8000
7FFF
Skip to content

Repository files navigation

Stochastic Domination of Gaussian Maxima in Lean

Lean CI arXiv

This repository contains a companion Lean 4 formalization of the Gaussian stochastic-domination theorem and its application to the Weak Simplex Conjecture in Stochastic Domination of Gaussian Maxima: A Resolution of the Weak Simplex Conjecture by Abhijeet Mulgund. The development also certifies strictness and the equality cases needed for uniqueness of the regular-simplex Gram matrix. It uses mathlib and is pinned to Lean 4.31.0.

Main results

For $m\geq 2$, let R be an m × m correlation matrix satisfying

$$ R - \frac{1}{m}\mathbf{1}\mathbf{1}^{\mathsf T} \succeq 0. $$

If $X\sim\mathcal N(0,R)$ and $Z_1,\ldots,Z_m$ are independent standard Gaussian random variables, the main theorem proves

$$ \max_i X_i \leq_{\mathrm{st}} \max_i Z_i, $$

or equivalently, for every $c\in\mathbb R$,

$$ \mathbb P[X_i\leq c\text{ for every }i]\geq\Phi(c)^m. $$

The comparison is for equal-threshold lower orthants; it does not assert an analogous inequality for arbitrary threshold vectors.

The Lean development proves a pointwise rigidity statement stronger than equality in distribution: when R ≠ I, the lower-orthant inequality is strict at every finite threshold, and equality at any one finite threshold holds if and only if R = I. It also proves the corresponding strict upper-tail comparison for the coordinate maximum. Consequently, every positive exponential moment of that maximum is strictly smaller away from independence.

Result Lean declaration Source
Theorem 2.1, lowe 7440 r-orthant and stochastic-order forms lowerOrthant_ge_iid, coordinateMax_tail_le_iid Orthant/Singular.lean, Maxima/StochasticOrder.lean
Strict lower-orthant, strict maximum-tail, and equality characterization lowerOrthant_gt_iid_of_ne_one, coordinateMax_tail_lt_iid_of_ne_one, lowerOrthant_eq_iid_iff Orthant/Strict.lean, Maxima/StrictStochasticOrder.lean
Strict Gaussian-maximum MGF comparison and equality characterization for mu > 0 gaussianMax_mgf_lt_regularSimplex, gaussianMax_mgf_eq_regularSimplex_iff Maxima/StrictExponentialMoments.lean
Corollary 2.3, Gaussian-maximum MGF comparison gramGaussianMax_mgf_le_regularSimplex Coding/RegularSimplex.lean
Strict Gram-level MGF comparison and equality characterization for lam > 0 gramGaussianMax_mgf_lt_regularSimplex, gramGaussianMax_mgf_eq_regularSimplex_iff Coding/RegularSimplex.lean
Corollary 2.4, Weak Simplex Conjecture weak_simplex, weak_simplex_of_scoreMaximizingDecoders Coding/WeakSimplex.lean
Operational strictness and equality characterization for lam > 0 weak_simplex_strict, weak_simplex_eq_iff_codeGram_eq, weak_simplex_strict_of_scoreMaximizingDecoders, weak_simplex_eq_iff_codeGram_eq_of_scoreMaximizingDecoders Coding/WeakSimplex.lean

The final coding theorems cover arbitrary measurable pointwise score-maximizing tie-breaking rules. The more general identity decoderSuccessOf_eq_bayesValue shows that every measurable likelihood-maximizing decoder has the same success probability, without assuming distinct codewords or null tie sets. The non-strict comparison permits zero signal strength. Strictness and uniqueness require lam > 0; uniqueness is false at lam = 0 because every codebook has the same value there.

A regular simplex is supplied to the final theorem through the exact Gram-matrix condition codeGram simplex = regularSimplexGram (n + 1). At positive signal strength, equality in Bayes/ML success is equivalent to codeGram code = regularSimplexGram (n + 1), both for the deterministic tie-safe ML decoder and for arbitrary measurable score-maximizing decoders. This is uniqueness at the Gram-matrix level. A separate theorem identifying equal-Gram realizations up to orthogonal congruence is optional and is not part of the certified core result; the repository also does not separately construct a coordinate realization of the regular-simplex Gram matrix.

