Optimization Methods for Sparse Approximation

TLDR: This thesis proposes a novel augmented Lagrangian method for solving the l1-norm relaxation problems of the original l0 minimization problems and applies it to the proposed formulation of sparse principal component analysis (PCA), and establishes some convergence results for both inner and outer methods.

Abstract: In the last two decades, there are numerous applications in which sparse solutions are concerned. Mathematically, all these applications can be formulated into the l0 minimization problems. In this thesis, we first propose a novel augmented Lagrangian (AL) method for solving the l1-norm relaxation problems of the original l0 minimization problems and apply it to our proposed formulation of sparse principal component analysis (PCA). We next propose penalty decomposition (PD) methods for solving the original l0 minimization problems in which a sequence of penalty subproblems are solved by a block coordinate descent (BCD) method. For the AL method, we show that under some regularity assumptions, it converges to a stationary point. Additionally, we propose two nonmonotone gradient methods for solving the AL subproblems, and establish their global and local convergence. Moreover, we apply the AL method to our proposed formulation of sparse PCA and compare our approach with several existing methods on synthetic, Pitprops, and gene expression data, respectively. The computational results demonstrate that the sparse principal components (PCs) produced by our approach substantially outperform those by other methods in terms of total explained variance, correlation of PCs, and orthogonality of loading vectors. For the PD methods, under some suitable assumptions, we establish some convergence results for both inner (the BCD method) and outer (the PD method) iterations, respectively. We test the performance of our PD methods by applying them to sparse logistic regression, sparse inverse covariance selection, and compressed sensing problems. The computational results demonstrate that when solutions of same cardinality are sought, our approach applied to the l0-based models generally has better solution quality and/or speed than the existing approaches that are applied to the corresponding l1-based models. Finally, we adapt the PD method to solve our proposed wavelet frame based image