Exact polynomial optimization strengthened with Fritz John conditions

TL;DR

This paper introduces a semidefinite programming approach that finitely converges to the exact minimum of polynomial optimization problems, leveraging Fritz John conditions and critical point analysis.

Contribution

It develops a novel method combining Fritz John conditions with semidefinite programming to exactly solve polynomial optimization problems over basic semi-algebraic sets.

Findings

Sequence of values from semidefinite programs converges finitely to the minimum.

Exact minimum can be computed for polynomials over standard sets like the unit ball, hypercube, and simplex.

Method extends to convex semi-algebraic sets with non-empty interior.

Abstract

Let $f,g_1,\dots,g_m$ be polynomials with real coefficients in a vector of variables $x=(x_1,\dots,x_n)$. Denote by $\text{diag}(g)$ the diagonal matrix with coefficients $g=(g_1,\dots,g_m)$ and denote by $\nabla g$ the Jacobian of $g$. Let $C$ be the set of critical points defined by \begin{equation} C=\{x\in\mathbb R^n\,:\,\text{rank}(\varphi(x))< m\}\quad\text{with}\quad\varphi:=\begin{bmatrix} \nabla g\\ \text{diag}(g) \end{bmatrix}\,. \end{equation} Assume that the image of $C$ under $f$, denoted by $f(C)$, is empty or finite. (Our assumption holds generically since $C$ is empty in a Zariski open set in the space of the coefficients of $g_1,\dots,g_m$ with given degrees.) We provide a sequence of values, returned by semidefinite programs, finitely converges to the minimal value attained by $f$ over the basic semi-algebraic set $S$ defined by \begin{equation} S:=\{x\in\mathbb R^n\,:\,g_j(x)\ge 0\,,\,j=1,\dots,m\}\,. \end{equation} Consequently, we can compute exactly the minimal value of any polynomial with real coefficients in $x$ over one of the following sets: the unit ball, the unit hypercube and the unit simplex. Under a slightly more general assumption, we extend this result to the minimization of any polynomial over a basic convex semi-algebraic set that has non-empty interior and is defined by the inequalities of concave polynomials.