Correction to: Kernel-based interpolation at approximate Fekete points

Toni Karvonen, Simo Särkkä, Ken'ichiro Tanaka

Introduction

This document serves as a formal correction (erratum) to the article "Kernel-based interpolation at approximate Fekete points," originally published in the journal Numerical Algorithms. The purpose of this correction is to rectify several typographical and notation errors found in the foundational mathematical sections of the original work. While these corrections do not alter the experimental results or the primary conclusions of the paper, they are essential for maintaining the mathematical rigor required for theoretical analysis and implementation of kernel-based interpolation methods.

Mathematical Preliminaries and Error Bounds

In Section 2 of the original paper, several notation errors were identified in the discussion of Reproducing Kernel Hilbert Spaces (RKHS) and pointwise error bounds.

Pointwise Error Bound (Equation 2.1)

The correction provides a revised version of Equation 2.1, which utilizes the Cauchy-Schwarz inequality to bound interpolation error. The relationship between the interpolation error, the RKHS norm of the function $\|f\|_{\mathcal{H}_K(\Omega)}$, and the "power function" $P_{X_n}(x)$ is now correctly represented. The power function is a critical metric in kernel methods, as it quantifies the uncertainty of the interpolant at any given point $x$.

Basis Expansions (Section 2.2)

The series representations of functions using orthonormal bases $\{\phi_\ell\}$ have been corrected. The original text omitted the necessary inner products required for reconstruction. The corrected version clarifies that a function $f$ is reconstructed from its inner products with the basis functions, ensuring consistency with standard Hilbert space theory.

Spectral Theory and RKHS Definitions

Section 3, which deals with spectral theory, contained errors regarding the definition of the RKHS and the operator $T(L^2(\mu))$.

Norm Weighting and Convergence

A vital correction was made to the definition of the RKHS norm in Section 3.3. The norm must properly include the reciprocal of the eigenvalues ($1/\lambda_\ell$). This weighting is fundamental: for a function to belong to the RKHS, its spectral coefficients must decay at a rate faster than the kernel's eigenvalues. The corrected definition ensures that the weighted sum of spectral coefficients squared is convergent, providing a rigorous basis for understanding how kernel smoothness influences convergence rates.

Inner Product Consistency

The "reproducing property" $f(x) = \langle f, K(\cdot, x) \rangle_{\mathcal{H}_K(\Omega)}$ is central to kernel methods. The correction ensures that the placement of the inner product in various series expansions throughout the text accurately reflects this identity.

Local Interpolation and Stability

In Section 4.3, an inequality involving local interpolants ($s_{i, g_i}$) was corrected. This fix is particularly relevant for researchers utilizing "Partition of Unity" methods or other localized kernel techniques. These methods decompose global interpolation problems into smaller, local sub-problems; the corrected inequality ensures the theoretical stability of these local approximations is correctly characterized.

Conclusion

These corrections provide the necessary mathematical clarity for the theoretical framework of kernel-based interpolation at approximate Fekete points. Researchers and practitioners are encouraged to reference these corrected equations to ensure the accuracy of their theoretical derivations and algorithmic implementations.

Summary updated

BibTeX
@article{karvonen2021correction,
  author = {Toni Karvonen and Simo Särkkä and Ken'ichiro Tanaka},
  title = {{Correction to: Kernel-based interpolation at approximate Fekete points}},
  journal = {Numer. Algorithms},
  year = {2021},
  doi = {10.1007/s11075-020-01034-0},
}
Figure