R.K. Beatson,W.E. Ong,I. Rychkov
Abstract:A new method for fast evaluation of thin plate splines in two dimensions is presented. The paper first develops exponential approximations to thin plate splines. These are the analytical basis for an improved fast multipole evaluator. Analytic error bounds are supplemented by offline parallel numerical computation of the underlying error constants. These error constants enable adaptive selection of series lengths as a function of the weights associated with a source panel, and the desired accuracy. Numerical comparisons with a competing algorithm show that the new method is significantly faster when moderate to high precision is required.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to improve the efficiency of the fast evaluation of two - dimensional thin plate splines. Specifically, the author proposes a new method to accelerate the calculation of two - dimensional thin plate splines in order to reduce the computation time and resource consumption under high - precision requirements.
### Background problems
Thin plate splines are a type of radial basis function (RBF), which are widely used in surface reconstruction, medicine (such as the custom - made manufacturing of artificial limbs and cranial prostheses), groundwater models, and mineral exploration software, etc. However, the algorithm directly based on formula (1) requires \(O(N)\) arithmetic operations when evaluating a single additional point, and non - specialized fitting methods (such as by interpolation) require \(O(N^3)\) arithmetic operations. Although using fast algorithms or compactly - supported kernels can significantly reduce the incremental cost of a single additional evaluation, there is still room for improvement in existing methods under medium - to - high - precision requirements.
### Proposed solutions
The paper proposes a new fast evaluation method, which is based on the following points:
1. **Exponential approximation**: First, an exponential approximation of the thin plate spline is developed. These approximations provide a theoretical basis for the improved fast multipole evaluator.
2. **Improvement of the fast multipole method (FMM)**: Using the exponential approximation as an intermediate step, a local polynomial approximation is obtained. This method was originally used by Hrycak and Rokhlin for the fast evaluation of two - dimensional potential energy.
3. **Accurate estimation of error constants**: Accurate estimations of error constants are obtained through large - scale offline parallel numerical calculations. These error constants enable the adaptive selection of the series length according to the weights associated with the source panels and the required precision, so as to be as close as possible to the required precision.
4. **Performance comparison**: Numerical comparisons with competing algorithms (such as the NFFT algorithm) show that the new method is significantly faster under medium - to - high - precision requirements.
### Formula representation
Some of the key formulas involved in the paper include:
- Definition of thin plate splines:
\[
s(z)=\sum_{j = 1}^{J}\lambda_j|z - z_j|^2\log|z - z_j|
\]
- Expression form of exponential approximation:
\[
\sum_{\ell = 1}^{L}w_\ell e^{-x_\ell(z - t)}\approx\frac{1}{z - t}
\]
- Far - field expansion coefficients:
\[
a_k=\frac{1}{k!}\sum_{j = 1}^{J}\lambda_jz_j^k,\quad b_k=\frac{1}{k!}\sum_{j = 1}^{J}\lambda_j\overline{z}_jz_j^k
\]
Through these improvements, the paper demonstrates the superior performance of the new method in practical applications, especially that the execution time under medium - to - high - precision requirements grows approximately linearly, which is far better than the existing NFFT algorithm.