We use quadratic penalty functions along with some recent ideas from linear l1 estimation to arrive at a new characterization of primal optimal solutions in linear programs. The algorithmic implications of this analysis ...