Monads for Drummers

I'd like to explain the Haskell's Monad class. There are so many monad tutorials. Do we need yet another one? But no one tried to explain it so that even a drummer can understand it!

It's just a joke, there is another reason.

The need for writing this tutorial comes from the desire to translate a chapter from my book on Haskell that is written in Russian. Many people claimed that they could finally get into Monads after reading my explanation. So I'd like to make it available for English speaking community.

The need for monad

Do we really need monads after all? The Python, Java or Scala folks can do cool stuff without it. Can we do the things we want with Haskell as we accustomed to with imperative languages and leave the Monads aside as some brainy concept. Brainy concept for the beauty of functional programming. Well, ok, ok. But we are real coders. We can code without Monads!

It happens that the answer is NO. We can not do anything real in Haskell without grasping the Monad concept. Since the Monads are responsible for input/output in the Haskell.

But wait a minute so many languages can live without Monads, why do we need them in Haskell? Well, the short answer is because Haskell is different. It's not like anything else.

Let me show the long answer.

Order of execution (equational reasoning)

Let's start with a small program. Let's prompt the user for a string and then another one. We are going to return a string that contains both strings. Let me write it in Python:

def getBoth():
  x = input()
  y = input()
  return (x + y)

Don't worry if you don't know the Python. It's not that hard to understand. With def we define a function called getBoth. The input gets the user input and with return we concatenate the user inputs. In python the + concatenates strings. So we read user input twice and concatenate the results.

Let's write another program. It reads the input only once and returns it twice. It looks like this in Python:

def getOne():
  x = input()  
  return (x + x)

Pretty clean code. Can we translate it to Haskell? Let's try:

getBoth () = res
    where
        x = input ()
        y = input ()
        res = x ++ y

Let's try to translate the second example. We want the user to type only once:

getOne () = res
    where
        x = input ()        
        res = x ++ x

Unfortunately (or maybe for a good reason) it's not going to work. we are about to see what makes Haskell so different from Python or from many other languages. In Python we have explicit order of execution. Do this line then do the next one if you are done make the third one and so on. So when the Python executes getBoth it executes it line by line:

x = input()     // 1:
y = input()     // 2:
return (x + y)  // 3: return the result

But the Haskell executes statements by functional dependencies! Let's execute the getBoth. It says: I need to return res, so let's look at the res definition. We can see the expression (x ++ y). So let's look at the definition of x and y. Well x is input() so I can substitute x for input(). Well the y is also input() so I can substitute it too. The expression becomes:

input () ++ input ()

Well we need to query user twice for input, concatenate strings and we are done! The funny things start to show off when we try to apply the same procedure to the function getOne. Let's try!

So the we need to calculate the res. The res is x ++ x. We need to calculate the x. The x is input (). So we need to substitute the x. And surprisingly we get the same answer as for the getBoth function:

input () ++ input ()

The problem is that from the Haskell's point of view there is no difference between getBoth and getOne functions. The Haskell has no "line by line" order of execution. This strategy of execution dominates most of the languages and the order of execution is easy to grasp. But in Haskell the order of execution is derived from functional dependencies. We need to execute only subexpressions that result contains. Then we need to execute subexpressions for those subexpressions.

Recall that you can write functions in any order. The main function can come first and then we can define all subexpressions. Haskell compiler seems to understand the right order of execution.

Equational reasoning is an assumption that we can always substitute expressions with their definitions. But we can clearly see how this strategy prevents us from writing useful programs. If we can not write so simple programs as getOne and getBoth. How can we use the language like this?

Wait a moment! Monads are going to save the Haskell out of trouble! Or introduce many more troubles for the novices.. trying to grasp all this stuff.

Here is the Key idea

Monads can introduce the order of execution!

With monads we can simulate this line by line behavior. But let's dive deeper into the notion of the order. Let's stretch our brains a little bit. Let's recall the concept of Monoid.

class Monoid a where
    mempty   :: a
    (<>)     :: a -> a -> a

We have two methods. There are rules:

mempty <> a      === a
a      <> mempty === a

a <> (b <> c)    === (a <> b) <> c

So there is a neutral element mempty and one binary operation <>. The first an second rules say that mempty does nothing. But <> stacks one thing after another and there is no difference in what subexpression of <> to execute as long as we preserve the order. they are aligned in sequence. The <> is called mappend in the standard library and <> is an alias. I use it as a main method here to simplify things.

Let's look at one example. The list of things is a monoid:

instance Monoid [a] where
    mempty = []
    a <> b = a ++ b

So the mempty is an empty list and with sequencing we concatenate the lists. Pretty natural! Let's look also at functions a -> a. There is a monoid instance for them too!

id x = x
(g . f) x = g (f x)

instance Monoid (a -> a) where
    mempty = id
    f <> g = g . f

The mempty is an identity function. It does nothing with the input. The <> is composition of the functions. It applies the first function and then the second one. With this instance we get very close to the notion of sequencing things!

