首页 - 学术活动The linear response eigenvalue problem associated with the block matrix H=[O K; M O], where K is positive semidefinite and M is positive definite, arises in a wide range of applications, including Bose–Einstein condensation, density functional theory, and classical mechanics. When K is singular, the zero eigenvalue of H is defective, rendering H non-diagonalizable. The resulting nontrivial Jordan structure may lead to a loss of accuracy, slow or stagnating convergence, or even divergence in numerical eigensolvers. Existing methods typically rely on a priori knowledge of the generalized nullspace or the spectral distribution. Such information, however, is often unavailable and can be computationally difficult to obtain. In this work, we develop a family of biorthogonal structure-preserving eigensolvers comprising a direct method and a subspace iteration method. Neither method requires a priori spectral information. The direct method employs standard matrix decompositions to explicitly construct the generalized nullspace and separate it from the invariant subspaces associated with nonzero eigenvalues, thereby enabling accurate computation of the complete eigenstructure. The iterative method projects the original problem onto a sequence of low-dimensional subspaces and solves each resulting projected eigenvalue problem using the proposed direct method. Through successive subspace refinement, the method automatically and progressively approximates both the generalized nullspace and the desired eigenpairs associated with positive eigenvalues, while preserving, at every iteration, the biorthogonality between approximate eigenspaces corresponding to eigenvalues of distinct magnitudes. Numerical experiments demonstrate that the proposed methods achieve high accuracy, rapid convergence, and strong robustness and compare favorably with state-of-the-art eigensolvers, including LOBP4dCG and BOSP.