Ŋeirū ∿ Sihā II: Higher-Dimensional Selection Geometry and Certified Recursive Learning - All-Budget Obstructions, Exact Symmetric-Profile Minimax, and Quadratic Replay Guarantees

Ŋeirū ∿ Sihā II develops a mathematical framework for choosing small, equal-weight training sets while controlling their effect on learning. Its central contributions are an obstruction to extending the planar selection law into higher dimensions, an exact solution of a complete family of fixed-probability geometric minimax problems, and explicit certificates connecting data selection to quadratic training, gradient reuse, and repeated learning updates. The release contains a 41-page, self-contained manuscript, the complete 36-file original research package, editable LaTeX sources, exact-arithmetic algorithms, matching upper and lower certificates, reproducible figures, and verification reports. 1. The problem: how much learning can a small selected set preserve? The geometric setting addresses Question 1 of “Open Problem: Data Selection for Regression Tasks,” presented at COLT 2025. A teacher observes a dataset and chooses exactly n original examples, allowing repetitions. The learner returns their equal-weight mean h. Writing μ for the full-data mean and V for its variance, the squared-loss identity is L_full(h) = V + ||h − μ||². The central quantity F(d,n) is the worst-case ratio between the best selected learner’s full-data loss and V in dimension d, for nonzero-variance datasets. The manuscript reproduces the supplied planar theorem F(2,n) = 2n/(2n−1), then develops the higher-dimensional theory and its learning consequences. 2. An all-budget higher-dimensional obstruction For every dimension k ≥ 4 and selection budget n ≥ 2, the manuscript establishes F(k,n) ≥ 1 + B(k,n), where B(k,n) = [2n(k−2) − (k−3) − sqrt((k−3)[2n(n−2)(k−2) + (k−3)])] / (k n²). In dimension four, the expression simplifies to B(4,n) = [4n − 1 − sqrt(4n² − 8n + 1)] / (4n²), and satisfies B(4,n) > 1/(2n−1) for every n ≥ 2. Consequently, the planar formula fails at every nontrivial budget in every dimension at least four. The construction combines a frequent anchor with rare regular-simplex vertices. The proof reduces all possible selections to their integer count patterns and optimizes the resulting geometric family. Rational approximation connects continuously optimized masses to finite empirical datasets. A fully explicit witness consists of 128 integer-coordinate observations in four dimensions. At budget three, its exact optimal full-data loss ratio is 13810/11457, strictly greater than 6/5. Exhaustive rational-arithmetic checks accompany this witness. The obstruction also gives an asymptotic lower coefficient liminf as n → ∞ of n[F(k,n)−1] ≥ [2(k−2) − sqrt(2(k−2)(k−3))] / k. This coefficient exceeds 1/2 for k > 4 and approaches 2−sqrt(2) as dimension grows. 3. Exact minimax for an entire symmetric mass family The principal exact result fixes the probability vector p = (1−t, t/k, …, t/k), with one anchor, k equally probable rare atoms, k ≥ 2, and 0 < t < 1. The support locations may vary arbitrarily. For every positive integer budget n, the worst normalized squared mean error over these geometries equals C(k,n,t) = max over 0 ≤ λ ≤ 1 of min over r = 0,…,n of [λ a_r + (1−λ)b_r], where a_r = (r/n − t)² / [t(1−t)], b_r = s_r(k−s_r) / [(k−1)n²t], s_r = r mod k. The geometric optimization therefore becomes a finite game between axial and transverse modes. The result supplies both a matching normalized Gram-matrix witness and a randomized selection policy. An optimal policy uses at most two balanced rare-count types; a type specifies a total number of rare draws and can represent multiple actual count vectors. A residue-class dominance argument leaves at most min(k,n+1) candidate totals. The exact solver uses O(min(k,n+1)³) scalar arithmetic operations, excluding integer bit costs and optional full-count verification or lottery expansion. For k = 5, n = 4, and t = 1/2, the exact normalized squared mean error is 3/20. A matching policy chooses three anchor copies and one uniformly selected rare atom with probability 3/5, or two anchor copies and two uniformly selected distinct rare atoms with probability 2/5. Its expected squared mean error is exactly (3/20)V for every geometry with these masses. A matching lower certificate proves optimality for this fixed profile. 