On thin plate spline interpolation

M. Loehndorf,J.M. Melenk

DOI: https://doi.org/10.1007/978-3-319-65870-4

2017-05-15

Abstract:We present a simple, PDE-based proof of the result [M. Johnson, 2001] that the error estimates of [J. Duchon, 1978] for thin plate spline interpolation can be improved by $h^{1/2}$. We illustrate that ${\mathcal H}$-matrix techniques can successfully be employed to solve very large thin plate spline interpolation problems

Numerical Analysis

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the error estimate of thin plate spline interpolation. Specifically, the author aims to prove M. Johnson's result by using methods from partial differential equation (PDE) theory, that is, for a function $f$ satisfying certain conditions, its interpolation error estimate can be improved by a factor of $h^{1/2}$ compared to J. Duchon's result. In addition, the paper also shows how to use H - matrix techniques to efficiently solve large - scale thin plate spline interpolation problems.

### Specific description of the problem

Thin plate spline interpolation is a classical interpolation method used to find a smooth function $v(x)$ at given data points such that $v(x_i)=f(x_i)$ while minimizing a certain seminorm $\lvert v \rvert_{H^m(\mathbb{R}^d)}$. Specifically, consider the following optimization problem:

\[
\minimize \lvert v \rvert_{H^m(\mathbb{R}^d)} \quad \text{subject to} \quad v(x_i) = f(x_i), \quad i = 1,\ldots, N
\]

where the seminorm $\lvert v \rvert_{H^m(\mathbb{R}^d)}$ is induced by the bilinear form $\langle v, w \rangle_m$:

\[
\langle v, w \rangle_m:=\sum_{|\alpha| = m}\frac{m!}{\alpha!}\int_{\mathbb{R}^d}D^\alpha vD^\alpha w\,dx
\]

### Main contributions

1. **Improvement of error estimate**: The author proves that for a function $f$ satisfying certain smoothness conditions, the interpolation error can be improved by a factor of $h^{1/2}$ compared to Duchon's result. Specifically, when $f\in B^{m + 1/2}_{2,1}(\Omega)$ and the boundary is smooth enough, the error estimate can be improved to:

\[
\lvert E_\Omega f - I f \rvert_{H^m(\mathbb{R}^d)}\leq C_1h^{1/2}\|f\|_{B^{m+1/2}_{2,1}(\Omega)}
\]

2. **Application of H - matrix techniques**: The paper shows how to use H - matrix techniques to efficiently solve large - scale thin plate spline interpolation problems. H - matrix techniques can approximate a dense matrix as a low - rank block matrix, thereby significantly reducing the computational complexity.

### Numerical experiments

To verify the above theoretical results, the author conducted numerical experiments. The experimental results show that for the classical thin plate spline interpolation with $m = d=2$, the error indeed converges at the expected rate. For example, when interpolating on uniformly distributed nodes in square and L - shaped domains, the order of convergence of the error is in line with the theoretical prediction.

### Conclusion

By introducing elliptic partial differential equation theory and H - matrix techniques, this paper successfully improves the error estimate of thin plate spline interpolation and provides an efficient numerical solution method. This not only deepens the understanding of the properties of thin plate spline interpolation but also provides a powerful tool for practical applications.