8 mins read

Learn Haskell 3: Algebraic Data Types

Category: Functional Programming Programming

A recent revelation I had is:

All I’m doing when I program is nothing but doing algebra - algebra on numbers and algebra on types.

Algebra on numbers are prevelant in functional programming, when we write a function, we mostly either do some algebra with the arguments, or call some other functions to do algebra for us. Afterall, functional programs are kind of like “executable mathamatics”.

What do I mean by “algebra on types”? Well, the algebra that I know, aka high school algebra, has additions and multiplications on numbers. “Doing algebra on types” simply means instead of doing additions and muliplications on numbers, we do “additions” and “multiplications” on types. Types that result from “doing algebra on types” are called Algebraic Data Types. The two fundamental algebraic data types are Sum type which corresponds to “adding” two types, and Product type which corresponds to “multiplying” two types.

Part 1: Sum (aka Coproduct)

If we think of types as sets of values, then a Sum type corresponds to the union1 of two types. In set theory, if we take two completely different sets, one has 5 elements and the other has 10, then the union of these two sets has 5 + 10 = 15 elements. What does this corresponds to in types? And what is a good example of a Sum type?

Well, if we think of True as a set, i.e., a singleton set that only has one element called True, and False as another set, i.e., a singleton set that only has one element called False, then Bool is actually a Sum type! Bool is exactly the union of the two singleton sets, the True set and the False set. The number of elements in Bool, as expected from set theory, is 1 + 1 = 2.

Actually, if we think of the sum type this way, all types are sum types. Integers are sum of all individual integers, Strings are sum of all possible strings:

data Bool = True | False
data Integer = 1 | 2 | 3 | ...
data String = "" | "a" | "b" | ... | "aa" | "ab" | ...

Of course, this is not how these primitive types are implemented in haskell, but at least conceptually, they are sums/unions of lots of different singleton sets.

A more classic example of a sum type is Maybe a (i.e. Option[A] in scala). For example, Maybe Char is a union of set Char and the singleton set Nothing. That is, an element of type Maybe Char is either Just Char (128 possible values, suppose we restrict Char to 128 ASCII characters) or Nothing (1 possible value). So, Maybe Char has 128 + 1 = 129 possible values which is how many elements a union is supposed to have.

In fact, in languages that have null, it just feels like all non-primitive types are Maybe types. For example in java/scala, although the type signature tells you that a variable is of type Boolean, but because null can appear anywhere, the variable can actually also be null. That is, when the compiler tells you it’s a Boolean, it’s actually not a Boolean, it maybe is a Boolean… I think that’s perhaps why null causes lots of run-time null pointer exceptions.

Another example of a sum type that I used a lot but I’m not aware of the fact that they are sum types is subclasses in OOP languages. Say there is an abstract class/trait called Animal and two subclasses Dog and Cat, then in the code where the type signature tells us it’s an animal, we know that it can either be a dog or a cat. That is, the set Animal is the union of set Dog and set Cat.

Part 2: Product

Let’s still think of types as sets. As discussed before, Bool is a set of two elements, True and False. Char is a set of 128 elements (suppose Char is restricted to ASCII characters). We can represent the Cartesian Product of these two sets as a new type which we denote as (Bool, Char). This new type, (Bool, Char), has 2 components: the first component is a Bool and it has 2 possible values, True and False, and the second component is a Char and it has 128 possible values. Therefore the new type has 2 * 128 = 256 possible values, which is exactly the number of elements that the Cartesian product is meant to have. This new type, usually called a pair or tuple depending on the language, is a classic example of a Product type. It represents the Cartesian product of two types.

What else are also product types? Well, records (aka dictionary) are also Product. The only difference it has with the pair type is how the components are named. For a pair, the components are named as as fst and snd in haskell, _1, _2 in scala, [0], [1] in python, whereas records/dictionaries are named using some user-defined names. Really, this is a superficial difference.