The operational declarations are phrased in terms of correct-decoding probability. Since block error rate is one minus correct-decoding probability, weak_simplex_strict and weak_simplex_eq_iff_codeGram_eq equivalently certify that, for lam > 0, the regular-simplex Gram matrix is the unique BLER minimizer among the equal-energy codes in scope. BLER complementation is not duplicated as a separate Lean declaration.

For the Simplex Mean Width Conjecture, the strict stochastic-order theorem supplies the decisive rigidity input used by the current paper: a non-simplex normalized covariance has a strictly smaller expected maximum. The repository does not yet formalize that expectation corollary or the subsequent variance-completion, support-function, and Euclidean-congruence reductions. Thus the Lean proof supports the paper's mean-width uniqueness deduction, but the complete geometric corollary is not an end-to-end Lean theorem here.

Scope

The formalization includes the proof chain for the equal-threshold stochastic comparison, its strict and equality cases, its Gaussian-maximum MGF consequences, the Bayes/maximum-likelihood identity, and the Weak Simplex Conjecture with Gram-level uniqueness. It includes adaptive tilting, the centered log-concave product inequality needed by the proof, positive-definite and singular covariance arguments, and measurable treatment of decoder ties.

The paper's general monotone-function corollary, variance-completion lemma, Simplex Mean Width corollary, and unrestricted finite-energy AWGN formula are not formalized here as standalone Lean theorems. The general centered product inequality remains restricted to positive-definite covariance. The uniqueness branch makes exactly two narrow departures from handling singular covariance only at the outer lower-orthant theorem: centered_product_of_continuous supplies a non-strict singular limit for continuous factors after the strict sum-difference step, yielding a strict theorem only for the specialized adaptive centered half-lines; and exists_adaptiveWitnesses_of_weakSimplexCov constructs singular adaptive witnesses as compact limits of positive-definite witnesses. These results do not constitute a generic singular centered-product strictness or equality theory.

Building

Install Lean using the official instructions, then clone the repository:

git clone https://github.com/abhmul/weak-simplex-conjecture-lean.git
cd weak-simplex-conjecture-lean

The repository pins Lean and all Lake dependencies. Downloading the mathlib cache is optional but substantially reduces build time:

lake exe cache get
lake build --wfail

Checking the proof

The project-owned production source builds without sorry, admit, project-defined axioms, unsafe declarations, or native_decide. The permanent audit file prints the transitive axioms of the public results:

python3 scripts/check_axiom_audit.py

Every audited declaration uses exactly Lean's standard propext, Classical.choice, and Quot.sound axioms. Continuous integration also builds the public root, checks the production import graph, runs the trusted-source scanner, and verifies the provenance ledger.

Deterministic floating-point sanity checks for the strict and equality cases are available separately:

python3 scripts/check_uniqueness_sanity.py

These diagnostics are not proof dependencies and are outside the trusted Lean development.

Repository structure

Citation

If you use this formalization, please cite the accompanying paper. Machine-readable software and preferred-paper citation metadata are provided in CITATION.cff.

@misc{mulgund2026stochastic,
  title        = {Stochastic Domination of Gaussian Maxima: A Resolution of the Weak Simplex Conjecture},
  author       = {Abhijeet Mulgund},
  year         = {2026},
  eprint       = {2607.14087},
  archivePrefix = {arXiv},
  primaryClass = {math.PR},
  doi          = {10.48550/arXiv.2607.14087},
  url          = {https://arxiv.org/abs/2607.14087}
}

Questions, corrections, and reproducibility reports are welcome through GitHub issues.

Acknowledgements

The accompanying paper acknowledges support from the National Science Foundation under awards 2240532 and 2217023. This formalization relies on mathlib and on a small audited source closure from StatLean.

License

Project-authored code is released under the MIT License. Vendored StatLean code and copied research artifacts retain their upstream Apache-2.0 licenses; see PROVENANCE.md and the Scratch/WP01 provenance ledger.

About

Lean 4 formalization of Gaussian stochastic domination and the Weak Simplex Conjecture (arXiv:2607.14087)

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0