Skip to content

Erdős Problem #284

Editorial depth: Reviewed explanatory article — evidence qualified.
Snapshot classification: proved (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.

Editorial authority: The protected GCL V2 admissions ledger alone determines publication status. This source-bounded explanation is not mathematical certification.

The question in plain language

For a positive integer \(k\), consider all ways to write one as a sum of exactly \(k\) distinct unit fractions:

\[ X_k=\left\{(x_1,\ldots,x_k)\in\mathbb Z_{>0}^k:x_1<\cdots<x_k,\ \sum_{i=1}^k\frac1{x_i}=1\right\}. \]

When \(X_k\) is nonempty, define \(f(k)\) to be the largest possible smallest denominator:

\[ f(k)=\max\left\{x_1:(x_1,\ldots,x_k)\in X_k\right\}. \]

The question is whether

\[ f(k)\sim\frac{k}{e-1}. \]

The quantifier matters. This means that for every positive \(\varepsilon\), there should be a \(K\) such that for every \(k\ge K\), the set \(X_k\) is nonempty and

\[ \left|\frac{f(k)}{k/(e-1)}-1\right|<\varepsilon. \]

It is an assertion about all sufficiently large \(k\), not merely a statement about a subsequence. The elementary splitting identity below shows that \(X_k\) is nonempty for every \(k\ge3\), so the domain issue can be settled separately. The difficult part is making the smallest denominator as large as the asymptotic predicts.

A small example

The identity

\[ \frac1n=\frac1{n+1}+\frac1{n(n+1)} \]

preserves the reciprocal sum and increases the number of terms by one. If it is applied to the current largest denominator, the two new denominators are larger than every denominator already present, so strict ordering and distinctness are preserved. Starting from

\[ 1=\frac12+\frac13+\frac17+\frac1{42}, \]

we can repeatedly split an offending denominator until every denominator is greater than \(5\). The following calculation uses exact rational arithmetic, not floating-point approximations.

from fractions import Fraction
from collections import Counter

a = [2, 3, 7, 42]
N = 5
while True:
    c = Counter(a)
    bad = [n for n in a if n <= N or c[n] > 1]
    if not bad:
        break
    n = min(bad)
    a.remove(n)
    a += [n + 1, n * (n + 1)]

a.sort()
print(a)
print(len(a), a[0], a[-1], len(a) == len(set(a)))
print(sum((Fraction(1, n) for n in a), Fraction(0)))

The output is a 24-term increasing list beginning with \(6\) and ending with \(3263442\); the checks return True for distinctness and 1 for the exact sum. This is a finite padding illustration. It does not estimate \(f(k)\) and does not prove the asymptotic theorem.

What is known

There is an immediate upper bound. If the smallest denominator is \(u\), then the reciprocal sum is at most the sum obtained from the \(k\) consecutive denominators \(u,u+1,\ldots,u+k-1\). Comparing that harmonic sum with an integral gives

\[ 1\le \log\left(1+\frac{k}{u}\right)+o(1), \]

and hence \(u\le(1+o(1))k/(e-1)\). Thus the conjectural constant is forced from above.

The published primary source is Ernest S. Croot III, “On unit fractions with denominators in short intervals,” Acta Arithmetica 99 (2001), 99–114, DOI 10.4064/aa99-2-1. On printed p. 99 Croot defines the same \(X_k\) and asks the maximum \(x_1\) question. His Main Theorem on printed p. 100 states that, for a fixed positive rational \(r\) and every \(N>1\), there are distinct integers

\[ N<x_1<\cdots<x_k\le\left(e^r+O_r\left(\frac{\log\log N}{\log N}\right)\right)N \]

whose reciprocal sum is \(r\). Taking \(r=1\) gives the relevant short-interval construction. The proof of that theorem reaches its conclusion on printed pp. 104–105 in the accessible journal PDF. This is a published external proof claim; inspecting the theorem and proof location is not the same as independently rechecking every argument, and it is not GCL certification.

