Boosting stability and performance in randomized SVD

Randomized singular value decomposition (rSVD) is a highly efficient technique for approximating the singular value decomposition of a resultant matrix using a randomized sampling methods, as opposed to traditional SVD algorithms. This study presents the truncated singular value decomposition (trSVD), specifically tailored for Kernel-Based Methods such as radial basis function (RBF) ones. This novel approach significantly enhances stability and reduces the condition number of the linear system. Our experimental results demonstrate the effectiveness of the truncation process in trSVD. Furthermore, we thoroughly evaluate the performance of trSVD across Gaussian and Multiquadric RBFs. Truncated randomized SVD improves stability and accuracy in kernel-based methods with promising results in diverse RBF settings.