A block-Lanczos method for large continuation problems (Q1817783)

When computing solution paths of systems of nonlinear parametric equations \[ G(u,\tau)=0, \] important problems may arise when bifurcation or turning points are encountered. In order to be able to detect such points and pass or traverse them along a desired path, a method is proposed here that solves iteratively the linear systems that arise in the application of Euler's method combined with Newton's method. The problems considered here produce symmetric Jacobian matrices and the iterative method is based on the block-Lanczos algorithm. The method allows the simultaneous computation of some small eigenvalues as well as its eigenvectors. This makes it easy to detect singular points, whose eigenvectors span the null space of the Jacobian matrix, which allows the computation of the least squares solution of the system \[ G_u*u_{\tau }+G_{\tau } =0, \] from which the derivative \(u_{\tau }\) for Euler's method is obtained. In a numerical example that illustrates the method, it is shown that it is more efficient than it would be with the application of conjugate gradients, that it is efficient for path following, and that it can be applied to very large problems, because it only requires the evaluation of matrix-vector products with Jacobian matrices.