
,
ν,at/2,1
n→∞ log(2n )
where QC ∈ D(C) such that Jν,at/2,1 (C)q = ν(QC )q Λ(QC )qat/2 (the existence of
those cubes is ensured by Jν,at/2,1 (C) ≤ 2n(−s+(d−2)t/2 for C ∈ DnN ) . It follows that
t 7→ τJD/N
ν,at/2,1
(q) is an increasing, convex function on (0, 2 dim∞ (ν)/(d − 2)) as lim sup
of convex functions. In particular, we have for each q ≥ 0 that limt→2 τJD/N
ν,at/2,1
(q) =
τJD/N
ν,a,1
(q). By Lemma 3.5 we deduce that qJNν,at/2,1 is the unique zero of q 7→ τJD/N
ν,at/2,1
(q).
Hence, for fixed t ∈ (0, 2 dim∞ (ν)/(d−2)), we have that q 7→ τJD/N
ν,at/2,1
(q) is decreasing
and has a unique zero given by qJNν,at/2,1 . Now, Lemma 6.5 implies limt↓2 qJNν,at/2,1 =
qJNν,a,1 .
is equivalent to sup x∈Rd ,%>0 %(2−d)/2 ν (Q ∩ B(x, %))1/t in the sense that there exist
c1 , c2 > 0 only depending on d and t such that
c1C ≤ sup %(2−d)/2 ν (Q ∩ B(x, %))1/t ≤ c2C.
x∈Rd ,%>0
The result for the case d > 2 is a corollary of Adams’ Theorem on Riesz potentials
(see e. g. [Maz11, p. 67]) and the case d = 2 is due to Maz’ya and Preobrazenskii
and can be found in [Maz11, p. 83] or [MP84]. The following lemma establishes an
alternative representation of the best equivalent constant in terms of dyadic cubes.
Lemma 6.7. Let Q ∈ D and v a finite Borel measure on Q. Then, for a < 0 and
√ a √ db
b > 0 and C1 B 2 d , C2 B 3 d 2−a ,
√ Let
Proof. Q ∈ DnN . Since a < 0 we√assume with out√loss of generality that
% < d2−n+1 . Then for m ≥ n − 1 with d2−(m+1) < % ≤ d2−m , and x ∈ Rd ,
b
X
% ν (Q ∩ B(x, %)) ≤ 2
a b −a
ν(Q ∩ Q ) 2−ma
0
N ,Q0 ∩Q∩B(x,%),∅
Q0 ∈Dm
√ db
≤ 3 d 2−a max ν(Q ∩ Q0 )b Λ(Q0 )a/d
N
Q0 ∈Dm
√ d
where we used the fact that B (x, %) ∩ Q can be covered by at most 3 d elements
N and if Q0 ∩ Q , ∅, then Q0 ⊂ Q for m ≥ n, and
of Dm
max ν(Q ∩ Q0 )b Λ(Q0 )a/d ≤ ν(Q)b Λ(Q)a/d = max ν(Q ∩ Q0 )b Λ(Q0 )a/d .
N
Q0 ∈Dn−1 Q0 ∈DnN
N with Q0 ⊂ Q and % B
√ −m+1
On the other hand, for Q0 ∈ Dm d2 we find x ∈ Rd
such that Q ⊂ B (x, %). Then
0
ν(Q0 )b log Λ(Q0 ) ≤ ν (Q ∩ B (x, %))b dm log(2)
√
≤ dν (Q ∩ B (x, %))b (m + 1) log(2) + log d
= dν (Q ∩ B (x, %))b log(%)
≤ d sup log(%) ν (Q ∩ B(x, %))b .
x∈Rd ,%>0
Corollary
6.8.
For d ≥ 2, t > 2 there exists a constant Ct,d > 0 such that for all
u ∈ C∞ R d and for every Borel measure ν on Q and Q ∈ D, we have
c
1/2
log(r) ν(Q ∩ B(x, r)) kukH 1 (Rd ) , d = 2,
2/t
kukL2 (Rd ) ≤ kukLt (Rd ) ≤ c1 sup
ν|Q ν|Q
x∈Rd ,0<r<1/2
and
1/2
kukL2 (Rd ) ≤ kukLt (Rd ) ≤ c2 sup %(2−d) ν (Q ∩ B(x, %))2/t kukH 1 (Rd ) , d > 2.
