Identifiability from Moments in Independent Component Analysis Mixture Models

Abstract

We study Independent Component Analysis (ICA) mixture models. These models consist of mixtures of ICA densities. An ICA density is associated with a random vector whose coordinates become mutually independent after an invertible linear transformation. ICA mixture models are nonparametric extensions of the celebrated family of Gaussian mixtures. We show that in many cases the ICA mixtures are identifiable from the collection of the observed moment tensors, that is, the mixing coefficients, linear transformations, and source moments of order up to $d$ can be identified from the moment tensors of order up to $d$ for generic parameters. We address the problem by studying the corresponding secant varieties and their algebraic properties such as dimensions and Gauss maps. In particular, for $n$-variate distributions, we establish that generically and up to the trivial scaling and permutation symmetries, when $d\geq 5$, the parameters can be recovered uniquely for asymptotically up to $n^{d-2}/d!$ mixture components. For $d=3,4$, we show identifiability up to a finite equivalence class for asymptotically up to $n^{d-2}/d!$ and $0.5 n^{d-2}/d!$ components respectively, where for $d=4$ the mixing coefficients need to be known a priori. The obtained bounds on the component numbers have the maximum possible orders in $n$, and for $d = 3$ and $d\geq 5$, they are also asymptotically sharp.