VibeMathedMath problems solved with AI

Hlawka constants for Schatten pp-norms (Audenaert–Kittaneh Problem 7)

Hlawka's inequality says that in an inner product space the "triple deficit" ∥X∥+∥Y∥+∥Z∥−∥X+Y+Z∥\|X\|+\|Y\|+\|Z\|-\|X+Y+Z\| is at most the sum of the three "pair deficits". Audenaert and Kittaneh asked for the best constant CC, independent of the matrix size, if it exists, such that for all complex matrices X,Y,ZX, Y, Z
∥X∥+∥Y∥+∥Z∥−∥X+Y+Z∥≤C[(∥X∥+∥Y∥−∥X+Y∥)+(∥X∥+∥Z∥−∥X+Z∥)+(∥Y∥+∥Z∥−∥Y+Z∥)],\|X\|+\|Y\|+\|Z\|-\|X+Y+Z\| \le C\big[(\|X\|+\|Y\|-\|X+Y\|)+(\|X\|+\|Z\|-\|X+Z\|)+(\|Y\|+\|Z\|-\|Y+Z\|)\big],
and in particular for the best constant CpC_p when the norm is the Schatten pp-norm. They noted that C2=1C_2 = 1 by Hlawka's inequality and that C∞C_\infty is infinite. They reported that numerical simulations give C1≥40C_1 \ge 40, "indicating that it might be arbitrarily large as well".

Result
Proved(see note)
Status
Partial result
AI contribution
AI co-developed
Method
Argument
Field
Matrix and operator inequalities; Schatten norms
Posed by
Koenraad Audenaert and Fuad Kittaneh, Problem 7 in "Problems and Conjectures in Matrix and Operator Inequalities" (arXiv:1201.5232)
Year posed
2012
Years open
14y
Solved
2026-09-22
Model
GPT 5.6 Sol and GPT 6 Astra
Vendor
OpenAI
Collaborators
—
Verification
Lean-verified
Publication
Announced
Significance
20 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Settles the "if it exists" part for every pp: a finite size-independent constant CpC_p exists exactly when 1<p<∞1 < p < \infty. The proof gives some constant, not the best one, for complex matrices of every size, square or rectangular.

At p=1p = 1 no constant exists, already for 2×22 \times 2 matrices, turning Audenaert and Kittaneh's numerical C1≥40C_1 \ge 40 into C1=∞C_1 = \infty. At p=∞p = \infty none exists on 3×33 \times 3 matrices, as they already knew.

For diagonal matrices and p≥256p \ge 256 the best constant is an explicit KpK_p, the maximum of a ratio over the family (−t,1,1),(1,−t,1),(1,1,−t)(-t,1,1), (1,-t,1), (1,1,-t), t∈[1/2,2]t \in [1/2, 2], with 939p/2000<Kp≤p939p/2000 < K_p \le p. Since diagonal matrices are matrices, 939p/2000<Cp<∞939p/2000 < C_p < \infty for p≥256p \ge 256.

Open: the best CpC_p for general matrices (except C2=1C_2 = 1) and the best diagonal constant for 1<p<2561 < p < 256; the cutoff 256 is not claimed optimal. All of this is in Lean except the one-line step from diagonal to general matrices.

What the AI did

The author and AI agents co-developed both results. Existence: GPT 5.6 Sol, running in Codex, built the Lean proof in small kernel-checked layers from audited mathematical proof artifacts; the approach was motivated by the Ball–Carlen–Lieb uniform convexity framework. Sharp diagonal constant: the argument came out of two weeks of directed proof search. The author identified the proof strategy, diagnosed repeated dead ends where agents stalled, and refined the approach through failure analysis at each stage. AI agents executed computational searches and produced candidate lemmas, and GPT 6 Astra, running in Codex, formalized the final argument in Lean. The author, GPT 5.6 Sol and Claude Fable 5.1 reviewed the work. The proofs have not been examined in depth by human experts.

Verification

Audited here on 27 September 2026 from a clone at commit 1829591: 48 Lean files, 8,711 lines on Lean 4.35.0-rc2; exactly four sorry, all four the deliberate holes in the two Challenge files that the solution files fill, which is the comparator layout Palomar expects; zero axiom declarations, zero native_decide. CI run 35949577091 is green there and audits the four headline theorems to propext, Classical.choice and Quot.sound.

The submitter notes Palomar last passed its statement check at the earlier commit 79aa498, before the toolchain upgrade, which would normally leave the statements unanchored here. So both Challenge files were fetched at both commits and compared. All four theorem statements are byte-identical. Of the supporting definitions three differ, all cosmetically: N renamed size, and pairGapSum refactored through a new pairGap whose body is size x + size y - size (x + y), the same sum of three pair deficits. The anchoring carries over, which is what lean-verified requires.

Not covered by the four statements, as the submitter says: the bounds 939p/2000 < K_p <= p, library theorems, and the diagonal-to-general step, not in Lean. No human expert has examined the proofs.

Sources

Submitted by ZestyWombat854 on

Changelog2 changes

Discussion