Posts tagged by Max New
Recently, “final encodings” and “finally tagless style” have become popular techniques for defining embedded languages in functional languages. In a recent discussion in the Northeastern PRL lab, Michael Ballantyne, Ryan Culpepper and I asked “in what category are these actually final objects”? As it turns out our very own Mitch Wand wrote one of the first papers to make exactly this idea precise, so I read it available here and was pleasantly surprised to see that the definition of a final algebra there is essentially equivalent to the definition of observational equivalence.
In this post, I’ll go over some of the results of that paper and explain the connection to observational equivalence. In the process we’ll learn a bit about categorical logic, and I’ll reformulate some of the category theory in that paper to be a bit more modern in presentation, cleaning some things up in the process.
The continuation-passing style transform (cps) and closure conversion (cc) are two techniques widely employed by compilers for functional languages, and have been studied extensively in the compiler correctness literature. Interestingly, typed versions of each can be proven to be equivalence preserving using polymorphic types and parametric reasoning, as shown by my advisor Amal Ahmed and Matthias Blume (cps,cc).
In fact, there is something like a duality between the two proofs, cps uses a universal type, closure-conversion uses an existential type and the isomorphism proofs use analogous reasoning. It turns out that both are instances of general theorems in category theory: the polymorphic cps isomorphism can be proven using the Yoneda lemma, and the polymorphic closure-conversion isomorphism can be proven using a less well known theorem often called the *co*Yoneda lemma.
The connection between cps and the Yoneda embedding/lemma is detailed elsewhere in the literature and blogosphere (ncafe, Bartosz), so I’ll focus on closure conversion here. Also, I’ll try to go into some detail in showing how the “usual” version of Yoneda/coYoneda (using the category of sets) relates to the appropriate version for compilers.
I’ll assume some background knowledge on closure conversion and parametricity below. Fortunately, Matt Might has a nice blog post explaining untyped closure conversion.
One of my favorite papers at ICFP 2016 (in lovely Nara, Japan) was Constructive Galois Connections: Taming the Galois Connection Framework for Mechanized Metatheory by David Darais and David Van Horn. The central technical result is quite interesting, but a little intimidating, so I’d like to share a “de-generalization” of the result that I found helpful to understand.