acatalepsie/content/posts/overloading-lambda.md

title	date	draft
Overloading the lambda abstraction in Haskell	2022-12-10	false

I've been dreaming about overloading the lambda abstraction in Haskell for a few years already. I am working on an EDSL with its very own notion of morphisms k a b from a to b. Thus I'm looking for a syntax that enables users to write said morphisms using just regular Haskell functions and the \x -> ... notation.

Until 5 days ago I thought it was impossible.

• Haskell's Arrow abstraction and associated syntactic sugar are supposed to do just that, but it just doesn't work as intended. It's a difficult thing to implement properly and if proposals to fix it have been come up, nothing concrete came out of it (yet).

  But even if the notation worked properly (and was convenient to use), many times your internal representation of morphisms k a b cannot possibly embed all Haskell functions a -> b. Therefore it's impossible to declare an instance of Arrow k.

• Now there is indeed this wonderful paper from Conal Elliott: "Compiling to Categories".
  In it, he recalls that any function in the simply typed lambda calculus corresponds to an arrow in any cartesian closed category (CCC). What he demonstrated with the concat GHC plugin is that this well-known result can be used to lift any monomorphic Haskell function into a Haskell arrow of any user-defined Haskell CCC.

  This is very cool, but also an experimental research project. It's very unstable and unreliable. It's not packaged on Hackage so fairly difficult to install as a library, and looks hard to maintain precisely because it's a compiler plugin.

  So probably not a sound dependency to carry around when trying to develop a little, self-contained, portable library yourself.

  And again, not every category is cartesian-closed, especially if exponentials have to be plain Haskell functions --- which concat enforces.

However --- I'm happy to report that "overloading" the lambda abstraction is completely doable. Not only that: it is actually very easy and doesn't require any advanced Haskell feature, nor any kind of metaprogramming. No. library. needed. It really is the perfect solution.

I cannot emphasize enough how powerful the approach presented here looks to me:

• You can define your very own proc notation for the category of your choice.

• Your category k a b only has to define operations (&&&), (.), and id. In particular, you can define this notation even for things that cannot provide an Arrow instance because of the arr function. You don't have to be able to embed pure Haskell functions in your category.

• And finally, this syntax is way more usable than the proc notation, because you can manipulate morphims k a b of your category as plain old Haskell functions, and can therefore compose them and apply them to variables just as any other regular Haskell function, with the same syntax.

It's in fact so simple that I suspect it must already be documented somewhere, but for the life of me I couldn't find anything. So here you go.

A toy DSL for flow diagrams

So let's say our EDSL is supposed to encode flow diagrams, with boxes and wires. Boxes have distinguished inputs and outputs, and wires flow from outputs of boxes to inputs of other boxes.

We can encode any flow diagram using the following operations:

{-# LANGUAGE GADTs #-}

data Flow a b where
  -- just a wire
  Id :: Flow a a
  -- putting two boxes one after the other
  Seq :: Flow a b -> Flow b c -> Flow a c
  -- putting two boxes one next to the other
  Par :: Flow a b -> Flow c d -> Flow (a, c) (b, d)
  -- box that duplicates its input
  Dup :: Flow a (a, a)
  -- box that gets rid of its input
  Void :: Flow a ()
  -- box that projects on first input
  Fst :: Flow (a, b) a
  -- box that projects on second input
  Snd :: Flow (a, b) b
  -- finally, we embed any pure function into a box
  Embed :: (a -> b) -> Flow a b

I think there are enough constructors here to claim it a cartesian category, but I didn't double check and it doesn't really matter.

Now I think we can agree that although this is the right abstraction to internally transform and reason about diagrams, what an awful, awful way to write them. Even if we provided a bunch of weird infix operators for some of the constructors, only a few Haskellers would be able to make sense of this gibberish.

No, what matters is that we can give an instance for both Category Flow and Arrow Flow:

import Control.Category
import Control.Arrow

instance Category Flow where id = Id ; g . f = Seq f g

instance Arrow Flow where
  arr     = Embed
  first f = Par f id
  second  = Par id
  (***)   = Par
  f &&& g = Seq Dup (Par f g)

We can even give a custom implementation for every Arrow operation. And yet, disappointment ensues when we attempt to use the proc notation: Haskell's desugarer for the Arrow syntactic sugar is really dumb.

{-# LANGUAGE Arrows #-}

t :: Flow (a, b) (b, a)
t = proc (x, y) -> returnA -< (y, x)

If you print the term, you get something like this:

Seq (Embed(_)) (Seq (Embed(_)) (Embed(_)))

We just wrote swap. But Haskell doesn't even try to use the operations we've just defined in Arrow Flow, and just lifts pure Haskell functions using arr. The terms you get are not inspectable, they are black boxes. Not. good.

