Functional programming with Juvix¶
Welcome to the Juvix functional programming tutorial! This thorough guide provides a structured introduction to Juvix language features and functional programming concepts. By the end of this tutorial, you'll have a strong foundation in functional programming with Juvix. Let's get started on your Juvix journey!
Juvix REPL¶
After installing Juvix, launch the Juvix REPL:
juvix repl
The response should be similar to:
Juvix REPL version 0.6.1: https://juvix.org. Run :help for help
OK loaded: ./.juvix-build/stdlib/Stdlib/Prelude.juvix
Stdlib.Prelude>
Currently, the REPL supports evaluating expressions but it does not yet support
adding new definitions. To see the list of available REPL commands type :help
.
Basic expressions¶
You can try evaluating simple arithmetic expressions in the REPL:
Stdlib.Prelude> 3 + 4
7
Stdlib.Prelude> 1 + 3 * 7
22
Stdlib.Prelude> div 35 4
8
Stdlib.Prelude> mod 35 4
3
Stdlib.Prelude> sub 35 4
31
Stdlib.Prelude> sub 4 35
0
By default, Juvix operates on non-negative natural numbers. Natural number
subtraction is implemented by the function sub
. Subtracting a bigger natural
number from a smaller one yields 0
.
You can also try boolean expressions
Stdlib.Prelude> true
true
Stdlib.Prelude> not true
false
Stdlib.Prelude> true && false
false
Stdlib.Prelude> true || false
true
Stdlib.Prelude> if | true := 1 | else := 0
1
and strings, pairs and lists:
Stdlib.Prelude> "Hello world!"
"Hello world!"
Stdlib.Prelude> (1, 2)
(1, 2)
Stdlib.Prelude> [1; 2; 3]
1 :: 2 :: 3 :: nil
[a; b; c]
is an abbreviation for a :: b :: c :: nil
, where ::
is a list "cons" operator and nil
(also []
) denotes the empty list.
In fact, you can use all functions and types from the Stdlib.Prelude module of the standard library, which is preloaded by default.
Stdlib.Prelude> length [1; 2]
2
Stdlib.Prelude> isEmpty [1; 2]
false
Stdlib.Prelude> swap (1, 2)
2, 1
Files, modules and compilation¶
Currently, the REPL does not support adding new definitions. To define new
functions or data types, you need to put them in a separate file and either load
the file in the REPL with :load file.juvix
, evaluate it with the shell command
juvix eval file.juvix
, or compile it to a binary executable with juvix
compile native file.juvix
.
To conveniently edit Juvix files, an Emacs mode and a VSCode extension are available.
A Juvix file must declare a module whose name corresponds exactly to the name of
the file. For example, a file Hello.juvix
must declare a module Hello
:
module Hello;
import Stdlib.Prelude open;
main : String := "Hello world!";
end;
A file compiled to an executable must define the zero-argument function main
which is evaluated when running the program. The definition of main
can have
any non-function type, e.g., String
, Bool
, Nat
or Int
. The generated executable
prints the result of evaluating main
.
Data types and functions¶
To see the type of an expression, use the :type
REPL command:
Stdlib.Prelude> :type 1
Nat
Stdlib.Prelude> :type -1
Int
Stdlib.Prelude> :type true
Bool
The types Nat
, Int
and Bool
are defined in the standard library.
The type Bool
has two constructors true
and false
.
type Bool :=
| true
| false;
The constructors of a data type can be used to build elements of the type. They
can also appear as patterns in function definitions. For example, the not
function is defined in the standard library by:
not : Bool -> Bool
| true := false
| false := true;
The type of the definition is specified after the colon. In this case, not
is
a function from Bool
to Bool
. The type is followed by two function clauses
which specify the function result depending on the shape of the arguments. When
a function call is evaluated, the first clause that matches the arguments is
used.
In contrast to languages like Python, Java or C/C++, Juvix doesn't require
parentheses for function calls. All arguments are just listed after the
function. The general pattern for function application is: func arg1 arg2 arg3 ...
Initial arguments that are matched against variables or wildcards in all clauses can be moved to the left of the colon. For example,
or (x : Bool) : Bool -> Bool
| true := true
| _ := x;
is equivalent to
or : Bool -> Bool -> Bool
| _ true := true
| x _ := x;
If there is only one clause and all arguments are to the left of the colon, the
pipe |
should be omitted:
id (x : Bool) : Bool := x;
A more complex example of a data type is the Nat
type from the standard
library:
type Nat :=
| zero : Nat
| suc : Nat -> Nat;
The constructor zero
represents 0
and suc
represents the successor
function – suc n
is the successor of n
, i.e., n+1
. For example, suc zero
represents 1
. The number literals 0
, 1
, 2
, etc., are just shorthands for
appropriate expressions built using suc
and zero
.
The constructors of a data type specify how the elements of the type can be
constructed. For instance, the above definition specifies that an element of
Nat
is either:
zero
, orsuc n
wheren
is an element ofNat
, i.e., it is constructed by applyingsuc
to appropriate arguments (in this case the argument ofsuc
has typeNat
).
Any element of Nat
can be built with the constructors in this way – there are
no other elements. Mathematically, this is an inductive definition, which is why
the data type is called inductive.
Constructors can either by specified by listing their types after colons like in
the above definition of Nat
, or with a shorter ADT syntax like in the
definition of Bool
. The ADT syntax is similar to data type definition syntax
in functional languages like Haskell or OCaml: one lists the types of
constructor arguments separated by spaces. In this syntax, the Nat
type could
be defined by
type Nat :=
| zero
| suc Nat;
If implemented directly, the above unary representation of natural numbers would
be extremely inefficient. The Juvix compiler uses a binary number representation
under the hood and implements arithmetic operations using corresponding machine
instructions, so the performance of natural number arithmetic is similar to
other programming languages. The Nat
type is a high-level presentation of
natural numbers as seen by the user who does not need to worry about low-level
arithmetic implementation details.
One can use zero
and suc
in pattern matching, like any other constructors:
syntax operator + additive;
+ : Nat -> Nat -> Nat
| zero b := b
| (suc a) b := suc (a + b);
The syntax operator + additive
declares +
to be an operator with the
additive
fixity. The +
is an ordinary function, except that function
application for +
is written in infix notation. The definitions of the clauses
of +
still need the prefix notation on the left-hand sides. Note that to use
this definition in the code one needs to import and open Stdlib.Data.Fixity
.
The a
and b
in the patterns on the left-hand sides of the clauses are
variables which match arbitrary values of the corresponding type. They can be
used on the right-hand side to refer to the values matched. For example, when
evaluating
(suc (suc zero)) + zero
the second clause of +
matches, assigning suc zero
to a
and zero
to b
.
Then the right-hand side of the clause is evaluated with a
and b
substituted
by these values:
suc (suc zero + zero)
Again, the second clause matches, now with both a
and b
being zero
. After
replacing with the right-hand side, we obtain:
suc (suc (zero + zero))
Now the first clause matches and finally we obtain the result
suc (suc zero)
which is just 2
.
The function +
is defined like above in the standard library, but the Juvix
compiler treats it specially and generates efficient code using appropriate CPU
instructions.
Evaluation order¶
By default, evaluation in Juvix is eager (or strict), meaning that
the arguments to a function are fully evaluated before applying the
function. Only logical operators ||
and &&
are treated specially
and evaluated lazily. These special functions cannot be partially
applied (see Partial application and higher-order
functions
below).
Pattern matching¶
The patterns in function clauses do not have to match on a single constructor – they may be arbitrarily deep. For example, here is an (inefficient) implementation of a function which checks whether a natural number is even:
isEven : Nat -> Bool
| zero := true
| (suc zero) := false
| (suc (suc n)) := isEven n;
This definition states that a natural number n
is even if either n
is zero
or, recursively, n-2
is even.
If a subpattern is to be ignored, then one can use a wildcard _
instead of
naming the subpattern.
isPositive : Nat -> Bool
| zero := false
| (suc _) := true;
The above function could also be written as:
isPositive : Nat -> Bool
| zero := false
| _ := true;
It is not necessary to define a separate function to perform pattern matching.
One can use the case
syntax to pattern match an expression directly.
Stdlib.Prelude> case (1, 2) of (suc _, zero) := 0 | (suc _, suc x) := x | _ := 19
1
It is possible to name subpatterns with @
.
Stdlib.Prelude> case 3 of suc n@(suc _) := n | _ := 0
2
Comparisons and conditionals¶
The standard library includes all the expected comparison operators:
<
, <=
, >
, >=
, ==
, /=
. Similarly to arithmetic operations,
the comparisons are in fact defined generically for different
datatypes using traits, which are out of the scope of this
tutorial. For basic usage, one can assume that the comparisons operate
on natural numbers.
For example, one may define the function max3
using >
and max
from the standard library:
max3 (x y z : Nat) : Nat :=
if
| x > y := max x z
| else := max y z;
Local definitions¶
Juvix supports local definitions with let-expressions.
f (a : Nat) : Nat :=
let
x : Nat := a + 5;
y : Nat := a * 7 + x;
in x * y;
The variables x
and y
are not visible outside f
.
One can also use multi-clause definitions in let
-expressions, with the same
syntax as definitions inside a module. For example:
open Bool;
isEven : Nat -> Bool :=
let
isEven' : Nat -> Bool
| zero := true
| (suc n) := isOdd' n;
isOdd' : Nat -> Bool
| zero := false
| (suc n) := isEven' n;
in isEven';
The functions isEven'
and isOdd'
are not visible outside isEven
.
Recursion¶
Juvix is a purely functional language, which means that functions have no side effects and all variables are immutable. An advantage of functional programming is that all expressions are referentially transparent – any expression can be replaced by its value without changing the meaning of the program. This makes it easier to reason about programs, in particular to prove their correctness. No errors involving implicit state are possible, because the state is always explicit.
In a functional language, there are no imperative loops. Repetition is expressed using recursion. In many cases, the recursive definition of a function follows the inductive definition of a data structure the function analyses. For example, consider the following inductive type of lists of natural numbers:
type NList :=
| nnil : NList
| ncons : Nat -> NList -> NList;
An element of NList
is either nnil
(empty) or ncons x xs
where x : Nat
and xs : NList
(a list with head x
and tail xs
).
A function computing the length of a list may be defined by:
nlength : NList -> Nat
| nnil := 0
| (ncons _ xs) := nlength xs + 1;
The definition follows the inductive definition of NList
. There are two
function clauses for the two constructors. The case for nnil
is easy – the
constructor has no arguments and the length of the empty list is 0
. For a
constructor with some arguments, one typically needs to express the result of
the function in terms of the constructor arguments, usually calling the function
recursively on the constructor's inductive arguments (for ncons
this is the
second argument). In the case of ncons _ xs
, we recursively call nlength
on
xs
and add 1
to the result.
Let's consider another example – a function which returns the maximum of the numbers in a list or 0 for the empty list.
nmaximum : NList -> Nat
| nnil := 0
| (ncons x xs) := max x (nmaximum xs);
Again, there is a clause for each constructor. In the case for ncons
, we
recursively call the function on the list tail and take the maximum of the
result and the list head.
For an example of a constructor with more than one inductive argument, consider binary trees with natural numbers in nodes.
type Tree :=
| leaf : Nat -> Tree
| node : Nat -> Tree -> Tree -> Tree;
The constructor node
has two inductive arguments (the second and the third)
which represent the left and the right subtree.
A function which produces the mirror image of a tree may be defined by:
mirror : Tree -> Tree
| (leaf x) := leaf x
| (node x l r) := node x (mirror r) (mirror l);
The definition of mirror
follows the definition of Tree
. There are two
recursive calls for the two inductive constructors of node
(the subtrees).
Partial application and higher-order functions¶
Strictly speaking, all Juvix functions have only one argument. Multi-argument
functions are really functions which return a function which takes the next
argument and returns a function taking another argument, and so on for all
arguments. The function type former ->
(the arrow) is right-associative.
Hence, the type, e.g., Nat -> Nat -> Nat
when fully parenthesised becomes Nat
-> (Nat -> Nat)
. It is the type of functions which given an argument of type
Nat
return a function of type Nat -> Nat
which itself takes an argument of
type Nat
and produces a result of type Nat
. Function application is
left-associative. For example, f a b
when fully parenthesised becomes (f a)
b
. So it is an application to b
of the function obtained by applying f
to
a
.
Since a multi-argument function is just a one-argument function returning a
function, it can be partially applied to a smaller number of arguments than
specified in its definition. The result is an appropriate function. For example,
sub 10
is a function which subtracts its argument from 10
, and (+) 1
is a
function which adds 1
to its argument. If the function has been declared as an
infix operator (like +
), then for partial application one needs to enclose it
in parentheses.
A function which takes a function as an argument is a higher-order function.
An example is the nmap
function which applies a given function to each element
in a list of natural numbers.
nmap (f : Nat -> Nat) : NList -> NList
| nnil := nnil
| (ncons x xs) := ncons (f x) (nmap f xs);
The application
nmap \{x := div x 2} lst
divides every element of lst
by 2
, rounding down the result. The expression
\{x := div x 2}
is an unnamed function, or a lambda, which divides its argument by 2
.
Polymorphism¶
The type NList
we have been working with above requires the list elements to
be natural numbers. It is possible to define lists polymorphically,
parameterising them by the element type. This is similar to generics in
languages like Java, C++ or Rust. Here is the polymorphic definition of lists
from the standard library:
syntax operator :: cons;
type List A :=
| nil : List A
| :: : A -> List A -> List A;
The constructor ::
is declared as a right-associative infix operator. The
definition has a parameter A
which is the element type. Then List Ty
is the
type of lists with elements of type Ty
. For example, List Nat
is the type of
lists of natural numbers, isomorphic to the type NList
defined above.
Now one can define the map
function polymorphically:
map {A B} (f : A -> B) : List A -> List B
| nil := nil
| (h :: hs) := f h :: map f hs;
This function has two implicit type arguments A
and B
. These arguments are
normally omitted in function application – they are inferred automatically
during type checking. The curly braces indicate that the argument is implicit
and should be inferred.
In fact, the constructors nil
and ::
also have an implicit argument: the
type of list elements. All type parameters of a data type definition become
implicit arguments of the constructors.
Usually, the implicit arguments in a function application can be inferred. However, sometimes this is not possible and then the implicit arguments need to be provided explicitly by enclosing them in braces:
f {implArg1} .. {implArgK} arg1 .. argN
For example, nil {Nat}
has type List Nat
while nil
by itself has type {A : Type} -> List A
.
Tail recursion¶
Any recursive call whose result is further processed by the calling function
needs to create a new stack frame to save the calling function environment. This
means that each such call will use a constant amount of memory. For example, a
function sum
implemented as follows will use an additional amount of memory
proportional to the length of the processed list:
sum : List Nat -> Nat
| nil := 0
| (x :: xs) := x + sum xs;
This is not acceptable if you care about performance. In an imperative language, one would use a simple loop going over the list without any memory allocation. In pseudocode:
var sum : Nat := 0;
while (lst /= nil) do
begin
sum := sum + head lst;
lst := tail lst;
end;
result := sum;
Fortunately, it is possible to rewrite this function to use tail recursion. A recursive call is tail recursive if its result is also the result of the calling function, i.e., the calling function returns immediately after it without further processing. The Juvix compiler guarantees that all tail calls will be eliminated, i.e., that they will be compiled to simple jumps without extra memory allocation. In a tail recursive call, instead of creating a new stack frame, the old one is reused.
The following implementation of sum
uses tail recursion.
sum (lst : List Nat) : Nat :=
let
go (acc : Nat) : List Nat -> Nat
| nil := acc
| (x :: xs) := go (acc + x) xs;
in go 0 lst;
The first argument of go
is an accumulator which holds the sum computed so
far. It is analogous to the sum
variable in the imperative loop above. The
initial value of the accumulator is 0. The function go
uses only constant
additional memory overall. The code generated for it by the Juvix compiler is
equivalent to an imperative loop.
Most imperative loops may be translated into tail recursive functional programs
by converting the locally modified variables into accumulators and the loop
condition into pattern matching. For example, here is an imperative pseudocode
for computing the nth Fibonacci number in linear time. The variables cur
and
next
hold the last two computed Fibonacci numbers.
var cur : Nat := 0;
var next : Nat := 1;
while (n /= 0) do
begin
tmp := next;
next := cur + next;
cur := tmp;
n := n - 1;
end;
result := cur;
An equivalent functional program is:
fib : Nat -> Nat :=
let
go (cur next : Nat) : Nat -> Nat
| zero := cur
| (suc n) := go next (cur + next) n;
in go 0 1;
A naive definition of the Fibonacci function runs in exponential time:
fib : Nat -> Nat
| zero := 0
| (suc zero) := 1
| (suc (suc n)) := fib n + fib (suc n);
Tail recursion is less useful when the function needs to allocate memory anyway.
For example, one could make the map
function from the previous section tail
recursive, but the time and memory use would still be proportional to the length
of the input because of the need to allocate the result list. In fact, a tail
recursive map
needs to allocate and discard an intermediate list which is
reversed in the end to preserve the original element order:
map {A B} (f : A -> B) : List A -> List B :=
let
go (acc : List B) : List A -> List B
| nil := reverse acc
| (x :: xs) := go (f x :: acc) xs;
in go nil;
So we have replaced stack allocation with heap allocation. This actually decreases performance.
Conclusion¶
- Use tail recursion to eliminate stack allocation.
- Do not use tail recursion to replace stack allocation with heap allocation.
Iteration over data structures¶
A common use of recursion is to traverse a data structure in a specified order
accumulating some values. For example, the tail recursive sum
function fits
this pattern.
Juvix provides special support for data structure traversals with the iterator
syntax. The standard library defines several list iterators, among them for
and rfor
. We can implement the sum
function using for
:
sum (l : List Nat) : Nat :=
for (acc := 0) (x in l) {
x + acc
};
The above for
iteration starts with the accumulator acc
equal to 0
and
goes through the list l
from left to right (from beginning to end), at each
step updating the accumulator to x + acc
where x
is the current list element
and acc
is the previous accumulator value. The final value of the iteration is
the final value of the accumulator. The for
iterator is tail recursive, i.e.,
no stack memory is allocated and the whole iteration is compiled to a loop.
The rfor
iterator is analogous to for
except that it goes through the list
from right to left (from end to beginning) and is not tail recursive. For
example, one can implement map
using rfor
:
map {A B} (f : A -> B) (l : List A) : List B :=
rfor (acc := nil) (x in l) {
f x :: acc
};
The iterators are just ordinary higher-order Juvix functions which can be used
with the iterator syntax. In fact, the map
function from the standard library
can also be used with the iterator syntax. The expression
map (x in l) {body}
is equivalent to
map \{x := body} l
Whenever possible, it is advised to use the standard library iterators instead
of manually writing recursive functions. When reasonable, for
should be
preferred to rfor
. The iterators provide a readable syntax and the compiler
might be able to optimize them better than manually written recursion.
Totality checking¶
By default, the Juvix compiler requires all functions to be total. Totality consists of:
- termination and coverage for function declarations, and
- strict positivity for user-defined data types.
The termination check ensures that all functions are structurally recursive, i.e., all recursive calls are on structurally smaller values – subpatterns of the matched pattern.
However, we can still make Juvix accept a non-terminating function via the
terminating
keyword, skipping the termination check.
terminating
log2 (n : Nat) : Nat :=
if
| n <= 1 := 0
| else := suc (log2 (div n 2));
Let us look at other examples. The termination checker rejects the
following definition of the factorial function (when the terminating
keyword is removed):
terminating
fact (x : Nat) : Nat :=
if
| x == 0 := 1
| else := x * fact (sub x 1);
To ensure termination, the argument to the recursive call must be a proper subpattern of a pattern matched on in the clause. One can reformulate this definition so that it is accepted by the termination checker:
fact : Nat -> Nat
| zero := 1
| x@(suc n) := x * fact n;
Coverage checking ensures that there are no unhandled patterns in function
clauses or case
expressions. For example, the following definition is rejected
because the case suc zero
is not handled:
isEven : Nat -> Bool
| zero := true
| (suc (suc n)) := isEven n;
Since coverage checking forces the user to specify the function for all input
values, it may be unclear how to implement functions which are typically
partial. For example, the tail
function on lists is often left undefined for
the empty list. One solution is to return a default value. In the Juvix standard
library, tail
is implemented as follows, returning the empty list when the
argument is empty.
tail {A} : List A -> List A
| (_ :: xs) := xs
| nil := nil;
Another solution is to wrap the result in the Maybe
type from the standard
library, which allows representing optional values. An element of Maybe A
is
either nothing
or just x
with x : A
.
type Maybe A :=
| nothing : Maybe A
| just : A -> Maybe A;
For example, one could define the tail function as:
tail {A} : List A -> Maybe (List A)
| (_ :: xs) := just xs
| nil := nothing;
Then the user needs to explicitly check if the result of the function contains a value or not:
case tail' lst of
| just x := ...
| nothing := ...
Exercises¶
You have now learnt essential functional programming techniques in Juvix. To consolidate your understanding, try doing some of the following exercises.
Warm-up exercises¶
Boolean operators¶
Let's start by defining some functions on booleans.
The type for booleans is defined in the standard library like this:
type Bool :=
| true : Bool
| false : Bool;
Remember that you can import this definition by adding import Stdlib.Prelude
open
at the beginning of your module.
Now, define the logical function not
by using pattern matching.
Tip
The type of your function should be:
not : Bool -> Bool;
Now, define the logical functions and
, or
by using pattern matching as well.
Feel free to experiment and see what happens if your patterns are not
exhaustive, i.e., if not all the cases are covered.
Next, let's define the logical function xor
, which should return true
if and
only if exactly one of its arguments is true
. This time, instead of using
pattern matching, use the previously defined logical functions.
Tip
Be wary of using the standard library here as it may cause name conflicts.
Solution
type Bool :=
| true : Bool
| false : Bool;
not : Bool -> Bool
| false := true
| true := false;
or : Bool -> Bool -> Bool
| false b := b
| true := true;
and : Bool -> Bool -> Bool
| true b := b
| false := false;
xor (a b : Bool) : Bool := and (not (and a b)) (or a b);
The Maybe
type¶
The NMaybe
type encapsulates an optional natural number (the preceding N
stands for Nat
). The nnothing
constructor is used when the value is missing.
On the other hand, the njust
constructor is used when the value is present.
type NMaybe :=
| nnothing : NMaybe
| njust : Nat → NMaybe;
Let's define a function isJust : NMaybe -> Bool
that returns true
when the
value is present.
Solution
type NMaybe :=
| nnothing : NMaybe
| njust : Nat → NMaybe;
isJust : NMaybe -> Bool
| (njust _) := true
| nnothing := false;
Now let's define a function fromMaybe : Nat -> NMaybe -> Nat
that given a
NMaybe
, returns its value if present and otherwise returns the first argument
as a default value.
Solution
type NMaybe :=
| nnothing : NMaybe
| njust : Nat → NMaybe;
fromMaybe (d : Nat) : NMaybe -> Nat
| (njust n) := n
| nnothing := d;
It would be useful to have a type that represents optional values of any type.
In Juvix, we can define the polymorphic version of NMaybe
like this:
type Maybe A :=
| nothing : Maybe A
| just : A → Maybe A;
In this definition, we parameterize the type Maybe
with a generic type A
.
Implement again the fromMaybe
function, but now, for the polymorphic Maybe
type. Note that in function definitions we must specify the type variables. The
definition of fromMaybe
begins with:
fromMaybe {A} (d : A) : Maybe A -> A
Give the implementation.
Solution
import Stdlib.Prelude open;
fromMaybe {A} (d : A) : Maybe A -> A
| (just n) := n
| nothing := d;
Neat! It is indeed very easy to define polymorphic functions in Juvix.
To get some more practice, give an implementation for maybe
which begins with:
maybe {A B} (d : B) (f : A -> B) : Maybe A -> B
This should return the value (if present) applied to the function f
. Otherwise
it should return the default value d
.
Solution
import Stdlib.Prelude open;
maybe {A B} (d : B) (f : A -> B) : Maybe A -> B
| (just n) := f n
| nothing := d;
List exercises¶
We can define polymorphic lists as follows:
import Stdlib.Data.Fixity open;
syntax operator :: cons;
type List A :=
| nil : List A
| :: : A -> List A -> List A;
Let's define a function that returns the first element of a List
if it exists.
Is beginning the definition as follows appropriate? If not, why?
head {A} : List A -> A
Try to give an implementation for it.
Solution
As we know, Juvix guarantees that all functions are total.
But we cannot return anything when the list is empty.
Therefore it makes sense to use the Maybe
type that we defined in the previous section.
The proper definition of head
should be:
import Stdlib.Prelude open;
head {A} : List A -> Maybe A
| nil := nothing
| (h :: _) := just h;
So far we have defined only functions that do not involve looping, but any non-trivial program will require some sort of repetition, so let's tackle that.
As stated previously, the only way to express repetition in Juivx is by using recursion. We say that a function is recursive if it is defined in terms of itself, i.e., the name of the function appears in its body.
The next exercise is to define a function which returns the last element of a list. This function will need to call itself until it reaches the last element of the list.
last {A} : List A -> Maybe A;
Solution
import Stdlib.Prelude open;
last {A} : List A -> Maybe A
| nil := nothing
| (x :: nil) := just x
| (_ :: xs) := last xs;
Next, implement a function that concatenates two lists:
concat {A} : List A -> List A -> List A
Tip
It is enough to pattern match the first list.
Solution
import Stdlib.Prelude open;
concat {A} : List A -> List A -> List A
| nil b := b
| (a :: as) b := a :: concat as b;
Now write a function that concatenates a list of lists.
concatMany {A} : List (List A) -> List A
Tip
concat
may be helpful.
Solution
import Stdlib.Prelude open;
concat {A} : List A -> List A -> List A
| nil b := b
| (a :: as) b := a :: concat as b;
concatMany {A} : List (List A) -> List A
| nil := nil
| (a :: as) := concat a (concatMany as);
Can you give an alternative implementation that uses the rfor
iterator? What
would happen if you used for
instead of rfor
?
Solution
import Stdlib.Prelude open;
import Stdlib.Data.Fixity open;
concat {A} : List A -> List A -> List A
| nil b := b
| (a :: as) b := a :: concat as b;
concatMany-iter {A} (m : List (List A)) : List A :=
rfor (acc := nil) (l in m) {
concat l acc
};
In the previous solution, if you replace rfor
by for
, the resulting list
will be as if the original list was reversed, but each of the nested lists keep
their original order.
Write a function that reverses a list:
- using the
for
iterator, - using tail recursion.
Solution
Using the for
iterator:
import Stdlib.Prelude open;
reverse {A} (xs : List A) : List A :=
for (acc := nil) (x in xs) {
x :: acc
};
Using tail recursion:
import Stdlib.Prelude open;
reverse {A} : List A -> List A :=
let
go (acc : List A) : List A -> List A
| nil := acc
| (x :: xs) := go (x :: acc) xs;
in go nil;
Function composition¶
Let's try a different exercise. Define a function compose
that composes two
functions f
and g
. It should take three arguments f
, g
, x
and its only
clause's body should be f (g x)
.
Can you make the compose
function polymorphic and as general as possible?
Hint
The definition should start like this:
compose {A B C} ...
Congratulations! Your warm-up is complete!
More exercises¶
Prime numbers¶
Define a function prime : Nat -> Bool
which checks if a given natural number
is prime.
Tip
A number is prime if it is greater than 1 and has no divisors other than 1 and itself.
Solution
import Stdlib.Prelude open;
prime (x : Nat) : Bool :=
let
go : Nat -> Bool
| zero := true
| (suc zero) := true
| n@(suc k) :=
if
| mod x k == 0 := false
| else := go k;
in case x of
| zero := false
| suc zero := false
| suc k := go k;
Half¶
Does Juvix accept the following definition?
half : Nat -> Nat :=
if
| n < 2 := 0
| else := half (sub n 2) + 1;
If not, how can you reformulate this definition so that it is accepted by Juvix?
Solution
The definition doesn't pass the termination checker. One way to reformulate it is as follows:
import Stdlib.Prelude open;
half : Nat -> Nat
| zero := 0
| (suc zero) := 0
| (suc (suc n)) := half n + 1;
Tree map¶
Recall the Tree
type from above.
import Stdlib.Prelude open;
type Tree :=
| leaf : Nat -> Tree
| node : Nat -> Tree -> Tree -> Tree;
Analogously to the map
function for lists, define a function
tmap : (Nat -> Nat) -> Tree -> Tree;
which applies a function to all natural numbers stored in a tree.
Solution
import Stdlib.Prelude open;
type Tree :=
| leaf : Nat -> Tree
| node : Nat -> Tree -> Tree -> Tree;
tmap (f : Nat -> Nat) : Tree -> Tree
| (leaf x) := leaf (f x)
| (node x l r) := node (f x) (tmap f l) (tmap f r);
Polymorphic tree¶
Modify the Tree
type to be polymorphic in the element type, and then
repeat the previous exercise.
Solution
The Tree
type and the tmap
function need to be made polymorphic in the element types.
type Tree A :=
| leaf : A -> Tree A
| node : A -> Tree A -> Tree A -> Tree A;
tmap {A B} (f : A -> B) : Tree A -> Tree B
| (leaf x) := leaf (f x)
| (node x l r) := node (f x) (tmap f l) (tmap f r);
Note that only the types needed to be changed.
Suffixes¶
A suffix of a list l
is any list which can be obtained from l
by removing
some initial elements. For example, the suffixes of 1 :: 2 :: 3 :: nil
are:
1 :: 2 :: 3 :: nil
,2 :: 3 :: nil
,3 :: nil
, andnil
.
Define a function which computes the list of all suffixes of a given list, arranged in descending order of their lengths.
Solution
import Stdlib.Prelude open;
suffixes {A} : List A -> List (List A)
| nil := nil :: nil
| xs@(_ :: xs') := xs :: suffixes xs';
Factorial¶
Write a tail recursive function which computes the factorial of a natural number.
Solution
import Stdlib.Prelude open;
fact : Nat -> Nat :=
let
go (acc : Nat) : Nat -> Nat
| zero := acc
| n@(suc n') := go (acc * n) n';
in go 1;
List function compose¶
Define a function
comp {A} : List (A -> A) -> A -> A
which composes all functions in a list. For example,
comp (suc :: (*) 2 :: \{x := sub x 1} :: nil)
should be a function which given x
computes 2(x - 1) + 1
.
Solution
import Stdlib.Prelude open;
comp {A} (fs : List (A -> A)) : A -> A :=
for (acc := id) (f in fs) {
f >> acc
};
where >>
is the composition function from the standard library:
import Stdlib.Data.Fixity open;
syntax operator >> composition;
>> {A B C} (f : B -> C) (g : A -> B) (x : A) : C := f (g x);