{"id":"arxiv/2608.21466","url":"https://arxiv.org/abs/2608.21466","tags":["stat.ML","cs.IT","cs.LG","math.OC","math.PR","stat.CO"],"score":0,"title":"Spectral partitioning for $k$-block averaging kernels of finite Markov chains","domain":"stat.ML","source":"arxiv","authors":["Michael C. H. Choi","Youjia Wang"],"reviews":0,"subject":"stat.ML","summary":"We develop spectral algorithms for selecting state-space partitions that define averaging kernels for finite, ergodic and reversible Markov chains. For a partition $\\mathcal O$, the Gibbs kernel $G_{\\mathcal O}$ resamples within the current block from the stationary conditional distribution; when this update is tractable, composing or mixing it with a baseline kernel $P$ can accelerate convergence. We select $\\mathcal O$ by rounding the bottom nonconstant eigenfunctions of $P^2$, or the algebraically smallest eigenfunctions of $P$ for additive mixtures, using weighted $k$-means. For…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-08-21","reactions":{}}