On the minimax-rate optimality of approximate Bayesian computation in nonparametric problems

Abstract

Approximate Bayesian computation (ABC) replaces likelihood evaluation in conventional Bayesian computation by simulation and comparison of observed and synthetic data. We demonstrate that ABC can be minimax-rate optimal in nonparametric settings. Our main result is a general contraction theorem for ABC posteriors based on summary statistics sieves and a localized prior-mass condition. We apply this theorem to Gaussian sequence estimation over Sobolev ellipsoids and to density estimation over bounded Sobolev-type classes and, under model-specific conditions, construct ABC procedures whose ideal posteriors and posterior means attain the corresponding minimax rates. Conditional on sampling from the specified priors, we show that the conventional Monte Carlo rejection ABC algorithm inherits the same rates when the number of simulation proposals grows at a sufficiently large exponential rate in the effective dimension.