On quantization and the classical variational principle for the metric mean dimension

Maria Carvalho, Gustavo Pessil · 2026

Read On quantization and the classical variational principle for the metric mean dimension by Maria Carvalho, Gustavo Pessil (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-17091.txt.

Introduction

Carvalho and Pessil link quantization of measures with the metric mean dimension of dynamical systems. This quantity gauges complexity when entropy is not finite. They define a mean quantization dimension for invariant measures. They prove a variational principle: the metric mean dimension of a system equals the maximum such value over all of them. This also allows limits and suprema to be swapped in the Lindenstrauss-Tsukamoto results. The Katok and Shapira entropies share this trait, and a best case always exists.

Catalog record

CREATORMaria Carvalho, Gustavo Pessil
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE74818
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.17091

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

On quantization and the classical variational principle for the metric mean dimension Maria Carvalho, Gustavo Pessil Abstract . For a homeomorphism T : X → X T\colon X\to X on a compact metric space ( X , d ) (X,d) , one can find in the literature several distinct notions of measure-theoretic ε \varepsilon -entropy maps F ⁡ ( μ , ε ) F(\mu,\varepsilon) , defined on the space of Borel T T -invariant probability measures of X X and satisfying mdim ¯ M ​ ( X , d , T ) = lim sup ε → 0 + sup μ ∈ 𝒫 T ​ ( X ) F ⁡ ( μ , ε ) log ⁡ ( 1 / ε ) \mathrm{\overline{mdim}_{M}}(X,d,T)\,=\,\limsup_{\varepsilon\,\to\,0^{+}}\,\sup_{\mu\,\in\,\mathcal{P}_{T}(X)}\,\frac{F(\mu,\varepsilon)}{\log\,(1/\varepsilon)} where mdim ¯ M ​ ( X , d , T ) \mathrm{\overline{mdim}_{M}}(X,d,T) stands for the upper metric mean dimension. A major question regards the change in the order in which lim sup ε \limsup_{\varepsilon} and sup μ ∈ 𝒫 T ​ ( X ) \sup_{\mu\,\in\,\mathcal{P}_{T}(X)} appear in the previous equality. We introduce the notion of mean quantization dimension of a measure and prove that mdim ¯ M ​ ( X , d , T ) = max μ ∈ 𝒫 T ​ ( X ) ⁡ mdim ¯ Q ​ ( μ ) . \mathrm{\overline{mdim}_{M}}(X,d,T)\,\,=\,\max_{\mu\,\in\,\mathcal{P}_{T}(X)}\,\mathrm{\overline{mdim}_{Q}}(\mu). This concept exhibits a fundamental property that yields a sufficient condition on the entropy map F F under which the aforementioned change in order is valid. 2010 Mathematics Subject Classification Primary: 37B99, 28D20, 37D35. 1. Introduction Let (X,d) be a compact metric space and T\colon X\to X be a homeomorphism. Denote by \mathcal{P}(X) the set of Borel probability measures on X with the weak ∗ -topology; by \mathcal{P}_{T}(X) its subset of T -invariant measures; by \mathcal{E}_{T}(X)\subset\mathcal{P}_{T}(X) the set of ergodic elements; and, given \mu\in\mathcal{P}(X) , \operatorname{{supp}}(\mu) stands for its support. Measure-theoretic and topological entropy are classical invariants in the theory of dynamical systems and are related by the classical variational principle h_{top}(T)\,\,\,=\,\sup_{\mu\,\in\,{\mathcal{P}}_{T}(X)}\,h_{\mu}(T)\,\,\,=\,\sup_{\mu\,\in\,{\mathcal{E}}_{T}(X)}\,h_{\mu}(T) (1) which sometimes provides a natural way to choose important elements from {\mathcal{P}}_{T}(X) . Entropy-dimensional concepts have been recently developed in order to estimate the complexity of infinite entropy systems from a better suited viewpoint. The metric mean dimension , introduced by E. Lindenstrauss and B. Weiss in [ 9 ] , is both a dynamical analogue of the box dimension and a dimensional analogue of the topological entropy. In what follows, we denote the upper and lower metric mean dimensions by \mathrm{\overline{mdim}_{M}}(X,d,T) and \mathrm{\underline{mdim}_{M}}(X,d,T) , respectively, thus emphasizing their dependence on the fixed metric d on the space X where the dynamics T acts. These concepts vanish if the topological entropy of T is finite; if, otherwise, T has infinite entropy, they quantify the speed at which the entropy at scale \varepsilon approaches +\infty as the scale goes to zero. The choice of the metric d has impact precisely on such rate of convergence. In [ 10 ] , E. Lindenstrauss and M. Tsukamoto established a connection between the upper metric mean dimension and the rate distortion function \mathcal{R}_{\mu} . More precisely, they showed that under the tame growth of covering numbers \mathrm{\overline{mdim}_{M}}(X,d,T)\,\,=\,\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\sup_{\mu\,\in\,\mathcal{P}_{T}(X)}\,{\mathcal{R}}_{\mu}(\varepsilon)}{\log\,(1/\varepsilon)}. Meanwhile, in [ 6 ] , Y. Gutman and A. Śpiewak proved a corresponding relation between the upper metric mean dimension and the Rényi entropy, namely \mathrm{\overline{mdim}_{M}}(X,d,T)\,=\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\sup_{\mu\,\in\,\mathcal{E}_{T}(X)}\,\,\inf_{\mathrm{diam}(P)\,\leq\,\varepsilon}\,h_{\mu}(P)}{\log\,(1/\varepsilon)}\, where the infimum is taken over the finite Borel partitions P of X with diameter at most \varepsilon and h_{\mu}(P) stands for the entropy of \mu with respect to P . Moreover, Problem 3 in [ 6 ] asked whether the upper metric mean dimension could be expressed in terms of the Brin-Katok local entropy. An affirmative answer to this problem was given by R. Shi in [ 15 ] : \mathrm{\overline{mdim}_{M}}(X,d,T)\,=\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\sup_{\mu\,\in\,\mathcal{E}_{T}(X)}\,\overline{h}^{BK}_{\mu}(\varepsilon)}{\log\,(1/\varepsilon)}\,. Apart from the ones mentioned above, many other distinct notions of measure-theoretic metric mean dimension were shown to exist and be inspired by some version of entropy, such as Katok metric entropy (cf. [ 15 ] ), Shapira entropy (cf. [ 15 ] ) and Pfister-Sullivan entropy (cf. [ 21 ] ), just to mention a few. A unified framework was provided by E. Chen, R. Yang and X. Zhou in [ 3 ] , where they consider the following family of entropy-like functions of a measure \mu at a given scale \varepsilon (see the precise definitions in [ 3 ] and Section 3 ): {\mathcal{D}}\,\,=\,\,\Bigg\{\begin{aligned} &\underline{h}_{\mu}^{K}(\varepsilon,\delta),\,\,\,\overline{h}_{\mu}^{K}(\varepsilon,\delta),\,\,\,\underline{h}_{\mu}^{K}(\varepsilon),\,\,\,\overline{h}_{\mu}^{K}(\varepsilon),\,\,\,\overline{h}_{\mu}^{BK}(\varepsilon),\,\,\,h_{\mu}^{S}(\varepsilon,\delta),\,\,\\ &PS_{\mu}(T,\varepsilon),\,\,\,{\mathcal{R}}_{\mu,\,L^{\infty}}(\varepsilon),\,\,\,\inf_{\mathrm{diam}(P)\,\leq\,\varepsilon}h_{\mu}(P),\,\,\,{\mathcal{P}}_{\mu}(T,d,\varepsilon)\end{aligned}\Bigg\}. (2) The authors prove that, for any ergodic probability measure \mu\in{\mathcal{E}}_{T}(X) , the limits \limsup_{\varepsilon\,\to\,0^{+}}\,\frac{F(\mu,\varepsilon)}{\log\,(1/\varepsilon)}\qquad\text{and}\qquad\liminf_{\varepsilon\,\to\,0^{+}}\,\frac{F(\mu,\varepsilon)}{\log\,(1/\varepsilon)} (3) do not depend on the choice of F\in{\mathcal{D}} (so they coincide with the upper and lower L^{\infty} -rate distortion dimensions, respectively) and that, for every F\in{\mathcal{D}} , \mathrm{\overline{mdim}_{M}}(X,d,T)\,=\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\sup_{\mu\,\in\,\mathcal{P}_{T}(X)}\,F(\mu,\varepsilon)}{\log\,(1/\varepsilon)}\,=\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\sup_{\mu\,\in\,\mathcal{E}_{T}(X)}\,F(\mu,\varepsilon)}{\log\,(1/\varepsilon)}. (4) We stress that, in equality ( 4 ), the maximization over the space of probability measures is done for a fixed scale, which only afterwards