Fundamental Question: Which of the options give useful surrogate for a non-convex function, the tightest convex lower or the tightest upper bound? (By useful, I mean with theoretical guarantees.) I know that minimizing a lower bound, gives a lower bound on the minimum of the original objective. ...