An algorithm framework of sparse minimization for positive definite quadratic forms

Many well-known machine learning and pattern recognition methods can be seen as special cases of sparse minimization of Positive Definite Quadratic Forms (PDQF). An algorithm framework of sparse minimization is proposed for PDQF. It is theoretically analyzed to converge to global minimum. The computational complexity is analyzed and compared with the state-of-the-art Fast Iterative ShrinkageThresholding Algorithm (FISTA). Some well-known machine learning and pattern recognition methods are illustrated to be optimized by the proposed algorithm framework. Illustrative experiments show that Sparse Representation Classification (SRC) and Least Absolute Shrinkage and Selection Operator (LASSO) via the proposed method converges much faster than several classical methods. & 2014 Elsevier B.V. All rights reserved.