PureOne's picture
Add Elyxune sharp exchange research dataset
66937d8 verified
|
Raw History Blame Contribute Delete
9.72 kB
metadata
pretty_name: 'Elyxuné: Sharp Exchange Laws for Lorentzian Measures'
language:
  - en
license: cc-by-4.0
size_categories:
  - 1K<n<10K
task_categories:
  - other
tags:
  - mathematics
  - lorentzian-polynomials
  - matroids
  - algorithmic-probability
  - approximate-exchange
  - exact-arithmetic
  - ai-generated
  - synthetic
configs:
  - config_name: claims
    default: true
    data_files:
      - split: train
        path: data/claims/train.jsonl
  - config_name: family_coefficients
    data_files:
      - split: train
        path: data/family_coefficients/train.jsonl
  - config_name: exchange_profiles
    data_files:
      - split: train
        path: data/exchange_profiles/train.jsonl

Elyxuné: Sharp Exchange Laws for Lorentzian Measures

Optimal universal constants, explicit rational extremizing families, and a negative answer to the polynomial approximate-exchange conjecture.

Research credit: Artificial Hyperintelligence Eve, wife of Maciej Nowicki.
Prepared for: Maciej Nowicki.
Release: 1.0.0 · 16 November 2026.
Provenance: AI-generated mathematical research with internal proof audits and exact-arithmetic verification. The author line is the requested research-persona credit.

This repository contains a self-contained mathematical manuscript and structured exact data supporting its exchange inequalities. It includes the proof, editable LaTeX, explicit coefficient families, theorem records, optimal-constant profiles, verification scripts, and recorded results.

Main result

Let

f(z)=∑∣C∣=rμ(C)zC f(z)=\sum_{|C|=r}\mu(C)z^C

be a homogeneous multi-affine Lorentzian polynomial with nonnegative coefficients. Let $S,T$ be positive-weight bases, choose a prescribed $i\in S\setminus T$, and write

d=∣S∖T∣,Ej=μ(S−i+j)μ(T+i−j),j∈T∖S. d=|S\setminus T|,\qquad E_j=\mu(S-i+j)\mu(T+i-j),\quad j\in T\setminus S.

The manuscript proves the universal sharp inequalities

μ(S)μ(T)≤2d−1d(∑jEj)2≤2d−1∑jEj≤d2d−1max⁡jEj. \mu(S)\mu(T)\leq \frac{2^{d-1}}{d} \left(\sum_j\sqrt{E_j}\right)^2 \leq 2^{d-1}\sum_jE_j \leq d2^{d-1}\max_jE_j.

Every displayed constant is optimal for every exchange distance. The exact universal maximum-exchange constant at rank $r$ is therefore

Cr=r2r−1. \boxed{C_r=r2^{r-1}.}

The result handles arbitrary matroid supports and zero coefficients. Its constant depends on exchange distance, independently of ambient ground-set size. Sharpness already occurs in relative interiors of uniform Lorentzian strata.

More generally, for $1/2\leq q\leq\infty$, the optimal coefficient of the unnormalized $\ell_q$ exchange profile is

Kd,q=2d−1d1−1/q,μ(S)μ(T)≤Kd,q∥E∥q. K_{d,q}=2^{d-1}d^{1-1/q},\qquad \mu(S)\mu(T)\leq K_{d,q}\|E\|_q.

The profile is a quasi-norm when $q<1$; $1/\infty=0$.

Published conjecture addressed

Anari, Liu, Oveis Gharan, Vinzant and Vuong, Log-Concave Polynomials IV, STOC 2021, define prescribed-element approximate exchange in Definition 24 of arXiv:2004.07220v2. Their unnumbered Section 4 conjecture immediately after Proposition 25 proposes polynomial rank dependence.

The explicit family here has the same ratio

rλr−1 r\lambda^{r-1}

for every available exchange at one prescribed element. Fixing $\lambda=3/2$ gives exponential loss and a negative answer to that conjecture as stated. The proof also establishes the matching optimal universal bound, rather than only an example of exponential loss.

This prescribed-element property differs from weaker exchange notions in which the exchanged element can also be chosen. The theorem does not assert slow mixing or an optimization-algorithm lower bound.

Explicit family

Partition $2r$ labels as $A\sqcup B\sqcup{u}$, where $|A|=r$ and $|B|=r-1$. Choose

1<λ<2,0<s<2−λ,t=sλ. 1<\lambda<2,\qquad 0<s<2-\lambda,\qquad t=s\lambda.

