Abstract: Recently, most manifold learning algorithms take advantage of distance to measure similarity of data, and obtain satisfactory results, but most of them can not handle subspace deviation caused by noise. To solve this problem, a global dimensionality reduction algorithm based on angle optimization is proposed in this paper. Theoretically it proves that the main factors of subspace deviation are the length of the center sample and the angle of deviation from the low-dimensional space by providing covariance matrix update mode of multi-sample incremental. Consequently, the algorithm solves the subspace deviation problem caused by noise. Compared with the principal component analysis, it can integrate better with other algorithms to solve small sample problems. Experiments carried out on handwork and real data sets show a clear improvement over the results of other linear algorithms.