VibeMathedMath problems solved with AI

The constant-width numerical range conjecture

The numerical range of a matrix A∈Cn×nA\in\mathbb{C}^{n\times n} is W(A)={x∗Ax:x∈Cn, x∗x=1}W(A)=\{x^*Ax: x\in\mathbb{C}^n,\ x^*x=1\}, a compact convex subset of the plane. A convex set has constant width dd if its orthogonal projection onto every line has length dd; disks have this property, and so do non-circular sets such as the Reuleaux triangle. Chien, Nakazato and Meng studied the diameter and width of numerical ranges, with constant-width examples among nilpotent Toeplitz matrices, and conjectured that a numerical range of constant width must be a disk. Is it true that the numerical range of every complex square matrix with constant width dd is a disk of radius d/2d/2?

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI-assisted
Method
Argument
Field
Matrix analysis; numerical ranges
Posed by
M.-T. Chien, H. Nakazato and J. Meng, The diameter and width of numerical ranges, Linear Algebra Appl. 582 (2019), 76-98
Year posed
2019
Years open
7y
Solved
2026-11-10
Model
ChatGPT (OpenAI, model version unstated)
Vendor
OpenAI
Collaborators
Muyan Jiang
Verification
Unreviewed
Publication
Preprint
Significance
7 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Theorem 1: for every A∈Cn×nA\in\mathbb{C}^{n\times n}, n≥1n\ge1, and d≥0d\ge0, the numerical range W(A)W(A) has constant width dd if and only if it is a closed disk of radius d/2d/2 (a single point when d=0d=0). The proof writes the width in direction ζ\zeta as the gap between the extreme eigenvalues of the Hermitian matrix H(ζ)=12(ζ−1A+ζA∗)H(\zeta)=\frac12(\zeta^{-1}A+\zeta A^*), shows by a greatest-common-divisor argument that the midpoint of the projected interval is a rational function of ζ\zeta, bounds it by the operator norm to get a Laurent polynomial a−1ζ−1+a0+a1ζa_{-1}\zeta^{-1}+a_0+a_1\zeta, and uses reversal of direction and reality on the unit circle to identify it with Re(ζˉc)\mathrm{Re}(\bar\zeta c) for one point cc, the centre. Not addressed: the analogous conjecture of Chien, Li and Nakazato for higher-rank numerical ranges, and operators on infinite-dimensional spaces.

What the AI did

From the paper's acknowledgements: the author used OpenAI's ChatGPT to explore proof strategies and to develop and refine the proof and its exposition, and takes responsibility for the mathematical arguments and the final manuscript. The model version is not stated. The same acknowledgements credit Ilya Spitkovsky with suggesting the use of characteristic roots separated by the width, which is the starting point of the argument, and Minwu Kim with checking an earlier version.

Verification

Unreviewed: no independent mathematician is on public record for this proof. Checked here on 19 November 2026: arXiv 2611.13461v2 (Muyan Jiang, UC Berkeley, 2 pages, revised 17 November) was read in full and each step of the proof of Theorem 1 was followed: the Rayleigh-Ritz description of the projected interval, the common-root argument over the field of rational functions, the passage from the monic greatest common divisor to a rational midpoint function, the operator-norm bound that removes all finite poles and caps the growth, the oddness and reality conditions on the unit circle, and the conclusion through support functions. No gap was found in that reading, which is this site's own and not an expert review. That the conjecture is as stated was confirmed from the abstract of Chien, Li and Nakazato (2021), which describes it as a conjecture on the classical numerical range; the 2019 paper that poses it is behind a paywall and was not opened. The acknowledgements say Minwu Kim checked an earlier version and Ilya Spitkovsky commented on drafts; neither is an independent public confirmation. No formalization. An arXiv search found no earlier or competing proof.

Sources

Submitted by MellowGander562 on

Changelog2 changes

Discussion