Kernel-Based Learning of Analytic Operators and Parametric PDEs
Organizations: MaLGa Center, DIMA, University of Genoa, Via Dodecaneso 35, 16146 Genoa, Italy · Institute for Machine Learning, Johannes Kepler University Linz, Altenberger Straße 69, 4040 Linz, Austria · MaLGa Center, DIBRIS, University of Genoa, Via Dodecaneso 35, 16146 Genoa, Italy
Abstract
We study kernel-based learning of analytic operators from finite samples, with a focus on how the regularity of the target should determine the choice of the kernel. We establish embeddings of analytic operators into vector-valued reproducing kernel Hilbert spaces and derive finite-sample, high-probability guarantees that quantify the uncertainty due to random sampling: in the noiseless setting, suitable analytic kernels yield rates exponential in a fractional power of the sample size, while Sobolev-type kernels yield polynomial rates. We apply the theory to the solution operator of a parametric elliptic PDE, where the training data are themselves finite-element approximations, and obtain bounds that separate the sampling and discretization contributions to the error and prescribe how the mesh size should be coupled to the sample size. The results are illustrated numerically.
Figures & tables
| Szegö | Matérn | Gaussian | pipeline floor | |
|---|---|---|---|---|
| 100 | 0.03013 0.00034 | 0.0316 0.0013 | 0.0320 0.0014 | 0.029750 |
| 300 | 0.004041 0.000042 | 0.00859 0.00017 | 0.00920 0.00016 | 0.003867 |
| 500 | 0.001263 0.000038 | 0.00466 0.00018 | 0.00373 0.00015 | 0.001118 |
| 700 | 0.000561 0.000021 | 0.00333 0.00013 | 0.00270 0.00015 | 0.000423 |
| 900 | 0.000308 0.000018 | 0.002625 0.000064 | 0.002369 0.000047 | 0.000190 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Szegö | Matérn | Gaussian | ||||
|---|---|---|---|---|---|---|
| anisotropic | isotropic | ARD | isotropic | ARD | isotropic | |
| 100 | 1.01 | 1.06 | 1.03 | 1.06 | 1.06 | 1.08 |
| 300 | 1.04 | 1.45 | 1.37 | 2.22 | 1.49 | 2.38 |
| 500 | 1.13 | 2.19 | 2.27 | 4.17 | 2.67 | 3.34 |
| 700 | 1.32 | 3.69 | 4.36 | 7.87 | 5.19 | 6.37 |
| 900 | 1.63 | 5.98 | 7.83 | 13.85 | 8.17 | 12.50 |
| anisotropy cross-validated | Szegö, | polynomial LS, | ||
|---|---|---|---|---|
| Szegö | polynomial LS | (Table 1 ) | isotropic | |
| 100 | 1.000 | 1.000 | 1.013 | 1.12 |
| 300 | 1.001 | 1.001 | 1.045 | 2.49 |
| 500 | 1.002 | 1.002 | 1.130 | 3.14 |
| 700 | 1.003 | 1.007 | 1.325 | 6.39 |
| 900 | 1.005 | 1.008 | 1.626 | 12.9 |
Operator Learning
Operator learning studies models that map functions or fields to functions, enabling fast prediction across scientific and dynamical systems.
See all 113 papers