Locking and restarting quadratic eigenvalue solvers (Q2706470)

Iterative methods are discussed for solving the quadratic eigenvalue problem NEWLINE\[NEWLINE Ku + i\omega C u -\omega^2Mu=0, NEWLINE\]NEWLINE where \(K, C\) and \(M\) are \(n\times n\) matrices and \(M\) is symmetric positive definite. The \(\omega\) is called an eigenvalue and \(u\) is the corresponding eigenvector. This problem can be written into a ``linearized'' form NEWLINE\[NEWLINE \left[\begin{matrix} K & 0\cr 0 & M \end{matrix}\right]\begin{pmatrix} u\cr \omega u \end{pmatrix}=\omega\left[\begin{matrix} -i c & M\cr M & 0 \end{matrix}\right]\begin{pmatrix} u\cr \omega u\end{pmatrix} NEWLINE\]NEWLINE and solved by a shift-and-invert Arnoldi method. Another approach is to tackle the quadratic eigenvalue problem directly by solving a sequence of the linear equation NEWLINE\[NEWLINE (K+i\omega C - \omega^2)y=r NEWLINE\]NEWLINE where \(\omega\) may change at each iteration step. An interesting linkage between these two approaches is established in this paper. Furthermore, the Schur form is extended to quadratic eigenvalues problems and proposed for the linearized problem in quadratic residual iteration and Jacobi-Davidson method. Thereafter, the author develops a locking and restarting strategy for computing a partial Schur form of the linearized problem. Extensive comparisons among the shift-and invert Arnoldi method, quadratic residual iteration and the Jacobi-Davidson method are performed and illustrated with numerical examples.