Curvature, concentration and error estimates for Markov chain Monte Carlo (Q606638): Difference between revisions

This interesting work contains the explicit nonasymptotic estimates for the convergence rate of Markov chains empirical means, together with a Gaussian or exponential control on the empirical means deviations. The main assumption can be seen geometrically as a ``positive Ricci curvature'' property of the Markov chain. The authors use this in the form of constraint on the Wasserstein distance between two probability measures from the associated transition kernel: \(W_1(P_x,P_y)\leq (1-k)d(x,y)\), where \(k>0\) and \(x, y\) belong to a Polish state space with metric \(d\). There are many situations which satisfy this assumption. The authors concretize the following: a heat bath for the Ising model; the \(M/M/\infty\) queueing process; Euler scheme for diffusions; the diffusions on positively curved manifolds; and a certain nonlinear state space model.