Superaccurate finite element eigenvalues via a Rayleigh quotient correction

Abstract

The consistent finite element formulation of the vibration problem generates upper bounds on the corresponding exact eigenvalues but requires the solution of the highly expensive general algebraic eigenproblem Kx = λ Mx with a global matrix M that is of the same sparsity pattern as the global stiffness K. The lumped, diagonal, mass matrix finite element formulation is no longer variationally correct but results in a simplified algebraic eigenproblem of comparable accuracy. We may write the mass matrix as a linear matrix function, M ( γ ) = M 1 + γ M 2 , of parameter γ such that M ( γ = 1 ) is the (diagonal) lumped mass matrix and M ( γ = 0 ) is the consistent mass matrix. It has been shown that an optimal γ exists between these two states which results in superaccurate eigenvalues. What detracts from the appeal of this approach is that the superior accuracy thus achieved comes at the hefty price of having to solve the still general algebraic eigenproblem with a nondiagonal mass matrix. In this note we show that the same superior accuracy can be had by first computing an eigenvector u from Ku = λ Du , in which D = M 1 + M 2 is the lumped, diagonal, mass matrix, and then obtaining the corresponding, superaccurate, eigenvalue from the Rayleigh quotient R [ u ] = u T Ku / u T M ( γ ) u , M ( γ ) = M 1 + γ M 2 for an optimal γ .