The sharp 3k-moment threshold for rational identifiability of Gaussian, gamma and inverse Gaussian mixtures
- نُشر
- الخادم
- Zenodo
- DOI
- 10.5281/zenodo.24432966
For every k ≥ 2, exactly 3k consecutive initial raw moments are needed to rationally identify over the complex numbers the unordered parameters of a generic mixture of k univariate Gaussians, gamma distributions or inverse Gaussians. This proves the rational-identifiability assertions conjectured by Améndola, Faugère and Sturmfels and by Henriksson, Seccia and Yu, and establishes sharpness. Earlier work gave the general upper bound 3k + 2. Geometrically, the unordered secant parametrization is birational onto the corresponding secant hypersurface in projective space of dimension 3k; in projective space of dimension 3k − 1 it is generically finite of degree greater than one. The same sharp threshold holds for negative binomial mixtures with freely varying shapes. For sufficiency, exponential tilting makes each cumulant the derivative of the previous one. A lemma on iterated derivatives then shows that the first 3k moments generate the field of all moments. For necessity, we send one component to infinity while its weight tends to zero and prove that every regular decomposition of the limiting system persists. This construction also gives product lower bounds for the generic numbers of complex decompositions. Exact rational certificates and a real version of the construction yield, for every k ≥ 2 in all four families, a nonempty Euclidean open set of probability mixtures with a different probability mixture sharing their first 3k − 1 moments.