This paper1 presents a novel pattern classification scheme: Class-wise Non-Principal Component Analysis (CNPCA), which utilizes the distribution characteristics of the samples in each class. The Euclidean distance in the subspace spanned by the eigenvectors associated with smallest eigenvalues in each class, named CNPCA distance, is adopted as the classification criterion. The number of the smallest eigenvalues is selected in such a way that the classification error in a given database is minimized. It is a constant for the database and can be determined by experiment. The CNPCA classification scheme usually outperforms other classification schemes under the situations of high computational complexity (associated with high dimensionality of features and/or calculation of inverse variance matrix) or high classification error rate (e.g., owing to the scattering of between-class being less than that of within-class). The experiments have demonstrated that this method is promising in practical applications.