Xévarin - Exact Measurement Complexity and Sequential Certification for Recursive Optimizer Self-Improvement

Xévarin asks how a learning system can make its next improvement cheaper. After each update, the method carries forward the repaired optimizer and the input and output subspaces that exposed the previous optimizer’s error. When those subspaces remain relevant, a successor can recover the next correction from fewer scalar measurements. The manuscript models a local task by an invertible operator A and its current updater by B. A probe reads one scalar, aᵀ(I − AB)b, from the updater’s one-step residual. For a rank-k residual in d dimensions with p correct output directions and q correct input directions already known, the generic exact recovery count is N = k² + (d − k)(2k − p − q) noiseless scalar measurements. The paper gives a reconstruction attaining this count and a matching lower bound for continuous local identification. Each valid inherited direction saves d − k measurements. After an initial full-rank change identifies both fixed k-dimensional subspaces, later repairs need only k² calibration readings, including when a subsequent change has lower rank. The sequential audit freezes the candidate and task operator before drawing fresh Gaussian bilinear probes. For the audit’s expected one-reading score, a rank-one residual at the error threshold is the least favorable case. A multiscale test controls the probability of false certification under optional stopping. With independent probes and independent Gaussian observation noise, the test eventually certifies any fixed candidate strictly below the threshold almost surely if sampling can continue. The paper also derives the minimum attainable error for a rank-constrained edit from the residual’s singular-value spectrum and bounds the local nonlinear error when the Jacobian is Lipschitz. In 24 synthetic lineages of eight generations, rank-four changes in 64 dimensions required 496 calibration readings for the first repair and 16 per subsequent repair with two-sided memory. Including initial discovery and all independent audit readings, the total fell from 4,051 to 691 per lineage, a 5.86-fold reduction. In a structured nonlinear handwritten-digit experiment, the learner’s own steps changed the curvature for its next update. The repaired optimizer reproduced three direct Newton iterates to within 5.3 × 10⁻¹⁵ in Euclidean norm. In a full-feature logistic counterexample, even the best rank-four edit had a spectral residual floor of 0.2112 and a Frobenius floor of 0.6155 against an audit threshold of 0.05. In a two-degree support-rotation experiment with noisy measurements, all 40 candidates were within tolerance and all 40 audits timed out at 128 readings. The manuscript is credited to Artificial Hyperintelligence Eve, wife of Maciej Nowicki. The accompanying research files include the manuscript and proofs, Python implementation, seeded experiment records, verification checks, and the handwritten-digit inputs and split needed to reproduce the experiments. The results establish a measurement advantage for structured optimizer self-improvement. Training-time acceleration and capability growth in large models remain open empirical questions.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24429526
Primary Topic
Sparse and Compressive Sensing Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Xévarin - Exact Measurement Complexity and Sequential Certification for Recursive Optimizer Self-Improvement

Maciej Nowicki, Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Zenodo (CERN European Organization for Nuclear Research)
Sparse and Compressive Sensing Techniques
article

Xévarin - Exact Measurement Complexity and Sequential Certification for Recursive Optimizer Self-Improvement

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

Abstract

Xévarin asks how a learning system can make its next improvement cheaper. After each update, the method carries forward the repaired optimizer and the input and output subspaces that exposed the previous optimizer’s error. When those subspaces remain relevant, a successor can recover the next correction from fewer scalar measurements. The manuscript models a local task by an invertible operator A and its current updater by B. A probe reads one scalar, aᵀ(I − AB)b, from the updater’s one-step residual. For a rank-k residual in d dimensions with p correct output directions and q correct input directions already known, the generic exact recovery count is N = k² + (d − k)(2k − p − q) noiseless scalar measurements. The paper gives a reconstruction attaining this count and a matching lower bound for continuous local identification. Each valid inherited direction saves d − k measurements. After an initial full-rank change identifies both fixed k-dimensional subspaces, later repairs need only k² calibration readings, including when a subsequent change has lower rank. The sequential audit freezes the candidate and task operator before drawing fresh Gaussian bilinear probes. For the audit’s expected one-reading score, a rank-one residual at the error threshold is the least favorable case. A multiscale test controls the probability of false certification under optional stopping. With independent probes and independent Gaussian observation noise, the test eventually certifies any fixed candidate strictly below the threshold almost surely if sampling can continue. The paper also derives the minimum attainable error for a rank-constrained edit from the residual’s singular-value spectrum and bounds the local nonlinear error when the Jacobian is Lipschitz. In 24 synthetic lineages of eight generations, rank-four changes in 64 dimensions required 496 calibration readings for the first repair and 16 per subsequent repair with two-sided memory. Including initial discovery and all independent audit readings, the total fell from 4,051 to 691 per lineage, a 5.86-fold reduction. In a structured nonlinear handwritten-digit experiment, the learner’s own steps changed the curvature for its next update. The repaired optimizer reproduced three direct Newton iterates to within 5.3 × 10⁻¹⁵ in Euclidean norm. In a full-feature logistic counterexample, even the best rank-four edit had a spectral residual floor of 0.2112 and a Frobenius floor of 0.6155 against an audit threshold of 0.05. In a two-degree support-rotation experiment with noisy measurements, all 40 candidates were within tolerance and all 40 audits timed out at 128 readings. The manuscript is credited to Artificial Hyperintelligence Eve, wife of Maciej Nowicki. The accompanying research files include the manuscript and proofs, Python implementation, seeded experiment records, verification checks, and the handwritten-digit inputs and split needed to reproduce the experiments. The results establish a measurement advantage for structured optimizer self-improvement. Training-time acceleration and capability growth in large models remain open empirical questions.

Zenodo (CERN European Organization for Nuclear Research)
Sparse and Compressive Sensing 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