Core Concepts
The h-lifted Kullback-Leibler (KL) divergence can be used to obtain O(1/√n) convergence rates for mixture density estimation on compact domains, without requiring the densities to be strictly positive.
Abstract
The key points of the content are:
The authors introduce the h-lifted Kullback-Leibler (KL) divergence as a generalization of the standard KL divergence, which allows for the analysis of density estimation problems where the target densities are not strictly positive.
Under the assumption that the target density f and the component densities φ(·; θ) are bounded above by constants c and b, respectively, and the lifting function h is bounded above and below by constants b and a, the authors prove an O(1/√n) bound on the expected h-lifted KL divergence between the estimated mixture density fk,n and the true density f.
This result extends the previous work of Li and Barron (1999) and Rakhlin et al. (2005), which required the densities to be strictly positive.
The authors develop a procedure for computing the corresponding maximum h-lifted likelihood estimators (h-MLLEs) using the Majorization-Maximization framework and provide experimental results supporting their theoretical bounds.
The h-lifted KL divergence is shown to be a Bregman divergence, which allows the authors to leverage existing results on greedy approximation sequences and Rademacher complexities to derive the convergence rates.
The key technical contributions include a uniform concentration bound for the h-lifted log-likelihood ratios and the use of bracketing numbers to bound the covering numbers of the component density class.
Stats
The following sentences contain key metrics or figures:
The authors prove an O(1/√n) bound on the expected estimation error when using the h-lifted KL divergence.
The authors assume that the target density f and the component densities φ(·; θ) are bounded above by constants c and b, respectively, and the lifting function h is bounded above and below by constants b and a.