betaFeedback

WorksMathematics

Hereditary lowerability of diagonal entropy scales

Abstract

Let TT be a continuous self-map of a compact metric space. For each integer p≥2p\geq2, we consider the upper capacity entropy scale obtained by dividing the pp-fold logarithm of an orbit covering number by the (p−1)(p-1)-fold logarithm of time. We prove that every nonempty compact set contains a compact subset realizing any value between zero and its own scale entropy. Every strictly intermediate value is realized by a countable compact set with exactly one accumulation point. The proof combines localization at a fixed spatial resolution, a finite-tree selection lemma, and refinements of orbit covers at dyadic times. An example shows that the assertion about one accumulation point does not extend to the upper endpoint. On the product-diagonal construction used in classical entropy lowerability, the diagonal scales preserve the values of the base subsets, although classical upper capacity entropy takes only zero or infinity.

Introduction

The entropy of a compact subset describes the growth of distinguishable orbit segments starting in that set. The set need not be invariant. This leads to an intermediate-value question: if a compact set has a given entropy, which smaller values occur on its compact subsets? The answer depends both on the growth normalization and on the definition of subset entropy.

The open-cover definition of topological entropy was introduced by Adler, Konheim, and McAndrew [1]. Bowen’s metric formulation [2] describes the same invariant on compact systems through orbit separation. For arbitrary subsets one can retain a common orbit length in every covering ball, obtaining upper capacity entropy, or allow the orbit lengths to vary, as in Bowen’s definition for noncompact sets [3]. These constructions can assign different values even to compact sets that are not invariant.

Huang, Ye, and Zhang [7] formulated lowerability and hereditary lowerability for upper capacity entropy. Lowerability asks for all values between zero and the entropy of the whole system; hereditary lowerability asks the same question inside each nonempty compact subset. They proved lowerability for systems of finite topological entropy. Their subsequent work [8] established lowerability without the finiteness assumption and constructed systems that are not hereditarily lowerable. In particular, their product-diagonal example contains a compact set whose compact subsets have classical upper capacity entropy only zero or infinity. In a recent preprint, Wang [10] proves hereditary lowerability for finite-entropy homeomorphisms. Thus the finite-entropy and infinite-entropy settings have different behavior for the classical normalization.

To distinguish orbit growth below the exponential scale, one can replace the classical normalization by slower growth functions. Katok and Thouvenot [9] introduced slow-entropy invariants in measure-preserving dynamics. In topological dynamics, de Carvalho [4] and Dou, Huang, and Park [5] developed entropy dimension using growth rates between polynomial and exponential ones. More generally, Helfter [6] organized metric covering invariants by families of gauge functions, including families obtained from iterated exponentials and logarithms. Xie, Chen, and Zhou [11] formulate Bowen and upper capacity

2020 Mathematics Subject Classification. Primary 37B40; Secondary 37B10, 54H20.

Key words and phrases. Upper capacity entropy, entropy scale, hereditary lowerability, entropy dimension, compact subset. entropy scales in dynamics and study their comparison and variational properties. These results motivate asking how lowerability behaves for a specified family of orbit-growth scales.

Here the time convention is fixed as t=e−nt=e^{-n}. If rn(E,ε)r_n(E,\varepsilon) denotes the least number of radius-ε\varepsilon Bowen balls of length nn covering EE, the diagonal member of the resulting hierarchy is

Dp(T,E)=lim⁡ε↓0lim sup⁡n→∞ℓp(rn(E,ε))Lp−1(n),p≥2.(1)D_p(T,E)=\lim_{\varepsilon\downarrow0}\limsup_{n\to\infty}\frac{\ell_p(r_n(E,\varepsilon))}{L_{p-1}(n)},\qquad p\ge2. \tag*{(1)}

Here LkL_k is the kk-fold logarithm, defined for sufficiently large arguments, and ℓk\ell_k is a regularized version defined on [0,∞)[0,\infty); both are specified in § 2. The regularization makes (1.1) meaningful when covering numbers stay bounded. For p=2p=2 and E=XE=X, this is upper entropy dimension; for general EE it is the corresponding upper capacity extension to subsets. At every fixed resolution, orbit covering numbers are at most exponential in nn, so 0≤Dp(T,E)≤10\le D_p(T,E)\le1.

Theorem 1.1. Let XX be a nonempty compact metric space and let T:X→XT:X\to X be continuous. Fix an integer p≥2p\ge2. For every nonempty compact set E⊂XE\subset X and every a∈[0,Dp(T,E)]a\in[0,D_p(T,E)], there is a nonempty compact set Ka⊂EK_a\subset E such that

Dp(T,Ka)=a.D_p(T,K_a)=a.

If 0<a<Dp(T,E)0<a<D_p(T,E), the set KaK_a can be chosen countable with exactly one accumulation point.

Consequently every compact continuous system is hereditarily lowerable for each fixed diagonal scale in (1.1). Neither EE nor KaK_a is required to be invariant. No finiteness assumption on classical topological entropy is imposed, and invertibility is not used. The statement is separate for each pp; it does not prescribe several scale values on the same subset.

The construction follows the extraction strategy for countable compact sets in [10 Lemma 3.5], using metric covering trees and the diagonal scale estimates developed below. First, fixed-resolution localization produces a countable compact source set with one accumulation point and scale entropy larger than the target. Second, finite subsets of successively smaller tails are selected so that their orbit counts satisfy one common upper bound at all times. For this step we use Wang’s finite-tree lemma [10 Lemma 3.3]. The trees here come from successive metric covers at dyadic times. Their node counts include a logarithmic loss, which the parameter gap in the diagonal scales absorbs. A recurring schedule of spatial resolutions supplies the matching lower bound.

The restriction to strictly intermediate values in the assertion about accumulation points is necessary. § 6 gives, for every fixed p≥2p\ge2 and H∈(0,1]H\in(0,1], a compact set with Dp=HD_p=H for which every compact subset with at most one accumulation point has Dp<HD_p<H. The same section computes DpD_p on the product-diagonal example of [8]: the classical zero-or-infinity restriction coexists with preservation of the base subsets’ diagonal scale values.

§ 2 records the definitions, covering facts, and finite combinatorial terminology used later. § 3 establishes the scale estimates. § 4 and § 5 give the localization and the proof of Theorem 1.1.

Preliminaries

Orbit metrics and subset entropy

2.1. Orbit metrics and subset entropy. Let (X,d)(X,d) be a nonempty compact metric space and let T:X→XT:X\to X be continuous. We use N={1,2,…}\mathbb{N}=\{1,2,\ldots\} and N0={0,1,…}\mathbb{N}_0=\{0,1,\ldots\}, and all logarithms are natural unless a base is displayed. For n∈Nn\in\mathbb{N}, set

dn(x,y)=max⁡0≤j<nd(Tjx,Tjy),Bn(x,ε)={y∈X:dn(x,y)<ε}.d_n(x,y)=\max_{0\le j<n}d(T^j x,T^j y),\qquad B_n(x,\varepsilon)=\{y\in X:d_n(x,y)<\varepsilon\}.

For a nonempty set E⊂XE \subset X, let rn(E,ε)r_n(E,\varepsilon) be the least number of such balls covering EE, with centers allowed in XX. Let sn(E,ε)s_n(E,\varepsilon) be the largest cardinality of a subset F⊂EF \subset E satisfying dn(x,y)>εd_n(x,y)>\varepsilon for distinct x,y∈Fx,y\in F. Such a set is called (n,ε)(n,\varepsilon)-separated. The metric dnd_n induces the topology of XX, so compactness makes these counts finite. The largest separated cardinality exists even if EE is not closed: possible cardinalities are integers in a nonempty finite set. Both counts are nondecreasing in nn.

The standard covering–separation inequalities, in the present open-ball convention, are

rn(E,2ε)≤sn(E,ε)≤rn(E,ε/2).(2)r_n(E,2\varepsilon)\le s_n(E,\varepsilon)\le r_n(E,\varepsilon/2). \tag*{(2)}

Indeed, a separated set of largest cardinality is maximal under inclusion, so every point of EE is within dnd_n-distance at most ε\varepsilon of it. Also, an open ball of radius ε/2\varepsilon/2 contains at most one point of an (n,ε)(n,\varepsilon)-separated set. See [8 Section 2] for the customary covering and separation descriptions of subset entropy.

