Sharp inverse statements for kernel interpolation

Tizian Wenzel

2024-09-12

Abstract:While direct statements for kernel based interpolation on regions $\Omega \subset \mathbb{R}^d$ are well researched, far less is known about corresponding inverse statements. The available inverse statements for kernel based interpolation so far are not sharp. In this paper, we derive sharp inverse statements for interpolation using finitely smooth kernels, such as popular radial basis function (RBF) kernels like the class of Matérn or Wendland kernels. In particular, the results show that there is a one-to-one correspondence between the smoothness of a function and its approximation rate via kernel interpolation: If a function can be approximated with a given rate, it has a corresponding smoothness and vice versa.

Numerical Analysis

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the inverse estimation problem in kernel interpolation. Specifically, the goal of the paper is to derive accurate inverse estimates when interpolating with finitely smooth kernels (such as the popular radial basis function (RBF) kernels, e.g., Matérn or Wendland kernels).

### Background and Problem Description

In approximation theory:
- **Direct Estimation**: It gives the error estimate of approximating an element (such as a function) using predefined basis elements (such as polynomials). The convergence rate usually depends on the properties of the element to be approximated (such as the smoothness of the function).
- **Inverse Estimation**: Conversely, it derives the properties of the approximated element (such as the smoothness of the function) based on the approximation rate.

For the kernel interpolation method, although direct estimation has been well - studied, the existing inverse estimates are not accurate. This paper aims to fill this gap and provide accurate inverse estimates for kernel interpolation.

### Main Contributions

The main contribution of the paper lies in proving the following theorem:

#### Theorem 1 (Main Result)
Consider a compact Lipschitz domain \(\Omega\subset\mathbb{R}^d\) and a continuous kernel \(k\) such that \(H_k(\Omega)\approx H_\tau(\Omega)\) for some \(\tau > d/2\). Let \(f\in C(\Omega)\), and for some constants \(c_f,\beta,h_0 > 0\), satisfy
\[
\|f - s_{f,X}\|_{L^2(\Omega)}\leq c_fh_X^\beta
\]
for all \(X\subset\Omega\) with \(h_X\leq h_0\) and \(\rho_X\leq44\) (quasi - uniform distribution). If \(\beta\in(0,\tau]\), then \(f\in H_{\beta'}(\Omega)\) for all \(\beta'\in(0,\beta)\). If \(\beta > 1\), then \(f\in H_k(\Omega)\approx H_\tau(\Omega)\).

This result significantly extends the inverse estimates in the existing literature. In particular, it provides the correct estimate of the function smoothness, thus establishing for the first time a one - to - one correspondence between the function smoothness and the kernel interpolation approximation rate.

### Methods and Techniques

To prove the above theorem, the author adopts the following methods and techniques:
1. **Discretization Estimation**: Transform the convergence rate of the continuous \(L^2(\Omega)\) norm into the convergence rate of the discrete \(L^2(X)\) norm.
2. **Cauchy Sequence Construction**: By constructing a series of nested and increasingly dense point sets, prove that these interpolation sequences form a Cauchy sequence.
3. **Completeness in Sobolev Spaces**: Utilize the completeness of Sobolev spaces to obtain the existence of the limit element.

### Significance and Impact

The results of this paper not only provide more accurate theoretical support for the kernel interpolation method but also fill the gap of inaccurate inverse estimates in the existing literature. This is of great significance for understanding the smoothness of functions and their relationship with the kernel interpolation approximation rate. Moreover, these results can be applied to various scientific computing and engineering fields, especially in cases involving high - dimensional data and complex geometric structures.

### Future Work

The paper also discusses future research directions, including:
- Generalizing the \(L^2(\Omega)\) error analysis to \(L^\infty(\Omega)\) or other \(L^p(\Omega)\) error statements.
- Exploring how to generalize the inverse estimates to adaptively distributed points (such as greedily selected points).
- Extending the results to the kernel interpolation problem on manifolds.

In conclusion, this paper provides an important theoretical progress for the inverse estimates of the kernel interpolation method and lays the foundation for further research.