Multivariate confluent Vandermonde with G-Arnoldi and applications

Lei-Hong Zhang, Ya-Nan Zhang, Linyi Yang, Yifu WuarXiv 2024 · DOI · arXiv

Introduction

Numerical approximation of multivariate functions and their derivatives is a cornerstone of scientific computing. However, standard polynomial bases often lead to ill-conditioned Vandermonde matrices, limiting accuracy. This paper presents the Multivariate Confluent Vandermonde with G-Arnoldi (MV+G-A) method, which extends the Vandermonde with Arnoldi framework to handle multiple variables and derivative data. By employing a flexible $G$-orthogonalization process and explicit recurrence relations, the method achieves machine-precision accuracy and perfect conditioning for least-squares fitting and solving PDEs on irregular domains.

Methodology

The MV+G-A framework relies on three core pillars: $G$-orthogonalization, structured ordering, and derivative recurrences.

G-Orthogonalization and Basis Construction

Instead of the standard Euclidean inner product, the authors utilize a $G$-inner product, defined as $\langle x, y \rangle_G = x^T G y$. This allows the Arnoldi process to generate a basis that is orthogonal with respect to a specific operator or domain geometry. By choosing an appropriate $G$ matrix, the resulting least-squares problem can be reduced to a condition number of 1, ensuring maximum numerical stability.

Multivariate Ordering and Recurrence

To manage multiple variables, the algorithm employs "graded reverse lexicographical" (Grevlex) ordering for monomials. This structured approach facilitates the derivation of recurrence relations for both function values and their partial derivatives.

Matrix representation of the recurrence relation for second-order partial derivatives in the MV+G-A framework.

As shown in the figure above, the framework uses explicit matrix-based recurrence relations to build the basis for derivatives. This avoids the instability of numerical differentiation and allows for efficient evaluation of the model at new coordinates. The process is divided into two stages: a Fitting Stage, where a discrete $G$-orthogonal polynomial basis is generated, and an Evaluation Stage, where recurrence relations are used to compute values at target points.

Numerical Experiments and Results

The method was tested on several challenging benchmarks, including unit disks, spheres, and irregular "holed" domains (e.g., an elliptical boundary with a circular cutout).

Key results include:

  • Perfect Conditioning: The coefficient matrices for least-squares problems maintained a condition number of 1, meaning the solution is as accurate as the computer's floating-point precision allows.
  • High Precision: Residual errors for complex Poisson equation problems reached as low as $10^{-14}$, effectively hitting machine epsilon.
  • Efficiency: The evaluation stage proved computationally lean, with complexity scaling quadratically with the number of basis functions.

Conclusion

The Multivariate Confluent Vandermonde with G-Arnoldi method offers a powerful solution for stable, high-precision function fitting and PDE solving. By integrating derivative information and customizable orthogonality, it overcomes the traditional limitations of Vandermonde-based methods. Future work will explore handling "serious breakdowns" in semi-definite cases and extending the framework to non-linear differential equations.

Summary updated

BibTeX
@misc{zhang2024multivariate,
  author = {Lei-Hong Zhang and Ya-Nan Zhang and Linyi Yang and Yifu Wu},
  title = {{Multivariate confluent Vandermonde with G-Arnoldi and applications}},
  year = {2024},
  eprint = {2404.09266},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.09266},
}
Figure