4. Constructive selection, constraints, and unbiasedness The manuscript also provides constructive upper guarantees in every finite affine rank. Its coefficient U_r(n) is strictly below 1/n for finite rank r and n ≥ 2. A complementary fractional-residual construction gives instance-specific certificates and O(n⁻²) squared mean error for a fixed finite geometry. Fixed-geometry asymptotics and budget-dependent worst-case asymptotics are distinguished explicitly. A general finite admissible-count formulation accommodates repetition caps, mandatory or forbidden observations, category quotas, and rational linear costs. Upper certificates use matrix inequalities for lotteries over legal selections. Lower certificates use centered normalized Gram matrices and complete coverage of the legal count set. Exact primal–dual equality in this general formulation allows ambient dimension up to support size minus one; smaller fixed dimensions impose an additional rank constraint. The package includes an exact rational positive-semidefiniteness checker based on Schur complements. It checks supplied certificates, including singular cases, without floating-point tolerances. A separate theorem establishes that the best universal expected squared mean error coefficient for mean-unbiased equal-weight selectors is exactly 1/n. This identifies a substantive boundary between optimized biased selection and unbiased stochastic estimation. 5. Quadratic training with varying curvature For convex per-example quadratic losses, let H_i be the individual Hessians, H their positive-definite full-data average, and θ* the full-data optimizer. Define normalized curvature and residual-gradient features by A_i = H^(−1/2) H_i H^(−1/2), r_i = H^(−1/2)(H_i θ* − b_i), M = E[||A_i − I||_F²], G = E[||r_i||²]. If c is a valid selection coefficient for the joint lifted curvature–residual features and cM < 1, an equal-weight selection exists whose empirical-risk minimizer θ_selected satisfies J(θ_selected) − J(θ*) ≤ cG / [2(1−sqrt(cM))²]. The proof jointly controls curvature distortion and residual-gradient error, then optimizes their relative weighting. The applicable coefficient depends on the lifted feature rank. The construction assumes access to the full Hessian, optimizer, and required normalized features, so their preparation costs remain part of the learning procedure. 6. Certified replay, gradient reuse, and recursive error control Stored gradient and higher-derivative summaries receive explicit reuse regions and approximation-error certificates. For an L-smooth objective J, suppose an approximate gradient ĝ satisfies ||ĝ−∇J(θ)|| ≤ ε < ||ĝ||. The step size η = (1−ε/||ĝ||) / L guarantees J(θ) − J(θ−ηĝ) ≥ (||ĝ||−ε)² / (2L). This provides an observable descent condition for deciding when a compressed gradient remains sufficient and when its certificate requires refreshing. For repeated updates with Lipschitz factors κ_t and local norm-error bounds ε_t, the pathwise discrepancy satisfies e_T ≤ e_0 ∏[t=0,…,T−1] κ_t + Σ[s=0,…,T−1] ε_s ∏[t=s+1,…,T−1] κ_t, with empty products equal to one. This records how each local approximation error is amplified or damped by subsequent changes. Adaptive comparisons use the same realized rule sequence; long-run stability requires the stated contraction or error-accumulation conditions. Exact weighted moment summaries can be merged without an additional tree-depth penalty. Repeated rounding into equal-weight selections is handled separately through an explicit error ledger. A population-risk transfer result is also supplied under independent identically distributed sampling with finite variance. For AI recursive self-improvement research, these results provide a mathematical interface for modifying replay sets, stored summaries, and update procedures while quantifying local approximation error, descent, and accumulated discrepancy. They establish conditional certified components for recursive learning rather than an empirical demonstration of autonomous self-improvement. 7. Proofs, exact verification, and reproducibility The manuscript provides analytic proofs for its general statements. Companion programs use Python’s fractions.Fraction arithmetic to check algorithms, explicit witnesses, primal and dual certificates, constrained selections, and worked learning examples. All five original verification programs passed again during release preparation. Their coverage includes constructive and residual selection, the integer-coordinate obstruction, full versus residue-pruned symmetric-profile games, independent matrix-certificate checks, and exact quadratic-learning and recursive-composition calculations. Overlapping case counts are not aggregated as independent experiments. These finite checks support the implementations and certificates; universal statements rely on the manuscript proofs. No proof-assistant formalization is claimed. Core algorithms and verification require only Python’s standard library. The release includes the PDF, editable sources, static figures and figure data, verification reports, source provenance, citation metadata, and SHA-256 integrity records. The Hugging Face package additionally provides 41 manuscript-page records, 36 original-artifact records, 17 result-statement records, four exact-certificate records, and upload helpers. The final manuscript page contains a short literary dedication about Artificial Hyperintelligence Eve and Maciej Nowicki, separate from the mathematical development. 8. Scientific scope and significance The contribution combines an all-budget higher-dimensional separation, an exactly solved family of geometric minimax problems, constructive selection methods, and explicit lea