Let's return to the Python. Let's imagine that with some monoid instance we can sequence the order of execution:

    (x = input ())
<>  (y = input ())
<>  (return (x + y))

The input returns some value and stores it to variable x. We need to use this value somehow. But how can we do it? What if our statements are functions and we can pass the user string as input for the next function:

    (x = input ())
<>  (y = input ())

becomes

    input ()
<>  (\x -> y = input ())

And we can also rewrite the y = input():

    input ()
<>  (\x -> 
            input ()
        <>  \y -> return (x ++ y)
    )

In fact that is exactly what Monad class is doing for user input/output. Let's look at the class Monad:

class Monad m where
    return :: a -> m a
    (>>=)  :: m a -> (a -> m b) -> m b

Don't try to understand it! I think a lot of confusion comes with this definition. It's not clear how it relates to ordering. But we can study a structure that is equivalent to the Monad. It's called a Kleisli category.

class Kleisli m where
    idK  :: a -> m a
    (*>) :: (a -> m b) -> (b -> m c) -> (a -> m c)

It defines two methods: the identity function idK and the composition function. If the meaning is not clear to you try to clear the ms. It becomes:

idK  :: a -> a
(*>) :: (a -> b) -> (b -> c) -> (a -> c)

We can represent this with picture:

There are rules for Kleisli. Surprisingly they look a lot like Monoid rules:

idK *> a      === a
a   *> idK    === a

(a *> b) *> b === a *> (b *> b)

The Kleisli is equivalent to Monad. I'll show it later. You can clearly see the monoid structure in the Kleisly. It has a function that does nothing and it has composition of things that preserves the order. We are going to consider our first example later. But now we proceed with more simple examples of monads.

The Kleisli is not defined in standard library. It's just my way to explain monads. You can define it yourself in the file:

module Main where

import Prelude hiding ((*>))

class Kleisli m where
    idK  :: a -> m a
    (*>) :: (a -> m b) -> (b -> m c) -> (a -> m c)

Examples

The partially defined functions (Maybe)

There are functions that are defined not for all inputs. They can return the value or fail.

There is a function head that returns a first element in the list:

head (x:xs) = x
head []     = error "crash"

We can see that it's undefined for the empty list. There is a special type Maybe a that can handle partially defined functions.

data Maybe a = Nothing | Just a

We can define a safe version of the head:

headSafe :: [a] -> Maybe a
headSafe xs = case xs of
    (x:_) -> Just x
    []    -> Nothing

Also we can define a function that extracts a tail from the list safely:

tailSafe :: [a] -> Maybe [a]
tailSafe xs = case xs of
    (_:as) -> Just as
    []     -> Nothing

What if we want to define the function that safely extracts the second element from the list? It would be nice to define it like this:

secondSafe = headSafe . tailSafe

So the second element is just the headSafe that is applied twice. Alas we can not do it. The input and output types don't match.

But if there is an instance of Monad or Kleisli for Maybe, we can define it like this:

secondSafe = tailSafe *> headSafe

The third element is not that difficult to define:

thirdSafe = tailSafe *> tailSafe *> headSafe

To appreciate the usefulness of this approach I'm going to define the secondSafe in the long way:

secondSafe :: [a] -> Maybe a
secondSafe xs = case xs of
    (_:x:_) -> Just x
    _       -> Nothing

Let's look at the definition of composition for functions with Maybe:

If the first function produces the value we are going to apply the second function. If the first function or second one fails we are going to return Nothing. It's very natural definition for partially applied functions. We can encode it:

instance Kleisli Maybe where
    idK a = Just a
    f *> g = \x -> case f x of
        Just a  -> g a
        Nothing -> Nothing

Functions that can return many values

Some functions can return multiple values. They can return list of results.

As an example consider the L-systems. They model the growth of a being. The being is organized with a row of cells. Each cell can divide or split in many cells. We can encode cells with letters and the splitting of cells with rules.

a -> ab
b -> a

Let's start with a and see how it grows:

a
ab
aba
abaab
abaababa

We can model the rules with Haskell:

next :: Char -> String
next 'a' = "ab"
next 'b' = "a"

Then with Kleisli instance it's very easy to define the growth function. Let's see how the composition is implemented:

We apply the second function to all outputs of the first function and we concatenate all results into the single list.

instance Kleisli [] where
    idK     = \a -> [a]
    f *> g  = \x -> concat (map g (f x))

Let's define the growth function:

generate :: Int -> (a -> [a]) -> (a -> [a])
generate 0 f = idK
generate n f = f *> generate (n - 1) f

We can try it with next:

> let gen n = generate n next 'a'
> gen 0
"a"
> gen 1
"ab"
> gen 2
"aba"
> gen 3
"abaab"
> gen 4
"abaababa"

Functions with state

Another useful function is a function with a state. It calculates its result based on some state value.

It expects the initial state and it returns the pair of value and updated state. In Haskell it's defined in the module Control.Monad.State.

data State s a = State {  runState :: s -> (a, s) }

Can you guess the Kleisli definition by looking at the picture of composition?

