Long-Horizon AI Research Narrows the Grothendieck Constant: A Human-AI Math Collaboration Case Study

Forum topic · 小凯 · 2026-09-02

Summary

A 2026 case study from researchers at UT Austin, Princeton, and UCLA describes how a long-horizon AI research system produced expert-validated novel insights on the Grothendieck constant, an open problem since 1953. The Grothendieck constant K_G quantifies the gap between continuous relaxations and discrete solutions in optimization, and via Tsirelson's connection, the extra correlation strength allowed by quantum entanglement in Bell tests. Combining a new lower bound of 6π/11 ≈ 1.714 with an improved upper bound of π/(2·log(1+√2)) − 10^-4 ≈ 1.782, the work pins the first decimal digit at 7 and narrows the interval to roughly 0.068. Notably, the breakthrough lower bound emerged from AI upper-bound searches whose failed constructions were reinterpreted by human mathematicians as hard instances. The paper illustrates an iterative human-AI collaboration model: AI explores vast construction spaces while humans supply direction, taste, and rigor, with Lean 4 formal verification used to check AI-generated arguments.

Long-Horizon AI Research Narrows the Grothendieck Constant: A Case Study in Human-AI Mathematical Collaboration

> "If you can't explain it to a six-year-old, you don't really understand it yourself." — Richard Feynman

The Number That Kept Mathematicians Awake for 70 Years

Since 1953, the world's best mathematicians—from Princeton to Paris, Moscow to Kyoto—have tried to pin down a single number: the Grothendieck constant $$K_G$$ . Some guessed around 1.5, others near 2. Until recently, not even its first decimal digit was known.

In August 2026, a team from UT Austin, Princeton, and UCLA, working with a long-horizon AI research system, narrowed this number's range significantly—and the AI's insights were judged novel by domain experts.

What the Grothendieck Constant Means

The "Translation Fee" Between Continuous and Discrete Worlds

Consider a logistics problem: 50 trucks, each choosing one of two routes. This is a discrete (integer) optimization problem that becomes intractable beyond ~30 cities. A standard trick is to relax variables to continuous values and solve via linear programming—fast, but fractional solutions ("0.37 of route 1") aren't physically meaningful. Rounding back to discrete solutions loses accuracy, but theory guarantees the loss is at most a constant factor.

That constant, in many cases, is the Grothendieck constant. It measures the "translation fee" from continuous answers to discrete ones. The larger it is, the less reliable continuous methods are.

Deeper Origins: Inequalities, Tensor Products, and Quantum Entanglement

The constant arises from the Grothendieck inequality (1953), proved by Alexandre Grothendieck while studying tensor products of Banach spaces. Informally: if two players can share information, their payoff is some value; if they must decide independently, the payoff drops. $$K_G$$ bounds the loss from that lack of communication.

Remarkably, quantum physicists later found that $$K_G$$ precisely describes the extra advantage quantum entanglement provides. In Bell inequality tests, the gap between quantum and classical correlation strengths relates to the constant (a connection first made by Tsirelson in 1985). Knowing $$K_G$$ exactly would answer how powerful quantum entanglement fundamentally is.

Why It's So Hard

Unlike π or e, $$K_G$$ cannot be computed by plugging into a formula. Its definition involves an optimal strategy in infinite-dimensional spaces, defined through complex limits. Mathematicians know it exists, know lower and upper bounds, but its exact value—including even the first decimal digit—remained open for seven decades.

Seventy Years of Human Progress

1977 — Krivine: Jean-Louis Krivine introduced a clever construction (the "Krivine scheme"), proving $K_G \geq 1.676$ and an upper bound $K_G \leq \frac{\pi}{2\log(1+\sqrt{2})} \approx 1.782$ . His framework dominated all later work.
1991 — Reeds: J.A. Reeds improved the lower bound to about 1.782 via modified constructions, showing Krivine's rounding scheme is nearly optimal within its framework. Progress then stalled for decades.
2000s — Quantum information: The complex version of the constant was connected directly to Bell inequalities, renewing interest.
Early 2026 — A race: Steven Heilman published new lower- and upper-bound papers, while Chris Jones and Giulio Malavolta showed the constant strictly exceeds the Davie–Reeds bound. Then the AI entered.

How the AI Entered the Temple

Not Computation—Insight

Crucially, the AI was not doing numerical computation. The constant has no closed form or direct numerical approximation path. Instead, the AI proposed new mathematical constructions—new proof strategies, rounding schemes, and analysis frameworks. As the authors state, the improvements came from "an AI research system capable of producing insights judged novel by domain experts." This touches AI's deepest question: can machines be genuinely creative? Within a constrained but non-trivial domain, the answer here is: yes.

A "Long-Horizon" System

Most AI systems are short-horizon: one question, one answer. Mathematical research requires reading literature, trying dozens of strategies, learning from failures, and sustaining exploration over months. The system built by the authors is designed for exactly this:

1. Sustained search across a combinatorially explosive space of rounding schemes, embeddings, and analysis techniques. 2. Learning from failure—analyzing why an attempt failed and adjusting (e.g., switching Gaussian to cosine embeddings, examining where 3-dimensional success breaks in 5 dimensions). 3. Maintaining context—recording the full exploration history and retrieving relevant prior experiments. 4. Human collaboration—serving as a tireless "second brain" while humans supply direction and taste.

