Core Concepts
Non-asymptotic theory provides deterministic approximations for KRR test error.
Abstract
The paper introduces a non-asymptotic theory for Kernel Ridge Regression (KRR) focusing on deterministic equivalents for the test error. It establishes that the test error can be accurately approximated by a closed-form estimate based on eigenvalues and target function coefficients. The study considers general spectral and concentration properties on the kernel eigendecomposition to provide non-asymptotic bounds. The theoretical predictions show excellent agreement with numerical simulations across various scenarios. Additionally, the generalized cross-validation (GCV) estimator is proven to concentrate uniformly on the test error over a range of ridge regularization parameters, including zero. The analysis relies on recent advances in random matrix functionals pioneered by Cheng and Montanari.
Stats
A recent string of work has empirically shown that the test error of KRR can be well approximated by a closed-form estimate derived from an 'equivalent' sequence model.
The equivalence holds for a general class of problems satisfying spectral and concentration properties on the kernel eigendecomposition.
The deterministic approximation for the test error of KRR only depends on eigenvalues and target function alignment to eigenvectors.
The GCV estimator concentrates on the test error uniformly over a range of ridge regularization parameters.