The storage capacity of the Ising perceptron: verification of the outstanding numerical conditions

Yitzchak Shmalo

Shmalo's paper verifies Ding–Sun Condition 1.2, including curvature and endpoint clauses, and Huang Condition 1.3 for the zero-margin Ising perceptron. Combining these with prior conditional and universality theorems yields M_N/N→α★≈0.8330786 in probability for Gaussian and fixed standardized subgaussian disorder, including Bernoulli.

Sources for this record

What was already known

Krauth and Mézard predicted α★ in 1989. Ding and Sun's lower bound and Huang's upper bound were conditional on global sign conditions; Huang had interval-verified other numerical inputs but not Condition 1.3, and earlier evidence for the outstanding signs used nonrigorous floating-point calculations.

AI involvement AI-autonomous

Shmalo set the objective and reviewed the mathematics. Claude Fable 5 developed the central analytic reduction and verification approach; Codex audited and repaired verification blocks. The paper does not identify a completed human check of the entire analytic argument.

The source explicitly says the systems operated for extended unattended sessions and selected mathematical approaches after the author set the objective; Fable 5 supplied the named central architecture.

Importance Major

Reports closure of the two outstanding conditions behind the Krauth–Mézard capacity prediction.

Limitations

The capacity statement is for the zero-margin Ising perceptron under Gaussian and fixed standardized subgaussian disorder. Interval certificates cover numerical conditions, while the analytic reduction is a paper proof without a documented complete check.

Technical verification Partial

Arb/python-flint interval certificates and source-bound verifiers cover the parameter rectangle and finite inequalities for both outstanding conditions, with a second-system audit and corrected cross-check errors. The moment reduction, convex duality and ray-majorant implications are paper proofs rather than machine-checked statements; the paper itself says the package was not independently reproduced or peer reviewed. A completed rigorous check of the whole analytic argument is not documented.

Human understanding Few

Shmalo reports supervision and mathematical review, but the central mechanism and initial explanation were model-produced; no named independent human explanation of it is documented. Few is the evidence default.

Correction and publication history

2026-10-18 · publication

Reports closure of the two outstanding conditions behind the Krauth–Mézard capacity prediction.

Read this published revision