# CIS 194, week 2: Algebraic Data Types

Posted by Manuel on Permalink

Updated on

Tags:

This is a post about week 2 of the CIS 194: Introduction to Haskell course by Brent Yorgey, from the Penn University of Pennsilvania.

This week is all about **Algebraic Data Types** (ADT), not to be confused with *Abstract* Data Types (also ADT) which are another topic.

Haskell has *enumeration types* (like Java, but still less verbose and more intuitive). An example:

data Food = Pizza | Bacon | Salad deriving Show

We have just declared a new (algebraic data) type called `Food`

, with three *data constructors* which are the *only* values of the type `Food`

.

We can now define new functions on the new data type using *pattern matching*:

isTempting:: Food -> Boolean isTempting Salad = False isTempting _ = True

But in Haskell enumeration types are only a special case of Algebraic Data Types. One common class of ADT is called sum type (a.k.a. tagged union, or type class in Scala). A simple example of ADT which is not an enumeration is this:

data OperationResult = OK | Error Integer deriving Show

Here we see that `Error`

is a data constructor that *takes an argument* of type `Integer`

. We can *construct* new `OperationResult`

values using the `Error`

data constructor:

success:: OperationResult success = OK failure:: OperationResult failure = Error 404

`OK`

is a value of type `OperationResult`

(since it's a data constructor with zero arguments), but `Error`

by itself it's not. We have to pass an `Integer`

value to it to build an `OperationResult`

with it.

We've just introduced *polymorphic data types*. Specifically, we can have *type signatures with variables* just as we can have function implementations with variables. The difference here is that while in actual code variables are symbols bound to *values*, in *type variables are bound to types of values*. In other words, types are actually values in type signatures. We're reasoning on a higher and more abstract level. Take a moment to contemplate this fact.

Formally, in Haskell an ADT is *a type with one or more data constructors, each one of them can have zero or more arguments*.

A general example that shows how to build values is:

isSafeDiv:: Double -> Double -> OperationResult isSafeDiv _ 0 = Error 1000 isSafeDiv _ _ = OK

We can also use pattern matching to make decisions based on the *structure* of the `OperationResult`

value and bind variables to the arguments:

isSuccessful:: OperationResult -> Boolean isSuccessful OK = True isSuccessful (Error n) = False

It's idiomatic in Haskell when you have an algebraic data type with a single data constructor, to name it like the data type itself. Example:

data Person = Person String String Int deriving Show

This can be done since *types and data constructors live in separate namespaces*.

## Pattern Matching

In general, pattern matching is a way to *know what data constructor has been used to create a value of a certain ADT, and to take apart its arguments*. Effectively, in Haskell this is *the only way to make decisions*.

pat ::= _ | var | var @ (pat) | (Constructor pat1 pat2 ... patn)In order:

`_`

is a wildcard.- We can pattern match against literal values (for example:
`OK`

). - We can pattern match against a pattern, and still bind the entire value to a variable.
- We can pattern match against a data constructor (even recursively).

It's worth noting that types like `Int`

can be viewed like an ADT:

data Int = 0 | 1 | 2 | ...

Indeed, we can pattern match on its values. But, perhaps obviously, they are not implemented like that in the compiler.

### Case expressions

A way (the only one actually) to do pattern matching is by using *case expressions*:

case exp of pat1 -> exp1 pat2 -> exp2 ...

For example, we could reimplement the `isSuccessful`

function from earlier using a case expression:

isSuccessful:: OperationResult -> Boolean isSuccessful op = case op of OK -> True (Error n) -> False

However it's more elegant to use the first version. Indeed, the syntax for doing pattern matching in a function definition is just *syntactic sugar* on case expressions.

## Recursive algebraic data types

It's interesting to note that ADTs can be *recursive*. For example, let's define a list of integers:

data IntList = Empty | Cons Int IntList

This definition can be read as: "an `IntList`

is either an `Empty`

one or an `Int`

value followed by an `IntList`

". This kind of definition is quite clear and elegant (see Church encoding). For example:

-- [1,2,3] can be represented as an IntList: l:: IntList l = Cons 1 (Cons 2 (Cons 3 Empty))

A recursive data ADT naturally leads to recursive functions. For example, to calculate the sum of all the values in an `IntList`

:

calcSum:: IntList -> Int calcSum Empty = 0 calcSum (Cons n l) = n + calcSum l

So, we've seen so far that type signatures can have variables, and can be recursive. Sounds like we could have a Turing-complete type system... indeed, we have one. Someone even implemented a LISP interpreter that completely runs on the Haskell type system!

That's all for this week. Remember: *do the exercises*!