Modified logarithmic Sobolev inequalities for some models of random walk (Q2485798)

The aim of this paper is to investigate distinct logarithmic (log) Sobolev inequalities for the special case of Markov chain convergence in discrete setting formulations, namely for the models of random walk. Section 2 introduces the notations and preliminary results relating log Sobolev inequalities for estimating rates of convergence of Markov chains to their stationary distributions. As an example, there are discussed previous estimates for modified log Sobolev inequalities on the asymmetric 2-point space. Section 3 presents the main results of this paper: modified logarithmic Sobolev inequalities for some models of random walk, including the random transposition shuffle and the top-random transposition shuffle on the symmetric group \(S_n\), and the shuffle generated by 3-cycles on the alternating group \(A_n\). As an application of these results, sharp bounds on the rates of convergence are derived. This section shows also that a generic \(r\)-regular graph has the modified log Sobolev constant much smaller than its spectral gap in the convergence results based on the classical Fourier analysis. Finally, the modified log Sobolev inequality for the random transposition model is examined. Section 4 outlines the method of comparison chains to device modified log Sobolev inequalities, and analyzes a perturbation of the top-random transposition shuffle that cannot be realized as a walk on a group, making it difficult to be studied by other methods. Section 5 presents the well-known connection between log Sobolev and concentration inequalities, and illustrates this relationship by some examples based on the inequalities derived in Section 3. A special useful remark is that modified log Sobolev inequalities can give better results than the classical log Sobolev inequality when applied to non-diffusion Dirichlet forms.