If at first you don't succeed...

About | Archive

Recursivized Types

So, a short break from \(\varepsilon—\delta\).

Anyway, I’m going to look at Algebraic types in this section. Specifically, I am going to use Haskell types.

The provided code is all actual Haskell, in fact, I’m going to have to de-import the standard libraries because some of this is actually defined there. You can find all the code here

Algebraic types

I’ll start with a quick go-over of Haskell types. If you already know Haskell, feel free to skip this section.

Enumeration types

Some types are enumerations of potential values.

A quick example is the trivial singleton type:

data Singleton = Single
    deriving Show

This is to be read as “there exists a type Singleton, which contains one element named Single.” The “deriving show” part is Haskell book keeping for those following along on their interpreters to be able to see things

Another example is

data Void

This is to be read as “there exists a type Void with no possible elements”

Yet another type is

data Bool = False | True
    deriving Show

This is to be read as “there exists a type Bool with the possible elements True and False.”

Structure types

Types can be used as as structures, as such:

data Wrapper a = Wrap a
    deriving Show

This is to be read as “for every type a there exists a type (Wrapper a) with a single type constructor Wrap that takes an a

A more complicated structure is

data Pair a b = ConsPair a b
    deriving Show

Which is similar to the previous one except that it contains two possible elements which can be of two different types.

Aside: Kinds

Algebraic types have kinds.

  1. Singleton, Void, and Bool are concrete types in their own right.
  2. Wrapper needs a type before it can become a type (nothing can be a Wrapper by itself).
  3. Pair needs two types before it can become a type (nothing can be a Pair or Pair Bool by itself).

We refer to the latter two cases as “higher-kinded types.”

We say that Singleton, Void, etc. have kind *. Types that need one type added on the end to be complete, such as Wrapper or Pair Bool have kind * -> *. Types such as Pair, which needs two types to be complete, have kind * -> * -> *.

Note also that types such as

data TwoOfSame a = ConsTwoOfSame a a

have kind * -> * rather than * -> * -> * because they only take one type argument. TwoOfSame Bool Singleton wouldn’t make any sense.

Mixing Structures and Enumerations

Algebraic types are a generalization of both systems, in which you can have multiple constructors, each of which can contain zero or multiple arguments.

A canonical example combines both forms to create something resembling the concept of null from Java or None from Python.

data Maybe a = Nothing | Just a
    deriving Show

Here, there is a combination of the enumeration and structure style. Interestingly, while, for example, Just True :: Maybe Bool (:: is to be read as “is of type”), Nothing :: Maybe a, which means that it can take on any Just type (this is key to what we will do later).

Recursive Algebraic types

Recursive types are similar to the examples shown, except that they can also contain themselves. For example, the not particularly efficient definition of the natural numbers (a natural number is either zero or the successor of another natural number):

data Nat = Zero | Succ Nat
    deriving Show

Another algebraic type is a stream, which is defined as follows, which is basically an infinite list.

data Stream a = ConsStream a (Stream a)
    deriving Show

This only works because you can define values as follows, where in each case the same variable name refers to the same value:

trues = ConsStream True trues
falses = ConsStream False falses
alternating = ConsStream True (ConsStream False alternating)

elementAt (ConsStream value rest) Zero = value
elementAt (ConsStream value rest) (Succ x) = elementAt rest x

So if you ask for the (fifth element of trues), that reduces to the (fourth element of (the rest of trues)) which reduces to (the fourth element of trues), etc., until you get to the value True.

A potentially terminating list can be represented as follows:

data List a = Nil | Cons a (List a)
    deriving Show

Here the list can possibly terminate because you can stop with Nil. For example, you can represent [True, False, False] as

exampleList = Cons True (Cons False (Cons False Nil))

You could also use stream-like circular definition for a list too, so streams are just a subset of lists.

Higher-kinded type parameters

OK, one final bit of craziness before we get to our actual example. Let’s say we want to create a type representing the shape of another type.

data Shape f = Structure (f Singleton)

If we, for example, look at Shape List, we get the values Structure Nil, Structure (Cons Single Nil), Structure (Cons Single (Cons Single Nil)), etc. Basically, we get a picture of what the type looks like independent of its elements.

Notice that on the right side of that equation we have (f Singleton) as a concrete value, meaning that f is of kind * -> *. Therefore, Shape is of kind (* -> *) -> *.

Another example of a (* -> *) -> * is the NullOf type, which filters a type for only its nullary constructors, those which do not take a value.

data NullOf f = Null (f Void)

For example, the only possible value of NullOf List is Null Nil and the only possible value of NullOf Just is Null Nothing.

A general encoding of Recursive Types

You might notice a correspondence between the types Maybe and Nat:

    data Maybe a = Nothing | Just a
    data Nat = Zero | Succ Nat

Basically, Nat is just Maybe Nat. We can encode the conversion from Maybe to Nat as follows:

data Recursive f = Recurse (f (Recursive f))

This definition means that for all types f of kind * -> *, there exists a type called Recursive f (let’s call this rf for now) which contains a single element whose type is f rf. In other words, Recursive takes a type and returns a recursive version of that type.

