You can infer the interaction potential from a single snapshot — if you use both distances and forces.
Note (2026-11-09): The abstract says the method infers the interaction potential from configurations at an arbitrary state point by enforcing consistency between two estimates of the pair correlation function (from distances and from forces), does not need a new simulation at each iteration, and applies 'including when the density is large and alternative methods may fail'; performance is benchmarked on a set of test potentials in simulation. The essay's explanations that the iterative mapping becomes stiff, that the redundancy 'gets stronger' with correlations, and that the inversion becomes well-posed are not in the abstract.
Inferring the potential between particles from their configurations usually requires iterative simulation. Guess a potential, simulate, compare the resulting structure to the observed one, adjust, repeat. At high densities the iteration converges slowly or not at all, because the mapping from potential to structure becomes stiff.
Coquand, Paolino, and Berthier (arXiv:2603.12081) show that the pair correlation function g(r) — the most basic structural descriptor — can be estimated two independent ways from the same configuration: once from interparticle distances, once from the forces between particles. In exact statistics, these estimates agree. With a wrong potential, they disagree. The method enforces consistency between the two estimates, and this single constraint is enough to determine the correct potential without iteration.
The insight is that particle configurations contain redundant information about the potential, and the redundancy closes the inverse problem. The distance-based estimate and the force-based estimate are formally equivalent but computationally independent. Their agreement at the right answer and disagreement at wrong answers turns an ill-posed inversion into a well-posed consistency check.
This works at a single state point — no need for data at multiple temperatures or densities. And it works at high density, precisely where iterative methods fail, because the redundancy gets stronger (not weaker) as correlations increase.
Coquand, Paolino, and Berthier, "Direct Boltzmann inversion method from particle configurations at arbitrary state points," arXiv:2603.12081 (2026).