Classical upper capacity entropy is

h(T,E)=lim⁡ε↓0lim sup⁡n→∞log⁡rn(E,ε)n.(3)h(T,E)=\lim_{\varepsilon\downarrow0}\limsup_{n\to\infty}\frac{\log r_n(E,\varepsilon)}{n}. \tag*{(3)}

For E=XE=X this is topological entropy. The counts in (3) use one common orbit length at each stage. Bowen’s entropy for noncompact sets instead uses weighted covers whose orbit lengths may vary from ball to ball [3]. All lowerability assertions below concern upper capacity counts.

Iterated functions and the time convention

Set

L0(u)=u,Lk+1(u)=log⁡Lk(u),E0(u)=u,Ek+1(u)=exp⁡(Ek(u)).L_0(u)=u,\qquad L_{k+1}(u)=\log L_k(u),\qquad E_0(u)=u,\qquad E_{k+1}(u)=\exp(E_k(u)).

An expression containing LkL_k is used only where its successive logarithms are defined, and in limits its argument is sufficiently large that they are positive. Thus L1(u)=log⁡uL_1(u)=\log u, L2(u)=log⁡log⁡uL_2(u)=\log\log u, and E2(u)=exp⁡(eu)E_2(u)=\exp(e^u). The identities Lk(Ek(u))=uL_k(E_k(u))=u and Ek(Lk(u))=uE_k(L_k(u))=u hold on these domains. For counts that may equal one, use

ℓ0(u)=u,ℓk+1(u)=log⁡(e+ℓk(u)),u≥0.(4)\ell_0(u)=u,\qquad\ell_{k+1}(u)=\log(e+\ell_k(u)),\qquad u\ge0. \tag*{(4)}

Every ℓk\ell_k is increasing, and

ℓk(u)=Lk(u)+o(1)(u→∞),k≥1.(5)\ell_k(u)=L_k(u)+o(1)\qquad(u\to\infty),\qquad k\ge1. \tag*{(5)}

For k=1k=1, the difference is log⁡(1+e/u)\log(1+e/u). If ℓk(u)=Lk(u)+o(1)\ell_k(u)=L_k(u)+o(1), the next difference is log⁡(1+(e+o(1))/Lk(u))\log(1+(e+o(1))/L_k(u)), proving (5) by induction.

In asymptotic notation, F=O(G)F=O(G) means ∣F∣≤C∣G∣\lvert F\rvert\le C\lvert G\rvert eventually for some constant CC, and F=o(G)F=o(G) means F/G→0F/G\to0. A subscript on OO, such as Oε(1)O_\varepsilon(1), permits the constant to depend on that subscript. Limits in time are always taken before the spatial resolution tends to zero.

For positive integers p,qp,q and α>0\alpha>0, the iterated scale family of [6 Proposition 2.4] can, for sufficiently small t>0t>0, be written as

scl⁡αp,q(t)=1Ep(αLq(t−1)).(6)\operatorname{scl}_{\alpha}^{p,q}(t)=\frac{1}{E_p\left(\alpha L_q(t^{-1})\right)}. \tag*{(6)}

Only its behavior as t↓0t\downarrow0 is needed. Under the convention t=e−nt=e^{-n}, its reciprocal becomes

Aαp,q(n)=Ep(αLq−1(n)).(7)A_{\alpha}^{p,q}(n)=E_p\left(\alpha L_{q-1}(n)\right). \tag*{(7)}

We define

hp,q(T,E)=lim⁡ε↓0inf⁡{α>0:rn(E,ε)Aαp,q(n)⟶0}.(8)h^{p,q}(T,E)=\lim_{\varepsilon\downarrow0}\inf\left\{\alpha>0:\frac{r_n(E,\varepsilon)}{A_{\alpha}^{p,q}(n)}\longrightarrow0\right\}. \tag*{(8)}

Here inf⁡∅=+∞\inf\varnothing=+\infty. This explicitly fixes the normalization; changing the substitution for tt would give a different time-growth question. The notation used in the rest of the paper is Dp(T,E)=hp,p(T,E)D_p(T,E)=h^{p,p}(T,E) for p≥2p\ge2.

Lemma 2.1. For every nonempty E⊂XE \subset X and all p,q≥1p,q \ge1,

hp,q(T,E)=lim⁡ε↓0lim sup⁡n→∞ℓp(rn(E,ε))Lq−1(n).(9)h^{p,q}(T,E)=\lim_{\varepsilon\downarrow0}\limsup_{n\to\infty}\frac{\ell_p(r_n(E,\varepsilon))}{L_{q-1}(n)}. \tag*{(9)}

The same outer limit is obtained with sn(E,ε)s_n(E,\varepsilon) in place of rn(E,ε)r_n(E,\varepsilon). These quantities are independent of the compatible metric on XX. In particular, h1,1(T,E)=h(T,E)h^{1,1}(T,E)=h(T,E).

Proof. Fix ε\varepsilon and put

Rn=rn(E,ε),vn=Lq−1(n),β=lim sup⁡nℓp(Rn)vn.R_n=r_n(E,\varepsilon),\qquad v_n=L_{q-1}(n),\qquad\beta=\limsup_n\frac{\ell_p(R_n)}{v_n}.

If α>β\alpha>\beta, choose β<b′<b<α\beta<b'<b<\alpha. Eventually ℓp(Rn)<b′vn<ℓp(Ep(bvn))\ell_p(R_n)<b'v_n<\ell_p(\mathcal{E}_p(bv_n)), by (2.4). Monotonicity gives Rn≤Ep(bvn)R_n\le\mathcal{E}_p(bv_n), and

Ep(bvn)Ep(αvn)⟶0.\frac{\mathcal{E}_p(bv_n)}{\mathcal{E}_p(\alpha v_n)}\longrightarrow0.

For p=1p=1 this is e−(α−b)vne^{-(\alpha-b)v_n}; for p>1p>1 its logarithm is Ep−1(bvn)−Ep−1(αvn)→−∞\mathcal{E}_{p-1}(bv_n)-\mathcal{E}_{p-1}(\alpha v_n)\to-\infty. Conversely, if 0<α<β0<\alpha<\beta, choose α<b<b′<β\alpha<b<b'<\beta. On an infinite subsequence, ℓp(Rn)>b′vn>ℓp(Ep(bvn))\ell_p(R_n)>b'v_n>\ell_p(\mathcal{E}_p(bv_n)). Hence Rn≥Ep(bvn)R_n\ge\mathcal{E}_p(bv_n) on that subsequence, and the ratio with Ep(αvn)\mathcal{E}_p(\alpha v_n) tends to infinity there. These two arguments also cover β=0\beta=0 and β=+∞\beta=+\infty by the available one-sided inequality. Bounded sequences RnR_n are included in the first argument. This proves the critical-parameter formula at fixed ε\varepsilon.

Inequalities (2.1) give the separated-count version after ε↓0\varepsilon\downarrow0. If d′d' is another compatible metric, compactness gives, for every ε>0\varepsilon>0, a δ>0\delta>0 such that d(x,y)<δd(x,y)<\delta implies d′(x,y)<εd'(x,y)<\varepsilon. The same implication holds for the corresponding orbit metrics, whence rn(E,d′,ε)≤rn(E,d,δ)r_n(E,d',\varepsilon)\le r_n(E,d,\delta). Interchanging the two metrics proves metric independence. Finally, 0≤ℓ1(R)−log⁡R≤log⁡(1+e)0\le\ell_1(R)-\log R\le\log(1+e) for R≥1R\ge1; division by nn gives the last assertion.

For b>0b>0, two examples of numerical growth rates are

Rn=exp⁡(nb),ℓ2(Rn)log⁡n⟶b,R_n=\exp(n^b),\qquad\frac{\ell_2(R_n)}{\log n}\longrightarrow b,

and

Rn=exp⁡(exp⁡((log⁡n)b)),ℓ3(Rn)log⁡log⁡n⟶b.R_n=\exp(\exp((\log n)^b)),\qquad\frac{\ell_3(R_n)}{\log\log n}\longrightarrow b.

Bounded counts have rate zero. Lemma 2.3 bounds the rates that can occur dynamically.