Constructing a value of Recursive * at first seems difficult. For example, let us look at Recursive Maybe.

To construct a type of Recursive Maybe we need to have an element of type Maybe (Recursive Maybe). The only element of this we have is Nothing :: Maybe a. We can therefore construct the value Recurse Nothing :: Recursive Maybe. Using this value we can construct the value Just (Recurse Nothing) :: Maybe (Recursive Maybe) and therefore the value Recurse (Just (Recurse Nothing)) :: Recursive Maybe. We can continue to build values in this manner:

zero' = Recurse Nothing
one' = Recurse (Just zero')
two' = Recurse (Just one')
three' = Recurse (Just two')

Notice that I used the names for numbers for these. This is to demonstrate their correspondence with the natural numbers:

zero = Zero
one = Succ Zero
two = Succ one
three = Succ two

Basically, Zero is Nothing and Succ is Recurse . Just, where . is function composition.

Interestingly, there is another possible value of these types

omega' = Recurse (Just omega')
omega = Succ omega

The name omega is used because this number is similar to one particular kind of infinity. because of the way that it reacts to this definition of comparison:

_ < Zero        = False
Zero < Succ _   = True
Succ x < Succ y = x < y

As you can see, x < omega is True for all x defined by Succ (Succ (Succ ... Succ (Zero))). Note that omega < omega will never evaluate since omega < omega reduces to Succ omega < Succ omega, which reduces to omega < omega.

Recursivizing Types

To begin with, let’s look at the entirely contrived types:

data Constant a = Const
    deriving (Show)
data Empty a

We can see that there is one possible value for Constant (Recursvie Constant), which is Const. Therefore, the only possible value for Recursive Constant is Recurse Const. Therefore, Recursive Constant is similar to the Singleton:

single = Single
single' = Recurse Const

Similarly, there is no value Empty (Recursive Empty) possible, so there is no possible value of Recursive Empty.

Let’s now look at a less contrived * -> * type, Wrapper, and see how it works. Recursive Wrapper requires a value of Wrapper (Recursive Wrapper) in order to initialize. There is in fact only one way to do this:

wrappedRecurse = Recurse (Wrap wrappedRecurse)

This type is also basically Singleton since it only has one value

We can also use the partial application of Pair to a single type variable a and then construct a recursion of the pair. We can then analyze values of Recurse (Pair Bool). At first, this doesn’t seem to work any better than Recurse Wrap because we need a value of Pair Bool (RPair Bool) before we can create a value of RPair Bool. However, since in this case each value also has a Bool, these form a stream:

trues' = Recurse (ConsPair True trues')
falses' = Recurse (ConsPair False falses')
alternating' = Recurse (ConsPair True (Recurse (ConsPair False alternating')))

Which is remarkably similar to our previous definitions:

trues = ConsStream True trues
falses = ConsStream False falses
alternating = ConsStream True (ConsStream False alternating)

Finding the base type for recursive types

So now we know (I’m using \(\equiv\) to mean that two types have the same structure):

  • Recurse Maybe \(\equiv\) Nat
  • Recurse Constant \(\equiv\) Singleton
  • Recurse Wrapper \(\equiv\) Singleton
  • Recurse Empty \(\equiv\) Void
  • Recurse Constant \(\equiv\) Stream

But what is List equivalent to? We can look again at the definition of a list:

data List a = Nil | Cons a (List a)

OK, so we need to have a type parameter representing the a and another to be eaten by Recurse.

data PreList a self = ???

And we need to have some element Nil.

data PreList a self = PreNil | ???

Now, we just need something to combine both the first and rest of the list

data PreList a self = PreNil | PreCons a self

Hey, what’s on the right of the list looks a lot like a Maybe with two values instead of one! Here’s a final implementation

data PreList a self = PreList (Maybe (Pair a self))
    deriving (Show)

type RecurseList a = Recursive (PreList a)

Redoing our [True, False, False] example, we get:

exampleList'
    = Recurse (PreList (Just (ConsPair
        True
        (Recurse (PreList (Just (ConsPair
            False
            (Recurse (PreList (Just (ConsPair
                False
                (Recurse (PreList Nothing)))))))))))))

No reasonable person (except (maybe (a (Lisp (fanatic))))) would say that the above code is pretty, but it is clear that Recursive (PreList a) \(\equiv\) List.

Exercises

  • Find the de-recursified forms of
data BinaryTree a = BNode (BinaryTree a) (BinaryTree a) | BLeaf a
data ListTree a = ListNode a (List (ListTree a))
  • What is a general algorithm for de-recursification?

The answers are given at the end of the attached source code.

Conclusion

OK, so this probably doesn’t have any practical applications. And given how ugly the output of such code is, you’re unlikely to ever see this in real life code, even in an academic language like Haskell.

On the other hand, I hope this hurt your brain a little; it certainly hurt mine. Really, to me, the best part is that with only a simple type system (e.g., non-recursive types), and a single recursive type, you can construct all other simply recursive types. This example shows to me how out of simplicity can grow complexity.

Also, abstract nonsense is Just fun :-).

comments powered by Disqus