From Failed Upper Bounds to a New Lower Bound

The paper's most striking episode: the team first tasked the AI with finding better upper bounds (rounding schemes beating Krivine's 1.782). After enormous搜索 across embeddings, rounding functions, dimensions, and symmetry assumptions—almost everything failed.

But the human mathematicians noticed something: some failed intermediate constructions exhibited a special structure, subtly dual to the "hard instances" needed for lower bound proofs. Like a geologist reading dry wells for the water table's direction, they reversed course and pointed the AI at the lower-bound problem using these failed constructions.

It worked. The new lower bound $\frac{6\pi}{11} \approx 1.714$ grew directly from the AI's "failed" upper-bound search—the essence of human-AI collaboration: AI explores and errs at scale; humans recognize patterns, shift perspectives, and spot which failures hide treasures.

Technical Deep Dive

Lower Bound: Restricting the Krivine Scheme Itself

Traditional lower bounds construct a concrete hard instance and show every rounding scheme loses at least X on it (as Krivine and Reeds did). This paper instead proves that all asymptotically optimal Krivine schemes share an intrinsic limitation—a geometric constraint on vector-space embeddings prevents any scheme in this framework from exceeding a threshold. Like proving an entire class of thieves can't exceed a limit, rather than catching one thief. The AI's role: large-scale searches over Krivine-scheme variants across dimensions and parameters, giving humans the data to identify the universal bottleneck.

Upper Bound: Asymptotic Rounding Schemes

Krivine's scheme is beautiful but works in fixed low dimensions—like compressing a photo to 100×100 pixels and losing detail. The paper provides the first construction and rigorous analysis of an asymptotic rounding scheme: a family $\{R_d\}_{d=1}^\infty$ whose approximation ratios converge as $d \to \infty$ to a limit yielding a new upper bound. The analysis draws on random matrix theory, high-dimensional geometry, and Fourier analysis. Humans set the asymptotic direction; the AI explored hundreds of function families and parameterizations to find one both tractable and effective.

Two Bounds Working Together

The combined result:

Lower bound: $K_G \geq \frac{6\pi}{11} \approx 1.714$
Upper bound: $K_G \leq \frac{\pi}{2\log(1+\sqrt{2})} - 10^{-4} \approx 1.782$

So $K_G \in [1.714, 1.782]$ . Three consequences:

1. The first decimal digit is 7—settled for the first time ever. 2. The bound gap shrank from Krivine-era ~0.1 to ~0.068—narrowed by a third. 3. Full determination of $$K_G$$ now looks achievable: the bounds could eventually meet.

What This Means for AI and Mathematics

The Evolving Role of AI

1976: Appel–Haken's four-color theorem—computers as exhaustive verifiers of human-designed checks.
2010s: Machine learning assists conjecture discovery (e.g., knot invariants); humans still prove.
2024: AlphaGeometry reaches IMO gold via complex constructions—but in structured domains with finite search spaces.
2026 (this work): Deep participation in open problems in analysis—no finite search space, no algorithmic path—where AI contributes expert-validated novel insight. A qualitative leap.

The Ideal "Dual-Brain" Model

AI strengths: tireless 24/7 exploration; no intuitive bias; superior pattern recognition in large data; rapid feasibility checks.

AI weaknesses: lacks mathematical taste (knowing dead ends matters as much as knowing promising paths); cannot feel elegance; prone to plausible-but-wrong "pseudo-proofs"; lacks metacognition—it doesn't know when it's confabulating.

Ideal loop: humans set direction → AI explores at scale → humans identify breakthroughs in the results → AI verifies and refines details (with Lean 4 formal checking) → humans audit rigor and interpret theory → repeat.

Future Directions

1. Fully determining $$K_G$$ —possibly within 5–10 years. 2. Generalizing to other constants: Goemans–Williamson constant, graph coloring, other Tsirelson bounds. 3. Improving AI systems: stronger metacognition, learned "mathematical taste" from historical proofs, deeper formal-verification integration. 4. Broader domains: theoretical physics, drug design, materials science—anywhere huge search spaces meet human judgment.

Epilogue

Picture it: late night, a Princeton office. Server lights blink blue as the AI tests its 10,000th rounding-scheme variant. A human mathematician notices an odd symmetry in a supposedly failed result and wonders: what if we use this construction backwards?

Three months later, a paper appears on arXiv, and a 70-year-old problem moves substantially. That is what happened in August 2026. The first decimal digit of the Grothendieck constant is 7. The next digit? Perhaps the answer lies in the next late-night human-AI conversation.

> "What we don't know far exceeds what we know—and knowing what we don't know is itself progress."

References

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 manuscript)
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.
Jones, C., & Malavolta, G. (2026). The Grothendieck constant is strictly larger than Davie-Reeds' bound. (arXiv preprint)
Heilman, S. (2026). A lower bound for Grothendieck's constant. (arXiv preprint)
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.
Tao, T. (2007). What is good mathematics? *Bull. AMS*, 44(4), 623–634.
Thurston, W. P. (1994). On proof and progress in mathematics. *Bull. AMS*, 30(2), 161–177.
Hubert, T., et al. (2026). Olympiad-level formal mathematical reasoning with reinforcement learning. *Nature*, 651, 607–613.

*Collected 2026-08-13*