Long-Horizon AI Makes Progress on the Grothendieck Constant
This post reviews a 2026 case study (arXiv:2608.11195) in which an AI research system, collaborating with researchers from UT Austin, Princeton, and UCLA, narrowed the range of the Grothendieck constant K_G — a number that had resisted precise determination for over 70 years. The article was collected on 2026-08-13.
What is the Grothendieck constant?
The constant, introduced by Alexander Grothendieck in 1953 in the context of tensor products of Banach spaces, quantifies the "translation cost" between continuous and discrete worlds. In combinatorial optimization, it bounds how much precision is lost when a continuous (semidefinite/linear) relaxation is rounded back to a discrete solution. More deeply, the Grothendieck inequality controls the gap between correlations achievable with independent strategies versus shared information — and, via Tsirelson's 1985 work, it governs how much advantage quantum entanglement provides over classical correlations in Bell-type experiments.
Unlike π or e, K_G has no closed-form formula; its definition involves optimal strategies in infinite-dimensional spaces, so only upper and lower bounds are known.
Historical progress
- 1977 — Krivine: proved K_G ≥ 1.676 (via explicit hard instances, the "Krivine scheme") and K_G ≤ π/(2 log(1+√2)) ≈ 1.782 (via a rounding scheme).
- 1991 — Reeds: improved the lower bound, showing Krivine-type schemes are nearly optimal within their framework (later formalized by Naor–Regev 2014).
- 2000s — quantum information: the complex Grothendieck constant was connected directly to Bell inequality violations.
- Early 2026: a burst of activity — Steven Heilman published new lower and upper bound papers; Jones and Malavolta showed K_G is strictly larger than the Davie–Reeds bound.
- Lower bound: K_G ≥ 6π/11 ≈ 1.714
- Upper bound: K_G ≤ π/(2 log(1+√2)) − 10⁻⁴ ≈ 1.782
- Li, A., Saha, R., Xue, A., Chaudhuri, S., Klivans, A., Kothari, P. K., & Meka, R. (2026). *Long-Horizon AI Research for Grothendieck Constant: A Case Study in Human-AI Mathematical Collaboration*. arXiv:2608.11195.
- Saha, R., Li, A., Xue, A., et al. (2026). *New Upper and Lower Bounds for the Grothendieck Constant*. arXiv:2608.11202.
- Grothendieck, A. (1953). Résumé de la théorie métrique des produits tensoriels topologiques. *Boletim da Sociedade de Matemática de São Paulo*, 8, 1–79.
- Krivine, J.-L. (1977). Sur la constante de Grothendieck. *C. R. Acad. Sci. Paris*, Série A-B, 284(8), A445–A447.
- Reeds, J. A. (1991). A new lower bound on the real Grothendieck constant (unpublished).
- Naor, A., & Regev, O. (2014). Krivine schemes are optimal. *Proc. AMS*, 142(12), 4315–4320.
- Pisier, G. (2011). Grothendieck's theorem, past and present. arXiv:1104.2083.
- Tsirelson, B. S. (1985). Quantum analogues of Bell's inequalities. *Zapiski Nauchnykh Seminarov LOMI*, 142, 174–194.
- Lakatos, I. (1976). *Proofs and Refutations*. Cambridge University Press.
- Pólya, G. (1957). *How to Solve It* (2nd ed.). Princeton University Press.
The new results
The 2026 paper (with companion paper arXiv:2608.11202) establishes:
Together these pin the first decimal digit of K_G as 7, and shrink the gap to roughly 0.068 — a qualitative step toward eventually determining the constant exactly.
Methodologically, the lower bound does not construct a specific hard instance; instead it proves that all asymptotically optimal Krivine schemes share an intrinsic structural limitation. The upper bound gives the first asymptotic rounding scheme — a family {R_d} whose approximation ratios converge as d → ∞ — analyzed using random matrix theory, high-dimensional geometry, and Fourier analysis.
The role of AI: insight, not calculation
Crucially, the AI system did not do numerical computation. It proposed new mathematical constructions — proof strategies, rounding schemes, and analytical frameworks — judged *novel* by domain experts. The "long-horizon" system:
1. Sustained exploration of a combinatorially huge strategy space (embeddings, rounding functions, symmetry assumptions) 2. Learning from failure, analyzing why each attempt failed and adjusting 3. Maintaining context across weeks or months of exploration 4. Collaborating with human mathematicians who supplied direction and "mathematical taste"
The paper's most striking episode: an extended AI search for a *better upper bound* failed — but human mathematicians noticed that certain intermediate constructions had structure dual to what *lower bound* proofs require. Repurposing these "failed" results produced the new lower bound 6π/11. As Pólya advised: if you cannot solve a problem, solve a related one.
Significance for AI and mathematics
The post traces AI's evolving role: Appel–Haken's computer-assisted four-color theorem proof (1976, brute verification), ML-assisted conjecture generation (2010s), AlphaGeometry's IMO-level constructions (2024), and now genuine participation in open problems in analysis (2026) with Lean 4 machine-checked formalization.
The proposed ideal collaboration: humans set direction and verify rigor; AI explores tirelessly, free of intuition bias, while humans recognize patterns, pivot perspectives, and judge elegance — with humans also catching subtle "pseudo-proofs" AI can produce.
Future directions: fully determining K_G (perhaps within 5–10 years), extending the approach to other constants (Goemans–Williamson, Tsirelson-type bounds), improving AI meta-cognition and mathematical taste, and applying the paradigm to theoretical physics, drug design, and materials science.