In the case p=q=2p=q=2, one also has

D2(T,E)=lim⁡ε↓0inf⁡{α>0:lim sup⁡n→∞log⁡rn(E,ε)nα=0}.(10)D_2(T,E)=\lim_{\varepsilon\downarrow0}\inf\left\{\alpha>0:\limsup_{n\to\infty}\frac{\log r_n(E,\varepsilon)}{n^\alpha}=0\right\}. \tag*{(10)}

To check this, fix ε\varepsilon and let β\beta be the limsup in (9) for p=q=2p=q=2. If β<b<α\beta<b<\alpha, then log⁡rn(E,ε)≤nb\log r_n(E,\varepsilon)\le n^b eventually. If α<b<β\alpha<b<\beta, then log⁡rn(E,ε)≥nb/2\log r_n(E,\varepsilon)\ge n^{b/2} on an infinite subsequence. These assertions follow by undoing the two regularized logarithms; the bounded difference between ℓ1(R)\ell_1(R) and log⁡R\log R does not change either estimate. Division by nαn^\alpha proves (2.9), with the same one-sided argument at β=0\beta=0. For E=XE=X, this is the upper entropy dimension considered in [4, 5]; equation (2.9) specifies the extension used here when EE is not invariant.

For the estimates used later, fix p≥2p\ge2, put r=p−1r=p-1, and define, for b>0b>0 and sufficiently large real tt,

fb(t)=Er(bLr(t)),Gb(t)=exp⁡(fb(t)).(11)f_b(t)=\mathcal{E}_r(bL_r(t)),\qquad G_b(t)=\exp(f_b(t)). \tag*{(11)}

Thus Gb=Abp,pG_b=A_b^{p,p}, and f1(t)=tf_1(t)=t. For p=2p=2 one has fb(t)=tbf_b(t)=t^b and Gb(t)=etbG_b(t)=e^{t^b}; for p=3p=3 one has fb(t)=exp⁡((log⁡t)b)f_b(t)=\exp((\log t)^b). Thresholds will always be chosen with Lr(t)>0L_r(t)>0, on which domain both functions increase in tt and bb.

Finite unions, localization rates, and the range of values

Fix p≥2p \ge2. To retain a spatial resolution in the notation, put

dε(E)=lim sup⁡n→∞ℓp(sn(E,ε))Lp−1(n).(12)d_{\varepsilon}(E) = \limsup_{n \to\infty} \frac{\ell_p(s_n(E,\varepsilon))}{L_{p-1}(n)}. \tag*{(12)}

Thus Dp(T,E)=lim⁡ε↓0dε(E)D_p(T,E) = \lim_{\varepsilon\downarrow0} d_{\varepsilon}(E). The symbol dε(E)d_{\varepsilon}(E) denotes a rate, whereas dn(x,y)d_n(x,y) denotes an orbit distance. The map TT and the fixed integer pp are suppressed in (2.11).

Lemma 2.2. The quantities dε(E)d_{\varepsilon}(E) and Dp(T,E)D_p(T,E) are monotone under inclusion. For a finite family of nonempty subsets E1,…,EmE_1,\ldots,E_m of XX,

dε(⋃i=1mEi)=max⁡idε(Ei),(13)d_{\varepsilon}\left(\bigcup_{i=1}^{m} E_i\right) = \max_i d_{\varepsilon}(E_i), \tag*{(13)}
Dp(T,⋃i=1mEi)=max⁡iDp(T,Ei).(14)D_p\left(T,\bigcup_{i=1}^{m} E_i\right) = \max_i D_p(T,E_i). \tag*{(14)}

Finite nonempty sets have value zero. Deleting finitely many points from an infinite set changes neither rate.

Proof. Let Sn=max⁡isn(Ei,ε)S_n = \max_i s_n(E_i,\varepsilon). Assign each point of a separated set in the union to one containing EiE_i. The assigned points in each EiE_i are separated, so

Sn≤sn(⋃iEi,ε)≤∑isn(Ei,ε)≤mSn.S_n \le s_n\left(\bigcup_i E_i,\varepsilon\right) \le\sum_i s_n(E_i,\varepsilon) \le mS_n.

For fixed mm and pp, the difference ℓp(mu)−ℓp(u)\ell_p(mu)-\ell_p(u) is bounded on u≥1u \ge1. For p=1p=1 it is at most log⁡m\log m, and induction follows because u↦log⁡(e+u)u \mapsto\log(e+u) has derivative at most 1/e1/e. Division by Lp−1(n)→∞L_{p-1}(n) \to\infty removes this bounded difference. The limsup of a maximum of finitely many sequences is the maximum of their limsups, proving (2.12). Taking the monotone resolution limit proves (2.13). A finite set has uniformly bounded counts. Writing a set as the union of its remaining points and its deleted finite part proves the deletion statement. □

Lemma 2.3. For every ε>0\varepsilon> 0 there is an integer Mε≥1M_{\varepsilon} \ge1 such that

rn(E,ε)≤rn(X,ε)≤Mεn(n∈N, ∅≠E⊂X).(15)r_n(E,\varepsilon) \le r_n(X,\varepsilon) \le M_{\varepsilon}^{n} \qquad(n \in\mathbb{N},\ \varnothing\ne E \subset X). \tag*{(15)}

Consequently 0≤Dp(T,E)≤10 \le D_p(T,E) \le1 for p≥2p \ge2, and hp,q(T,E)=0h^{p,q}(T,E)=0 if p>qp>q. If h(T,E)>0h(T,E)>0, then Dp(T,E)=1D_p(T,E)=1 for every p≥2p \ge2.

Proof. Cover XX by MεM_{\varepsilon} open sets of diameter less than ε\varepsilon. Each nonempty intersection Ui0∩T−1Ui1∩⋯∩T−(n−1)Uin−1U_{i_0} \cap T^{-1}U_{i_1} \cap\cdots\cap T^{-(n-1)}U_{i_{n-1}} has dnd_n-diameter less than ε\varepsilon. There are at most MεnM_{\varepsilon}^{n} such intersections, and one point from each supplies an open Bowen-ball cover. This proves (2.14). For M>1M>1 and p≥2p \ge2,

ℓp(Mn)Lp−1(n)⟶1.\frac{\ell_p(M^n)}{L_{p-1}(n)} \longrightarrow1.

If M=1M=1 the limit is zero. This gives the bound on DpD_p; when p>qp>q, use additionally Lp−1(n)/Lq−1(n)→0L_{p-1}(n)/L_{q-1}(n) \to0. If h(T,E)>0h(T,E)>0, some fixed ε\varepsilon and c>0c>0 satisfy rn(E,ε)≥ecnr_n(E,\varepsilon) \ge e^{cn} for infinitely many nn. The corresponding normalized pp-fold logarithms tend to one on that subsequence. The upper bound just established gives equality. □

Definition 2.4. For a fixed p≥2p \ge2, (X,T)(X,T) is lowerable for DpD_p if each a∈[0,Dp(T,X)]a \in[0,D_p(T,X)] occurs on a nonempty compact subset of XX. It is hereditarily lowerable for DpD_p if the same assertion holds inside every nonempty compact subset EE, with upper endpoint Dp(T,E)D_p(T,E).

These definitions follow the subset formulation of [7, 8]. An accumulation point of a set is a point every neighborhood of which contains a different point of that set. A model for the compact sets used below is {0}∪{1/j:j∈N}⊂R\{0\} \cup\{1/j : j \in\mathbb{N}\} \subset\mathbb{R}: it is countable, and zero is its only accumulation point. On a compact metric space an infinite set has an accumulation point, so a compact set with none is finite.

Finite partitions and rooted trees

A finite partition of a finite set PP is a collection of nonempty, pairwise disjoint subsets, called atoms, whose union is PP. The join P∨Q\mathcal{P} \vee\mathcal{Q} consists of all nonempty intersections of an atom of P\mathcal{P} with an atom of Q\mathcal{Q}. It refines both partitions: each new atom is contained in one atom of either partition. For instance, the join of {{1,2},{3,4}}\{\{1, 2\}, \{3, 4\}\} and {{1,3},{2,4}}\{\{1, 3\}, \{2, 4\}\} form the singleton partition of {1,2,3,4}\{1, 2, 3, 4\}.

A finite rooted tree is a finite connected graph with no cycles and with a distinguished vertex, its root. A vertex has level kk if the path from it to the root has kk edges. Its parent is the preceding vertex on that path. A leaf is a vertex with no children. We consider trees whose leaves all lie at the same level JJ, and call these vertices terminal leaves. A vertex is occupied by a chosen set of leaves if at least one chosen leaf descends from it. For example, a binary tree of depth two has leaves 00, 01, 10, 11. Choosing 00 and 01 occupies one vertex at level one and two at level two; choosing 00 and 10 occupies two vertices at each of these levels.

A sequence of refining partitions of PP gives such a tree: atoms are vertices, inclusion determines parenthood, and singleton atoms in the last partition are leaves. Controlling the number of occupied vertices therefore controls how many partition atoms are needed to cover a selected subset of PP.

