Skip to content

cffls/py2hs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

23 Commits
 
 
 
 
 
 
 
 

Repository files navigation

py2hs

Introduction

I find it an interesting practice to learn and reinforce a new programming language by comparing it with a familiar language, and adding the distinct features and concepts from the new language on top of my existing knowledge system. This is a reference guide I created when I was learning Haskell.

This project is inspired by py2rs.

This is:

  • a complementary resource that helps python programmers to learn Haskell as a new language
  • a place for beginners to learn and reinforce some of the core concepts in Haskell

This is not:

  • an one-stop-shop that teaches everything about Haskell
  • an encyclopedia that aims at covering all similarities and differences between Python and Haskell

Table of contents

General

Definition Python Haskell
Programming paradigm Imperative Purely functional
Type System Dynamically typed Statically typed
First appeared 1989 1990
File extensions .py, .pyw, .pyc .hs, .lhs
Programming guidelines PEP8 Haskell programming guidelines

Getting started with Haskell

Official Haskell documentation provides plenty of tutorials and books to start with.

https://www.haskell.org/documentation/

Code format

All examples are either code snippet or interactions with python shell or Haskell GHCi. Python shell examples starts with >>> , and Haskell GHCi examples starts with λ> . Please note that all Python codes are written in Python 3.


Functions

Basic function syntax

Python

>>> def multiply(x, y):
...     return x * y
...
>>> multiply(2, 3)
6

Haskell

λ> multiply x y = x * y
λ> multiply 2 3
6

Flow control

Python

def bigger(x, y):
    if x > y:
        return x
    else:
        return y

Haskell

bigger x y = if x > y
             then x
             else y

Explicit typing

Python

def bigger(x: int, y: int) -> int:
    if x > y:
        return x
    else:
        return y

Haskell

Approach 1: explicitly specify all inputs and the output
bigger :: Int -> Int -> Int
bigger x y = if x > y
             then x
             else y
Approach 2: set a type constraint Integral for a
bigger :: (Integral a) => a -> a -> a 
bigger x y = if x > y
             then x
             else y

Lambda function

Python

>>> list(map(lambda x: x + 1, [1, 2, 3]))
[2, 3, 4]

Haskell

λ> map (\x -> x + 1) [1, 2, 3]
[2,3,4]

Comments

Python

# This is a python comment
'''
Python
multiline
comment
'''

Haskell

-- This is a haskell comment
{-
haskell
multiline
comment
-}

List

List creation

Python

>>> [1, 2, 3, 4]
[1, 2, 3, 4]
>>> list(range(1, 5)) # Use range
[1, 2, 3, 4]

Haskell

Syntax sugar:

λ> [1, 2, 3, 4]
[1,2,3,4]
λ> [1..4] -- Use range
[1,2,3,4]

Without syntax sugar:

λ> 1:2:3:4:[]
[1,2,3,4]

List concatenation

Python

>>> [1, 2] + [3, 4]
[1, 2, 3, 4]

Haskell

λ> [1, 2] ++ [3, 4]
[1,2,3,4]

Adding element to list

Append

Python

>>> my_list = [1, 2, 3, 4]
>>> my_list.append(5)
>>> my_list
[1, 2, 3, 4, 5]

Haskell

λ> [1, 2, 3, 4] ++ [5]
[1,2,3,4,5]
Prepend

Python

>>> my_list = [1, 2, 3, 4]
>>> my_list.insert(0, 5)
>>> my_list
[5, 1, 2, 3, 4]

Haskell

λ> 5:[1, 2, 3, 4]
[5,1,2,3,4]

Indexing

Python

>>> my_list = [1, 2, 3, 4]
>>> my_list[2]
3

Haskell

λ> let myList = [1, 2, 3, 4]
λ> myList !! 2
3

List comprehension

Python

>>> [i * 2 for i in range(1, 11)]
[2, 4, 6, 8, 10, 12, 14, 16, 18, 20]

Haskell

λ> [x*2 | x <- [1..10]]
[2,4,6,8,10,12,14,16,18,20]
List comprehension with conditions

Python

>>> [i * 2 for i in range(1, 11) if i % 2 == 0]
[4, 8, 12, 16, 20]

Haskell

λ> [x*2 | x <- [1..10], even x]
[4,8,12,16,20]
More complex list comprehension

Python

>>> [i * j for i in range(1, 5) for j in range(1, 5)]
[1, 2, 3, 4, 2, 4, 6, 8, 3, 6, 9, 12, 4, 8, 12, 16]

>>> [[a, b, c] for c in range(1, 11)
...     for b in range(1, c+1)
...     for a in range(1, b+1) if a**2 + b**2 == c**2]
[[3, 4, 5], [6, 8, 10]]

Haskell

λ> [ x*y | x <- [1..4], y <- [1..4]]
[1,2,3,4,2,4,6,8,3,6,9,12,4,8,12,16]

λ> [ [a,b,c] | c <- [1..10], b <- [1..c], a <- [1..b], a^2 + b^2 == c^2]
[[3,4,5],[6,8,10]]

Zipping

Python

>>> list(zip(range(1,5), range(2,6)))
[(1, 2), (2, 3), (3, 4), (4, 5)]

