Erdős Problem #312
Editorial depth: Reviewed explanatory article — evidence qualified.
Snapshot classification: open (not automatically the live result).
Subject areas: number theory, unit fractions.
Read this status correctly
The status above is reproduced from GCL's immutable external-catalogue snapshot. It is not a new claim of current mathematical resolution, a source-semantic review, an independent proof check, or GCL certification. A published proof and a GCL replay are separate events.
The question in plain language
Problem 312 asks how closely a finite collection of unit fractions can approach \(1\) from below. The collection is a multiset: a denominator may occur more than once. If \(A\) is such a multiset, write
For a submultiset \(S\) whose reciprocal sum is at most \(1\), define the gap to be
The problem asks whether there is an absolute constant \(c>0\) such that, for every \(K>1\), every sufficiently large finite multiset of positive integers with \(R(A)>K\) has a submultiset satisfying
The phrase “sufficiently large” is part of the quantifier issue. The pinned Lean statement makes the threshold \(N_0\) depend on \(K\), then requires the assertion for every indexed multiset of length at least \(N_0\). It also models a multiset by a function a : Fin n → ℕ, so it permits zero entries unless positivity is added separately. In Lean, the inverse of zero is zero; that is a semantic difference from the maintained problem’s positive-integer wording.
The original issue is MATH-PROGRAMME #1582. Its frozen catalogue packet identifies the question as open, in number theory and unit fractions, and records a statement-formalized but not formally solved entry.
A small example
Take the multiset \(A=\{2,3,3\}\). Its total reciprocal mass is
Use denominator \(6\) as a common denominator. The three available contributions are the integer weights \(3,2,2\). Enumerating all submultisets gives numerator sums
The largest numerator not exceeding \(6\) is \(5\), so the largest reciprocal subsum below \(1\) is \(5/6\) and the exact gap is \(\varepsilon(A)=1/6\). There is no exact representation of \(1\) in this example because \(6\) is absent from the displayed set of attainable numerators.
This is an exact finite calculation, not a numerical approximation. It illustrates the optimization problem and the role of repeated denominators. It does not test the asymptotic assertion: the example has no “sufficiently large” threshold, and without a value of the conjectural constant \(c\) it cannot decide the exponential inequality. A short integer-subset enumeration reproduces the numerator set exactly, so no floating-point or unverified computational output is being used here.
What is known
The maintained Erdős Problems entry, last edited 20 January 2026, labels the problem open and says that Erdős and Graham knew the weaker estimate with \(e^{-cK}\) replaced by \(c/K^2\). Its displayed source key is [ErGr80], referring to P. Erdős and R. L. Graham, Old and New Problems and Results in Combinatorial Number Theory (1980), the monograph listed as 128 pages in the catalogue. The exact page in that monograph was not independently read during this bounded audit, so the original page-level attribution remains a source boundary rather than a certified citation.
There is now a relevant primary research document: Samuel Korsky, “Approximating 1 by Sums of Unit Fractions,” arXiv:2607.04157v2, submitted 5 July 2026 and revised 8 October 2026. Its introduction defines the same gap and states Theorem 1.1: for absolute constants \(c>0\) and \(K_0\), every finite multiset of positive integers with \(R(A)>K\geq K_0\) satisfies
The paper describes this as an improvement on the Erdős–Graham polynomial estimate. This is materially stronger than \(K^{-2}\) for large \(K\), but it is weaker than the requested \(e^{-cK}\) bound. Thus it is progress toward Problem 312, not a resolution of its stated target. The associated forum post from 24 June 2026 is an unverified report of the same stretched-exponential direction and explicitly says that parts of its Fourier details warrant inspection; it is not used as proof evidence.
GCL's current state
The frozen catalogue revision is teorth/erdosproblems@536408716724a2402bdfa7c1d764bf83d64a8c89, with the Problem 312 Git blob 4e60263a80af6207aca494d2e1388e44fa372889. At that revision, the informal and general status fields are open, last updated 2025-08-31; the formal-status field is unformalized, while statement-formalized metadata is yes.
The pinned formal source is google-deepmind/formal-conjectures, commit 85f863718beeec7b58a3a1926ee92e3472bc2020, file FormalConjectures/ErdosProblems/312.lean, blob 6324d3b70f0c0347c80cd849f1e9a82825f530f8. The theorem is tagged research/open and has the shape answer(sorry) ↔ ... := by sorry. That records a formal interface, not a proof, replay, or semantic-equivalence certificate.
The formal statement does correctly expose the existential constant, the \(K>1\) quantifier, the cardinality threshold depending on \(K\), indexed multiplicities, and the strict lower/non-strict upper inequalities. The missing positivity hypothesis on entries should be reviewed before treating it as a faithful theorem interface. GCL has supplied no mathematical certification here; V2 ledger decisions, rather than this article, determine accepted coverage.
What remains
The immediate bounded successor is a source-and-interface review. First, obtain a readable copy of the Erdős–Graham monograph and locate the exact Problem 312 passage and the hypotheses behind the \(c/K^2\) estimate. Next, compare that passage line by line with the maintained wording and the Lean quantifiers, explicitly deciding whether “sufficiently large” is cardinality-dependent on \(K\) and whether positive entries must be enforced. Finally, record the Korsky theorem as a dated external result with its exact stretched-exponential scope, without upgrading it to the open exponential target.
The acceptance test should be a source table containing the original page, exact hypotheses, theorem conclusion, date/version, and a resolved or explicitly retained positivity/threshold discrepancy. Falsification is triggered by a primary passage that changes the quantifier order, by a proof conclusion that is only polynomial, or by an external source that actually proves the full \(e^{-cK}\) claim. Stop when the source table is complete or when the monograph remains inaccessible after the bounded retrieval attempts. Do not launch a Lean build or a full computational campaign on the basis of this audit.
Sources and claim boundaries
The original issue is the controlling assignment and currently has no posted RESULT/1; the original SOURCE RESULT/1 citation is authenticated original RESULT/1. The frozen catalogue and formal hashes above are exact packet identities, not claims that either artifact is mathematically correct. The live problem page supplies current editorial status and the weaker historical claim; its eight forum comments are user reports and are not proofs. Korsky’s arXiv v2 is the accessible primary research source for the stretched-exponential theorem; its retrieved PDF had SHA-256 51bba38930f3197c4d09b42ac711693ad75224af9aacbb82a4be1b38f5eace08, recorded as an observation rather than a GCL pin.
Accordingly, this article claims only: the normalized question; the exact finite example; the pinned metadata; the existence and stated scope of the arXiv result; and the remaining source-semantic work. It does not claim that Problem 312 is solved, that the arXiv proof has been independently audited, that the Lean theorem is proved, or that GCL has certified any mathematical statement.
Catalogue and evidence provenance
- Pinned catalogue source: teorth/erdosproblems at
536408716724. - GCL issue: #1582 — protected issue/assignment identity; its labels and results are live, not reproduced automatically here.
- GCL triage: PROTECTED_READY_BACKLOG.
- Formal-statement coverage: A provider statement was included in the pinned snapshot. Presence does not mean the statement or proof has been checked.
- External mathematical status: consult current problem page and its dated bibliography.