Proving equivalence of sequence definitions from Applicative and Monad - list

How can I properly prove that
sequenceA :: (Traversable t, Applicative f) => t (f a) -> f (t a)
sequenceA [] = pure []
sequenceA (x:xs) = pure (:) <*> x <*> sequenceA xs
is essentially the same to monad inputs as
sequenceA' :: Monad m => [m a] -> m [a]
sequenceA' [] = return []
sequenceA' (x:xs) = do
x' <- x
xs' <- sequenceA' xs
return (x':xs')
In spite of the constraint Applicative and Monad of course.

Here's a proof sketch:
Show that
do
x' <- x
xs' <- sequenceA' xs
return (x' : xs')
is equivalent to
do
f1 <- do
cons <- return (:)
x' <- x
return (cons x')
xs' <- sequenceA' xs
return (f1 xs')
This involves desugaring (and resugaring) do notation and applying the Monad laws.
Use the definition of ap:
ap m1 m2 = do { x1 <- m1; x2 <- m2; return (x1 x2) }
to turn the above code into
do
f1 <- return (:) `ap` x
xs' <- sequenceA' xs
return (f1 xs')
and then
return (:) `ap` x `ap` sequenceA' xs
Now you have
sequenceA' [] = return []
sequenceA' (x:xs) = return (:) `ap` x `ap` sequenceA' xs
Assume that pure and <*> are essentially the same as return and `ap`, respectively, and you're done.
This latter property is also stated in the Applicative documentation:
If f is also a Monad, it should satisfy
pure = return
(<*>) = ap

Since the Functor-Applicative-Monad proposal, implemented in GHC 7.10, Applicative is a superclass of Monad. So even though your two functions can't be strictly equivalent, since sequenceA's domain includes sequenceA''s domain, we can look at what happens in this common domain (the Monad typeclass).
This paper shows an interesting demonstration of desugaring do notation to applicative and functor operations (<$>, pure and <*>). If the expressions on the right hand side of your left-pointing arrows (<-) don't depend on each other, as is the case in your question, you can always use applicative operations, and therefore show that your hypothesis is correct (for the Monad domain).
Also have a look at the ApplicativeDo language extension proposal, which contains an example that's just like yours:
do
x <- a
y <- b
return (f x y)
which translates to:
(\x y -> f x y) <$> a <*> b
Substituting f for (:), we get:
do
x <- a
y <- b
return (x : y)
... which translates to...
(\x y -> x : y) <$> a <*> b
--And by eta reduction
(:) <$> a <*> b
--Which is equivalent to the code in your question (albeit more general):
pure (:) <*> a <*> b
Alternatively, you can make GHC's desugarer work for you by using the ApplicativeDo language extension and by following this answer to the SO question "haskell - Desugaring do-notation for Monads". I'll leave this exercise up to you (as it honestly goes beyond my capacities!).

My own two cents
There is no do notation for applicatives in Haskell. It can be seen specifically in this segment.
return and pure do exactly the same, but with different constraints, right?, so this part pure (:) and this part return (x:xs) are essentially the same.
Then, here x <- act you are getting the value of act, and then the value of the recursion xs <- seqn acts, to finally wrap it with the return.
And that's what pure (:) <*> x <*> sequenceA xs is essentially doing.

Related

Haskell: concat a list of tuples with an element and a list: [(a,[b])] -> [(a,b)]

