Pseudo-Riemannian optimization for Boolean satisfiability

We generalize the convexity inequality “function above tangent” to a class of functions that are convex in a pseudo-Riemannian sense. We propose a new descent algorithm for such functions and provide a regret analysis. Finally, we explain how to apply that method to design a 3-SAT solver based on continuous relaxation.