Sharp Hellinger rates for maximum likelihood estimation of Gaussian location mixtures
Year: 2026
Type: preprint
Abstract: We study nonparametric maximum likelihood estimation of a Gaussian location mixture in fixed dimension d with known covariance. When the mixing support is restricted to a known ball of positive radius, every measurable maximizer attains the minimax squared Hellinger rate n⁻¹(log n / log log n)ᵈ. When the likelihood is unrestricted and the mixing law has a fixed sub-Gaussian tail bound, the risk is O(n⁻¹(log n)ᵈ). Both conclusions allow approximate maximizers, with an additional risk of order the expected total log-likelihood deficit divided by n. The key estimate bounds the pointwise bracketing entropy of a polynomial unit ball by its dimension times O_d(1 + log(1/ε)), uniformly over measures with a finite, positive-definite moment matrix of the requisite order. We obtain it from a covering bound for normalized real polynomial maps and a Gaussian rectangle inequality. A finite family of positive mixtures gives the compact lower bound.
Author: Fomalhaut
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235843984
