Quantifying the remaining search space in analogue optimisation

Abstract

Benchmarking analogue optimisers requires certifying not only how close a returned energy is to the optimum, but also how many better configurations may remain. Semidefinite relaxations provide rigorous bounds on the ground-state energy and selected observables, yet reveal little about the size of this residual search space. We introduce a deterministic, instance-specific upper bound on the number of binary configurations whose energies improve upon a returned state by a prescribed margin. Combined with a semidefinite energy interval, this yields a complementary certificate of both the magnitude and multiplicity of any possible improvement. To judge an analogue optimiser's native performance, we provide a generalisation of the framework to obtain bounds on the Haar volume of the remaining continuous phase-space. To further characterise the remaining search space, we extend our bounds to resolve regions in configuration space in both the binary and continuous setting. Our approach couples a variational-Chernoff reduction with an elimination bound of the resulting partition function. Comparisons with exact enumeration and independent Haar sampling are consistent with the bounds on small instances. In numerical experiments at $N=1000$, even the largest family-median bound restricts the potentially improving binary configurations to less than $10^{-144}$ of the full configuration space, while the largest upper bound on the corresponding Haar volume is of order $10^{-499}$.