Ensemble Markov Chain Monte Carlo with Teleporting Walkers

We introduce an ensemble Markov chain Monte Carlo approach to sampling from a probability density with known likelihood. This method upgrades underlying by allowing of such chains interact via process in which one chain’s state is cloned as another’s deleted. effective teleportation states can overcome issues metastability the chain, scheme enjoys rapid mixing once modes target have been populated. derive mean-field limit for evolution ensemble. analyze global and local convergence this limit, showing asymptotic independent spectral gap moreover we interpret limiting gradient flow. explain how interaction be applied selectively subset variables order maintain advantage on very high-dimensional problems. Finally, present application our methodology Bayesian hyperparameter estimation Gaussian regression.