Lately I’ve been auditing a course in program analysis and synthesis and the material presented in it keeps reminding me of category theory. This is my attempt to describe a categorical intuition for static analysis (or a static analysis co-intuition of category theory).
Disclaimer: Since my goal is developing intuition, I won’t attempt any rigorous definitions or proofs; nor will I dwell too much on expository material.
Static analysis tries to answer questions about a program’s behavior without actually executing it, by reasoning about program semantics. Since computing the exact concrete semantics of a program is undecidable, abstract interpretation uses sound over-approximations of semantics.
Two important aspects of the abstract interpretation scheme are how to abstract program properties and transformers (selecting abstract domains); and how to abstract program semantics (moving between abstract domains). We can think of those as what and how abstractions, respectively.
Abstract domains and order-preserving elementary program transformers are modeled with complete semi-lattices. Semantic abstractions themselves are composed using monotone functions.
We can interpret (ha!) the lattice of abstract domains as a category, where semantic states, or configurations, are objects and an arrow exists between two configurations if there’s an abstract transformer from to . In this category there’s correspondence between lattice properties and some categorical limits:
- Top/Bottom are Terminal/Initial objects, respectively.
- Meets/joins are products/coproducts, respectively.
Now we notice that the monotone functions between semantic representations are functors between abstract domain categories. Thus, abstractions are functors. Since we don’t want to abstract indefinitely – eventually we would like to come back to the “real world” – we have concretizations, or functors going in the opposite direction. And this is where we get to the crux of the matter: Abstraction and concretization are adjoint functors.
Why is this cool?
For one, this shows why adjunctions “generalized” inverses: because abstraction followed by concretization gets us, approximately, back in the same place.
In addition, when the domain selection enables it, we can use Galois connections to compute best abstract transformers. Adjunctions are generalize Galois connections, so this is an example of how adjunctions capture the idea of “optimal solutions to a problem”.
Finally, this new setting gives rise to some interesting questions:
- Adjoint functors induce an “abstract interpretation monad”. Can we can “lift” programs to this monadic representation to easily utilize abstract interpretation?
- Semantic abstractions themselves form a lattice, so can we generalize further and get a higher order category?
- Axiomatic predicates of strongest post-condition and weakest pre-condition are also adjoint functors, so this may hint to a nice correspondence between operational and axiomatic semantics (and likewise for denotational semantics).
- What are other categorical limits/universal properties (pullbacks, equalizers etc.) in those categories?
Trying to understand static analysis in categorical terms actually helped me gain better intuition about category theory, since I’m not versed enough in algebraic geometry to relate to “common” intuitions. I was surprised to find very little research in this direction and I’m curious how these category theory insights can be put to practical use.