Second Derivatives for Optimizing Eigenvalues of Symmetric Matrices

Let A denote an n n real symmetric matrix-valued function depending on a vector of real parameters, x 2 < m. Assume that A is a twice continuously diierentiable function of x, with the second derivative satisfying a Lipschitz condition. Consider the following optimization problem: minimize the largest eigenvalue of A(x). Let x denote a minimum. Typically, the maximum eigen-value of A(x) is multiple, so the objective function is not diierentiable at x , and straightforward application of Newton's method is not possible. Nonetheless, the formulation of a method with local quadratic convergence is possible. The main idea is to minimize the maximum eigenvalue subject to a constraint that this eigenvalue has a certain multiplicity. The manifold of matrices with such multiple eigenvalues is parameterized using a matrix exponential representation, leading to the deenition of an appropriate Lagrangian function. Consideration of the Hessian of this Lagrangian function leads to the second derivative matrix used by Newton's method. The convergence proof is nonstandard because the parameterization of is explicitly known only in the limit. In the special case of multiplicity one, the maximum eigenvalue is a smooth function and the method reduces to a standard Newton iteration.