Importance sampling is a mainstay of scientific computing. Often, one tries to choose a tractable proposal distribution to make the variance of the estimate small. These are well known “long tailed” problems and the variance is sometimes a bad idea. We have a usable alternative criteria that is necessary and sufficient for convergence. It also offers a simple alternative to things like “effective sample size” which is notoriously unreliable. Minimizing the new criteria suggests new problems for the many good ideas that have been developed classically.
This is joint work with Sourav Chatterjee.
The Statistics Seminars are held in Sequoia Hall, Room 200, at 4:30pm on Tuesdays.
Refreshments are served at 4pm in the Lounge on the first floor.