Convergence analysis of Lanczos-type methods for the linear response eigenvalue problem (Q1946190)

The convergence of two different Lanczos methods is analyzed for the linear response eigenvalue problem \[ H = \left[\begin{matrix} 0 & K \\ M & 0 \end{matrix}\right] \] where \( K \) and \( M \) are \( n\times n \) real symmetric and \( M \) is positive definite. The first method is based on the following initial result of the authors, and can be viewed as a natural extension of the classical Lanczos method for the symmetric eigenvalue problem. Theorem 1. Given \( 0\neq v_0\in\mathbb{R}^n \) and \( 0\neq u_0\in\mathbb{R}^n \) such that \[ u_0=\gamma Mv_0, \] then there exist nonsingular \( U,V\in\mathbb{R}^{n\times n} \) such that \( Ve_1=\alpha v_0 \) and \( Ue_1=\beta u_0 \) for some \( \alpha,\beta\in\mathbb{R} \), and \[ U^TV=I,\quad U^TKU=T,\quad V^TMV = D,\tag{*} \] where \( T \) is tridiagonal and \( D \) is diagonal. The second method, proposed by \textit{E. V. Tsiper} [``A classical mechanics technique for quantum linear response'', J. Phys. B: At. Mol. Opt. Phys. 34, No. 12, L401--L407 (2001); ``Varational procedure and generalized Lanczos recursion for small-amplitude classical oscillations'', JEPT Lett. 70, No. 11, 751--755 (1999)], is based on a decomposition for \( K \) and \( M \) similar to~(*), but for \( T \) and \( D \) both tridiagonal, and for given \( 0\neq v_0\in\mathbb{R}^n \) and \( 0\neq u_0\in\mathbb{R}^n \) such that \( u_0^Tv_0\neq0 \) and \( u_0\neq\gamma Mv_0 \) for any \( \gamma \). For both methods, bounds are derived on the errors for the eigenvalue and eigenvector approximations obtained by means of the reduced matrix \[ H_{2k} = \left[\begin{matrix} 0 & T_k \\ D_k & 0 \end{matrix}\right]\in\mathbb{R}^{2k\times 2k}. \] These bounds suggest that the approximations obtained by the second method may converge (much) slowlier than those obtained by the first method. The paper concludes with some numerical examples that support this claim.