How does auto-differentiation apply to loop indices? Is this really any more computationally powerful than techniques used to solve integer linear programmes?