The Limit of Convexity Based Isoperimetry: Sampling Harmonic-Concave Functions

TL;DR

This paper extends the class of functions with good isoperimetry beyond logconcave functions to (1/(n-1))-harmonic concave functions, providing new theoretical insights and efficient sampling algorithms for these functions, including the Cauchy distribution.

Contribution

It characterizes (1/(n-1))-harmonic concave functions as the largest class with guaranteed isoperimetry and develops the first efficient sampling algorithms for this class.

Findings

Characterization of harmonic-concave functions with good isoperimetry

Development of efficient sampling algorithms for these functions

Matching mixing time bounds for Cauchy distribution sampling

Abstract

Logconcave functions represent the current frontier of efficient algorithms for sampling, optimization and integration in R^n. Efficient sampling algorithms to sample according to a probability density (to which the other two problems can be reduced) relies on good isoperimetry which is known to hold for arbitrary logconcave densities. In this paper, we extend this frontier in two ways: first, we characterize convexity-like conditions that imply good isoperimetry, i.e., what condition on function values along every line guarantees good isoperimetry? The answer turns out to be the set of (1/(n-1))-harmonic concave functions in R^n; we also prove that this is the best possible characterization along every line, of functions having good isoperimetry. Next, we give the first efficient algorithm for sampling according to such functions with complexity depending on a smoothness parameter. Further, noting that the multivariate Cauchy density is an important distribution in this class, we exploit certain properties of the Cauchy density to give an efficient sampling algorithm based on random walks with a mixing time that matches the current best bounds known for sampling logconcave functions.