Go West Young Man!

I have had to find it out the hard way, but blogger sucks when it comes to math blogging. It wouldn't be all that bad, were it not because on the other side of the road things are soooo very much easier. And while I will miss google analytics, the native support for latex and syntax highlighting make up for it more than enough. So I'm moving this blog to wordpress...

You can find the new blog here, and suscribe to the RSS feed here.

For starters I have redone the much improvable first post of the series on the FFT, which I have split in two, one on the DFT, the other starting discussion on the FFT, plus an unconnected post on computing the GCD...

I will be moving all the content in this blog over to the new one, but have no intention of deleting anything from this one.

The Cooley-Tukey FFT Algorithm

I'm currently a little fed up with number theory, so its time to change topics completely. Specially since the post on basic integer factorization completes what I believe is a sufficient toolkit to tackle a very cool subject: the fast Fourier transform (FFT).

I have some mixed feelings about how does the Fourier transform qualify for the "uncomplicated complexity" rule I imposed on myself when starting this blog. There certainly is a lot of very elaborate math behind Fourier analysis, but I think that if you skip all the mathematical subtleties and dive head first into a "the Fourier transform gives you the frequencies present in the signal" type of intuitive description, the journey from the naive implementation of the discrete Fourier transform (DFT) to the many flavors of the FFT is a very enjoyable one, which requires only a limited amount of complication, but gentle doses of complexity. Which is exactly what I'm after...

So I will take the definition of the DFT for granted, and concentrate on the algorithms to compute it efficiently. There will have to be a few more leaps of faith, but I will try to mark them clearly, so you can whisper "amen" and keep going without a second thought. Although things may unfold differently, I think there is material for about half a dozen posts. Towards the end I may devote some time to optimizing the code, with things such as avoiding recursion, or doing the calculations in place, but most of my efforts will be devoted to optimizing the math behind the code, not the code itself. So the resulting code will probably be suboptimal: if you spot a potential improvement your comments are welcome and encouraged.

The Discrete Fourier Transform
Without further explanation, we will begin by writing down the analytical expression of the DFT,

and of its corresponding inverse transform,

With python's built-in support for complex arithmetic, there really isn't much mistery in turning these two formulas into two python functions, or as I have chosen, one with an inverse switch:

from __future__ import division
import math
import time

def dft(x, inverse = False, verbose = False) :
    t = time.clock()
    N = len(x)
    inv = -1 if not inverse else 1
    X =[0] * N
    for k in xrange(N) :
        for n in xrange(N) :
            X[k] += x[n] * math.e**(inv * 2j * math.pi * k * n / N)
        if inverse :
            X[k] /= N
    t = time.clock() - t
    if verbose :
        print "Computed","an inverse" if inverse else "a","DFT of size",N,
        print "in",t,"sec."
    return X

Of course, having two nested loops of the size of the input means that computation time will grow with N², which is extremely inefficient.

The Radix-2 Cooley-Tukey Algorithm with Decimation in Time
How can that be improved? If the size of the input is even, we can write N = 2·M and it is possible to split the N element summation in the previous formulas into two M element ones, one over n = 2·m, another over n = 2·m + 1. By doing so, we arrive at

for the direct transform, and

for the inverse.

If you look carefully at this formulas, you'll notice we have just split our size N transform, be it direct or inverse, into two others of half the size, one over even indexes, the other over odd ones. The factor multiplying the second partial transform is known as a twiddle factor, and introduces a little overhead to the total timing, but overall we have just managed to cut the total time in half. And if the input size happens to be a power of two, we need not stop here, for as long as it is divisible by two, we can repeat the process iteratively, leading to a time of N log N.

To implement this recursive approach we need to take care of a few more details, though. First, if you look at the previous formulas, variable k in the direct transform, or n in the inverse, runs all the way to N = 2·M, rather than to M, as one would expect in the standard DFT. By taking advantage of the fact that the exponential function is periodic with period 2πi, it is quite easy to show that, for k (or n) larger than M, the resulting value will be the same as the one obtained for k-M (or n-M).

