ABSTRACT：Dimensionality reduction is a big problem for pattern recognition, especially for small sample size. The normal LDA approach suffers plenty of problems for dimensionality reduction because it the singularity problem for data set. And the pseudoinverse method also has drawbacks because of the inverse is not always possible for dataset. In this paper, we propose a method which add perturbation into the matrix automatically and makes the inverse of the matrix is always possible. The experiment results show that the proposed method provide reliable performance for dimension reduction for recognition tasks .