>>> list(zip(range(1, 5), [8, 9, 10]))
[(1, 8), (2, 9), (3, 10)]

Haskell

λ> zip [1..4] [2..5]
[(1,2),(2,3),(3,4),(4,5)]

λ> zip [1..4] [8,9,10]
[(1,8),(2,9),(3,10)]

Class -> data

In Python, and most OOP languages, Classes provide a means that bundles two things:

  • data
  • functions that retrieve, modify, or manipulate the data

Let's start with defining a class in Python.

from dataclasses import dataclass

@dataclass
class Square:
    side: float

    def area(self):
        return self.side ** 2
>>> my_square = Square(5)
>>> my_square.area()
25
>>> my_square.side = 6
>>> my_square.area()
36

Class Square bundles a float number, side, and a function, area, which calculates the area of the square. Attribute side could be directly updated. The functions bundled with the class are stateful, because the outputs of the same function could be different depending on the state of the class. Being stateful is like holding a double-edge sword. On one hand, you can make a class very flexible and adaptive to new changes. On the other hand, however, the behavior of functions could be non-deterministic with respect to their inputs, making them unpredictable, and therefore, difficult to debug.

Now let's see a way to define a Square and a function that calculates any Square's area in Haskell.

data is a keyword in Haskell. It defines a type of data structure, similar to Class in Python.

data Square = Square Float -- Defining "Square" as a data structure that holds a Float

area :: Square -> Float -- Declaring function "area" as a function that takes a "Square" and returns a Float
area (Square side) = side ^ 2 -- Implement the area function
λ> let mySquare = Square 5
λ> area mySquare
25.0

What we saw above, is a type (Square) that holds a float, and a function (area) that operates on this specific type, while not being bundled to the data structure itself. The separation of data and function, is one of the most important characteristics of Haskell. What's good about the separation of data and function? Stateless! Instead of depending on some "internal" state, functions are deterministic and predictable purely from their inputs.


Interface -> Type class

Area does not only apply to square, but all types of shapes. Let's start with defining an interface, area, in Python.

from abc import ABC, abstractmethod

class Shape(ABC):

    @abstractmethod
    def area(self):
        pass

Then we define a couple of shapes that implements area.

import math
from dataclasses import dataclass

@dataclass
class Square(Shape):
    side: float

    def area(self):
        return self.side ** 2

@dataclass
class Rectangle(Shape):
    height: float
    width: float

    def area(self):
        return self.height * self.width

@dataclass
class Circle(Shape):
    radius: float

    def area(self):
        return math.pi * self.radius ** 2
>>> my_shapes = [Square(5), Rectangle(5, 6), Circle(3)]
>>> [shape.area() for shape in my_shapes]
[25, 30, 28.274333882308138]

How do to do so in Haskell?

In Haskell, data can define more than one constructor.

data Shape = Square Float | Rectangle Float Float | Circle Float

Here is one way to write our area function in Haskell:

area :: Shape -> Float
area (Square side) = side ^ 2
area (Rectangle height width) = height * width
area (Circle radius) = pi * radius ^ 2
λ> map area [Square 5, Rectangle 5 6, Circle 3]
[25.0,30.0,28.274334]

You might already observe a problem here: whenever a new type of "Shape" is created, we will need to add that type to the implementation of "area".

For instance, if we create a new type that represents continents and we want them to support calculation of their areas. If we continue to use the same area function, things become ugly.

First, we need to add more types to shape:

data Shape = Square Float
    | Rectangle Float Float
    | Circle Float
    | Africa
    | Antarctica

Then, we need to modify area function as well

area :: Shape -> Float
area (Square side) = side ^ 2
area (Rectangle height width) = height * width
area (Circle radius) = pi * radius ^ 2
area (Africa) = 1.117e7 -- in the unit of square miles
area (Antarctica) = 5.483e6

In other words, the implementation of area is coupled with Shape.

Here is where type classes become handy. Type class is similar to the concept of interface in object-oriented programming. Let's try to define our "interface" using a type class.

class MyObject a where
    area :: a -> Float

What we are are seeing, is a type class, MyObject, that declares a behavior (or interface), area, which returns a Float when a type of MyObject is given. Notice that the class in Haskell is not the same as the "class" in Python. Instead, it is more similar to an abstract class in Python.

Now, we can define Shapes and Continents separately and make them support area calculation at the same time.

Define shapes
data Shape = Square Float | Rectangle Float Float
instance MyObject Shape where
    area (Square side) = side ^ 2
    area (Rectangle height width) = height * width

A new Haskell keyword, instance, is used here. instance declares that a data type is to implement all "interfaces" defined in a type class. In our case, instance MyObject Shape could be read as MyObject will implement all functions declared in type cass Shape. In Python, it will look something like this class MyObject(Shape):.

λ> let a = Square 4
λ> area a
16.0
λ> let b = Rectangle 4 5
λ> area b
20.0
Define continents
data Continent = Africa | Antarctica
instance MyObject Continent where
    area Africa = 1.117e7
    area Antarctica = 5.483e6
λ> area Africa
1.117e7
λ> area Antarctica
5483000.0

The codes of Shape and Continent are completely independent, but we can still apply the same function (area) on them.

This concept is actually nothing new in Python. Remember data model? Yes, type class is a more power version of data model, which is a list of pre-defined interfaces by Python, while type class can define any possible interfaces.