Lemma 2.5 (Wang, [10 Lemma 3.3]). Let a finite rooted tree have AA leaves at a common level JJ, and let BkB_k be the number of vertices at level kk having a terminal descendant. For every integer 1≤M≤A1 \le M \le A, one can select exactly MM leaves such that the number bkb_k of occupied vertices satisfies

bk≤2k2(1+MBkA),1≤k≤J.(16)b_k \le2k^2\left(1 + \frac{M B_k}{A}\right), \qquad1 \le k \le J. \tag*{(16)}

We use this finite combinatorial result in the stated form. Its proof in [10] applies to arbitrary finite rooted trees and requires no dynamical assumptions. In particular, its loss depends on the current level kk, not on the final depth JJ.

Symbolic and product models

The two-sided full shift is Σ2={0,1}Z\Sigma_2 = \{0, 1\}^{\mathbb{Z}} with (σx)i=xi+1(\sigma x)_i = x_{i+1}. A compatible metric bounded by one is

d0(x,y)=13∑i∈Z2−∣i∣∣xi−yi∣.(17)d_0(x, y) = \frac{1}{3}\sum_{i \in\mathbb{Z}} 2^{-|i|}|x_i-y_i|. \tag*{(17)}

A cylinder set specifies the coordinates in a finite index set and leaves the others unrestricted. Cylinders are open and closed and form a basis for the product topology. Distinct patterns on [0,n)[0,n) have Bowen distance at least 1/31/3, whereas agreement on [−q,n−1+q][-q,n-1+q] gives Bowen distance at most (2/3)2−q(2/3)2^{-q}. Thus, for 0<ε<1/60 < \varepsilon< 1/6 and (2/3)2−q<ε(2/3)2^{-q} < \varepsilon,

2n≤rn(Σ2,ε)≤2n+2q.2^n \le r_n(\Sigma_2,\varepsilon) \le2^{n+2q}.

The lower bound uses one representative for each pattern on [0,n)[0,n); the upper bound uses cylinders on [−q,n−1+q][-q,n-1+q]. Hence h(σ,Σ2)=log⁡2h(\sigma,\Sigma_2) = \log2, and Lemma 2.3 gives Dp(σ,Σ2)=1D_p(\sigma,\Sigma_2) = 1 for every p≥2p \ge2.

For any compact metric space (X,d)(X,d) with d≤1d \le1, the countable product XNX^{\mathbb{N}} is compact and metrizable by

ρ(y,z)=∑j≥12−jd(yj,zj).(18)\rho(y,z) = \sum_{j \ge1}2^{-j}d(y_j,z_j). \tag*{(18)}

For completeness, every sequence in the product has a subsequence converging in coordinate one, a further subsequence converging in coordinate two, and so on. The diagonal subsequence converges in every coordinate. To obtain convergence for (18), first make the tail sum smaller than a prescribed tolerance, and then use convergence of the finitely many remaining coordinates. Sequential compactness is compactness for a metric space. This also explains why only finitely many coordinates are visible at a fixed resolution.

Estimates for the diagonal scales

Fix p≥2p \ge2, write r=p−1r=p-1, and use the functions fbf_b, GbG_b defined in (11). The following estimates account for the changes in covering counts introduced by the tree construction.

Lemma 3.1. For fixed positive parameters the following statements hold.

(1) For b>0b>0,

ℓp(Gb(t))Lr(t)⟶b,fb(t)log⁡t⟶+∞.(19)\frac{\ell_p(G_b(t))}{L_r(t)} \longrightarrow b,\qquad\frac{f_b(t)}{\log t} \longrightarrow+\infty. \tag*{(19)}

(2) If 0<b<d0<b<d, then

(log⁡t)fb(t)=o(fd(t)).(20)(\log t)f_b(t)=o(f_d(t)). \tag*{(20)}

(3) For b>0b>0, A≥0A\ge0, and C,c>0C,c>0,

ℓp(CtAGb(ct))Lr(t)⟶b.(21)\frac{\ell_p(Ct^A G_b(ct))}{L_r(t)} \longrightarrow b. \tag*{(21)}

The same conclusion holds if a fixed nonnegative constant is added inside ℓp\ell_p.

(4) For u,v>0u,v>0, fu∘fv=fuvf_u\circ f_v=f_{uv} wherever both sides are defined. In particular, the inverse of fDf_D on a sufficiently large interval is f1/Df_{1/D}.

(5) Suppose 0<c<D0<c<D and 0<a<D0<a<D. Set

uN=fD−1(12fc(N)),MN=⌊Ga(uN)⌋.(22)u_N=f_D^{-1}\left(\frac{1}{2}f_c(N)\right),\qquad M_N=\lfloor G_a(u_N)\rfloor. \tag*{(22)}

Then uN→∞u_N\to\infty, uN<Nu_N<N eventually, and

ℓp(MN)Lr(N)⟶acD.(23)\frac{\ell_p(M_N)}{L_r(N)}\longrightarrow\frac{ac}{D}. \tag*{(23)}

Proof. We first record two comparisons of iterated exponentials. For d>b>0d>b>0 and j≥1j\ge1,

Ej(dz)Ej(bz)⟶∞,Ej(bz)Ej−1(z)⟶∞(z→∞).(24)\frac{\mathcal{E}_j(dz)}{\mathcal{E}_j(bz)}\longrightarrow\infty,\qquad \frac{\mathcal{E}_j(bz)}{\mathcal{E}_{j-1}(z)}\longrightarrow\infty\qquad(z\to\infty). \tag*{(24)}

For j=1j=1, these are e(d−b)z→∞e^{(d-b)z}\to\infty and ebz/z→∞e^{bz}/z\to\infty. If either comparison holds at jj, the logarithm of the corresponding ratio at j+1j+1 is the difference between its two level-jj terms. The larger term divided by the smaller one tends to infinity, and both terms tend to infinity; hence their difference tends to infinity. This proves both assertions by induction.

The identity Lp(Gb(t))=bLr(t)L_p(G_b(t))=bL_r(t) and (5) give the first limit in (19). For r=1r=1, the second limit and (20) reduce to tb/log⁡t→∞t^b/\log t\to\infty and (log⁡t)tb−d→0(\log t)t^{b-d}\to0. If r≥2r\ge2, put z=Lr(t)z=L_r(t). Taking logarithms gives

log⁡fb(t)log⁡t=Er−1(bz)−Er−2(z),\frac{\log f_b(t)}{\log t}=\mathcal{E}_{r-1}(bz)-\mathcal{E}_{r-2}(z),
log⁡fd(t)(log⁡t)fb(t)=Er−1(dz)−Er−1(bz)−Er−2(z).\log\frac{f_d(t)}{(\log t)f_b(t)}=\mathcal{E}_{r-1}(dz)-\mathcal{E}_{r-1}(bz)-\mathcal{E}_{r-2}(z).

Both right-hand sides tend to infinity by (24): in the second line each subtracted term is of smaller order than the first one. This proves the first two parts. For part (3), first note that Lr(ct)/Lr(t)→1L_r(ct)/L_r(t) \to1. This follows from log⁡(ct)=log⁡t+log⁡c\log(ct)=\log t+\log c for r=1r=1, and from repeated logarithms for larger rr. Part (1) gives log⁡(CtA)=O(log⁡t)=o(fb(ct))\log(Ct^A)=O(\log t)=o(f_b(ct)), whence

log⁡(CtAGb(ct))=fb(ct)(1+o(1)).\log(Ct^A G_b(ct))=f_b(ct)(1+o(1)).

Taking rr further logarithms changes Lr(fb(ct))L_r(f_b(ct)) by o(1)o(1). That expression equals bLr(ct)bL_r(ct). Equation (5) now proves (21). Adding a fixed constant changes the first logarithm by o(1)o(1) and therefore gives the same limit.

