Optimally Tuned Iterative Reconstruction Algorithms for Compressed Sensing
read the original abstract
We conducted an extensive computational experiment, lasting multiple CPU-years, to optimally select parameters for two important classes of algorithms for finding sparse solutions of underdetermined systems of linear equations. We make the optimally tuned implementations available at {\tt sparselab.stanford.edu}; they run `out of the box' with no user tuning: it is not necessary to select thresholds or know the likely degree of sparsity. Our class of algorithms includes iterative hard and soft thresholding with or without relaxation, as well as CoSaMP, subspace pursuit and some natural extensions. As a result, our optimally tuned algorithms dominate such proposals. Our notion of optimality is defined in terms of phase transitions, i.e. we maximize the number of nonzeros at which the algorithm can successfully operate. We show that the phase transition is a well-defined quantity with our suite of random underdetermined linear systems. Our tuning gives the highest transition possible within each class of algorithms.
This paper has not been read by Pith yet.
Forward citations
Cited by 1 Pith paper
-
Approximate Message Passing for Indoor THz Channel Estimation
Hard-thresholding AMP for indoor THz channel estimation outperforms prior methods and approaches oracle performance.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.