In python, a class can implement __eq__ so we can apply == operator to it:

class Rectangle:
    def __init__(self, height, width):
        self.height = height
        self.width = width

    def __eq__(self, other):
        if isinstance(other, Rectangle):
            return self.height == other.height and self.width == other.width
>>> Rectangle(5, 6) == Rectangle(5, 6)
True
>>> Rectangle(5, 6) == Rectangle(6, 7)
False

Unsurprisingly, == belongs to type class Eq in Haskell.

This is the definition of Eq:

class Eq a where
    (==) :: a -> a -> Bool
    (/=) :: a -> a -> Bool
    x == y = not (x /= y)  
    x /= y = not (x == y)   

Let's try implement Eq for our rectangle.

data Rectangle = Rectangle Float Float
instance Eq Rectangle where
    Rectangle a b == Rectangle x y = a == x && b == y
λ> Rectangle 5 6 == Rectangle 5 6
True
λ> Rectangle 5 6 == Rectangle 6 7
False
λ> Rectangle 5 6 /= Rectangle 6 7
True

Notice that we only need to implement either == or /=, and the other operator will be automatically usable. This is achieved by the default implementation of Eq, where one is the negation of the other:

    x == y = not (x /= y)
    x /= y = not (x == y)

Higher order functions

Higher order functions are functions that takes functions as inputs or return functions as outputs. Let's start with some of the most common higher order functions: map, filter, and reduce.

Map

Python

>>> list(map(lambda x: x*2, range(1, 5)))
[2, 4, 6, 8]

Haskell

λ> map (*2) [1..4]
[2,4,6,8]

Filter

Python

>>> list(filter(lambda x: x >= 3, range(1, 5)))
[3, 4]

Haskell

λ> filter (>=3) [1..4]
[3,4]

Reduce

Python

>>> from functools import reduce
>>> reduce(lambda x, a: x + a, range(1,5), 100)
110

Reduce is called "fold" in Haskell

λ> foldl (+) 100 [1..4]
110

Partial functions

In Python, a function can be passed as an argument to another function, which is called higher order function. For example, we can create a higher order function that supplies default arguments to other functions. Below, we have function add and multiply that both take two input arguments.

def add(x, y):
    return x + y

def multiply(x, y):
    return x * y

Now, we can create a higher order function that applies default input argument to the two functions above.

def partial(func, *partial_args):
    def apply(*remaining_args):
        return func(*partial_args,  *remaining_args)
    return apply

Now, we can use a combination of partial and another function to construct interesting functions.

>>> increment = partial(add, 1)
>>> increment(2)
3
>>> increment(3)
4
>>> double = partial(multiply, 2)
>>> double(2)
4
>>> double(3)
6

In fact, Python standard library provides partial as a utility function in module functools.

from functools import partial
>>> from functools import partial
>>> increment = partial(add, 1)
>>> increment(3)
4
>>> double = partial(multiply, 2)
>>> double(2)
4

Can we do something similar in Haskell? Yes, and even simpler!

λ> increment = (+) 1
λ> increment 3
4
λ> increment 4
5
λ> double = (*) 2
λ> double 2
4
λ> double 3
6

In fact, all functions could be partially applied in Haskell by default. In Haskell, we call partially applied functions as "curried functions", named after Haskell Curry.


Functor

In Python, map is very useful when we want to transform an iterable.

For example, we can transform a list of integers to a list of strings with map.

>>> list(map(str, [1, 2, 3, 4, 5]))
['1', '2', '3', '4', '5']

Here, list is a "container" that holds the elements, each of which is mapped to a new element.

A limitation of map is that the container has to be an iterable.

Say we have a tree structure like below:

class Tree:
    def __init__(self, left=None, right=None, val=0):
        self.left = left
        self.right = right
        self.val = val

    def __repr__(self):
        return f'Tree(val={self.val}, left={self.left}, right={self.right})'
>>> my_tree = Tree(left=Tree(val=1), right=Tree(val=2))
>>> my_tree
Tree(val=0, left=Tree(val=1, left=None, right=None), right=Tree(val=2, left=None, right=None))

Clearly, map can't transform the values in the tree structure. In this case, it is easier to implement our own map.

def map_tree(my_func, tree):
    if tree:
        return Tree(left=map_tree(my_func, tree.left), 
                    right=map_tree(my_func, tree.right),
                    val=my_func(tree.val))
>>> map_tree(lambda x: x*2, my_tree)
Tree(val=0, left=Tree(val=2, left=None, right=None), right=Tree(val=4, left=None, right=None))

>>> map_tree(lambda x: 'transformed!', my_tree)
Tree(val=transformed!, left=Tree(val=transformed!, left=None, right=None), right=Tree(val=transformed!, left=None, right=None))

Now, we have a default map function from Python standard library, and we also have our own implementation of map_tree, implemented for Tree structure. Unfortunately, there is not a straightforward way to unify these two functions in Python. By unify, we mean a universal mapper that can apply a function to list, Tree, and essentially any type of data structures.

Thanks to the robustness of type class, this universal mapper is possible in Haskell, and it is provided as part of Haskell's standard library. It is called "Functor".