What I'm gonna show you is how to get the following, amazing syntax:

t :: Flow (a, b) (b, a)
t = flow \(Tup x y) -> Tup y x

That reduces to the following term:

t = Seq Dup (Par (Seq Id Snd) (Seq Id Fst))

How beautiful! Sure, it's not the most optimal representation, and a traversal over the term could simplify it by removing Seq Id, but at the very least it's inspectable.

The solution

Now my solution stems from this truly amazing paper from Jean-Philippe Bernardy and Arnaud Spiwack: Evaluating Linear Functions to Symmetric Monoidal Categories (SMCs).

In this paper, they explain that if CCCs are models of the simply typed lambda calculus, it is "well-known" that SMCs are models of the linear simply-typed lambda calculus. And thus, they explain how they are able to evaluate linear Haskell functions (as in, Linear Haskell linear functions) into arrows in any target SMC.

They even released a library for that: linear-smc. I implore you to go and take a look at both the paper and the library, it's very very smart. Sadly, it seems to have gone mostly unnoticed.

I found out about this paper 5 days ago. Because my target category is actually cartesian (ergo, I can duplicate values), I suspected that I could remove almost all the complex machinery they had to employ to go from a free cartesian category to a monoidal one. I was hopeful that I could do without Linear Haskell, too.

And, relieved as I am, I can tell you that yes: not only can all of this be simplified (if you don't care about SMCs), but everything can be implemented in a handful lines of Haskell code.

Interface

So, the first thing we're gonna look at is the interface exposed to the users of your EDSL. These are the magic primitives that we will have to implement later.

type Port r a

First there is this (abstract) type Port r a. It is meant to represent the output of a box (in our flow diagrams) carrying information of type a.

Because the definition of Port r a is not exported, there is crucially no way for the user to retrieve a value of type a from it. Therefore, to use this variable in any meaningful way, they can only use the operations on ports that you --- the library designer --- export.

And now, the two other necessary primitives:

encode :: Flow a b -> Port r a -> Port r b
decode :: (forall r. Port r a -> Port r b) -> Flow a b

• encode transforms a morphism from your category --- here a diagram Flow a b --- into a Haskell functions on ports. By exporting this function, you enable users to apply any morphism in your category to the "port variables" they have at their disposal. And precisely only those operations.

• decode does the reverse translation. It takes any Haskell function on ports, and converts it back into a morphism in your category. Well, not any function. Only functions that are closed with respect to port variables. That's why there is this type variable r in Port r a. Because all the operations on ports exported by this interface share the same r between all inputs and outputs, there is no way to use a port variable defined outside of a function over ports, and still have the function be quantified over this r.

  The same kind of trick, I believe, is used in Control.Monad.ST

  This is really really neat.

And finally, because inputs of your diagrams are nested tuples, it's important to provide ways to combine ports to construct new ports carrying tuples.

pair :: Port r a -> Port r b -> Port r (a, b)
unit :: Port r ()

And... that's all you need! Really? Really!

---

Now of course you're free to include in your API your own morphisms converted into functions over ports. This would allow you to fully hide from sight your internal representation of diagrams. And to do so, you only need to use the previous primitives:

flow = decode

fst :: Port r (a, b) -> Port r a
fst = encode Fst

snd :: Port r (a, b) -> Port r b
snd = encode Snd

split :: Port r (a, b) -> (Port r a, Port r b)
split p = (fst p, snd p)

pattern Tup x y <- (split -> (x, y))
  where Tup x y = pair x y

void :: Port r a -> Port r ()
void = encode Void

box :: (a -> b) -> Port r a -> Port r b
box f = encode (Embed f)

(>>) :: Port r a -> Port r b -> Port r b
x >> y = snd (pair x y)

You can see we use PatternSynonyms and ViewPatterns to define this Tup syntax that appeared earlier. Because we defined this (>>) operator to discard its first port, we can use QualifiedDo to have an even nicer syntax for sequencing ports:

box1 :: Port r Text -> Port r (Int, Int)
box2 :: Port r Text -> Port r Bool
box3 :: Port r Int -> Port r ()
box4 :: Port r Int -> Port r ()

diag :: Flow (Text, Text) Bool
diag = flow \(Tup x y) -> Flow.do
  let Tup a b = box1 x
  box3 a
  box4 b
  box2 y

Because we defined (>>) in such a way, box1, box3 and box4 will appear in the resulting diagram even though their output gets discarded. Were we to define x >> y = y, they would simply disappear altogether. This technique gives a lot of control over how the translation should be specified.

