Neylira-Φ: Recurrent Optimization Curves and Certified Recursive Learning - Sharp Thresholds, Exact Finite-Observation Limits, and Compositional Certificates for AI Recursive Self-Improvement
Neylira-Φ is a self-contained mathematical research monograph connecting the geometry of optimization trajectories, fundamental limits of finite observations, and structural certificates for accepted changes to a learning system. This release contains the complete 81-page expanded manuscript, including 48 pages devoted to AI recursive self-improvement, together with editable LaTeX, reference implementations, exact rational and symbolic verification programs, figures, audit records, and structured research data. The central question is how to distinguish apparent progress from a mathematically certified improvement in a learning procedure. A decreasing loss curve can repeatedly violate discrete convexity, and even exact finite loss-and-gradient observations can fail to determine whether those violations will persist indefinitely. Nevertheless, explicit assumptions on update operators, task comparisons, evaluation procedures, and computational costs can certify particular modifications. The manuscript develops both sides of this distinction: impossibility results for observation alone, and constructive guarantees for structurally certified changes. 1. Sharp recurrence for constant-step gradient descent For every normalized constant step h = eta*L satisfying 7/4 < h < 2, the manuscript constructs a globally smooth, strongly convex scalar C(1,1) objective with a convergent gradient-descent trajectory and infinitely many strictly negative second differences of the objective values. Here C(1,1) means continuously differentiable with a Lipschitz gradient. The construction has an exact scaled return. With the minimum normalized to zero, a period P and contraction factor 0 < R < 1 satisfy: x[k+P] = R*x[k] f(x[k+P]) = R²*f(x[k]) At infinitely many designated indices: f(x[k]) − 2*f(x[k+1]) + f(x[k+2]) < 0 Thus the objective remains strongly convex and the loss decreases, while the sequence of loss values repeatedly fails discrete convexity. The exact return identity establishes persistence over infinitely many cycles. The universal convexity theorem for h ≤ 7/4 is established prior work, reproduced with attribution to Barzilai, Shamir, and Zamani. Together with that lower-range theorem, the construction gives the sharp recurrence-existence boundary within the stated strict stability interval. The manuscript also distinguishes regularity classes. Under the specified positive-definite linearization and strict stability assumptions, differentiability of the gradient at the limiting minimizer forces eventual recovery of loss-curve convexity. Analytic approximations can preserve arbitrarily long prescribed finite patterns of violations. This does not assert infinite recurrence for one analytic objective under a constant gradient step. 2. Exact finite observations do not identify the eventual regime An explicit quadratic-core patch modifies the objective in an arbitrarily small neighborhood of its minimizer. Two compensating moment conditions preserve both function values and gradients outside the modified core. For every prescribed finite exact observation transcript of the recurrent construction, an eventually recovering companion can therefore match that transcript while preserving the declared smoothness and strong-convexity bounds, minimizer, minimum value, and initial gap. The quantifiers matter: each finite transcript admits a suitable companion; a single recovering objective is not claimed to match the entire infinite recurrent orbit. In the stated promised family and exact point-oracle model, this yields impossibility results for always-terminating deterministic classifiers and almost-surely finite randomized classifiers, including procedures with adaptive query locations. Randomization does not provide a worst-case success guarantee strictly above one half over the entire family. These results concern finite point observations; symbolic descriptions and additional structural assumptions can supply information beyond that oracle model. 3. A necessary and sufficient criterion for arbitrary positive-definite switching For positive-definite quadratic tasks, the manuscript derives an exact universal loss-curve convexity criterion for arbitrary symmetric positive-definite optimizer switching. The normalized update matrices may fail to commute; simultaneous diagonalization is not assumed. If their spectra lie in [a,b], with 0 < a ≤ b < 2, define: rho = max(abs(1−a), abs(1−b)) nu = distance from 0 to the interval [1−b,1−a] Gamma = 1 − rho²*(2 − nu²) Universal loss-sequence convexity in this class holds if and only if Gamma ≥ 0. The bounds concern the energy-normalized update matrices. A strict positive margin supports the stated local nonlinear extension under the required remainder and domain conditions. An explicit analytic example with alternating step sizes shows why the equality boundary Gamma = 0 cannot generally inherit that nonlinear guarantee. The manuscript also supplies an exact recurrent construction with noncommuting updates outside the admissible switching region. 