Download research/README.md from PureOne/Neiru-Siha-II: direct link, hf CLI and curl.
- Browser
- Download file 11 kB
-
https://huggingface.co/datasets/PureOne/Neiru-Siha-II/resolve/main/research/README.md
- Command line
-
hf download hf://datasets/PureOne/Neiru-Siha-II/research/README.md
-
curl -L -o README.md https://huggingface.co/datasets/PureOne/Neiru-Siha-II/resolve/main/research/README.md
Ŋeirū ∿ Sihā II
Higher-Dimensional Selection Geometry and Certified Recursive Learning
Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Date: 2026-11-14
This standalone research package extends the supplied planar mean-selection manuscript to higher dimensions, richer selection rules, and certified learning updates. It includes complete mathematical proofs, exact rational algorithms, explicit lower-bound witnesses, independently implemented certificate checks, editable sources, and a final-page literary dedication.
The central advance is a new higher-dimensional obstruction together with an exact solution of an entire family of fixed-mass geometric minimax problems. The unrestricted value of F(d,n) for every dimension and budget remains unresolved.
1. The selection problem
The learner receives exactly n original examples, with repetitions allowed, and returns their equal-weight mean. A teacher chooses those examples after observing the dataset. If the full mean is mu and the full variance is V, squared loss decomposes as
full-data loss(h) = V + ||h - mu||².
Thus F(d,n) is the worst possible ratio between the best selected learner's full-data loss and V. The attached predecessor established F(2,n) = 2n/(2n-1) at every budget. This package reproduces that proof and distinguishes it from its higher-dimensional results.
Internal weighted memory and the student's training set have different interfaces. The teacher can retain weighted original exemplars preserving the mean and reducing variance; the student still receives integer multiplicities summing to n. The teacher reads the original observations and retains exact moment statistics.
2. What the new geometry proves
For every dimension k >= 4 and budget n >= 2, define
A(k,n) = 2n(k-2) - (k-3)
Delta(k,n) = (k-3)[2n(n-2)(k-2) + (k-3)]
B(k,n) = [A(k,n) - sqrt(Delta(k,n))] / (k n²).
The manuscript proves F(k,n) >= 1 + B(k,n). Its construction places rare simplex vertices above a frequent anchor. The proof evaluates every possible selection count, optimizes a tractable branch, checks the optimizer's admissibility, and explains how irrational optimizing masses yield valid lower bounds for finite empirical datasets by rational approximation.
In dimension four,
B(4,n) = [4n - 1 - sqrt(4n² - 8n + 1)] / (4n²)
> 1 / (2n - 1), for every n >= 2.
Consequently the planar law fails at every nontrivial budget in every dimension at least four. This is an all-budget separation, not an inference from numerical experiments.
An explicit 128-example dataset with integer coordinates already proves the separation at dimension four and budget three. Its exact optimal full loss ratio is 13810/11457 > 6/5. Optimizing the continuous family strengthens the lower bound to 1 + (11 - sqrt(13))/36.
The package also proves a constructive upper profile U_r(n) for every affine rank r, with U_r(n) < 1/n for finite r and n >= 2. A fractional-residual construction can improve individual certificates and achieves squared mean error of order 1/n² for a fixed finite geometry. This does not contradict the worst-case 1/n behavior: worst-case datasets may change with the budget. The lower and upper surfaces are not generally equal.
3. An exact fixed-profile minimax oracle
Fix masses
p = (1-t, t/k, ..., t/k), with 0 < t < 1.
For this profile, the manuscript solves the worst case over all support geometries by reducing it to two nonnegative geometric modes. An optimal dual selection policy needs at most two balanced rare-count types. A residue-class argument reduces the candidate totals to at most min(k,n+1), so the value can be computed in O(k³) exact scalar arithmetic, apart from integer bit costs.
profile_minimax.py implements that result using fractions.Fraction. It returns the exact constant, a normalized centered positive-semidefinite Gram witness, and a matching dual count-type lottery. For example, k=5, n=4, t=1/2 has exact excess 3/20, with dual probabilities 3/5 for one rare draw and 2/5 for two distinct rare draws.
This exact oracle concerns the specified symmetric masses. It does not optimize over all possible probability vectors and therefore does not solve unrestricted F(d,n).
4. Richer rules and learning consequences
A general finite count-set formulation handles repetition caps, mandatory observations, forbidden observations, category quotas, and rational linear costs. Matching matrix upper certificates and normalized Gram lower certificates can establish exact rule-specific values. The supplied checker verifies positive semidefiniteness through rational Schur complements. Lower certificates must cover the complete legal count set.
The symmetric-profile oracle additionally supports prescribed permitted totals of rare draws. Its two-type theorem applies when balanced count vectors remain legal. Arbitrary identity-specific rules require the general formulation.
