Dynamic condensation techniques have been broadly applied to the domains of testanalysis-model correlation, vibration control, damage detection and so on to reduce the structural matrices (sti!ness, mass and/or damping matrices) of "nite element models. Based on the subspace iteration method in the eigenproblem, a dynamic condensation approach is derived in this paper. It is iterative. Comparing almost all the iterative schemes for dynamic condensation proposed in the past, the present approach has three advantages: (1) The convergence is much faster than all these methods, especially when the eigenpairs of the reduced model are close to those of the full model. ( 2) The convergence of the iterative scheme can be proved simply. (3) It is computationally e$cient since it is unnecessary to calculate the sti!ness and mass matrices as well as the eigensolutions of the condensed model in every iteration. Two iterative schemes, which are based on the convergence of the eigenvalues of the reduced model and the column vectors of the dynamic condensation matrix, respectively, are given in this paper. Not only the accuracy of eigenvalues, but also that of eigenvectors are considered in every iteration. Numerical examples are also presented to show the e$ciency of the proposed method.