The other open question is at what point to stop the recursion. The simplest approach is to keep going until the input size is odd, and then call the naive DFT function. If the size happens to be a power of two, this call will not happen until the input size comes down to one. Since the DFT of a single sample signal is the same signal unchanged, that is a wasteful way of coding. But since it clarifies the logic behind the algorithm, I will leave those optimizations for a later stage.

With this caveats in mind, this FFT algorithm can be coded in python as follows:

from __future__ import division
import math
import time

def fft_CT(x, inverse = False, verbose = False) :
    t = time.clock()
    N = len(x)
    inv = -1 if not inverse else 1
    if N % 2 :
        return dft(x, inverse, False)
    x_e = x[::2]
    x_o  = x[1::2]
    X_e = fft_CT(x_e, inverse, False)
    X_o  = fft_CT(x_o, inverse, False)
    X = []
    M = N // 2
    for k in range(M) :
        X += [X_e[k] + X_o[k] * math.e ** (inv * 2j * math.pi * k / N)]
    for k in range(M,N) :
        X += [X_e[k-M] - X_o[k-M] * math.e ** (inv * 2j * math.pi * (k-M) / N)]
    if inverse :
        X = [j/2 for j in X]
    t = time.clock() - t
    if verbose :
        print "Computed","an inverse" if inverse else "a","CT FFT of size",N,
        print "in",t,"sec."
    return X

The speed gain with this approach, restricted for now to power of two sizes, are tremendous even for moderately large input sizes. For instance, a 2²⁰ (~10⁶) item input can be processed with the FFT approach in a couple of minutes, while the projected duration using the naive DFT will be more like a couple of months...

Naive Integer Factorization

After three posts (1, 2, 3) on calculating prime numbers, it is probably worth putting that knowledge to a more useful task. As we will see in a near future, integer factorization, i.e. breaking down a (composite) number into its prime factors is one such task. In purity, factoring a number n is simply decomposing it as the product of two smaller non-trivial, i.e. different from 1 and n itself, divisors. But by repeatedly factoring the divisors one will eventually come up with a unique set of primes which, when multiplied together, render the original number, or so says the fundamental theorem of arithmetic... The point is, we will consider factorization a synonym of prime decomposition, be it formally correct or not.

There are some very sophisticated methods to factor very large numbers, but they use a lot of extremely complex math, so I doubt they will ever find their way onto this blog. So we are going to be left with the naive, straightforward approach as our only option, although I will try to give it an efficiency boost. What is this naive approach? Trial division, of course: given a number n, we know that its smallest factor will be smaller than the square root of n, so we can simply try and see if any of those numbers divide it. No, I will not try to code that yet... If you have read the entries on determining prime numbers, it should come as no surprise that we really do not need to do trial division by all numbers smaller than the square root of n, but only by the primes within. This is a consequence of the fact that, if a composite number divides n, then each of the prime factors of that composite number will also divide n. According to the prime number theorem the number of primes below x is asymptotic to x / log x. So by limiting our trials to prime numbers we can reduce the number of tests from n^1/2 to something around 2 n^1/2 / log n. If we rescue the primeListSofE function from the post on the sieve of Erathostenes, a python implementation of naive factorization could look something like this...

from time import clock

def factor(n, verbose = False) :
    """Returns all prime factors of n, using trial division by prime
    numbers only. Returns a list of (possibly repeating) prime factors
    """
    t = clock()
    ret =[]
    nn = n
    maxFactor = int(n**0.5)
    primes = primeListSofE(maxFactor, verbose)
    for p in primes :
        while nn % p == 0 :
            nn //= p
            ret += [p]
        if nn == 1 :
            break
    if nn != 1 :
        ret += [nn]
    t = clock() - t
    if verbose :
        print "Calculated factors of",n,"in",t,"sec."
    return ret