ν|Q ν|Q d x∈R ,%>0
ForR every t > 2 there exists T t > 0 such that for all Q ∈ D and
Lemma 6.9.
u ∈ Cb Q with Q u dΛ = 0 we have
∞
Proof. Combining Lemma 2.1 and Corollary 6.8, we have for all u ∈ Cb∞ (Q)
52 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
kukL2 (Q) =
EQ (u)
L2 (Rd )
ν|Q ν|Q
=
ιRd ,ν (EQ (u))
L2 (Rd )
ν|Q
≤ Ct,d Jν,2/d−1,2/t (Q)1/2
EQ (u)
H 1 (Rd )
Z 2 1/2
Ct,d 1/2 kEQ k 1
· k∇ukL2 (Q) +
2
≤ Jν,2/d−1,2/t (Q) u dΛ
DQ DQ Λ Λ(Q) Q
|{z}
CT t
kEQ k
= T t Jν,2/d−1,2/t (Q)1/2 · k∇uk2L2 (Q) .
DQ Λ
Corollary 6.10. If (♠) holds, i. e. dim∞ (ν) > d − 2, then
sD ≤ sN ≤ hJν ≤ qJNν .
In particular, in the case d = 2, we have sN ≤ 1.
Proof. Note that dim∞ (ν) > d − 2 implies that for all t ∈ (2, dim∞ (ν) /(d − 2)),
Jν,2/d−1,2/t is non-negative, monotone and uniformly vanishing on D. Combining
Lemma 6.9, Proposition 6.1 and Proposition 6.6 we obtain sN ≤ hJν,t(2/d−1)/2,1 ≤
2 qJν,t(2/d−1)/2,1 , for all t ∈ (2, dim∞ (ν) /(d − 2)). The claim follows by letting t & 2
t N
7. Lower bounds
7.1. Lower bound on the spectral dimension. Recall, for n ∈ N and α > 0,
D/N
D/N D/N
n o
Nα,J (n) = card Mα,J (n) with Mα,J (n) B C ∈ DnD/N : J(C) ≥ 2−αn .
As before, for s > 0, we let hQi s denote the cube centred and parallel with respect
to Q such that Λ(Q) = s−d Λ hQi s , s > 0. In the following we always assume
dim∞ (ν) > d − 2.
Proposition 7.1. Assume the conditions of Theorem 1.2 are fulfilled. Then for fixed
α > 0 and for x > 0 large, we have
D n N n
Nα,J
Nα,J α,x α,x x
!
D
− 1 ≤ N (x) and −1≤ N N
,
5d 2 · 5d DQ
with nα,x B log2 (x) /α .
D/N
Proof. For n ∈ N large enough, i. e. Mα,J (n) , ∅, we construct by a finite
D/N
j D/N k
induction a subset En of Mα,J (n) of cardinality en B card (En ) ≥ Nα,J (n) /5d
D E D E
such that for all cubes Q, Q0 ∈ En with Q , Q0 we have Q̊ ∩ Q̊0 = ∅. At the
3 3
D/N
initial step of the induction we set D(0) B Mα,J (n). Assume we have constructed
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 53
D E
D(0) ⊃ D(1) ⊃ · · · ⊃ D( j−1) such that the following condition holds Q̊ j ∩ Q̊ , ∅,
5
for some Q, Q j ∈ D( j−1) with Q , Q j . Then we set
n D E o n o
D( j) B C ∈ D( j−1) : C̊ ∩ Q̊ j = ∅ ∪ Q j .
5
D E
By this construction, we have card D < card D( j−1) , since Q̊ ∩ Q̊ j , ∅. If
( j)
D E 5
Q̊ ∩ Q̊0 = ∅, for all Q, Q0 ∈ D( j−1) with Q , Q0 , then we set En = D( j−1) . In each
5
( j−1)
inductive step, we remove at most 5d − 1 elements j D/N of D k , while one element,
namely Q j , is kept. This implies card (En ) ≥ Nα,J (n) /5d .
