Skip to main content

Manon Michel: Using piecewise deterministic Markov processes in MCMC sampling

Time: Wed 2021-10-27 15.15 - 16.00

Location: Kräftriket, house 6, room 306

Lecturer: Manon Michel (CNRS, Université Clermont-Auvergne)


During this talk, I will discuss the main concepts and ideas that have been underlying and pushing the development of MCMC sampling through piecewise deterministic Markov processes (PDMP). First, I will introduce PDMP, which are composed of a random sequence of ballistic trajectories, and the associated formalism. I will then present the evolution and generalisation of PDMP-based MCMC methods since their first implementations in multiparticle systems. The general progress line can be framed into a constant effort of replacing the usually-enforced time reversibility, as done in detailed-balanced schemes, by symmetries of the sampled probability distribution itself. I will finally explain how PDMP can be used to reduce the computational complexity and how this reduction scheme can be extended to standard discrete-time Markov processes.