RT @guillefix: is there any consensus about whether convex relaxation or projected gradient decent is better for constrained optimisation?