CholeskyQR is an efficient algorithm for QR factorization with several advantages compared with orhter algorithms. In order to improve its orthogonality, CholeskyQR2 is developed \cite{2014}\cite{error}. To deal with ill-conditioned matrices, a shifted item $s$ is introduced and we have Shifted CholeskyQR3 \cite{Shifted}. In many problems in the industry, QR factorization for sparse matrices is very common, especially for some sparse matrices with special structures. In this work, we discuss the property of CholeskyQR-type algorithms for sparse matrices. We introduce new definitions for the input sparse matrix $X$ and divide them into two types based on column properties. We provide better sufficient conditions for $\kappa_{2}(X)$ and better shifted item $s$ for CholeskyQR-type algorithms under certain element-norm conditiones(ENCs) compared with the original ones in \cite{Shifted}\cite{error}, together with an alternative error analysis for the algorithm. The steps of analysis utilize the properties of the $g$-norm of the matrix which is given in the previous work. Moreover, a new three-step CholeskyQR-type algorithm with two shifted items called 3C is developed for sparse matrices with good orthogonality. We do numerical experiments with some typical real examples to show the advantages of improved algorithms compared with the original ones in the previous works.
翻译:暂无翻译