Bernoulli shifts with bases of equal entropy are isomorphic (Q2162732)

Let \((L,\lambda)\) be a standard probability space (i.e., an interval equipped with Lebesgue measure, an at-most countable collection of atoms, or the union of both). Any group \(G\) can be made to act on the left on the probability space \((L^G, \lambda^G)\) by the rule that an element \(g \in G\) takes a function \(\varphi\colon G \to L\) to the function \(g.\varphi\) defined by the rule \(g.\varphi(t) = \varphi(g^{-1}t)\). This is the \textit{Bernoulli shift} over \(G\) with base \((L,\lambda)\). Let \((2, u_2)\) and \((3, u_3)\) denote standard probability spaces on a set of two or three atoms with equal measure respectively. An old question of von Neumann asked whether the Bernoulli shifts over \(\mathbb{Z}\) with bases \((2, u_2)\) and \((3, u_3)\) are isomorphic. More generally, classifying Bernoulli shifts over countably infinite groups up to isomorphism is a major open problem in dynamics. This paper is a contribution to this problem. To answer von Neumann's question in the negative, \textit{A. N. Kolmogorov} [Dokl. Akad. Nauk SSSR 124, 754--755 (1959; Zbl 0086.10101)] and \textit{Ya. G. Sinai} [Sov. Math., Dokl. 3, 1725--1729 (1962; Zbl 0205.13501); translation from Dokl. Akad. Nauk SSSR 147, 797--800 (1962)] developed a notion of \textit{entropy} for probability-measure-preserving actions of \(\mathbb{Z}\) on probability spaces like Bernoulli shifts. Later a notion of entropy was defined for probability-measure-preserving actions of \textit{sofic} groups. For a Bernoulli shift over a countable sofic group \(G\) with base \((L,\lambda)\), this notion of entropy is equal to the \textit{Shannon entropy} \(H(L,\lambda)\) of the base space, which is defined to be \[ H(L,\lambda) = \sum_{\ell \in L} -\lambda(\ell) \log \lambda(\ell) \] when the support of \(\lambda\) is countable, and infinity otherwise. Since isomorphic Bernoulli shifts must have equal entropy, it follows that for a sofic group \(G\) like \(\mathbb{Z}\), the Bernoulli shifts over \(G\) with bases \((2,u_2)\) and \((3,u_3)\) are not isomorphic. It remains unknown whether all countable groups are sofic, so for general \(G\), von Neumann's problem remains open. The contribution of this paper is the following. Suppose the standard probability spaces \((L,\lambda)\) and \((H,\kappa)\) has equal Shannon entropy. Then for any countably infinite group \(G\), the Bernoulli shifts over \(G\) with bases \((L,\lambda)\) and \((K,\kappa)\) are isomorphic. In the special cases where \(L\) and \(K\) are either finite or countably infinite with infinite entropy, then the isomorphism so constructed may be taken to be \textit{finitary.} The proof of this theorem is inspired by work of \textit{L. Bowen} [Contemp. Math. 567, 67--78 (2012; Zbl 1277.37016)], who proves a similar statement under the more restrictive assumption that \((L,\lambda)\) and \((K,\kappa)\) are supported on at least three points. It also relies on earlier work of \textit{A. M. Stepin} [Sov. Math., Dokl. 16, 886--889 (1975; Zbl 0326.28026); translation from Dokl. Akad. Nauk SSSR 223, 300--302 (1975)] and \textit{D. Ornstein} [Adv. Math. 4, 337--352 (1970; Zbl 0197.33502)], which establishes the statement of the theorem under the assumption that the group \(G\) contains \(\mathbb{Z}\) as a subgroup. The main technical innovation of the proof appears to be the definition of a certain relation \(R_\Gamma\) on standard probability spaces, where \(\Gamma\) denotes a nontrivial finite group. The relation \(R_\Gamma\) is not transitive, but the author shows that if \(\Gamma\) has at least five elements, then two standard probability spaces have equal Shannon entropy if and only if they are related in the transitive closure of \(R_\Gamma\). The author then shows that if \((L,\lambda)\) and \((K,\kappa)\) are actually related by \(R_\Gamma\), then one can cleverly construct a copy of \(\mathbb{Z}\) (not in \(G\) but preserving the orbit-equivalence relation on a certain positive-measure subset of an ergodic probability space \((M^G,\mu)\)) and apply Ornstein's isomorphism theorem. The paper is well written and provides a very useful account of the history of the problem. The definition of the relation \(R_\Gamma\) made this reviewer uncomfortable: for \((L,\lambda)\) and \((K,\kappa)\) to be related by \(R_\Gamma\), part of the definition requires the \emph{sets} \(L\) and \(K\) to not be disjoint. This assumption appears not to be very onerous from the perspective of the transitive closure of \(R_\Gamma\), since one can cleverly construct new probability spaces from old ones. On the other hand, it makes the property of being related by \(R_\Gamma\) not an invariant of probability spaces. It might have been nice to have instead posited the existence of a bijection from a subset of \(L\) to a subset of \(K\) with certain desired properties.