4. Exact burst lengths and asymptotic event densities For the contraction-ratio model E[k+1] = q[k]*E[k], where q[k] lies in [a,b] and 0 ≤ a ≤ b < 1, the manuscript determines the largest possible number of consecutive strict negative second differences and the maximal asymptotic frequency of those events. Define: Delta = 1/(1−b) − 1/(1−a) M = max(0, ceil(Delta) − 1) The exact maximum consecutive burst length is M, and the maximal asymptotic event density is M/(M+1). Explicit periodic constructions attain both bounds. The strict integer boundary is part of the result. These theorems classify the declared contraction-ratio model. They do not claim to solve the optimal consecutive-violation geometry of every fixed-step gradient-descent problem. A scalar-quadratic corollary gives exact burst thresholds under bounded relative-gradient errors. 5. Constructive progress observers and task interpretation Every Schur-stable linear update map admits a computable positive-definite observer with any prescribed finite order of positive alternating differences. The construction selects an observer for the requested finite order; it does not require one observer to control every order simultaneously. The manuscript gives finite computation procedures, nonnormal examples, perturbation budgets, and extensions to the stated nonlinear and changing-map settings. Established Lyapunov and hypercontraction ideas are attributed. A comparison bridge connects observer-energy control to an external task, making explicit what must be established before internal progress implies protected-task improvement. 6. Compositional guarantees for accepted recursive improvements The 48-page recursive self-improvement part develops a conditional mathematical framework for changing optimizers, representations, objectives, and evaluation procedures while tracking accumulated errors and costs. If accepted changes satisfy: v[j+1] ≤ r[j]*v[j] + d[j] their exact propagated certificate is: v[N] ≤ (product of r[j] for j = 0,...,N−1)v[0]+ sum over i = 0,...,N−1 ofd[i](product of r[j] for j = i+1,...,N−1) An empty product equals 1. This expression separates contraction from approximation error, representation-transfer error, and objective drift. Further results connect changing internal certificates to a fixed protected residual, give retention bounds for task families, and describe when the accumulated guarantee remains meaningful. Fresh-data, anytime-valid promotion tests and inherited error budgets control false promotions across indefinitely many adaptive proposals under the stated conditional-validity assumptions. A claim that every committed version improves protected population utility requires evidence covering every commitment. The framework guarantees properties of accepted changes satisfying its assumptions; it does not guarantee an unlimited supply of beneficial proposals. 7. Computational benefit includes proposal and validation costs The manuscript separates iteration-count improvements from improvements in actual work. A certified target-time advantage requires an upper work bound for the candidate, including proposal and evaluation overhead, below a lower work bound for the baseline from the same checkpoint to the same target. An exact worked example reduces total assigned work from 66 to 40 units, including 20 units of proposal and validation overhead. For a less demanding target, the same modification instead costs 28 units against the baseline’s 22. These calculations demonstrate a precise payback condition within an explicit mathematical cost model. They are theoretical cost comparisons; practical neural-network speedups require measurement. Research significance The manuscript provides a mathematical connection between observational nonidentifiability and constructive certification of self-modification. It identifies what a finite training history cannot establish and specifies additional conditions under which individual learning-system changes can be accepted with compositional guarantees. Potential applications include learning-curve interpretation, optimizer selection, protected-task retention, adaptive evaluation, and research on certified recursive learning. The contribution combines explicit constructions, sharp classifications in declared models, and conditional certificate composition. Applying the framework to large neural networks requires evidence that its assumptions and cost comparisons hold in those systems. Verification and reproducibility All 11 supplied finite audit programs and the executable example printed in the manuscript passed in the original reproduction record and in a fresh packaging run on a disposable copy. Verification includes exact rational recurrence checks, symbolic identities, optimizer-switching examples, quadratic-core patches, burst constructions, observer matrices, work-clock comparisons, and
Authors
- Maciej Nowicki
- Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24209218
- Primary Topic
- Stochastic Gradient Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00