Values and Functions

Introduction

In OCaml, functions are treated as values, so you can use functions as arguments to functions and return them from functions. This tutorial introduces the relationship between expressions, values, and names. The first four sections address non-function values. The following sections, starting at Function as Values, address functions.

We use UTop to understand these concepts by example. You are encouraged to modify the examples to gain a better understanding.

What is a Value?

Like most functional programming languages, OCaml is an expression-oriented programming language. That means programs are expressions. Actually, almost everything is an expression. In OCaml, statements don't specify actions to be taken on data. All computations are made through expression evaluation. Computing expressions produce values. Below, you'll find a few examples of expressions, their types, and the resulting values. Some include computation and some do not:

# "Every expression has a type";;
- : string = "Every expression has a type"

# 2 * 21;;
- : int = 42

# int_of_float;;
- : float -> int = <fun>

# int_of_float (3.14159 *. 2.0);;
- : int = 6

# fun x -> x * x;;
- : int -> int = <fun>

# print_endline;;
- : string -> unit = <fun>

# print_endline "Hello!";;
Hello!
- : unit

An expression's type (before evaluation) and its resulting value's type (after computation) are the same. This allows the compiler to avoid runtime type checks in binaries. In OCaml, the compiler removes type information, so it's not available at runtime. In programming theory, this is called subject reduction.

Global Definitions

Every value can be named. This is the purpose of the let … = … statement. The name is on the left; the expression is on the right.

• If the expression can be evaluated, it is.
• Otherwise, the expression is turned into a value as-is. That's the case of function definition.

This is what happens when writing a definition in UTop:

# let the_answer = 2 * 3 * 7;;
val the_answer : int = 42

Global definitions are those entered at the top level. Here, the_answer is defined globally.

Local Definitions

Local definitions bind a name inside an expression:

# let d = 2 * 3 in d * 7;;
- : int = 42

# d;;
Error: Unbound value d

Local definitions are introduced by the let … = … in … expression. The name bound before the in keyword is only bound in the expression after the in keyword. Here, the name d is bound to 6 inside the expression d * 7.

A couple of remarks:

• No global definition is introduced in this example, which is why we get an error.
• Computation of 2 * 3 will always take place before d * 7.

Local definitions can be chained (one after another) or nested (one inside another). Here is an example of chaining:

# let d = 2 * 3 in
  let e = d * 7 in
  d * e;;
- : int = 252

# d;;
Error: Unbound value d
# e;;
Error: Unbound value e

Here is how scoping works:

• d is bound to 6 inside let e = d * 7 in d * e
• e is bound to 42 inside d * e

Here is an example of nesting:

# let d =
    let e = 2 * 3 in
    e * 5 in
  d * 7;;
- : int = 210

# d;;
Error: Unbound value d
# e;;
Error: Unbound value e

Here is how scoping works:

• e is bound to 6 inside e * 5
• d is bound to 30 inside d * 7

Arbitrary combinations of chaining or nesting are allowed.

In both examples, d and e are local definitions.

Pattern Matching in Definitions

Some definitions can introduce more than one name or no name.

Pattern Matching on Tuples

A common case is tuples. It allows the creation of two names with a single let.

# List.split;;
- : ('a * 'b) list -> 'a list * 'b list

# let (x, y) = List.split [(1, 2); (3, 4); (5, 6); (7, 8)];;
val x : int list = [1; 3; 5; 7]
val y : int list = [2; 4; 6; 8]

The List.split function turns a list of pairs into a pair of lists. Here, each resulting list is bound to a name.

Pattern Matching on Records

We can pattern match on records:

# type name = { first : string; last: string };;
type name = { first : string; last : string; }

# let robin = { first = "Robin"; last = "Milner" };;
val robin : name = {first = "Robin"; last = "Milner"}

# let { first; last } = robin;;
val first : string = "Robin"
val last : string = "Milner"

Pattern Matching on unit

A special case of combined definition and pattern matching involves the unit type:

# let () = print_endline "ha ha";;
ha ha

Note: As explained in the Tour of OCaml tutorial, the unit type has a single value (), which is pronounced "unit."

Above, the pattern does not contain any identifier, meaning no name is defined. The expression is evaluated and the side effect takes place (printing ha ha to standard output).

Note: In order for compiled files to only evaluate an expression for its side effects, you must write them after let () =.

Pattern Matching on User-Defined Types

This also works with user-defined types.

# type live_person = int * name;;
type live_person = int * name

# let age ((years, { first; last }) : live_person) = years;;
val age : live_person -> int = <fun>

Discarding Values Using Pattern Matching

As seen in the last example, the catch-all pattern (_) can be used in definitions.

