Functors

CIS 194 Week 9
18 March 2013

Motivation

Over the past weeks we have seen a number of functions designed to “map” a function over every element of some sort of container. For example:

map :: (a -> b) -> [a] -> [b]
treeMap :: (a -> b) -> Tree a -> Tree b
In Homework 5 many people ended up doing a similar thing when you had to somehow apply eval :: ExprT -> Int to a Maybe ExprT in order to get a Maybe Int.

maybeEval :: (ExprT -> Int) -> Maybe ExprT -> Maybe Int

maybeMap :: (a -> b) -> Maybe a -> Maybe b

There’s a repeated pattern here, and as good Haskell programmers we want to know how to generalize it! So which parts are the same from example to example, and which parts are different?

The part that is different, of course, is the container being “mapped over”:

thingMap :: (a -> b) -> f a -> f b

But what sort of things are these “containers”? Can we really assign a type variable like f to them?

A brief digression on kinds

Just as every expression has a type, types themselves have “types”, called kinds. (Before you ask: no, there’s not another level beyond kinds—not in Haskell at least.) In ghci we can ask about the kinds of types using :kind. For example, let’s ask for the kind of Int:

Prelude> :k Int
Int :: *

We see that Int has kind *. In fact, every type which can actually serve as the type of some values has kind *.

Prelude> :k Bool
Bool :: *
Prelude> :k Char
Char :: *
Prelude> :k Maybe Int
Maybe Int :: *

If Maybe Int has kind *, then what about Maybe? Notice that there are no values of type Maybe. There are values of type Maybe Int, and of type Maybe Bool, but not of type Maybe. But Maybe is certainly a valid type-like-thing. So what is it? What kind does it have? Let’s ask ghci:

Prelude> :k Maybe
Maybe :: * -> *

ghci tells us that Maybe has kind * -> *. Maybe is, in a sense, a function on types — we usually call it a type constructor. Maybe takes as input types of kind *, and produces another type of kind *. For example, it can take as input Int :: * and produce the new type Maybe Int :: *.

Are there other type constructors with kind * -> *? Sure. For example, Tree, or the list type constructor, written [].

Prelude> :k []
[] :: * -> *
Prelude :k [] Int
[] Int :: *
Prelude> :k [Int]  -- special syntax for [] Int
[Int] :: *
Prelude> :k Tree
Tree :: * -> *

What about type constructors with other kinds? How about JoinList from Homework 7?

data JoinList m a = Empty
                  | Single m a
                  | Append m (JoinList m a) (JoinList m a)

Prelude> :k JoinList
JoinList :: * -> * -> *

This makes sense: JoinList expects two types as parameters and gives us back a new type. (Of course, it is curried, so we can also think of it as taking one type and giving back something of kind * -> *.) Here’s another one:

Prelude> :k (->)
(->) :: * -> * -> *

Tthis tells us that the function type constructor takes two type arguments. Like any operator, we use it infix:

Prelude> :k Int -> Char
Int -> Char :: *

But we don’t have to:

Prelude> :k (->) Int Char
(->) Int Char :: *

OK, what about this one?

data Funny f a = Funny a (f a)

Prelude> :k Funny
Funny :: (* -> *) -> * -> *

Funny takes two arguments, the first one a type of kind * -> *, and the second a type of kind *, and constructs a type. (How did GHCi know what the kind of Funny is? Well, it does kind inference just like it also does type inference.) Funny is a higher-order type constructor, in the same way that map is a higher-order function. Note that types can be partially applied, too, just like functions:

Prelude> :k Funny Maybe
Funny Maybe :: * -> *
Prelude> :k Funny Maybe Int
Funny Maybe Int :: *

Functor

The essence of the mapping pattern we saw was a higher-order function with a type like

thingMap :: (a -> b) -> f a -> f b

where f is a type variable standing in for some type of kind * -> *. So, can we write a function of this type once and for all?

thingMap :: (a -> b) -> f a -> f b
thingMap h fa = ???

Well, not really. There’s not much we can do if we don’t know what f is. thingMap has to work differently for each particular f. The solution is to make a type class, which is traditionally called Functor:

class Functor f where
  fmap :: (a -> b) -> f a -> f b

(Functor is defined in the standard Prelude. Note that the name “functor” comes from category theory, and is not the same thing as functors in C++ (which are essentially first-class functions).) Now we can just implement this class in a way specific to each particular f. Note that the Functor class abstracts over types of kind * -> *. So it would make no sense to write

instance Functor Int where
  fmap = ...

Indeed, if we try, we get a very nice kind mismatch error:

[1 of 1] Compiling Main             ( 09-functors.lhs, interpreted )

09-functors.lhs:145:19:
    Kind mis-match
    The first argument of `Functor' should have kind `* -> *',
    but `Int' has kind `*'
    In the instance declaration for `Functor Int'

If we understand kinds, this error tells us exactly what is wrong.

However, it does make sense (kind-wise) to make a Functor instance for, say, Maybe. Let’s do it. Following the types makes it almost trivial:

instance Functor Maybe where
  fmap _ Nothing  = Nothing
  fmap h (Just a) = Just (h a)

How about lists?

instance Functor [] where
  fmap _ []     = []
  fmap f (x:xs) = f x : fmap f xs
  -- or just
  -- fmap = map

Easy peasy. What about IO? Could it make sense to create an instance of Functor for IO?

Sure. fmap :: (a -> b) -> IO a -> IO b results in the IO action which first runs the IO a action, then applies the function to transform the result before returning it. We can implement this without too much trouble:

instance Functor IO where
  fmap f ioa = ioa >>= (\a -> return (f a))

or even

instance Functor IO where
  fmap f ioa = ioa >>= (return . f)

Now let’s try something a bit more mind-twisting:

instance Functor ((->) e) where

What!? Well, let’s follow the types: if f = (->) e then we want

fmap :: (a -> b) -> (->) e a -> (->) e b

or, with (->) written infix:

fmap :: (a -> b) -> (e -> a) -> (e -> b)

Hmm, this type signature seems familiar…

instance Functor ((->) e) where
  fmap = (.)

Crazy! What does this mean? Well, one way to think of a value of type (e -> a) is as a “e-indexed container” with one value of a for each value of e. To map a function over every value in such a container corresponds exactly to function composition: to pick an element out of the transformed container, we first we apply the (e -> a) function to pick out an a from the original container, and then apply the (a -> b) function to transform the element we picked.

Generated 2013-03-21 14:41:58.488748