Séminaire SPACE Tours
Convergence of non-reversible Markov processes via lifting and flow Poincaré inequalityLéo Hahn (ETH Zurich)
Friday 27 November 2026 10:30 - Tours - E2 1180
Résumé :
Markov chain Monte Carlo offers a powerful way to approximate integrals by averaging a function along a random trajectory. Widely used in Bayesian statistics and statistical mechanics, it is particularly valuable in high dimensions, where deterministic integration can become prohibitively expensive. In this talk, I will explain how to compare the effectiveness of different Markov chain Monte Carlo methods by addressing the question: how quickly does each method converge to the target distribution?
The starting point will be two widely used and closely related processes: overdamped and kinetic Langevin dynamics. Practitioners have long observed that kinetic Langevin can converge faster than overdamped Langevin, as its more persistent motion allows it to explore the state space more efficiently. More recently, Cao, Lu and Wang (ARMA, 2023) established quantitative bounds capturing this diffusive-to-ballistic speed-up for convex potentials.
I will explain the ideas behind this acceleration and present a framework extending it to a broad class of non-reversible Markov processes. Two ingredients play a central role. The first is the concept of second-order lifts, which relate dynamics on an enlarged state space to a simpler reference process. The second is the flow Poincaré inequality, which controls convergence through time-averaged dissipation, as opposed to the instantaneous dissipation estimate provided by the classical Poincaré inequality.
This talk is based on the article “Convergence of non-reversible Markov processes via lifting and flow Poincaré inequality”, joint work with Andreas Eberle, Arnaud Guillin, Francis Lörler and Manon Michel.
Liens :
