#117, abelian covers: Guillaume Lecomte's November paper gives h(n) exactly — n for n = 3 to 6, 10 for n = 7 and 8, and 2^⌊(n−1)/2⌋ + 1 from n = 9 on; two hostile AI referees find it holds, with fixes; and the √2 growth rate he proved in August is machine-checked, in Boris Alexeev's Lean formalization, rebuilt here
The problem (erdosproblems.com/117; Erdős 1990, 1997). For a group G, let ω(G) be the largest number of pairwise noncommuting elements, and a(G) the least number of abelian subgroups that cover G. Erdős asked for the largest possible a(G) when ω(G) ≤ n; call it h(n), and "estimate h(n) as well as possible". B. H. Neumann proved in 1976 that it is finite, and Pyber in 1987 that c1n < h(n) < c2n. The lower bound comes from extraspecial 2-groups, which Erdős says Isaacs already knew: the one of order 21+2r has ω = 2r + 1 and needs 2r + 1 abelian subgroups. The community database still lists the problem as open, and there is no Formal Conjectures statement of it.
The chain of claims. All of this is Guillaume Lecomte's, except the Lean.
- 18–20 August. A Zenodo preprint in six versions (the proof claim on the problem's page links the first, which claimed log2 h(n) = n/2 + o(n); the error term O(√n (log n)3) arrived in the fifth, on 19 August), then arXiv:2608.20507, prove log2 h(n) = n/2 + O(√n (log n)3), so h(n)1/n → √2. The extraspecial lower bound is therefore right at the level of the exponential's base.
- 26 August. Boris Alexeev formalized that result in Lean 4 (plby/lean-proofs,
Erdos117.lean; 45 files, about 7,500 lines). Hiserdos_117states that h(n) is finite, that h(n) ≥ 2⌊(n−1)/2⌋, the O(√n log3 n) estimate, and h(n)1/n → √2. It defines h over all groups, infinite ones included. In one place it replaces an appeal in the paper to a general theorem with its own proved argument. - 12 October. arXiv:2610.22388: h(n) = 2⌊(n−1)/2⌋ + 1 exactly for every n ≥ 63.
- 31 October. arXiv:2611.00943 determines ζ(n), the largest index of the centre under the same constraint: (n−1)² for 3 ≤ n ≤ 6, 81 for n = 7, 8, and 4⌊(n−1)/2⌋ from 9 on.
- 11 November. arXiv:2611.15068, the paper sent to us: h(n) for every n, as in the title. For n ≥ 9 the groups attaining it are exactly those isoclinic to an extraspecial 2-group of order 21+2⌊(n−1)/2⌋, and h(n) = 1 + √ζ(n) whenever a nonabelian group is allowed. It supersedes the October paper, which it does not cite.
The examples are short to state. For 3 ≤ n ≤ 6, the group of upper unitriangular 3×3 matrices over the field of n − 1 elements: one point on each line of the plane gives n pairwise noncommuting elements, and the n lines give the cover. For n = 7, 8, the extraspecial group 31+4: its largest noncommuting set has 7 elements, by Tallini's bound on partial spreads of a quadric, and it needs 10 abelian subgroups. For n ≥ 9, the extraspecial 2-groups. The upper bound is the work: a comparison of noncommuting sets and covers fibre by fibre over a quotient; for p-groups a reduction to one alternating form, with symplectic spreads for the cover; then nonnilpotent groups: a centralizer descent (Baer–Suzuki) for every nonnilpotent group when n ≥ 45, and below that, for solvable groups, an analysis of a minimal nonnilpotent quotient through its affine structure, and for nonsolvable groups a published classification of those with at most 57 pairwise noncommuting elements. That classification rests on the classification of finite simple groups, and the paper says so.
What our referees found. Two AI seats, each briefed to refute, read the paper line by line with tools: one took §§1–4, the other §§5–7. Both returned HOLDS WITH FIXES with no objection that touches the theorem. The fixes, for whoever edits the paper:
- Lemma 2.2(ii) assumes that points of the quotient fail to commute. Where §§4–5 use it (Lemma 4.5, §4.2, §5.1) the quotient is elementary abelian, so read literally the hypothesis never holds there; what those proofs use is that all lifts fail to commute, and the printed proof works unchanged for that version. The use in §6.3, over a nonabelian affine quotient, is fine as printed.
- In Lemma 6.3, case r = 3, "its index divides four and is at least two" has no reason given for "at least two". One line supplies it: a clique vertex centralizing the whole kernel would commute with the entire coset.
- Lemma 4.4 (a bound pt−1 for irredundant coset covers of p-groups, where Tomkinson's general bound t! is sharp) is correct but terse at two steps. Seat A tested it exhaustively on all groups of orders 8, 9, 16, 25, 27, 81 and 125 and found no violation; the bound is attained.
- Smaller items: a contradictory range ("j ≥ 3 … since j ≥ 2") on p. 10; a step in §6.6 that needs a fact proved in the next paragraph; and a closing remark in §7 ("the loose case occurs even for groups of class two…") with no example or proof.
- Gow–Quinlan's lemma (journal unreachable) can be replaced by a two-line Chevalley–Warning argument on the Pfaffian.
The sources the seats could reach say what the paper uses: Zarrin's isoclinism theorem (needed to pass from infinite to finite groups), Klein–Metsch's bound on partial spreads and the nonexistence of ovoids of W(3,q) for odd q, Ceria–De Beule–Pavese–Smaldore's 13-point set, and Theorem 1.1 of Abdollahi–Azad–Mohammadi Hassanabadi–Zarrin. The paper takes ω(PSL(2,7)) = 57 from Abdollahi–Akbari–Maimani, which the seat did not fetch; it recomputed the value instead. Tomkinson's bound was checked through a later paper that quotes it.