# let (_, y) = List.split [(1, 2); (3, 4); (5, 6); (7, 8)];;
val y : int list = [2; 4; 6; 8]

The List.split function returns a pair of lists. We're only interested in the second list, we give it the name y and discard the first list by using _.

Scopes and Environments

Without oversimplifying, an OCaml program is a sequence of expressions or global let definitions.

Execution evaluates each item from top to bottom.

At any time during evaluation, the environment is the ordered sequence of available definitions. The environment is also known as context in other languages.

Here, the name twenty is added to the top-level environment.

# let twenty = 20;;
val twenty : int = 20

The scope of twenty is global. This name is available anywhere after the definition.

Here, the global environment is unchanged:

# let ten = 10 in 2 * ten;;
- : int = 20

# ten;;
Error: Unbound value ten

Evaluating ten results in an error because it hasn't been added to the global environment. However, in the expression 2 * ten, the local environment contains the definition of ten.

Although OCaml is an expression-oriented language, it has a few statements. The global let modifies the global environment by adding a name-value binding.

Top-level expressions are also statements because they are equivalent to let _ = definitions.

# (1.0 +. sqrt 5.0) /. 2.0;;
- : float = 1.6180339887498949

# let _ = (1.0 +. sqrt 5.0) /. 2.0;;
- : float = 1.6180339887498949

Inner Shadowing

Once you create a name, define it, and bind it to a value, it does not change. That said, a name can be defined again to create a new binding:

# let i = 21;;
val i : int = 21

# let i = 7 in i * 2;;
- : int = 14

# i;;
- : int = 21

The second definition shadows the first. Inner shadowing is limited to the local definition's scope. Therefore, anything written after will still take the previous definition, as shown above. Here, the value of i hasn't changed. It's still 21, as defined in the first expression. The second expression binds i locally, inside i * 2, not globally.

Same-Level Shadowing

Another kind of shadowing takes place when there are two definitions with the same name at the same level.

# let h = 2 * 3;;
val h : int = 6

# let e = h * 7;;
val e : int = 42

# let h = 7;;
val h : int = 7

# e;;
- : int = 42

There are now two definitions of h in the environment. The first h is unchanged. When the second h is defined, the first one becomes unreachable.

Function as Values

In OCaml, functions are values. This is the key concept of functional programming. In this context, it is also possible to say that OCaml has first-class functions.

Applying Functions

When several expressions are written side by side, the leftmost one should be a function. All the others are arguments. In OCaml, no parentheses are needed to express passing an argument to a function. Parentheses serve a single purpose: associating expressions to create subexpressions.

# max (21 * 2) (int_of_string "713");;
- : int = 713

The max function returns the largest of its two arguments, which are:

• 42, the result of 21 * 2
• 713, the result of int_of_string "713"

When creating subexpressions, using begin ... end is also possible. This is the same as using brackets ( ... ). As such, the above could also be rewritten and get the same result:

# max begin 21 * 2 end begin int_of_string "713" end;;
- : int = 713

# String.starts_with ~prefix:"state" "stateless";;
- : bool = true

Some functions, such as String.starts_with have labelled parameters. Labels are useful when a function has several parameters of the same type; naming arguments allows to guess their purpose. Above, ~prefix:"state" indicates "state" is passed as labelled argument prefix.

Labelled and optional parameters are detailed in the Labelled Arguments tutorial.

There are two alternative ways to apply functions.

The Application Operator

The application operator @@ operator.

# sqrt 9.0;;
- : float = 3.

# sqrt @@ 9.0;;
- : float = 3.

The @@ application operator applies an argument (on the right) to a function (on the left). It is useful when chaining several calls, as it avoids writing parentheses, which creates easier-to-read code. Here is an example with and without parentheses:

# int_of_float (sqrt (float_of_int (int_of_string "81")));;
- : int = 9

# int_of_float @@ sqrt @@ float_of_int @@ int_of_string "81";;
- : int = 9

The Pipe Operator

The pipe operator (|>) also avoids parentheses but in reversed order: function on right, argument on left.

# "81" |> int_of_string |> float_of_int |> sqrt |> int_of_float;;
- : int = 9

This is just like a Unix shell pipe.

Anonymous Functions

Functions don't have to be bound to a name unless they are recursive. Take these examples:

# fun x -> x;;
- : 'a -> 'a = <fun>

# fun x -> x * x;;
- : int -> int = <fun>

# fun s t -> s ^ " " ^ t ;;
- : string -> string-> string = <fun>

# function [] -> None | x :: _ -> Some x;;
- : 'a list -> 'a option = <fun>

Function values not bound to names are called anonymous functions.

In order, here is what they are:

• The identity function, which takes anything and returns it unchanged
• The square function, which takes an integer and returns it squared
• The function that takes two strings and returns their concatenation with a space character in between
• The function that takes a list and either returns None, if the list is empty, or returns its first element.

Anonymous functions are often passed as arguments to other functions.

# List.map (fun x -> x * x) [1; 2; 3; 4];;
- : int list = [1; 4; 9; 16]

Defining Global Functions

You can globally bind a function to a name using a global definition.

# let f = fun x -> x * x;;
val f : int -> int = <fun>

The expression, which happens to be a function, is turned into value and bound to a name. Here is another way to do the same thing:

# let g x = x * x;;
val g : int -> int = <fun>

The former explicitly binds the anonymous function to a name. The latter uses a more compact syntax and avoids the fun keyword and the arrow symbol.

Defining Local Functions

A function may be defined locally.

# let sq x = x * x in sq 7 * sq 7;;
- : int = 2401

# sq;;
Error: Unbound value sq

Calling sq gets an error because it was only defined locally.

The function sq is only available inside the sq 7 * sq 7 expression.

Although local functions are often defined inside the function's scope, this is not a requirement.

Closures

This example illustrates a closure using Same-Level Shadowing

# let j = 2 * 3;;
val j : int = 6

# let k x = x * j;;
val k : int -> int = <fun>

# k 7;;
- : int = 42

# let j = 7;;
val j : int = 7

# k 7;; (* What is the result? *)
- : int = 42

Here is how this makes sense:

1. Constant j is defined, and its value is 6.
2. Function k is defined. It has a single parameter x and returns the value of x * j.
3. Compute k of 7, and its value is 42
4. Create a new definition j, shadowing the first one
5. Compute k of 7 again, the result is the same: 42

Although the new definition of j shadows the first one, the original remains the one the function k uses. The k function's environment captures the first value of j, so every time you apply k (even after the second definition of j), you can be confident the function will behave the same.

However, all future expressions will use the new value of j (7), as shown here:

# let m = j * 3;;
val m : int = 21

Partially applying arguments to a function also creates a new closure.

# let max_42 = max 42;;
val max_42 : int -> int = <fun>

Inside the max_42 function, the environment contains an additional binding between the first parameter of max and the value 42.

Recursive Functions

In order to perform iterated computations, a function may call itself. Such a function is called recursive.

# let rec fibo n =
    if n <= 1 then n else fibo (n - 1) + fibo (n - 2);;
val fibo : int -> int = <fun>

# let u = List.init 10 Fun.id;;
val u : int list = [0; 1; 2; 3; 4; 5; 6; 7; 8; 9]

# List.map fibo u;;
- : int list = [0; 1; 1; 2; 3; 5; 8; 13; 21; 34]

This is a classic (and very inefficient) way to compute Fibonacci numbers. The number of recursive calls created doubles at each call, which creates exponential growth.

In OCaml, recursive functions must be defined and explicitly declared by using let rec. It is not possible to accidentally create a recursive function, and recursive functions can't be anonymous.

Note: List.init is a standard library function that allows you to create a list by applying a given function to a sequence of integers, and Fun.id is the identity function, which returns its argument unchanged. We created a list with the numbers 0 - 9 and named it u. We applied the fibo function to every element of the list using List.map.

This version does a better job:

# let rec fib_loop m n i =
    if i = 0 then m else fib_loop n (n + m) (i - 1);;
val fib_loop : int -> int -> int -> int = <fun>

# let fib = fib_loop 0 1;;
val fib : int -> int = <fun>

# List.init 10 Fun.id |> List.map fib;;
- : int list = [0; 1; 1; 2; 3; 5; 8; 13; 21; 34]

The first version fib_loop has two extra parameters: the two previously computed Fibonacci numbers.

The second version fib uses the first two Fibonacci numbers as initial values. There is nothing to be computed when returning from a recursive call, so this enables the compiler to perform an optimisation called tail call elimination.

Note: Notice that the fib_loop function has three parameters m n i but when defining fib only two arguments were passed 0 1, using partial application.

Functions with Multiple Parameters

Defining Functions with Multiple Parameters

To define a function with multiple parameters, each must be listed between the name of the function (right after the let keyword) and the equal sign, separated by space:

# let sweet_cat x y = x ^ " " ^ y;;
val sweet_cat : string -> string -> string = <fun>

# sweet_cat "kitty" "cat";;
- : string = "kitty cat"

Anonymous Functions with Multiple Parameters

We can use anonymous functions to define the same function in a different way:

# let sour_cat = fun x -> fun y -> x ^ " " ^ y;;
val sour_cat : string -> string -> string = <fun>

# sour_cat "kitty" "cat";;
- : string = "kitty cat"

Observe sweet_cat and sour_cat have the same body: x ^ " " ^ y. They only differ in the way parameters are listed:

1. As x y between name and = in sweet_cat
2. As fun x -> fun y -> after = in sour_cat (and nothing but name before =)

Also observe that sweet_cat and sour_cat have the same type: string -> string -> string.

If you check the assembly code generated using compiler explorer, you'll see it is the same for both functions.

The way sour_cat is written corresponds more explicitly to the behaviour of both functions. The name sour_cat is bound to an anonymous function having parameter x and returning an anonymous function having parameter y and returning x ^ " " ^ y.

The way sweet_cat is written is an abbreviated version of sour_cat. Such a way of shortening syntax is called syntactic sugar.

Partial Application and Closures

We want to define functions of type string -> string that appends "kitty " in front of its arguments. This can be done using sour_cat and sweet_cat

# let sour_kitty x = sour_cat "kitty" x;;
val sour_kitty : string -> string = <fun>

# let sweet_kitty = fun x -> sweet_cat "kitty" x;;
val sweet_kitty : string -> string = <fun>

# sour_kitty "cat";;
- : string = "kitty cat"

# sweet_kitty "cat";;
- : string = "kitty cat"

However, both definitions can be shortened using something called partial application

# let sour_kitty = sour_cat "kitty";;
val sour_kitty : string -> string = <fun>

# let sweet_kitty = sweet_cat "kitty";;
val sweet_kitty : string -> string = <fun>

Since a multiple-parameter function is a series of nested single-argument functions, you don't have to pass all arguments at once.

Passing a single argument to sour_kitty or sweet_kitty returns a function of type string -> string. The first argument, here "kitty", is captured and the result is a closure.

These expressions have the same value:

• fun x -> sweet_cat "kitty" x
• sweet_cat "kitty"

Types of Functions of Multiple Parameters

Let's look at the types here:

# let dummy_cat : string -> (string -> string) = sweet_cat;;
val dummy_cat : string -> string -> string = <fun>

Here the type annotation : string -> (string -> string) is used to explicitly state the type of dummy_cat.

However, OCaml answers claiming the fresh definition has type string -> string -> string. This is because types string -> string -> string and string -> (string -> string) are the same.

With parentheses, it is obvious that a multiple-argument function is a single-parameter function that returns an anonymous function with one parameter removed.

Putting the parentheses the other way does not work:

# let bogus_cat : (string -> string) -> string = sweet_cat;;
Error: This expression has type string -> string -> string
       but an expression was expected of type (string -> string) -> string
       Type string is not compatible with type string -> string

Functions having type (string -> string) -> string take a function as a parameter. The function sweet_cat has a function as a result, not a function as a parameter.

The type arrow operator associates to the right. Function types without parentheses should be treated as if they have parentheses to the right in the same way that the type of dummy_cat was declared above. Except they are not displayed.

Tuples as Function Parameters

In OCaml, a tuple is a data structure used to group a fixed number of values, which can be of different types. Tuples are surrounded by parentheses, and the elements are separated by commas. Here's the basic syntax to create and work with tuples in OCaml:

# ("felix", 1920);;
- : string * int = ("felix", 1920)

It is possible to use the tuple syntax to specify function parameters. Here is how it can be used to define yet another version of the running example:

# let spicy_cat (x, y) = x ^ " " ^ y;;
val spicy_cat : string * string -> string = <fun>

# spicy_cat ("hello", "world");;
- : string = "hello world"

It looks like two arguments have been passed: "hello" and "world". However, only one, the ("hello", "world") tuple, has been passed. Inspection of the generated assembly would show it isn't the same function as sweet_cat. It contains some more code. The contents of the tuple passed to spicy_cat (x and y) must be extracted before evaluation of the x ^ " " ^ y expression. This is the role of the additional assembly instructions.

In many imperative languages, the spicy_cat ("hello", "world") syntax reads as a function call with two arguments; but in OCaml, it denotes applying the function spicy_cat to a tuple containing "hello" and "world".

Currying and Uncurrying

In the previous sections, two kinds of multiple-parameter functions have been presented.

• Functions returning a function, such as sweet_cat and sour_cat
• Functions taking a tuple as a parameter, such as spicy_cat

Interestingly, both kinds of functions provide a way to pass several pieces of data while being functions with a single parameter. From this perspective, it makes sense to say: “All functions have a single argument.”

This goes even further. It is always possible to translate back and forth between functions that look like sweet_cat (or sour_cat) and functions that look like spicy_cat.

These translations have names:

• Currying goes from the spicy_cat form into the sour_cat (or sweet_cat) form.
• Uncurrying goes from the sour_cat (or sweet_cat) form into the spicy_cat form.

It also said that sweet_cat and sour_cat are curried functions whilst spicy_cat is uncurried.

Functions with the following types can be translated back and forth:

• string -> (string -> string) — curried function type
• string * string -> string — uncurried function type

These translations are attributed to the 20th-century logician Haskell Curry.

Here, this is shown using string as an example, but it applies to any group of three types.

You can change the curried form into the uncurried form when refactoring, or the other way round.

However, it is also possible to implement one from the other to have both forms available:

# let uncurried_cat (x, y) = sweet_cat x y;;
val uncurried_cat : string * string -> string = <fun>

# let curried_cat x y = uncurried_cat (x, y);;
val curried_cat : string -> string -> string = <fun>

In practice, curried functions are the default because:

• They allow partial application
• No parentheses or commas
• No pattern matching over a tuple takes place

Functions With Side Effects

To explain side effects, we need to define what domain and codomain are. Let's look at an example:

# string_of_int;;
- : int -> string = <fun>

For the function string_of_int:

• Its domain is int, the type of its parameters
• The codomain is string, the type of its results

In other words, the domain is left of the -> and the codomain is on the right. These terms help avoid saying the "type at the right" or "type at the left" of a function's type arrow.

Some functions operate on data outside of their domain or codomain. This behaviour is called an effect, or a side effect.

Doing input and output (I/O) with the operating system is the most common form of side effects. The result of functions returning random numbers (such as Random.bits does) or the current time (such as Unix.time does) is influenced by external factors, which is also called an effect.

Similarly, any observable phenomena triggered by the computation of a function is an out-of-codomain output.

In practice, what is considered an effect is an engineering choice. In most circumstances, system I/O operations are considered as effects, unless they are ignored. The heat emitted by the processor when computing a function isn't usually considered a relevant side-effect, except when considering energy-efficient design.

In the OCaml community, as well as in the wider functional programming community, functions are often said to be either pure or impure. The former does not have side effects, the latter does. This distinction makes sense and is useful. Knowing what the effects are, and when are they taking place, is a key design consideration. However, it is important to remember this distinction always assumes some sort of context. Any computation has effects, and what is considered a relevant effect is a design choice.

Since, by definition, effects lie outside function types, a function type can't reflect a function's possible effects. However, it is important to document a function's intended side effects. Consider the Unix.time function. It returns the number of seconds elapsed since January 1, 1970.

# Unix.time ;;
- : unit -> float = <fun>

Note: If you're getting an Unbound module error in macOS, run this first: #require "unix";;.

The result of the Unix.time function is determined only by external factors. To perform the side effect, the function must be applied to an argument. Since no data needs to be passed, the argument is the () value.

Consider print_endline. It prints the string it was passed to standard output, followed by a line termination.

# print_endline;;
- : string -> unit = <fun>

Since the purpose of the function is only to produce an effect, it has no meaningful data to return; it returns the () value.

This illustrates the relationship between functions that have side effects and the unit type. The presence of the unit type does not indicate the presence of side effects. The absence of the unit type does not indicate the absence of side effects. But when no data needs to be passed as input or can be returned as output, the unit type is used.

What Makes Functions Different From Other Values

Functions are like other values; however, there are restrictions:

1. Function values cannot be displayed in interactive sessions. The placeholder <fun> is displayed instead. This is because there is nothing meaningful to print. Once parsed and typed-checked, OCaml discards the function's source code and nothing remains to be printed.

# sqrt;;
- : float -> float = <fun>

2. Equality between functions can't be tested.

# pred;;
- : int -> int = <fun>

# succ;;
- : int -> int = <fun>

# pred = succ;;
Exception: Invalid_argument "compare: functional value".

There are two main reasons explaining this:

• There is no algorithm that takes two functions and determines if they return the same output when provided the same input.
• Assuming it was possible, such an algorithm would declare that implementations of quicksort and bubble sort are equal. That would mean one could replace the other, and that may not be wise.

Conclusion

At the heart of OCaml lies the concept of the environment. The environment works as an ordered, append-only, key-value store. This means that items cannot be removed. Furthermore, it maintains order by preserving the sequence of available definitions.

When we use a let statement, we introduce zero, one, or more name-value pairs into the environment. Similarly, when applying a function to some arguments, we extend the environment by adding names and values corresponding to its arguments.