Information geometry for approximate Bayesian computation

  • 2018-12-05 17:30:03
  • Konstantinos Spiliopoulos
  • 4

Abstract

The goal of this paper is to explore the basic Approximate BaeysianComputation (ABC) algorithm via the lens of information theory. ABC is a widelyused algorithm in cases where the likelihood of the data is hard to work withor intractable, but one can simulate from it. We use relative entropy ideas toanalyze the behavior of the algorithm as a function of the thresholdingparameter and of the size of the data. Relative entropy here is data driven asit depends on the values of the observed statistics. We allow differentthresholding parameters for each different direction (i.e. for differentobserved statistic) and compute the weighted effect on each direction. Thelatter allows to find important directions via sensitivity analysis leading topotentially larger acceptance regions, which in turn brings the computationalcost of the algorithm down for the same level of accuracy. In addition, we alsoinvestigate the bias of the estimators for generic observables as a function ofboth the thresholding parameters and the size of the data. Our analysisprovides error bounds on performance for positive tolerances and finite samplesizes. Simulation studies complement and illustrate the theoretical results.

 

Introduction (beta)

None

 

Conclusion (beta)

None