Work

Exact certification of irrational global minimizers from inexact sum-of-squares certificates

Year: 2026
Type: preprint
Abstract: Sum-of-squares programming can certify polynomial inequalities, but in practice its certificates are computed numerically, in floating-point arithmetic. A common way to make them exact is to round them to rational numbers, which typically needs some slack. Bounds that certify optimality at an irrational minimizer have essentially no slack, so rounding alone is generally not enough. We describe one way to handle this case and apply it to the Thomson problem: placing N equal charges on the unit sphere to minimize their Coulomb energy. We round a numerical three-point semidefinite certificate and keep the rounding error as an explicit residual, which gives an energy bound within 10⁻⁵⁰ of the candidate configuration. A quantitative form of complementary slackness, a finite case analysis and a local Hessian bound are then used to turn this nearly tight bound into an exact result. In this way we obtain a computer-assisted proof that the tricapped trigonal prism (N = 9) and the capped square antiprism (N = 10) are the unique minimizers up to rotations, reflections and relabeling. Both results are also formalized in the Lean 4 proof assistant.
Language: English
Open Access status: green
OpenAlex ID: W7236659347
Cites: 28
Cited by:
Related to: 10