Finally and most importantly, box is the only way to introduce pure Haskell functions a -> b into a black box in our diagram. For this reason, we can be sure that Embed will never show up if box isn't used explicitely by the user.

Implementation

And now, the reveal: Ports r a are just morphisms from r to a in your category! This is the implementation of the primitives from earlier.

newtype Port r a = P { unPort:: Flow r a }

encode :: Flow a b -> Port r a -> Port r b
encode f (P x) = P (f . x)

decode :: (forall r. Port r a -> Port r b) -> Flow a b
decode f = unPort (f (P id))

pair :: Port r a -> Port r b -> Port r (a, b)
pair (P x) (P y) = P (x &&& y)

unit :: Port r ()
unit = P Void

And indeed, if you consider Port r a to represent paths from some input r to output a in a diagram, then by squinting your eyes a bit the meaning (in your category) of this interface makes sense.

• For encode, surely if you now how to go from a to b, and from r to a, then you know how to go from r to b: that's just composition, hence (.).

• For decode, you have as assumption that from any point r in a diagram, you know how to transform a path from r to a into a path from r to b. In particular, because this holds for any r, this holds for a. And you know a path from a to a, it's the identity! This gets you a path from a to b.

It would appear encode and decode are just another instance of the Yoneda lemma, but I'm not a categorician so I cannot give any further explanation, my apologies.

Still, the sheer simplicity of this technique is mesmerizing to me.

---

Some things to be wary of

Ok, so now we've seen how to successfully "overload" the lambda abstraction. But there is one quirk that I think you should be aware of.

Preventing (too much) duplication in the diagram

If you look at the definition of split p, you can see that the input p gets duplicated in the two output ports. If both of them are used in a given diagram, p will appear at least twice in this diagram. Now in the abstract setting of category theory, duplication in your diagram doesn't matter because of the equivalences and laws that equate diagrams.

However, we're trying to encode useful things. My original motivation for going to this deep end was modeling flow diagrams of effectful computations, for the new version of achille. You could imagine changing the Embed constructor of Flow to the following:

Embed :: (a -> IO b) -> Flow a b

From then on, encoded flow diagrams are meant to be executed --- and even parallelised for free, how neat --- but I hope you see now why we should never duplicate any such embed box anymore. In the paper, that's why they argue it's important to restrict yourself to monoidal categories and linear functions, and make copy an explicit operation in your syntax.

copy :: Port r a %1 -> (Port r a, Port r a)

Indeed, the type system will prevent you to use a variable more than once without explicitly inserting copy. However, everything has to be made linear and this has huge implications on the user-friendliness of the overall syntax. I've also found error messages from Linear Haskell to be pretty uninformative --- and let bindings are not linear, what??.

Instead, I came up with a solution I'm fairly happy with. There is a trick: we introduce a new primitive.

(>>=) :: Port r a -> (Port r a -> Port r b) -> Port r b

Now of course, you could just implement it as x >>= f = f x, but that's precisely what we want to avoid. No, instead, I force the evaluation of x, and then evaluate f applied to a box that simply returns the pre-computed value, side-effect free. So a box gets duplicated, sure, but this box does nothing. What I am now noticing while writing this down, is that introducing this Bind operator makes the diagram non-inspectable again, oops. I can still run it of course, but this is a problem for future me.

Anyway, the benefit of this (>>=) operator is that it's very convenient with QualifiedDo:

diag :: Flow FilePath Bool
diag = flow \src -> Flow.do
  Tup a b <- box1 src
  box3 a
  box4 (box2 b)

You just have to stop using let and you're good to go. What's remarkable in the syntax is that these effectful boxes are used as plain old functions. Therefore there is no need to clutter the syntax with <$>, >>=, pure, you just compose and apply regular functions (that do side effects).

Compile vs. Evaluate

Now the paper is called "Evaluating [...]" and says there is a runtime cost to pay for the translation. Considering how simple this implementation is, and how the target is fully first-order (well, without (>>=)), I wouldn't be surprised if GHC is actually able to fully unfold and translate to category morphisms at compile-time. But I'm not an expert Haskell developer by any means and have no clue how one would go about checking this, so if anyone does, please tell me!

Expressive power

I'm very curious to see which kind of categories you can "compile" Haskell functions to using this technique. I didn't expect converting to cartesian category's morphisms to be so straightforward, and I'm looking forward to see whether there is any fundamental limitation preventing us to do the same for CCCs.

---

Thank you for reading! I hope this was or will be useful to some of you. I am also very grateful for Bernardy and Spiwack's fascinating paper and library, it quite literally made my week.

Till next time, perhaps when I manage to release the next version of achille, using this very trick that I've been searching for in the past 2 years.