On the nonparametric maximum likelihood estimator for Gaussian location mixture densities with application to Gaussian denoising (Q2196192)

The article discusses thoroughly the nonparametric maximum likelihood estimation of Gaussian mixture densities from independent multi-dimensional observations. The substantial difference compared to ordinary likelihood methods is that convex optimization is employed for fitting mixtures. Finite sample accuracy results (based on the squared Hellinger loss function) are quantified and found to be near parametric in precision. The article concludes with the application and assessment (at nearly optimal level) of the estimates developed herein, as empirical Bayesian estimates of the oracle estimator in the Gaussian denoising problem.