Functional Patterns: Interfaces and Functors

This is part 3 of a series of articles entitled Functional Patterns.

Make sure to check out the rest of the articles!

1. The Monoid
2. Compositions and Implicitness

Generics and Typeclasses

To be correct, a function must type-check, and is therefore provable. But in the case of generalized functions, meant to deal with various types, this immediately shows as a pain point. To make a double function work across types, we would have to define them separately!

doubleInt :: Int -> Int
doubleChar :: Char -> Char
doubleFloat :: Float -> Float
-- ...

And for any self-respecting programmer, you should already be finding yourself absolutely appalled by this. We'd just learned about a pattern for building case-handling using partial application but we can't really apply it here since our type signatures won't allow that, and our function has to type-check.

Thankfully, this is already a feature in most modern programming languages. We are allowed to define a generic type. A hypothetical type that only has to verify matching positions in the function signature or variable declarations.

// c++
template <typename T>
T double(T x) {
    return x*2;
}

// rust
fn double<T>(x: T) -> T {
    return x*2;
}

-- haskell
double :: a -> a
double = (*2)       -- partially applied multiplication

And that should solve our problem! As long as the compiler is given these generics, it can figure out what types it has to use at run-time (Rust actually still does this inference at compile-time!).

However, even though there is merit in this implementation— there is still a glaring flaw, that actually gets pointed out by the Haskell compiler, as the above Haskell code actually raises an error.

No instance for ‘Num a’ arising from a use of ‘*’...

We've defined a type, but we aren't always going to be sure this type has the capacity to be doubled. Sure, this immediately works on numbers, but what's stopping the user from calling double on a String? A list? Without a predefined method for doubling these types, they should not be allowed as arguments, in the first place.

So contrary to the name of generics, we're going to have to get a bit more specific, but still general.

This is where typeclasses come in, or also known more commonly in the imperative world as interfaces. Again, if you're using any language that has been made later than C++, you should have access to some implementation of interfaces.

Interfaces, compared to generics, specify some sort of capability of types that can be categorized under it.

Here is a fixed version of our previous code.

double :: (Num a) => a -> a     -- a has to be of typeclass Num
double = (*2)

or in Go:

// We first create an interface that is the union of floats and integers.
type Num interface {
    ~int | ~float64
    // ... plus all other num types
}

func double[T Num](a T) T {
    return a * 2
}

For brevity's sake we'll say that Haskell doesn't really deal with embedded state in their interfaces, such as Typescript's and Go's interfaces (a constraint brought upon by pure functional rules). So even though you might be able to define required attributes of a type to be under an interface, know that pure interfaces only have to define functions or capabilities of the type.

And by capabilities in this context, we are talking about if the type has a dependency in the form of a doubling function— is the compiler taught how to double it?

import Control.Monad (join)

class CanDouble a where
  double :: a -> a

instance CanDouble Int where
  double = (* 2)

instance CanDouble Float where
  double = (* 2)

-- we tell the compiler that doubling a string is concatenating it to itself.
instance CanDouble String where 
  double = join (++)    -- W-combinator, f x = f(x)(x)

And now we're pretty much back to where we were at the start when it comes to code repetition, isn't that funny?

But this fine-grained control of implementation is actually where the power of this comes in. If you've ever heard of the Strategy pattern before, this is pretty much it, in the functional sense.

quadruple :: (CanDouble a) => a -> a
quadruple = double . double

leftShift :: (CanDouble a) => Int -> a -> a
leftShift n e
  | e <= 0 = n
  | otherwise = leftShift (double n) $ e - 1

These functions type-check now, all because we taught the compiler how double types under the CanDouble typeclass.

We can achieve something similar in Go, a big caveat being that we can only define interface methods on non-primitive types. Meaning, we have to define wrapper structs to primitive types.

type CanDouble interface {
    double() CanDouble
}

type String string
type Number interface {
    ~int | ~float64
    // ... plus all other num types
}

type Num[T Number] struct {
    v T
}

func (s String) double() String {
    return s + s
}

func (n Num[T]) double() Num[T] {
    return Num[T]{n.v * 2}
}

func quadruple(n CanDouble) CanDouble {
    return n.double().double()
}

func leftShift(n CanDouble, e uint) CanDouble {
    for i := uint(0); i < e; i++ {
        n = n.double()
    }

    return n
}

This honestly is kind of a bummer, but no worries, as most of the time you're going to be dealing with interfaces will be with custom types and structs.

Categories

Category theory is a general theory of mathematical structures and their relations.

We've briefly brushed upon category theory back in The Monoid, and we'd like to keep it that way, only close encounters. I will be referencing it here and there, but rest assured: you won't need to have a background in it to grasp whatever follows.

However, there is no doubt that we have encountered sets before.

As a brief recap, Sets can be thought of as a collection of elements. These elements can be absolutely anything.

