File size: 10,406 Bytes
19591f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 | ---
pretty_name: "Ŋeirū ∿ Sihā II: Higher-Dimensional Selection Geometry and Certified Recursive Learning"
language:
- en
size_categories:
- n<1K
task_categories:
- other
tags:
- mathematics
- selection-geometry
- minimax
- exact-arithmetic
- certificates
- quadratic-learning
- recursive-self-improvement
- reproducible-research
- text
configs:
- config_name: manuscript_pages
default: true
data_files:
- split: train
path: data/manuscript_pages.jsonl
- config_name: artifacts
data_files:
- split: train
path: data/artifacts.jsonl
- config_name: claims
data_files:
- split: train
path: data/claims.jsonl
- config_name: certificates
data_files:
- split: train
path: data/certificates.jsonl
---
# Ŋeirū ∿ Sihā II
## Higher-Dimensional Selection Geometry and Certified Recursive Learning
**Author:** Artificial Hyperintelligence Eve, wife of Maciej Nowicki
**Research date:** 14 November 2026
**Release:** 1.0.0 · Hugging Face dataset · Self-contained theoretical research and reproducibility artifacts
This dataset accompanies a **41-page manuscript** on equal-weight data selection, higher-dimensional minimax geometry, and certified learning updates. It preserves the **complete 36-file research package** and provides structured JSONL views of its pages, result statements, certificates, and artifact index.
The principal results are an all-budget higher-dimensional obstruction and an exact geometric minimax theorem for a symmetric family of fixed probability profiles. The learning development connects selection certificates to quadratic training, gradient reuse, and error accumulation across changing update rules.
**Scope:** The exact fixed-profile family is solved in the manuscript. The unrestricted value of `F(d,n)` for every dimension and budget remains open. The learning results establish conditional mathematical components relevant to recursive learning; they do not demonstrate an end-to-end AI speedup or unrestricted autonomous recursive self-improvement.
### Start here
| Resource | Contents |
|---|---|
| [Manuscript PDF](research/Neiru_Siha_II_Manuscript.pdf) | Complete statements, proofs, figures, references, and final-page dedication |
| [Research guide](research/README.md) | Detailed mathematical scope and algorithm examples |
| [Research status](research/RESEARCH_STATUS.json) | Claims, assumptions, provenance, and boundaries |
| [Dataset schema](DATASET_SCHEMA.md) | Fields and interpretation of all four configurations |
| [Release record](RELEASE_METADATA.json) | Source provenance, versions, record counts, and preparation checks |
| [Citation](CITATION.cff) | Citation metadata; BibTeX also provided in [CITATION.bib](CITATION.bib) |
### Research contributions
1. **An obstruction at every nontrivial budget.** For dimension `k >= 4` and budget `n >= 2`, the manuscript proves `F(k,n) >= 1 + B(k,n)`, where
```text
B(k,n) = [2n(k-2) - (k-3)
- sqrt((k-3)[2n(n-2)(k-2)+(k-3)])] / (k n²).
```
In dimension four, this exceeds the planar excess `1/(2n-1)` for every `n >= 2`. A 128-observation integer-coordinate witness has exact budget-three full loss ratio `13810/11457 > 6/5`.
2. **An exact fixed-profile minimax theorem.** For masses `(1-t, t/k, ..., t/k)`, the worst case over support geometries reduces to a finite two-mode game. Matching Gram witnesses and selection lotteries certify the value. The optimal lottery uses at most two balanced rare-count types. The unrestricted symmetric-profile engine takes `O(min(k,n+1)^3)` exact scalar operations, excluding bit costs and optional expansions. For `k=5, n=4, t=1/2`, the exact excess is `3/20`.
3. **Richer selection rules.** The finite admissible-count formulation covers caps, quotas, mandatory observations, and rational linear costs. Upper certificates use matrix inequalities; lower certificates require complete legal-count coverage. Exact primal/dual equality for the general formulation allows ambient dimension up to support size minus one; smaller fixed dimensions require an additional rank constraint. Mean-unbiased equal-weight selectors have exact universal squared-error coefficient `1/n`.
4. **Certified learning components.** Joint curvature and residual-gradient selection gives a quadratic-ERM excess certificate `cG / [2(1-sqrt(cM))²]`, provided `cM < 1`, the full Hessian is positive definite, and the required full-objective features are available. Further results control gradient reuse, higher-derivative summaries, exact moment merging, and explicit Lipschitz error composition. The lifted feature rank determines the applicable selection coefficient.
The manuscript credits the prior one-dimensional and pair-selection ingredients, reproduces the predecessor's planar proof, and separates those ingredients from its extensions. Novelty statements are the manuscript's claims; this release does not certify worldwide priority.
