A novel optimal discriminant principle in high dimensional spaces

A novel optimal discriminant principle and algorithm in high dimensional space are presented in this paper. The new optimal discriminant vectors have the property: in their spanned space, the within-class distance of training samples equals to zero while the between-class distance doesn´t equal to zero. We also illustrate how many optimal discriminant vectors satisfying property above can be obtained. We apply this method to the face recognition and the experimental result shows the performance is superior to the existed methods.