I want to concat in a list of tuples of elements and Strings, each char with the element.
For example: [(True, "xy"), (False, "abc")] 􏰁-> [(True,’x’),(True,’y’),(False,’a’), (False,’b’),(False,’c’)]
I do have a solution, but im wondering if there is a better one:
concatsplit :: [(a,[b])] -> [(a,b)]
concatsplit a = concatMap (\(x,y)-> concatsplit' (x,y)) a
concatsplit' :: (a,[b]) -> [(a,b)]
concatsplit' y = map (\x -> ((fst y),x)) (snd y)
Why not a simple list comprehension?
List comprehensions can often do as much as higher order functions, and I think that if you don't need to change the data but only "unpack" it, they are pretty clear. Here's a working example:
concS :: [(a,[b])] -> [(a,b)]
concS ls = [(a,b) | (a,x) <- ls, b <- x]
The other answers show how to do this idiomatically from scratch, which I like a lot. It might also be interesting to show how you might polish what you've already got. Here it is again as a reminder:
concatsplit a = concatMap (\(x,y)-> concatsplit' (x,y)) a
concatsplit' y = map (\x -> ((fst y),x)) (snd y)
The first thing I'd consistently change is called "eta reduction", and it is when you turn something of the shape \x -> foo x into just foo. We can do this in the argument to concatMap to get
concatsplit a = concatMap concatsplit' a
and then again in the argument to concatsplit to get:
concatsplit = concatMap concatsplit'
Looking at concatsplit', the thing I like least is the use of fst and snd instead of pattern matching. With pattern matching, it looks like this:
concatsplit' (a,bs) = map (\x -> (a,x)) bs
If you really wanted to practice your eta reduction, you might notice that (,) can be applied prefix and change this to
concatsplit' (a,bs) = map (\x -> (,) a x) bs
= map ((,) a) bs
but I think I'm just as happy one way or the other. At this point, this definition is small enough that I'd be tempted to inline it into concatsplit itself, getting:
concatsplit = concatMap (\(a,bs) -> map ((,) a) bs)
This looks like a pretty good definition to me, and I'd stop there.
You might be bugged by the almost-eta-reduction here: it's almost the right shape to drop the bs. An advanced user might go on, noticing that:
uncurry (\a bs -> map ((,) a) bs) = \(a,bs) -> map ((,) a) bs
Hence with some eta reduction and other point-free techniques, we could transition this way:
concatsplit = concatMap (uncurry (\a bs -> map ((,) a) bs))
= concatMap (uncurry (\a -> map ((,) a)))
= concatMap (uncurry (map . (,)))
But, personally, I find this less readable than where I declared I would stop above, namely:
concatsplit = concatMap (\(a,bs) -> map ((,) a) bs)
sequenceA makes it really easy:
concatsplit = concatMap sequenceA
Or generalize it even further:
concatsplit = (>>= sequenceA)
Details:
sequenceA has type (Applicative f, Traversable t) => t (f a) -> f (t a). This means that if you have a type with a Traversable on the "outside" and an Applicative on the "inside", you can call sequenceA on it to turn it inside-out, so that the Applicative is on the "outside" and the Traversable is on the "inside".
In this case, (True, "xy") has type (Bool, [Char]), which desugars to (,) Bool ([] Char). There's an instance of Traversable for ((,) a), and there's an instance of Applicative for []. Thus, you can call sequenceA on it, and the result will be of type [] ((,) Bool Char), or [(Bool, Char)] with sugar.
As it turns out, not only does sequenceA have a useful type, but it does exactly the thing you need here (and turns out to be exactly equivalent to concatsplit').
concatMap has type Foldable t => (a -> [b]) -> t a -> [b]. (>>=) has type Monad m => m a -> (a -> m b) -> m b. When specialized to lists, these become the same, except with their arguments in the opposite order.
You can also use the monad instance for lists:
concatSplit l = l >>= \(a,x) -> x >>= \b -> return (a,b)
Which can be simplified to:
concatSplit l = l >>= \(a,x) -> map ((,) a) x
And reformatted to do notation
concatSplit l = do
(a,x) <- l
map ((,) a) x

