On classical analogues of free entropy dimension (Q2461228): Difference between revisions

Using his own notion of free entropy \({\mathcal X}\), \textit{D. Voiculescu} [Invent.\ Math.\ 118, No. 3, 411--440 (1994; Zbl 0820.60001)] introduced the free entropy dimension \(\delta\) of a non-commutative law. As a matter of fact, if \(X_1, \dots, X_n\) \(\in\) \(( M, \tau)\) are self-adjoint non-commutative random variables in a tracial \(W^*\)-probability space, then \[ \delta ( X_1, \dots, X_n ) = n + \limsup_{t \to 0} \frac{ {\mathcal X} ( X_1^t, \dots, X_n^t)}{ | \log t |}, \tag{1} \] where \(X_j^t = X_j + t S_j\) and \(S_1, \dots, S_n\) form a free semicircular family, free from \(X_1, \dots, X_n\). The authors of the present paper define a classical probability analogue (= the classical probability entropy dimension of a probability measure on \({\mathbb R}^n\)) of Voiculescu's free entropy dimension. The purpose of this paper is to show the various relations of the classical probability entropy dimension of a measure to diverse other notions of dimension, such as fractal dimension and non-microstates free entropy dimension, etc. More precisely, let a probability measure \(\mu\) on \({\mathbb R}^n\) be the law of \(n\) real random variables \(X_1, \dots, X_n\), and let \(\nu_t\) be the Gaussian law of independent Gaussian random variables \(G_1, \dots, G_n\). Then the measure \(\mu_t = \mu * \nu_t\) is the law of \(X_1^t, \dots, X_n^t\) with \(X_j^t=X_j + t G_j\), \(( j=1, \dots, n)\). The authors define the classical analogue of free entropy dimension as \[ \delta_c(\mu) = n - \liminf_{ t \to 0} \frac{ H(\mu_t)}{| \log t |}, \tag{2} \] where for a non-negative Lebegue absolutely continuous measure \(p(x) d x\), \(H\) is given by \[ H( p(x) dx) = \int p(x) \log p(x) dx. \] First, it is shown that the dimension \(\delta_c(\mu)\) can be viewed as a kind of average fractal dimension of the measure \(\mu\). In particular, the authors prove that \(\delta_c(\mu)\) remains the same if \(\mu\) is replaced by a push-forward by a Lipschitz function. Note that the value of \(\delta_c(\mu)\) may change if one pushes forward \(\mu\) by a continuous or measurable function. For a probability measure \(\mu\) on \({\mathbb R}\) and the function \[ d_t(x) = - \log \mu[x - \frac{t}{2}, x + \frac{t}{2} ] / | \log t |, \] we have \[ \delta_c(\mu) = \limsup_{t \to 0} \int d_t(y) d \mu(y). \tag{3} \] The authors also show that there is a connection between \(\delta_c\) and the entropy dimension \(h^*\) recently studied by \textit{A. Batakis} and \textit{Y. Heurteaux} [Asian J.\ Math.\ 6, No. 3, 399--408 (2002; Zbl 1046.28003)]. Here, for a probability measure \(m\) on \([0, 1)\) and an integer \(\ell > 1\), \[ \begin{aligned} I_i( \ell^n) &= [ i \ell^{-n}, ( i+1) \ell^{-n} ), \\ h_n(m) &= - \frac{1}{n \log \ell } \sum_{i=1}^{\ell^n -1} m( I_i(\ell^n)) \cdot \log\{ m( I_i(\ell^n)) \}, \quad \text{and} \\ h^*(m) & \equiv h_{\ell}^*(m) = \limsup_{n \to \infty} h_n(m). \end{aligned} \] As a matter of fact, the following assertion holds: Proposition 1. For each \(\ell, n\) and \(\epsilon\), let \[ X(\ell, \epsilon, n) = \bigcup_k \left( \frac{ (1- \epsilon) (k+1)}{\ell^n}, \frac{(k+1)}{\ell^n} \right]. \] Then we have (i) If the \(\limsup\) in (3) is attained along the sequence \(\ell^{-n}\), then \(h_{\ell}^*(m) \geq \delta_c(m)\). (ii) \(h_{\ell}^*(m) \leqslant \delta_c(m) + \inf_{\epsilon} \limsup_n m( X(\ell, \epsilon, n))\). Moreover, they also show that \(\delta_c\) is dominated by the packing dimension \(\text{Dim}^*\), namely, when we denote by \(\text{Dim}(E)\) the packing dimension of the set \(E\) and \[ \text{Dim}^*(\mu) = \inf \{ \text{Dim}(E) : \mu(E) =1 \}, \] if the \(\liminf\) in the definition (2) of \(\delta_c\) is a limit, then \(\delta_c(\mu) \leqslant \text{Dim}^*(\mu)\). In the second part of the paper, the rate of increase of optimal constants in an ad hoc notion of Bochner's inequality for measures is related with entropy dimension. A probability measure \(\mu\) is said to satisfy Bochner's inequality with constants \((n, K(n))\) \(\in\) \(( {\mathbb R}^+ )^2\) if for all smooth \(f\), \[ \mu( \Gamma_2(f,f) ) \geq \frac{1}{n} \mu( ( \varDelta f)^2) - K(n) \cdot \mu( \Gamma(f,f)), \tag{4} \] where \(\Gamma(f,f)\) (resp. \(\Gamma_2(f,f)\)) is the carré du champ (resp. carré du champ itéré) respectively. Intuitively, the parameter \(n\) can be regarded as the dimension of the support of \(\mu\) and \(K(n)\) as an estimate for the smallest eigenvalue of the Ricci curvature of the support, because for the case of \(\mu = \delta_x\), the classical Bochner inequality at the point \(x\) can be recovered with the dimension \(n\) of the manifold where \(x\) lives and a lower bound on the Ricci curvature, cf. \textit{D. Bakry} and \textit{Z. Qian} [Adv.\ Math.\ 155, No. 1, 98--153 (2000; Zbl 0980.58020)]. Actually, when one regards the microstates \[ \Gamma_N(\mu, \epsilon) = \left\{ x_1, \dots, x_N \in {\mathbb R}^N : d\left( \frac{1}{N} \sum_{i=1}^N \delta_{x_i}, \mu \right) < \epsilon \right\} \] as a submanifold of \({\mathbb R}^N\) with some dimension \([n N ]\) and Ricci curvature bounded below by \(- K(n)\), the passage to the limits \(N \to \infty\) and \(\epsilon \to 0\) gives the inequality (4). Now let us replace \(\mu\) with \(\mu_{\epsilon}\) \(=\) \(\mu* \nu_{\epsilon}\) and let \(K(n, \epsilon)\) (for a non-negative real number \(n\)) be the extended non-negative real number such that \(\mu_{\sqrt{ \epsilon}}\) satisfies Bochner's inequality (4) with constants \((n, K(n, \epsilon))\), and also define \[ \delta^{\square}(\mu) = 1 - \inf_{ n \geq 0} \left\{ \liminf_{\epsilon \to 0} \frac{ \int_{\epsilon}^1 K(n, y) d y}{ | \log \epsilon |} + 1 \right\} n. \tag{5} \] Then the second main theorem asserts that \(\delta^{\square} = \delta_c\). Finally, the authors study the free non-commutative analogue of the inequality (4), discuss the related free entropy dimension quantity, and show that it is less than or equal to the non-microstates free entropy dimension.