Understanding Constructive Galois Connections

:: icfp, galois connection, adjunction, category theory, math

By: Max New

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.

History

I won’t go into much of the details of the paper, because I think it is quite well written, but here’s a short overview. The paper is about how to do verified static analysis while taking advantage of the calculational approach of Abstract Interpretation. The problem is that the Galois connections people use for abstract domains are not always computable. Darais and Van Horn show however that there is a very useful class of Galois connections that is computable, and they show how they can exploit this to write verified static analyses that more closely follow the “on-paper” proofs, and offload much of the details to the proof assistant as mere calculation.

David Darais told me about these results when we were at POPL 2016 (in less lovely but much more convenient St. Petersburg, Florida) and in particular about the central theorem of the paper, which shows that two different classes of Galois connections they define, “Kleisli” and “Constructive” Galois connections, are actually constructively equivalent. I was really surprised by the result when he explained it to me, and so I hoped to find if there was a known generalization of the result for adjunctions of categories, rather than Galois connections of posets.

Eventually, my usual trawling of Mathoverflow and nlab led me to a not-quite generalization to categories and interestingly a de-generalization to sets that helped me immensely to understand the theorem.

Since I know that the original theorem is a bit technical, I’ll explain the de-generalization to sets here, which I hope will help to understand their theorem.

Functions and Relations

Let’s start with the “Kleisli Arrows”, which are monotone functions \(f : A \to P(B) \) where \(A,B \) are posets and \(P(B)\) represents the poset of downward-closed subsets of \(B \).

Now to “de-posetize” this, we’ll take sets \(X,Y \) and let \(P(Y) \) mean the powerset of \(Y\), that is the set of all subsets of \(Y \). Then a function \(f : X \to P(Y) \) is actually exactly the same thing as a relation \(R \subset X \times Y \). From \(f : X \to P(Y) \) we can take \(R = \{(x,y) \in X\times Y | y\in f(x)\} \) and from \(R\) we can construct \(f(x) = \{y \in Y | (x,y) \in R \}\).

Furthermore, the “Kleisli composition” is the same as composition of relations. If \(R \subset X \times Y \) and \(Q \subset Y \times Z \), then the composition is defined as \[ (R;Q) = \{(x,z) \in X \times Z | \exists y\in Y. (x,y) \in R \land (y,z) \in Q\}\]

Then the next thing we need to understand is what is the de-generalization of “Kleisli Galois connection”? Well, Galois connections are an instance of what’s called an adjunction in category theory, which is usually formulated in terms of categories, functors and natural transformations. However, you can interpret the definition of adjunction in any “universe” that acts like the universe of categories, functors and natural transformations and it turns out we have such a universe. The universe I’m talking about is called \(\texttt{Rel}\), and it consists of sets, relations between sets and inclusion of relations, i.e. that one relation is a subset of another.

Then what does it mean to have an adjunction between two relations \(R \subset X \times Y, Q \subset Y \times X\)? Taking apart the definition it just means

\begin{align}\tag{1} \Delta(X) \subset R;Q \end{align} \begin{align}\tag{2} Q;R \subset \Delta(Y) \end{align}

where \(\Delta \) means the diagonal, or equality relation on the set:

\[\Delta(X) = \{(x_1,x_2) \in X | x_1 = x_2 \} \]

So we just need to unravel what (1) and (2) mean above. Unwinding (1), we get that for any \(x \in X\), there exists a \(y \in Y \) such that \((x,y) \in R \) and \((y,x) \in Q\). This tells us for one that \(R \) is a “right-total” relation and \(Q \) is a “left-total” relation. Every \(x \) is related to some \( y\) by \( R \) and \( Q\).

If we unwind (2), we get that for any \(y,y' \in Y\) if there’s some \(x \in X \) such that \((x,y) \in R \) and \((y',x) \in Q \) then actually \(y = y')\). This one is a bit more mysterious, but first, let’s see what this tells us about the relationship between \(R\) and \(Q \).

If \((x,y) \in R \), then by (1) there’s some \(y' \in Y\) so that \((x,y') \in R \) and \((y',x) \in Q\). Then, by (2) we know that \(y = y'\), so we’ve shown that if \((x,y) \in R \) then \((y,x) \in Q\). Then a completely symmetric argument shows that if \((y,x) \in Q \) then \((x,y)\in R\)! So we’ve discovered that actually \(Q \) is just the opposite relation of \(R \).

Then if we look at (2) again but replace the \(Q\)’s by flipped \(R\)’s we get that for any \(y,y' \in Y\), if there’s some \(x \in X\) such that \((x,y) \in R \) and \((x,y')\in R\) then \(y = y'\), which tells us that \(R \) is a partial function, i.e., that every \(x \) is related to at most one \(y \) by \(R \).

You may recognize it now, our \(R \subset X \times Y \) is just a function, and saying \(R, Q\) are adjoint is exactly the same as saying that \(Q = R^{\text{op}}\) and \(R \) is a function. Adjunctions are so pervasive you saw them back in pre-algebra!

Constructive Galois Connections

Back to constructive Galois connections, I hope if you read the paper you can see that their theorem is a generalization of the above argument, where instead of relations we have “monotone relations”, i.e., downward-closed \(R \subset A^{\text{op}} \times B \). Then you can interpret the definition of adjunction in that universe and get that it’s the same as a Kleisli Galois connection and that a similar argument to the above shows that the “left adjoint” is represented by a monotone function \(f : A \to B \):

\[R = \{(x,y) | y \le f(x) \} \]

Which shows that every Kleisli Galois connection is actually a constructive Galois connection! The details are in their paper, and I hope they are easier to follow now.

In fact, we get a little extra from what’s mentioned in their paper, which is that the “right adjoint” is represented by \(f \) as well but in the opposite way:

\[Q = \{(y,x) | f(x) \le y \}\]

Category Theory Post Scriptum

If you’re interested in Category theory, here’s a more technical addendum.

Remembering from Category Theory class, sets are just posets where objects are only less than themselves and posets are (basically) categories where there is at most 1 arrow between objects, so we might naturally ask, does this theorem extend to categories?

Well, first we need a generalization from relations to downward-closed relations to what are called distributors or profunctors. Then we can also generalize inclusion of relations to morphisms of distributors and ask, is every left adjoint distributor represented by a functor?

The answer is, at least in full generality, no! For it to be true we need a special property on the codomain of the left adjoint \(R : C \not\to D \), which is called (for mind-boggling reasons) Cauchy completeness. Viewing sets and posets as special categories, it turns out that they always have this property, and that’s why the theorem worked out for those adjunctions.