Tensor decomposition for learning Gaussian mixtures from moments (Q2133926)

Given a form \(T\) in some polynomial ring (over \(\mathbb R\)), the construction of a Waring decomposition of \(T\) is relevant to determine several properties of the form. The authors propose a new iterative method to determine a decomposition, which uses the Hankel operator \(H^{k,d-k}_T\) of \(T\). \(H^{k,d-k}_T\) is defined by sending a form \(p\) to the contraction by \(T\) of some derivatives of \(p\). The authors prove a criterion which guarantees the uniqueness of Waring decompositions of \(T\), based on the computation of the rank of \(H^{k,d-k}_T\). The criterion holds for moment forms built from spherical Gaussian mixtures. When the uniqueness holds, the authors propose an iterative method for the computation of the Waring decomposition, which uses the Singular Value Decomposition of \(H^{k,d-k}_T\) to determine a starting point of the algorithm. The authors evaluate, in specific examples, the efficiency of the new method in comparison with other state-of-the-art approaches.