Here is the definition of functor:

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

f is a type of container that "contains" data. It could be a list, a Tree, or anything that holds data. (a -> b) means a function that transform a value of type a to a value of type b. f a is a concrete type that means container f is holding values of type a. f b means container f holds values of type b.

All together, the definition of Functor in plain words, is something that implements a function fmap, that, given a function ((a -> b)) and a container f that holds values of type a, returns a container f that holds values of type b. fmap is the universal map function we've been looking for.

Let's look at the implementation of Functor for list.

instance Functor [] where  
    fmap = map  

It is really straightforward. The Functor implementation of list is simply map.

λ> fmap (*2) [1,2,3]
[2,4,6]

Now, let's create a Tree type and see how Functor can work in such context.

λ> data Tree a = EmptyTree | Node a (Tree a) (Tree a) deriving (Show, Read, Eq)
λ> tree = Node 0 (Node 1 EmptyTree EmptyTree) (Node 2 EmptyTree EmptyTree)
λ> tree
Node 0 (Node 1 EmptyTree EmptyTree) (Node 2 EmptyTree EmptyTree)

Implementing fmap for Tree:

instance Functor Tree where  
    fmap f EmptyTree = EmptyTree  
    fmap f (Node x left right) = Node (f x) (fmap f left) (fmap f right)  
λ> fmap (*2) tree
Node 0 (Node 2 EmptyTree EmptyTree) (Node 4 EmptyTree EmptyTree)
λ> fmap (\x -> "Transformed!") tree
Node "Transformed!" (Node "Transformed!" EmptyTree EmptyTree) (Node "Transformed!" EmptyTree EmptyTree)

We can see that fmap could be applied to any type of "container" as long as it is an instance of Functor.


Applicative

Applicative is a more powerful type of Functor.

In the section of Functor, we learn how to apply a function to a Tree with fmap.

What if we have two trees and we want to overlay the two to generate a new Tree?

  Tree 1          Tree 2              New Tree

    1       +       5       --->         6                
   / \             /                    /
  2   3           6                    8

What if we want to apply three trees to a function?

 
   Tree 1          Tree 2         Tree 2              New Tree
 _                        _
|    1               5     |        5                    30                
|   / \      +      /      |   *   /        --->        /
|  2   3           6       |      6                   48
 -                        - 

First, let's see how to implement this in python.

We can't use map_tree created previously, because it only operates on one tree. We can create a functions that can operates on any number of trees:

def map_trees(trees, func):
    if all(trees):
        return Tree(val=func(*[tree.val for tree in trees]), 
                    left=map_trees([tree.left for tree in trees], func), 
                    right=map_trees([tree.right for tree in trees], func))
    else:
        return None
>>> tree1 = Tree(val=1, left=Tree(val=2), right=Tree(val=3))
>>> tree2 = Tree(val=5, left=Tree(val=6))
>>> map_trees([tree1, tree2], add)
Tree(val=6, left=Tree(val=8, left=None, right=None), right=None)
>>> map_trees([tree1, tree2], pow)
Tree(val=1, left=Tree(val=64, left=None, right=None), right=None)
>>> map_trees([tree1, tree2, tree2], lambda a, b, c: (a + b) * c)
Tree(val=30, left=Tree(val=48, left=None, right=None), right=None)

Intuitively, map_trees, which accepts any numbers of trees as inputs, is a more powerful version of map_tree, which takes only two trees as inputs. Similarly, applicative is a more powerful version of functor, in a way that it could be passed to a function that has more than one input. In Haskell, applicative is defined as below:

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

First, the type constraint tells us that applicative has to be a Functor in the first place. Second, applicative has two functions. pure is basically wrapping a value in a functor (or "container"). <*> is similar to fmap. A quick reminder of fmap:

fmap :: Functor f => (a -> b) -> f a -> f b

The only difference between fmap and <*> lies in their first argument, where functor is looking for a function, (a -> b), while applicative is looking for a function that is wrapped in a functor, f (a -> b).

Seems weird? Let's look how we can implement map_trees with applicative in Haskell.

instance Applicative Tree where
    pure x = Node x (pure x) (pure x)
    Node a aLeft aRight <*> Node b bLeft bRight = Node (a b) (aLeft <*> bLeft) (aRight <*> bRight) 
    _ <*> _ = EmptyTree

What pure does is simply placing a value inside a container (a tree in the case above). For <*>, the values of two Trees are combined by applying b to a recursively when both inputs are not empty. For the rest of cases, simply return an EmptyTree. With Tree being an applicative, we can now apply a function takes X number of inputs to X number of trees.

λ> tree1 = Node 1 (Node 2 EmptyTree EmptyTree) (Node 3 EmptyTree EmptyTree)
λ> tree2 = Node 5 (Node 6 EmptyTree EmptyTree) EmptyTree
λ> pure (+) <*> tree1 <*> tree2
Node 6 (Node 8 EmptyTree EmptyTree) EmptyTree
λ> my_func a b c = (a + b) * c
λ> pure my_func <*> tree1 <*> tree2 <*> tree2
Node 30 (Node 48 EmptyTree EmptyTree) EmptyTree
λ> my_func a b c d = a * b + c * d
λ> pure my_func <*> tree1 <*> tree2 <*> tree2 <*> tree2
Node 30 (Node 48 EmptyTree EmptyTree) EmptyTree

In the code above, we can pass multiple trees as inputs to a pure function by connecting them with <*>, e.g. pure (+) <*> tree1 <*> tree2. How and why does this work exactly?