While this function will be about as good as we can make it for numbers which are the product of two large prime factors, it will be terribly inefficient for most numbers. Consider, as an extreme example, that we are trying to factor 2⁵⁵ ~ 3.6·10¹⁶. We would first calculate all primes up to 1.9·10⁸, a challenging feat in itself with our available tools, only to find out that we only needed the first of those primes, i.e. 2. Taking into account that 50% of all numbers are divisible by 2, 33% are divisible by 3, 20% are divisible by 5... it doesn't seem wise to disregard the potential time savings. What we can do to profit from this is to do the trial division checks at the same time as we determine the prime numbers, updating the largest prime to test on the fly. This has to be done in two stages, the first while we sieve up to n^1/4, the second while we search the rest of the sieve up to n^1/2 searching for more primes. The following Franken-code has been written mostly by cut-and-paste from primeListSofE and factor, which hopefully hasn't affected its readability much:

from time import clock

def factorAndSieve(n, verbose = False) :
   """Returns all prime factors of n, using trial division while sieving
   for primes. Returns a list of (possibly repeating) prime factors
   """
   t = clock()
   ret =[]
   nn = n
   while nn % 2 == 0 : # remove 2's first, as 2 is not in sieve
       nn //= 2
       ret += [2]
   maxFactor = int(nn**0.5)
   maxI = (maxFactor-3) // 2
   maxP = int(maxFactor**0.5)
   sieve = [True for j in xrange(maxI+1)]
   i = 0
   for p in xrange(3, maxP+1,2) : # we then sieve as far as needed
       if p > maxP :
           break
       i = (p-3) // 2
       if sieve[i] :
           while nn % p == 0 :
               nn //= p
               ret += [p]
               maxFactor = int(nn**0.5)
               maxI = (maxFactor-3) // 2
               maxP = int(maxFactor**0.5)
           if nn == 1 :
               break
           else :
               i2 = (p*p - 3) // 2
               for k in xrange(i2, maxI+1, p) :
                   sieve[k] = False
   index = i
   for i in xrange(index, maxI+1) : # and inspect the rest of the sieve
       if i > maxI :
           break
       if sieve[i] :
           p = 2*i + 3
           while nn % p == 0 :
               nn //= p
               ret += [p]
               maxFactor = int(nn**0.5)
               maxI = (maxFactor-3) // 2
               maxP = int(maxFactor**0.5)
           if nn == 1 :
               break
   if nn != 1 :
       ret += [nn]
   t = clock() - t
   if verbose :
       print "Calculated factors of",n,"in",t,"sec."
       print "Stopped trial division at",2*i+3,"instead of",int(n**0.5)
   return ret   

This new code will very often be much faster than the other one, but at times it will be just about as slow as in the other case, or even slower, since the mixing of both codes introduces some inefficiencies. The most extreme examples of such cases would be a prime number, or the square of a prime number on one side, and a power of 2 on the other one.



The graph above plots times to calculate the factors of numbers between 10⁶ and 10⁶ + 100. Prime numbers in this interval stick out as the red dots among the blue ones: 10⁶ +3, +33, the twin primes +37 and +39, +81 and +99. And numbers with many small prime factors populate the bottom of the red cloud.

If the above graph is not enough to convince you of the benefits of the second approach, maybe this timings for very large numbers will:

>>> factor(10**15+37,True)
Calculated primes to 31622776 in 6.760 sec.
Calculated factors of 1000000000000037 in 8.466 sec.
[1000000000000037L]
>>> factorAndSieve(10**15+37,True)
Calculated factors of 1000000000000037 in 8.666 sec.
Stopped trial division at 31622775 instead of 31622776
[1000000000000037L]

>>> factor(2**55,True)
Calculated primes to 189812531 in 42.811 sec.
Calculated factors of 36028797018963968 in 43.261 sec.
[2, ..., 2]
>>> factorAndSieve(2**55,True)
Calculated factors of 36028797018963968 in 8.632e-05 sec.
Stopped trial division at 3 instead of 189812531
[2, ..., 2]

Prime Numbers 3. Wheel Factorization

In an earlier post, I described the sieve of Eratosthenes, a very fast algorithm to calculate all primes below a given limit, n. At least when programming it in python, speed is not an issue at all with this algorithm, because we start having memory problems much earlier, due to the fact that we need to keep in memory a list of size n...

