On Accelerated Mixing of the No-U-turn Sampler

  • 2025-07-17 16:09:04
  • Stefan Oberdörster
  • 0

Abstract

Recent progress on the theory of variational hypocoercivity established thatRandomized Hamiltonian Monte Carlo -- at criticality -- can achieve pronouncedacceleration in its convergence and hence sampling performance over diffusivedynamics. Manual critical tuning being unfeasible in practice has motivatedautomated algorithmic solutions, notably the No-U-turn Sampler. Beyond itsempirical success, a rigorous study of this method's ability to achieveaccelerated convergence has been missing. We initiate this investigationcombining a concentration of measure approach to examine the automatic tuningmechanism with a coupling based mixing analysis for Hamiltonian Monte Carlo. Incertain Gaussian target distributions, this yields a precise characterizationof the sampler's behavior resulting, in particular, in rigorous mixingguarantees describing the algorithm's ability and limitations in achievingaccelerated convergence.

 

Quick Read (beta)

loading the full paper ...