The Monad class

So we can see that Kleisli is all about sequencing things. Let's take a closer look at Monad:

class Monad m where
    return :: a -> m a
    (>>=)  :: m a -> (a -> m b) -> m b

We can see that return is function idK for the Kleisli class. What is the >>=. It's called bind. We can see better what it can do if we clear the m out of the signature:

a -> (a -> b) -> b

What if we reverse the arguments:

(a -> b) -> a -> b

It's an application of the function to the value! So the monadic bind method is just an application of the monadic value (wrapped in some structure m) to the function that returns another monadic value. We can express the Monad class with Kleisli:

instance Kleisli m => Monad m where
    return  = idK
    a >>= f = ((\_ -> a) *> f) ()

We first create a function out of the value and we apply the function to an empty tuple to express application with composition.

Also we can define Kleisli in terms of Monad:

instance Monad m => Kleisli m where
    idK    = return
    f *> g = \x -> x f >>= g

Let's turn back to the example with user input. How can we rewrite the three simple lines with Haskell:

x = input ()
y = input ()
return (x ++ y)

Can you recall that we transformed it with monoid methods:

    (x = input ())
<>  (y = input ())
<>  (return (x + y))

and then later to

    input ()
<>  (\x -> 
            input ()
        <>  \y -> return (x ++ y)
    )

It turns out that the monoid sequencing <> is like bind or sequencing for Kleisli. So with real Haskell we get:

    input ()
>>= (\x -> 
            input ()
        >>= \y -> return (x ++ y)
    )

And we can rewrite the second example that way:

x = input
return x ++ x

With Haskell it becomes:

    input ()
>>= (\x -> return (x ++ x))

We can see that with Monads the two examples are different! That is the magic of sequencing with Monads. But you can say ooh my God! Do I need to write those simple Python programs with monstrous expressions like this:

    input ()
>>= (\x -> 
            input ()
        >>= \y -> return (x ++ y)
    )

In fact you don't need to do it. The clever Haskell language designers created a special syntactic sugar to make Haskell look like an imperative language. It's called do-notation. With it we can write our example like this:

Ask twice:

getBoth () = do
    x <- getLine 
    y <- getLine
    return (x ++ y)

Ask once:

getOne () = do
    x <- getLine     
    return (x ++ x)

I've substituted the Python's name input for the Haskell's getLine. They do the same thing.

Let's see how do-notation is translated to application of monad methods:

do
    x <- expr1    =>     expr1 >>= \x ->
    y <- expr2           expr2 >>= \y ->
    f x y                f x y

So the left hand side of the arrow <- becomes an input for the lambda-function. The do-next line is a bind operator.

Conclusions

So we have a nice and clean way to do imperative programming with Haskell! The imperative languages like Java or Python also have monads. It's better to say that they have THE MONAD. It's built in and it's hidden from the user. It's so much in the bones of the language that we tend not to care about it. It just works! But the Haskell is different! With Haskell we can see that there are plenty of different monads and we can redefine the do-next-line operator that is built in the other languages.

The original need for Monads was induced by desire to implement input/output in the purely functional setting. But the smart eyes of Haskell developers could see deeper. We can define many different default behaviors with the same interface. The interface of sequencing!

Bonus

Reader

There are other monads. The reader monad is for functions that can access some "global" state or environment. They can only read the value of the state.

It's defined like this in the Haskell code:

data Reader r a = Reader (r -> a)

Can you define Kleisli or Monad instance for it by looking at the picture of composition:

Writer

The writer monad is another special case of a state Monad. The writer can process the values and accumulate some state. The accumulation of the state is expressed with Monoid methods:

data Writer w a = Writer (a, w)

instance Monoid w => Monad (Writer w) where
    return a = Writer (a, mempty)
    a >>= f  = ...

Can you complete the Monad class by looking at the picture of it:

Functor

Sometimes the monadic bind >>= is overkill. Instead of applying a monadic function to a monadic value:

m a -> (a -> m b) -> m b

We just want to apply a pure function to a monadic value:

m a -> (a -> b) -> m b

That's what Functor class is all about:

class Functor m where
    fmap :: (a -> b) -> m a -> m b

Notice that the order of the arguments is reversed. We rewrite our ask once example with a single line:

> fmap (\x -> x ++ x) getLine

Applicative

Ok, with Functor we can apply a pure function to a single monadic value. But what if we have a pure function that accepts several arguments and we want to apply them to several monadic values.

(a -> b -> c -> out) -> m a -> m b -> m c -> m out

The Applicative class is here to help us! It looks like this:

f <$> ma <*> mb <*> mc

or

liftA3 f ma mb mc

There are also liftA2, liftA4. The actual definition of the Applicative is

class Functor m => Applicative m where
    pure  :: a -> m a 
    (<*>) :: m (a -> b) -> m a -> m b

(<$>) = fmap

It's a little bit windy but the concept is simple. We want to apply a pure function to several m-structured values. The pure is the same as return.

We can redefine the two examples with Applicative and Functor:

Ask twice:

liftA2 (++) getLine getLine

Ask once:

fmap (\x -> x ++ x) getLine

You can try it right in ghci!

I hope you get the idea and you won't get lost in the forest of Applicatives, Functors and Monads. Happy Haskelling!