Elementary bounds on Poincaré and log-Sobolev constants for decomposable Markov chains (Q1769410)

The paper considers finite-state Markov chains that can be naturally decomposed into smaller (sub)chains, called ``restriction'' and ``projection'' ones. The state-space partition of the original Markov chains induces a number of restriction chains, in which transitions are restricted to occur within blocks of the partition, and a projection chain, whose states are the blocks themselves. The authors obtain an elementary, self-contained basic decomposition result which, in the context of inductively defined Markov chains, is able to provide inverse polynomial bounds on the Poincaré constant \(\lambda\) for the corresponding Poincaré inequality. Expressions for Poincaré constants of the initial Markov chain are given in terms of Poincaré constants of the projection and restriction chains, together with an additional parameter. The Markov chain decomposition result for the Poincaré constant is shown to carry over directly to the log-Sobolev inequality constant \(\alpha\), being obtained for the first time a general decomposition for log-Sobolev inequalities, with improved, tighter bounds of the Poincaré and log-Sobolev inequality constants (and by consequence, on the mixing time of the Markov chains). The finite-state Markov chain results, with potential applications to Markov chain Monte Carlo (MCMC) algorithms, can be naturally extended to Markov chain frameworks based on countably (or even uncountably) infinite state spaces.