Local Kernel Dimension Reduction in Approximate Bayesian Computation

TL;DR

This paper introduces Local Gradient Kernel Dimension Reduction (LGKDR), a novel method for constructing low-dimensional, sufficient summary statistics in Approximate Bayesian Computation, improving efficiency especially in sequential Monte Carlo methods.

Contribution

LGKDR implicitly considers all nonlinear transformations of summary statistics to identify a sufficient subspace without strong distributional assumptions.

Findings

LGKDR performs competitively with simple rejection ABC.

LGKDR significantly improves results in sequential Monte Carlo ABC.

Method effectively reduces Monte Carlo errors.

Abstract

Approximate Bayesian Computation (ABC) is a popular sampling method in applications involving intractable likelihood functions. Without evaluating the likelihood function, ABC approximates the posterior distribution by the set of accepted samples which are simulated with parameters drawn from the prior distribution, where acceptance is determined by the distance between the summary statistics of the sample and the observation. The sufficiency and dimensionality of the summary statistics play a central role in the application of ABC. This paper proposes Local Gradient Kernel Dimension Reduction (LGKDR) to construct low dimensional summary statistics for ABC. The proposed method identifies a sufficient subspace of the original summary statistics by implicitly considers all nonlinear transforms therein, and a weighting kernel is used for the concentration of the projections. No strong assumptions are made on the marginal distributions nor the regression model, permitting usage in a wide range of applications. Experiments are done with both simple rejection ABC and sequential Monte Carlo ABC methods. Results are reported as competitive in the former and substantially better in the latter cases in which Monte Carlo errors are compressed as much as possible.