The Sherrington-Kirkpatrick model with no external field gives each configuration of spins the energy , with independent standard Gaussian couplings . Its ground-state energy per spin is the limit of as grows. An order parameter is a nonnegative, nondecreasing, right-continuous function on with finite integral; the measure it defines encodes a hierarchy of overlap scales. Auffinger and Chen proved the zero-temperature Parisi formula: the ground-state energy is the minimum of the Parisi functional over order parameters, and a minimizer exists.
The question is whether every scale below one occurs, that is, whether the support of a minimizer is all of . Parisi’s theory of replica symmetry breaking predicts it, and it is the hypothesis under which Montanari and El Alaoui, Montanari and Sellke built algorithms that find near-optimal configurations. Auffinger, Chen and Zeng proved that the support of a minimizer is infinite, and noted that it does not end in a stretch on which is constant; whether the support can have a gap in the interior of remained open.
OpenAI’s preprint Full support of the zero-temperature Sherrington-Kirkpatrick order parameter (September 27, 2026) excludes such a gap: every minimizer has support , so it is strictly increasing. With the theorem of El Alaoui, Montanari and Sellke, the preprint obtains, for every , a randomized algorithm that, in operations with depending only on , finds a configuration within of the maximum energy, with probability tending to one.
OpenAI formalized the support theorem in Lean 4 with Mathlib, as the Lean library SKSupport in its openai/math repository. The results on these pages are those of the Lean library, each stated as the Lean library states it, with the definitions it rests on, and linked to the Lean declaration behind it. Lean has checked every one of them, so the Lean library is the reference: where the preprint states a result differently, the statement on the page is the Lean library’s, and the preprint’s is given beside it. The preprint gives the results their order and their names. The Lean library does not prove that a minimizer exists, which Auffinger and Chen proved; it takes this as a hypothesis of its main theorem, Theorem 1.1. It does not contain the algorithm. The proof of Theorem 1.1 in §6 combines the other results.
SKSupport is a Lean library: 17 items, in six groups. Each has a page written by OctoInform, listed below.
No item matches.