Linea discriminant analysis(LDA) is one of the most classical subspace learning and supervised learning methods. Inspired by manifold learning, many improved methods based on LDA have been proposed in recent years. Although the motivations of these methods are different, they are all based on the Euclidean distance to measure the spatial dispersion of the samples. The non-linear characteristic of Eucilidean distance brings about two problems: 1) these methods are too sensitive to noise and outlier; 2) the essential structure would be destructed, due to the overemphasis of the points which has a large local dispersion in manifold or multimodal datasets. To solve these problems, a new dimension reduction method based on nonparametric discriminant analysis (NDA) is proposed, called a dynamic weighted nonparametric discriminant analysis (DWNDA).Then DWNDA uses the dynamic weighted distance to caluculate the within-class and between-class scatters. It can not only retain the essential geometrical structure of multimodal datasets, but also make better use of the discriminant information between marginal point pairs. Hence, the DWNDA shows better robustness to noise and outlier than other methods, which is also demonstrated in experiments. Besides, the DWNDA also shows excellent performance for face and handwrting classification.