Identifiability for mixtures of centered Gaussians and sums of powers of quadratics

Abstract

We consider the inverse problem for the polynomial map which sends an $m$-tuple of quadratic forms in $n$ variables to the sum of their $d$-th powers. This map captures the moment problem for mixtures of $m$ centered $n$-variate Gaussians. In the first non-trivial case $d=3$, we show that for any $n\in \mathbb{N}$, this map is generically one-to-one (up to permutations of $q_1,\dots ,q_m$ and third roots of unity) in two ranges: $m\le \left(\genfrac{}{}{0ex}{}{n}{2}\right)+1$ for $n\le 16$ and $m\le \left(\genfrac{}{}{0ex}{}{n+5}{6}\right)/\left(\genfrac{}{}{0ex}{}{n+1}{2}\right)-\left(\genfrac{}{}{0ex}{}{n+1}{2}\right)-1$ for $n>16$, thus proving generic identifiability for mixtures of centered Gaussians from their (exact) moments of degree at most $6$. The first result is obtained by studying the explicit geometry of the tangential contact locus of the variety of sums of cubes of quadratic forms at concrete points, while the second result is accomplished using a link between secant non-defectivity with identifiability. The latter approach generalizes also to sums of $d$-th powers of $k$-forms for $d\ge 3$ and $k\ge 2$.

14 pages. Code for the base case computations can be found on GitHub

References in corpus (1)

• On the Expressive Power of Deep Polynomial Neural Networks

Cited by in corpus (2)

• Generalized identifiability of sums of squares
• Moment varieties of the inverse Gaussian and gamma distributions are nondefective