Forbidden subset sums in a linear range: an inverse theorem and the Erdős-Graham irregularity problem

Mateo Matijasevick

Matijasevick proves that the maximum size of a subset of the integers from 1 to the integer part of cn avoiding subset ∑n\sum n, divided by n, converges exactly when c is at least 1. For 0<c<10 < c < 1 the limit superior is c/2c/2 and the sequence has at least two limit points.

What was already known

Erdős and Graham asked about irregularity. Alon’s work gives the even subsequence limit in the special case c=1/2c = 1/2 and motivates using divisibility classes.

AI involvement AI-centered

Matijasevick credits a GPT-5 and Claude research pipeline with the ideas behind the paper and credits earlier problem discussion for specific constructions and data.

Importance Notable

This settles the convergence question throughout the linear range and separates it from the finer task of classifying all subsequence limits.

Technical verification Formal (Lean 4)

The project reports a clean Lean build and axiom audit. Its formal non-convergence proof uses a bounded zero-sum argument; the development also checks the regular range and the limit superior.

Formal checking

The project reports a clean Lean build and axiom audit. Its formal non-convergence proof uses a bounded zero-sum argument; the development also checks the regular range and the limit superior.

Checker
Lean 4 in Matijasevick’s reported build
Scope
Checked the named non-convergence, regular-range and limit-superior theorems against the corresponding statements.
Basis
Reported by the linked checker

Human understanding Few

The disclosure does not assign the explanation of the central proof to a named human separately from the model-generated ideas.

Correction and publication history