{ 0, 1, 2, 3, ... }             -- the set of natural numbers
{ a, b, c, ..., z}              -- the set of lowercase letters
{ abs, min, max, ... }          -- the set of `Math` functions in Javascript
{ {0, 1}, {a, b}, {abs, min} }  -- the set of sets containing the first 2 elements of the above sets

Adding on to that, we have these things called morphisms, which we can think of a mapping between elements.

Very big omission here on the definitions of morphisms, in that they are relations between elements, and not strictly functions/mappings,
you can look it up if you are curious.

We can say a function like toUpper() is a morphism between lowercase letters to uppercase letters, just like how we can say double = (*2) is a morphism from numbers to numbers (specifically even numbers).

And if we group these together, the set of elements and their morphisms, we end up with a category.

Again, omission, categories have more constraints such as a Composition partial morphism and identities. But these properties are not that relevant here.

If you have a keen eye for patterns you'd see that there is a parallel to be drawn between categories and our interfaces! The objects (formal name for a category's set of elements) of our category are our instances, and our implementations are our morphisms!

class CanDouble a where
    double :: a -> a

-- `Int` is our set of elements { ... -1, 0, 1, ... }
-- `(* 2)` is a morphism we defined
-- ... (other omissions)
-- ...
-- Therefore, `CanDouble Int` is a Category.
instance CanDouble Int where
    double = (* 2)

Functors

Man, that was a lot to take in. Here's a little bit more extra:

A Functor is a type of a function (also known as a mapping) from category to another category (which can include itself, these are called endofunctors).

What this essentially means, is that it is a transformation on some category that maps every element to a corresponding element, and every morphism to a corresponding morphism. An output category based on the input category.

In Haskell, categories that can be transformed by a functor is described by the following typeclass (which also makes it a category in of itself, that's for you to ponder):

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

f here is what we call a type constructor. By itself it isn't a concrete type, until it is accompanied by a concrete type. An example of this would be how an array isn't a type, but an array of Int is. The most common form of a type constructor is as a data type (a struct).

From this definition we can surmise that all we need to give to this function fmap is a function (a -> b) (which is our actual functor, don't think about the naming too much), and this would transform a type f a to type f b, a different type in the same category.

Yes, this means Haskell's Functor typeclass is actually a definition for endofunctors, woops!

If all of that word vomit was scary, a very oversimplified version for the requirement of the Functor typeclass is that you are able to map values to other values in the same category.

Arguably the most common Functor we use are arrays:

instance Functor [] where
--  fmap f [] = []
--  fmap f (a:as) = f a : fmap as

    -- simplified
    fmap :: (a -> b) -> [a] -> [b]
    fmap f arr = map f arr

We are able to map an array of [a] to [b] using our function (or functor) f. The typeconstructor of [] serves as our category, and so our functor is a transformation from one type of an array to another.

So, formally: the map function, though commonly encountered nowadays in other languages and declarative frameworks such as React, is simply the application of an endofunctor on the category of arrays.

Wow. That is certainly a description.

Here are more examples of functors in action:

// Go
type Functor[T any] interface {
    fmap(func(T) T) Functor[T]
}

type Pair[T any] struct {
    a T
    b T
}

type List[T any] struct {
    get []T
}

// Applying a functor to a Pair is applying the function
// to both elements
func (p *Pair[T]) fmap(f func(T) T) Pair[T] {
    return Pair[T]{     // apply f to both a and b
        f(p.a),
        f(p.b),
    }
}

func (a *List[T]) fmap(f func(T) T) List[T] {
    res := make([]T, len(a.get))    // create an array of size len(a.get)

    for i, v := range a.get {
        res[i] = f(v)
    }

    return List[T]{res}
}

-- haskell
data Pair t = P (t, t)

instance Functor Pair where
    fmap f (P (x, y)) = P (f x, f y)

So all that it takes to fall under the Functor (again, endofunctor), interface is to have a definition on how to map the contents of the struct to any other type (including its own).

This is another simplifcation, functors also need to have property of identity and composition.

To put simply, whenever you do a map, you're not only transforming the elements of your array (or struct), you're also transforming the functions you are able to apply on this array (or struct). This is what we mean by mapping both objects and morphisms to different matching objects and morphisms in the same category.

This is important to note as even though we end up in the same category (in this context, we map an array, which results in another array), these might have differing functions or implementations available to them (though most of them will be mapped to their relatively equivalent functions, such as a reverse on an array of Int to reverse on an array of Float).

This is where the oversimplifcation kinda messes us up a bit, because if we follow only our definition, we could say that reducing functions such as sum and concat are functors from the category of arrays to atoms, but this isn't necessarily true. As functors also require that you preserve the categorical structure, which won't be covered in this article series as that's way too deeply rooted in category theory.

Apologies if this article contained way more math than applications, but understanding these definitions will help us greatly in understanding the harder patterns later in this series, namely Applicatives and finally Monads.

A monad is a monoid in the category of endofunctors.

We're getting there! :>