Question

I’m trying to create prime number generator in one-line of Python just as a fun exercise.

The following code works as expected, but it is too slow:

primes = lambda q: (i for i in xrange(1,q) if i not in [j*k for j in xrange(1,i) for k in xrange(1,i)])
for i in primes(10):
   print i,

So I I tried to do it by only checking up to the square-root of j and k:

primes = lambda q: (i for i in xrange(1,q) if i not in [j*k for j in xrange(1,int(round(math.sqrt(i)+1))) for k in xrange(1,int(round(math.sqrt(i)+1)))])
for i in primes(10):
   print i,

But it outputs: 2 3 5 6 7 8

So there must be something wrong with my indices j and k, but I haven’t got a clue.

Answer 1

That’s not the Sieve of Eratosthenes, even though it looks like it is. It is in fact much worse. The Sieve is the best algorithm for finding primes.

See http://en.wikipedia.org/wiki/Sieve_of_Eratosthenes

edit: I’ve modified https://stackoverflow.com/a/9302299/711085 to be a one-liner (originally it was not the real Sieve, but now it is… probably…):

reduce( (lambda r,x: r-set(range(x**2,N,x)) if (x in r) else r), 
        range(2,N), set(range(2,N)))

Demo:

>>> primesUpTo(N): lambda N: reduce(...)
>>> primesUpTo(30)
{2, 3, 5, 7, 11, 13, 17, 19}

---

Sadly I think that while this would be efficient in a functional programming language, it might not be as efficient in python due to non-persistent (shared-state and immutable) data structures, and any sieve in python would need to use mutation to achieve comparable performance. We can still cram it into a one-liner if we desperately wanted to. But first…

Normal sieve:

>>> N = 100
>>> table = list(range(N))
>>> for i in range(2,int(N**0.5)+1):
...     if table[i]:
...         for mult in range(i**2,N,i):
...             table[mult] = False
... 
>>> primes = [p for p in table if p][1:]
>>> primes
[2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97]

We can now define and call anonymous functions on the same line, as well as the hack of [...].__setitem__ to do inline mutation, and the hack of ... and foo to evaluate ... while returning foo:

>>> primesUpTo = lambda N: (lambda table: [[table.__setitem__(mult,False) for mult in range(i**2,N,i)] for i in range(2,int(N**0.5)+1) if table[i]] and [p for p in table if p][1:])(list(range(N)))
>>> primesUpTo(30)
[2, 3, 5, 7, 11, 13, 17, 19, 23, 29]

Proceed to cringe in horror, the one-liner expanded (oddly beautiful because you could almost directly translate the control flow, yet a terrible abuse of everything):

lambda N:
    (lambda table: 
        [[table.__setitem__(mult,False) for mult in range(i**2,N,i)] 
            for i in range(2,int(N**0.5)+1) if table[i]] 
        and [p for p in table if p][1:]
    )(list(range(N)))

This one-liner mutating version gave up at around 10⁸ on my machine, while the original mutating version gave up at around 10⁹, running out of memory (oddly).

The original reduce version gave up at 10⁷. So perhaps it is not that inefficient after all (at least for numbers you can deal with on your computer).

edit2 It seems you can abuse side-effects more concisely as:

reduce( (lambda r,x: (r.difference_update(range(x**2,N,x)) or r)
                     if (x in r) else r), 
        range(2,N), set(range(2,N)))

It gives up at around 10⁸, the same as the one-liner mutating version.

edit3: This runs at O(N) empirical complexity, whereas without the difference_update it ran at O(n^2.2) complexity.

Limiting the range that is reduced over, to the sqrt of the upper limit, and working with odds only, both result in additional speed-ups (2x and 1.6x correspondingly):

reduce( (lambda r,x: (r.difference_update(range(x*x,N,2*x)) or r)
                     if (x in r) else r), 
        range(3, int((N+1)**0.5+1), 2),
        set([2] + range(3,N,2)))