Markov chain Monte Carlo (MCMC) methods are widely used for sampling from complex distributions, and practitioners typically run diagnostics to check whether the chain has converged. A new paper on arXiv proves that this practice has a fundamental limitation: no finite checklist of diagnostics can certify convergence.

Drawing on the long-standing folklore that passing convergence diagnostics does not guarantee convergence, the authors show this is an impossibility for any finite list of tests. In other words, for any finite set of diagnostics, there exists a Markov chain that passes all of them yet has not converged to the target distribution.

The abstract begins to note that running more chains and longer chains are among the strategies considered, but the sentence is cut off in the source. The result underscores that convergence assessment in MCMC remains inherently heuristic.