Randomly aggregated least squares for support recovery

Abstract We study the problem of exact support recovery: given an (unknown) vector ? * ? { ? 1 , 0 } D with known sparsity k = ? we are access to noisy measurement y X + ? where R N × is a (known) Gaussian matrix and noise vector. How small can be for reliable recovery ? present RAWLS (Randomly Aggregated unWeighted Least Squares Support Recovery): main idea take random subsets equations, perform least squares over this reduced bit information, average many subsets. show that proposed procedure provably recover approximation demonstrate its use through numerical examples. simulations beneficial task recovery. Finally, observe at par several strong baselines in low information regime (i.e. or large).