A pathological MCMC algorithm and its use as a benchmark for convergence assessment technique chains (Q1297845)

The author examines the behavior of a particular Metropolis-type algorithm for simulation of a beta \(Be(\alpha,\beta)\) variable. It occurs that this algorithm which is rather simple to describe (and to program) is unimodal and unidimensional. It does not require extreme parameter values, has known stationary distribution \(Be(\alpha,1)\), but is not seen to converge after millions of iterations. Thus this algorithm can be used as an ultimate benchmark for testing standard or new convergence control techniques and other assessments of Markov chain Monte Carlo (MCMC) algorithms. This is illustrated by certain examples. Comparison with other criteria is presented.