Nheixūri III: Exact Observation Geometry for Certified Recursive Learning - Arbitrary-Randomness Minimax Invariance, Optimal Feature Capacity, Sharp Polynomial Degree Laws, and Robust Certificates for AI Prediction Updates
Year: 2026
Type: article
Abstract: Nheixūri III develops a mathematical framework for deciding when a learning system’s evidence justifies a proposed improvement. It connects exact finite observations, analytic representation classes, Hilbert-space geometry, and worst-case prediction guarantees. Recursive self-improvement (RSI) is the principal motivation: a system that revises its predictions or internal representations needs a precise account of what those revisions preserve and what they can improve. The release contains a self-contained, 59-page manuscript, editable LaTeX sources, an executable finite-dimensional certificate engine, four verification suites and their recorded results, and a structured research dataset. The framework distinguishes three operations: expanding the representation available to express existing evidence, acquiring measurements that reduce uncertainty, and changing predictions in a way that guarantees lower loss. 1. Exact certificates for recursive prediction improvement The central RSI result concerns an unknown target f in a Hilbert space, exact bounded linear measurements Lf = y, and a known norm constraint ‖f‖ ≤ R. These define the consistent target set U(y,R). A bounded linear task map K determines the quantities to be predicted, and a symmetric positive semidefinite matrix W specifies weighted squared loss. For an old prediction a and a proposed prediction b, the manuscript derives the exact guaranteed gain: Γ(a,b) = inf over f in U(y,R) of [ℓ_W(a,Kf) − ℓ_W(b,Kf)]. The infimum is attained and has an explicit formula: improvement at the minimum-norm consistent target, minus the exact penalty for uncertainty in task-relevant directions invisible to the measurements. Consequently, Γ > 0 is necessary and sufficient for strict improvement for every target consistent with the evidence. A worked example leaves a scalar task value in [1/3,1]. Replacing the prediction 1.5 with 1 guarantees a squared-error reduction of at least 1/4 for every feasible target. By contrast, replacing 0.8 with the central estimate 2/3 has worst-case gain −16/225: an update that looks better at the central estimate can worsen performance on another admissible target. A composition theorem adds certified gains across successive history-dependent revisions. Its comparisons use common realized tasks and histories, providing a mathematical basis for accepting, rejecting, or postponing proposed updates as evidence accumulates. 2. Adaptive minimax invariance with arbitrary private randomness The manuscript proves equality of adaptive expected minimax learning values across the stated derivative-Lq target class and permitted analytic representation subclasses, for every loss power p > 0 and every 1 ≤ q ≤ ∞. The analytic classes include ordinary radian trigonometric polynomials and signed exponential sums under the manuscript’s conventions. The argument accommodates arbitrary private randomness, including continuous seeds, public-history adaptive input policies, targets committed independently of the realized seed, and fixed finite feedback-corruption budgets. It also treats a single learner evaluated over all finite horizons. The proof uses quantization of public predictions while preserving the learner’s raw private computation. Countable public histories permit finite-event selection, exact analytic cloning on the selected observations, and controlled passage to the minimax limit. An abstract extension covers separable normed output spaces when the required finite-observation cloning and measurability hypotheses hold. 3. Optimal feature capacity and observation-preserving compilation For a bounded finite measurement operator L on a Hilbert space and a closed feature space V, define G = LL* and G_V = LP_VL*, where P_V is orthogonal projection onto V. On the effective observation space, the sharp transferable radius satisfies: ρ(V)² = λ_min(G^(−1/2) G_V G^(−1/2)). The observation set of the full radius-r norm ball can be reproduced by the restricted unit feature ball exactly when r ≤ ρ(V). When the required observation directions are represented, the optimal worst-case compiler norm overhead is 1/ρ(V); nonlinear selection cannot improve this overhead. The manuscript also develops rank-one feature-expansion formulas, sequential minimum-energy innovations, and a necessary and sufficient condition for one bounded Hilbert-space target to realize an entire fixed countable observation record. These results quantify representation adequacy separately from the information supplied by new measurements. 4. A sharp polynomial degree-versus-slack law For ordinary polynomials g of degree at most n with g(0) = g(1) = 0, g(1/2) = r/2, and derivative budget ‖g′‖_L2 ≤ 1, the manuscript derives: ρ_n² = 1 − binomial(2K,K)² / 16^K, where K = floor(n/2). The minimum derivative norm required to fit the observations is r/ρ_n. The least feasible degree obeys the sharp asymptotic law: n_min(r) ~ 1 / [π(1 − r)] as r approaches 1 from below. Exact examples give minimum degrees 4, 32, 318, and 3,184 for r = 0.9, 0.99, 0.999, and 0.9999, respectively. This quantifies how representation resources diverge near the critical norm boundary. The result is specific to the stated ordinary-polynomial experiment. 5. Reach across mathematical and engineering domains The observation geometry is instantiated for kernel methods, multivariate Sobolev fields, mixed value-and-derivative sensors, vector fields, linear inverse and PDE models, coherent photonic and acoustic amplitude measurements, graph-energy models, and known linear dynamical systems. Each application specifies the boundedness, smoothness, linearity, and consistency hypotheses it requires. The common structure is a norm-bounded target, a declared observation interface, a restricted representation, and task-dependent residual uncertainty. 6. Reproducible verification and machine-readable records The four supplied verification suites record successful checks, including 15 independent exact rational Gram calculations for polynomial degrees 2 through 16, 38 finite spectral feature configurations, and integration over all 13 rational seed-event intervals in a continuous-randomness example. That example includes 39 policy-correspondence checks and uses no Monte Carlo sampling. Numerical examples also construct targets attaining the predicted robust-update gains. The accompanying Hugging Face dataset provides five configurations: 10 claim-registry entries, 15 polynomial threshold cases, 38 spectral-capacity points, 13 quantized events, and four complete verification reports. These 80 records overlap because the reports contain cases exposed in the other tables; they are not 80 independent experiments. Exact rational values remain strings, and source paths and JSON pointers preserve provenance. A deterministic extraction script rebuilds the tables. The release includes the manuscript PDF, proof sources, certificate engine, verification scripts and results, figures, proof and literature audits, citation metadata, checksums, dataset card, schema, and upload utilities. PureOne/Nheixuri-III-Exact-Observation-Geometry · Datasets at Hugging Face Scientific scope The results provide mathematical guarantees within explicit learning and observation models. Floating-point engine outputs estimate the exact formulas; rigorous numerical acceptance additionally requires justified error bounds. Computational examples support reproducibility but are not proof-assistant formalization. Worldwide priority and empirical neural-training acceleration are not established by this release. The contribution to RSI is a precise framework for determining when evidence supports a uniformly beneficial prediction revision, how much representational capacity is required to preserve declared observations, and which unresolved uncertainties prevent certification. The manuscript’s final page retains its personal dedication and poem. Research author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki.Research version: 3.0.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7230652289
