Quantitative non-geometric convergence bounds for independence samplers (Q539523)

The authors derive upper and lower bounds for the speed of convergence of independence sampler MCMC algorithms. In particular, their results apply for chains that are not geometrically ergodic. In this case, no quantitative bounds were previously known. The authors define convergence time as the smallest integer \(n\) such that the total variation distance to stationarity after \(n\) steps is less than \(0.01\). Their main results are an upper and a lower bound for \(n\). In order to prove the results, a coupling technique is employed. The results are applied to three test cases which allow to draw qualitative conclusions. Firstly, convergence can be very slow even for seemingly very simple Markov chains. Secondly, slight changes in the parameters can have enormous effects on the convergence time, i.e., changing from between 24 and 50 iterates to between \(4\cdot10^9\) and \(14\cdot10^9\) iterates or from between \(4\cdot10^3\) and \(8\cdot10^3\) to \(5\cdot10^{32}\) and \(10^{34}\) in their examples. Here, the numbers correspond to the derived upper and lower bounds on the time to convergence. Finally, the authors note that their results while applicable to certain cases of independence samplers are not expected to provide useful results in all situations.