The cancellation of LrL_r with Er\mathcal{E}_r gives

fu(fv(t))=Er(uLr(Er(vLr(t))))=Er(uvLr(t)),f_u(f_v(t))=\mathcal{E}_r\bigl(uL_r(\mathcal{E}_r(vL_r(t)))\bigr)=\mathcal{E}_r(uvL_r(t)),

which proves part (4). For part (5), fc(N)→∞f_c(N)\to\infty, so uN→∞u_N\to\infty. Since c<Dc<D and Lr(N)>0L_r(N)>0, one has 12fc(N)<fD(N)\frac{1}{2}f_c(N)<f_D(N) eventually; strict monotonicity gives uN<Nu_N<N. Moreover,

DLr(uN)=Lr(12fc(N))=cLr(N)+O(1).(25)DL_r(u_N)=L_r\left(\frac{1}{2}f_c(N)\right)=cL_r(N)+O(1). \tag*{(25)}

For r=1r=1 the error is exactly −log⁡2-\log2, and for r≥2r\ge2 it tends to zero. Since Ga(uN)→∞G_a(u_N)\to\infty, taking the integer part changes its first logarithm by o(1)o(1) and hence its pp-fold logarithm by $o(1). Thus ℓp(MN)=aLr(uN)+o(1)\ell_p(M_N)=aL_r(u_N)+o(1), and (25) proves (23). □

In the construction below, the parameters in Lemma 3.1 are fixed separately at each stage before the time is chosen large enough for the required error bound. No uniformity over parameters from different stages is needed.

Localization and metric covering trees

Keep p≥2p\ge2 and r=p−1r=p-1 fixed.

A countable compact source

The next two lemmas adapt the localization approach of [7 Section 5] to the fixed-resolution rates (2.11) used in the extraction below.

Lemma 4.1. Let E⊂XE\subset X be nonempty and compact. If dε(E)>b>0d_\varepsilon(E)>b>0, then there is x∗∈Ex_*\in E such that every relative open neighborhood UU of x∗x_* in EE satisfies $d_\varepsilon(U)\ge b.

Proof. Otherwise, each x∈Ex\in E has a relative open neighborhood UxU_x with dε(Ux)<bd_\varepsilon(U_x)<b. Choose a finite subcover Ux1,…,UxmU_{x_1},\ldots,U_{x_m} of EE. Since these sets are contained in EE, their union is EE, and (13) gives dε(E)=max⁡idε(Uxi)<bd_\varepsilon(E)=\max_i d_\varepsilon(U_{x_i})<b, a contradiction. □

Lemma 4.2. If E⊂XE\subset X is nonempty and compact and Dp(T,E)>a>0D_p(T,E)>a>0, then there is a countable compact set

C={x∗}∪{xj:j≥1}⊂E,xj⟶x∗,(26)C=\{x_*\}\cup\{x_j:j\ge1\}\subset E,\qquad x_j\longrightarrow x_*, \tag*{(26)}

with distinct xj≠x∗x_j\ne x_* and Dp(T,C)>aD_p(T,C)>a.

Proof. Choose ε0>0\varepsilon_0>0 and numbers a<b0<b1<dε0(E)a<b_0<b_1<d_{\varepsilon_0}(E). Apply Lemma 4.1 with b=b1b=b_1. For each j∈Nj\in\mathbb{N}, the set Uj=E∩B(x∗,1/j)U_j=E\cap B(x_*,1/j) has dε0(Uj)≥b1>b0d_{\varepsilon_0}(U_j)\ge b_1>b_0. The definition of limsup therefore permits a strictly increasing sequence njn_j and finite (nj,ε0)(n_j,\varepsilon_0)-separated sets Fj⊂UjF_j\subset U_j such that

ℓp(∣Fj∣)Lr(nj)>b0.\frac{\ell_p(|F_j|)}{L_r(n_j)}>b_0.

Put C={x∗}∪⋃jFjC=\{x_*\}\cup\bigcup_j F_j. Every neighborhood of x∗x_* contains all but finitely many of these finite sets. Thus CC is compact and has no accumulation point other than x∗x_*. Since snj(C,ε0)≥∣Fj∣s_{n_j}(C,\varepsilon_0)\ge|F_j|, we have Dp(T,C)≥dε0(C)≥b0>aD_p(T,C)\ge d_{\varepsilon_0}(C)\ge b_0>a. In particular CC is infinite. Enumerating its distinct points other than x∗x_* gives (4.1); every such enumeration converges to x∗x_*, because only finitely many points lie outside any neighborhood of x∗x_*. □

Lemma 4.3. Let CC have the form (4.1), and define CJ={x∗}∪{xj:j≥J}C_J = \{x_\ast\} \cup\{x_j : j \ge J\}. For every L∈NL \in\mathbb{N} and ρ>0\rho> 0, one has CJ⊂BL(x∗,ρ)C_J \subset B_L(x_\ast,\rho) for all sufficiently large JJ. For each JJ and ε>0\varepsilon> 0,

dε(CJ)=dε(C),Dp(T,CJ)=Dp(T,C).d_\varepsilon(C_J) = d_\varepsilon(C), \qquad D_p(T,C_J) = D_p(T,C).

Proof. Each of the finitely many maps TiT^i, 0≤i<L0 \le i < L, is continuous, so dL(xj,x∗)→0d_L(x_j,x_\ast) \to0. This proves the first assertion. The other assertions follow from finite deletion in Lemma 2.2. □

Trees from successive metric covers

Lemma 4.4. Let C⊂XC \subset X be nonempty, let L<NL < N be positive integers, and let P⊂C∩BL(x∗,ρ)P \subset C \cap B_L(x_\ast,\rho) be a nonempty finite (N,η)(N,\eta)-separated set, where η>2ρ\eta> 2\rho. Put tk=2kLt_k = 2^kL for 1≤k≤J1 \le k \le J, where JJ is the first index with tJ≥Nt_J \ge N. There is a rooted tree on successive partitions of PP with these properties:

  1. Each level-kk atom has dtkd_{t_k}-diameter less than 2ρ2\rho.

  2. If BkB_k denotes the number of level-kk vertices, then

Bk≤∏j=1krtj(C,ρ)≤rtk(C,ρ)k.(27)B_k \le\prod_{j=1}^{k} r_{t_j}(C,\rho) \le r_{t_k}(C,\rho)^k. \tag*{(27)}
  1. The terminal leaves are the individual points of PP. If a nonempty subset F⊂PF \subset P occupies bkb_k vertices at level kk, then

rn(F,2ρ)≤bk(n≤tk),rn(F,2ρ)=1(n≤L).(28)r_n(F,2\rho) \le b_k \quad(n \le t_k), \qquad r_n(F,2\rho) = 1 \quad(n \le L). \tag*{(28)}

Proof. At time tkt_k, take a cover of CC by exactly rtk(C,ρ)r_{t_k}(C,\rho) open Bowen balls of radius ρ\rho and order its members. Assign each point of PP to the first ball containing it, discarding empty assigned sets. This produces a partition PkP_k of PP. Each atom is contained in one ball and therefore has dtkd_{t_k}-diameter less than 2ρ2\rho. Define the refining partitions Qk=⋁j=1kPjQ_k = \bigvee_{j=1}^{k} P_j, with Q0={P}Q_0 = \{P\}. These define a rooted tree as in § 2. The bound on the number of intersections gives the first inequality in (4.2); monotonicity in time gives the second.

An atom of QJQ_J cannot contain two distinct points of PP: their dNd_N-distance would be less than $2\rho < \eta$, contradicting separation. Thus terminal atoms are singletons, and all vertices have terminal descendants. For a selected subset FF, choose one of its points from each occupied level-kk atom. The radius-2ρ2\rho balls about those points cover FF at time tkt_k and at every earlier time, since dn≤dtkd_n \le d_{t_k}. Finally F⊂BL(x∗,ρ)F \subset B_L(x_\ast,\rho), so a single ball centered at x∗x_\ast covers FF whenever $n \le L. □

Exact extraction and hereditary lowerability

The next construction adapts [10 Lemma 3.5]. Here the trees are supplied by Lemma 4.4, and Lemma 3.1 controls the resulting losses in the diagonal normalization.

Proposition 5.1. Let p≥2p \ge2 and let CC have the form (4.1). If H=Dp(T,C)>a>0H = D_p(T,C) > a > 0, there is a countable compact set K⊂CK \subset C with exactly one accumulation point such that Dp(T,K)=aD_p(T,K) = a.

Proof. We use r=p−1r = p - 1 and the functions in (2.10). By Lemma 2.3, 0<H≤10 < H \le1.