Haskell List Comprehensions with Arbitrary number of Generators

So what I have so far is something like this:
combs :: [[Char]]
combs = [[i] ++ [j] ++ [k] ++ [l] | i <- x, j <- x, k <- x, l <- x]
where x = "abc"
So this is the working function for n = 4, is there any way to make this work for an arbitrary number of generators? I could program in for n = 1, 2, 3 etc.. but ideally need it to work for any given n. For reference, x is just an arbitrary string of unique characters. I'm struggling to think of a way to somehow extract it to work for n generators.
You can use replicateM:
replicateM :: Applicative m => Int -> m a -> m [a]
E.g.:
generate :: Num a => Int -> [[a]]
generate = flip replicateM [1,2,3]
to generate all possiible lists of a given length and consisting of elements 1..3.
As far as I know, you can not construct list comprehension with an arbitrary number of generators, but usually if you do something with arbitrary depth, recursion is the way to do it.
So we have to think of solving this, in terms of itself. If you want all possible strings that can be generated with the characters in x. In case n = 0, we can generate exactly one string: the empty string.
combs 0 = [""]
so a list with one element [].
Now in case we want to generate strings with one characters, we can of course simply return x:
combs 1 = x
and now the question is what to do in case n > 1. In that case we can obtain all the strings with length n-1, and and for each such string, and each such character in x, produce a new string. Like:
combs n = [ (c:cs) | c <- x, cs <- combs (n-1) ]
Note that this makes the second case (n = 1) redundant. We can pick a character c from x, and prepend that to the empty string. So a basic implementation is:
combs :: Int -> [[Char]]
combs 0 = [""]
combs n = [(c:cs) | c <- x, cs <- combs (n-1)]
where x = "abc"
Now we can still look for improvements. List comprehensions are basically syntactical sugar for the list monad. So we can use liftA2 here:
import Control.Applicative(liftA2)
combs :: Int -> [[Char]]
combs 0 = [""]
combs n = liftA2 (:) x (combs (n-1))
where x = "abc"
we probably also want to make the set of characters a parameter:
import Control.Applicative(liftA2)
combs :: [Char] -> Int -> [[Char]]
combs _ 0 = [""]
combs x n = liftA2 (:) x (combs (n-1))
and we do not have to restrict us to characters, we can produce a certesian power for all possible types:
import Control.Applicative(liftA2)
combs :: [a] -> Int -> [[a]]
combs _ 0 = [[]]
combs x n = liftA2 (:) x (combs (n-1))
First I would translate the comprehension as a monadic expression.
x >>= \i -> x >>= \j -> x >>= \k -> x >>= \l -> return [i,j,k,l]
With n = 4 we see we have 4 x's, and generally will have n x's. Therefore, I am thinking about a list of x's of length n.
[x,x,x,x] :: [[a]]
How might we go from [x,x,x,x] to the monadic expression? A first good guess is foldr, since we want to do something with each element of the list. Particularly, we want to take an element from each x and form a list with these elements.
foldr :: (a -> b -> b) -> b -> [a] -> b
-- Or more accurately for our scenario:
foldr :: ([a] -> [[a]] -> [[a]]) -> [[a]] -> [[a]] -> [[a]]
There are two terms to come up with for foldr, which I will call f :: [a] -> [[a]] -> [[a]] and z :: [[a]]. We know what foldr f z [x,x,x,x] is:
foldr f z [x,x,x,x] = f x (f x (f x (f x z)))
If we add parentheses to the earlier monadic expression, we have this:
x >>= \i -> (x >>= \j -> (x >>= \k -> (x >>= \l -> return [i,j,k,l])))
You can see how the two expressions are looking similar. We should be able to find an f and z to make them the same. If we choose f = \x a -> x >>= \x' -> a >>= \a' -> return (x' : a') we get:
f x (f x (f x (f x z)))
= (\x a -> a >>= \a' -> x >>= \x' -> return (x' : a')) x (f x (f x (f x z)))
= f x (f x (f x z)) >>= \a' -> x >>= \x' -> return (x' : a')
= f x (f x (f x z)) >>= \a' -> x >>= \l -> return (l : a')
= (f x (f x z) >>= \a' -> x >>= \k -> return (k : a')) >>= \a' -> x >>= \l -> return (l : a')
= f x (f x z) >>= \a' -> x >>= \k -> x >>= \l -> return (l : k : a')
Note that I have reversed the order of i,j,k,l to l,k,j,i but in context of finding combinations, this should be irrelevant. We could try a' ++ [x'] instead if it was really of concern.
The last step is because (a >>= \b -> c) >>= \d -> e is the same as a >>= \b -> c >>= \d -> e (when accounting for variable hygiene) and return a >>= \b -> c is the same as (\b -> c) a.
If we keep unfolding this expression, eventually we will reach z >>= \a' -> … on the front. The only choice that makes sense here then is z = [[]]. This means that foldr f z [] = [[]] which may not be desirable (preferring [] instead). Instead, we might use foldr1 (for non-empty lists, and we might use Data.NonEmpty) or we might add a separate clause for empty lists to combs.
Looking at f = \x a -> x >>= \x' -> a >>= \a' -> return (x' : a') we might realise this helpful equivalence: a >>= \b -> return (c b) = c <$> a. Therefore, f = \x a -> x >>= \x' -> (x' :) <$> a. Then also, a >>= \b -> c (g b) = g <$> a >>= \b -> c and so f = (:) <$> x >>= \x' -> x' <$> a. Finally, a <*> b = a >>= \x -> x <$> b and so f = (:) <$> x <*> a.
The official implementation of sequenceA for lists is foldr (\x a -> (:) <$> x <*> a) (pure []), exactly what we came up with here too. This can be further shortened as foldr (liftA2 (:)) (pure []) but there is possibly some optimisation difference that made the implementors not choose this.
Last step is to merely come up with a list of n x's. This is just replicate replicate n x. There happens to be a function which does both replication and sequencing, called replicateM replicateM n x.

