Skip to content

Erdős Problem #316

Editorial depth: Reviewed explanatory article — evidence qualified.
Snapshot classification: disproved (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 316 asks whether a finite set of distinct positive integers, none equal to \(1\), can always be split into two groups whose reciprocal sums are both strictly below \(1\), provided the reciprocal sum of the whole set is below \(2\). In symbols, for a finite set \(A\subseteq\mathbb{N}\setminus\{1\}\), the question is whether

\[ \sum_{n\in A}\frac{1}{n}<2 \]

forces a partition \(A=A_1\sqcup A_2\) with

\[ \sum_{n\in A_i}\frac{1}{n}<1\qquad(i=1,2). \]

The word “set” matters: an integer may occur only once. A multiset version is easier to refute because repeated small denominators can put too much weight into either part. The strict inequalities also matter. A partition with one side exactly \(1\) is not a successful partition.

The original source is Erdős and Graham’s Old and New Problems and Results in Combinatorial Number Theory (1980), where the question appears on printed page 41. The source asks about distinct denominators and displays the two complementary reciprocal sums explicitly. This fixes the ambient object and the quantifier order more reliably than a short catalogue label.

A small example

The maintained catalogue records the counterexample

\[ A=\{2,3,4,5,6,7,10,11,13,14,15\}. \]

Its total reciprocal sum is

\[ \sum_{n\in A}\frac{1}{n}=\frac{120047}{60060}=1.998784548\ldots<2. \]

That inequality alone does not show failure: in principle the terms might still be arranged into two groups below \(1\). A finite exact check tests every possibility. The following calculation uses rational arithmetic, not floating-point comparisons.

from fractions import Fraction

A = [2, 3, 4, 5, 6, 7, 10, 11, 13, 14, 15]
v = [Fraction(1, n) for n in A]
total = sum(v, Fraction())
successes = []
for mask in range(1 << len(A)):
    left = sum((v[i] for i in range(len(A)) if mask >> i & 1), Fraction())
    right = total - left
    if left < 1 and right < 1:
        successes.append((left, right))
print(total, 1 << len(A), len(successes))

The actual output is 120047/60060 2048 0. Thus this particular set satisfies the hypothesis, while no one of its \(2048\) subset/complement pairs satisfies both strict upper bounds. The check establishes a concrete finite counterexample; it does not, by itself, establish that the set is minimal, prove Sándor’s general theorem, or certify any formal development.

What is known

The live Erdős Problems entry reports that Csaba Sándor resolved the question negatively in “On a problem of Erdős,” Journal of Number Theory 63 (1997), 203–210. It says that the proper divisors of \(120\), excluding \(1\), already form a counterexample, and that Sándor proved a stronger statement: for every integer \(n\geq2\), some finite set with reciprocal sum below \(n\) cannot be partitioned into \(n\) parts all having reciprocal sum below \(1\).

The proper-divisor claim also has a short exact check. For

\[ B=\{2,3,4,5,6,8,10,12,15,20,24,30,40,60\}, \]

the sum is \(239/120<2\). Enumerating its \(2^{14}=16384\) subsets gives zero successful subset/complement pairs. This agrees with the reported counterexample, but it is not a substitute for reading Sándor’s proof. The publisher metadata identifies the article and its open-access status, but the article’s full text was not retrievable in this audit environment; theorem and page locations for the general result therefore remain unverified here.

The source evidence supports a careful status distinction. The two-part problem is disproved by an explicit finite set, and the finite arithmetic is reproducible. The stronger statement for all values of \(n\) is reported by the maintained page and formal source comments, but its published proof has not been inspected in full during this audit. “Minimal counterexample” is likewise a maintained-site assertion, not a claim established by the small enumeration above.

GCL's current state

The original audit issue is MATH-PROGRAMME #1586, assignment ERDOS-CATALOGUE-0316-S01. Its frozen packet identifies the teorth/erdosproblems catalogue commit 536408716724a2402bdfa7c1d764bf83d64a8c89, entry number: "316", and blob 4e60263a80af6207aca494d2e1388e44fa372889. At that snapshot, the informal status is disproved, the formal status is Lean, and the formalized flag is yes, with the last catalogue update dated 2025-09-02.

The pinned formal provider is google-deepmind/formal-conjectures commit 85f863718beeec7b58a3a1926ee92e3472bc2020; FormalConjectures/ErdosProblems/316.lean has blob ee6f980ebeca0b87bab9a7b455f798c08b2c9514. Its theorem erdos_316 has the exact finite-set hypotheses, proves answer(False), constructs the 11-element set above, and closes the finite obstruction with kernel-decided computation. This is source evidence about that Lean file, not GCL mathematical certification.

The same file contains a separate generalized theorem for \(n\geq2\) ending in sorry. It must not be conflated with the completed two-part theorem. The live page was last edited 5 March 2026 and displays “DISPROVED (LEAN)”; that label records maintained-site status, while V2 ledger evidence decides accepted coverage. The original SOURCE RESULT/1 citation is authenticated original RESULT/1.

What remains

The most useful bounded successor is source closure for Sándor’s result. It should obtain a readable copy of the 1997 article, record its exact hash, identify the theorem and page proving the two-part counterexample and the extension to all values of \(n\), and compare its hypotheses with the finite-set Lean theorem. In particular, it should check whether “partition into \(n\) parts” means exactly \(n\) nonempty parts, permits empty parts, or uses another convention, and whether the published construction has the same exclusion of \(1\) and the same strict inequality.

A second, narrower task is a formal-source review: mark the completed two-part theorem separately from the generalized declaration with sorry, then test whether the generalized statement is definitionally or merely informally aligned with Sándor’s formulation. No Lean replay is needed to identify the visible sorry boundary. The finite rational checks are reusable regression tests, not a replacement for the missing source reading.

Sources and claim boundaries

The boundaries are deliberate: published-source reports, exact finite arithmetic, catalogue metadata, and Lean source status are different evidence classes. This article supplies no independent mathematical certification, publication decision, or claim that the generalized sorry theorem has been proved.

Catalogue and evidence provenance

  • Pinned catalogue source: teorth/erdosproblems at 536408716724.
  • GCL issue: #1586 — protected issue/assignment identity; its labels and results are live, not reproduced automatically here.
  • GCL triage: OUTSIDE_STRICT_OPEN_TRIAGE.
  • 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