74
21
v1v2v3 (latest)

Maximum Likelihood Estimation for Totally Positive Log-Concave Densities

Abstract

We study nonparametric maximum likelihood estimation for two classes of multivariate distributions that imply strong forms of positive dependence; namely log-supermodular (MTP2_2) distributions and log-L#L^\#-concave (LLC) distributions. In both cases we also assume log-concavity in order to ensure boundedness of the likelihood function. Given nn independent and identically distributed random vectors in Rd\mathbb R^d from one of our distributions, the maximum likelihood estimator (MLE) exists a.s. and is unique a.e. with probability one when n3n\geq 3. This holds independently of the ambient dimension dd. We conjecture that the MLE is always the exponential of a tent function. We prove this result for samples in {0,1}d\{0,1\}^d or in R2\mathbb{R}^2 under MTP2_2, and for samples in Qd\mathbb{Q}^d under LLC. Finally, we provide a conditional gradient algorithm for computing the maximum likelihood estimate.

View on arXiv
Comments on this paper