Apply a function to every second element in a list

I'd like to apply a function to every second element in a list:
> mapToEverySecond (*2) [1..10]
[1,4,3,8,5,12,7,16,9,20]
I've written the following function:
mapToEverySecond :: (a -> a) -> [a] -> [a]
mapToEverySecond f l = map (\(i,x) -> if odd i then f x else x) $ zip [0..] l
This works, but I wonder if there is a more idiomatic way to do things like that.
I haven't written very much Haskell, but here's the first thing that came into mind:
func :: (a -> a) -> [a] -> [a]
func f [] = []
func f [x] = [x]
func f (x:s:xs) = x:(f s):(func f xs)
It is a little ulgy, since you have to not only take care of the empty list, but also the list with one element. This doesn't really scale well either (what if you want every third, or
One could do as #Landei points out, and write
func :: (a -> a) -> [a] -> [a]
func f (x:s:xs) = x:(f s):(func f xs)
func f xs = xs
In order to get rid of the ugly checks for both [] and [x], though, IMHO, this makes it a little harder to read (at least the first time).
Here is how I would do it:
mapOnlyOddNumbered f [] = []
mapOnlyOddNumbered f (x:xs) = f x : mapOnlyEvenNumbered f xs
mapOnlyEvenNumbered f [] = []
mapOnlyEvenNumbered f (x:xs) = x : mapOnlyOddNumbered f xs
Whether this is "idiomatic" is a matter of opinion (and I would have given it as a comment if it would fit there) , but it may be useful to see a number of different approaches. Your solution is just as valid as mine, or the ones in the comments, and easier to change into say mapOnlyEvery13nd or mapOnlyPrimeNumbered
mapToEverySecond = zipWith ($) (cycle [id, (*2)])
Is the smallest I can think of, also looks pretty clear in my opinion. It also kinda scales with every nth.
Edit: Oh, people already suggested it in comments. I don't want to steal it, but I really think this is the answer.
Here's how I would probably do it:
mapToEverySecond f xs = foldr go (`seq` []) xs False
where
go x cont !mapThisTime =
(if mapThisTime then f x else x) : cont (not mapThisTime)
But if I were writing library code, I'd probably wrap that up in a build form.
Edit
Yes, this can also be done using mapAccumL or traverse.
import Control.Applicative
import Control.Monad.Trans.State.Strict
import Data.Traversable (Traversable (traverse), mapAccumL)
mapToEverySecond :: Traversable t => (a -> a) -> t a -> t a
-- Either
mapToEverySecond f = snd . flip mapAccumL False
(\mapThisTime x ->
if mapThisTime
then (False, f x)
else (True, x))
-- or
mapToEverySecond f xs = evalState (traverse step xs) False
where
step x = do
mapThisTime <- get
put (not mapThisTime)
if mapThisTime then return (f x) else return x
Or you can do it with scanl, which I'll leave for you to figure out.
This is more a comment to #MartinHaTh's answer. I'd slightly optimize his solution to
func :: (a -> a) -> [a] -> [a]
func f = loop
where
loop [] = []
loop [x] = [x]
loop (x:s:xs) = x : f s : loop xs
Not very elegant, but this is my take:
mapToEverySecond f = reverse . fst . foldl' cmb ([], False) where
cmb (xs, b) x = ((if b then f else id) x : xs, not b)
Or improving on MartinHaTh's answer:
mapToEverySecond f (x : x' : xs) = x : f x' : mapToEverySecond f xs
mapToEverySecond _ xs = xs