If we look at pure (+) carefully, it is an applicative that contains a function with type a -> a -> a.

λ> :t pure (+)
pure (+) :: (Applicative f, Num a) => f (a -> a -> a)

Intuitively, pure (+) could be visualized as an abstract container whose element is (+). When the container is Tree, for example, it could be visualized as an infinity Tree where each node contains (+) a function, as visualized below. One thing to be clear is that this "addition tree" is not fully constructed in memory. Instead, the nodes will be lazily created as we apply another Tree to it with <*>.

Applicative Tree of (+)

         (+)
        /  \
      (+)  (+)
      / \  / \
    (+)(+)(+)(+)
   . . . . . . . .  

Now if we apply a Tree to pure (+) using <*>, we end up with a concrete
Tree applicative that contains a function of type a -> a. The type wrapped in the applicative has now changed from a -> a -> a (addition function, (+)) to a -> a, because the first argument of the addition function has been filled with tree1.

λ> tree1 =  Node 1 (Node 2 EmptyTree EmptyTree) (Node 3 EmptyTree EmptyTree)
λ> :t pure (+) <*> tree1
pure (+) <*> tree1 :: Num a => Tree (a -> a)

pure (+) <*> tree1 could be visualized as below:

     (+)1
     /  \
    /    \
  (+)2   (+)3

With pure (+) <*> tree1, we can apply it with another Tree using <*> again, and finally we get a Tree that contains just numbers.

λ> tree2 = Node 5 (Node 6 EmptyTree EmptyTree) EmptyTree
λ> result = pure (+) <*> tree1 <*> tree2
λ> result
Node 6 (Node 8 EmptyTree EmptyTree) EmptyTree
λ> :t result
result :: Num b => Tree b

pure (+) <*> tree1 <*> tree2 could be intuitively visualized like this:

    pure (+) <*> tree1      <*>       tree 2       ->       result

          (+)1                          5                    6
          /  \                         /                    /
         /    \                       /                    /
      (+)2   (+)3                    6                    8


Monoid

In the section of Applicative, we learnt how a series of functors could be passed as arguments to a regular function. Now, let's consider a similar but slightly different problem: how can we apply a function to reduce or fold a tree recursively? e.g. calculating the sum or product of values in a tree.

There are many of ways to implement this in python. Here is one way:

def reduce_tree(tree, func, default=1):
    if tree:
        reduced_left = func(reduce_tree(tree.left, func, default=default), tree.val)
        reduced = func(reduced_left, reduce_tree(tree.right, func, default=default))
        return reduced
    else:
        return default
>>> tree = Tree(val=1, left=Tree(val=2), right=Tree(val=4))
>>> reduce_tree(tree, lambda a,b: a+b, default=0)
7
>>> reduce_tree(tree, lambda a,b: a*b, default=1)
8
>>> reduce_tree(tree, min, default=float("inf"))
1
>>> reduce_tree(tree, max, default=-float("inf"))
4

We can also make the Tree iterable, so we can directly use reduce and make our code a bit more functional.

class Tree:
    def __init__(self, left=None, right=None, val=0):
        self.left = left
        self.right = right
        self.val = val

    def __repr__(self):
        return f'Tree(val={self.val}, left={self.left}, right={self.right})'

    def __iter__(self):
        if self.left:
            yield from self.left
        yield self.val
        if self.right:
            yield from self.right
>>> tree = Tree(val=1, left=Tree(val=2), right=Tree(val=4))
>>> list(tree)
[2, 1, 4]
>>> from functools import reduce
>>> reduce(lambda a,b: a+b, tree, 0)
7
>>> reduce(lambda a,b: a*b, tree, 1)
8
>>> reduce(min, tree, float("inf"))
1
>>> reduce(max, tree, -float("inf"))
4

There is an important assumption in implementations above. That is, all functions being applied to the tree has to be associative. Associativity means that the order of function application doesn't determine the final output value. To define it more strictly, a binary function f is associative if the following condition satisfy for all x, y, z in S, where S is a set of values.

f(x, f(y, z)) = f(f(x,y), z)

It is not hard to tell that +, *, min, and max are all associative. That's why we can apply them to a Tree without worrying about the order in which they are applied internally.

An example of non-associative function is average. e.g. average(average(a, b), c) is not equal to average(a, average(b, c)). If we apply average to a tree, the final value will be different depending on whether left child or right child is reduced with the root first.

def reduce_tree(tree, func, default=1):
    if tree:
        reduced_left = func(reduce_tree(tree.left, func, default=default), tree.val)
        reduced = func(reduced_left, reduce_tree(tree.right, func, default=default))
        return reduced
    else:
        return default

def reduce_tree2(tree, func, default=1):
    if tree:
        reduced_right = func(tree.val, reduce_tree2(tree.right, func, default=default))
        reduced = func(reduce_tree2(tree.left, func, default=default), reduced_right)
        return reduced
    else:
        return default
>>> tree
Tree(val=1, left=Tree(val=2, left=None, right=None), right=Tree(val=4, left=None, right=None))
# Addition is not determined by the order in which values are applied
>>> reduce_tree(tree, lambda a,b: a+b, default=0)
7
>>> reduce_tree2(tree, lambda a,b: a+b, default=0)
7
# Average is determined by the order in which values are applied
>>> reduce_tree(tree, lambda a,b: (a+b)/2, default=0)
0.875
>>> reduce_tree2(tree, lambda a,b: (a+b)/2, default=0)
0.75

Why are we talking about associativity? Because it is the essence of monoids. In Haskell, Monoid is a type class that has an associative binary function and an identity value, A, which, when applied the binary function along with any value, B, always yields B as the output.