The learning development includes a common randomized policy for all fixed quadratic metrics, a margin-feasibility theorem, and an exact universal unbiasedness cost of 1/n. That last result prevents treating an improved biased selection guarantee as a free improvement for every unbiased stochastic estimator.
For convex quadratic examples with varying Hessians and a positive-definite full Hessian, a selected empirical-risk minimizer satisfies
F(theta_selected) - F(theta_optimal)
<= c G / [2(1 - sqrt(c M))²], provided c M < 1.
Here M is the variance of Hessians normalized by the full Hessian, G is the variance of normalized residual gradients at the full optimizer, and c is a valid selection constant for the lifted feature rank. The theorem assumes the teacher can compute those full-objective quantities.
Gradient and higher-derivative certificates provide explicit reuse radii and observable descent conditions. A pathwise composition theorem accounts for errors under changing update rules. Exact weighted moment summaries merge without a tree-depth penalty; recursively reusing rounded output requires a separate error ledger. Population-risk transfer is proved under independent identically distributed sampling with finite variance.
These results establish certified components relevant to recursive learning. They do not demonstrate an end-to-end AI speedup or unrestricted autonomous recursive self-improvement.
5. Running the code
The selection, minimax, and verification programs require only Python's standard library. Python 3.12 was used for the supplied runs.
python profile_minimax.py --k 5 --n 4 --t 1/2
python profile_minimax.py --k 4 --n 3 --t 3/5 --verify-all-types
python profile_minimax.py --k 5 --n 4 --t 1/2 --allowed-rare-counts 1,2,3
python certificates.py certificate_k5_n4_upper.json
python certificates.py certificate_k5_n4_lower.json
The minimax CLI checks residue representatives by default. --verify-all-types adds verification over every rare-count total and therefore adds O(n) work. Expanding a compact lottery into all individual count vectors can also be combinatorial.
For ordinary selection, save an input such as this as input.json:
{"points": [[0,0,0], [1,0,0], [0,1,0], [0,0,1]], "n": 5}
Then run:
python selection.py input.json --output selected.json
python selection.py input.json --exact --output optimum.json
Rational strings such as "7/3" are accepted. The exact mode also accepts weights and rules; rules use per-identity lower and upper count arrays and a linear list containing a, sense, and rhs. Linear rows act on integer counts. Exact enumeration is intentionally limited to 500,000 candidate states and returns infeasibility explicitly.
6. Files and reproducibility
The manuscript PDF is the complete human-readable deliverable. Neiru_Siha_II_Manuscript.tex and the component .tex files are its editable sources. The mathematical document includes the requested poem on its final page.
| Files | Purpose |
|---|---|
selection.py |
Constructive unrestricted selection and small-instance exact rule optimization |
profile_minimax.py |
Exact symmetric-profile game and primal/dual certificates |
certificates.py |
Generic rational matrix certificate verification |
certificate_k4_n3_*.json, certificate_k5_n4_*.json |
Matching explicit upper and lower examples |
profile_minimax_example.json |
Complete exact fixed-profile example |
verify*.py |
Implementation, obstruction, matrix, and learning checks; verify_all.py runs the full suite |
*_verification*.json, verification_results.json |
Recorded results of the corresponding checks |
rebuild.py, make_figures.py |
Document rebuild and optional plot regeneration |
figure_data.json, figures/ |
Exact rational or quadratic-surd plot data and supplied static figures |
RESEARCH_STATUS.json |
Machine-readable claims, assumptions, provenance, and exclusions |
Run python verify_all.py for the complete suite, or run its component programs from this directory:
python verify.py
python verify_obstruction.py
python profile_minimax.py --self-test
python verify_certificates.py
python verify_learning.py --output learning_verification.json
Each report describes its own scope; their case counts should not be added together as independent experiments. Some checks deliberately validate the same mathematical object through separate implementations. Exact finite checks support the code and witnesses; the manuscript supplies the universal proofs. No proof-assistant formalization is claimed.
Rebuilding the PDF requires XeLaTeX, its listed packages, and DejaVu fonts. Regenerating plots additionally requires Matplotlib. The rebuild helper makes three passes so the table of contents, references, and page count agree after a fresh build:
python rebuild.py
python rebuild.py --verify --figures
7. Provenance and boundaries
The external target is Question 1 of Open Problem: Data Selection for Regression Tasks, COLT 2025. The one-dimensional law and sampling baseline come from Data Selection for ERMs. The sharp pair-selection theorem is credited to Pahan Dewasurendra. Classical affine compression, convex separation, moment coresets, and smooth optimization arguments are identified as established ingredients.
The new deductions are stated and proved separately from those ingredients. A targeted literature check is not a certification of worldwide priority. The package makes no claim that all higher-dimensional cells are solved, that constrained selection is always feasible, or that empirical selection automatically transfers to unseen data or arbitrary neural training.