Block-Exact Sparse Kernel Regression via Closed-Form Laplace Integral Indexing and Cross-Bandwidth Index Sharing
Bandwidth selection in Nadaraya-Watson and kernel ridge regression incurs quadratic time and memory costs on large datasets. Low-rank Nyström and block-mean summary methods cut this cost, yet both attenuate local high-frequency peaks. Inverted-file centroid routing preserves local detail inside selected blocks, while isotropic distance scoring ignores anisotropic variation within blocks. Laplace-IndexShare projects data into univariate-ordered blocks and scores them with a closed-form Gaussian Laplace convolution. Exponential calls are avoided during routing by sharing top-k block indices across cross-validation bandwidth grids. Evaluating uncompressed kernel weights on 2.48% of pairs yields 41x sparsity with an exact 2y_inf epsilon_tail pointwise error bound. An explicit four-dimensional construction exhibits a 119x kernel-mass inversion of isotropic routing that Laplace scoring resolves. Across 30 ablation tasks, closed-form scoring avoids iterative overflow and recovers 90.2% of the block-mass oracle gain. Across 30 unseen tasks, Laplace-IndexShare outperforms relevance-weighted inverted-file routing at p_holm = 8.0x10^-5 and rank-R Nyström regression at p_holm = 9.3x10^-4. At n = 15,000, test mean squared error drops 43.9% below isotropic routing with a 1.82x speedup over dense cross-validation, while small-sample isotropic differences are not significant at p_holm = 0.060.
Authors
- Felipe Provezano Coutinho (ORCID: https://orcid.org/0000-0001-7206-5856)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24374092
- Primary Topic
- Machine Learning and ELM
- Type
- preprint