For example, 0 is the identity value with respect to binary function +, and 1 is the identity value with respect to binary function *.

Here is the formal definition of Monoid type class:

class Monoid m where  
    mempty :: m  
    mappend :: m -> m -> m  
    mconcat :: [m] -> m  
    mconcat = foldr mappend mempty  

mempty is the identity value, and mappend is the associative binary function.
mconcat is a function that "concatenates" a list of monoids into one, and we get it for free with a default implementation.

Now, let's first see how to make a foldable tree in haskell without using monoids.

import qualified Data.Foldable as F  

data Tree a = EmptyTree | Node a (Tree a) (Tree a) deriving (Show, Read, Eq)

instance F.Foldable Tree where
    foldr f z EmptyTree = z
    foldr f z (Node val left right) = f (f val rightResult) leftResult
	    where
	    rightResult = foldr f right z
	    leftResult = foldr f left z

The code seems fine in the first look. We pass the reduced results from right and left to function f, and we get the final reduced result . However, GHC complains!

    • Couldn't match expected type ‘a’ with actual type ‘b’

  |
7 |     foldr f z (Node val left right) = f (f val rightResult) leftResult
  |                                          ^^^^^^^^^^^^^^^^^

Well, it turns that function f (whose type is a -> b -> b) is expecting its first argument to be type a, but it got type b, which is the output of f val rightResult. In fact, we don't really have a way of turning type b into type a. We are stuck here.

Fortunately, Foldable provides a second interface, foldMap, which is easier to implement:

foldMap :: (Monoid m, Foldable t) => (a -> m) -> t a -> m  

Now, instead of taking a function of type a -> b -> b, foldMap takes a function of type a -> m, which transforms a type, a, to a monoid, m.

Here is the implementation of foldMap for Tree:

instance F.Foldable Tree where  
    foldMap f EmptyTree = mempty  
    foldMap f (Node x left right) = mconcat [foldMap f l, f x, foldMap f r]

Now we can see the advantage of using monoid here. F.foldMap f left, f x, and F.foldMap f right all return a monoid as outputs, which could be "concatenated" into
a single monoid with mconcat. Now, we just need to find a function that turns a value of a Node to a monoid.

A simple way is to create a newtype as a monoid for numbers.

newtype Product a =  Product { getProduct :: a }  
    deriving (Eq, Ord, Read, Show, Bounded)  

instance Num a => Monoid (Product a) where  
    mempty = Product 1  
    Product x `mappend` Product y = Product (x * y)  
λ> getProduct $ Product 3 `mappend` Product 4
12

Now, we can use Product as a monoid for foldMap.

λ> tree = Node 2 (Node 1 EmptyTree EmptyTree) (Node 4 EmptyTree EmptyTree)
λ> getProduct $ foldMap Product tree
8

Not surprisingly, Product is defined as one of the monoids in standard module Data.Monoid, along with Sum, Min, Max, and many more.

λ> getSum $ foldMap Sum tree
7

Monad

Monad is one of the most important concepts in functional programming, but it is not well known in the realm of object-oriented programming. Nevertheless, we will try to implement a monad in Python, use it in an example, and then show how it could be equivalently implemented and used in Haskell.

Programs can fail with exceptions, which can usually cause side effects in control flows.

For example, in Python, a division function will throw a ZeroDivisionError when the divisor is 0.

>>> def div(x, y):
...     return x / y
...
>>> div(4, 2)
2.0
>>> div(3, 2)
1.5
>>> div(3, 0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 2, in div
ZeroDivisionError: division by zero

When a ZeroDivisionError exception is thrown and not handled within div, the exception will be propagated to the caller of div, causing a side effect. If we want to get rid of such side effect, or in other words, make it a pure function, we need to catch and handle any possible exception within it. This could be implemented by creating a wrapper class. Let's call it Maybe, which stores a successfully returned value or a "failure" state when an exception is thrown. We also need a wrapper function that safely calls the division
function and wraps the returned value into the wrapper. See the implementation below.

class Maybe():
    def __init__(self, value=None, failed=False):
        self.value = value
        self.failed = failed

    def __repr__(self):
        if self.failed:
            return "Nothing"
        else:
            return f"Just {self.value}"

def safe_div(x, y):
    try:
        value = x / y
        return Maybe(value=value)
    except Exception as e:
        print(f"Got exception: {e}") # IO is also a side effect, but we will keep it here to help us understand
                                     # the behavior of this function in examples below.
        return Maybe(failed=True)

With class Maybe and function safe_div, we can both eliminate its side effects and retain its pureness.