Let us now first consider the Dirichlet case. Since for each Q ∈ DnD , we have
∂Q ∩ Q = ∅, it follows that hQi ˚ 3 ⊂ Q and therefore ψQ ∈ Cc∞ (Q̊). Now, with
nα,x B log2 (x) /α we have for each Q ∈ Enα,x that
R 2
∇ψQ dΛ 1
≤ ≤ 2αnα,x ≤ x.
ψQ dν
R
2 J(Q)
Hence, the ψQ : Q ∈ Enα,x C fi : i = 1, . . . , enα,x are mutually orthogonal both in
Lν2 and in H01 , and we obtain that span fi : i = 1, . . . , eα,x is an enα,x -dimensional
Pen ?
subspace of H01 . For every (c1 , . . . , cenα,x ) ∈ Renα,x with i=1α,x ci fi ∈ H01 , we
obtain
enα,x Penα,x R
X
i=1 ci
2
|∇ fi |2 dΛ
RH 1 ci fi ≤ P enα,x 2 R ≤ x.
0
i=1 i=1 ci fi2 dν
Hence, we deduce from the min-max principle from Proposition 2.9
D
nα,x /5d − 1 ≤ enα,x ≤ N D (x) .
Nα,J
In the Neumann case, we proceed similarly. For fixed α > 0 set nα,x = log2 (x) /α
and write En = E1 , · · · , Ecard(En ) . For each i = 1, . . . , ben /2c C N we define
fi B RQ̊ a2i−1 ψE2i−1 + a2i ψE2i ∈ C∞ b Q ,
D E
where we choose (a2i−1 , a2i ) ∈ R2 \ {(0, 0)} such that Q fi dΛ = 0. Since E̊ j ∩
R
D E 3
E˚k = ∅ for j , k and the properties of mediants, we obtain
3
7.2. Lower bound on the embedding constant. In the following we assume (♠),
that is dim∞ (ν) > d − 2. We need a slight modification of Jν for the case d = 2.
We define Jν (Q) = supQ∈D(Q) ν(Q)Λ(Q)2/d−1 for Q ∈ D. Hence, in the case d = 2,
we have Jν (Q) = ν(Q). Clearly, we again have dim∞ (Jν ) > d − 2, τJD/N
ν
= τJD/N by
ν
D/N D/N
Proposition 3.15 and for d > 2, F JD/N = F JD/N
ν
and FJ = F Jν . The case d = 2 is
ν ν
covered by the following lemma.
Lemma 7.2. In the case d = 2, we have
D/N D/N
F JD/N = F JD/N
ν
and F J = F Jν .
ν ν
and, using dim∞ (ν) > d − 2, we obtain for every δ > 1 and n ∈ N large enough
This leads to
D/N n D/N
o
ν(Q 0 0 −αn
ν(C) −αδn
.
C ∈ D : sup )| log(Λ(Q ))| ≥ 2 ⊂ C ∈ D : ≥ 2
n
n
Q0 ∈D(C)
Proposition 7.3. There exists a constant K > 0 such that every Q ∈D DE with
Jν (Q) > 0 there exists a function ψQ ∈ C∞ d
c (R ) with support contained in Q̊ 3 and
ϕQ
L2 > 0 such that
ν
2
2
ψQ
L2 ≥ KJ (Q)
∇ψQ
L2 Rd .
ν ν ( )Λ
Proof. Since dim∞ (Jν ) > 0, it follows that for each Q ∈ D there exists C Q ∈ D(Q)
such that Jν (Q) = ν(C Q )Λ(C Q )2/d−1 . Now, choose ψQ B ϕhC Q i ,3 as in Lemma
3
2.13. Then ψQ · 1C Q = 1C Q , supp(ψQ ) ⊂ C˚Q ⊂ Q̊ and
D E D E
3 3
D
E 1−2/d
Λ
2
CQ 3
R
∇ψQ dΛ −2 1−2/d 1/3
R 2 ≤ C2 3 D
E
ψQ dν ν CQ 3
1/3
1−2/d
Λ CQ
= C2−2 31−2/d = C2−2 31−2/d Jν (Q).
ν CQ
Proposition 7.4. For fixed α > 0 and for x > 0 large, we have
D (n )
Nα,J α,x
ν
− 1 ≤ N D (xK)
5d
D
with nα,x B log2 (x) /α . In particular, F JDν ≤ sD and F Jν ≤ sD .
