Theory for Eigenvalue Sensitivity Analysis
In simplified terms, the eigenvalue problem solved by COMSOL Multiphysics can be written
(2-8)
where λ is the eigenvalue and u is the eigenmode, while E, D, and K are the eliminated mass, damping, and stiffness matrices, respectively. These matrices are also sometimes referred to as Ec, Dc, and Kc. Taking the derivative of Equation 2-8 with respect to a set of controls, ξ, yields
(2-9)
Multiplying this equation with the adjoint mode, u, and rearranging yields
(2-10)
so the adjoint equation can be identified as
(2-11)
If the problem is symmetric, it is self-adjoint; that is, u = u. Otherwise, the left eigenvectors have to be computed. In either case, the sensitivity of the eigenvalue can be computed as
(2-12)
In COMSOL Multiphysics this is referred to as the adjoint gradient method. The implementation assumes that the controls do not enter any (nonweak) constraints. Moreover, the eigenvalue problem can be a linearization of a nonlinear problem, and the implementation assumes that the linearization point is independent of the controls.