Scaling entropy and automorphisms with purely point spectrum

Consider the new measure M_{\mu,\rho} on the space X\times X and the measurable partition into the classes of points C^x=\{(x,*)\}\subset X\times X, and use the classical theorem on the existence of conditional measures (see, e.g., [Ro]), which implies the desired formula and the uniqueness \bmod 0 of the family of conditional measures. Now consider the space (X,\rho,\mu^x) for a fixed x; the metric \rho on this space is admissible since the admissibility of a metric is obviously preserved under replacing a measure with an equivalent one.

The fact that an admissible metric satisfies the condition in question was observed above. Now assume that this condition is satisfied. We must prove that the space (X,\rho) (if \rho is a metric) or the quotient space X_{\rho} (if \rho is a semimetric) is separable. It suffices to consider the case of a metric.

The condition stated in the proposition implies that for every \epsilon there exists \delta=\delta(\epsilon) with \lim_{\epsilon \to 0} \delta(\epsilon)=0 such that some set of measure >1-\delta contains a finite \epsilon-net. But this means exactly that the space is separable and the measure is concentrated on a \sigma-compact set.

Assume that the diameter of the compact space does not exceed 1. Assume that the measure of a set X' is greater than 1-\varepsilon and X'\subset \bigcup_{i=1}^k V_{\varepsilon}(x_i). Thus the points x_1,\dots, x_k form an \varepsilon-net in X'. Consider the discrete measure \nu supported by the points x_1,\dots, x_k with charges \nu(x_i) equal to the measures \mu(V(x_i)) of the corresponding balls (if two balls have a nonempty intersection, then we distribute the measure of the intersection proportionally between their centers). Choose an arbitrary point x_{\infty} and set its charge equal to 1-\mu(X'). Now consider the Monge–Kantorovich transportation problem with input measure \mu and output measure \nu. It is easy to see that we have in fact determined an admissible plan \Psi for this problem: the transportation from a point x\in X' goes to the points x_i for which x\in V_{\varepsilon}(x_i), and the remaining part of the measure \mu on the set X\setminus X' goes to the point x_{\infty}. It is easy to compute the cost of this plan; this gives a bound on the Kantorovich distance between the measures \nu and \mu:

On the other hand, we have H(\nu)\leq \ln k =H'_{\rho,\varepsilon}(\mu).

Since an admissible (semi)metric lies in the space L^1(X\times X, \mu \times \mu), the tensor square of the operator U_T, which corresponds to the automorphism T\times T, has nonconstant eigenfunctions. This can happen only if the spectrum of the unitary operator U_T contains a discrete component, and the first claim is proved. In other words, T is an extension of some quotient automorphism with discrete spectrum, which may coincide with T itself. This means that T is a skew product over a base with discrete spectrum. Denote by H \subset L^2(X,\mu) the subspace spanned by all eigenfunctions of U_T. All invariant functions of the operator U_T\otimes U_T belong to the tensor square H\otimes H; hence these functions, regarded as functions of two variables, do not change when the argument ranges over a fiber of the skew product. But if these fibers are not single point sets, it follows that the metric does not distinguish points in fibers and hence is a semimetric. Thus if \rho is a metric, then each fiber necessarily consists of a single point, the spectrum of T is purely discrete, and, by the ergodicity, T is isomorphic to a translation on a group.

Condition 1 is exactly equivalent to the condition of Lemma 1. It is useful to give a direct proof that this condition is necessary. Assume that it is not satisfied, i.e., for all elements g from some set of positive Haar measure, \phi_{\rho}(z)=\phi_{\rho}(g+z) for almost all z. Therefore, the measurable function \phi is constant on cosets of the subgroup generated by g. However, every set of positive Haar measure in a nondiscrete Abelian compact group on which there is an ergodic translation contains an element g that generates a dense cyclic subgroup. Then, since the function \phi_{\rho} is measurable, it follows that it is constant almost everywhere. But if the function \phi_{\rho} is constant, then the metric \rho is also constant and hence not admissible.

Let us prove that Condition 2 of the theorem is necessary. Consider the function \phi_{\rho} for an invariant admissible metric \rho. Assume that \theta=\operatorname{ess\,inf} from Condition 2 is positive; then, using the invariance of \rho, we can construct a continuum of points lying at a fixed positive \rho-distance greater than \theta, which contradicts the admissibility. For the same reason, the set of values of \phi_{\rho} is dense in some neighborhood of the zero. The fact that Condition 2 is sufficient follows from Proposition 1.

The “if” part is proved above; the “only if” part follows from the equality \phi_{\hat\rho}(r)=\int_G \rho(x,x+r)dm(x) and the previous proposition.

Assume that the metric \rho is not semicontinuous at zero and there exists a positive number c>0 such that

Assume that the metric is normalized so that the diameter of the space X is equal to 1. Then it follows from our assumption that for sufficiently small (i.e., belonging to a small neighborhood of the zero in the group G) r there exists a (depending on r) subset in X of measure \alpha, which does not depend on r and is greater than \frac{c}{2}, on which \rho(x,x+r)>\frac{c}{2}. Indeed, 1 \cdot \alpha +(1-\alpha)\frac{c}{2}>\int_G \rho(x,x+r)dm(x)> c. But since the group is compact, there is a set of positive measure for all points x of which the inequality \rho(x,x+r)>\frac{c}{2} holds for all sufficiently small r from some set of positive measure. This in turn contradicts the admissibility of the metric \rho in the formulation of Proposition 1: arbitrarily small values \phi(r) for small r cannot interlace with values greater than \frac{c}{2} by the triangle inequality. Thus we have proved that \rho is semicontinuous.

Since the set B_{\varepsilon}=\phi^{-1}([0,\varepsilon]) is, by definition, the ball of radius \varepsilon centered at \textbf{0}\in G in the metric \hat \rho, which is admissible by assumption, it follows that m B_{\varepsilon}>0, since the metric is nondegenerate (see the definition of an admissible metric). But the sum B_{\varepsilon}+B_{\varepsilon}, like the sum A+A for every set A of positive Haar measure in a locally compact group, contains a neighborhood V of the zero in the standard topology (see, e.g., [We]). It follows from the triangle inequality that B_{\varepsilon}+B_{\varepsilon}\subset B_{2\varepsilon}, so that V \subset B_{2\varepsilon}. Since \varepsilon>0 is arbitrary, we see that the topology on G determined by the average metric \hat\rho coincides with the standard topology, i.e., in the topology determined by \hat \rho, the group G is compact and contains a finite \varepsilon-net for every \varepsilon.

The pointwise a.e. convergence

which follows from the pointwise ergodic theorem, implies that the number of points in an \varepsilon-net for (G,\rho_n) tends to the number of points in an \varepsilon-net for (G,\hat \rho). This means that the sequence H_{\rho_n,\varepsilon}(X) converges to H_{\hat \rho,\varepsilon}(G). From the inequalities of Lemma 3 we see that the sequence H_{\rho_n,\varepsilon}(\mu) is bounded as n \to \infty, and thus the scaling sequence for the automorphism T, which acts on the metric triple (G,\rho, m), is bounded.

Above we have proved that if an automorphism has a discrete spectrum and the average metric is admissible, then the scaling sequence is bounded. But the average metric is always admissible provided that the original metric is bounded and admissible.

Assume that the scaling sequence is bounded for an automorphism T and an admissible metric \rho. Recall that the average metric is indeed a metric (and not a semimetric). Consequently, the space (X,\hat \rho) is precompact, and hence the metric \hat \rho is admissible. Since it is T-invariant, it follows from Theorem 1 that T has a purely discrete spectrum.