>>> safe_div(4, 2)
Just 2.0
>>> safe_div(4, 1)
Just 4.0
>>> safe_div(4, 0)
Got exception: division by zero
Nothing

However, with this design, we will inevitably run into a problem when more than one safe_div are chained together:

>>> x = safe_div(4, 2)
>>> y = safe_div(2, 1)
>>> x
Just 2.0
>>> y
Just 2.0
>>> safe_div(x, y)
Got exception: unsupported operand type(s) for /: 'Maybe' and 'Maybe'
Nothing

safe_div(x, y) is supposed to return Just 1.0, but it ends up returning Nothing! The problem is that both x and y returned from safe_div are in type Maybe, but operator "/" doesn't know how to operate on Maybe, so an exception is thrown by div and caught inside safe_div, which returned Nothing.

A simple way to fix this is to create another function safe_div_maybe, which takes inputs in Maybe types and calls safe_div internally.

def safe_div_maybe(x: Maybe, y: Maybe) -> Maybe:
    if not x.failed and not y.failed:
        return safe_div(x.value, y.value)
    else:
        return Maybe(failed=True)
>>> x = safe_div(4, 2)
>>> y = safe_div(2, 1)
>>> x
Just 2.0
>>> y
Just 2.0
>>> safe_div_maybe(x, y)
Just 1.0

However, with this approach, we need to create two functions (safe_div and safe_div_maybe) that do pretty much the same thing except that their input types are different. In addition, for every new function, e.g. func_x, we will need to create safe_func_x and its counter part safe_func_x_maybe. This doesn't seem to be an elegant solution.

The root problem is that function safe_div is not composable, because its input type (int) and its output type (Maybe) are different. Therefore, we cannot simply pass an output value from a safe_div as an input argument to another safe_div. In comparison, the original operator "/" is composable, because we can pass the output value from a "/" to another "/" without any problem, as shown below:

>>> x = 4 / 2
>>> y = 2 / 1
>>> x / y
1.0

Is there a wonderland where we can enjoy both the composability of "/" and the pureness of safe_div? Yes, this is where monad shines.

Let's create a magical function, named "bind".

from typing import Callable

def bind(x: Maybe, some_safe_call: Callable[..., Maybe]) -> Maybe:
    if x.failed:
        return Maybe(failed=True)
    else:
        return some_safe_call(x.value)

bind takes two inputs: a Maybe object, and a function that takes a raw value and returns a Maybe object. First, it checks if the value stored in the first input is failed. If so, it simply returns a failed Maybe. Otherwise, it will pass its internally stored value to function some_safe_call, and directly return the output of it. With bind, we can now chain multiple safe_div's together:

>>> maybe_x = safe_div(4, 2)
>>> maybe_y = safe_div(4, 1)
>>> bind(maybe_x, lambda x: bind(maybe_y, lambda y: safe_div(x, y)))
Just 0.5

How does it work? Let's first look at the most inner call safe_div(x, y). We are already familiar with this. x and y are passed to safe_div as raw integers (not Maybe type). The more interesting part is the two lambda functions introduced. The first bind is called with input maybe_x and a lambda function. The purpose of the lambda function is to, by using bind, create an environment/context where the actual value x could be extracted from maybe_x and passed to function safe_div. Here, bind could also be seen as a variable assignment, which assigns the actual value stored in maybe_x to a temporary variable x. The same idea also applies to maybe_y, whose actual value is extracted by the second bind and passed along with x to safe_div.

Now, if we tweak the code with some new lines in between, it actually looks imperative in a way that each bind call acts like a reversed variable assignment, where the variable shows after its assigned value. This format is not a good style but it helps us to understand the binding relations.

bind(         maybe_x,      lambda x:
  ^              ^                 ^
assignment     value           variable
>>> maybe_x = safe_div(4, 2) # Just 2
>>> maybe_y = safe_div(4, 1) # Just 4
>>> bind(maybe_x, lambda x:  # bind the value stored in maybe_x to x
... bind(maybe_y, lambda y:  # bind the value stored in maybe_y to y
... safe_div(x, y)           # x=2 and y=4 are passed to safe_div
... )                        # ) matches ( from second bind
... )                        # ) matches ( from first bind
Just 0.5

# Notice that we are free to change the order in which "x" and "y" are defined

>>> maybe_x = safe_div(4, 2) # Just 2
>>> maybe_y = safe_div(4, 1) # Just 4
>>> bind(maybe_y, lambda y:  # bind the value stored in maybe_y to y (y <- 4)
... bind(maybe_x, lambda x:  # bind the value stored in maybe_x to x (x <- 2)
... safe_div(x, y)           # x=2 and y=4 are passed to safe_div
... )
... )
Just 0.5

With the magic function bind, we are now able to compose any number of safe_div together.

