Identifiability of Finite Mixtures of Elliptical Distributions

Citation

• Authors: Hajo Holzmann, Axel Munk, Tilmann Gneiting
• Year: 2006
• Venue: Scandinavian Journal of Statistics 33(4), 753-763
• arXiv: None
• URL: https://doi.org/10.1111/j.1467-9469.2006.00505.x

One-Line Takeaway

Holzmann, Munk, and Gneiting extend classical Gaussian-mixture identifiability to broad multivariate elliptical families, including cases with varying shape parameters, location vectors, and scatter matrices.

Technical Summary

• Setup: the component family is elliptical on $\mathbb R^p$, with densities of the form $$f_{\theta,p}(x)=|\Sigma|^{-1/2}\,f_p\!\left((x-\mu)^\top \Sigma^{-1}(x-\mu);k\right),$$ where location $\mu$, scatter $\Sigma$, and an additional shape parameter $k$ may all vary across mixture components.
• Transform-based criterion: the paper gives a general identifiability theorem for parametric families of characteristic generators. If there is a total ordering on the relevant parameter pairs and, whenever $\theta_1\prec\theta_2$, the ratio of the corresponding characteristic generators tends to $0$ in the tail, then finite mixtures over that family are identifiable.
• Density-generator criterion: an analogous theorem is stated directly in terms of density generators, again requiring an ordering for which the tail ratio of the "smaller" generator to the "larger" one vanishes; this prevents different components from matching the same radial profile asymptotically.
• Scope: the criteria cover broad multivariate families beyond Gaussians, including multivariate $t$, symmetric stable, exponential-power, Kotz, and some normal scale-mixture classes, while permitting the shape parameter to vary in the mixture.

Assumptions And Limits

The result does not say that every elliptical family is identifiable; the generator conditions matter. The paper is about population identifiability, not finite-sample computational recovery or robustness near degenerate parameter configurations.

Relevance To This Project

This paper shows that identifiability can survive fairly rich component classes when there is still enough analytic rigidity in the family. For the repo's latent-strategy questions, that is the useful middle ground between overly narrow classical mixtures and expressive mixture-closed component classes: one can sometimes have nontrivial latent structure without giving up identifiability entirely.

Comparison Notes

Compared with Teicher (1963), this is a modern extension from classical one-parameter families to richer multivariate elliptical families. Compared with Hsu and Kakade (2013), the contribution is identifiability of the model class itself rather than an efficient recovery algorithm.

Related Notes

• Background on identification and learning of mixture distributions
• Identifiability of Finite Mixtures
• Learning mixtures of spherical Gaussians: moment methods and spectral decompositions