On the hardness of deciding the finite convergence of Lasserre hierarchies

Abstract: A polynomial optimization problem (POP) asks for minimizing a polynomial function given a finite set of polynomial constraints (equations and inequalities). This problem is well-known to be hard in general, as it encodes many hard combinatorial problems. The Lasserre hierarchy is a sequence of semidefinite relaxations for solving (POP). Under the standard archimedean condition, this hierarchy is guaranteed to converge asymptotically to the optimal value of (POP) (Lasserre, 2001) and, moreover, finite convergence holds generically (Nie, 2012). In this paper, we aim to investigate whether there is an efficient algorithmic procedure to decide whether the Lasserre hierarchy of (POP) has finite convergence. We show that unless P=NP there cannot exist such an algorithmic procedure that runs in polynomial time. We show this already for the standard quadratic programs. Our approach relies on characterizing when finite convergence holds for the so-called Motzkin-Straus formulation (and some variations of it) for the stability number of a graph.