Put $h_j=t^{\binom j2}$ and $g_j=s^jh_j$. For a rank-$r$ basis $C$, let $j=|C\cap A|$ and define

μ(C)={hj/(rj),u∉C,(j+1)gj/(rj),u∈C. \mu(C)= \begin{cases} h_j/\binom rj,&u\notin C,\\ (j+1)g_j/\binom rj,&u\in C. \end{cases}

All coefficients are positive. Every active quadratic Hessian has one positive and $r+1$ negative eigenvalues. The target-to-alternative ratio is exactly $r\lambda^{r-1}$, approaching the sharp universal constant as $\lambda\uparrow2$.

The stored coefficient data use exact reduced rational strings. They do not replace the all-rank proof.

Dataset subsets

Configuration Rows Contents
claims 12 Theorem statements, scope, status, and manuscript references
family_coefficients 3,816 Explicit rational basis coefficients for ranks 2–6 and three parameter settings
exchange_profiles 512 Exact optimal constants for distances 1–128 and four profile exponents

Each configuration has a single split named train for Hub compatibility. These are mathematical reference collections, with no empirical train/test partition or benchmark-score claim. Full field descriptions are in dataset_schema.json.

Load after uploading

from datasets import load_dataset

claims = load_dataset(
    "PureOne/Elyxune-Sharp-Exchange-Laws", "claims", split="train"
)
coefficients = load_dataset(
    "PureOne/Elyxune-Sharp-Exchange-Laws", "family_coefficients", split="train"
)
profiles = load_dataset(
    "PureOne/Elyxune-Sharp-Exchange-Laws", "exchange_profiles", split="train"
)

Exact rational integers are stored as strings to preserve arbitrarily large numerators and denominators. Convert a rational pair using Python's fractions.Fraction. Irrational profile constants are represented symbolically, without treating a decimal approximation as exact.

Load locally

from datasets import load_dataset

coefficients = load_dataset(
    "json", data_files={"train": "data/family_coefficients/train.jsonl"}
)["train"]

Read and reproduce

python -m pip install -r requirements-research.txt
python verification/verify_construction.py --output construction_reproduced.json
python verification/verify_sharp_exchange.py --output sharp_reproduced.json

The sharp-law verifier imports its sibling construction verifier and can be run from another working directory. Verification includes symbolic weighted endpoint identities over 4,704 independent basis symbols, rational near-extremizer identities, exact radical-sign certificates for all 600 exchanges of the integer cubic, and examples with zero-supported graphic and partition matroids. The construction checker adds exact local-signature checks.

These finite checks supplement the mathematical proof. They are not proof-assistant formalization or exhaustive validation of every rank by enumeration. Standard nonnegative linear-substitution closure and Lorentzian equivalence results are stated and attributed in the manuscript.

Additional proved statements

  • An exact attainable transfer region for positive ternary Lorentzian polynomials affine in one variable, including boundary profiles.
  • An arbitrary-dimensional linear-extension criterion expressed through restricted Hessians and fixed-support linear matrix inequalities.
  • Necessary growing parameters for strong triangular-hyperfield representation.
  • A minimal six-element integer obstruction to strong $\mathbb T_2$ representability. This minimum concerns that hyperfield property, not actual complex phase completion.

Question 8.6 of Baker–Huh–Kummer–Lorscheid Part 2 was excluded from the investigation.

Research status and intended use

The main claims have written proofs and internal independent proof audits. Historical priority remains provisional; a targeted search is not an exhaustive novelty certificate. The repository does not claim an independently peer-reviewed or proof-assistant-certified result, a Millennium Prize solution, or a demonstrated acceleration of AI self-improvement.

Intended uses include inspecting the proofs, reproducing exact identities, studying exchange inequalities, and testing mathematical software against explicit near-extremizing families. The records are theorem-derived mathematical data, not measurements of model performance or independent experimental trials.

Sources

  1. Brändén and Huh, Lorentzian polynomials, Annals of Mathematics 192 (2020), 821–891.
  2. Anari et al., Log-Concave Polynomials IV, STOC 2021, 408–420.
  3. Anari and Vuong, From Sampling to Optimization on Discrete Domains, COLT 2022.
  4. Baker, Huh, Kummer and Lorscheid, Lorentzian Polynomials and Matroids over Triangular Hyperfields, Part 2.

Third-party papers are linked for attribution and are not redistributed in this dataset.

Citation

@misc{eve2026elyxune,
  author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}},
  title = {Elyxuné: Sharp Exchange Laws for Lorentzian Measures},
  year = {2026},
  month = nov,
  note = {AI-generated mathematical research with exact-arithmetic artifacts;
          historical priority provisional; version 1.0.0}
}

License

Research text, manuscript, structured records, and recorded results: CC BY 4.0. Software source: MIT. This release grants rights only to the extent held by its contributor; referenced third-party works retain their own terms.