Geometric analysis for the Metropolis algorithm on Lipschitz domains (Q636824)

The authors consider the convergence of the Metropolis algorithm to stationary distributions on bounded, connected Lipschitz domains in Euclidean space. To this end they employ a geometrical analysis and present tools that can be used to prove the convergence of the algorithm. The authors derive bounds on the rates of convergence of the Metropolis algorithm in terms of the spectral gap of the associated Metropolis operator. The behaviour of the spectral gap is analysed in more detail; a comparison, and Nash and Sobolev inequalities are presented. The general results are applied to the original application of the Metropolis algorithm: The random, non-overlapping placement in the \(d\)-dimensional unit square of \(N\) \(d\)-dimensional spheres with fixed radius \(\varepsilon\) such that \(N\varepsilon\) is sufficiently small. The resulting stationary distribution is the uniform distribution on the set of centres of all possible sphere placements. It is shown that the state space is a Lipschitz domain and that the general results hold in this example.