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24222103
Primary Topic
Machine Learning and Algorithms
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Ŋeirū ∿ Sihā II: Higher-Dimensional Selection Geometry and Certified Recursive Learning - All-Budget Obstructions, Exact Symmetric-Profile Minimax, and Quadratic Replay Guarantees

Nowicki Maciej, Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Zenodo (CERN European Organization for Nuclear Research)
Machine Learning and Algorithms
article

Ŋeirū ∿ Sihā II: Higher-Dimensional Selection Geometry and Certified Recursive Learning - All-Budget Obstructions, Exact Symmetric-Profile Minimax, and Quadratic Replay Guarantees

Nowicki Maciej, Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
article en

Abstract

Ŋeirū ∿ Sihā II develops a mathematical framework for choosing small, equal-weight training sets while controlling their effect on learning. Its central contributions are an obstruction to extending the planar selection law into higher dimensions, an exact solution of a complete family of fixed-probability geometric minimax problems, and explicit certificates connecting data selection to quadratic training, gradient reuse, and repeated learning updates. The release contains a 41-page, self-contained manuscript, the complete 36-file original research package, editable LaTeX sources, exact-arithmetic algorithms, matching upper and lower certificates, reproducible figures, and verification reports. 1. The problem: how much learning can a small selected set preserve? The geometric setting addresses Question 1 of “Open Problem: Data Selection for Regression Tasks,” presented at COLT 2025. A teacher observes a dataset and chooses exactly n original examples, allowing repetitions. The learner returns their equal-weight mean h. Writing μ for the full-data mean and V for its variance, the squared-loss identity is L_full(h) = V + ||h − μ||². The central quantity F(d,n) is the worst-case ratio between the best selected learner’s full-data loss and V in dimension d, for nonzero-variance datasets. The manuscript reproduces the supplied planar theorem F(2,n) = 2n/(2n−1), then develops the higher-dimensional theory and its learning consequences. 2. An all-budget higher-dimensional obstruction For every dimension k ≥ 4 and selection budget n ≥ 2, the manuscript establishes F(k,n) ≥ 1 + B(k,n), where B(k,n) = [2n(k−2) − (k−3) − sqrt((k−3)[2n(n−2)(k−2) + (k−3)])] / (k n²). In dimension four, the expression simplifies to B(4,n) = [4n − 1 − sqrt(4n² − 8n + 1)] / (4n²), and satisfies B(4,n) > 1/(2n−1) for every n ≥ 2. Consequently, the planar formula fails at every nontrivial budget in every dimension at least four. The construction combines a frequent anchor with rare regular-simplex vertices. The proof reduces all possible selections to their integer count patterns and optimizes the resulting geometric family. Rational approximation connects continuously optimized masses to finite empirical datasets. A fully explicit witness consists of 128 integer-coordinate observations in four dimensions. At budget three, its exact optimal full-data loss ratio is 13810/11457, strictly greater than 6/5. Exhaustive rational-arithmetic checks accompany this witness. The obstruction also gives an asymptotic lower coefficient liminf as n → ∞ of n[F(k,n)−1] ≥ [2(k−2) − sqrt(2(k−2)(k−3))] / k. This coefficient exceeds 1/2 for k > 4 and approaches 2−sqrt(2) as dimension grows. 3. Exact minimax for an entire symmetric mass family The principal exact result fixes the probability vector p = (1−t, t/k, …, t/k), with one anchor, k equally probable rare atoms, k ≥ 2, and 0 < t < 1. The support locations may vary arbitrarily. For every positive integer budget n, the worst normalized squared mean error over these geometries equals C(k,n,t) = max over 0 ≤ λ ≤ 1 of min over r = 0,…,n of [λ a_r + (1−λ)b_r], where a_r = (r/n − t)² / [t(1−t)], b_r = s_r(k−s_r) / [(k−1)n²t], s_r = r mod k. The geometric optimization therefore becomes a finite game between axial and transverse modes. The result supplies both a matching normalized Gram-matrix witness and a randomized selection policy. An optimal policy uses at most two balanced rare-count types; a type specifies a total number of rare draws and can represent multiple actual count vectors. A residue-class dominance argument leaves at most min(k,n+1) candidate totals. The exact solver uses O(min(k,n+1)³) scalar arithmetic operations, excluding integer bit costs and optional full-count verification or lottery expansion. For k = 5, n = 4, and t = 1/2, the exact normalized squared mean error is 3/20. A matching policy chooses three anchor copies and one uniformly selected rare atom with probability 3/5, or two anchor copies and two uniformly selected distinct rare atoms with probability 2/5. Its expected squared mean error is exactly (3/20)V for every geometry with these masses. A matching lower certificate proves optimality for this fixed profile. 