A Semantics for Approximate Program Transformations

Abstract: An approximate program transformation is a transformation that can change the semantics of a program within a specified empirical error bound. Such transformations have wide applications: they can decrease computation time, power consumption, and memory usage, and can, in some cases, allow implementations of incomputable operations. Correctness proofs of approximate program transformations are by definition quantitative. Unfortunately, unlike with standard program transformations, there is as of yet no modular way to prove correctness of an approximate transformation itself. Error bounds must be proved for each transformed program individually, and must be re-proved each time a program is modified or a different set of approximations are applied. In this paper, we give a semantics that enables quantitative reasoning about a large class of approximate program transformations in a local, composable way. Our semantics is based on a notion of distance between programs that defines what it means for an approximate transformation to be correct up to an error bound. The key insight is that distances between programs cannot in general be formulated in terms of metric spaces and real numbers. Instead, our semantics admits natural notions of distance for each type construct; for example, numbers are used as distances for numerical data, functions are used as distances for functional data, an polymorphic lambda-terms are used as distances for polymorphic data. We then show how our semantics applies to two example approximations: replacing reals with floating-point numbers, and loop perforation.