Croot’s own introduction says that the two Erdős–Graham questions are solved “for infinitely many \(k\).” That wording is the central source exception. His theorem is quantified over the lower cutoff \(N\), whereas Problem 284 asks for a value for every sufficiently large cardinality \(k\). A proposed bridge is to choose \(N\) as a function of \(k\), ensure Croot’s representation has at most \(k\) terms, and then use largest-denominator splitting to reach exactly \(k\) terms without changing the smallest denominator. The published statement, as checked here, does not by itself display the required cardinality bound or the resulting argument for every sufficiently large \(k\).

GCL's current state

The authenticated original issue is #1554, whose complete source packet identifies the problem page, the pinned catalogue, the absence of an admitted FormalConjectures statement, and the source prerequisites. Its actual authenticated RESULT/1 comment has disposition STATUS_CONFLICT. It records the normalized statement, Croot’s theorem, the exact finite check above, and the unresolved bridge between “infinitely many \(k\)” and “all sufficiently large \(k\).” That return is the controlling GCL evidence for this article.

The pinned catalogue snapshot records proved, last updated 2025-08-31, with formal_status: unformalized and formalized: no; its full-file Git blob is 4e60263a80af6207aca494d2e1388e44fa372889. The current catalogue entry remains textually the same at commit 76848eddd25687dfc69bb26bec740c35c4a9c6b4 (file blob 97ae243a7fc056bb9c6413ce665b575eadd27ff3). “Proved” is catalogue metadata, not a GCL proof certificate.

The pinned FormalConjectures provider is commit 85f863718beeec7b58a3a1926ee92e3472bc2020. A lookup for FormalConjectures/ErdosProblems/284.lean at that SHA returned 404, as did the current main lookup. Therefore #284 has no Lean statement, no sorry-bearing #284 theorem, and no external Lean proof claim to report or promote.

Two neighboring audits illustrate why the interfaces must not be blurred. For #276, the weak English condition “no fixed divisor divides every term” is not automatically equivalent to the stronger pinned Lean condition \(\forall d>1\;\exists k\;\gcd(d,a_k)=1\). For #285, the pinned Lean use of Fin k.succ shifts the represented quantity by one relative to \(f(k)\) and its theorem bodies contain sorry. Neither neighboring record supplies formal evidence for #284.

What remains

The bounded mathematical-documentary task is to prove or refute the bridge. One must state a function \(N(k)\), extract from Croot a representation with cardinality at most \(k\), and show that repeated splitting of its largest denominator reaches exactly \(k\) terms while retaining a smallest denominator of \((1-o(1))k/(e-1)\). The error term \(O(\log\log N/\log N)\) must be carried through an explicit \(\varepsilon\)–\(K\) argument. Eventual nonemptiness is easier: split the largest term of \(1=1/2+1/3+1/6\) repeatedly.

If Croot’s theorem cannot provide the needed cardinality control, the correct result is an unresolved semantic bridge, not a silent upgrade from “infinitely many” to “all large.” A source-faithful follow-up should stop at that blocker and should not describe the catalogue label as a proof of the stronger quantifier.

Sources and claim boundaries

The original problem is attributed to P. Erdős and R. L. Graham, Old and New Problems and Results in Combinatorial Number Theory (1980), pp. 33–34; an accessible scan is available here. The exact source packet and its live status discussion are preserved in issue #1554; the exact returned evidence is the comment at 6135376451. The maintained #284 page and history returned HTTP 403 in this audit environment, so their status wording is reported through the authenticated return rather than claimed as an independent live-page readback.

Croot’s accessible published PDF supports the theorem title, definition, printed theorem page, proof pages, and DOI cited above. No additional resolving paper was needed, and no unavailable proof paper is silently treated as checked. The missing #284 FormalConjectures file is an unavailable formal artifact, not evidence of either truth or falsity. External catalogue status, a published proof attribution, a source-audit return, and GCL internal certification are four different claims: this draft supports the first three only to the limits stated here and asserts no GCL certification, no accepted editorial status, and no completed bridge for all sufficiently large \(k\).

Catalogue and evidence provenance

  • Pinned catalogue source: teorth/erdosproblems at 536408716724.
  • GCL issue: #1554 — 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.

Back to catalogue · Compendium introduction