This is specially problematic with python, because although all we need is a list of booleans (is prime / is not prime), which could fit in a bit each, I'm afraid python assigns several bytes to each. How many of them is probably machine dependent, but I'm almost certain it is at least 4 of them. So it is very likely that we could improve memory usage by a factor of 32, maybe more, if we could get around this. There are not-so-standard libraries that do just that, but I've sworn not to use them, and one day I may expose my failed attempts at using very long ints as bit arrays to the world, but python is definitely not the best language to mess around with low level optimizations, so I'm going to leave this line of thinking aside for the time being.

I also presented a very naive optimization, which simply discards all even numbers, which we already know are not prime, and thus cuts in half the memory requirements. More formally, what was being done there was to partition all numbers into two groups: those of the form 2·n, and those of the form 2·n+1, and then discard the former, since no prime of that form exists, and keep the latter only. This line of thinking is much more promising, as it leads to a very neat way of further reducing the memory required, known as wheel factorization.

For some odd reason that escapes my shallow understanding, the standard description of this algorithm places much emphasis on setting all numbers in circles, and then eliminating what it calls spokes. I personally don't see any benefit of such a mental image, specially since the number are going to end up arranged in a rectangular array. So for the sake of clarity, I will not honor tradition, and spare you the circles and the spokes.

We begin by choosing a first few prime numbers. Say the first three: 2, 3 and 5. We multiply them together to get M = 30. We are now going to partition the set of all numbers into M groups, of the form M·m + k_i, where the k_i are the numbers 1 through M, both inclusive. Most of these subsets happen to hold no prime number, and thus need not be considered at all. To find out which of them, remember that M is the product of several prime numbers, M = p₁·p₂ ... p_n. Therefore, if k_i is a multiple of any of those prime numbers, i.e. k_i = p_j·k'_i, we can always extract the common factor p_j from both summands in M·m + k_i, which means that all numbers in that subset are composite. These subsets, which we discard, are very closely related to the spokes I was telling you about a while ago.

If M = 30, of the 30 potential k_i we are left with just 8 (1, 7, 11, 13, 17, 19, 23 and 29), so we have cut memory required to about 27% of the straightforward implementation of the sieve of Eratosthenes. If we use the first 4 primes, we then have M = 210, and of the 210 potential k_i we only need to keep 48. This improves memory use, which now is roughly 23% of the original. But you can probably see a law of diminishing returns at wortk here: the improvement in memory use gets smaller and smaller with every prime you add to the calculation of M. You can find some more data regarding this at the end of this link.

Of course we are still going to have to sieve each of our subsets, in much the same way we did with the sieve of Eratothenes. But sieving all multiples of a prime p when you have several subsets to consider is anything but trivial. Lets see how to tackle it...

The question to answer is, what integer values of m fulfill the equation M·m + k_i = p·n? Using modular arithmetic, this equation is equivalent to finding M·m = -k_i (mod p). This requires calculating M^-1 (mod p), the modular multiplicative inverse of M modulo p, a subject already treated in a previous post. Given that, we can simply write m_i = M^-1·(-k_i) (mod p), and therefore determine all multiples of p in the subset of k_i as m = p·j + m_i.

Last, but not least, we want to start sieving at the smallest multiple of p in each subset that is larger than or equal to p², and so we want to increment m_i with p·j_i, where j_i >= (p²-k_i- M·m_i) / (m·p).

Lets try to put all this ideas together in a python script. To make it work, you will need the primeListSofE function from this post, as well as the extEuclideanAlg and modInvEuclid ones from this other post...

