Maximum Likelihood Estimation for Totally Positive Log-Concave Densities

We study nonparametric maximum likelihood estimation for two classes of multivariate distributions that imply strong forms of positive dependence; namely log-supermodular (MTP) distributions and log--concave (LLC) distributions. In both cases we also assume log-concavity in order to ensure boundedness of the likelihood function. Given independent and identically distributed random vectors in from one of our distributions, the maximum likelihood estimator (MLE) exists a.s. and is unique a.e. with probability one when . This holds independently of the ambient dimension . We conjecture that the MLE is always the exponential of a tent function. We prove this result for samples in or in under MTP, and for samples in under LLC. Finally, we provide a conditional gradient algorithm for computing the maximum likelihood estimate.
View on arXiv