Homotopy method for the eigenvalues of symmetric tridiagonal matrices (Q455904)

The authors present a homotopy method for separately finding eigenvalues of symmetric, tridiagonal matrices. It is assumed that \(A\) has distinct diagonal elements and nonzero offdiagonal elements. Let \(D\) be the diagonal of \(A\) and \(E = A-D\). Define the matrix-valued function \(H(t) = D + t(A-D)\). Then \(H\) is a smooth function, \(H(0) = D\) and \(H(1) = A\). Theorems from perturbation theory tell us that there are curves in \(\mathbb R^2\) which give the eigenvalues of \(H(t)\) at any time. By assumption, \(H(t)\) is always symmetric and has \(n\) distinct eigenvalues. A bound on the timestep is shown to guarantee the convergence of Newton's method. This result takes the error in the approximation of the eigenpair into account. An improvement to the method is presented which involves using a different starting matrix.