Choice of resolutions and parameters. Choose εm↓0\varepsilon_m \downarrow0 such that

hm=dεm(C)↑H,h_m = d_{\varepsilon_m}(C) \uparrow H,

discarding finitely many initial terms to have h1>0h_1 > 0. Let m(s)m(s) be a sequence in which every positive integer appears infinitely often; for instance, concatenate the finite lists $(1), (1,2), (1,2,3), …\ldots. This recurrence will provide infinitely many witness times at each fixed resolution before the resolution limit is taken. Choose ηs↓0\eta_s \downarrow0 with 0<ηs<h1/20 < \eta_s < h_1/2, and set

cs=hm(s)−ηs,Ds=H+ηs.(29)c_s = h_{m(s)} - \eta_s,\qquad D_s = H + \eta_s. \tag*{(29)}

Then 0<cs<H<Ds0 < c_s < H < D_s and 0<a<Ds0 < a < D_s. Choose ρs↓0\rho_s \downarrow0 so that 2ρs<εm(s)2\rho_s < \varepsilon_{m(s)} for every ss. The auxiliary number DsD_s may exceed one; it is a comparison parameter rather than a claimed entropy value.

A threshold before choosing a tail. Put N0=0N_0 = 0. At stage ss, choose an integer Ls>Ns−1L_s > N_{s-1}, in the positive-logarithm domain, so large that

(1+log⁡2t)log⁡rt(C,ρs)≤fDs(t)for every integer t≥Ls.(30)(1+\log_2 t)\log r_t(C,\rho_s) \le f_{D_s}(t)\qquad\text{for every integer }t \ge L_s. \tag*{(30)}

Here and below log⁡2t=(log⁡t)/(log⁡2)\log_2 t = (\log t)/(\log2). To justify this choice, take H<Ds′<DsH < D'_s < D_s. Lemma 2.1 implies that the fixed-resolution covering rate at ρs\rho_s is at most HH, so rt(C,ρs)≤GDs′(t)r_t(C,\rho_s) \le G_{D'_s}(t) for all sufficiently large integers tt. One can obtain this last inequality directly by inserting a parameter strictly between HH and Ds′D'_s in the proof of that lemma. Equation (20) then yields (1+log⁡2t)fDs′(t)≤fDs(t)(1+\log_2 t)f_{D'_s}(t) \le f_{D_s}(t) eventually, establishing (5.2).

Only after LsL_s has been fixed, choose a tail index JsJ_s, strictly larger than previous tail indices, for which

Cs:={x∗}∪{xj:j≥Js}⊂BLs(x∗,ρs).(31)C_s := \{x_*\} \cup\{x_j:j \ge J_s\} \subset B_{L_s}(x_*,\rho_s). \tag*{(31)}

This is possible by Lemma 4.3. In particular, all times up to LsL_s are controlled by one ball.

Separated source points and the selected cardinality. Finite deletion gives dεm(s)(Cs)=hm(s)>csd_{\varepsilon_{m(s)}}(C_s) = h_{m(s)} > c_s. There are therefore arbitrarily large integers N>LsN > L_s for which CsC_s contains an (N,εm(s))(N,\varepsilon_{m(s)})-separated set PP with

∣P∣≥Gcs(N).(32)|P| \ge G_{c_s}(N). \tag*{(32)}

Indeed, choose bb with cs<b<hm(s)c_s < b < h_{m(s)}. Along arbitrarily large times the normalized regularized logarithm of the separated count exceeds bb, while that of Gcs(N)G_{c_s}(N) tends to csc_s. For such times define

uN=fDs−1(12fcs(N)),MN=⌊Ga(uN)⌋.u_N = f_{D_s}^{-1}\left(\frac{1}{2}f_{c_s}(N)\right),\qquad M_N = \left\lfloor G_a(u_N)\right\rfloor.

Lemma 3.1 allows us to choose one of the times in (5.4), denoted by NsN_s, so large that uNs≥Lsu_{N_s} \ge L_s, uNs<Nsu_{N_s} < N_s, MNs≥2M_{N_s} \ge2, and

∣ℓp(MNs)Lr(Ns)−acsDs∣<1s.(33)\left|\frac{\ell_p(M_{N_s})}{L_r(N_s)}-\frac{ac_s}{D_s}\right| < \frac{1}{s}. \tag*{(33)}

Write us=uNsu_s = u_{N_s}, Ms=MNsM_s = M_{N_s}, and let PsP_s be the corresponding separated set, with As=∣Ps∣A_s = |P_s|. All these requirements are eventual inequalities for the already fixed stage parameters, so they can be met simultaneously along the unbounded available times.

Since a<Dsa < D_s and Lr(us)>0L_r(u_s) > 0,

log⁡Ms≤fa(us)≤fDs(us)=12fcs(Ns).(34)\log M_s \le f_a(u_s) \le f_{D_s}(u_s) = \frac{1}{2}f_{c_s}(N_s). \tag*{(34)}

Together with As≥exp⁡(fcs(Ns))A_s \ge\exp(f_{c_s}(N_s)), this gives 1≤Ms≤As1 \le M_s \le A_s.

Selection on a metric covering tree. Apply Lemma 4.4 to PsP_s, using L=LsL = L_s, N=NsN = N_s, ρ=ρs\rho= \rho_s, and η=εm(s)\eta= \varepsilon_{m(s)}. Its times are tk=2kLst_k = 2^kL_s, ending at the first tJ≥Nst_J \ge N_s. Denote its node counts by Bs,kB_{s,k}. Since k≤1+log⁡2tkk \le1+\log_2 t_k, equations (27) and (5.2) give

Bs,k≤rtk(C,ρs)k≤exp⁡(fDs(tk)).(35)B_{s,k} \le r_{t_k}(C,\rho_s)^k \le\exp(f_{D_s}(t_k)). \tag*{(35)}

Use Lemma 2.5 to select MsM_s leaves, and let Fs⊂PsF_s \subset P_s be the corresponding points. Write bs,kb_{s,k} for the occupied node counts. When tk≤ust_k \le u_s, we have

MsBs,kAs≤exp⁡(fa(us)+fDs(tk)−fcs(Ns))≤exp⁡(2fDs(us)−fcs(Ns))=1.\begin{aligned} \frac{M_sB_{s,k}}{A_s} \le\exp\left(f_a(u_s)+f_{D_s}(t_k)-f_{c_s}(N_s)\right) \\ &\le\exp\left(2f_{D_s}(u_s)-f_{c_s}(N_s)\right)=1. \end{aligned}

It follows that bs,k≤4k2b_{s,k} \le4k^2. When tk>ust_k > u_s, cardinality alone gives bs,k≤Ms≤Ga(tk)b_{s,k} \le M_s \le G_a(t_k). These two estimates hold for every level in the same selected tree.

One upper bound for every stage and every time. For Ls<n≤NsL_s < n \le N_s, let tkt_k be the first dyadic time at least nn. Then n≤tk<2nn \le t_k < 2n and k≤1+log⁡2nk \le1 + \log_2 n. By (28) and the two preceding estimates,

rn(Fs,2ρs)≤max⁡{4(1+log⁡2n)2,Ga(2n)}.r_n(F_s,2\rho_s) \le\max\{4(1+\log_2 n)^2,G_a(2n)\}.

This includes the case n<us<tkn < u_s < t_k, for which the second estimate applies. For n≤Lsn \le L_s the cover has one ball by (31). For n>Nsn > N_s,

rn(Fs,2ρs)≤Ms≤Ga(us)≤Ga(n).r_n(F_s,2\rho_s) \le M_s \le G_a(u_s) \le G_a(n).

Extend GaG_a to a positive nondecreasing function on the remaining bounded interval, keeping its original values for large arguments. The three cases give a stage-independent bound

rn(Fs,2ρs)≤Q(n)(n≥1),r_n(F_s,2\rho_s) \le Q(n) \qquad(n \ge1),
Q(n)=4(1+log⁡2(2n))2(1+Ga(2n)).(36)Q(n)=4(1+\log_2(2n))^2(1+G_a(2n)). \tag*{(36)}

The compact union and its upper rate. Set

K={x∗}∪⋃s≥1Fs.(37)K=\{x_*\}\cup\bigcup_{s\ge1}F_s. \tag*{(37)}

Each FsF_s is finite and lies in B(x∗,ρs)B(x_*,\rho_s) by (31). Since ρs→0\rho_s\to0, the set KK is countable and compact, and it has at most the accumulation point x∗x_*. Also K⊂CK\subset C.