Proof. This follows from Proposition 7.3, Lemma 7.1 and Lemma 7.2.
In the same way we obtain the following proposition for the Neumann case.
Proof of Theorem 1.3. The first inequalities follow from Proposition 8.1, Proposi-
tion 7.5 and Proposition 7.4. Furthermore, by Proposition 3.15, we always in the
case d = 2 that qJNν = 1. For the last claim note that ν(Q̊) > 0, implies that there
exists an open cube Q ⊂ Q̊ with ν(Q) > 0 and dim∞ (ν|Q ) > d − 2 = 0. Hence, we
obtain
N D D
1 = qJNν| ≤ F Jν| = F Jν| ≤ F Jν ≤ sD ≤ sN = 1.
Q Q Q
Set τN\D (q) B lim supn 1/ log 2n log Jν (Q)q . Then for q ≥ 0
P
Q∈DnN \DnD
9. Examples
In this last section we present some example illuminating the critical case with
dim∞ (ν) = d − 2 and the possibility of non-existing spectral dimension.
9.1. Critical cases. We give three examples of measures νi in dimension d = 3 for
the critical case, i. e. dim∞ (νi ) = d − 2 = 1, i = 1, 2, 3. In the first example the
Kreı̆n–Feller operator exists but has no compact resolvent and we therefore have
no orthonormal basis of eigenvectors. In the second example the operator we has a
compact resolvent and we are able to determine the spectral dimension sD/N = 3/2.
In the third example the operator cannot be defined via our form approach as there
is no continuous embedding of the Sobolev space into Lν2 .
58 SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS
H 1 (R) ,→ LΛ
6
(R) ,→ Lν2i (R) , (9.2)
where the embedding constant C1 of the first embedding is independent of R (see
[Ada75, Lemma 5.10]).
Example 9.1. For i = 1 the embedding is continuous but not compact. We ob-
R −n 2/3
2
Rn L3/2 =
−1 = log 2 2/3 . This observation and (9.1)
serve that
f |
2−n−1
z dΛ
Λ
combined give
Z XZ
f |
u|
2 = log 22/3
u|
2
X
X
|u|2 dν1 = |u|2 dν1 ≤ Rn L3/2 Rn L 6 Rn L6
Rn Λ Λ Λ
n n n
u|
2 = log 22/3 C kuk2
X
≤ log 2 2/3 C1
Rn H 1 1 H1
n
Now consider the limes superior as m, n are tending to infinity and then let ` → ∞.
This proves that (un ) is Cauchy in Lν22 and convergent there as well. This shows
that the embedding H 1 ,→ Lν22 is compact. To finally determine the
spectral
dimension in this case, in view of Theorem 1.1 we choose J (Q) B
f |Q
L3/2
Λ
with Q ∈ D and prove κJ ≤ 3/2: For every cube Q ∈ DnN for n > 1 lying
in Rn,k B (x, y, z) ∈ Q : k2−n < z < (k + 1) 2−n with 2n−1 ≥ k ≥ 1 (these are
(k + 1)2 k2 many with ν(Q) > 0), we have
3/2 Z −3 Z (k+1)2−n
−2
f |Q
3/2 = −2n
z−3 − log z −2 dz
L
z − log z dΛ ≤ 2
Λ Q k2−n
Z (k+1)2−n
−2
≤ 2−2n k2−n z−1 − log z −2 dz
k2−n
!
1 1 1
= 2 −
k − log(k) + log(2n ) − log(k + 1) + log(2n )
1
≤ 3 .
k log(2 /(k + 1))2
n
X
q 2q
! X 2n−1
X −1 card Q ∈ DN : Q ∩ R
n n,k , ∅, ν(Q) > 0 k−2q
f |
Q L3/2 ≤ ζ +
Q∈D
Λ 3 n∈N k=1
log(2n /(k + 1))4q/3
2q
!
