Formalizing progress on Epoch AI's FrontierMath open problem on hypergraphs.
This repository formalizes lower bounds for the extremal hypergraph function $H(n)$ introduced by Will Brian and Paul B. Larson in Choosing between incompatible ideals. The sequence arises as the finitary core of a Ramsey-theoretic problem about incompatible ideals and the simultaneous convergence of infinite series.
A hypergraph $(V,\mathcal H)$ is said to contain a partition of size $n$ if there are $D \subseteq V$ and $\mathcal P \subseteq \mathcal H$ with $|D| = n$ such that every member of $D$ belongs to exactly one edge of $\mathcal P$. The extremal function $H(n)$ is the largest $k$ for which there is a hypergraph with $k$ vertices, no isolated vertices, and no partition of size greater than $n$.
The FrontierMath page presents warm-up, single-challenge, and full-problem variants. This repository addresses the full-problem variant: it formalizes an explicit construction showing
$$ H(n) \ge \frac{26}{25},k_n \qquad (n \ge 15), $$
where Brian–Larson's recursive benchmark is given by
$$ k_1 = 1, \qquad k_n = \lfloor n/2 \rfloor + k_{\lfloor n/2 \rfloor} + k_{\lfloor (n+1)/2 \rfloor}. $$
Thus the constant-factor improvement is already in effect at $n=15$, exactly as requested in the FrontierMath full problem.
Asymptotic consequence. The repository also formalizes GPT-5.4 Pro's Lubell-family lower bound, which implies
$$ \liminf_{n\to\infty} \frac{H(n)}{k_n} \ge 2\ln 2. $$
Combined with Brian–Larson's upper bound (not formalized here)
$$ H(n) < n\ln n + \gamma n + \tfrac12, $$
this yields the sharp asymptotic formula
$$ \lim_{n\to\infty} \frac{H(n)}{n\ln n} = 1. $$
Note: on the finite side, the formalized construction gives $H(20) \ge 65$. So this development resolves the FrontierMath full-problem prompt, while stopping one vertex short of the single-challenge target $H(20) \ge 66$.
This Lean development is based on an informal proof produced by GPT-5.4 Pro. The prompters of that informal proof were Kevin Barreto and Liam Price; the original paper output is archived on Google Drive here.
The substitution theorem underlying the recursive witness constructions is formalized in Lean.
Explicit uniform lower bound
$$ H(n) \ge \frac{26}{25},k_n \qquad (n \ge 15). $$
$$ H(n) \ge \frac{h_t-1}{\log_2 t}n\log_2 n - O_t(n) \qquad \left(t \ge 2 \text{ fixed},\qquad h_t:=\sum_{n=1}^{t}\frac{1}{n}\right). $$
$$ \liminf_{n\to\infty} \frac{H(n)}{k_n} \ge 2\ln 2, \qquad k_n = \frac12 n\log_2 n + O(n). $$
The core development is organized as follows:
FrontierMathHypergraphs/Basic.lean — hypergraph definitions, partitions, the extremal function H, and the benchmark sequence k.FrontierMathHypergraphs/Substitution.lean — support patterns, frames, substitution hypergraphs, and the substitution theorem.FrontierMathHypergraphs/Uniform.lean — the finite bootstrap and the uniform $26/25$ lower bound.FrontierMathHypergraphs/Lubell.lean — Lubell frames and the asymptotic theory.blueprint/src/content.tex — blueprint source.frontier.tex and paper/input.tex — paper source.Compile the Lean files (requires Lean):
lake exe cache get && lake build
Build the blueprint PDF (requires uv):
uvx leanblueprint pdf
Build and serve the blueprint website:
uvx leanblueprint web && uvx leanblueprint serve
1 commits
Lean
70.7%
TeX
29.3%
Formalizing progress on Epoch AI's FrontierMath open problem on hypergraphs.
This repository formalizes lower bounds for the extremal hypergraph function $H(n)$ introduced by Will Brian and Paul B. Larson in Choosing between incompatible ideals. The sequence arises as the finitary core of a Ramsey-theoretic problem about incompatible ideals and the simultaneous convergence of infinite series.
A hypergraph $(V,\mathcal H)$ is said to contain a partition of size $n$ if there are $D \subseteq V$ and $\mathcal P \subseteq \mathcal H$ with $|D| = n$ such that every member of $D$ belongs to exactly one edge of $\mathcal P$. The extremal function $H(n)$ is the largest $k$ for which there is a hypergraph with $k$ vertices, no isolated vertices, and no partition of size greater than $n$.
The FrontierMath page presents warm-up, single-challenge, and full-problem variants. This repository addresses the full-problem variant: it formalizes an explicit construction showing
$$ H(n) \ge \frac{26}{25},k_n \qquad (n \ge 15), $$
where Brian–Larson's recursive benchmark is given by
$$ k_1 = 1, \qquad k_n = \lfloor n/2 \rfloor + k_{\lfloor n/2 \rfloor} + k_{\lfloor (n+1)/2 \rfloor}. $$
Thus the constant-factor improvement is already in effect at $n=15$, exactly as requested in the FrontierMath full problem.
Asymptotic consequence. The repository also formalizes GPT-5.4 Pro's Lubell-family lower bound, which implies
$$ \liminf_{n\to\infty} \frac{H(n)}{k_n} \ge 2\ln 2. $$
Combined with Brian–Larson's upper bound (not formalized here)
$$ H(n) < n\ln n + \gamma n + \tfrac12, $$
this yields the sharp asymptotic formula
$$ \lim_{n\to\infty} \frac{H(n)}{n\ln n} = 1. $$
Note: on the finite side, the formalized construction gives $H(20) \ge 65$. So this development resolves the FrontierMath full-problem prompt, while stopping one vertex short of the single-challenge target $H(20) \ge 66$.
This Lean development is based on an informal proof produced by GPT-5.4 Pro. The prompters of that informal proof were Kevin Barreto and Liam Price; the original paper output is archived on Google Drive here.
The substitution theorem underlying the recursive witness constructions is formalized in Lean.
Explicit uniform lower bound
$$ H(n) \ge \frac{26}{25},k_n \qquad (n \ge 15). $$
$$ H(n) \ge \frac{h_t-1}{\log_2 t}n\log_2 n - O_t(n) \qquad \left(t \ge 2 \text{ fixed},\qquad h_t:=\sum_{n=1}^{t}\frac{1}{n}\right). $$
$$ \liminf_{n\to\infty} \frac{H(n)}{k_n} \ge 2\ln 2, \qquad k_n = \frac12 n\log_2 n + O(n). $$
The core development is organized as follows:
FrontierMathHypergraphs/Basic.lean — hypergraph definitions, partitions, the extremal function H, and the benchmark sequence k.FrontierMathHypergraphs/Substitution.lean — support patterns, frames, substitution hypergraphs, and the substitution theorem.FrontierMathHypergraphs/Uniform.lean — the finite bootstrap and the uniform $26/25$ lower bound.FrontierMathHypergraphs/Lubell.lean — Lubell frames and the asymptotic theory.blueprint/src/content.tex — blueprint source.frontier.tex and paper/input.tex — paper source.Compile the Lean files (requires Lean):
lake exe cache get && lake build
Build the blueprint PDF (requires uv):
uvx leanblueprint pdf
Build and serve the blueprint website:
uvx leanblueprint web && uvx leanblueprint serve
1 commits
Lean
70.7%
TeX
29.3%