What’s more revealing is that, if you think about it, a (immutable) class in an OOP language is in essence also a Product type! They are just like records or dictionaries, the components of the product are member variables (aka, fields, instance variables, etc) and methods. Actually, this type of Product (products with with custom names and functions and without mutation), is so useful that scala has a special syntax for them: case class. In fact, all case classes automatically extend Product in scala.

Part 3: Algebra on Sum and Product

with addition and multiplication, we can actually do some simple algebra already. List, perhaps not surprisingly, is simply the result of performing a few simple additions and multiplications on some generic type. The classic definition of a list is as follows:

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

It’s a recursive definition, which basically says that List a is either Nil (empty list) or it’s a pair whose first element is of type a and the second element is another List a. If we translate this into the “type algebra”, it’s something like this:

\[ \begin{align} List(a) = Nil + a \times List(a) \end{align} \]

Well, we have one equation and one unknown - \(List(a)\), maybe we can try to solve for it:

\[ \begin{align} List(a) & = Nil + a \times List(a) \\ & = Nil + a \times (Nil + a \times List(a)) \\ & = Nil + a \times Nil + a \times a \times List(a) \\ & = Nil + a \times Nil + a \times a \times (Nil + a \times List(a)) \\ & = Nil + a \times Nil + a \times a \times Nil + a \times a \times List(a) \\ & = Nil + a \times Nil + a \times a \times Nil + a \times a \times a \times List(a) \\ & = ... \\ & = \sum_{i=0}^\infty{ a^i \times Nil } \end{align} \]

Wow, the resulting equation perhaps reveals more substance of a list: a list contains either no element of type a and Nil (\(a^0 \times Nil\)), or 1 element of type a and Nil (\(a^1 \times Nil\)), or 2 elements of type a and Nil (\(a^2 \times Nil\)), or 3 elements of type a and Nil, and so on… Looks like the distributive law in high school algebra also applies to type algebra…

Part 4: Exponential

So, now it looks like that pretty much any data type I use when I do any sort of programming, from primitive types like integers, booleans, to more complex types like pair (aka tuple), Maybe (aka Option) and list, to user defined data types like case class and classes in class hierachies, they are all algebraic data types!!! That is, they are simply some algebra on primitive types! In some sense, when I program in a functional style (without mutation), all I’m doing is nothing but algebra - algebra on numbers and algebra on types.

There’s one thing that’s omitted here - functions. Can functions also fit in to this story of doing algebra? aren’t they some different beasts altogether? Well, it turns out that they are called exponential in category theory. Here’s why from a layman’s perspective.

Let’s say a function is of type Bool -> Char. Let’s still think of types as sets, so in this case, one of the possible elements in the set Bool -> Char is:

myFunction b = if b then 'a' else 'b'
:t myFunction
myFunction True
myFunction False
## myFunction :: Bool -> Char
## 'a'
## 'b'

myFunction takes a Bool b and returns 'a' if the argument is True or 'b' if it is False. How many distinct such functions can there possibly be? That is, how many elements are there in the set Bool -> Char?

Again let’s suppose Char is ASCII character and it has only 128 elements. When b is True, there’re 128 different functions, one for each of the 128 possible ASCII characters; when b is False, there’re also 128 different functions that each return different ASCII characters. So in total, we have \(128 \times 128 = 128^2\) such functions in the set of Bool -> Char. In category theory, mathematicians call this type an Exponential, denoted as \(Char^{Bool}\). And it sort of describes the cardinality of the type.

So, the boundary between function types and data types are sort of blurred from the point of view of catgegory theory.

Actually, the boundary between function and data is also sort of blurred. We can think of functions as a lookup table that stores the data. Unlike elements in a data type which are themselves data, elements in a function type, i.e. a function, is like a database, we need to give it an argument, or a query, in order to get some value back.

Also, in one of the previous post, we can actually use pure functions to implement boolean values.

Conclusion

That’s my recent revelation. I don’t know where it will lead me to lol.


  1. Actually to be more precisely, the sum type corresponds to tagged unions in set theory.