1 X 2n−1 X−1 k−2q+2
≤ζ + +
3 log(2n /(2))q4/3 n∈N k=1 log(2n /(k + 1))4q/3
>1
! X
2q X 1
=ζ + k−2q+2
3 (n log(2) − log(k + 1))4q/3
k∈N n≥dlog2 (k)+1e+1
! X
2q X 1
≤ζ + k−2q+2
3 (n − log2 (k) + 1 )4q/3 log 2 4q/3
k∈N n≥dlog2 (k)+1e+1
where we used log2 (k + 1) ≤ log2 (k) + 1 for all k ≥ 1 and ζ denotes the Riemann
ζ-function. Since for all q > 3/2 the right-hand side is finite, we find sD ≤ sN ≤
sN ≤ κJ ≤ 3/2 as a consequence of Theorem 1.1 with regard to (9.3). Moreover,
SPECTRAL DIMENSIONS OF KREĬN–FELLER OPERATORS 61
note that there exists an open sub-cube Q ⊂ Q such that Q ⊂ Q̊ with ν(Q) > 0. Since
f |Q is bounded, Lemma 2.11 gives 3/2 = sνD2 |Q ≤ sD and we have 3/2 = sD = sN .
References
[Ada75] R. A. Adams. Sobolev spaces. Pure and Applied Mathematics, Vol. 65.
Academic Press [A subsidiary of Harcourt Brace Jovanovich, Publish-
ers], New York-London, 1975, pp. xviii+268.
[Arz14] P. Arzt. Eigenvalues of Measure Theoretic Laplacians on Cantor-like
Sets. PhD thesis, Universität Siegen, 2014.
[Arz15] P. Arzt. Measure theoretic trigonometric functions. J. Fractal Geom. 2.2
(2015), 115–169. doi: 10.4171/JFG/18.
[BC86] J. Brossard and R. Carmona. Can one hear the dimension of a fractal?
Comm. Math. Phys. 104.1 (1986), 103–122.
[Ber79] M. V. Berry. Distribution of modes in fractal resonators. Structural
stability in physics (Proc. Internat. Symposia Appl. Catastrophe Theory
and Topological Concepts in Phys., Inst. Inform. Sci., Univ. Tübingen,
Tübingen, 1978). Vol. 4. Springer Ser. Synergetics. Springer, Berlin,
1979, pp. 51–53. doi: 10.1007/978-3-642-67363-4_7.
[Ber80] M. V. Berry. Some geometric aspects of wave motion: wavefront dislo-
cations, diffraction catastrophes, diffractals. Geometry of the Laplace
operator (Proc. Sympos. Pure Math., Univ. Hawaii, Honolulu, Hawaii,
1979). Proc. Sympos. Pure Math., XXXVI. Amer. Math. Soc., Provi-
dence, R.I., 1980, pp. 13–28.
[Bor71] V. V. Borzov. “Quantitative characteristics of singular measures”. Spec-
tral Theory and Wave Processes. Ed. by M. S. Birman. Boston, MA:
Springer US, 1971, pp. 37–42. doi: 10.1007/978- 1- 4684- 8926-
2_5.
[BS67] M. v. Birman and M. Z. Solomjak. Piecewise polynomial approxima-
tions of functions of classes W p α . Mat. Sb. (N.S.) 73 (115) (1967), 331–
355.
[BS70] M. v. Birman and M. Z. Solomjak. The principal term of the spectral
asymptotics for ”non-smooth” elliptic problems. Funkcional. Anal. i
Priložen. 4.4 (1970), 1–13.
[Dav95] E. B. Davies. Spectral Theory and Differential Operators. Cambridge
Studies in Advanced Mathematics. Cambridge University Press, 1995.
doi: 10.1017/CBO9780511623721.
[DeV87] R. A. DeVore. “A note on adaptive approximation”. Proceedings of
China-U.S. Joint Conference on Approximation Theory (Hangzhou,
1985). Vol. 3. 4. 1987, pp. 74–78.
[DKS20] O. Davydov, O. Kozynenko, and D. Skorokhodov. Optimal approxima-
tion order of piecewise constants on convex partitions. J. Complexity 58
(2020), 101444, 14. doi: 10.1016/j.jco.2019.101444.
[DN15] D.-W. Deng and S.-M. Ngai. Eigenvalue estimates for Laplacians on
measure spaces. J. Funct. Anal. 268.8 (2015), 2231–2260. doi: 10 .
1016/j.jfa.2014.12.019.
REFERENCES 63