More examples:

>>> maybe_x = safe_div(4, 2) # Just 2
>>> maybe_y = safe_div(4, 1) # Just 4
>>> maybe_z = safe_div(1, 2) # Just 0.5
>>> maybe_n = safe_div(1, 0) # Nothing
Got exception: division by zero
>>> bind(maybe_n, lambda n:  # bind the value stored in maybe_n to n
... bind(maybe_y, lambda y:  # bind the value stored in maybe_y to y
... safe_div(n, y)
... )
... )
Nothing

>>> bind(maybe_x, lambda x:                 # bind the value stored in maybe_x to x
... bind(maybe_y, lambda y:                 # bind the value stored in maybe_y to y
... bind(maybe_z, lambda z:                 # bind the value stored in maybe_z to z
... bind(safe_div(y, x), lambda y_div_x:    # bind the value stored in the output of safe_div(y, x) to y_div_x
... safe_div(y_div_x, z)
... )
... )
... )
... )
Just 4.0

In all examples above, whenever a safe function, e.g. safe_div, need to access the raw value stored in a maybe object as its input, we can create a lambda function, whose input is "bind" to the raw value, and pass the function along with the Maybe object to bind, as a pattern looks roughly like this: bind(maybe_x, lambda x: ...).

bind is what makes Maybe a monad. It brings composability to safe_div, and essentially any functions that takes raw value as inputs and returns Maybe outputs.

Now that we have a good understanding of monad, and saw its implementation in Python, let's look at how it is implemented in Haskell.

First, let's create a Maybe data type in Haskell.

data Maybe a = Just a | Nothing deriving Show

In fact, Maybe is already implemented by Haskell's standard library and imported as a part of prelude.

Maybe has two constructors, Just and Nothing. Just takes one value, which could be any type. Nothing does not have input argument.

λ> Just 1
Just 1
λ> Just 2
Just 2
λ> Nothing
Nothing

With Maybe, we can create a function that performs division and returns a Maybe type, and use Nothing to represent the failed status of an operation.

safeDiv Int :: Int -> Int -> Maybe Int
safeDiv _ 0 = Nothing
safeDiv x y = Just (x `div` y)

The implementation is a bit different from Python example. In Python, we use try...except to catch exceptions. In Haskell, we directly check the value of divisor before dividing the numbers to avoid any exception. The purpose is to simplify the Haskell example so readers can focus more on the concept of Monad.

Now, let's try a couple of examples.

λ> safeDiv 4 2
Just 2
λ> safeDiv 4 1
Just 4
λ> safeDiv 4 0
Nothing

If we pass Maybe types to safeDiv, Haskell will complain:

λ> maybeX = safeDiv 4 2
λ> maybeY = safeDiv 4 1
λ> safeDiv maybeX maybeY

<interactive>:14:9: error:
     Couldn't match expected type Int with actual type Maybe Int
     In the first argument of safeDiv, namely maybeX
      In the expression: safeDiv maybeX maybeY
      In an equation for it’: it = safeDiv maybeX maybeY

<interactive>:14:16: error:
     Couldn't match expected type Int with actual type Maybe Int
     In the second argument of safeDiv, namely maybeY
      In the expression: safeDiv maybeX maybeY
      In an equation for it’: it = safeDiv maybeX maybeY

This is because safeDiv expects input type Int, but the actual input type Maybe Int. This is exactly the same problem we encountered in Python example, except that the error is thrown because of mismatched types.

We will use the same solution to solve this problem by introducing the bind function.

Let's define Monad more rigidly as a type class.

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

Symbol "(>>=)" is called "bind" in haskell. It has the exactly same purpose as bind function in our previous Python example. "(>>=)" takes two inputs. The first input is of type "m a", which will be matched to "Maybe Int" in the context of our example. The second input is a function, whose type is (a -> m b). In our case, it will be a lambda function.

instance Monad Maybe where
    Nothing  >>= f = Nothing
    (Just x) >>= f = f x

The implementation is almost identical to the one we saw in Python. We check if the first object is "Nothing", if so, we ignore the second argument, function f, and directly return "Nothing". Otherwise, we take the actual value stored in the first input, and pass it to function f and return the result.

Now, let's try to use "(>>=)".

λ> maybeX = safeDiv 4 2
λ> maybeY = safeDiv 4 1
λ> maybe_x >>= (\x -> (maybe_y >>= (\y -> safeDiv y x)))
Just 2

It seems to work! We are basically doing the same thing as we did in Python examples. Comparing the code side by side:

Python bind(maybe_x, lambda x: bind(maybe_y, lambda y: safe_div(y, x)))
Haskell maybe_x >>= (\x -> (maybe_y >>= (\y -> safeDiv y x)))

The haskell's expression could be written in multiple lines when wrapped in a function:

doWork :: Maybe Int -> Maybe Int -> Maybe Int
doWork maybeX maybeY = maybeX >>= \x ->
                       maybeY >>= \y ->
                       safeDiv x y
λ> maybeX = safeDiv 4 2
λ> maybeY = safeDiv 4 1
λ> doWork maybeY maybeX
Just 2

Does it look familiar? Yes, it looks really close to our Python implementation except that function "bind" is now replaced with the infix operator "(>>=)".

Because binding is used quite often in Haskell, the language provides a syntax sugar to free us from creating lambda functions and typing (>>=) over and over again. The syntax sugar starts with a Haskell keyword "do" and followed by expressions of "a <- Ma":

doWork :: Maybe Int -> Maybe Int -> Maybe Int
doWork maybeX maybeY = do
    x <- maybeX
    y <- maybeY
    safeDiv x y
λ> maybeX = safeDiv 4 2
λ> maybeY = safeDiv 4 1
λ> doWork maybeY maybeX
Just 2