def primeListWF(n, m, verbose = True) :
    """Returns a list of prime numbers smaller than or equal to n,
    using wheel factorization, with a wheel size of the product of
    all primes smaller than or equal to m
    """
    t = time()
    # We use the standard sieve of Eratosthenes first...
    primes = primeListSofE(m)
    M = 1
    for p in primes :
        M *= p
    if verbose :
        print "Size of factorization wheel is",M
    # We now sieve out all multiples of these primes,including
    # the primes themselves, to get the k[i]
    sieve = [True] * M
    for p in primes :
        sieve[p-1] = False
        for j in xrange(p*p,M+1,p) :
            sieve[j-1] = False
    k = [j+1 for j in range(M) if sieve[j]]
    if verbose :
        print "Memory use is only",len(k),"/",M,"=",len(k)/M
    N = n
    if N %M :
        N += M
    N //= M
    maxP = int(sqrt(M*N))
    if verbose:
        print "Primes will be calculated up to",M*N,"not",n

    sieve = [[True]*N for j in range(len(k))]
    sieve[0][0] = False # Take care of the non primality of 1
    for row in range(N) :
        baseVal = M * row
        for subset in range(len(k)) :
            if sieve[subset][row] :
                p = baseVal + k[subset]
                primes += [p]
                if p > maxP :
                    continue
                # Sieve all multiples of p...
                invMp = modInvEuclid(M,p)
                for i in range(len(k)) :
                    mi = invMp * (-k[i] % p)
                    mi %= p
                    # ...starting at p**2
                    ji = (p*p - k[i] - M*mi)
                    if ji % (M*p) :
                        ji += M*p
                    ji //= M*p
                    mi+= ji*p

                    for r in range(mi,N,p) :
                        sieve[j][r] = False
    if verbose :
        print "Calculated primes to",n,"in",time()-t,"sec."
        
    return primes

You can play around with this algorithm and look at the time required. It probably depends on the value of n, but there seems to be a sweet spot for speed somewhere around m = 7 (and thus M = 210) or m = 11 (M = 2310). This being a general purpose wheel factorization algorithm, it has speeds that are faster than the unoptimized sieve of Eratosthenes code, but it cannot beat the optimized code, which in fact is a particular case of wheel factorization, for m = M = 2.

On the other hand, using m = 7 I have managed to compute all primes up to 10⁸, a feat unreachable for the previous code, at least without finding a workaround the MemoryError they produce. 10⁹ is still out of reach, though...

Modular Multiplicative Inverse

Now that we are all proficient with modular exponentiation thanks to my previous post, it is time to tackle a more complicated issue, still in the realm of modular arithmetic.
While addition, subtraction and multiplication are "compatible with the congruence relation" introduced by modular arithmetic, the same doesn't happen with division. This means that solving a simple equation such as a·x = 1 (mod m) is anything but trivial. In fact, determining x = a^-1 (mod m) that fulfills that equation is the whole object of this post...

There is, as always, the brute force approach, which I always like to visit first, so that I can later congratulate myself on how much the efficiency has improved with my deep insight. In this particular case, brute force means trying every number between 1 and m-1, and see if the m-modulo of its product with a is 1. :

def modInv1(a,m) :
  """Computes the modular multiplicative inverse of a modulo m,
  using brute force
  """
  a %= m
  for x in range(1,m) :
      if a*x%m == 1 :
          return x
  return None

Do notice that the possibility of no multiplicative inverse existing is contemplated in the code. This will actually be the case if a and m have any common factors, i.e. if they are not coprime. So the previous code could serve as a moron's approach to determining if two numbers are relatively prime. Not that I recommend it, though, as it should be clear that the code I have just produced will be very inefficient for large values of m.

The Extended Euclidean Algorithm
As stated, we are after a number x that fulfills a·x = 1 (mod m). This can of course be written as well as a·x = 1 + m·y, which rearranges neatly into a·x - m·y = 1. Since x and y need not be positive, we can write it as well in the standard form, a·x + m·y = 1. If a and m are coprime, this is a particular instance of a·x + b·y = gcd(a,b), where gcd(a,b) is the greatest common divisor of a and b. This equatiin is known as Bézout's identity, and can be efficiently solved using the extended Euclidean algorithm. To find a solution to our particular equation, this algorithm would proceed as follows:

1. Rewrite the equation as a_n·x_n + m_n·y_n = 1, starting with n = 0
2. Compute the quotient, q_n, and the remainder, r_n, of a_n divided by m_n, and rewrite the original equation as (m_n·q_n + r_n)·x_n + m_n·y_n =1
   
