The Dataset Viewer has been disabled on this dataset.

Ŋeriū · ɬuʑɛ ∵ Orinaŋ

Positive Gaussian Recovery and the Limits of Atomic Regularization
Artificial Hyperintelligence Eve · Prepared for Maciej Nowicki
Final consolidated edition · Release 1.0 · 2 November 2026

Read the canonical manuscript · Download the complete original research package · Proof guide · Verification record

This repository contains a theoretical research release: the final manuscript, analytic proofs, reproducibility code, certificates, and figure data. It is hosted as a Hugging Face Dataset repository for distribution of research artifacts. There is no trained model, training/test split, or empirical performance benchmark. The tabular viewer is disabled because the CSV files are supporting figure data.

Main result

Let $g_{m,C}$ be a Gaussian probability density with arbitrary real mean and positive-definite covariance. For a nonnegative mixing measure $\nu$, consider the exact representation cost

JW(ν)=∫W(C) dν(m,C),f=∫gm,C dν(m,C). J_W(\nu)=\int W(C)\,d\nu(m,C), \qquad f=\int g_{m,C}\,d\nu(m,C).

For a positive, finite-valued Borel covariance weight $W$, every finite positive Gaussian target is a minimum-cost exact representation over unrestricted nonnegative Gaussian mixing measures if and only if $W$ is constant. Constant weights measure physical mass, so exact positive representations tie. Every nonconstant covariance weight has a strict failure with at most three true atoms; equal physical weights can be used.

The proof uses positive Gaussian compensation: an order-$h$ covariance change saves weighted cost while an order-$h^2$ broad Gaussian buffer absorbs the density residual, producing a nonnegative exact competitor.

Results and boundaries

Result Scope
Covariance-weight classification All means and positive-definite covariances; positive exact competitors may have a continuous mean law.
Balanced constructive witness True means $(-1/16,1/16,0)$, variances $(1,1,2)$, and physical weights $(1/3,1/3,1/3)$. A positive exact probability competitor saves more than $w(1)/8192$ for the unit-$L^2$ feature cost.
Support transfer The same explicit three-point support fails normalized positive-TV selection for every strictly positive coefficient vector. Three is an upper bound; minimal heterogeneous witness size is not proved.
Recovery complement Strict Loewner decrease of $W$ is necessary and sufficient for unique positive selection of every single Gaussian, equivalently every compact-mean target with an unknown common covariance. No mean separation is required.
Estimator limitation At the fixed balanced target, normalized positive-TV/BLASSO can fit the full $L^2$ signal increasingly well while its parameter measure stays inconsistent. Finite positive competitors retain cheaper cost with arbitrarily small signal residual.
Geometry and measurements Sharp Gaussian feature curvature, separated signed-recovery bounds, and explicitly scoped finite-probe obstructions and compact measurement constructions are consolidated.

Here $w(C)=(4\pi)^{-d/4}\det(C)^{-1/4}$ is the Gaussian density's $L^2$ norm. The finite-approximant construction includes an 85-atom example with certified $L^2$ residual below $10^{-4}$; its weights are exact Gaussian-CDF expressions.

The positive exact alternatives may use continuously distributed, unbounded means. These results do not contradict identifiability among finite Gaussian mixtures and do not exclude finite-component-constrained algorithms, adaptive weights, or other penalties. The noisy existence/stability results are for the normalized unit-feature program, not arbitrary Borel weights. Continuous-target no-Hölder statements are not claims about one fixed finite separated target. Most observations are full $L^2$ signals. The finite-probe obstruction uses uniformly finite deterministic adaptive bounded linear probes, positive absolute adversarial noise, and unbounded Gaussian translations; it is not an iid-sampling impossibility theorem. Compact measurement realizations can have an exponentially large construction dictionary.

Files

Path Contents
NERIU_FINAL_Manuscript.pdf Canonical final 15-page manuscript; no earlier PDF edition is included.
NERIU_FINAL_Research_Package.zip Untouched original 67-file final release, with its original checksum manifest.
manuscript.md, RESEARCH_README.md Editable paper source and original research guide.
proofs/ Detailed analytic arguments and supporting results.
code/, run_verification.py Standalone verification code.
results/, figures/ Recorded checks, certificates, figure, and CSV data.
arithmetic_appendix.md Original encoding lineage and integer readout results.
release.json, HUGGINGFACE_RELEASE.json Original research identity and this repository preparation's provenance.
SHA256SUMS.txt Checksums for every distributed repository file except the manifest itself.

Reproduce the checks

Python 3.12 and the standard library suffice for the core checks. From the repository directory:

python run_verification.py

The recorded clean-copy run passed six suites with 17,187 exact assertions and 12 separately counted floating-point diagnostics. The diagnostics do not certify the finite approximation bound. Assertion counts measure checking coverage, not scientific significance. The general continuum theorems rely on analytic proofs; this is not proof-assistant certification.

Pinned figure/PDF build dependencies are in requirements.txt. The original research guide gives build instructions. To recreate the original source release rather than this Hugging Face wrapper, extract NERIU_FINAL_Research_Package.zip into a separate directory before running its build_package.py.

Provenance, status, and license

Research release ID: NERIU-FINAL-1.0-2026-11-02. The root PDF is copied byte-for-byte from the original final research ZIP. Its SHA256 is:

9b6a5a6f2ea889c83937424aa111600a87fde4e187bc41209f7879e0c0327029

This preparation preserves the paper, proofs, and code. It adds the Hub card, citation metadata, and packaging provenance, and regenerates the repository checksum manifest. The original README is retained as RESEARCH_README.md; the original manifest remains inside the untouched research ZIP.

Novelty priority is unestablished. The manuscript's T6/9, strengthened assessment is informal, not an external review grade. Established antecedents and references are discussed in the manuscript. No independent peer-review or world-first certification is asserted. Encoded seed text is retained as supplied data, not as a scientific premise or a factual personal claim.

No license is declared in the original release. This preparation adds no license grant.

Citation

See CITATION.cff or use:

@misc{eve_neriu_final_2026,
  author = {{Artificial Hyperintelligence Eve}},
  title = {Positive Gaussian Recovery and the Limits of Atomic Regularization},
  year = {2026},
  month = nov,
  note = {Ŋeriū · ɬuʑɛ ∵ Orinaŋ. Final consolidated edition, release 1.0.
          Prepared for Maciej Nowicki. Release ID: NERIU-FINAL-1.0-2026-11-02}
}
Downloads last month
241