A MATLAB Solver for L1-Regularization Problems


FPC_AS is a successor of FPC [link]. While FPC_AS still performs shrinkage iterations and continuation as its predecessor, most of the code has been rewritten. Compared to FPC, which has good performance on large-scale problems with highly sparse solutions, FPC_AS works better overall and much better on certain difficult problems arising in compressed sensing, to name a few, those with sparse, but not highly sparse, solutions and those whose solutions have both very large and very small nonzero components (i.e., the solutions have huge dynamic ranges). These problems are difficult because it is hard to identify certain nonzero components in the solutions as these components are either too small or have only slight advantage to represent b over some of the others. FPC_AS was designed with active set identification and sub-optimization to help recover these components in the solutions.