blog/content/posts/2023-03-26-inductive-types.md

+++ title = "Inductive Types" date = 2023-03-26 tags = ["type-theory"] math = true draft = true toc = true

language_switcher_languages = ["ocaml", "python"] +++

There's a feature common to many functional languages, the ability to have algebraic data types. It might look something like this:

{{< language-switcher >}}

type bool =
| True
| False

For those who are unfamiliar with the syntax, I'm defining a type called bool that has two constructors, or ways to create this type. The constructors are True and False. This means:

1. Any time I use the value True, it's understood to have type bool.
2. Any time I use the value False, it's understood to have type bool.
3. In addition, there are no other ways to create values of bool other than combining True and False constructors.

---

from typing import Literal
MyBool = Literal[True] | Literal[False]

{{</ language-switcher >}}

Note: I'm using an experimental language switcher. It's implemented in pure CSS using a feature called the :has pseudo-class. As of writing, all major browsers except Firefox has it implemented and enabled by default. For Firefox there does exist a feature flag in about:config, but your mileage may vary.

---

Many languages have this feature, under different names. Tagged unions, variant types, enumerations, but they all reflect a basic idea: a type with a limited set of variants.

Now, in type theory, one of the interesting things to know about a type is its cardinality. For example, the type Boolean is defined to have cardinality 2. That's because there's only one constructor, so if at any point you have some unknown value of type Boolean, you know it can only take one of two values.

Note about Booleans

There's actually nothing special about boolean itself. I could just as easily define a new type, like this:

{{< language-switcher >}}

type WeirdType =
| Foo
| Bar

---

from typing import Literal
WeirdType = Literal['foo'] | Literal['bar']

{{</ language-switcher >}}

Because this type can only have two values, it's semantically equivalent to the Boolean type. I could use it anywhere I would typically use Boolean.

I would have to define my own operators such as AND and OR separately, but those aren't properties of the Boolean type itself, they are properties of the Boolean algebra, which has several algebraic properties such as associativity, commutativity, distributivity, and several others. Think of it as a sort of interface, where if you can implement that interface, your type qualifies as a Boolean algebra!

You can make any finite type like this: just create an algebraic data type with unit constructors, and the result is a type with a finite cardinality. If I wanted to make a unit type for example:

{{< language-switcher >}}

type unit =
| Unit

---

from typing import Literal
Unit = Literal[None]

{{</ language-switcher >}}

There's only one way to ever construct something of this type, so the cardinality of this type would be 1.

Doing things with types

Creating types and making values of those data types is just the first part though. It would be completely uninteresting if all we could do is create types. So, the way we typically use these types is through pattern matching (also called structural matching in some languages).

Let's see an example. Suppose I have a type with three values, defined like this:

{{< language-switcher >}}

type direction =
| Left
| Middle
| Right

---

from typing import Literal
Direction = Literal['left'] | Literal['middle'] | Literal['right']

{{</ language-switcher >}}

If I was given a value with a type of direction, but I wanted to do different things depending on exactly which direction it was, I could use pattern matching like this:

{{< language-switcher >}}

let do_something_with (d : direction) =
  match d with
  | Left -> do_this_if_left
  | Middle -> do_this_if_middle
  | Right -> do_this_if_right

---

def do_something_with(d : Direction) -> str:
  match inp:
    case 'left': return do_this_if_left
    case 'middle': return do_this_if_middle
    case 'right': return do_this_if_right
    case _: assert_never(inp)

Note: the assert_never is a static check for exhaustiveness. If we missed a single one of the cases, a static type checker like pyright could catch it and tell us which of the remaining cases there are.

{{</ language-switcher >}}

This gives me a way to discriminate between the different variants of direction.

Most languages have a built-in construct for discriminating between values of the Boolean type, called if-else. What would if-else look like if you wrote it as a function in this pattern-matching form?

Constructing larger types

Finite-cardinality types like the ones we looked at just now are nice, but they're not super interesting. If you had a programming language with nothing but those, it would be very painful to write in! This is where type constructors come in.

When I say type constructor, I mean a type that can take types and build other types out of them. There's several ways this can be done, but the one I want to discuss today is called inductive types.

If you don't know what induction is, the Wikipedia article on it is a great place to start!

The general idea is that we can build types using either base cases (variants that don't contain themselves as a type), or inductive cases (variants that do contain themselves as a type).

You can see an example of this here:

{{< language-switcher >}}

type nat =
| Suc of nat
| Zero

{{</ language-switcher >}}

These are the natural numbers, which are defined inductively. Each number is just represented by a data type that wraps 0 that number of times. So 3 would be Suc (Suc (Suc Zero)).

At this point you can probably see why these have infinite cardinality: with the Suc case, you can keep wrapping nats as many times as you want!

One key observation here is that although the cardinality of the entire type is infinite, it only uses two constructors to build it. This also means that if we want to talk about writing any functions on nat, we just have to supply 2 cases instead of an infinite number of cases:

let rec is_even = fun (n : nat) ->
  match n with
  | Zero -> true
  | Suc n1 -> not (is_even n1)

Try it for yourself

If you've got an OCaml interpreter handy, try a couple values for yourself and convince yourself that this accurately represents the naturals and an even testing function:

utop # is_even Zero;;
- : bool = true
utop # is_even (Suc Zero);;
- : bool = false

This is a good way of making sure the functions you write make sense!

Induction principle

Let's express this in the language of mathematical induction. If I have any natural number:

• If the natural number is the base case of zero, then the is_even relation automatically evaluates to true.
• If the natural number is a successor, invert the induction hypothesis (which is what is_even evaluates to for the previous step, a.k.a whether or not the previous natural number is even), since every even number is succeeded by an odd number and vice versa.

Once these rules are followed, by induction we know that is_even can run on any natural number ever. In code, this looks like:

let is_even
    (n_zero : bool)
    (n_suc : nat -> bool)
    (n : nat)
  : bool =
  match n with
  | Zero -> n_zero
  | Suc n1 -> n_suc n1

where n_zero defines what to do with the zero case, and n_suc defines what to do with the successor case.

You might've noticed that this definition doesn't actually return any booleans. That's because this is not actually the is_even function! This is a general function that turns any natural into a boolean. In fact, we can go one step further and generalize this to all types:

let nat_transformer
    (n_zero : 'a)
    (n_suc : 'a -> 'a)
    (n : nat)
  : 'a =
  match n with
  | Zero -> n_zero
  | Suc n1 -> n_suc n1

Let's say I wanted to write a function that converts from our custom-defined nat type into an OCaml integer. Using this constructor, that would look something like this:

let convert_nat = nat_transformer 0 (fun x -> x + 1)

TODO: Talk about https://counterexamples.org/currys-paradox.html