Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Claim. is a Sidon set meeting every infinite arithmetic progression in , so its complement contains no infinite arithmetic progression and the answer is no.
Each term exceeds twice the one before, which makes two-term sums distinct, so is Sidon; and for and the term with lies in the progression , since it equals and divides . The library's construction page writes the proof out.
Postings. The claimant is Google DeepMind, whose system AlphaProof found
the construction. The site's page credits it to AlphaProof and writes it out;
the construction is absent from a web archive capture of the page of
2024-11-07 and present in one of 2025-05-30. Its first dated records are in the
formal-conjectures catalog, Google DeepMind's project: an issue of 2025-05-13
asks to update the catalog's Problem 198 with the information added to the
site after contact with the curator, and a pull request opened on 2025-05-15
and merged on 2025-05-27, linked above with the catalog file at the merge,
names the counterexample that AlphaProof found, reports that the site's answer
changed from yes to no after Google DeepMind reported it to the curator, and
adds the constructions and to the
catalog; the page is named by the pull request's date, the first record naming
AlphaProof. On 2025-11-24 Alexeev posted in the site's discussion thread a Lean
formalization of this construction, produced from a ChatGPT exposition by
Aristotle with the catalog's statement erdos_198 as target, in Alexeev's
lean-proofs repository, linked above; its header names Baumgartner as the
original human prover, citing his unrelated 1975 paper on canonical partition
relations, and notes that the proof used is AlphaProof's. The catalog's
198.lean marks erdos_198 solved with two formal_proof links, one to that
file, which it describes as Alexeev's Lean formalization made with Aristotle,
and one into the GitHub user XC0R's fork of formal-conjectures at a pinned
commit that GitHub does not serve. The catalog's pull request of 2026-04-13
that added that link carried the proof in its first commit, which GitHub
serves and which is linked above: it proves erdos_198 without sorry from
the set , which the pull request names as AlphaProof's
construction, and the pull request says Claude assisted the Lean translation.
The catalog separately records the
variant , which it says AlphaProof found and proved, with a
formal_proof link into a fork whose file at the pinned commit, linked above,
proves that variant in full: the set is Sidon and meets every infinite
arithmetic progression. The library's
source record
separates the text of Alexeev's file from the builds its postings report; this
corpus has built none of these Lean files, so they give no formalized
evidence.
Acceptance. Reviewed: the site's curator, Thomas Bloom, labels the
problem disproved, credits the construction to AlphaProof, thanks its team and
writes the construction out on the problem page, a documented acceptance
outside this project and independent of the claimant. The formal-conjectures
catalog marks its statement solved with the formalizations linked, but it is
Google DeepMind's own project, so its mark is not independent acceptance. No
refereed publication exists, so refereed is not listed; the Lean files were
not built or audited by this corpus, so formalized is not listed although
the site's label reads DISPROVED (LEAN).
Depends on. No wiki page; the claim rests on the construction stated above.