Stability-Constrained Approximation in Spline KANs: Exact Layer Balancing and Budget-Compatible Saturation

DOI10.48550/arXiv.2609.17619MaRDI QIDQ7353391arXivFDO

Authors Aleksander Tankman

Publication date 14 September 2026

Published in arXiv

Tldr We study stability-constrained spline KANs and derive exact layerwise balancing for finite-depth networks under a Lipschitz budget constraint. We establish a constructive spline discretisation theorem with controlled slack, and provide lower bounds on classes constrained in both first and third derivative norms, showing that accumulation of error is not artefactual.

Full work available at URL https://doi.org/10.48550/arXiv.2609.17619

arXiv classification

cs.LGmath.NAstat.ML

Cites work

• A lower bound for the second moment of Schoenberg operator
• An identity for spline functions with applications to variation diminishing spline approximation
• Lower bounds for the approximation with variation-diminishing splines
• Lower estimates for linear operators with smooth range
• On the Betti Numbers of Real Varieties
• Optimally scaled matrices
• Saturation of positive operators
• Spline approximation by quasiinterpolants
• Stability of discrete-time switching systems with constrained switching sequences
• The approximation of continuous functions by positive linear operators
• The Collatz-Wielandt quotient for pairs of nonnegative operators
• The Perron-Frobenius theorem for multi-homogeneous mappings

This page was built for publication: Stability-Constrained Approximation in Spline KANs: Exact Layer Balancing and Budget-Compatible Saturation

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7353391)