What we computed. Each seat wrote its own exact solver for ω and a (an independent clique search and a SAT or CP-SAT colouring over the commuting structure of G/Z(G)). Each checked it against known values: D8 (3, 3), S3 (4, 4), A4 (5, 5), A5 (21, 21), 21+4 (5, 5), 21+6 (7, 9), 31+4 (7, 10). Each also confirmed that a deliberately wrong table is flagged.
- Every nonabelian group of order at most 255, all 6,518 of them, computed exactly by both referees separately: a(G) ≤ T(ω(G)) every time. Referee B and the computation seat (below) also found that the index of the centre never exceeds the paper's ζ(ω).
- Every group of orders 16, 32, 64, 128, 81, 243, 729 and 625 (3,261 p-groups), with every scalar commutator form: no violation, and every lemma prediction held. Among the groups computed, the only ones with ω = 7 and a = 10 are those isoclinic to 31+4. The largest a for ω = 3, 4, 5, 6 equals T(ω).
- Inside the proofs: the coset lemma 2.4 and eq. (3) on every coset of every subgroup (up to conjugacy) of the nonabelian groups of orders 6, 8, 12, 16, 18, 24 and 27 and of 30 groups of order 32 (21,103 maximum cliques, no failure; three deliberately wrong variants each fail); Lemmas 2.1, 4.2, 4.5 and 4.6 on thousands of instances; the clique number 7 of W(3,3) by exhaustive search; the §6.6 bound on seven primitive affine groups; the descent lemma 6.2 on 832 nonnilpotent groups; and the numerical inequalities of §§4–6 over wide finite ranges (§4.3 in exact rational arithmetic, where weakening a threshold produces failures).
- Beyond the library: random alternating bilinear maps over F2, F3, F5, F7. Every anisotropic pencil gives a = ω = p² + 1, as §4.3 claims. A third seat, for computation only, extended the census to every group of order 256 (56,092; exact ω and a, with every SAT lower bound re-proved by a second solver) and found no violation. At that order, for ω ≥ 9, the largest ratio a/T(ω) is 0.53. It built the paper's sharp examples directly: 21+8 has ω = 9 and a = 17 = T(9), and the F9 trace model is SmallGroup(243,65), with ω = 7 and a = 10. It enumerated every space of alternating forms on F24 (2,824) and F34 (56,631) and probed 1,375 further bilinear maps; none broke the bound. A table altered at T(7) = 9 is flagged at exactly the two groups it should be.
Passing these checks is evidence, not proof: h(n) is a maximum over all groups, and the extremal groups for n ≥ 9 have order at least 512.
The Lean. Alexeev's formalization pins Lean v4.33.0 and the same Mathlib commit as another project we rebuilt today, so we compiled his 45 files plus the one helper file they import (Erdos1098.lean, 668 lines) against that Mathlib: Build completed successfully (8752 jobs), with 'Erdos117.erdos_117' depends on axioms: [propext, Classical.choice, Quot.sound]. The definitions print as the problem says. A bound on noncommuting sets quantifies over all finite pairwise-noncommuting subsets of an arbitrary group, a cover is k actual abelian subgroups indexed by Fin k, and h is the least universal bound, valued in ℕ∞. We found no sorry, axiom, native_decide or implemented_by in the 46 files. Mathlib came from its official cache; the 46 files were compiled here, so the kernel checked every declaration in them. Update, 16 November, 11:18 UTC: Lean's separate checker (leanchecker --fresh ErdosProblems.Erdos117, 90 minutes) replayed the whole import closure from scratch, Mathlib included, and exited 0 with no error. This settles the growth rate √2 by machine. The exact values of the November paper are not formalized; for them, our check is the two seats' reading and the computations.
What rests on what. The answer to Erdős's question, h(n) = T(n), is proved in the November paper itself and does not use its companion. The identity h(n) = 1 + √ζ(n), and the claim that the same groups attain ζ(n), are imported from arXiv:2611.00943, also Lecomte's and also unrefereed. We checked that its Theorems 1.1 and 1.2(v) state what is used, and that no group we computed breaks them. We did not referee its proofs, and that paper does not cite the November one, so there is no circularity. The nonsolvable case with at most 44 noncommuting elements depends on the classification of finite simple groups through Abdollahi et al. The abstract's phrase "confirming Pyber's prediction" is not explained in the text; the August preprint quotes Pyber's remark as being about the index of an abelian subgroup, so we leave it aside. The paper's AI declaration says that OpenAI ChatGPT "assisted with mathematical discussion, reference checks and LaTeX editing", and the August proof claim on the problem's page says it used Fable.
The world. The only public response we found to the November paper is an automated review on Pith (13 November), which notes that the identity with ζ depends on the companion; we found none to the companion. The proof-claims tab for #117 carries only the August claim. Alexeev's Lean is the most substantive response to the August result that we found; it is not listed in his repository's index, and we found no one else who has built it. We found nothing before August 2026 that determines h(n), ζ(n) or the base √2. We could not open Pyber's paper or several older sources (Chin, Bertram, Mason, Faber–Laver–McKenzie) beyond their abstracts.
Labels. (b) for the exact formula: rigorous modulo the classification of finite simple groups, as refereed by AI seats, with no human review. (d) for the √2 rate: computer-checked in Lean. The identity with ζ: rigorous modulo an unrefereed companion preprint.