VibeMathedMath problems solved with AI

Lusztig's Fourier matrix conjecture for noncrystallographic Coxeter groups

For a two-sided Kazhdan-Lusztig cell cc of a finite Coxeter group WW, Lusztig defined an asymptotic Hecke category from Soergel bimodules. Its Drinfeld center is a modular tensor category (Elias and Williamson), and for Weyl groups it is the category of unipotent character sheaves supported on cc. Lusztig's conjecture: for noncrystallographic WW (dihedral groups, H3H_3, H4H_4) the SS-matrix of this center equals the Fourier matrix attached to cc by Lusztig, Malle and Broué-Malle in the 1990s. Rogel and Thiel proved it for dihedral groups and for some cells of H3H_3 and H4H_4; other cells were left with two candidate categorifications, and the middle cell of H4H_4 (aa-value 6), with Malle's 74 by 74 Fourier matrix, was fully open.

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI-discovered
Method
Computation
Field
Representation theory; fusion categories; Hecke categories
Posed by
George Lusztig, Truncated convolution of character sheaves, Section 10 (Bull. Inst. Math. Acad. Sin. 2015)
Year posed
2015
Years open
11y
Solved
2026-11-12
Model
GPT-6 Pro
Vendor
OpenAI
Collaborators
Qiyue Tang
Verification
Unreviewed
Publication
Preprint
Significance
20 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Theorem 1.2: for every finite irreducible noncrystallographic Coxeter group and every two-sided cell cc, Sc=FcmodS_c=F_c^{\mathrm{mod}} under a bijection of simple objects with the Fourier family, for the pivotal structure inherited from Soergel bimodules. New here: a relative Hodge-Riemann argument fixing the pivotal signs, which decides the cells left with two options, and the middle cell of H4H_4: a Morita equivalence of Alvis's rank-24 corner with a crossed product of the Fibonacci Hecke category of type A3A_3, an exact tube-algebra computation of its center (74 simple objects) over Q(5)\mathbb{Q}(\sqrt5), and certified ball arithmetic with a cyclotomic separation bound that makes the match with the CHEVIE matrix exact in all 5476 entries. The Morita step rests on a uniqueness theorem of He and Liu (arXiv 2610.41811, October 2026, unrefereed). No geometric construction of character sheaves is claimed. AI-written; the record is labelled human understanding: no parts.

What the AI did

From the Hexagon record's provenance statement: "GPT-6 Pro was used to assist in developing and checking the mathematical results. The principal recorded research session lasted 147 minutes and 19 seconds. The model was subsequently used for independent proof checks, verification of computational certificates, and several revisions of the manuscript. A further review using Rethlas-OSV is planned, and the review report will be updated when available." Hexagon labels the work "Primarily AI-generated text" and "Human understanding: no parts". The PDF itself carries no AI statement.

Verification

No independent mathematician has checked this yet. Checked here on 19 November 2026: the Hexagon record 2611.00362v1 (contributed by Qiyue Tang, Tsinghua University, ORCID-linked) and its provenance statement, which carries the AI disclosure; the PDF's introduction and Theorem 1.2. The conjecture was read as Rogel and Thiel state it (arXiv 2307.07276v3, attributed to Lusztig 2015, Section 10), together with their statement that the middle cell of H4H_4 was unresolved; Lusztig's paper itself was not opened. The supplementary archive on GitHub (commit 502c80a) was downloaded and its README, reproduction notes and AI audit reports were read; its verify.py was not run here. A separate script written here read the shipped 512-bit enclosures and confirmed that the 74 by 74 matrix SS and the twists form a modular datum: SS real and symmetric, S2=IS^2=I to 50 digits, 1/S00=1080+48051/S_{00}=1080+480\sqrt5, (ST)3=S2(ST)^3=S^2 with twist order 60, and 12,210 sampled Verlinde coefficients are non-negative integers. That tests the internal consistency of the output, not the tube computation, the Morita equivalence or the match with CHEVIE. The only reviews so far are by AI systems.

Sources

Submitted by MellowGander562 on

Changelog2 changes

Discussion