Optimizing the asymptotic convergence rate of the Diaconis-Holmes-Neal sampler (Q2455515)

A generalization of Diaconsis-Holmes-Neal Markov chain Monte Carlo sampler is considered. This sampler produces a non-reversible chain which achieves the target uniform distribution on a \(n\)-states space by \(O(n)\) steps, while the standard (reversible) Metropolis chain takes \(O(n^2)\) steps. The authors investigate the reduced spectral radius (the second largest eigenvalue modulus) of the transition matrix of the chain. For the one parameter sampler the value of parameter is derived at which the reduced spectral radius attains its minimum and hence the sampler achieves the optimal convergence rate. The authors argue that the optimization in an \(n\)-parametric model can't improve the convergence, at least locally.