Learning $k$-Modal Distributions via Testing
Abstract
A -modal probability distribution over the discrete domain is one whose histogram has at most "peaks" and "valleys." Such distributions are natural generalizations of monotone () and unimodal () probability distributions, which have been intensively studied in probability theory and statistics. In this paper we consider the problem of \emph{learning} (i.e., performing density estimation of) an unknown -modal distribution with respect to the distance. The learning algorithm is given access to independent samples drawn from an unknown -modal distribution , and it must output a hypothesis distribution such that with high probability the total variation distance between and is at most Our main goal is to obtain \emph{computationally efficient} algorithms for this problem that use (close to) an information-theoretically optimal number of samples. We give an efficient algorithm for this problem that runs in time . For , the number of samples used by our algorithm is very close (within an factor) to being information-theoretically optimal. Prior to this work computationally efficient algorithms were known only for the cases \cite{Birge:87b,Birge:97}. A novel feature of our approach is that our learning algorithm crucially uses a new algorithm for \emph{property testing of probability distributions} as a key subroutine. The learning algorithm uses the property tester to efficiently decompose the -modal distribution into (near-)monotone distributions, which are easier to learn.
Cite
@article{arxiv.1107.2700,
title = {Learning $k$-Modal Distributions via Testing},
author = {Constantinos Daskalakis and Ilias Diakonikolas and Rocco A. Servedio},
journal= {arXiv preprint arXiv:1107.2700},
year = {2014}
}
Comments
28 pages, full version of SODA'12 paper, to appear in Theory of Computing