Fixed Point Continuation (FPC). General l1-regularized minimization problems of the form (1) min ||x||1 + μ f(x), where f is a convex, but not necessarily strictly convex, function, can be solved with a globally-convergent fixed-point iteration scheme. In addition, q-linear rates of convergence can be achieved under mild conditions. Problems in the form of (1) are often of interest when x is expected to be sparse, or contain outliers. In compressed sensing signal reconstruction, f(x) is a weighted least-squares term. In this case, q-linear convergence rates can be shown as long as a certain reduced Hessian is full rank, or a strict complementarity condition holds. In order to obtain good practical performance, the basic fixed-point iterations should be augmented with a continuation approach. In brief, the continuation approach consists of solving (1) for an increasing sequence of μ values, using the solution at the last μ value as the starting point for the next μ value. Thus, Fixed-Point Continuation (FPC).
Keywords for this software
References in zbMATH (referenced in 2 articles )
Showing results 1 to 2 of 2.
- Li, Chong-Jun; Zhong, Yi-Jun: A pseudo-heuristic parameter selection rule for (l^1)-regularized minimization problems (2018)
- Karimi, Sahar; Vavasis, Stephen: IMRO: A proximal quasi-Newton method for solving (\ell_1)-regularized least squares problems (2017)