Lists defined as Maybe in Haskell? Why not?

You don't offen see Maybe List except for error-handling for example, because lists are a bit Maybe themselves: they have their own "Nothing": [] and their own "Just": (:).
I wrote a list type using Maybe and functions to convert standard and to "experimental" lists. toStd . toExp == id.
data List a = List a (Maybe (List a))
deriving (Eq, Show, Read)
toExp [] = Nothing
toExp (x:xs) = Just (List x (toExp xs))
toStd Nothing = []
toStd (Just (List x xs)) = x : (toStd xs)
What do you think about it, as an attempt to reduce repetition, to generalize?
Trees too could be defined using these lists:
type Tree a = List (Tree a, Tree a)
I haven't tested this last piece of code, though.
All ADTs are isomorphic (almost--see end) to some combination of (,),Either,(),(->),Void and Mu where
data Void --using empty data decls or
newtype Void = Void Void
and Mu computes the fixpoint of a functor
newtype Mu f = Mu (f (Mu f))
so for example
data [a] = [] | (a:[a])
is the same as
data [a] = Mu (ListF a)
data ListF a f = End | Pair a f
which itself is isomorphic to
newtype ListF a f = ListF (Either () (a,f))
since
data Maybe a = Nothing | Just a
is isomorphic to
newtype Maybe a = Maybe (Either () a)
you have
newtype ListF a f = ListF (Maybe (a,f))
which can be inlined in the mu to
data List a = List (Maybe (a,List a))
and your definition
data List a = List a (Maybe (List a))
is just the unfolding of the Mu and elimination of the outer Maybe (corresponding to non-empty lists)
and you are done...
a couple of things
Using custom ADTs increases clarity and type safety
This universality is useful: see GHC.Generic
Okay, I said almost isomorphic. It is not exactly, namely
hmm = List (Just undefined)
has no equivalent value in the [a] = [] | (a:[a]) definition of lists. This is because Haskell data types are coinductive, and has been a point of criticism of the lazy evaluation model. You can get around these problems by only using strict sums and products (and call by value functions), and adding a special "Lazy" data constructor
data SPair a b = SPair !a !b
data SEither a b = SLeft !a | SRight !b
data Lazy a = Lazy a --Note, this has no obvious encoding in Pure CBV languages,
--although Laza a = (() -> a) is semantically correct,
--it is strictly less efficient than Haskell's CB-Need
and then all the isomorphisms can be faithfully encoded.
You can define lists in a bunch of ways in Haskell. For example, as functions:
{-# LANGUAGE RankNTypes #-}
newtype List a = List { runList :: forall b. (a -> b -> b) -> b -> b }
nil :: List a
nil = List (\_ z -> z )
cons :: a -> List a -> List a
cons x xs = List (\f z -> f x (runList xs f z))
isNil :: List a -> Bool
isNil xs = runList xs (\x xs -> False) True
head :: List a -> a
head xs = runList xs (\x xs -> x) (error "empty list")
tail :: List a -> List a
tail xs | isNil xs = error "empty list"
tail xs = fst (runList xs go (nil, nil))
where go x (xs, xs') = (xs', cons x xs)
foldr :: (a -> b -> b) -> b -> List a -> b
foldr f z xs = runList xs f z
The trick to this implementation is that lists are being represented as functions that execute a fold over the elements of the list:
fromNative :: [a] -> List a
fromNative xs = List (\f z -> foldr f z xs)
toNative :: List a -> [a]
toNative xs = runList xs (:) []
In any case, what really matters is the contract (or laws) that the type and its operations follow, and the performance of implementation. Basically, any implementation that fulfills the contract will give you correct programs, and faster implementations will give you faster programs.
What is the contract of lists? Well, I'm not going to express it in complete detail, but lists obey statements like these:
head (x:xs) == x
tail (x:xs) == xs
[] == []
[] /= x:xs
If xs == ys and x == y, then x:xs == y:ys
foldr f z [] == z
foldr f z (x:xs) == f x (foldr f z xs)
EDIT: And to tie this to augustss' answer:
newtype ExpList a = ExpList (Maybe (a, ExpList a))
toExpList :: List a -> ExpList a
toExpList xs = runList xs (\x xs -> ExpList (Just (x, xs))) (ExpList Nothing)
foldExpList f z (ExpList Nothing) = z
foldExpList f z (ExpList (Just (head, taill))) = f head (foldExpList f z tail)
fromExpList :: ExpList a -> List a
fromExpList xs = List (\f z -> foldExpList f z xs)
You could define lists in terms of Maybe, but not that way do. Your List type cannot be empty. Or did you intend Maybe (List a) to be the replacement of [a]. This seems bad since it doesn't distinguish the list and maybe types.
This would work
newtype List a = List (Maybe (a, List a))
This has some problems. First using this would be more verbose than usual lists, and second, the domain is not isomorphic to lists since we got a pair in there (which can be undefined; adding an extra level in the domain).
If it's a list, it should be an instance of Functor, right?
instance Functor List
where fmap f (List a as) = List (f a) (mapMaybeList f as)
mapMaybeList :: (a -> b) -> Maybe (List a) -> Maybe (List b)
mapMaybeList f as = fmap (fmap f) as
Here's a problem: you can make List an instance of Functor, but your Maybe List is not: even if Maybe was not already an instance of Functor in its own right, you can't directly make a construction like Maybe . List into an instance of anything (you'd need a wrapper type).
Similarly for other typeclasses.
Having said that, with your formulation you can do this, which you can't do with standard Haskell lists:
instance Comonad List
where extract (List a _) = a
duplicate x # (List _ y) = List x (duplicate y)
A Maybe List still wouldn't be comonadic though.
When I first started using Haskell, I too tried to represent things in existing types as much as I could on the grounds that it's good to avoid redundancy. My current understanding (moving target!) tends to involve more the idea of a multidimensional web of trade-offs. I won't be giving any “answer” here so much as pasting examples and asking “do you see what I mean?” I hope it helps anyway.
Let's have a look at a bit of Darcs code:
data UseCache = YesUseCache | NoUseCache
deriving ( Eq )
data DryRun = YesDryRun | NoDryRun
deriving ( Eq )
data Compression = NoCompression
| GzipCompression
deriving ( Eq )
Did you notice that these three types could all have been Bool's? Why do you think the Darcs hackers decided that they should introduce this sort of redundancy in their code? As another example, here is a piece of code we changed a few years back:
type Slot = Maybe Bool -- OLD code
data Slot = InFirst | InMiddle | InLast -- newer code
Why do you think we decided that the second code was an improvement over the first?
Finally, here is a bit of code from some of my day job stuff. It uses the newtype syntax that augustss mentioned,
newtype Role = Role { fromRole :: Text }
deriving (Eq, Ord)
newtype KmClass = KmClass { fromKmClass :: Text }
deriving (Eq, Ord)
newtype Lemma = Lemma { fromLemma :: Text }
deriving (Eq, Ord)
Here you'll notice that I've done the curious thing of taking a perfectly good Text type and then wrapping it up into three different things. The three things don't have any new features compared to plain old Text. They're just there to be different. To be honest, I'm not entirely sure if it was a good idea for me to do this. I provisionally think it was because I manipulate lots of different bits and pieces of text for lots of reasons, but time will tell.
Can you see what I'm trying to get at?

