Gaussian Mixture Models and Characteristic Function Estimation

Fourier-based density estimation, empirical characteristic function methods, and mixture models for neuroimaging

2026-06-07 18:16 PDT

Overview

Gaussian mixture models (GMMs) and their estimation occupy a long thread in this research program, beginning with thesis work in 1982 on Fourier-based decomposition of two-component normal mixtures and continuing through current work on empirical characteristic function (ECF) methods, with applied extensions to Alzheimer’s disease neuroimaging data.

The unifying methodological challenge is parameter estimation for normal mixture distributions when the likelihood surface is pathological. Maximum likelihood estimation of normal mixture parameters is undermined by an infinite spike in the likelihood at each sample point – a well-known theoretical flaw that motivates alternative estimation strategies. The Fourier cosine series and the empirical characteristic function provide principled alternatives that avoid the degenerate likelihood surface while achieving competitive efficiency.

Background: the 1982 Fourier approach

The original thesis document, Parametric Estimation for Normal Mixtures Based on Fourier Density Estimation Methods, addressed the problem of decomposing a two-component normal mixture

f(x) = \frac{p}{\sqrt{2\pi\sigma_1^2}}\,e^{-\frac{(x-\mu_1)^2}{2\sigma_1^2}} + \frac{1-p}{\sqrt{2\pi\sigma_2^2}}\,e^{-\frac{(x-\mu_2)^2}{2\sigma_2^2}}

estimating the mixing proportion p and the four component parameters. The approach represents the density as a truncated Fourier cosine series, following Kronmal (1966), and matches the empirical cosine coefficients \hat{A}_k to the theoretical coefficients implied by the mixture model. This avoids the degenerate likelihood and outperforms both Pearson’s method of moments and the Quandt-Ramsey moment generating function method in simulation.

Papers in development

  • Empirical characteristic function methods for parametric estimation: analytic evaluation-point selection, closed-form estimators for normal and exponential families, and extension to stable distributions.

  • Characteristic function distances as an alternative to Jensen-Shannon divergence in generative adversarial networks.

  • Normal mixture models applied to MRI-derived distributions in Alzheimer’s disease: decomposing multimodal brain-volume and cortical-thickness distributions across aging and disease stages.

  • Research opportunities in Fourier-based mixture methods for contemporary machine learning and time-series applications.

Methods

Fourier cosine series density estimation (Kronmal 1966); empirical characteristic function estimation with analytic evaluation-point selection; minimum-distance estimation for stable distributions; normal and exponential closed-form ECF estimators; Monte Carlo simulation for bias and efficiency assessment; Gaussian mixture models via mclust for neuroimaging data; characteristic function GAN loss functions.

Publications

Early work on characteristic function and density estimation methods is accessible through the full publications list by filtering on biostatistics or statistical-computing.