Spring Seminar Series - February 6, 2003
University of Minnesota
School of Statistics
College of Liberal Arts
Statistical Models for Monte Carlo Integration
Zhiqiang Tan
Statistics Department
The University of Chicago
Thursday, February 6, 2003
4:00 PM, 115
Ford Hall
Minneapolis, East Bank Campus
Social at 3:30 PM, 300
Ford Hall
Abstract
Nontrivial integration arises in many statistical analyses, like evaluating
the likelihood for spatial models, mixed-effect models, missing data problems,
and genetic pedigree analysis, and calculating the Bayes factor for Bayesian
model selection. In this talk, I will show a novel approach which formulates
Monte Carlo integration as a statistical model using simulated observations
as data, and demonstrate the soundness and usefulness of the approach.
The key idea of the approach is to ignore part of the information about
the baseline measure and treat the measure as a parameter in a semiparametric
model, which is estimated by maximum likelihood. The integrals of interest
are estimated as linear functionals. This formulation is of theoretical
importance because it provides a unified framework to study Monte Carlo
estimation procedures. I will show several results about the maximum likelihood
estimator and its comparison to other estimators. The formulation is also
of practical importance because it provides an effective way to construct
new Monte Carlo estimators. I will show a new method to estimate integrals
by iterative simulation. For integrands proportional to the stationary density,
the estimator converges at a rate faster than the square root rate.