Faster log-concave sampling via algorithmic warm starts

The problem of sampling from a log-concave distribution is a key algorithmic component of fields such as Bayesian inference, yet non-asymptotic computational guarantees for this task have only emerged recently, within the last decade. In this talk, I’ll discuss recent progress on understanding one of the most popular samplers, the Metropolis-adjusted Langevin algorithm (MALA), by first showing refined mixing time bounds under a warm start, and then showing how to algorithmically obtain the warm start via the underdamped Langevin process.

• Speaker: Sinho Chewi (Institute for Advanced Study)
• Friday 06 October 2023, 14:00–15:00
• Venue: MR12, Centre for Mathematical Sciences.
• Series: Statistics; organiser: Qingyuan Zhao.