Proving My Point
Previous: Simple Linear Regression in One Pass
A coworker of mine has been kind enough to spend his scant spare time guiding a small group of future Haskellers through Hutton’s Programming in Haskell, assigning weekly problems and judging the solutions submitted along the way. What follows here today^{1} is my smartass solution to my colleague’s Problem 1.
Executing Programs on Paper
Week 1 of the reading group was to read Chapter 1 in Hutton and complete a few simple problems based on the reading. Chapter 1 is mostly an inspirational overview, but it does show how one can write simple, recursive functions on lists and how one can evaluate such functions on paper. For example, Hutton defines sum
as follows.
He then shows how one may evaluate sum [1, 2, 3]
on paper.
Week 1 Problem 1
Using the above example as a guide, said group leader assigned what would be the first problem for week 1.
Problem 1: Prove that for all x
, sum [x]
is x
.
This is a great problem, and I’m sure the intended solution was to evaluate the program on paper for a general value \(x\). By that, I mean we assume that \(x\) is some fixed value and we proceed to evaluate \(\texttt{sum}\,[x]\) on paper as though we knew what value \(x\) were.
\[\begin{align*} \texttt{sum}\,[x] &= \texttt{sum}\,(x : []) \\ &= x + \texttt{sum}\,[] \\ &= x + 0 \\ &= x \end{align*}\]Since our answer did not depend on any specific characteristics of \(x\), the reasoning is valid no matter what value \(x\) were, and so we know that this same conclusion can be drawn about any such value. That is, we have shown that for all x
, sum [x]
is x
. (Notice that the above evaluation could not have been carried out to conclusion had the formula for sum
directed us to pattern match \(x\). Also, we must assume a lawful Num
instance, as we tacitly use \(x + 0 = x\).)
Proofs that Compile
The above is certainly a fine proof, and—I stress again—surely the intended solution to Problem 1.
But I was bored. And I wanted to be extra, superduper sure of the veracity of the claim.
To this very end, I wrote a little program in Haskell. And as we know about Haskell, if it compiles, it works.
We’re going to prove Problem 1, but for typelevel lists, using classes as predicates on types. Our typelevel list is implemented using Cons
“cells” terminating with a Nil
. For example, the type level list consisting of Int
, Double
, and String
(in that order) is quite simply Cons Int (Cons Double (Cons String Nil))
. This type has no inhabitants, but that’s fine—we won’t be needing any.
We’ve added a class to represent the predicate The type x
is the sum of the type list
. That is, x
is the sum of list
if and only if the compiler can find an instance Sum list x
. The syntax  list > x
means that attempting to create two instances Sum list x1
and Sum list x2
is a compiler error. We want this, because the sum of list
(if a sum exists) should be uniquely and unambiguously determined by list
. The FunctionalDependencies
language extension enables this syntax and the associated check. The extension also aids type inference, since—in light of the restriction imposed by  list > x
—the type x
can now be inferred from the type list
.
The class method sum
is notable for both its signature and its implementation. Its signature Proxy list > Proxy x
tells us that sum
gives us a valuelevel way to find the type x
that is the sum of the type list
.
To see how, we need to talk about Proxy
. Proxy
is defined as data Proxy a = Proxy
. So, Proxy Int
has the one inhabitant Proxy :: Proxy Int
. And Proxy Bool
has the one inhabitant Proxy :: Proxy Bool
. And Proxy [IO String]
has the one inhabitant Proxy :: Proxy [IO String]
. In fact, no matter what type is used for the parameter, we have just one inhabitant, Proxy
, that carries no data. What gives?
Well, the interesting thing about Proxy
isn’t the value, it’s the type parameter. In code, the data constructor Proxy
is often given an explicit type annotation as a way to pass types as arguments into Haskell functions. This is exactly how we’re going to use it with our sum
function.
If we have some type list
in mind and we want to know its sum, all we have to do is ask the compiler to inferred type of sum (Proxy :: Proxy list)
. The compiler will infer a type Proxy x
if it can find an instance Sum list x
, from which we see that type x
is the sum of type list
, or else it will reject the program if it fails to find an instance. This is how we get the compiler to perform typelevel computations for us (such as summing lists) at compile time.
The implementation sum _ = Proxy
works because the compiler doesn’t need to inspect the argument (it has no data anyway) as it already knows the type of the argument, and from that it can infer the type of the result Proxy
value.
Moving on, we still have to create instances of class Sum
. To see this viscerally, let’s try to get the compiler to infer the type of sum (Proxy :: Proxy (Cons 1 (Cons 2 (Cons 3 Nil))))
.
We get a compiler error:
No instance for (Sum (Cons 1 (Cons 2 (Cons 3 Nil))) x0)
arising from a use of ‘sum’
Although we humans know the type of sum123
should be Proxy 6
, the compiler error is actually quite reasonable: we haven’t written any instances of class Sum
yet!
The instances are where we teach the compiler how to compute sums. In other words, instances are the actual programming in this flavor of typelevel programming.
Commenting out sum123
for suspense, we’ve ported to the type system the standard recursive definition of sum
that we encountered at the start of this post. We have two instances, a base case and a recursive case. Our base case instance Sum Nil 0
tells the compiler that the sum of an empty list is 0
. Our recursive case instance (Sum rest n2, n ~ (n1 + n2)) => Sum (Cons n1 rest) n
tells the compiler that the sum of Cons n1 rest
is n
, provided that the sum of rest
is n2
and that n
equals n1 + n2
.
Found type wildcard ‘_’ standing for ‘Proxy 6’
Uncommenting sum123
, we see the compiler now infers the correct type, namely Proxy 6
. That is, the sum of the type Cons 1 (Cons 2 (Cons 3 Nil))
is the type 6
.
Let’s fill the underscore with Proxy 6
in our code and press on to the main event.
And, would you look at that! problem1
says that sum
has the type forall (n :: Nat). Proxy (Cons n Nil) > Proxy n
, which is what we wanted to show. It compiles, so it works ;)
Pulling Back on the Reins
If you found this post confusing and maybe even a little scary, don’t worry. Programming in Haskell is really nothing like this. For that reason, I won’t go out of my way to share the above “solution” with the reading group. I did it simply because I, personally, found it to be an enjoyable puzzle on which to pass an hour.
If, on the other hand, you did enjoy this post, you might want to take a look at this lovely story called Typing the Technical Interview, by Aphyr, from which I learned this style of programming. And it’s a riot, to boot. You might also enjoy the excellent talk 4 Programming Paradigms in 45 Minutes by Aja Thagomizer Hammerly. Finally, we provide the above program (and some bonus features) conveniently as a single file.
Thanks for reading!

I know I owe you a retro on the linear regression post, including a clever observations made by a user on Twitter.com and a constantspace solution. (My solution uses memory proportional to the size of the input—yuck!) I haven’t forgotten. ↩
Previous: Simple Linear Regression in One Pass