Erdős Problem #300
Editorial depth: Reviewed explanatory article — evidence qualified.
Snapshot classification: solved (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
For a positive integer \(N\), look at the numbers \(1,2,\ldots,N\). We seek the largest set \(A\) of these numbers with a strong avoidance property: no selection of distinct members of \(A\) has reciprocals adding to \(1\). Write this maximum size as \(A(N)\). The original Erdős–Graham question asks how large \(A(N)\) can be compared with \(N\).
The quantifiers matter. The forbidden selection is any finite subset \(S\) of \(A\), including a one-element subset, and the equation is an exact rational equation, not an approximation. Thus a set containing \(1\) is immediately disallowed, while a set may have a large total reciprocal sum provided no sub-selection sums exactly to \(1\).
The early guess was that almost every integer could be retained: \(A(N)=(1+o(1))N\). Croot showed that this is false. The later theorem of Liu and Sawhney identifies the correct density:
Here \(e\) is the base of the natural logarithm, so the limiting proportion is about \(0.632\). This is an asymptotic statement: it describes the ratio as \(N\) grows, not an exact formula for any particular small \(N\).
A small example
An exact finite check illustrates the definition without pretending to address the asymptotic theorem. For \(N=5\), the set \(\{2,3,4,5\}\) is admissible. Exhaustively checking its nonempty subcollections gives the reciprocal sums
none of which equals \(1\). The only five-element subset of \(\{1,2,3,4,5\}\) contains \(1\), so it is not admissible. An exact enumeration of all subsets therefore gives \(A(5)=4\), with \(\{2,3,4,5\}\) a maximizer. A reproducible check using rational arithmetic is:
from fractions import Fraction
from itertools import combinations
def good(B):
return all(sum((Fraction(1, n) for n in S), Fraction(0)) != 1
for k in range(1, len(B) + 1)
for S in combinations(B, k))
valid = [B for k in range(6)
for B in combinations(range(1, 6), k) if good(B)]
print(max(map(len, valid)), [B for B in valid if len(B) == 4])
The actual output is 4 [(2, 3, 4, 5)]. This verifies one small value by finite exact calculation. It does not prove the limiting density, establish the analytic estimates in the research paper, or validate any formal proof.
What is known
The primary modern source is Yang P. Liu and Mehtaab Sawhney, “On further questions regarding unit fractions,” arXiv:2404.07113v1, submitted 10 April 2024. Their Theorem 1.3 states that for every fixed \(ε\) with \(0<ε<1/2\), and for all sufficiently large \(N\) depending on \(ε\), every \(A⊆[1,N]\) satisfying \(|A|≥(1−1/e+ε)N\) contains a subset \(A'\) whose reciprocal sum is \(1\). This is the upper bound on the avoidance maximum. Their discussion immediately gives the matching obstruction: for fixed \(γ>0\), the tail beginning around \((1/e+γ)N\) has reciprocal sum below \(1\), so it itself contains no forbidden sub-sum. Together these bounds yield the stated asymptotic.
The paper’s proof is not a numerical search. It uses Fourier analysis on reciprocal sums, smooth-number decompositions, and an inverse argument controlling modular residues. The paper explains that Theorem 1.3 is obtained from its Proposition 5.2 after restricting to a suitable range and removing numbers with unsuitable smoothness or too many prime factors. This is the external mathematical proof evidence used here; it is not a GCL certification.
The historical source is Erdős and Graham, “Old and new problems and results in combinatorial number theory” (1980), where the question appears in the unit-fraction problems. Croot’s “On a coloring conjecture about unit fractions,” Annals of Mathematics 157 (2003), 545–556, supplies the earlier obstruction to density tending to \(1\). These references explain the provenance of the question and the first qualitative disproof, while Liu–Sawhney supplies the sharp asymptotic resolution.
GCL's current state
The original source-audit issue is MATH-PROGRAMME #1570, assignment ERDOS-CATALOGUE-0300-S01. At audit time it had the required direct-editorial labels and no comments, so there was no existing RESULT/1 to duplicate.
The frozen catalogue evidence is teorth/erdosproblems commit 536408716724a2402bdfa7c1d764bf83d64a8c89, file blob 4e60263a80af6207aca494d2e1388e44fa372889. Its entry records informal status “solved,” then status “solved (Lean),” with formalization metadata dated 20 September 2026. The live Erdős Problems page, last edited 23 January 2026, still says “Formalised statement? No.” That is a real metadata conflict, not evidence against the Liu–Sawhney theorem.
An upstream FormalConjectures statement is now readable at commit e0eca2b35e9fbf4977c32711ca8d9c8ae82bfbbb, file blob 281737f2f5a0dcf22151a1d35367d9bad471d74d. It defines the same finite maximum and states the limit, but the wrapper theorem is literally “by sorry.” The linked external Lean source at plby/lean-proofs commit 8822f7ddef30fadbd92e1c6ab4ed897af356af5e, blob 4cd2a809cb08cda88d11360a0353601209f2db99, contains a completed theorem source, but it was not replayed here. The issue’s admitted FC-GDM-85F86371 snapshot says no statement file is present; that snapshot boundary remains distinct from the newer upstream file.
What remains
The main mathematical question is externally resolved, but a bounded successor task remains useful: reconcile the exact statement across the catalogue, live page, FormalConjectures wrapper, and linked Lean development. The task should compare the finite-set predicates, the treatment of \(N=0\), the normalization of rational reciprocals, and the meaning of the asymptotic limit. It should then record which source is admitted for GCL coverage and which metadata is merely external evidence.
No new proof search is justified by this audit. A suitable acceptance test is a line-by-line semantic comparison plus an independently recorded replay result for the exact pinned Lean commit, if a coordinator authorizes that replay. Falsification would be a quantifier mismatch, a different denominator interval, a changed forbidden-subset predicate, or a proof source that fails at the cited commit. The stopping rule is reached when those comparisons and the formalization-status drift are documented; mathematical certification, publication, and any claim that V2 ledger coverage is accepted remain outside this article.
Sources and claim boundaries
The source locks used here are the issue URL above; the frozen catalogue commit and blob above; Liu–Sawhney arXiv v1 PDF, whose observed SHA-256 is 29f6ea8f73bb17a9084bec115af4b3d8015d9b6d5e3c803a3caa8f79a287d91b; and the Erdős–Graham PDF, whose observed SHA-256 is 0cbf0c32f0ab1e1c71db5121a88bac905bf976c4a6ab6bb6d7d9cf9ddd184ed3. The exact primary citation for Liu–Sawhney is arXiv:2404.07113; the historical scan is Erdős–Graham (1980); and Croot is Annals of Mathematics 157 (2003).
The external paper establishes the asymptotic theorem. The finite Python calculation establishes only the displayed value \(A(5)=4\). The FormalConjectures file is a theorem statement with a sorry; the linked Lean source is external evidence whose replay and axioms were not independently checked here. Catalogue fields and the live page are status metadata, not proof. GCL mathematical certification is not supplied, and accepted coverage is decided by the V2 ledger alone. Original SOURCE RESULT/1 citation: authenticated original RESULT/1.
Catalogue and evidence provenance
- Pinned catalogue source: teorth/erdosproblems at
536408716724. - GCL issue: #1570 — protected issue/assignment identity; its labels and results are live, not reproduced automatically here.
- GCL triage: OUTSIDE_STRICT_OPEN_TRIAGE.
- Formal-statement coverage: No statement was confirmed in GCL's pinned formal-provider snapshot. Presence does not mean the statement or proof has been checked.
- External mathematical status: consult current problem page and its dated bibliography.