The Ordered Weighted $\ell_1$ Norm: Atomic Formulation, Dual Norm, and Projections

The ordered weighted $\ell_1$ norm (OWL) was recently proposed, with two different motivations: because of its good statistical properties as a sparsity promoting regularizer, and as generalization of the so-called {\it octagonal shrinkage and clustering algorithm for regression} (OSCAR). The OSCAR is a convex group-sparsity inducing regularizer, which does not require the prior specification of the group structure. Also recently, much interest has been raised by the atomic norm formulation of several regularizers, not only because it provides an new avenue for their theoretical characterization, but also because it is particularly well suited to a type of method known as {\it conditional gradient} (CG), or Frank-Wolfe, algorithm. In this paper, we derive the atomic formulation of the OWL and exploit this formulation to show how Tikhonov regularization schemes can be handled using state-of-the-art proximal splitting algorithms, while Ivanov regularization can be efficiently implemented via the Frank-Wolfe algorithm.