Skip to main content
Browse by:
GROUP

Until further notice, in-person public events have been canceled. This includes recruitment events, tours, student programs, reunions, performances, conferences and social events.
Event listings include how to access online content. Contact event sponsor with questions.
Please note that all visitors to campus must comply with Duke’s community safety measures, which include wearing a mask,
check https://returnto.duke.edu/campus-visitors/ before coming to campus, and follow direction provided by campus personnel.

Sampling Multi-modal Distributions using Simulated Tempering Langevin Monte Carlo

Event Image
Icon calendar
Thursday, February 14, 2019
Icon time
3:15 pm - 4:15 pm
Icon speaker
Holden Lee (Princeton University)
Icon series
Probability Seminar

A fundamental problem in Bayesian statistics is sampling from distributions that are only specified up to a partition function (constant of proportionality). In particular, we consider the problem of sampling from a distribution given access to the gradient of the log-pdf. For log-concave distributions, classical results due to Bakry and Emery show that natural continuous-time Markov chains called Langevin diffusions mix in polynomial time. But in practice, distributions are often multi-modal and hence non-log-concave, and can take exponential time to mix.

We address this problem by combining Langevin diffusion with simulated tempering. The result is a Markov chain that mixes in polynomial rather than exponential time by transitioning between different temperatures of the distribution. We prove fast mixing for any distribution that is close to a mixture of gaussians of equal variance.

For the analysis, we bound the spectral gap using a novel Markov chain decomposition theorem. Previous approaches rely on decomposing the state space as a partition of sets, while our approach can be thought of as decomposing the stationary measure as a mixture of distributions (a "soft partition").

Based on the paper https://arxiv.org/abs/1812.00793

Joint work with Rong Ge (Duke) and Andrej Risteski (MIT).