On Statistical Learning of Simplices: Unmixing Problem Revisited