analiza aiqbench
This arXiv preprint (cs.LG) proposes a finite-dimensional framework for building data-adapted dictionaries and sparse representations, with a method called CCDS that extends matching-pursuit-style sparse coding. Its claims are mostly theoretical (orbit independence, interpolation bounds, stability), supported by experiments on synthetic geometry and standard classification benchmarks, where it reports competitive but not clearly superior predictive performance. Readers should check how the results compare against established sparse-coding and classification baselines, and whether the guarantees hold beyond the controlled settings tested.
A Constructive Orbit-Based Theory of Learnability
arXiv:2610.41641v1 Announce Type: new Abstract: We develop a finite-dimensional orbit framework for constructing data-adapted dictionaries, sparse realizations, and predictive extensions. The method uses wrapped weighted shifts to generate orbit dictionaries and local interpolation maps from independent orbit representatives, turning the learning problem into three explicit tasks: scaffold discovery, sparse realization on the discovered dictionary, and prediction through learned sparse coordinates. We introduce CCDS as the main constructive ingredient: a deviation-controlled sparse realization procedure that recovers MP/OMP-type behavior at the minimal-deviation boundary while allowing feasible interior paths with prescribed residual geometry that is inaccessible to MP/OMP. We prove generic orbit independence, exact block interpolation with norm control, observable bounds on sparse approximation distortion, sparse realization guarantees on selected supports, and fixed-path stability under admissible perturbations of the block interpolants. These results connect sparse approximation, interpolation, and finite-dimensional operator dynamics through the construction and analysis of data-induced orbit dictionaries. Experiments on synthetic geometry, controlled realization tests, and standard classification benchmarks show that selected orbit scaffolds have substantially better matched-rank projection behavior than random orbit or unstructured alternatives, that CCDS attains near-zero realization error at the minimal-deviation boundary, and that predictors trained from the sparse CCDS certificates give competitive performance.
Abstract:We develop a finite-dimensional orbit framework for constructing data-adapted dictionaries, sparse realizations, and predictive extensions. The method uses wrapped weighted shifts to generate orbit dictionaries and local interpolation maps from independent orbit representatives, turning the learning problem into three explicit tasks: scaffold discovery, sparse realization on the discovered dictionary, and prediction through learned sparse coordinates. We introduce CCDS as the main constructive ingredient: a deviation-controlled sparse realization procedure that recovers MP/OMP-type behavior at the minimal-deviation boundary while allowing feasible interior paths with prescribed residual geometry that is inaccessible to MP/OMP. We prove generic orbit independence, exact block interpolation with norm control, observable bounds on sparse approximation distortion, sparse realization guarantees on selected supports, and fixed-path stability under admissible perturbations of the block interpolants. These results connect sparse approximation, interpolation, and finite-dimensional operator dynamics through the construction and analysis of data-induced orbit dictionaries. Experiments on synthetic geometry, controlled realization tests, and standard classification benchmarks show that selected orbit scaffolds have substantially better matched-rank projection behavior than random orbit or unstructured alternatives, that CCDS attains near-zero realization error at the minimal-deviation boundary, and that predictors trained from the sparse CCDS certificates give competitive performance.
Machine Learning (cs.LG)
arXiv:2610.41641 [cs.LG]
(or arXiv:2610.41641v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.41641 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 286 cuvinte
Subiecte conexe
Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents
arXiv · 27 oct. 2026
CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks
arXiv · 7 aug. 2026
Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization
arXiv · 22 oct. 2026
GS-KAN: Parameter-Efficient Kolmogorov-Arnold Networks via Sprecher-Type Shared Basis Functions
arXiv · 8 iun. 2026