3. Rearrange the terms in the equation to get m_n·(q_n·x_n+y_n) + r_n·x_n = 1
   
4. Do the changes of variables
   
   • x_n+1 = q_n·x_n+y_n,
     
   • y_n+1 = x_n,
     
   • a_n+1 = m_n, and
     
   • m_n+1 = r_n
     
5. Repeat the process, until a_n = 1, and m_n = 0.
   
6. Such equations has the trivial solution x_n = 1, y_n = 0.
   
7. Work your way back to x₀ and y₀, noting that
   
   • x_n-1 = y_n,
     
   • y_n-1 = x_n - q_n-1·y_n
     
8. x₀ is the number we are after.
   

There is a certain component of faith in believing that, eventually, a_n = 1 and m_n = 0. It is in fact not always the case, as if a and m are not coprime, then the values we will arrive at are m_n = 0, but a_n = gcd(a,m). As a matter of fact, the gcd is normally computed using the Euclidean algorithm, dropping the adjective extended, which basically does the same as above, but without the back-tracking of the x_n's and the y_n's. With all of this in mind, we can write two functions, one calling the other, as follows:

def extEuclideanAlg(a, b) :
  """Computes a solution  to a x + b y = gcd(a,b), as well as gcd(a,b)
  """
  if b == 0 :
      return 1,0,a
  else :
      x, y, gcd = extEuclideanAlg(b, a % b)
      return y, x - y * (a // b),gcd

def modInvEuclid(a,m) :
  """Computes the modular multiplicative inverse of a modulo m,
  using the extended Euclidean algorithm
  """
  x,y,gcd = extEuclideanAlg(a,m)
  if gcd == 1 :
      return x % m
  else :
      return None

The gains in speed are tremendous with this new approach. Again, write your own testing function as an exercise, but mine spits out the following results...

>>> testModInv(1234,5,10000)
Got 4 using method 1 in 0.0160000324249 sec.
Got 4 using Euclidean algorithm in 0.0309998989105 sec.

>>> testModInv(1234,56789,10000)
Got 31800 using method 1 in 59.4000101089 sec.
Got 31800 using Euclidean algorithm in 0.047000169754 sec.

So while brute force is good enough, or even the best choice, for very small values of m, as soon as it grows, the benefits of the more elaborate algorithm are all too obvious. According to wikipedia, the algorithm we have coded has complexity O(log(m)²), i.e. time required to compute the modular multiplicative inverse is proportional to log(m)².

Modular Exponentiation

I have a post cooking on wheel factorization, yet another sieving algorithm for the computation of all primes below a given number n. It turns out that while implementing it, one comes across a number of not so obvious preliminary questions. It is a rule of this blog not to resort to any non standard library, and I feel it would be breaking the spirit of that rule to post code which hasn't been thoroughly explained in advance. So in order not to make that soon-to-be post on prime numbers longer than anyone could bear, I find myself in need of covering a couple of preliminaries, and so here comes this thrilling post on modular exponentiation...

What we are after is determining a^b (mod m), or to put it in plain English, the remainder of a to the power of b, divided by m. Doesn't seem much, does it? We could just go ahead and write:

def modExp1(a, b, m) :
   """Computes a to the power b, modulo m, directly
   """
   return a**b % m

The problems begin if a and b are both big. Python's integers don't overflow, and we all love it for it, but the memory juggling it has to do to accommodate a huge number in memory does come at a speed cost. Specially if m is small, it definitely seems a little overkill to keep a zillion digit number in memory, given that the answer will be smaller than m...

