QR-based Parallel Set-Valued Approximation with Rational Functions

Simon Dirckx,Karl Meerbergen,Daan Huybrechs

2023-12-16

Abstract:In this article a fast and parallelizable algorithm for rational approximation is presented. The method, called (P)QR-AAA, is a set valued variant of the Adaptive Antoulas Anderson (AAA) algorithm. It builds on the Set-Valued AAA framework from [16], accelerating it by using an approximate orthogonal basis obtained from a truncated QR decomposition. We demonstrate both theoretically and numerically this method's robustness. We show how it can be parallelized while maintaining the desired accuracy, with minimal communication cost.

Numerical Analysis

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to improve the computational efficiency and parallelization ability of set - valued (or vector - valued) rational approximation while maintaining the required precision. Specifically:

1. **Accelerating set - valued rational approximation**: The paper proposes a method named (P)QR - AAA, which is an improvement on the existing adaptive Antoulas - Anderson (AAA) algorithm. By using the approximately orthogonal basis obtained from the truncated QR decomposition, (P)QR - AAA can accelerate the computational process of set - valued rational approximation.

2. **Parallelization ability**: The paper shows how to parallelize this method to reduce communication costs while maintaining the required precision. This makes the method suitable for approximation problems of large - scale data sets and matrix functions.

3. **Theoretical analysis and verification**: The author not only theoretically proves the robustness of this method but also verifies its effectiveness through numerical experiments. In particular, this method shows significant advantages when dealing with the near - field wavenumber - dependent approximation problem in the boundary element method (BEM).

### Specific problem description

Set - valued rational approximation is a powerful tool for simultaneously approximating matrix - valued functions and large sets of functions. The traditional AAA algorithm constructs rational approximation through a greedy iterative scheme, but it has a high computational cost when dealing with large - scale data. For this reason, this paper proposes a parallelization method based on QR decomposition, aiming to solve the following problems:

- **Low computational efficiency**: The traditional method has high computational complexity when dealing with large - scale data, resulting in too long a computational time.

- **Difficult to parallelize**: The traditional method is difficult to parallelize effectively, which limits its application in high - performance computing environments.

- **Precision guarantee**: It is necessary to ensure that the approximation precision does not decrease during the parallelization process while minimizing communication overhead.

By introducing the (P)QR - AAA method, the paper successfully solves these problems and provides an efficient and scalable solution for set - valued rational approximation.