Rectangular polar quadrature in 1D and its error analysis

Abstract

This paper presents a one-dimensional analog of the Rectangular-Polar (RP) integration strategy and its convergence analysis for weakly singular convolution integrals. The key idea of this method is to break the whole integral into integral over non-overlapping patches (subdomains) and achieve convergence by increasing the number of patches while approximating the integral on patches accurately using a fixed number of quadrature points. The non-singular integrals are approximated to high-order using Fej\'er first quadrature, and a specialized integration strategy is designed and incorporated for singular integrals where the kernel singularity is resolved by mean of Polynomial Change of Variable (PCV). We prove that for high-order convergence, it is essential to compute integral weights accurately, and the method's convergence rate depends critically on the degree of the PCV and the singularity of the integral kernel. Specifically, for kernels of the form $|x-y|^{-\alpha}(\alpha>-1)$, the method achieves high-order convergence if and only if $p(1-\alpha) \in \mathbb{N}$. This relationship highlights the importance of choosing an appropriate degree $p$ for the PCV. A new error estimate in a framework where convergence is achieved by reducing the approximating domain size while keeping the number of discretization points fixed is derived for the Fej\'er first quadrature, decay rate of Chebyshev coefficients, and the error in approximating continuous Chebyshev coefficients with discrete ones. Numerical experiments corroborate the theoretical findings, showcasing the effectiveness of the RP strategy for accurately solving singular integral equations. As an application of the method, numerical solutions of the surface scattering problem in the two dimensions are computed for complex domains, and the algorithm's efficacy is demonstrated for large-scale problems.