Abstraction Functions as Types

Abstract: Modular development of programs relies on the principle that library code may be freely replaced without affecting client behavior. While an interface mediating this interaction should require a precise behavior of its implementations, allowing for downstream verification of client code, it should do so in a manner that allows private algorithmic and representation choices to vary freely. In this work we demonstrate how such modularity can be achieved in dependent type theory using a phase distinction between private algorithmic content and public client-facing behavior. We observe that modalities associated with such a phase distinction and their corresponding theorems, particularly noninterference and fracture, give rise to precise descriptions of common constructions surrounding algorithms and data structures. Using a modal construction to classify types that sufficiently restrict client-facing behavior, we use the noninterference property for the phase to state and prove a modularity property guaranteeing that implementations may be freely replaced without affecting behavior. We then cast the fracture property in the light of abstraction functions, showing internally that every type consists of a private algorithmic component, a public behavioral component representing an abstract data type, and an abstraction function between them that is uniformly activated by the behavioral phase for streamlined verification of client correctness. Finally, we use phased quotient types to ergonomically mark private data for behavioral deletion. We situate this development in a univalent adaptation of Calf, a dependent type theory for cost analysis, in order to amplify these points: beyond hiding private implementation details, we treat cost as a private matter that may be varied freely without affecting the behavior of clients.