Complete quadrics: Schubert calculus for Gaussian models and semidefinite programming (Q6566460)

This paper is concerned with the study of inverse varieties of linear spaces of symmetric matrices (called linear concentration models): this is obtained by inverting all the matrices belonging to the linear space, and then by taking the Zariski closure of the set of all such inverses. In particular, the paper is interested in the characterization of the degrees of such inverse varieties.\N\NThe main motivation comes from the maximum likelihood estimation in statistics, where one is usually interested in maximising products of densities of Gaussians under positive definite constraints restricted to inverse images of linear spaces of symmetric matrices. The main result is the solution to a conjecture of Sturmfels and Uhler about closed formulas for the ML-degree of the inverse variety; the result states that such degree is polynomial on the size of the matrices.\N\NThe interesting part is the relation with semidefinite programming (SDP), a natural generalization of linear programming to the space of real symmetric matrices. The ML-degrees and algebraic degrees of SDP are related and can be computed through intersection theory (Prop. 3.5 and Cor. 3.6). The main result of the paper proves both polynomiality of the two degrees with respect to the matrix size.