Isoperimetry for Gibbs measures (Q1872206)

This very clearly written article builds up on earlier work of \textit{S. G. Bobkov} [Ann. Probab. 25, 206-214 (1997; Zbl 0883.60031)] and \textit{D. Bakry} and \textit{M. Ledoux} [Invent. Math. 123, 259-281 (1996; Zbl 0855.58011)]. In both of these previous articles, the following functional inequality was considered: \[ \mathcal{I} (\mu f)\leq \mu\left(\mathcal{I}^2 (f) + C\cdot |\nabla f|^2\right)^{1/2} , \] where \(\mathcal{I} = \gamma\circ\Gamma^{-1}\) is the Lévy-Gromov isoperimetric function of a standard Gaussian distribution \(\Gamma\) with density \(\gamma\), and \(\mu F\) denotes the expectation of a function \(F\) with respect to a probability measure \(\mu\). Using the fact that this inequality may be ``tensorised'' then leads to similar statements for infinite-dimensional tensor products of the original measure \(\mu\). The main aim of the present article is to investigate the validity of some more general inequalities in an infinite-dimensional setting; these inequalities are more general in the sense that an arbitrary concave, nonnegative function \(\mathcal{U}\) is used instead of \(\mathcal{I}\), while the probability measure \(m\) under consideration need not be an infinite-dimensional tensor product. It is established, both in the discrete setting of Bobkov (loc. cit.) and in the Riemannian setting of Bakry and Ledoux (loc. cit.), that \(m\) satisfies such an inequality as soon as its finite-dimensional conditional expectations do and some regularity and mixing conditions are met. Considering specifically Gibbs measures, the author then shows that such assumptions are fulfilled in the Dobrushin-Shlosman regime. In summary, these results may be used in order to establish isoperimetric and log-Sobolev inequalities for Gibbs measures, but also for other types of (non-product) probability measures in infinite dimensions, e.g. in a Markov chain context. Such methods could also be used in order to seek for better constants in certain known functional inequalities.