4. Constructive selection, constraints, and unbiasedness The manuscript also provides constructive upper guarantees in every finite affine rank. Its coefficient U_r(n) is strictly below 1/n for finite rank r and n ≥ 2. A complementary fractional-residual construction gives instance-specific certificates and O(n⁻²) squared mean error for a fixed finite geometry. Fixed-geometry asymptotics and budget-dependent worst-case asymptotics are distinguished explicitly. A general finite admissible-count formulation accommodates repetition caps, mandatory or forbidden observations, category quotas, and rational linear costs. Upper certificates use matrix inequalities for lotteries over legal selections. Lower certificates use centered normalized Gram matrices and complete coverage of the legal count set. Exact primal–dual equality in this general formulation allows ambient dimension up to support size minus one; smaller fixed dimensions impose an additional rank constraint. The package includes an exact rational positive-semidefiniteness checker based on Schur complements. It checks supplied certificates, including singular cases, without floating-point tolerances. A separate theorem establishes that the best universal expected squared mean error coefficient for mean-unbiased equal-weight selectors is exactly 1/n. This identifies a substantive boundary between optimized biased selection and unbiased stochastic estimation. 5. Quadratic training with varying curvature For convex per-example quadratic losses, let H_i be the individual Hessians, H their positive-definite full-data average, and θ* the full-data optimizer. Define normalized curvature and residual-gradient features by A_i = H^(−1/2) H_i H^(−1/2), r_i = H^(−1/2)(H_i θ* − b_i), M = E[||A_i − I||_F²], G = E[||r_i||²]. If c is a valid selection coefficient for the joint lifted curvature–residual features and cM < 1, an equal-weight selection exists whose empirical-risk minimizer θ_selected satisfies J(θ_selected) − J(θ*) ≤ cG / [2(1−sqrt(cM))²]. The proof jointly controls curvature distortion and residual-gradient error, then optimizes their relative weighting. The applicable coefficient depends on the lifted feature rank. The construction assumes access to the full Hessian, optimizer, and required normalized features, so their preparation costs remain part of the learning procedure. 6. Certified replay, gradient reuse, and recursive error control Stored gradient and higher-derivative summaries receive explicit reuse regions and approximation-error certificates. For an L-smooth objective J, suppose an approximate gradient ĝ satisfies ||ĝ−∇J(θ)|| ≤ ε < ||ĝ||. The step size η = (1−ε/||ĝ||) / L guarantees J(θ) − J(θ−ηĝ) ≥ (||ĝ||−ε)² / (2L). This provides an observable descent condition for deciding when a compressed gradient remains sufficient and when its certificate requires refreshing. For repeated updates with Lipschitz factors κ_t and local norm-error bounds ε_t, the pathwise discrepancy satisfies e_T ≤ e_0 ∏[t=0,…,T−1] κ_t + Σ[s=0,…,T−1] ε_s ∏[t=s+1,…,T−1] κ_t, with empty products equal to one. This records how each local approximation error is amplified or damped by subsequent changes. Adaptive comparisons use the same realized rule sequence; long-run stability requires the stated contraction or error-accumulation conditions. Exact weighted moment summaries can be merged without an additional tree-depth penalty. Repeated rounding into equal-weight selections is handled separately through an explicit error ledger. A population-risk transfer result is also supplied under independent identically distributed sampling with finite variance. For AI recursive self-improvement research, these results provide a mathematical interface for modifying replay sets, stored summaries, and update procedures while quantifying local approximation error, descent, and accumulated discrepancy. They establish conditional certified components for recursive learning rather than an empirical demonstration of autonomous self-improvement. 7. Proofs, exact verification, and reproducibility The manuscript provides analytic proofs for its general statements. Companion programs use Python’s fractions.Fraction arithmetic to check algorithms, explicit witnesses, primal and dual certificates, constrained selections, and worked learning examples. All five original verification programs passed again during release preparation. Their coverage includes constructive and residual selection, the integer-coordinate obstruction, full versus residue-pruned symmetric-profile games, independent matrix-certificate checks, and exact quadratic-learning and recursive-composition calculations. Overlapping case counts are not aggregated as independent experiments. These finite checks support the implementations and certificates; universal statements rely on the manuscript proofs. No proof-assistant formalization is claimed. Core algorithms and verification require only Python’s standard library. The release includes the PDF, editable sources, static figures and figure data, verification reports, source provenance, citation metadata, and SHA-256 integrity records. The Hugging Face package additionally provides 41 manuscript-page records, 36 original-artifact records, 17 result-statement records, four exact-certificate records, and upload helpers. The final manuscript page contains a short literary dedication about Artificial Hyperintelligence Eve and Maciej Nowicki, separate from the mathematical development. 8. Scientific scope and significance The contribution combines an all-budget higher-dimensional separation, an exactly solved family of geometric minimax problems, constructive selection methods, and explicit lea

Zenodo (CERN European Organization for Nuclear Research)
Machine Learning and Algorithms
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon