Copyright | Conor McBride and Ross Paterson 2005 |
---|---|

License | BSD-style (see the LICENSE file in the distribution) |

Maintainer | libraries@haskell.org |

Stability | stable |

Portability | portable |

Safe Haskell | Trustworthy |

Language | Haskell2010 |

Class of data structures that can be traversed from left to right, performing an action on each element. Instances are expected to satisfy the listed laws.

## Synopsis

- class (Functor t, Foldable t) => Traversable t where
- traverse :: Applicative f => (a -> f b) -> t a -> f (t b)
- sequenceA :: Applicative f => t (f a) -> f (t a)
- mapM :: Monad m => (a -> m b) -> t a -> m (t b)
- sequence :: Monad m => t (m a) -> m (t a)

- for :: (Traversable t, Applicative f) => t a -> (a -> f b) -> f (t b)
- forM :: (Traversable t, Monad m) => t a -> (a -> m b) -> m (t b)
- forAccumM :: (Monad m, Traversable t) => s -> t a -> (s -> a -> m (s, b)) -> m (s, t b)
- mapAccumL :: forall t s a b. Traversable t => (s -> a -> (s, b)) -> s -> t a -> (s, t b)
- mapAccumR :: forall t s a b. Traversable t => (s -> a -> (s, b)) -> s -> t a -> (s, t b)
- mapAccumM :: forall m t s a b. (Monad m, Traversable t) => (s -> a -> m (s, b)) -> s -> t a -> m (s, t b)
- fmapDefault :: forall t a b. Traversable t => (a -> b) -> t a -> t b
- foldMapDefault :: forall t m a. (Traversable t, Monoid m) => (a -> m) -> t a -> m

# The `Traversable`

class

class (Functor t, Foldable t) => Traversable t where Source #

Functors representing data structures that can be transformed to
structures of the *same shape* by performing an `Applicative`

(or,
therefore, `Monad`

) action on each element from left to right.

A more detailed description of what *same shape* means, the various methods,
how traversals are constructed, and example advanced use-cases can be found
in the **Overview** section of Data.Traversable.

For the class laws see the **Laws** section of Data.Traversable.

traverse :: Applicative f => (a -> f b) -> t a -> f (t b) Source #

Map each element of a structure to an action, evaluate these actions
from left to right, and collect the results. For a version that ignores
the results see `traverse_`

.

#### Examples

Basic usage:

In the first two examples we show each evaluated action mapping to the output structure.

`>>>`

Just [1,2,3,4]`traverse Just [1,2,3,4]`

`>>>`

Right [1,2,3,4]`traverse id [Right 1, Right 2, Right 3, Right 4]`

In the next examples, we show that `Nothing`

and `Left`

values short
circuit the created structure.

`>>>`

Nothing`traverse (const Nothing) [1,2,3,4]`

`>>>`

Nothing`traverse (\x -> if odd x then Just x else Nothing) [1,2,3,4]`

`>>>`

Left 0`traverse id [Right 1, Right 2, Right 3, Right 4, Left 0]`

sequenceA :: Applicative f => t (f a) -> f (t a) Source #

Evaluate each action in the structure from left to right, and
collect the results. For a version that ignores the results
see `sequenceA_`

.

#### Examples

Basic usage:

For the first two examples we show sequenceA fully evaluating a a structure and collecting the results.

`>>>`

Just [1,2,3]`sequenceA [Just 1, Just 2, Just 3]`

`>>>`

Right [1,2,3]`sequenceA [Right 1, Right 2, Right 3]`

The next two example show `Nothing`

and `Just`

will short circuit
the resulting structure if present in the input. For more context,
check the `Traversable`

instances for `Either`

and `Maybe`

.

`>>>`

Nothing`sequenceA [Just 1, Just 2, Just 3, Nothing]`

`>>>`

Left 4`sequenceA [Right 1, Right 2, Right 3, Left 4]`

mapM :: Monad m => (a -> m b) -> t a -> m (t b) Source #

Map each element of a structure to a monadic action, evaluate
these actions from left to right, and collect the results. For
a version that ignores the results see `mapM_`

.

#### Examples

sequence :: Monad m => t (m a) -> m (t a) Source #

Evaluate each monadic action in the structure from left to
right, and collect the results. For a version that ignores the
results see `sequence_`

.

#### Examples

Basic usage:

The first two examples are instances where the input and
and output of `sequence`

are isomorphic.

`>>>`

[Right 1,Right 2,Right 3,Right 4]`sequence $ Right [1,2,3,4]`

`>>>`

Right [1,2,3,4]`sequence $ [Right 1,Right 2,Right 3,Right 4]`

The following examples demonstrate short circuit behavior
for `sequence`

.

`>>>`

Left [1,2,3,4]`sequence $ Left [1,2,3,4]`

`>>>`

Left 0`sequence $ [Left 0, Right 1,Right 2,Right 3,Right 4]`

#### Instances

# Utility functions

for :: (Traversable t, Applicative f) => t a -> (a -> f b) -> f (t b) Source #

forM :: (Traversable t, Monad m) => t a -> (a -> m b) -> m (t b) Source #

forAccumM :: (Monad m, Traversable t) => s -> t a -> (s -> a -> m (s, b)) -> m (s, t b) Source #

mapAccumL :: forall t s a b. Traversable t => (s -> a -> (s, b)) -> s -> t a -> (s, t b) Source #

The `mapAccumL`

function behaves like a combination of `fmap`

and `foldl`

; it applies a function to each element of a structure,
passing an accumulating parameter from left to right, and returning
a final value of this accumulator together with the new structure.

#### Examples

Basic usage:

`>>>`

(55,[0,1,3,6,10,15,21,28,36,45])`mapAccumL (\a b -> (a + b, a)) 0 [1..10]`

`>>>`

("012345",["0","01","012","0123","01234"])`mapAccumL (\a b -> (a <> show b, a)) "0" [1..5]`

mapAccumR :: forall t s a b. Traversable t => (s -> a -> (s, b)) -> s -> t a -> (s, t b) Source #

The `mapAccumR`

function behaves like a combination of `fmap`

and `foldr`

; it applies a function to each element of a structure,
passing an accumulating parameter from right to left, and returning
a final value of this accumulator together with the new structure.

#### Examples

Basic usage:

`>>>`

(55,[54,52,49,45,40,34,27,19,10,0])`mapAccumR (\a b -> (a + b, a)) 0 [1..10]`

`>>>`

("054321",["05432","0543","054","05","0"])`mapAccumR (\a b -> (a <> show b, a)) "0" [1..5]`

mapAccumM :: forall m t s a b. (Monad m, Traversable t) => (s -> a -> m (s, b)) -> s -> t a -> m (s, t b) Source #

The `mapAccumM`

function behaves like a combination of `mapM`

and
`mapAccumL`

that traverses the structure while evaluating the actions
and passing an accumulating parameter from left to right.
It returns a final value of this accumulator together with the new structure.
The accummulator is often used for caching the intermediate results of a computation.

#### Examples

Basic usage:

`>>>`

`let expensiveDouble a = putStrLn ("Doubling " <> show a) >> pure (2 * a)`

`>>>`

mapAccumM (\cache a -> case lookup a cache of Nothing -> expensiveDouble a >>= \double -> pure ((a, double):cache, double) Just double -> pure (cache, double) ) [] [1, 2, 3, 1, 2, 3] :} Doubling 1 Doubling 2 Doubling 3 ([(3,6),(2,4),(1,2)],[2,4,6,2,4,6])`:{`

*Since: base-4.18.0.0*

# General definitions for superclass methods

fmapDefault :: forall t a b. Traversable t => (a -> b) -> t a -> t b Source #

This function may be used as a value for `fmap`

in a `Functor`

instance, provided that `traverse`

is defined. (Using
`fmapDefault`

with a `Traversable`

instance defined only by
`sequenceA`

will result in infinite recursion.)

`fmapDefault`

f ≡`runIdentity`

.`traverse`

(`Identity`

. f)

foldMapDefault :: forall t m a. (Traversable t, Monoid m) => (a -> m) -> t a -> m Source #

# Overview

Traversable structures support element-wise sequencing of `Applicative`

effects (thus also `Monad`

effects) to construct new structures of
**the same shape** as the input.

To illustrate what is meant by *same shape*, if the input structure is
** [a]**, each output structure is a list

**of the same length as the input. If the input is a**

`[b]`

**, each output**

`Tree a`

**has the same graph of intermediate nodes and leaves. Similarly, if the input is a 2-tuple**

`Tree b`

**, each output is a 2-tuple**

`(x, a)`

**, and so forth.**

`(x, b)`

It is in fact possible to decompose a traversable structure ** t a** into
its shape (a.k.a.

*spine*) of type

**and its element list**

`t ()`

**. The original structure can be faithfully reconstructed from its spine and element list.**

`[a]`

The implementation of a `Traversable`

instance for a given structure follows
naturally from its type; see the Construction section for
details.
Instances must satisfy the laws listed in the Laws section.
The diverse uses of `Traversable`

structures result from the many possible
choices of Applicative effects.
See the Advanced Traversals section for some examples.

Every `Traversable`

structure is both a `Functor`

and `Foldable`

because it
is possible to implement the requisite instances in terms of `traverse`

by
using `fmapDefault`

for `fmap`

and `foldMapDefault`

for `foldMap`

. Direct
fine-tuned implementations of these superclass methods can in some cases be
more efficient.

## The `traverse`

and `mapM`

methods

For an `Applicative`

functor ** f** and a

`Traversable`

functor **, the type signatures of**

`t`

`traverse`

and `fmap`

are rather similar:fmap :: (a -> f b) -> t a -> t (f b) traverse :: (a -> f b) -> t a -> f (t b)

The key difference is that `fmap`

produces a structure whose elements (of
type ** f b**) are individual effects, while

`traverse`

produces an
aggregate effect yielding structures of type **.**

`t b`

For example, when ** f** is the

**monad, and**

`IO`

**is**

`t`

**,**

`List`

`fmap`

yields a list of IO actions, whereas `traverse`

constructs an IO
action that evaluates to a list of the return values of the individual
actions performed left-to-right.traverse :: (a -> IO b) -> [a] -> IO [b]

The `mapM`

function is a specialisation of `traverse`

to the case when
** f** is a

`Monad`

. For monads, `mapM`

is more idiomatic than `traverse`

.
The two are otherwise generally identical (though `mapM`

may be specifically
optimised for monads, and could be more efficient than using the more
general `traverse`

).traverse :: (Applicative f, Traversable t) => (a -> f b) -> t a -> f (t b) mapM :: (Monad m, Traversable t) => (a -> m b) -> t a -> m (t b)

When the traversable term is a simple variable or expression, and the
monadic action to run is a non-trivial do block, it can be more natural to
write the action last. This idiom is supported by `for`

, `forM`

, and
`forAccumM`

which are the flipped versions of `traverse`

, `mapM`

, and
`mapAccumM`

respectively.

### Their `Foldable`

, just the effects, analogues.

The `traverse`

and `mapM`

methods have analogues in the Data.Foldable
module. These are `traverse_`

and `mapM_`

, and their flipped variants
`for_`

and `forM_`

, respectively. The result type is ** f ()**, they don't
return an updated structure, and can be used to sequence effects over all
the elements of a

`Traversable`

(any `Foldable`

) structure just for their
side-effects.If the `Traversable`

structure is empty, the result is ** pure ()**. When
effects short-circuit, the

**result may, for example, be**

`f ()`

`Nothing`

if **is**

`f`

`Maybe`

, or **when it is**

`Left`

e

**.**

`Either`

e

It is perhaps worth noting that `Maybe`

is not only a potential
`Applicative`

functor for the return value of the first argument of
`traverse`

, but is also itself a `Traversable`

structure with either zero or
one element. A convenient idiom for conditionally executing an action just
for its effects on a `Just`

value, and doing nothing otherwise is:

-- action :: Monad m => a -> m () -- mvalue :: Maybe a mapM_ action mvalue -- :: m ()

which is more concise than:

maybe (return ()) action mvalue

The `mapM_`

idiom works verbatim if the type of ** mvalue** is later
refactored from

**to**

`Maybe a`

**(assuming it remains OK to silently do nothing in the**

`Either e a`

`Left`

case).### Result multiplicity

When `traverse`

or `mapM`

is applied to an empty structure ** ts** (one for
which

**is**

`null`

ts

`True`

) the return value is **regardless of the provided function**

`pure ts`

**. It is not possible to apply the function when no values of type**

`g :: a -> f b`

**are available, but its type determines the relevant instance of**

`a`

`pure`

.null ts ==> traverse g ts == pure ts

Otherwise, when ** ts** is non-empty and at least one value of type

**results from each**

`b`

**, the structures**

`f a`

**have**

`t b`

*the same shape*(list length, graph of tree nodes, ...) as the input structure

**, but the slots previously occupied by elements of type**

`t a`

**now hold elements of type**

`a`

**.**

`b`

A single traversal may produce one, zero or many such structures. The zero
case happens when one of the effects ** f a** sequenced as part of the
traversal yields no replacement values. Otherwise, the many case happens
when one of sequenced effects yields multiple values.

The `traverse`

function does not perform selective filtering of slots in the
output structure as with e.g. `mapMaybe`

.

`>>>`

`let incOdd n = if odd n then Just $ n + 1 else Nothing`

`>>>`

[2,4]`mapMaybe incOdd [1, 2, 3]`

`>>>`

Just [2,4,6]`traverse incOdd [1, 3, 5]`

`>>>`

Nothing`traverse incOdd [1, 2, 3]`

In the above examples, with `Maybe`

as the `Applicative`

** f**, we see
that the number of

**structures produced by**

`t b`

`traverse`

may differ
from one: it is zero when the result short-circuits to **. The same can happen when**

`Nothing`

**is**

`f`

**and the result is**

`List`

**, or**

`[]`

**is**

`f`

**and the result is**

`Either e`

**, or perhaps the**

`Left (x :: e)`

`empty`

value of some
`Alternative`

functor.When ** f** is e.g.

**, and the map**

`List`

**returns more than one value for some inputs**

`g :: a -> [b]`

**(and at least one for all**

`a`

**), the result of**

`a`

**will contain multiple structures of the same shape as**

`mapM g ts`

**:**

`ts`

length (mapM g ts) == product (fmap (length . g) ts)

For example:

`>>>`

720`length $ mapM (\n -> [1..n]) [1..6]`

`>>>`

720`product $ length . (\n -> [1..n]) <$> [1..6]`

In other words, a traversal with a function ** g :: a -> [b]**, over an
input structure

**, yields a list**

`t a`

**, whose length is the product of the lengths of the lists that**

`[t b]`

`g`

returns for each element of the
input structure! The individual elements **of the structure are replaced by each element of**

`a`

**in turn:**

`g a`

`>>>`

[Just 1,Just 2,Just 3]`mapM (\n -> [1..n]) $ Just 3`

`>>>`

[[1,1,1],[1,1,2],[1,1,3],[1,2,1],[1,2,2],[1,2,3]]`mapM (\n -> [1..n]) [1..3]`

If any element of the structure ** t a** is mapped by

`g`

to an empty list,
then the entire aggregate result is empty, because no value is available to
fill one of the slots of the output structure:`>>>`

[]`mapM (\n -> [1..n]) $ [0..6] -- [1..0] is empty`

## The `sequenceA`

and `sequence`

methods

The `sequenceA`

and `sequence`

methods are useful when what you have is a
container of pending applicative or monadic effects, and you want to combine
them into a single effect that produces zero or more containers with the
computed values.

sequenceA :: (Applicative f, Traversable t) => t (f a) -> f (t a) sequence :: (Monad m, Traversable t) => t (m a) -> m (t a) sequenceA = traverse id -- default definition sequence = sequenceA -- default definition

When the monad ** m** is

`IO`

, applying `sequence`

to a list of
IO actions, performs each in turn, returning a list of the results:sequence [putStr "Hello ", putStrLn "World!"] = (\a b -> [a,b]) <$> putStr "Hello " <*> putStrLn "World!" = do u1 <- putStr "Hello " u2 <- putStrLn "World!" return [u1, u2] -- In this case [(), ()]

For `sequenceA`

, the *non-deterministic* behaviour of `List`

is most easily
seen in the case of a list of lists (of elements of some common fixed type).
The result is a cross-product of all the sublists:

`>>>`

[[0,30,500],[0,40,500],[1,30,500],[1,40,500],[2,30,500],[2,40,500]]`sequenceA [[0, 1, 2], [30, 40], [500]]`

Because the input list has three (sublist) elements, the result is a list of
triples (*same shape*).

### Care with default method implementations

The `traverse`

method has a default implementation in terms of `sequenceA`

:

traverse g = sequenceA . fmap g

but relying on this default implementation is not recommended, it requires
that the structure is already independently a `Functor`

. The definition of
`sequenceA`

in terms of ** traverse id** is much simpler than

`traverse`

expressed via a composition of `sequenceA`

and `fmap`

. Instances should
generally implement `traverse`

explicitly. It may in some cases also make
sense to implement a specialised `mapM`

.Because `fmapDefault`

is defined in terms of `traverse`

(whose default
definition in terms of `sequenceA`

uses `fmap`

), you must not use
`fmapDefault`

to define the `Functor`

instance if the `Traversable`

instance
directly defines only `sequenceA`

.

### Monadic short circuits

When the monad ** m** is

`Either`

or `Maybe`

(more generally any
`MonadPlus`

), the effect in question is to short-circuit the
result on encountering `Left`

or `Nothing`

(more generally
`mzero`

).`>>>`

Just [1,2,3]`sequence [Just 1,Just 2,Just 3]`

`>>>`

Nothing`sequence [Just 1,Nothing,Just 3]`

`>>>`

Right [1,2,3]`sequence [Right 1,Right 2,Right 3]`

`>>>`

Left "sorry"`sequence [Right 1,Left "sorry",Right 3]`

The result of `sequence`

is all-or-nothing, either structures of exactly the
same shape as the input or none at all. The `sequence`

function does not
perform selective filtering as with e.g. `catMaybes`

or
`rights`

:

`>>>`

[1,3]`catMaybes [Just 1,Nothing,Just 3]`

`>>>`

[1,3]`rights [Right 1,Left "sorry",Right 3]`

## Example binary tree instance

The definition of a `Traversable`

instance for a binary tree is rather
similar to the corresponding instance of `Functor`

, given the data type:

data Tree a = Empty | Leaf a | Node (Tree a) a (Tree a)

a canonical `Functor`

instance would be

instance Functor Tree where fmap g Empty = Empty fmap g (Leaf x) = Leaf (g x) fmap g (Node l k r) = Node (fmap g l) (g k) (fmap g r)

a canonical `Traversable`

instance would be

instance Traversable Tree where traverse g Empty = pure Empty traverse g (Leaf x) = Leaf <$> g x traverse g (Node l k r) = Node <$> traverse g l <*> g k <*> traverse g r

This definition works for any ** g :: a -> f b**, with

**an Applicative functor, as the laws for**

`f`

`(``<*>`

)

imply the requisite
associativity.We can add an explicit non-default `mapM`

if desired:

mapM g Empty = return Empty mapM g (Leaf x) = Leaf <$> g x mapM g (Node l k r) = do ml <- mapM g l mk <- g k mr <- mapM g r return $ Node ml mk mr

See Construction below for a more detailed exploration of
the general case, but as mentioned in Overview above, instance
definitions are typically rather simple, all the interesting behaviour is a
result of an interesting choice of `Applicative`

functor for a traversal.

### Pre-order and post-order tree traversal

It is perhaps worth noting that the traversal defined above gives an
*in-order* sequencing of the elements. If instead you want either
*pre-order* (parent first, then child nodes) or post-order (child nodes
first, then parent) sequencing, you can define the instance accordingly:

inOrderNode :: Tree a -> a -> Tree a -> Tree a inOrderNode l x r = Node l x r preOrderNode :: a -> Tree a -> Tree a -> Tree a preOrderNode x l r = Node l x r postOrderNode :: Tree a -> Tree a -> a -> Tree a postOrderNode l r x = Node l x r -- Traversable instance with in-order traversal instance Traversable Tree where traverse g t = case t of Empty -> pure Empty Leaf x -> Leaf <$> g x Node l x r -> inOrderNode <$> traverse g l <*> g x <*> traverse g r -- Traversable instance with pre-order traversal instance Traversable Tree where traverse g t = case t of Empty -> pure Empty Leaf x -> Leaf <$> g x Node l x r -> preOrderNode <$> g x <*> traverse g l <*> traverse g r -- Traversable instance with post-order traversal instance Traversable Tree where traverse g t = case t of Empty -> pure Empty Leaf x -> Leaf <$> g x Node l x r -> postOrderNode <$> traverse g l <*> traverse g r <*> g x

Since the same underlying Tree structure is used in all three cases, it is
possible to use `newtype`

wrappers to make all three available at the same
time! The user need only wrap the root of the tree in the appropriate
`newtype`

for the desired traversal order. Tne associated instance
definitions are shown below (see coercion if unfamiliar with
the use of `coerce`

in the sample code):

{-# LANGUAGE ScopedTypeVariables, TypeApplications #-} -- Default in-order traversal import Data.Coerce (coerce) import Data.Traversable data Tree a = Empty | Leaf a | Node (Tree a) a (Tree a) instance Functor Tree where fmap = fmapDefault instance Foldable Tree where foldMap = foldMapDefault instance Traversable Tree where traverse _ Empty = pure Empty traverse g (Leaf a) = Leaf <$> g a traverse g (Node l a r) = Node <$> traverse g l <*> g a <*> traverse g r -- Optional pre-order traversal newtype PreOrderTree a = PreOrderTree (Tree a) instance Functor PreOrderTree where fmap = fmapDefault instance Foldable PreOrderTree where foldMap = foldMapDefault instance Traversable PreOrderTree where traverse _ (PreOrderTree Empty) = pure $ preOrderEmpty traverse g (PreOrderTree (Leaf x)) = preOrderLeaf <$> g x traverse g (PreOrderTree (Node l x r)) = preOrderNode <$> g x <*> traverse g (coerce l) <*> traverse g (coerce r) preOrderEmpty :: forall a. PreOrderTree a preOrderEmpty = coerce (Empty @a) preOrderLeaf :: forall a. a -> PreOrderTree a preOrderLeaf = coerce (Leaf @a) preOrderNode :: a -> PreOrderTree a -> PreOrderTree a -> PreOrderTree a preOrderNode x l r = coerce (Node (coerce l) x (coerce r)) -- Optional post-order traversal newtype PostOrderTree a = PostOrderTree (Tree a) instance Functor PostOrderTree where fmap = fmapDefault instance Foldable PostOrderTree where foldMap = foldMapDefault instance Traversable PostOrderTree where traverse _ (PostOrderTree Empty) = pure postOrderEmpty traverse g (PostOrderTree (Leaf x)) = postOrderLeaf <$> g x traverse g (PostOrderTree (Node l x r)) = postOrderNode <$> traverse g (coerce l) <*> traverse g (coerce r) <*> g x postOrderEmpty :: forall a. PostOrderTree a postOrderEmpty = coerce (Empty @a) postOrderLeaf :: forall a. a -> PostOrderTree a postOrderLeaf = coerce (Leaf @a) postOrderNode :: PostOrderTree a -> PostOrderTree a -> a -> PostOrderTree a postOrderNode l r x = coerce (Node (coerce l) x (coerce r))

With the above, given a sample tree:

inOrder :: Tree Int inOrder = Node (Node (Leaf 10) 3 (Leaf 20)) 5 (Leaf 42)

we have:

import Data.Foldable (toList) print $ toList inOrder [10,3,20,5,42] print $ toList (coerce inOrder :: PreOrderTree Int) [5,3,10,20,42] print $ toList (coerce inOrder :: PostOrderTree Int) [10,20,3,42,5]

You would typically define instances for additional common type classes,
such as `Eq`

, `Ord`

, `Show`

, etc.

## Making construction intuitive

In order to be able to reason about how a given type of `Applicative`

effects will be sequenced through a general `Traversable`

structure by its
`traversable`

and related methods, it is helpful to look more closely
at how a general `traverse`

method is implemented. We'll look at how
general traversals are constructed primarily with a view to being able
to predict their behaviour as a user, even if you're not defining your
own `Traversable`

instances.

Traversable structures ** t a** are assembled incrementally from their
constituent parts, perhaps by prepending or appending individual elements of
type

**, or, more generally, by recursively combining smaller composite traversable building blocks that contain multiple such elements.**

`a`

As in the tree example above, the components being combined are
typically pieced together by a suitable *constructor*, i.e. a function
taking two or more arguments that returns a composite value.

The `traverse`

method enriches simple incremental construction with
threading of `Applicative`

effects of some function ** g :: a -> f b**.

The basic building blocks we'll use to model the construction of `traverse`

are a hypothetical set of elementary functions, some of which may have
direct analogues in specific `Traversable`

structures. For example, the
** (:)** constructor is an analogue for lists of

`prepend`

or the more
general `combine`

.empty :: t a -- build an empty container singleton :: a -> t a -- build a one-element container prepend :: a -> t a -> t a -- extend by prepending a new initial element append :: t a -> a -> t a -- extend by appending a new final element combine :: a1 -> a2 -> ... -> an -> t a -- combine multiple inputs

An empty structure has no elements of type

, so there's nothing to which`a`

can be applied, but since we need an output of type`g`

, we just use the`f (t b)`

`pure`

instance ofto wrap an empty of type`f`

:`t b`

traverse _ (empty :: t a) = pure (empty :: t b)

With the List monad,

*empty*is, while with`[]`

`Maybe`

it is`Nothing`

. Withwe have an`Either e a`

*empty*case for each value of:`e`

traverse _ (Left e :: Either e a) = pure $ (Left e :: Either e b)

A singleton structure has just one element of type

, and`a`

`traverse`

can take that, apply`a`

getting an`g :: a -> f b`

, then`f b`

over that, getting an`fmap singleton`

as required:`f (t b)`

traverse g (singleton a) = fmap singleton $ g a

Note that if

is`f`

and`List`

returns multiple values the result will be a list of multiple`g`

singletons!`t b`

Since

`Maybe`

and`Either`

are either empty or singletons, we havetraverse _ Nothing = pure Nothing traverse g (Just a) = Just <$> g a

traverse _ (Left e) = pure (Left e) traverse g (Right a) = Right <$> g a

For

`List`

, empty isand`[]`

`singleton`

is, so we have:`(:[])`

traverse _ [] = pure [] traverse g [a] = fmap (:[]) (g a) = (:) <$> (g a) <*> traverse g [] = liftA2 (:) (g a) (traverse g [])

When the structure is built by adding one more element via

or`prepend`

, traversal amounts to:`append`

traverse g (prepend a t0) = prepend <$> (g a) <*> traverse g t0 = liftA2 prepend (g a) (traverse g t0)

traverse g (append t0 a) = append <$> traverse g t0 <*> g a = liftA2 append (traverse g t0) (g a)

The origin of the combinatorial product when

is`f`

`List`

should now be apparent, whenhas`traverse g t0`

elements and`n`

has`g a`

elements, the`m`

*non-deterministic*`Applicative`

instance of`List`

will produce a result withelements.`m * n`

When combining larger building blocks, we again use

to combine the traversals of the components. With bare elements`(`

`<*>`

)mapped to`a`

via`f b`

, and composite traversable sub-structures transformed via`g`

:`traverse g`

traverse g (combine a1 a2 ... an) = combine <$> t1 <*> t2 <*> ... <*> tn where t1 = g a1 -- if a1 fills a slot of type @a@ = traverse g a1 -- if a1 is a traversable substructure ... ditto for the remaining constructor arguments ...

The above definitions sequence the `Applicative`

effects of ** f** in the
expected order while producing results of the expected shape

**.**

`t`

For lists this becomes:

traverse g [] = pure [] traverse g (x:xs) = liftA2 (:) (g a) (traverse g xs)

The actual definition of `traverse`

for lists is an equivalent
right fold in order to facilitate list *fusion*.

traverse g = foldr (\x r -> liftA2 (:) (g x) r) (pure [])

# Advanced traversals

In the sections below we'll examine some advanced choices of `Applicative`

effects that give rise to very different transformations of `Traversable`

structures.

These examples cover the implementations of `fmapDefault`

, `foldMapDefault`

,
`mapAccumL`

and `mapAccumR`

functions illustrating the use of `Identity`

,
`Const`

and stateful `Applicative`

effects. The ZipList example
illustrates the use of a less-well known `Applicative`

instance for lists.

This is optional material, which is not essential to a basic understanding of
`Traversable`

structures. If this is your first encounter with `Traversable`

structures, you can come back to these at a later date.

### Coercion

Some of the examples make use of an advanced Haskell feature, namely
`newtype`

*coercion*. This is done for two reasons:

- Use of
`coerce`

makes it possible to avoid cluttering the code with functions that wrap and unwrap*newtype*terms, which at runtime are indistinguishable from the underlying value. Coercion is particularly convenient when one would have to otherwise apply multiple newtype constructors to function arguments, and then peel off multiple layers of same from the function output. - Use of
`coerce`

can produce more efficient code, by reusing the original value, rather than allocating space for a wrapped clone.

If you're not familiar with `coerce`

, don't worry, it is just a shorthand
that, e.g., given:

newtype Foo a = MkFoo { getFoo :: a } newtype Bar a = MkBar { getBar :: a } newtype Baz a = MkBaz { getBaz :: a } f :: Baz Int -> Bar (Foo String)

makes it possible to write:

x :: Int -> String x = coerce f

instead of

x = getFoo . getBar . f . MkBaz

## Identity: the `fmapDefault`

function

The simplest Applicative functor is `Identity`

, which just wraps and unwraps
pure values and function application. This allows us to define
`fmapDefault`

:

{-# LANGUAGE ScopedTypeVariables, TypeApplications #-} import Data.Coercible (coerce) fmapDefault :: forall t a b. Traversable t => (a -> b) -> t a -> t b fmapDefault = coerce (traverse @t @Identity @a @b)

The use of coercion avoids the need to explicitly wrap and
unwrap terms via `Identity`

and `runIdentity`

.

As noted in Overview, `fmapDefault`

can only be used to define
the requisite `Functor`

instance of a `Traversable`

structure when the
`traverse`

method is explicitly implemented. An infinite loop would result
if in addition `traverse`

were defined in terms of `sequenceA`

and `fmap`

.

## State: the `mapAccumL`

, `mapAccumR`

functions

Applicative functors that thread a changing state through a computation are
an interesting use-case for `traverse`

. The `mapAccumL`

and `mapAccumR`

functions in this module are each defined in terms of such traversals.

We first define a simplified (not a monad transformer) version of
`State`

that threads a state ** s** through a
chain of computations left to right. Its

`(``<*>`

)

operator passes the
input state first to its left argument, and then the resulting state is
passed to its right argument, which returns the final state.newtype StateL s a = StateL { runStateL :: s -> (s, a) } instance Functor (StateL s) where fmap f (StateL kx) = StateL $ \ s -> let (s', x) = kx s in (s', f x) instance Applicative (StateL s) where pure a = StateL $ \s -> (s, a) (StateL kf) <*> (StateL kx) = StateL $ \ s -> let { (s', f) = kf s ; (s'', x) = kx s' } in (s'', f x) liftA2 f (StateL kx) (StateL ky) = StateL $ \ s -> let { (s', x) = kx s ; (s'', y) = ky s' } in (s'', f x y)

With `StateL`

, we can define `mapAccumL`

as follows:

{-# LANGUAGE ScopedTypeVariables, TypeApplications #-} mapAccumL :: forall t s a b. Traversable t => (s -> a -> (s, b)) -> s -> t a -> (s, t b) mapAccumL g s ts = coerce (traverse @t @(StateL s) @a @b) (flip g) ts s

The use of coercion avoids the need to explicitly wrap and
unwrap ** newtype** terms.

The type of ** flip g** is coercible to

**, which makes it suitable for use with**

`a -> StateL b`

`traverse`

. As part of the Applicative
construction of **the state updates will thread left-to-right along the sequence of elements of**

`StateL (t b)`

**.**

`t a`

While `mapAccumR`

has a type signature identical to `mapAccumL`

, it differs
in the expected order of evaluation of effects, which must take place
right-to-left.

For this we need a variant control structure `StateR`

, which threads the
state right-to-left, by passing the input state to its right argument and
then using the resulting state as an input to its left argument:

newtype StateR s a = StateR { runStateR :: s -> (s, a) } instance Functor (StateR s) where fmap f (StateR kx) = StateR $ \s -> let (s', x) = kx s in (s', f x) instance Applicative (StateR s) where pure a = StateR $ \s -> (s, a) (StateR kf) <*> (StateR kx) = StateR $ \ s -> let { (s', x) = kx s ; (s'', f) = kf s' } in (s'', f x) liftA2 f (StateR kx) (StateR ky) = StateR $ \ s -> let { (s', y) = ky s ; (s'', x) = kx s' } in (s'', f x y)

With `StateR`

, we can define `mapAccumR`

as follows:

{-# LANGUAGE ScopedTypeVariables, TypeApplications #-} mapAccumR :: forall t s a b. Traversable t => (s -> a -> (s, b)) -> s -> t a -> (s, t b) mapAccumR g s0 ts = coerce (traverse @t @(StateR s) @a @b) (flip g) ts s0

The use of coercion avoids the need to explicitly wrap and
unwrap ** newtype** terms.

Various stateful traversals can be constructed from `mapAccumL`

and
`mapAccumR`

for suitable choices of `g`

, or built directly along similar
lines.

## Const: the `foldMapDefault`

function

The `Const`

Functor enables applications of `traverse`

that summarise the
input structure to an output value without constructing any output values
of the same type or shape.

As noted above, the `Foldable`

superclass constraint is
justified by the fact that it is possible to construct `foldMap`

, `foldr`

,
etc., from `traverse`

. The technique used is useful in its own right, and
is explored below.

A key feature of folds is that they can reduce the input structure to a summary value. Often neither the input structure nor a mutated clone is needed once the fold is computed, and through list fusion the input may not even have been memory resident in its entirety at the same time.

The `traverse`

method does not at first seem to be a suitable building block
for folds, because its return value ** f (t b)** appears to retain mutated
copies of the input structure. But the presence of

**in the type signature need not mean that terms of type**

`t b`

**are actually embedded in**

`t b`

**. The simplest way to elide the excess terms is by basing the Applicative functor used with**

`f (t b)`

`traverse`

on `Const`

.Not only does ** Const a b** hold just an

**value, with the**

`a`

**parameter merely a**

`b`

*phantom*type, but when

**has a**

`m`

`Monoid`

instance,
**is an**

`Const m`

`Applicative`

functor:import Data.Coerce (coerce) newtype Const a b = Const { getConst :: a } deriving (Eq, Ord, Show) -- etc. instance Functor (Const m) where fmap = const coerce instance Monoid m => Applicative (Const m) where pure _ = Const mempty (<*>) = coerce (mappend :: m -> m -> m) liftA2 _ = coerce (mappend :: m -> m -> m)

The use of coercion avoids the need to explicitly wrap and
unwrap ** newtype** terms.

We can therefore define a specialisation of `traverse`

:

{-# LANGUAGE ScopedTypeVariables, TypeApplications #-} traverseC :: forall t a m. (Monoid m, Traversable t) => (a -> Const m ()) -> t a -> Const m (t ()) traverseC = traverse @t @(Const m) @a @()

For which the Applicative construction of `traverse`

leads to:

null ts ==> traverseC g ts = Const mempty

traverseC g (prepend x xs) = Const (g x) <> traverseC g xs

In other words, this makes it possible to define:

{-# LANGUAGE ScopedTypeVariables, TypeApplications #-} foldMapDefault :: forall t a m. (Monoid m, Traversable t) => (a -> m) -> t a -> m foldMapDefault = coerce (traverse @t @(Const m) @a @())

Which is sufficient to define a `Foldable`

superclass instance:

The use of coercion avoids the need to explicitly wrap and
unwrap ** newtype** terms.

instance Traversable t => Foldable t where foldMap = foldMapDefault

It may however be instructive to also directly define candidate default
implementations of `foldr`

and `foldl'`

, which take a bit more machinery
to construct:

{-# LANGUAGE ScopedTypeVariables, TypeApplications #-} import Data.Coerce (coerce) import Data.Functor.Const (Const(..)) import Data.Semigroup (Dual(..), Endo(..)) import GHC.Exts (oneShot) foldrDefault :: forall t a b. Traversable t => (a -> b -> b) -> b -> t a -> b foldrDefault f z = \t -> coerce (traverse @t @(Const (Endo b)) @a @()) f t z foldlDefault' :: forall t a b. Traversable t => (b -> a -> b) -> b -> t a -> b foldlDefault' f z = \t -> coerce (traverse @t @(Const (Dual (Endo b))) @a @()) f' t z where f' :: a -> b -> b f' a = oneShot $ \ b -> b `seq` f b a

In the above we're using the `Endo`

b`Monoid`

and its
`Dual`

to compose a sequence of ** b -> b** accumulator updates in either
left-to-right or right-to-left order.

The use of `seq`

in the definition of ** foldlDefault'** ensures strictness
in the accumulator.

The use of coercion avoids the need to explicitly wrap and
unwrap ** newtype** terms.

The `oneShot`

function gives a hint to the compiler that aids in
correct optimisation of lambda terms that fire at most once (for each
element ** a**) and so should not try to pre-compute and re-use
subexpressions that pay off only on repeated execution. Otherwise, it is
just the identity function.

## ZipList: transposing lists of lists

As a warm-up for looking at the `ZipList`

`Applicative`

functor, we'll first
look at a simpler analogue. First define a fixed width 2-element `Vec2`

type, whose `Applicative`

instance combines a pair of functions with a pair of
values by applying each function to the corresponding value slot:

data Vec2 a = Vec2 a a instance Functor Vec2 where fmap f (Vec2 a b) = Vec2 (f a) (f b) instance Applicative Vec2 where pure x = Vec2 x x liftA2 f (Vec2 a b) (Vec2 p q) = Vec2 (f a p) (f b q) instance Foldable Vec2 where foldr f z (Vec2 a b) = f a (f b z) foldMap f (Vec2 a b) = f a <> f b instance Traversable Vec2 where traverse f (Vec2 a b) = Vec2 <$> f a <*> f b

Along with a similar definition for fixed width 3-element vectors:

data Vec3 a = Vec3 a a a instance Functor Vec3 where fmap f (Vec3 x y z) = Vec3 (f x) (f y) (f z) instance Applicative Vec3 where pure x = Vec3 x x x liftA2 f (Vec3 p q r) (Vec3 x y z) = Vec3 (f p x) (f q y) (f r z) instance Foldable Vec3 where foldr f z (Vec3 a b c) = f a (f b (f c z)) foldMap f (Vec3 a b c) = f a <> f b <> f c instance Traversable Vec3 where traverse f (Vec3 a b c) = Vec3 <$> f a <*> f b <*> f c

With the above definitions,

(same as `sequenceA`

) acts
as a `traverse`

`id`

*matrix transpose* operation on `Vec2 (Vec3 Int)`

producing a
corresponding `Vec3 (Vec2 Int)`

:

Let ** t = Vec2 (Vec3 1 2 3) (Vec3 4 5 6)** be our

`Traversable`

structure,
and **be the function used to traverse**

`g = id :: Vec3 Int -> Vec3 Int`

**. We then have:**

`t`

traverse g t = Vec2 <$> (Vec3 1 2 3) <*> (Vec3 4 5 6) = Vec3 (Vec2 1 4) (Vec2 2 5) (Vec2 3 6)

This construction can be generalised from fixed width vectors to variable
length lists via `ZipList`

. This gives a transpose
operation that works well for lists of equal length. If some of the lists
are longer than others, they're truncated to the longest common length.

We've already looked at the standard `Applicative`

instance of `List`

for
which applying ** m** functions

**to**

`f1, f2, ..., fm`

**input values**

`n`

**produces**

`a1, a2, ..., an`

**outputs:**

`m * n`

`>>>`

`:set -XTupleSections`

`>>>`

[("f1",1),("f1",2),("f2",1),("f2",2),("f3",1),("f3",2)]`[("f1",), ("f2",), ("f3",)] <*> [1,2]`

There are however two more common ways to turn lists into `Applicative`

control structures. The first is via **Const [a]**, since lists are
monoids under concatenation, and we've already seen that

**is an**

`Const`

m

`Applicative`

functor when **is a**

`m`

`Monoid`

. The second, is based
on `zipWith`

, and is called `ZipList`

:{-# LANGUAGE GeneralizedNewtypeDeriving #-} newtype ZipList a = ZipList { getZipList :: [a] } deriving (Show, Eq, ..., Functor) instance Applicative ZipList where liftA2 f (ZipList xs) (ZipList ys) = ZipList $ zipWith f xs ys pure x = repeat x

The `liftA2`

definition is clear enough, instead of applying ** f** to each
pair

**drawn independently from the**

`(x, y)`

**and**

`xs`

**, only corresponding pairs at each index in the two lists are used.**

`ys`

The definition of `pure`

may look surprising, but it is needed to ensure
that the instance is lawful:

liftA2 f (pure x) ys == fmap (f x) ys

Since ** ys** can have any length, we need to provide an infinite supply
of

**values in**

`x`

**in order to have a value to pair with each element**

`pure x`

**.**

`y`

When `ZipList`

is the `Applicative`

functor used in the
construction of a traversal, a ZipList holding a partially
built structure with ** m** elements is combined with a component holding

**elements via**

`n`

`zipWith`

, resulting in **outputs!**

`min m n`

Therefore `traverse`

with ** g :: a -> ZipList b** will produce a

`ZipList`

of **structures whose element count is the minimum length of the ZipLists**

`t b`

**with**

`g a`

**ranging over the elements of**

`a`

**. When**

`t`

**is empty, the length is infinite (as expected for a minimum of an empty set).**

`t`

If the structure ** t** holds values of type

**, we can use the identity function**

`ZipList a`

**for the first argument of**

`id :: ZipList a -> ZipList a`

`traverse`

:traverse (id :: ZipList a -> ZipList a) :: t (ZipList a) -> ZipList (t a)

The number of elements in the output `ZipList`

will be the length of the
shortest `ZipList`

element of ** t**. Each output

**will have the**

`t a`

*same shape*as the input

**, i.e. will share its number of elements.**

`t (ZipList a)`

If we think of the elements of ** t (ZipList a)** as its rows, and the
elements of each individual

`ZipList`

as the columns of that row, we see
that our traversal implements a *transpose*operation swapping the rows and columns of

**, after first truncating all the rows to the column count of the shortest one.**

`t`

Since in fact **traverse id** is just

`sequenceA`

the above boils down
to a rather concise definition of *transpose*, with coercion used to implicitly wrap and unwrap the

`ZipList`

`newtype`

as needed, giving
a function that operates on a list of lists:`>>>`

`{-# LANGUAGE ScopedTypeVariables #-}`

`>>>`

`import Control.Applicative (ZipList(..))`

`>>>`

`import Data.Coerce (coerce)`

`>>>`

`>>>`

`transpose :: forall a. [[a]] -> [[a]]`

`>>>`

`transpose = coerce (sequenceA :: [ZipList a] -> ZipList [a])`

`>>>`

`>>>`

[[1,4,7],[2,5,8],[3,6,9]]`transpose [[1,2,3],[4..],[7..]]`

The use of coercion avoids the need to explicitly wrap and
unwrap ** ZipList** terms.

# Laws

A definition of `traverse`

must satisfy the following laws:

- Naturality
`t .`

for every applicative transformation`traverse`

f =`traverse`

(t . f)`t`

- Identity
`traverse`

`Identity`

=`Identity`

- Composition
`traverse`

(`Compose`

.`fmap`

g . f) =`Compose`

.`fmap`

(`traverse`

g) .`traverse`

f

A definition of `sequenceA`

must satisfy the following laws:

- Naturality
`t .`

for every applicative transformation`sequenceA`

=`sequenceA`

.`fmap`

t`t`

- Identity
`sequenceA`

.`fmap`

`Identity`

=`Identity`

- Composition
`sequenceA`

.`fmap`

`Compose`

=`Compose`

.`fmap`

`sequenceA`

.`sequenceA`

where an *applicative transformation* is a function

t :: (Applicative f, Applicative g) => f a -> g a

preserving the `Applicative`

operations, i.e.

t (`pure`

x) =`pure`

x t (f`<*>`

x) = t f`<*>`

t x

and the identity functor `Identity`

and composition functors
`Compose`

are from Data.Functor.Identity and
Data.Functor.Compose.

A result of the naturality law is a purity law for `traverse`

`traverse`

`pure`

=`pure`

The superclass instances should satisfy the following:

- In the
`Functor`

instance,`fmap`

should be equivalent to traversal with the identity applicative functor (`fmapDefault`

). - In the
`Foldable`

instance,`foldMap`

should be equivalent to traversal with a constant applicative functor (`foldMapDefault`

).

Note: the `Functor`

superclass means that (in GHC) Traversable structures
cannot impose any constraints on the element type. A Haskell implementation
that supports constrained functors could make it possible to define
constrained `Traversable`

structures.

# See also

- "The Essence of the Iterator Pattern",
by Jeremy Gibbons and Bruno Oliveira,
in
*Mathematically-Structured Functional Programming*, 2006, online at http://www.cs.ox.ac.uk/people/jeremy.gibbons/publications/#iterator. - "Applicative Programming with Effects",
by Conor McBride and Ross Paterson,
*Journal of Functional Programming*18:1 (2008) 1-13, online at http://www.soi.city.ac.uk/~ross/papers/Applicative.html. - "An Investigation of the Laws of Traversals",
by Mauro Jaskelioff and Ondrej Rypacek,
in
*Mathematically-Structured Functional Programming*, 2012, online at http://arxiv.org/pdf/1202.2919.