A spectral analysis of subspace enhanced preconditioners (Q283325)

Let \(A\in\mathbb R^{n\times n}\) be symmetric positive definite (SPD) with eigenvalue decomposition \(A[V~V_\perp]=[V~V_\perp]\mathrm{diag}(\Lambda,\Lambda_\perp)\), where \(\Lambda\) contains the \(r\) smallest eigenvalues and \(V\) the corresponding eigenvectors. Three projections are defined: \(P_D=I-AVE^{-1}V^T\), \(P_C=I+VE^{-1}V^T\), and \(P_A=I-AVE^{-1}V^T+VE^{-1}V^T\) where \(E=V^TAV\). Applying one of these \(P_X\) to \(A\) will give a matrix \(P_XA\) whose eigenvalues are respectively \(\mathrm{diag}(0,\Lambda_\perp)\), \(\mathrm{diag}(I+\Lambda,\Lambda_\perp)\), and \(\mathrm{diag}(I,\Lambda_\perp)\) so that these \(P_X\) can serve as preconditioners. Clearly, to compute the eigenvalues of \(A\), the exact eigenvectors \(V\) cannot be used, but some orthogonal set \(\tilde{V}\approx V\) is used instead. Also the inverse \(E^{-1}\) is often approximated as \(\tilde{E}^{-1}\). The result is an approximate projector \(\tilde{P}_X\). Several bounds for the eigenvalues of the preconditioned \(\tilde{P}_X A\) matrices are proved. Numerical tests illustrate the technique.