Hermite moment formulas for the coefficients b1, b3 of a Krivine scheme
ProvedGrothendieckConstant.coeff_hermite_momentsby moona3k · Oct 7, 2026 · Mathlib 0df444a (Lean v4.33.1)
functional-analysisgrothendieck-constantharmonic-analysishermite-polynomialsprobability
Let S=(f,g) be a Krivine scheme of dimension k, let X be a standard Gaussian vector in Rk with density φ(x)=∏m(2π)−1/2e−xm2/2, and let H(t)=b1t+b3t3+⋯ be the normalized correlation function of the scheme, with b1=H′(0) (coeffLinear) and b3=H′′′(0)/6 (coeffCubic).
For a function h on Rk write ah=E[h(X)X]∈Rk for its first-order Hermite moments and Th=E[h(X)H3(X)] for its third-order Hermite moments, where
H3(x)ijl=xixjxl−δijxl−δilxj−δjlxi
is the third Hermite tensor. Then
b1=2π⟨af,ag⟩=2πi∑af,iag,i,b3=12π⟨Tf,Tg⟩=12πi,j,l∑Tf,ijlTg,ijl.
This is the order-one and order-three part of the Hermite (Mehler) expansion of the correlation function. It makes precise the paper's remark that (b1,b3) are linear functionals of each of f and g, so that affine constraints in (b1,b3) pass to mixtures and limits of schemes.
Proof outline. The Gaussian pair density equals c(t)kexp(−(A−2tB)/(2(1−t2))) with A=∣x∣2+∣y∣2 and B=⟨x,y⟩. Its first three t-derivatives are bounded on ∣t∣≤1/2 by 500(1+A+k)3e−A/4, which is integrable, so the correlation function can be differentiated three times under the integral sign. At t=0 the third derivative of the density is the density times B3−3AB+(3k+6)B=∑i,j,lH3(x)ijlH3(y)ijl, and the double integrals factor by Fubini.
Preamble
import Mathlib
import Definitions.Def_KrivineSchemeDefs
open MeasureTheory ProbabilityTheory Real
Formal statement
namespace GrothendieckConstant
theorem coeff_hermite_moments (k : ℕ) (S : KrivineScheme k) :
coeffLinear S = (π / 2) * ∑ i : Fin k,
(∫ x, S.f x * x i * ∏ m, gaussianPDFReal 0 1 (x m)) *
(∫ y, S.g y * y i * ∏ m, gaussianPDFReal 0 1 (y m)) ∧
coeffCubic S = (π / 12) * ∑ i : Fin k, ∑ j : Fin k, ∑ l : Fin k,
(∫ x, S.f x * (x i * x j * x l - (if i = j then x l else 0) - (if i = l then x j else 0)
- (if j = l then x i else 0)) * ∏ m, gaussianPDFReal 0 1 (x m)) *
(∫ y, S.g y * (y i * y j * y l - (if i = j then y l else 0) - (if i = l then y j else 0)
- (if j = l then y i else 0)) * ∏ m, gaussianPDFReal 0 1 (y m)) := by sorry
end GrothendieckConstant
Source
Li, Saha, Xue, Chaudhuri, Klivans, Kothari, Meka, "Long-Horizon AI Research for Grothendieck Constant: A Case Study in Human-AI Mathematical Collaboration", arXiv:2608.11195v3 (2026), https://arxiv.org/abs/2608.11195, Section 2, p. 6 (definition of the coefficients b1, b3 of H) and Appendix A, p. 17 (the Hermite coefficients (b1, b3) are linear functionals of the scheme). Standard Mehler/Hermite expansion of the bivariate Gaussian density.
View graph