Skip to content

Erdős Problem #319

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

Erdős Problem 319 asks how large a minimally balanced collection of reciprocals can be. Fix a positive integer \(N\). Choose a set \(A\) of distinct integers from \(1\) through \(N\), and give each chosen integer a sign \(delta_n\) equal to \(1\) or \(-1\). The total signed reciprocal must vanish:

\[ \sum_{n\in A}\frac{\delta_n}{n}=0. \]

The balance must be irreducible: whenever \(A'\) is a nonempty proper subset of \(A\), its signed reciprocal sum is not zero:

\[ \sum_{n\in A'}\frac{\delta_n}{n}\ne 0. \]

The problem asks for the largest possible \(|A|\) as a function of \(N\). It is useful to call this extremal quantity \(c(N)\). The signs are part of the witness, while minimality is checked against every proper subcollection using the same signs. This is not merely a question about finding any relation among reciprocals; it asks for a relation with no smaller zero-sum relation hidden inside it.

The original issue is MATH-PROGRAMME #1589. Its pinned catalogue snapshot is teorth/erdosproblems commit 536408716724a2402bdfa7c1d764bf83d64a8c89, entry 319, with Git blob 4e60263a80af6207aca494d2e1388e44fa372889. Its catalogue and informal status fields say open, last updated 2025-08-31. The entry records number theory and unit fractions, no prize, and statement formalized metadata, not a proof.

A small example

Take \(N=6\), \(A=\{2,3,6\}\), and set \(\delta_2=1\), \(\delta_3=-1\), and \(\delta_6=-1\). Exact arithmetic gives

\[ \frac12-\frac13-\frac16=\frac{3-2-1}{6}=0. \]

The six nonempty proper subsets have signed sums

\[ \frac12,\quad -\frac13,\quad -\frac16,\quad \frac16,\quad \frac13,\quad -\frac12, \]

in the order \(\{2\}\), \(\{3\}\), \(\{6\}\), \(\{2,3\}\), \(\{2,6\}\), and \(\{3,6\}\). None is zero, so this is a verified admissible witness and proves only the finite lower bound \(c(6)\ge 3\). It does not identify \(c(6)\), predict the asymptotic size of \(c(N)\), or say anything by itself about the unresolved exact extremal function. The calculation is reproducible with integer numerators over the common denominator \(6\); no floating-point approximation is involved.

What is known

The maintained problem page attributes a substantial lower bound to an observation of Sarosh Adenwalla using Ernest S. Croot III's paper, “On unit fractions with denominators in short intervals”, Acta Arithmetica 99 (2001), 99–114. Croot's primary source is also available as arXiv:math/9904181v1. In the introduction and Main Theorem, Croot proves that for a fixed rational \(r>0\) and sufficiently large \(N\), one can represent \(r\) as a sum of distinct unit fractions whose denominators lie just above \(N\) and whose largest denominator is at most \(\left(e^r+O_r(\log\log N/\log N)\right)N\). His proof starts from almost all reciprocals in an interval, removes a controlled set, and completes the residual.

The catalogue explains the transfer as follows. Apply the unit-fraction construction with the lower endpoint scaled near \(N/e\). The resulting denominators lie below \(N\) and the construction supplies a set \(B\) with reciprocal sum \(1\) and the asymptotic size reported on the maintained page as \((1-1/e+o(1))N\). Then \(A=B\cup\{1\}\), with sign \(-1\) on \(B\) and sign \(+1\) on \(1\), has total signed sum zero. Because all elements of \(B\) are larger than \(1\), every nonempty proper subset omitting \(1\) has negative sum, while a proper subset containing \(1\) has sum \(1-\sum_{b\in B'}1/b>0\). Thus the transfer has the required minimality mechanism. The exact counting transfer is recorded here as the catalogue's source-backed consequence; it is not a new proof or GCL certification.

The strongest safe status statement is therefore that the exact extremal question remains listed open, while a linear lower bound is externally reported and the trivial inclusion \(A\subseteq\{1,\ldots,N\}\) gives \(c(N)\le N\). The pinned formal source is google-deepmind/formal-conjectures commit 85f863718beeec7b58a3a1926ee92e3472bc2020, file FormalConjectures/ErdosProblems/319.lean, blob 4ab11d3f0e80356ae760316796721f95e8e42582. Its main theorem uses Finset.Icc 1 N, rational sums, a global sign function restricted to the selected set, and strict proper-subset quantification. It ends with answer(sorry) and by sorry; the file is a formal statement/interface, not a checked proof.

GCL's current state

This source audit is SOURCE_LOCK_READY for the bounded documentary task. The original issue has no comments and therefore no pre-existing RESULT/1 to duplicate. The live Erdős Problems entry, accessed 2026-10-13, preserves the same question, tags, bibliography, Croot lower-bound note, and a link to the formal statement. Its history shows a 2025-10-20 revision of the explanatory text. The live page also exposes one public proof claim asserting a density-one asymptotic, but it explicitly labels that claim unrefereed and says its appearance is no guarantee of correctness. It is an unverified external report, not accepted coverage or a solved-status change.

The formal-conjectures file has since drifted beyond the pinned issue packet: the current file at commit 15bd12092e91f8fc527df87a4ea5171d2f33248d, blob 2f565fec125e145c72cdb63240d908994fd6d773, marks the Theta variant solved with answer \(N\) and removes the older tautological Big-O and little-o variants, while retaining the main theorem as open and sorry-backed. That later state is useful drift evidence but does not silently replace the pinned source lock. The V2 admissions ledger alone decides accepted compendium coverage. This article does not confer mathematical certification, publication authority, or acceptance of any external proof claim.

What remains

A bounded successor task should audit the lower-bound bridge and the exact target separately. Its input is Croot's primary paper, the pinned catalogue entry, and the pinned Lean file. Its target is one source-locked lemma: for all sufficiently large \(N\), construct a set \(B\subseteq\{\lceil(1/e-o(1))N\rceil,\ldots,N\}\) with \(\sum_{b\in B}1/b=1\), \(|B|\ge(1-1/e+o(1))N\), and verify that \(B\cup\{1\}\) satisfies every proper-subset condition. Acceptance requires matching the quantifiers, distinctness, interval endpoints, asymptotic meaning, and sign convention to the primary source and reproducing the finite minimality argument exactly. Falsification is any missing count estimate, endpoint mismatch, or proper subset with zero signed sum. Stop after this lemma and its evidence ledger; do not infer an upper bound, exact formula, density-one theorem, Lean proof, or GCL certification from it.

Sources and claim boundaries

The source packet is the original issue, the frozen catalogue commit and blob above, and the frozen formal commit and blob above. The Croot arXiv PDF used for the primary-source reading had local SHA-256 bb28a4369e025e37aab746593ce46afe1b545aefe2c9659d30867e1aafc038f6 for the retrieved PDF; the published citation is the DOI-linked Acta Arithmetica article. The original bibliography entry is Erdős and Graham, Old and New Problems and Results in Combinatorial Number Theory (1980), shown live as [ErGr80]; the exact page for Problem 319 was not asserted from an unreadable scan.

The external Croot theorem and the maintained-page lower-bound transfer are separated from the unrefereed density-one proof claim, from the sorry-backed Lean theorem, and from any GCL mathematical certification. The finite witness above is independently checked but has only finite scope. Original SOURCE RESULT/1 citation: authenticated original RESULT/1. V2 ledger admission, protected state, and any later mathematical certification remain outside this article's authority.

Catalogue and evidence provenance

  • Pinned catalogue source: teorth/erdosproblems at 536408716724.
  • GCL issue: #1589 — 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.

Back to catalogue · Compendium introduction