Implementing filter using HoF in Haskell

I'm trying to write a function that takes a predicate f and a list and returns a list consisting of all items that satisfy f with preserved order. The trick is to do this using only higher order functions (HoF), no recursion, no comprehensions, and of course no filter.
You can express filter in terms of foldr:
filter p = foldr (\x xs-> if p x then x:xs else xs) []
I think you can use map this way:
filter' :: (a -> Bool) -> [a] -> [a]
filter' p xs = concat (map (\x -> if (p x) then [x] else []) xs)
You see? Convert the list in a list of lists, where if the element you want doesn't pass p, it turns to an empty list
filter' (> 1) [1 , 2, 3 ] would be: concat [ [], [2], [3]] = [2,3]
In prelude there is concatMap that makes the code simplier :P
the code should look like:
filter' :: (a -> Bool) -> [a] -> [a]
filter' p xs = concatMap (\x -> if (p x) then [x] else []) xs
using foldr, as suggested by sclv, can be done with something like this:
filter'' :: (a -> Bool) -> [a] -> [a]
filter'' p xs = foldr (\x y -> if p x then (x:y) else y) [] xs
You're obviously doing this to learn, so let me show you something cool. First up, to refresh our minds, the type of filter is:
filter :: (a -> Bool) -> [a] -> [a]
The interesting part of this is the last bit [a] -> [a]. It breaks down one list and it builds up a new list.
Recursive patterns are so common in Haskell (and other functional languages) that people have come up with names for some of these patterns. The simplest are the catamorphism and it's dual the anamorphism. I'll show you how this relates to your immediate problem at the end.
Fixed points
Prerequisite knowledge FTW!
What is the type of Nothing? Firing up GHCI, it says Nothing :: Maybe a and I wouldn't disagree. What about Just Nothing? Using GHCI again, it says Just Nothing :: Maybe (Maybe a) which is also perfectly valid, but what about the value that this a Nothing embedded within an arbitrary number, or even an infinite number, of Justs. ie, what is the type of this value:
foo = Just foo
Haskell doesn't actually allow such a definition, but with a slight tweak we can make such a type:
data Fix a = In { out :: a (Fix a) }
just :: Fix Maybe -> Fix Maybe
just = In . Just
nothing :: Fix Maybe
nothing = In Nothing
foo :: Fix Maybe
foo = just foo
Wooh, close enough! Using the same type, we can create arbitrarily nested nothings:
bar :: Fix Maybe
bar = just (just (just (just nothing)))
Aside: Peano arithmetic anyone?
fromInt :: Int -> Fix Maybe
fromInt 0 = nothing
fromInt n = just $ fromInt (n - 1)
toInt :: Fix Maybe -> Int
toInt (In Nothing) = 0
toInt (In (Just x)) = 1 + toInt x
This Fix Maybe type is a bit boring. Here's a type whose fixed-point is a list:
data L a r = Nil | Cons a r
type List a = Fix (L a)
This data type is going to be instrumental in demonstrating some recursion patterns.
Useful Fact: The r in Cons a r is called a recursion site
Catamorphism
A catamorphism is an operation that breaks a structure down. The catamorphism for lists is better known as a fold. Now the type of a catamorphism can be expressed like so:
cata :: (T a -> a) -> Fix T -> a
Which can be written equivalently as:
cata :: (T a -> a) -> (Fix T -> a)
Or in English as:
You give me a function that reduces a data type to a value and I'll give you a function that reduces it's fixed point to a value.
Actually, I lied, the type is really:
cata :: Functor T => (T a -> a) -> Fix T -> a