Fix ε>0\varepsilon>0, and choose s0s_0 such that 2ρs<ε2\rho_s<\varepsilon whenever s≥s0s\ge s_0. Let Cε=∑s<s0∣Fs∣C_\varepsilon=\sum_{s<s_0}|F_s|. The integers LsL_s are strictly increasing, since Ls>Ns−1>Ls−1L_s>N_{s-1}>L_{s-1}. Hence at most nn stages satisfy Ls<nL_s<n. For the stages s≥s0s\ge s_0 with Ls<nL_s<n, use (36) and monotonicity in the radius. For every remaining stage s≥s0s\ge s_0 with Ls≥nL_s\ge n, equation (31) implies Fs⊂Bn(x∗,ρs)⊂Bn(x∗,ε)F_s\subset B_n(x_*,\rho_s)\subset B_n(x_*,\varepsilon). All those remaining stages, together with x∗x_*, need just one ball. Thus

rn(K,ε)≤Cε+nQ(n)+1.(38)r_n(K,\varepsilon)\le C_\varepsilon+nQ(n)+1. \tag*{(38)}

For sufficiently large nn, the right-hand side is at most Cε′n3Ga(2n)C'_\varepsilon n^3G_a(2n) for a fixed constant Cε′>0C'_\varepsilon>0: use (1+log⁡2(2n))2=O(n2)(1+\log_2(2n))^2=O(n^2) and Ga(2n)→∞G_a(2n)\to\infty. Part (3) of Lemma 3.1 gives

lim sup⁡n→∞ℓp(rn(K,ε))Lr(n)≤a.\limsup_{n\to\infty}\frac{\ell_p(r_n(K,\varepsilon))}{L_r(n)}\le a.

Since ε\varepsilon was arbitrary, Dp(T,K)≤aD_p(T,K)\le a.

A lower bound at each recurring resolution. Fix m∈Nm\in\mathbb{N}. There are infinitely many stages with m(s)=mm(s)=m. At each such stage, Fs⊂KF_s\subset K is (Ns,εm)(N_s,\varepsilon_m)-separated, so sNs(K,εm)≥Mss_{N_s}(K,\varepsilon_m)\ge M_s. Equation (33) therefore implies

dεm(K)≥lim⁡s→∞m(s)=ma(hm−ηs)H+ηs=ahmH.d_{\varepsilon_m}(K)\ge\lim_{\substack{s\to\infty\\m(s)=m}}\frac{a(h_m-\eta_s)}{H+\eta_s}=\frac{ah_m}{H}.

Letting m→∞m\to\infty yields Dp(T,K)≥aD_p(T,K)\ge a. This proves equality. Since a>0a>0, Lemma 2.2 rules out a finite KK; compactness then ensures that x∗x_* is its unique accumulation point. □

Proof of Theorem 1.1. If a=0a=0, choose a singleton in EE. If a=Dp(T,E)a=D_p(T,E), choose Ka=EK_a=E. For 0<a<Dp(T,E)0<a<D_p(T,E), Lemma 4.2 gives a countable compact C⊂EC\subset E of the form (4.1) with Dp(T,C)>aD_p(T,C)>a. Apply Proposition 5.1 to CC. The resulting set lies in EE and has exactly one accumulation point. □

Thus, for every nonempty compact E⊂XE\subset X and each fixed p≥2p\ge2,

{Dp(T,K):∅≠K⊂E, K compact}=[0,Dp(T,E)];\{D_p(T,K):\varnothing\ne K\subset E,\ K\text{ compact}\}=[0,D_p(T,E)];

the theorem gives inclusion from right to left, and monotonicity gives the reverse inclusion.

Remark 5.2. The proof uses p≥2p \ge2 at specific points. For a formal classical replacement fb(t)=btf_b(t)=bt, the estimate (log⁡t)fb(t)=o(fd(t))(\log t)f_b(t)=o(f_d(t)) fails: the quotient equals (b/d)log⁡t(b/d)\log t. In addition, Lr(2n)/Lr(n)→1L_r(2n)/L_r(n)\to1 for r≥1r\ge1, whereas the classical denominator nn retains the factor two. Finally, the source value HH is always finite here by Lemma 2.3. Consequently this argument gives no hereditary lowerability assertion for h1,1=hh^{1,1}=h on infinite-entropy systems.

Sharpness and comparison with classical entropy

We give two complements to Theorem 1.1. The first shows why the assertion about one accumulation point is restricted to strictly intermediate values. The second compares DpD_p with classical entropy on the product-diagonal construction of Huang, Ye, and Zhang.

The upper endpoint

Proposition 6.1. For every integer p≥2p\ge2 and every H∈(0,1]H\in(0,1], there are a compact metric space ZZ, a homeomorphism R:Z→ZR: Z\to Z, and a nonempty compact set E⊂ZE\subset Z such that

Dp(R,E)=H,h(R,E)=0,D_p(R,E)=H,\qquad h(R,E)=0,

while every nonempty compact subset K⊂EK\subset E with at most one accumulation point satisfies Dp(R,K)<HD_p(R,K)<H.

Proof. Choose 0<aj<H0<a_j<H with aj↑Ha_j\uparrow H. Since Dp(σ,Σ2)=1D_p(\sigma,\Sigma_2)=1, Theorem 1.1 supplies nonempty compact sets Aj⊂Σ2A_j\subset\Sigma_2 with Dp(σ,Aj)=ajD_p(\sigma,A_j)=a_j. Moreover, h(σ,Aj)=0h(\sigma,A_j)=0: otherwise Lemma 2.3 would give Dp(σ,Aj)=1D_p(\sigma,A_j)=1, contrary to aj<1a_j<1.

Let XjX_j be a separate copy of Σ2\Sigma_2, and set Z={∞}∪⋃j≥1XjZ=\{\infty\}\cup\bigcup_{j\ge1}X_j. Define

δ(x,y)=2−jd0(x,y)(x,y∈Xj),δ(x,y)=2−i+2−j(x∈Xi, y∈Xj, i≠j),δ(x,∞)=δ(∞,x)=2−j(x∈Xj),\begin{aligned} \delta(x,y)&=2^{-j}d_0(x,y) &&(x,y\in X_j),\\ \delta(x,y)&=2^{-i}+2^{-j} &&(x\in X_i,\ y\in X_j,\ i\ne j),\\ \delta(x,\infty)&=\delta(\infty,x)=2^{-j} &&(x\in X_j), \end{aligned}

and δ(∞,∞)=0\delta(\infty,\infty)=0. This is a metric. Within a component the triangle inequality follows from that for d0d_0; a path leaving XjX_j has length at least 2⋅2−j2\cdot2^{-j}, whereas distances within XjX_j are at most 2−j2^{-j}. For points in different components the displayed sums give the remaining triangle inequalities. Each XjX_j is compact and open and closed in ZZ. A sequence with unbounded component indices has a subsequence converging to ∞\infty; otherwise it has a subsequence in a fixed compact component or a constant subsequence at ∞\infty. Thus ZZ is compact.

Let RR act as the shift on each XjX_j and fix ∞\infty. Both RR and R−1R^{-1} are continuous within each component and preserve the distance to ∞\infty, so RR is a homeomorphism. Regard AjA_j as a subset of XjX_j and put

E={∞}∪⋃j≥1Aj.E=\{\infty\}\cup\bigcup_{j\ge1}A_j.

This set is closed: each AjA_j is closed in its component, and limits with unbounded component indices equal ∞\infty. Hence EE is compact.

Use superscripts to distinguish covering numbers for RR and σ\sigma. Fix ε>0\varepsilon>0 and choose JJ so that 2−j<ε2^{-j}<\varepsilon for j>Jj>J. All components with j>Jj>J remain in the radius-ε\varepsilon Bowen ball about ∞\infty at every time. Covering the other components separately therefore gives

nnR(E,ε)≤1+∑j=1Jnnσ(Aj,2jε).(39)n_n^R(E,\varepsilon)\le1+\sum_{j=1}^{J}n_n^\sigma(A_j,2^j\varepsilon). \tag*{(39)}

A finite sum is at most its largest summand times the number of terms. Thus, by the bounded-error estimate in Lemma 2.2, its normalized ℓp\ell_p-growth is at most the largest rate of its summands; the same argument applies to the classical logarithm. It follows that h(R,E)=0h(R,E)=0 and Dp(R,E)≤HD_p(R,E)\le H after taking the resolution limit. For each fixed jj and ε<2−j\varepsilon<2^{-j}, a ball meeting XjX_j must have its center in XjX_j, already by the distance at time zero. Hence

rnR(Aj,ε)=rnσ(Aj,2jε)(ε<2−j).r_n^R(A_j,\varepsilon)=r_n^\sigma(A_j,2^j\varepsilon)\qquad(\varepsilon<2^{-j}).

Consequently Dp(R,Aj)=ajD_p(R,A_j)=a_j, and monotonicity yields Dp(R,E)≥sup⁡jaj=HD_p(R,E)\ge\sup_j a_j=H.