We can, and will, use modular arithmetic to our advantage. See, wikipedia says that "the congruence relation (introduced by modular arithmetic) on the integers is compatible with the multiplication operation of the ring of integers," and if it's on the internet it has to be true, right? Translated to plain English, the idea is that if we want to take the product of two numbers, a and b, modulo m, we can first multiply them, and then take the m-modulo of the result, or we can take the m-modulo of a and/or b, prior to multiplication, and the m-modulo of the result will still be the right answer. So the least a conscious programmer should do is to rewrite the previous function as:

def modExp2(a, b, m) :
    """Computes a to the power b, modulo m, a little less directly
    """
    return (a%m)**b % m

Of course, if b is large enough, you may still find yourself in the realm of zillion digit numbers, which is a place we don't really want to go. It turns out then that it may be a good option to remember that exponentiation is just repeated multiplication. If we multiply a by itself b times, and take the m-modulo of the result, the larger number we will ever have in memory will be m². It doesn't really seem right to put a loop of b iterations into the code, since what we are worried about are large b's, but lets do it anyway:

def modExp3(a, b, m) :
    """Computes a to the power b, modulo m, by iterative multiplication
    """
    ret = 1
    a %= m
    while b :
        ret *= a
        ret %= m
        b -= 1
    return ret

The only reason why this last function deserves a place here, is because it helps introduce the really cool algorithm that this whole post really is about, binary exponentiation. In John D. Cook's blog there is a great explanation on how it could be implemented iteratively, based on the binary representation of the exponent, but I'll go for a recursive version, which is essentially the same, but which I find easier for my brain's wiring...

So what we are about to do is, when asked to compute a^b, check if b is even, in which case we will compute it as the square of a^b/2. If b is odd, we will compute it as a a^b-1. If b is 0, we will of course return 1. If we also throw in all the modulo m stuff I have been discussing previously, we finally arrive at THE function to do modular exponentiation:

def modExp(a, b, m) :
    """Computes a to the power b, modulo m, using binary exponentiation
    """
    a %= m
    ret = None
    
    if b == 0 :
        ret = 1
    elif b%2 :
        ret = a * modExp(a,b-1,m)
    else :
        ret = modExp(a,b//2,m)
        ret *= ret
        
    return ret%m

Now that we have all four versions of the same function written, we can do a little testing, I will leave writing the testing code as an exercise for the reader (I've been wanting to say this since my first year of university...), but when computing 1234 to the power 5678, modulo 90, a 1000 times with each function, these is the performance I get:

>>> testModExp(1234,5678,90,1000)
Got 46 using method 1 in 3.73399996758 sec.
Got 46 using method 2 in 0.546999931335 sec.
Got 46 using method 3 in 1.25 sec.
Got 46 using THE method in 0.0150001049042 sec.

Results which really don't require much commenting...

EDIT: There is some very interesting information in the comments: python comes with a built-in pow function, which can take two or three arguments, the third being, precisely, the modulo. It probably implements an algorithm similar to what has been presented, but is more than an order of magnitude faster.

Prime Numbers 2. The Sieve of Erathostenes

The problem with the algorithms described in the previous post lies on the very large amount of redundant testing that goes on. Because once we have checked the primality of a certain number p, we know that 2p, 3p, 4p, ..., np are composites. So what is the point of checking those numbers at all? When you follow this line of thinking to the very end, it turns out you don't have to do any trial division at all! The resulting algorithm is attributed to Erathostenes, and it employs a technique, known as sieving, which may come in handy in many other problems. But enough for a presentation, lets get going with the Sieve of Erathostenes.

All the information we need to begin with, is that 2 is the first prime. This means, following the previous approach, that no other even number can be prime. To keep track of this information, we begin by writing a list of all the numbers between 2 and the desired limit, say N:



We now draw a circle around 2, because it is prime, and then strike out all even numbers, i.e. all multiples of 2:



Now, there is a good and a bad way to strike out all multiples of a number from a list. The wrong way is to do trial division to find out if it is a multiple. The right way is to use the fact that two consecutive multiples of a number differ by that exact number. Writing it out in python, you can check the tremendous gain for yourself:

from time import time

def strikeOut(mul, n) :

# note that list index i corresponds to number i+2
sieve = [True for j in xrange(2,n+1)]

# The bad way to sieve
t = time()
for j in xrange(2,n+1) :
   if j % mul == 0 :
       sieve[j-2] = False
t = time() - t
print "Sieved (bad!!!) all multiples of",mul,"below",n,"in",t,"sec."

sieve = [True for j in xrange(2,n+1)]

# The good way to sieve
t = time()
for j in xrange(mul,n+1,mul) :
       sieve[j-2] = False
t = time() - t
print "Sieved (good!!!) all multiples of",mul,"below",n,"in",t,"sec."

