Dimensionality Reduction and Optimal Union of Subspace Models for Data Clustering