An inverse iteration method for eigenvalue problems with eigenvector nonlinearities (Q6486765)

The problem is to solve an eigenvalue problem \(A(v)v=\lambda v\), where the matrix \(A\) is symmetric but depends on the eigenvector. Moreover, \(A(v)\) satisfies \(A(\alpha v)=A(v)\) for \(\alpha\in\mathbb{R}\setminus\{0\}\). The iteration used is \(v_{k+1}=w_{k+1}/\|w_{k+1}\|\) with \(w_{k+1}=(J(v_k)-\sigma I)^{-1}v_k\) and \(J\) the derivative \(\frac{\partial}{\partial v}(A(v)v)\). The classical case where \(A\) is constant (hence \(J=A\)) is well documented. The authors show that for this general case the (local) convergence properties are basically the same as in the classical situation. Other side-results are given as well. For example an iteration, including an update of the shift, is described as a Newton step solving a fixed point problem. Moreover, (for a particular choice of the shift) it is shown to correspond to a discretization of an ordinary differential equation. Numerical simulations solving the Gross-Pitaevskii equation (a variant of the Schrödinger equation) illustrate these results.