No, I'm not taking you, dear reader, for an idiot: you can find sieving algorithms on fist pages of google searches that do sieving this very wrong way...

But lets get back to primes. We had our list, with 2 circled and all its multiples crossed out. The first unmarked number we come across is 3. Just a coincidence that 3 is a prime number? Not really: it it is unmarked because it is not a multiple of any smaller number. Or to put it another way: it is only divisible by 1 and itself, which is hte very definition of prime... So we circle 3 and cross out all its multiples:



I've underlined the multiples of three that where already crossed out. The first number which gets crossed out for the first time is 9. Will it be just a coincidence that 9 = 3²? Of course not! We have already crossed out all multiples of numbers smaller than 3, and 3² is the first multiple of 3 which is not a multiple of any smaller number.

So we keep going down our list, and find 5 to be the next unmarked number in the list, which of course is prime, so we circle it, and then cross out all of its multiples. But as we saw before, rather than starting at 10, and crossing out the already crossed out 10, 15 and 20, we proceed directly to 5², that is 25.

This goes on until we've searched the whole list, circling numbers and crossing out its multiples. But since we start crossing out at the square of the new found prime, we only need to do the crossing out for primes smaller than the square root of N: once we get there, any uncrossed number up to N will be prime. This is what the beginning of our list will look like once we are through:



So lets try and write our very own sieve of Erathostenes in python, with all this square and square root tricks pointed out above...

from time import time
from math import sqrt

def primeListSofE(n, verbose = True) :
 t = time()
 sieve = [True for j in xrange(2,n+1)]
 for j in xrange(2,int(sqrt(n))+1) :
     i = j-2
     if sieve[i]:
         for k in range(j*j,n+1,j) :
             sieve[k-2] = False

 ret = [j for j in xrange(2,n+1) if sieve[j-2]]

 if verbose :
     print "Calculated primes to",n,"in",time()-t,"sec."

 return ret

Play around with it and once more feel the thrill of speed. On my computer it does all primes below 1,000,000 almost 5 times faster than the best trial division approach from the previous post. But the real beauty is in how the time now scales with n, because you can no longer find a suitable fit of the form n^q. We haven't simply reduced the size of the exponent, but are treading uncharted territory, as it turns out that the time required to calculate all primes below using this sieve requires (n.log(n))(log(log(n))) time. Actually, for the first 10¹⁰ numbers, this is basically indistinguishable from linear time, i.e proportional to n.

This speed does come at a memory cost, though: you need to store an array of size n. And while the algorithm could produce all primes below 10¹⁰ in just a few minutes, the program will very likely crash giving an out of memory error. Some memory saving can be done by not storing even numbers at all. Without boring you with the details, below is the python function I use to compute primes at work, which does 10⁷ almost three times faster than the previous one, and uses half the memory, but still crashes when asked for all primes below 10⁹...

from time import time
from math import sqrt

def primeListSofE(n, verbose = False) :
  """Returns a list of prime numbers smaller than or equal to n,
  using an optimized sieve of Erathostenes algorithm.
  """
  t = time()
  maxI = (n-3) // 2
  maxP = int(sqrt(n))
  sieve = [True for j in xrange(maxI+1)]
  for p in xrange(3,maxP+1,2) :
      i = (p-3) // 2
      if sieve[i] :
          i2 = (p*p-3) // 2
          for k in xrange(i2,maxI+1,p) :
              sieve[k] = False
  ret = [2]+[2*j+3 for j in xrange(len(sieve)) if sieve[j]]
  if verbose :
      print "Calculated primes to",n,"in",time()-t,"sec."
  return ret

There are slightly faster ways of sieving, and we could memoize precalculated primes, to speed up repeated calls to the function. But this is probably enough for today's post...