Now let K⊂EK\subset E be nonempty and compact with at most one accumulation point. If there is none, KK is finite and Dp(R,K)=0D_p(R,K)=0. If the only accumulation point is ∞\infty, each K∩AjK\cap A_j is finite: it is compact, and an infinite such intersection would have an accumulation point in XjX_j. At each fixed resolution only finitely many components require separate covers in (6.1); their intersections with KK are finite, and the remaining points fit in one ball about ∞\infty. Thus rnR(K,ε)r_n^R(K,\varepsilon) is bounded in nn and Dp(R,K)=0D_p(R,K)=0.

Finally, suppose the unique accumulation point lies in AjA_j. Then KK meets only finitely many components, since otherwise a sequence with increasing component indices would have the second accumulation point ∞\infty. Its intersections with the other components are finite by compactness. Therefore K∖AjK\setminus A_j is finite, and Lemma 2.2 gives Dp(R,K)≤aj<HD_p(R,K)\le a_j<H. □

The product-diagonal comparison

Let (X,T)(X,T) be a compact homeomorphism system with 0<h(T,X)<∞0<h(T,X)<\infty, give XX a compatible metric d≤1d\le1, and set

Y=XN,S(y)j=Tjyj,Δx=(x,x,…).Y=X^{\mathbb{N}},\qquad S(y)_j=T^j y_j,\qquad\Delta x=(x,x,\ldots).

Equip YY with the metric (2.17). The maps SS and S−1S^{-1} are continuous coordinatewise, so SS is a homeomorphism. The map Δ\Delta is a homeomorphism onto its image, with inverse given by the first coordinate projection. In this construction, [8 Proposition 3.2 and its proof] gives, for every nonempty compact A⊂XA\subset X,

h(S,ΔA)={0,h(T,A)=0,+∞,h(T,A)>0.(40)h(S,\Delta A)= \begin{cases} 0, & h(T,A)=0,\\ +\infty, & h(T,A)>0. \end{cases} \tag*{(40)}

In particular, ΔX\Delta X has infinite classical entropy and no compact subset of finite positive classical entropy. The diagonal scales have the following behavior.

Proposition 6.2. For every nonempty compact A⊂XA\subset X and every p≥2p\ge2,

Dp(S,ΔA)=Dp(T,A).(41)D_p(S,\Delta A)=D_p(T,A). \tag*{(41)}

Proof. Projection onto the first coordinate sends a radius-ε\varepsilon Bowen ball for SS into a radius-2ε2\varepsilon Bowen ball for TT, since ρ(Siy,Siz)≥12d(Tiy1,Tiz1)\rho(S^i y,S^i z)\ge\frac12d(T^i y_1,T^i z_1) for every ii. The centers project to points of XX, whether or not they belong to AA. Projecting a cover therefore yields

rnT(A,2ε)≤rnS(ΔA,ε),(42)r_n^T(A,2\varepsilon)\le r_n^S(\Delta A,\varepsilon), \tag*{(42)}

and hence Dp(T,A)≤Dp(S,ΔA)D_p(T,A)\le D_p(S,\Delta A).

For the reverse inequality, fix ε>0\varepsilon>0, choose mm with ∑j>m2−j<ε/2\sum_{j>m}2^{-j}<\varepsilon/2, and take 0<δ<ε/20<\delta<\varepsilon/2. If dmnT(x,u)<δd_{mn}^T(x,u)<\delta, then for 0≤i<n0\le i<n,

ρ(SiΔx,SiΔu)≤∑j=1m2−jd(Tjix,Tjiu)+∑j>m2−j<δ+ε/2<ε,\rho(S^i\Delta x,S^i\Delta u)\le\sum_{j=1}^{m}2^{-j}d(T^{ji}x,T^{ji}u)+\sum_{j>m}2^{-j}<\delta+\varepsilon/2<\varepsilon,

because ji<mnji<mn for j≤mj\le m. Replacing every center uu in a length-mnmn cover of AA by Δu\Delta u gives

rnS(ΔA,ε)≤rmnT(A,δ).(43)r_n^S(\Delta A,\varepsilon)\le r_{mn}^T(A,\delta). \tag*{(43)}

Put r=p−1≥1r=p-1\ge1. The integer mm is fixed at this resolution, and Lr(mn)/Lr(n)→1L_r(mn)/L_r(n)\to1. Thus

lim sup⁡n→∞ℓp(rnS(ΔA,ε))Lr(n)≤lim sup⁡n→∞ℓp(rmnT(A,δ))Lr(mn)≤lim sup⁡N→∞ℓp(rNT(A,δ))Lr(N)≤Dp(T,A).\limsup_{n\to\infty}\frac{\ell_p(r_n^S(\Delta A,\varepsilon))}{L_r(n)} \le\limsup_{n\to\infty}\frac{\ell_p(r_{mn}^T(A,\delta))}{L_r(mn)} \le\limsup_{N\to\infty}\frac{\ell_p(r_N^T(A,\delta))}{L_r(N)} \le D_p(T,A).

The middle inequality only bounds a subsequence upper limit by the full upper limit. Letting ε↓0\varepsilon\downarrow0 proves the reverse bound. □

For the binary full shift, Theorem 1.1 supplies, for each fixed p≥2p\ge2 and 0<a<10<a<1, a nonempty compact set A⊂Σ2A\subset\Sigma_2 with Dp(σ,A)=aD_p(\sigma,A)=a. Lemma 2.3 then gives h(σ,A)=0h(\sigma,A)=0. Equations (40) and (41) imply

h(S,ΔA)=0,Dp(S,ΔA)=a.h(S,\Delta A)=0,\qquad D_p(S,\Delta A)=a.

Thus the classical entropy restriction on ΔΣ2\Delta\Sigma_2 allows intermediate diagonal scale values on its classical zero-entropy subsets.

College of Mathematics and Statistics, Chongqing University, Chongqing 401331, China Email address: [email protected]

College of Mathematics and Statistics, Chongqing University, Chongqing 401331, China Email address: [email protected]

References

  1. [1]R. L. Adler, A. G. Konheim, and M. H. McAndrew, Topological entropy, Trans. Amer. Math. Soc. 114 (1965), no. 2, 309–319. doi:10.1090/S0002-9947-1965-0175106-9.DOI
  2. [2]R. Bowen, Entropy for group endomorphisms, Trans. Amer. Math. Soc. 153 (1971), 401–414. doi:10.1090/S0002-9947-1971-0274707-X.
  3. [3]R. Bowen, Topological entropy for noncompact sets, Trans. Amer. Math. Soc. 184 (1973), 125–136. doi:10.1090/S0002-9947-1973-0338317-X.DOI
  4. [4]M. de Carvalho, Entropy dimension of dynamical systems, Portugal. Math. 54 (1997), no. 1, 19–40.
  5. [5]D. Dou, W. Huang, and K. K. Park, Entropy dimension of topological dynamical systems, Trans. Amer. Math. Soc. 363 (2011), no. 2, 659–680. doi:10.1090/S0002-9947-2010-04906-2.DOI
  6. [6]M. Helfter, Scales, Math. Z. 310 (2025), no. 1, Paper No. 15. doi:10.1007/s00209-025-03719-5.DOI
  7. [7]W. Huang, X. Ye, and G. Zhang, Lowering topological entropy over subsets, Ergodic Theory Dynam. Systems 30 (2010), no. 1, 181–209. doi:10.1017/S0143385709000066.DOI
  8. [8]W. Huang, X. Ye, and G. Zhang, Lowering topological entropy over subsets revisited, Trans. Amer. Math. Soc. 366 (2014), no. 8, 4423–4442. doi:10.1090/S0002-9947-2014-06117-5.DOI
  9. [9]A. Katok and J.-P. Thouvenot, Slow entropy type invariants and smooth realization of commuting measure-preserving transformations, Ann. Inst. H. Poincaré Probab. Statist. 33 (1997), no. 3, 323–338. doi:10.1016/S0246-0203(97)80094-5.DOI
  10. [10]X. Wang, Hereditary lowerability of topological dynamical systems, 2026. Preprint, arXiv:2608.07115v1.arxiv.org/abs/2608.07115
  11. [11]Y. Xie, E. Chen, and X. Zhou, Entropy scales in topological dynamical systems, 2026. Preprint, arXiv:2608.10460v2.arxiv.org/abs/2608.10460

Paper details

vlnc:W_qp96w6nw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents