Ŋhoryūn-Ω III: Sharp Gradient-Descent Curvature Thresholds, Arbitrarily Long Finite Concave Runs on Analytic Strongly Convex Objectives, Complete Finite-History Prediction, and Verifiable Recursive Learning

Ŋhoryūn-Ω III: Certificates, Curvature, and Recursive Learning is a self-contained, 103-page mathematical monograph and reproducibility dataset connecting the geometry of optimization trajectories with finite-history inference, information-theoretic limits, and conditional guarantees for recursive learning. The release contains complete proofs, editable LaTeX sources, exact-arithmetic verification programs, a recursive comparison controller, figures, structured mathematical records, and a claim ledger linking the principal statements to their assumptions, proof locations, and computational evidence. The central question is what an observed learning trajectory can establish about future improvement. A decreasing loss need not improve at a decreasing rate, and an apparently weak initial improvement need not certify that little useful progress remains. Conversely, a finite record can impose sharp restrictions on future behavior without uniquely identifying the underlying objective. The manuscript develops these possibilities, guarantees, and limitations within one framework, then specifies how validated improvements can compose when a system changes its own learning procedure. Research version: 3.0.0.Release date: 13 November 2026. 1. TWO SHARP CURVATURE BOUNDARIES For exact constant-step gradient descent, define: xₖ₊₁ = xₖ − η∇f(xₖ) Dₖ = f(xₖ) − f(xₖ₊₁) κₖ = Dₖ − Dₖ₊₁ h = ηL Here, L is a global gradient-Lipschitz bound. Negative κₖ means that the next loss decrease is larger than the current one. The manuscript separates three distinct transitions. The established universal convexity boundary for the raw loss sequence is h = 7/4, attributed to Barzilai, Shamir, and Zamani. Beyond that boundary, an isolated negative second difference can occur, but stronger restrictions on successive improvements can remain valid. The exact compensation endpoint is: hᵢₛₒ = (19 − √17)/8 = 1.859611796797792… For 0 < h ≤ hᵢₛₒ, every trajectory in the stated smooth convex class satisfies: Dₖ₊₁ + Dₖ₊₂ ≤ 2Dₖ The manuscript proves this universal inequality and constructs counterexamples for every h > hᵢₛₒ. Its sharpness constructions admit real-analytic, strongly convex realizations. The resulting weighted curvature law also yields convex filtered losses and finite-horizon bounds on cumulative improvement. A second, strictly larger endpoint governs whether two adjacent second differences can both be negative. Denote this endpoint by h★: h★ = 1.863425364363717532… This constant is defined exactly as the unique root in the interval: 1.863425364 < h★ < 1.863425365 of the polynomial equation: 5000h⁷ − 33602h⁶ + 97161h⁵ − 175306h⁴ + 235672h³ − 224304h² + 126080h − 33536 = 0 A positive algebraic weight, r★ = 1.879888219435024417…, gives the universal certificate: r★κₖ + κₖ₊₁ ≥ 0 through h★. Above h★, the manuscript constructs realizable trajectories with both second differences strictly negative and transfers them to analytic strongly convex objectives. The exact root and weight are specified algebraically; their decimal values are explanatory approximations. These results distinguish failure of the compensation inequality from the onset of adjacent concavity. In particular, the interval between hᵢₛₒ and h★ permits compensation failure while still excluding adjacent negative curvatures. Both phenomena occur below the stable-descent endpoint h = 2. The positive proofs use exact interpolation identities and positive semidefinite certificates. The negative proofs establish full interpolation feasibility for the counterexamples. 2. ARBITRARILY LONG FINITE CONCAVE RUNS WITH EXPLICIT CONDITIONING For every integer m ≥ 1, the manuscript constructs a two-dimensional real-analytic, strongly convex objective with m consecutive strictly negative second differences. Writing M = m + 1, the optimal global constants are: L = 1 μ = 1/(128M) L/μ = 128M At the explicit normalized step: h = 2 − 1/(128M), the trajectory satisfies: κₖ ≤ −3/(128M²), for 0 ≤ k < m. The full orbit nevertheless converges geometrically to a unique minimizer. The construction therefore combines a globally convergent optimization process with an arbitrarily long prescribed finite interval of increasing improvements. The first decrement obeys: D₀ ≤ 3/(64M), whereas the cumulative improvement over the certified continuation interval is at least: 3/512. Across this family, the first observed improvement can tend to zero while a fixed positive amount of improvement remains available over a longer finite horizon. This provides a concrete obstruction to stopping rules that treat a small initial decrement alone as universal evidence of negligible remaining progress. The analytic regularization is quantitative. One explicit Gaussian smoothing width is: σ = 1 / [16M√(2 log(4096M²))] The proof controls smoothing errors along the perturbed orbit and preserves a strict negative-curvature margin. The objective and step may depend on m. The theorem does not assert an infinite increasing tail of positive decrements on a lower-bounded objective, nor does it claim that the stated conditioning tradeoff is optimal. 3. A GENERAL ANALYTIC REALIZATION PRINCIPLE Beyond the explicit family, the manuscript proves that every finite collection of strict continuous conditions on a gradient-descent transcript of a globally smooth convex objective can be reproduced arbitrarily closely by an analytic strongly convex objective, at the same dimension, initial point, and step, when the positive strong-convexity parameter is allowed to become sufficiently small. Consequently, the feasible finite strict sign patterns of decrements and second differences coincide between the smooth convex class and its analytic strongly convex subclass when no fixed positive conditioning requirement is imposed. Analyticity and unspecified strong convexity alone do not eliminate finite strict trajectory phenomena. The statement does not preserve exact interpolation equalities, exact derivative jets, a prescribed positive strong-convexity lower bound, or an infinite family of conditions simultaneously. The explicit long-run construction supplies quantitative margins and conditioning bounds beyond this general approximation principle. 4. COMPLETE FINITE-HISTORY PREDICTION AND CHECKABLE DECISIONS For prescribed finite first-order coefficients, the manuscript develops an exact finite Gram-matrix and interpolation model of compatible observations and future trajectories. Under the stated bounded normalization, compatible transcripts form a compact feasible set. Every affine future query has attained lower and upper endpoints, and every intermediate value is realizable by an objective in the declared class. This turns prediction into a precisely stated compatibility problem. A threshold decision is justified when the entire compatible interval lies on the required side of the threshold. For rational bounded-model data, every strictly valid decision margin admits a finite rational certificate with explicit allowances for coefficient rounding and semidefinite residuals. The framework also distinguishes observation errors from errors in the parameter updates. It derives explicit defect terms and robustness budgets for inexact or variable-step behavior. A successful finite consistency check does not establish global convexity, and arbitrary perturbations cannot preserve exact curvature signs without sufficient margins. The smooth interpolation and performance-estimation foundations are established prior work and are attributed accordingly. The manuscript develops their finite-decision, realization, robustness, and certificate consequences. It does not assert a universal certificate for every exact boundary case or a uniform practical complexity bound for certificate search. 5. SHARP INFORMATION LIMITS The manuscript constructs objectives sharing an arbitrarily long common value-and-gradient history while leaving the entire remaining-gap interval unresolved: [v²/(2L), v²/(2μ)] Here, v is the terminal gradient magnitude. The common-history construction assumes: 0 < ημ < 1 ηL < 2 and does not impose an additional fixed initial-distance bound. Its unresolved interval gives the sharp absolute-error obstruction: (v²/4)(1/μ − 1/L) and corresponding continuation and stopping-regret consequences. A separate analytic construction establishes that exact observations of derivatives through any prescribed finite order at finitely many trajectory points still need not identify the next loss. Analytic strongly convex objectives can agree on every requested finite jet at those points and produce different subsequent values. The separation is strictly positive; no uniform order-one separation is claimed. These results identify information that additional confidence in extrapolation cannot supply. Depending on the compatible objectives, a justified continuation or stopping decision may require new observations, stronger assumptions, or an explicit uncertainty allowance. 6. CONDITIONAL GUARANTEES FOR RECURSIVE LEARNING The recursive-learning component compares candidate systems through a declared common risk: R(s) ∈ [0,1] A candidate may change a model, architecture, optimizer, training procedure, or the mechanism that proposes subsequent modifications. Each incumbent and candidate remain fixed within their comparison. Observations must be fresh or conditionally mean-valid under the declared sampling model. For bounded paired observations, one explicit simultaneous confidence radius is: r(t,n; δ) = √[(2/n) log(2t(t+1)n(n+1)/δ)] Here, t indexes proposal rounds, n indexes paired observations within a comparison, and δ is the whole-run error budget. This allocation controls adaptive proposal rounds and sample stopping times under a single whole-run error budget. The controller adds declared numerical or measurement allowances and accepts a candidate only when its lower confiden

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24208588
Primary Topic
Stochastic Gradient Optimization Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Ŋhoryūn-Ω III: Sharp Gradient-Descent Curvature Thresholds, Arbitrarily Long Finite Concave Runs on Analytic Strongly Convex Objectives, Complete Finite-History Prediction, and Verifiable Recursive Learning

Maciej Nowicki, Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
4 links
Zenodo (CERN European Organization for Nuclear Research)
Stochastic Gradient Optimization Techniques
article

Ŋhoryūn-Ω III: Sharp Gradient-Descent Curvature Thresholds, Arbitrarily Long Finite Concave Runs on Analytic Strongly Convex Objectives, Complete Finite-History Prediction, and Verifiable Recursive Learning

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

Abstract

Ŋhoryūn-Ω III: Certificates, Curvature, and Recursive Learning is a self-contained, 103-page mathematical monograph and reproducibility dataset connecting the geometry of optimization trajectories with finite-history inference, information-theoretic limits, and conditional guarantees for recursive learning. The release contains complete proofs, editable LaTeX sources, exact-arithmetic verification programs, a recursive comparison controller, figures, structured mathematical records, and a claim ledger linking the principal statements to their assumptions, proof locations, and computational evidence. The central question is what an observed learning trajectory can establish about future improvement. A decreasing loss need not improve at a decreasing rate, and an apparently weak initial improvement need not certify that little useful progress remains. Conversely, a finite record can impose sharp restrictions on future behavior without uniquely identifying the underlying objective. The manuscript develops these possibilities, guarantees, and limitations within one framework, then specifies how validated improvements can compose when a system changes its own learning procedure. Research version: 3.0.0.Release date: 13 November 2026. 1. TWO SHARP CURVATURE BOUNDARIES For exact constant-step gradient descent, define: xₖ₊₁ = xₖ − η∇f(xₖ) Dₖ = f(xₖ) − f(xₖ₊₁) κₖ = Dₖ − Dₖ₊₁ h = ηL Here, L is a global gradient-Lipschitz bound. Negative κₖ means that the next loss decrease is larger than the current one. The manuscript separates three distinct transitions. The established universal convexity boundary for the raw loss sequence is h = 7/4, attributed to Barzilai, Shamir, and Zamani. Beyond that boundary, an isolated negative second difference can occur, but stronger restrictions on successive improvements can remain valid. The exact compensation endpoint is: hᵢₛₒ = (19 − √17)/8 = 1.859611796797792… For 0 < h ≤ hᵢₛₒ, every trajectory in the stated smooth convex class satisfies: Dₖ₊₁ + Dₖ₊₂ ≤ 2Dₖ The manuscript proves this universal inequality and constructs counterexamples for every h > hᵢₛₒ. Its sharpness constructions admit real-analytic, strongly convex realizations. The resulting weighted curvature law also yields convex filtered losses and finite-horizon bounds on cumulative improvement. A second, strictly larger endpoint governs whether two adjacent second differences can both be negative. Denote this endpoint by h★: h★ = 1.863425364363717532… This constant is defined exactly as the unique root in the interval: 1.863425364 < h★ < 1.863425365 of the polynomial equation: 5000h⁷ − 33602h⁶ + 97161h⁵ − 175306h⁴ + 235672h³ − 224304h² + 126080h − 33536 = 0 A positive algebraic weight, r★ = 1.879888219435024417…, gives the universal certificate: r★κₖ + κₖ₊₁ ≥ 0 through h★. Above h★, the manuscript constructs realizable trajectories with both second differences strictly negative and transfers them to analytic strongly convex objectives. The exact root and weight are specified algebraically; their decimal values are explanatory approximations. These results distinguish failure of the compensation inequality from the onset of adjacent concavity. In particular, the interval between hᵢₛₒ and h★ permits compensation failure while still excluding adjacent negative curvatures. Both phenomena occur below the stable-descent endpoint h = 2. The positive proofs use exact interpolation identities and positive semidefinite certificates. The negative proofs establish full interpolation feasibility for the counterexamples. 2. ARBITRARILY LONG FINITE CONCAVE RUNS WITH EXPLICIT CONDITIONING For every integer m ≥ 1, the manuscript constructs a two-dimensional real-analytic, strongly convex objective with m consecutive strictly negative second differences. Writing M = m + 1, the optimal global constants are: L = 1 μ = 1/(128M) L/μ = 128M At the explicit normalized step: h = 2 − 1/(128M), the trajectory satisfies: κₖ ≤ −3/(128M²), for 0 ≤ k < m. The full orbit nevertheless converges geometrically to a unique minimizer. The construction therefore combines a globally convergent optimization process with an arbitrarily long prescribed finite interval of increasing improvements. The first decrement obeys: D₀ ≤ 3/(64M), whereas the cumulative improvement over the certified continuation interval is at least: 3/512. Across this family, the first observed improvement can tend to zero while a fixed positive amount of improvement remains available over a longer finite horizon. This provides a concrete obstruction to stopping rules that treat a small initial decrement alone as universal evidence of negligible remaining progress. The analytic regularization is quantitative. One explicit Gaussian smoothing width is: σ = 1 / [16M√(2 log(4096M²))] The proof controls smoothing errors along the perturbed orbit and preserves a strict negative-curvature margin. The objective and step may depend on m. The theorem does not assert an infinite increasing tail of positive decrements on a lower-bounded objective, nor does it claim that the stated conditioning tradeoff is optimal. 3. A GENERAL ANALYTIC REALIZATION PRINCIPLE Beyond the explicit family, the manuscript proves that every finite collection of strict continuous conditions on a gradient-descent transcript of a globally smooth convex objective can be reproduced arbitrarily closely by an analytic strongly convex objective, at the same dimension, initial point, and step, when the positive strong-convexity parameter is allowed to become sufficiently small. Consequently, the feasible finite strict sign patterns of decrements and second differences coincide between the smooth convex class and its analytic strongly convex subclass when no fixed positive conditioning requirement is imposed. Analyticity and unspecified strong convexity alone do not eliminate finite strict trajectory phenomena. The statement does not preserve exact interpolation equalities, exact derivative jets, a prescribed positive strong-convexity lower bound, or an infinite family of conditions simultaneously. The explicit long-run construction supplies quantitative margins and conditioning bounds beyond this general approximation principle. 4. COMPLETE FINITE-HISTORY PREDICTION AND CHECKABLE DECISIONS For prescribed finite first-order coefficients, the manuscript develops an exact finite Gram-matrix and interpolation model of compatible observations and future trajectories. Under the stated bounded normalization, compatible transcripts form a compact feasible set. Every affine future query has attained lower and upper endpoints, and every intermediate value is realizable by an objective in the declared class. This turns prediction into a precisely stated compatibility problem. A threshold decision is justified when the entire compatible interval lies on the required side of the threshold. For rational bounded-model data, every strictly valid decision margin admits a finite rational certificate with explicit allowances for coefficient rounding and semidefinite residuals. The framework also distinguishes observation errors from errors in the parameter updates. It derives explicit defect terms and robustness budgets for inexact or variable-step behavior. A successful finite consistency check does not establish global convexity, and arbitrary perturbations cannot preserve exact curvature signs without sufficient margins. The smooth interpolation and performance-estimation foundations are established prior work and are attributed accordingly. The manuscript develops their finite-decision, realization, robustness, and certificate consequences. It does not assert a universal certificate for every exact boundary case or a uniform practical complexity bound for certificate search. 5. SHARP INFORMATION LIMITS The manuscript constructs objectives sharing an arbitrarily long common value-and-gradient history while leaving the entire remaining-gap interval unresolved: [v²/(2L), v²/(2μ)] Here, v is the terminal gradient magnitude. The common-history construction assumes: 0 < ημ < 1 ηL < 2 and does not impose an additional fixed initial-distance bound. Its unresolved interval gives the sharp absolute-error obstruction: (v²/4)(1/μ − 1/L) and corresponding continuation and stopping-regret consequences. A separate analytic construction establishes that exact observations of derivatives through any prescribed finite order at finitely many trajectory points still need not identify the next loss. Analytic strongly convex objectives can agree on every requested finite jet at those points and produce different subsequent values. The separation is strictly positive; no uniform order-one separation is claimed. These results identify information that additional confidence in extrapolation cannot supply. Depending on the compatible objectives, a justified continuation or stopping decision may require new observations, stronger assumptions, or an explicit uncertainty allowance. 6. CONDITIONAL GUARANTEES FOR RECURSIVE LEARNING The recursive-learning component compares candidate systems through a declared common risk: R(s) ∈ [0,1] A candidate may change a model, architecture, optimizer, training procedure, or the mechanism that proposes subsequent modifications. Each incumbent and candidate remain fixed within their comparison. Observations must be fresh or conditionally mean-valid under the declared sampling model. For bounded paired observations, one explicit simultaneous confidence radius is: r(t,n; δ) = √[(2/n) log(2t(t+1)n(n+1)/δ)] Here, t indexes proposal rounds, n indexes paired observations within a comparison, and δ is the whole-run error budget. This allocation controls adaptive proposal rounds and sample stopping times under a single whole-run error budget. The controller adds declared numerical or measurement allowances and accepts a candidate only when its lower confiden

Zenodo (CERN European Organization for Nuclear Research)
Stochastic Gradient Optimization Techniques
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