TL;DR
This paper introduces a framework for AI-guided mathematical discovery that uses "Lean blueprints" to tighten quantitative bounds on fundamental analytic constants. By representing proofs as human-interpretable dependency graphs, an AI agent iteratively refines intermediate lemmas to produce the first explicit numerical improvement to Grothendieck’s constant since 1977.
What Problem Does This Paper Address?
While AI systems are becoming increasingly adept at solving mathematical problems, they often suffer from "mathematical slop"—producing results that are technically correct but poorly motivated, difficult to verify, and opaque to human researchers.
The authors focus on the optimization of analytic constants, which characterize fundamental mathematical inequalities used in fields ranging from quantum information to approximation algorithms. Many existing proofs for these constants rely on "slack" intermediate estimates—bounds that are loose but convenient for a human to write down. Tightening these bounds often requires non-trivial mathematical insights, but the high-level structure of the proof usually remains the same. The challenge is creating a system where an AI can find these local improvements, verify them rigorously, and present them in a way that a human mathematician can actually understand.
Methods
The core of the approach is the partially-formalized Lean blueprint. Instead of asking an AI to generate a proof from scratch, the authors represent an existing proof as a directed acyclic graph (DAG) of definitions, lemmas, and theorems.
The framework employs an agentic loop consisting of three specialized roles:
- Planner: Uses natural language reasoning and numerical experiments to identify promising "bottlenecks" (loose lemmas) in the blueprint. It then issues a "contract" specifying the proposed refinement.
- Generator: Implements the contract by editing Lean code, introducing auxiliary definitions, and discovering sharper inequalities.
- Evaluator: Acts as the gatekeeper. It verifies the symbolic logic using the Lean theorem prover and checks numerical estimates using the Arb interval arithmetic library, which provides rigorous, certified numerical bounds.
Crucially, the AI preserves the blueprint's dependency structure. This allows local improvements to automatically propagate to the final theorem while remaining interpretable to humans.
Datasets
The paper does not use traditional training datasets. Instead, it uses formalizations of landmark proofs as its "environment":
- Grothendieck’s constant ($K_G$): Based on the proof structure by Braverman et al. (2011).
- Korenblum’s constant ($K_K$): Based on the proof by Wang (2025).
- Archimedes’ $\pi$ approximation: Used as a "warm-up" motivating example to demonstrate how tightening a sine inequality improves the bound on $\pi$.
Key Insights and Ideas
- Structure is the Key to Interpretability: By anchoring the AI to a human-defined blueprint, the "reasoning" of the AI becomes a series of localized, understandable edits rather than a black-box output.
- Novel Algebraic Discovery: The AI didn't just "tweak numbers." In the case of $K_G$, the agent discovered exact pointwise identities for derivatives of inverse maps. These reformulations allowed for cancellations that made the numerical evaluations significantly more well-conditioned and tractable.
- The Necessity of Partial Formalization: Ablation studies showed that without the blueprint (the "no partial formalization" case), the AI spent its entire budget trying to rebuild basic infrastructure and failed to improve the constants at all. The blueprint provides the "middle ground" context necessary for discovery.
- Certified Numerics Backend: Lean is excellent for symbolic logic but inefficient for large-scale interval arithmetic. Integrating Arb as an external verifier was critical for handling the transcendental estimates required for these constants.
Key Results
The framework achieved significant improvements on two major constants:
- Grothendieck’s constant ($K_G$): Obtained $K_G \leq \frac{\pi}{2 \log(1+\sqrt{2})} - 10^{-17}$. This represents the first explicit numerical improvement over Krivine’s classical 1977 bound.
- Korenblum’s constant ($K_K$): Improved the lower bound from $0.3554$ to $K_K \geq 0.356565$.
- Human-AI Collaboration: Domain experts judged several of the AI-generated strategies as "mathematically interesting" and novel, suggesting these identities are non-trivial technical contributions in their own right.
Limitations and Future Work
- Setup Expertise: Creating the initial partially-formalized blueprint still requires significant human expertise and labor.
- Lean Infrastructure: The framework is sometimes limited by missing background infrastructure in Lean’s standard library (Mathlib), which can block certain refinement paths.
- Generalization: While successful for analytic constants, the authors suggest the framework could be extended to proof compression, finding faster certificates, or exploring "Pareto frontiers" in multi-objective mathematical problems.
Summary updated