Towards a Mathematical Theory of Super-Resolution