Elementary Symmetric Polynomial Inequalities for Centered Vectors and Matrices
Yanjun Han, Jonathan Niles-Weed
math.CO, cs.IT, math.PR
2026-07-27
A combinatorics paper proves new bounds for elementary symmetric polynomials of centered vectors and matrices; the abstract states GPT-5.5 Pro produced the proofs of both main theorems, while the authors refined them and wrote the applications.
This is a combinatorics and information theory paper, not an AI paper, but it carries a footnote that matters to anyone tracking frontier models. The abstract states that the proof ideas for the two main theorems were produced by the GPT-5.5 Pro model, and the acknowledgments thank Yuzhou Gu of OpenAI for using GPT-5.5 Pro to produce the proofs of Theorems 1.1 and 1.2. The authors refined the arguments, derived all the applications, and wrote the paper. Yanjun Han and Jonathan Niles-Weed are researchers in probability and statistics at NYU's Courant Institute.
The math concerns new upper bounds for elementary symmetric polynomials (ESPs).
An ESP of degree k sums all products of k distinct entries of a vector. The paper bounds ESPs for "centered" vectors (entries summing to zero) and for matrices with zero row and column sums. Theorem 1.1 shows that for a centered complex vector with normalized squared magnitude, |ek(z)| is at most C times the square root of (n choose k), with C no larger than 24. Theorem 1.2 gives a matrix analogue that carries an extra exponential factor B^k, B at most 8e squared. Both turn "how large can the ESP get in the worst case" from uncontrolled into an explicit constant bound. The proof lines came from GPT-5.5 Pro; the authors refined and filled them in.
The inequalities are not idle. They plug into two applications. Corollary 1.4 caps the chi-squared distance between a permutation mixture and its mean-field (independent) approximation, with the bound governed by overlap eigenvalues. Corollary 1.5 gives a chi-squared version of the finite de Finetti theorem for small alphabets: the k-way marginal of the sequence is statistically close to an exchangeable approximation once n grows large enough, specifically k = o(n / sqrt(T)). These are long-standing questions in information theory and probability that previously lacked explicit, broadly applicable bounds.
For the mathematics itself, that is for specialists to weigh. For the AI community the concrete fact is this: a paper signed by serious researchers and posted to arXiv has its core theorem proofs produced directly by a frontier model, with humans doing the refinement, applications, and writing. Whether that counts as doing research depends on your definition, but in this one paper the proof step is the model's.
The authors note that the matrix inequality (Theorem 1.2) carries an exponential B^k factor they could not remove, so it is likely not tight. The constants C and B are unoptimized. The paper does not describe how GPT-5.5 Pro arrived at the proofs or how correctness was checked; from the text alone one can only confirm that the proof ideas came from the model and the authors made them stand. Final correctness rests on the authors' peer review.