A new justification of the Jacobi-Davidson method for large eigenproblems (Q886146): Difference between revisions

The author derives a new justification of the Jacobi-Davidson method for large eigenproblem, in a way that explains his robustness. The first part presents the large and sparse eigenvalue problem \[ AX=\lambda x\tag{1} \] or more generally the nonlinear eigenproblem \[ T(\lambda)x=0 \tag{2} \] where \(A\in{\mathbb C}^{n\times n}\) and \(T:D\rightarrow{\mathbb C}^{n\times n}\), \(D\subset{\mathbb C}\) being a family of sparse matrices, and the associated numerical methods. For the linear problem (1) iterative projection methods are very efficient if a small number of eigenvalues and eigenvectors are desired. For truly large eigenproblems, methods like the Lanczos algorithm and the Arnoldi algorithm are very sensitive to inexact solutions of \((A-\sigma I)x=y,\) \(\sigma\) denoting a shift which is close to the wanted eigenvalues. The Jacobi-Davidson iterative projection method is more robust to inexact expansions of the search spaces, and his robustness is justified by the present approach. The second part contains a geometric derivation of a robust search space expansion. The author characterizes the maximal possible acute angle between the exact and inexact planes. The third part is devoted to the Jacobi-Davidson method and his robust behaviour. The last part concerns inexact Krylov subspace methods.