Low-rank updates of balanced incomplete factorization preconditioners (Q509632)

Suppose we know a balanced incomplete factorization (BIF) preconditioner for the nonsingular matrix \(A\) to solve the linear system \(Ax=b\). The problem is to update this preconditioner to solve \(Bx=b\) where \(B\) is a rank \(k\) update of \(A\), i.e., \(B=A+PQ^T\), \(P,Q\in\mathbb{R}^{n\times k}\). Use is made of the inverse Sherman-Morrison (ISM) decomposition of the matrix and its transpose and its relation to the LDU factorization of the matrix and its inverse (see [\textit{R. Bru} et al., SIAM J. Sci. Comput. 25, No. 2, 701--715 (2003; Zbl 1048.65042)]). For the preconditioner of the updated matrix, this ISM formula is applied to the equivalent augmented problem \[ \begin{bmatrix} A&P\\-Q^T&I\end{bmatrix}\begin{bmatrix} x\\\mathbb Q^Tx\end{bmatrix}=\begin{bmatrix} b\\0\end{bmatrix}. \] The computational cost of the method and the approximation properties of the preconditioned matrix are derived and the method is extensively tested on numerical examples.