But the principle is the same. Notice, T is only parameterized over the type of the recursion sites, so the Functor part is really saying "Give me a way of manipulating all the recursion sites".
Then cata can be defined as:
cata f = f . fmap (cata f) . out
This is quite dense, let me elaborate. It's a three step process:
First, We're given a Fix t, which is a difficult type to play with, we can make it easier by applying out (from the definition of Fix) giving us a t (Fix t).
Next we want to convert the t (Fix t) into a t a, which we can do, via wishful thinking, using fmap (cata f); we're assuming we'll be able to construct cata.
Lastly, we have a t a and we want an a, so we just use f.
Earlier I said that the catamorphism for a list is called fold, but cata doesn't look much like a fold at the moment. Let's define a fold function in terms of cata.
Recapping, the list type is:
data L a r = Nil | Cons a r
type List a = Fix (L a)
This needs to be a functor to be useful, which is straight forward:
instance Functor (L a) where
fmap _ Nil = Nil
fmap f (Cons a r) = Cons a (f r)
So specializing cata we get:
cata :: (L x a -> a) -> List x -> a
We're practically there:
construct :: (a -> b -> b) -> b -> L a b -> b
construct _ x (In Nil) = x
construct f _ (In (Cons e n)) = f e n
fold :: (a -> b -> b) -> b -> List a -> b
fold f m = cata (construct f m)
OK, catamorphisms break data structures down one layer at a time.
Anamorphisms
Anamorphisms over lists are unfolds. Unfolds are less commonly known than there fold duals, they have a type like:
unfoldr :: (b -> Maybe (a, b)) -> b -> [a]
As you can see anamorphisms build up data structures. Here's the more general type:
ana :: Functor a => (a -> t a) -> a -> Fix t
This should immediately look quite familiar. The definition is also reminiscent of the catamorphism.
ana f = In . fmap (ana f) . f
It's just the same thing reversed. Constructing unfold from ana is even simpler than constructing fold from cata. Notice the structural similarity between Maybe (a, b) and L a b.
convert :: Maybe (a, b) -> L a b
convert Nothing = Nil
convert (Just (a, b)) = Cons a b
unfold :: (b -> Maybe (a, b)) -> b -> List a
unfold f = ana (convert . f)
Putting theory into practice
filter is an interesting function in that it can be constructed from a catamorphism or from an anamorphism. The other answers to this question (to date) have also used catamorphisms, but I'll define it both ways:
filter p = foldr (\x xs -> if p x then x:xs else xs) []
filter p =
unfoldr (f p)
where
f _ [] =
Nothing
f p (x:xs) =
if p x then
Just (x, xs)
else
f p xs
Yes, yes, I know I used a recursive definition in the unfold version, but forgive me, I taught you lots of theory and anyway filter isn't recursive.
I'd suggest you look at foldr.
Well, are ifs and empty list allowed?
filter = (\f -> (>>= (\x -> if (f x) then return x else [])))
For a list of Integers
filter2::(Int->Bool)->[Int]->[Int]
filter2 f []=[]
filter2 f (hd:tl) = if f hd then hd:filter2 f tl
else filter2 f tl
I couldn't resist answering this question in another way, this time with no recursion at all.
-- This is a type hack to allow the y combinator to be represented
newtype Mu a = Roll { unroll :: Mu a -> a }
-- This is the y combinator
fix f = (\x -> f ((unroll x) x))(Roll (\x -> f ((unroll x) x)))
filter :: (a -> Bool) -> [a] -> [a]
filter =
fix filter'
where
-- This is essentially a recursive definition of filter
-- except instead of calling itself, it calls f, a function that's passed in
filter' _ _ [] = []
filter' f p (x:xs) =
if p x then
(x:f p xs)
else
f p xs