Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated


Claim. On 2026-05-02 Lech Mazur posted in the thread of Problem 1194 a lower bound that, in Mazur's words, GPT-5.5 Pro, Codex and a verification harness found; the full write-up names GPT-5.5 Pro and Mazur as its authors. Here ana_n is the larger member of the unique representation n=an−bnn=a_n-b_n. With λ=1/(2log⁡2)\lambda=1/(2\log2) and B1=12+γ−log⁡log⁡2B_1=\tfrac12+\gamma-\log\log2, for every fixed B>B1B>B_1 infinitely many nn satisfy an>λn2/(log⁡n−log⁡log⁡n+B)a_n>\lambda n^2/(\log n-\log\log n+B). Consequently lim sup⁡n→∞an(log⁡n−log⁡log⁡n+B1)/n2≥λ\limsup_{n\to\infty}a_n(\log n-\log\log n+B_1)/n^2\ge\lambda; the endpoint B=B1B=B_1 is claimed only in this limsup form. The full proof and a sketch are PDFs in the claimant's repository, which also holds a Lean development whose Challenge.lean states the two results as FloorSaving.floor_saving_lower_bound and FloorSaving.endpoint_limsup.

Submission note. Posted to the site's forum by Lech Mazur on 2 May 2026:

GPT-5.5 Pro + Codex + a verification harness found the following improvement.

Let

λ=12log⁡2,B1=12+γ−log⁡(log⁡2).\lambda=\frac{1}{2\log 2}, \qquad B_1=\frac12+\gamma-\log(\log 2).

For

every fixed B>B1B>B_1, infinitely many nn satisfy

an>λn2log⁡>n−log⁡log⁡n+B.a_n> \lambda\frac{n^2}{\log > n-\log\log n+B}.

Consequently,

lim sup⁡n→∞an(log⁡>n−log⁡log⁡n+B1)n2≥λ=12log⁡2.\limsup_{n\to\infty} \frac{a_n(\log > n-\log\log n+B_1)}{n^2} \ge \lambda = \frac{1}{2\log 2}.

The endpoint

B=B1B=B_1 is only claimed in this limsup form, not as a pointwise infinitely-often inequality.

Lean repository: https://github.com/lechmazur/erdos_1194

Proof sketch: floor_saving_proof_sketch.pdf

Full proof: floor_saving_lower_bound_final_version.pdf

Covers. A lower bound of order n2/log⁡nn^2/\log n for ana_n along infinitely many nn. It does not determine how fast an/na_n/n must grow.

Standing. A reader's AI check, posted, reported one minor issue in the write-up, noted that the write-up's length made the check less reliable, and judged that the Lean appears to formalize the result. This corpus has not built the Lean development or audited its definitions, so the Lean is a link, not formalized evidence. The claim is unrefereed and stays claimed.

Depends on. No page of this wiki.