Local and Global Contraction Principles for MCMC Mixing
Researchers have developed a new framework for analyzing Markov chain Monte Carlo (MCMC) algorithms, focusing on contraction principles. This framework utilizes global and local contraction coefficients under the Eγ-divergence to prove mixing-time bounds. The approach offers direct proofs of exponential convergence for projected Langevin Monte Carlo and provides warm-start convergence bounds for Metropolis-Hastings algorithms, even in heavy-tailed regimes. AI
IMPACT This research introduces a novel theoretical framework for analyzing MCMC algorithms, potentially improving their efficiency and applicability in various AI domains.