## Thursday, August 30, 2007

### Y Combinator in Python I just finished reading The Little Schemer. What an awesome book! I love the Q&A format of this book and can't wait to read the next book in the series, The Seasoned Schemer.

Chapter 9 and 10 (the last two chapters) were really good. Chapter 10 is about writing a small scheme interpreter. But its Chapter 9 that I want to discuss in this blog post because Chapter 9 introduces the Y Combinator.

I've tried to understand it in the past and failed (the wikipedia page is completely unintelligible) , but this book explains it in a brilliant way. In fact it derives the Y combinator from the ground up, making sure each step is easy to understand.

As an exercise, I tried to write it in Python. This is how it came out:
```def Y(le):
def _anon(cc):
return le(lambda x: cc(cc)(x))
return _anon(_anon)```
If you don't know the Y combinator already, that probably looks quite cryptic.

To give a simplified explanation, the Y combinator is a function that takes one function as input and creates a recursive version as an output. Maybe an example will illustrate the point better:
```def _1(factorial):
def _fn(n):
if n == 0: return 1
else:
return n*factorial(n-1)
return _fn```
Take a look this function. The function name is _1. The function body has something that looks like a recursive factorial implementation, except that it never calls itself (remember, the function name is _1). Instead, it recurses on "factorial" which is a parameter of the function _1.

Basically, the "factorial" parameter is a function, so the meaning of _1 is:
• If n is zero, return one
• else, take the factorial parameter, and call that function with parameter n-1, and after that returns, multiply with n and return the value
Exploring a bit more:
```def error(n): raise Exception

f = _1(error) # passing function "error" as the parameter
f(0)   # prints 1
f(1)   # Exception```
The above function f is passing error as a parameter to _1. So if n is zero, it returns 1. Otherwise, it goes to the else part and calls the function that we passed as parameter, in this case error, which raises an exception.

In order for this to be recursive, we don't want it to call error, we want it to call the same function again. What if we passed the same function as a parameter? So in the else part, instead of calling error, it would call itself. Something like this
```f = _1(_1(error))
f(0)   # prints 1
f(1)   # prints 1
f(2)   # Exception

f = _1(_1(_1(_1(error))))
f(0)   # prints 1
f(1)   # prints 1
f(2)   # prints 2
f(3)   # prints 6
f(4)   # Exception```
Hmm, in each case, the recursion finally stops with an exception when it encounters the error function. A truly recursive version would be like
`f = _1(_1(_1(_1(_1...... forever`
Well, that is basically what the Y combinator does. Using some function passing magic, it converts _1 into the forever recursive version. How?? Well thats the magic!! Its a bit complicated to explain here. Get the book and read Chapter 9. But it works! Check this out:
```f = Y(_1)
f(0)   # prints 1
f(1)   # prints 1
f(5)   # prints 120
f(10)   # prints 3628800```
Amazing isn't it? And the even more amazing thing is that it is not specific to factorial. See this
```def _2(length):
def _fn(alist):
if not alist: return 0
else:
return 1 + length(alist[1:])
return _fn

f = Y(_2) # calculate length of a list
f([])   # prints 0
f([1,2,3,4,5])   # prints 5```
Woohoo!! On a roll now..
```def _3(reverse):
def _fn(alist):
if not alist: return []
else:
return reverse(alist[1:]) + [alist]
return _fn

f = Y(_3) # reverse a list
f([])   # prints []
f([1,2,3])   # prints [3,2,1]```
You can take any recursive function, and rewrite it in the above style and the Y combinator will make a recursive version of it. How cool is that?

This post is a part of the selected archive.

Antoine said...

For those of you wondering "Well, I guess that's cool, but what's the point?" check this out:

One advantage to writing recursive functions in this style is that it allows you to combine the primitive recursive step with various wrappers, and then "fix" the result into a normal recursive function, allowing you in insert behavior "in between" recursive calls. Anonymous said...

I am pretty sure that is incorrect. You are defining anon with a name which you cannot do with a pure y combinator.

Siddhi said...

Well that is how it is normally defined in a functional language. Since python doesn't have multiline lambdas, this way is more readable. For a single line version without named functions, see this comment on reddit

Leo Richard Comerford said...

Fortunately, Chapter 9 is the free sample chapter. (It's in Postscript format.)

_ said...

just to pique your interest :)
the y combinator in joy is

dup i

Winterstream said...

Nice Siddhi! It helps to see it used so clearly. Anonymous said...

def Y(f):
x = lambda: f(x)
return x() Anonymous said...

Y is always invoked with some function, in the above examples, _1, _2, or _3.

These "argument-functions" (i.e., functions which have a form similar to _1, _2, _3)... do they have a name?

If there a term which means "a function which can serve as the argument to the Y-combinator?"

I'm trying to tease out for myself exactly what these argument-functions have in common, and I can't quite put my finger on it. I think having a name for them would help.

Unknown said...

Y = lambda le: le(lambda x: Y(le)(x))

here is my research + upgrade of the topic one Y

BTW thanks for the great post, it makes me understand Y-combinator and python lambda functions much more.

Grizly: Your version doesn't just apply Y when creating the recursive function, but uses it every time the recursion happens. A minimal version using named functions would be:

def Y(f):
....def Yf(*args):
........return f(Yf)(*args)
....return f(Yf)

siva said...
Farhan.Jee said...

The training of models can be done by employing various methods. While some of these methods are simple, like regression, the others are complex, like neural nets. While discovering patterns out of data, machine learning tries to look for some patterns or search for some data associations in a situation where tagged data is absent. data science course syllabus

Ashok said...

Good to become visiting your weblog again, it has been months for me. Nicely this article that i've been waited for so long. I will need this post to total my assignment in the college, and it has exact same topic together with your write-up. Thanks, good share.

data science course in India

Ashok said...

I just got to this amazing site not long ago. I was actually captured with the piece of resources you have got here. Big thumbs up for making such wonderful blog page!
Artificial Intelligence Course