is made to go to zero. A recurring question in the literature asks whether we may exchange the order of \limsup_{\varepsilon} and \sup_{\mu} in ( 4 ). An affirmative answer to this question would suggest natural definitions of measure-theoretic metric mean dimension, namely \mathrm{\overline{mdim}_{M}}(X,d,T,F)(\mu)\,=\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{F(\mu,\varepsilon)}{\log\,(1/\varepsilon)} and provide a classical variational principle for the metric mean dimension in terms of F ; moreover, if \sup_{\mu} could be replaced by \max_{\mu} , this principle would single out a measure which captures the complexities of the system over all scales \varepsilon . It is known that this order can be changed under the marker property (cf. [ 11 ] ), but there are relevant systems that do not satisfy this condition, as shown in [ 16 , 18 ] . The following examples shed some light on this problem. Example 1.1 . [ 10 , Section VIII] Consider the set A=\{0\}\cup\{1/n\colon n\in\mathbb{N}\} , X=A^{\mathbb{Z}} endowed with the metric d(x,y)\,=\,\sum_{n\,\in\,\mathbb{Z}}\,2^{-|n|}\,|x_{n}-y_{n}| and T\colon X\to X the standard shift. Then \mathrm{\overline{mdim}_{M}}(X,d,T)=1/2 (cf. [ 20 , 7 ] ) and \forall\,\mu\in{\mathcal{P}}_{T}(X)\quad\quad\lim_{\varepsilon\,\to\,0^{+}}\,\frac{\mathcal{R}_{\mu}(\varepsilon)}{\log\,(1/\varepsilon)}\,=\,0. So, \mathrm{\overline{mdim}_{M}}(X,d,T)\,\,>\,\sup_{\mu\,\in\,\mathcal{P}_{T}(X)}\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\mathcal{R}_{\mu}(\varepsilon)}{\log\,(1/\varepsilon)}. Example 1.2 . [ 17 , Theorem 1.4] In [ 17 ] Tsukamoto constructs a homeomorphism T on a compact metric space (X,d) with the following properties: \bullet For any ergodic measure \nu\in{\mathcal{P}}_{T}(X) we have \,\displaystyle\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\mathcal{R}_{\nu}(\varepsilon)}{\log\,(1/\varepsilon)}\,=\,0. \bullet There exists \eta\in{\mathcal{P}}_{T}(X) such that \,\displaystyle\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\mathcal{R}_{\eta}(\varepsilon)}{\log\,(1/\varepsilon)}\,=\,+\infty. Consequently, one has \mathrm{\overline{mdim}_{M}}(X,d,T)=+\infty and \mathrm{\overline{mdim}_{M}}(X,d,T)\,\,>\,\sup_{\mu\,\in\,\mathcal{E}_{T}(X)}\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\mathcal{R}_{\mu}(\varepsilon)}{\log\,(1/\varepsilon)}. Yet, it is still true for this example that \mathrm{\overline{mdim}_{M}}(X,d,T)\,\,=\,\sup_{\mu\,\in\,\mathcal{P}_{T}(X)}\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\mathcal{R}_{\mu}(\varepsilon)}{\log\,(1/\varepsilon)}. (5) Why do they behave differently with respect to the change in the order in which \limsup_{\varepsilon} and \sup_{\mu\,\in\,{\mathcal{P}}_{T}(X)} appear? It turns out that a major role in this issue is played by the behavior of the measure-theoretic \varepsilon -entropy under ergodic decompositions. On the one hand, if the upper metric mean dimension is finite, then the upper rate distortion dimension \overline{R}(X,d,T,\mu)\,=\,\limsup_{\varepsilon\,\to\,0^{+}}\,\frac{\mathcal{R}_{\mu}(\varepsilon)}{\log\,(1/\varepsilon)} is a convex function of \mu\in\mathcal{P}_{T}(X) with respect to its ergodic decomposition (cf. [ 17 , Theorem 1.1] ). Thus, in Example 1.1 , the change in order between \limsup_{\varepsilon} and \sup_{\mu\,\in\,{\mathcal{P}}_{T}(X)} produces a strict inequality. On the other hand, the system described in Example 1.2 has \mathrm{\overline{mdim}_{M}}(X,d,T)=+\infty and the convexity no longer holds. This is the reason why ( 5 ) is possible. The purpose of this paper is to explain that we can understand this issue more clearly by comparing the maps F\in{\mathcal{D}} with a new concept that we introduce in the next section. 2. Main results After recalling the

Read the complete text