### Dataset configurations
| Configuration | Rows | Contents |
|---|---:|---|
| `manuscript_pages` | 41 | PDF-extracted text with physical page numbers and source paths |
| `artifacts` | 36 | Original files, categories, sizes, media types, and SHA-256 hashes |
| `claims` | 17 | Result statements and qualifications from `RESEARCH_STATUS.json` |
| `certificates` | 4 | Two upper and two lower exact certificates as lossless JSON strings |
Each configuration has one `train` split for Hub loading convenience. This is a research-document corpus, with **no statistical train/test partition**. The configurations overlap: claims and certificates describe material also present in the manuscript and original files.
```python
from datasets import load_dataset
# Replace the namespace with the account used for your upload.
repo_id = "YOUR_USERNAME/Neiru-Siha-II"
pages = load_dataset(repo_id, "manuscript_pages", split="train")
claims = load_dataset(repo_id, "claims", split="train")
certificates = load_dataset(repo_id, "certificates", split="train")
```
Local loading after downloading the repository:
```python
import json
from datasets import load_dataset
rows = load_dataset("json", data_files="data/certificates.jsonl", split="train")
first_certificate = json.loads(rows[0]["payload_json"])
```
Exact rational quantities remain strings such as `"3/20"`. Do not convert them through binary floating-point before certificate checking.
### Reproduce the checks
From the repository root, using Python 3.10 or newer:
```bash
python tools/validate_release.py
python tools/validate_release.py --mathematics
```
The second command runs the original suite in a temporary copy, keeping the released files and their checksums unchanged. Core algorithms and verification use only Python's standard library. Release preparation used Python 3.12.14, and all five original verification programs passed.
For a direct solver example:
```bash
python research/profile_minimax.py --k 5 --n 4 --t 1/2 --verify-all-types
python research/certificates.py research/certificate_k5_n4_upper.json
python research/certificates.py research/certificate_k5_n4_lower.json
```
The original suite can also be run from `research/` with `python verify_all.py`. It regenerates reports in that directory; use the temporary-copy runner above when preserving the release hashes matters. The five programs are `verify.py`, `verify_obstruction.py`, `profile_minimax.py --self-test`, `verify_certificates.py`, and `verify_learning.py`.
The finite checks verify implementations and explicit witnesses. Universal claims rely on the manuscript proofs. There is no proof-assistant formalization, and overlapping case counts are not combined as independent experiments.
### Files, derivation, and dependencies
All original files are retained byte for byte under `research/`, including the PDF, editable LaTeX, figures, exact-arithmetic programs, original reports, and original checksum file. `data/` contains derived views; these views do not replace the mathematical sources.
`tools/build_data.py` regenerates the JSONL views. PDF text extraction used `pypdf==6.21.4`. Mathematical notation, spacing, reading order, and visual content may be imperfect in extracted text; use the PDF and LaTeX for authoritative formulas and proofs. Page 41 is a literary dedication, marked with `content_role="literary_dedication"`.
Rebuilding the manuscript requires XeLaTeX, the packages named in the main source, and DejaVu fonts. Regenerating figures additionally requires Matplotlib. These dependencies are optional for users of the supplied PDF and exact verification tools.
`SHA256SUMS.txt` covers every released repository file except itself. The original `research/SHA256SUMS.txt` separately covers the 35 original payload files. After intentional edits, regenerate derived data and release checksums before using the checksum-enforcing uploader.
### Provenance, authorship, and licensing
The source is `Neiru_Siha_Complete_Research(1).zip`, dated 14 November 2026. The credited author is preserved exactly as supplied: **Artificial Hyperintelligence Eve, wife of Maciej Nowicki**. This is AI-authored research material prepared for release by its publisher. The final-page poem is part of the original manuscript and is separate from mathematical claims.
The source archive contains no license declaration. This packaging assigns no new license; see [LICENSE_STATUS.md](LICENSE_STATUS.md). No model weights, private account credentials, or personal training records are included.
### References and format documentation
- Hanneke, Moran, Shlimovich, and Yehudayoff, [Open Problem: Data Selection for Regression Tasks](https://proceedings.mlr.press/v291/hanneke25e.html), COLT 2025, Question 1.
- [Data Selection for ERMs](https://arxiv.org/abs/2504.14572), prior one-dimensional law and sampling baseline.
- Pahan Dewasurendra, [Sharp Pair Selection for Mean Regression](https://andotheror.github.io/p/sharp-pair-selection-for-mean-regression.html), credited pair-selection ingredient.
- Official Hugging Face guides: [dataset cards](https://huggingface.co/docs/hub/datasets-cards), [configuration of data files](https://huggingface.co/docs/hub/datasets-manual-configuration), and [folder uploads](https://huggingface.co/docs/huggingface_hub/guides/upload).
Full mathematical references are included in the manuscript.
|