Spectral partitioning speeds up convergence in finite Markov chains, paper details algorithm

michaelchchoi · x · 2026-08-25

This paper develops spectral algorithms for selecting state-space partitions that define averaging kernels for finite, ergodic and reversible Markov chains. It selects partitions by rounding the bottom nonconstant eigenfunctions of the transition matrix using weighted k-means. The authors derive exact trace and normalized-cut representations, showing the objective equals the Pearson χ² mutual information.

Original post →

More from Research

Research channel →