Question

Yesterday i had this interview question, which I couldn’t fully answer:

Given a function f() = 0 or 1 with a perfect 1:1 distribution, create a function f(n) = 0, 1, 2, ..., n-1 each with probability 1/n

I could come up with a solution for if n is a natural power of 2, ie use f() to generate the bits of a binary number of k=ln_2 n. But this obviously wouldn’t work for, say, n=5 as this would generate f(5) = 5,6,7 which we do not want.

Does anyone know a solution?

Answer 1

You can build a rng for the smallest power of two greater than n as you described. Then whenever this algorithm generates a number larger than n-1, throw that number away and try again. This is called the method of rejection.

Addition

The algorithm is

Let m = 2^k >= n where k is is as small as possible.
do
   Let r = random number in 0 .. m-1 generated by k coin flips
while r >= n
return r

The probability that this loop stops with at most i iterations is bounded by 1 - (1/2)^i. This goes to 1 very rapidly: The loop is still running after 30 iterations with probability less than one-billionth.

You can decrease the expected number of iterations with a slightly modified algorithm:

Choose p >= 1
Let m = 2^k >= p n where k is is as small as possible.
do
   Let r = random number in 0 .. m-1 generated by k coin flips
while r >= p n
return floor(r / p)

For example if we are trying to generate 0 .. 4 (n = 5) with the simpler algorithm, we would reject 5, 6 and 7, which is 3/8 of the results. With p = 3 (for example), pn = 15, we’d have m = 16 and would reject only 15, or 1/16 of the results. The price is needing four coin flips rather than 3 and a division op. You can continue to increase p and add coin flips to decrease rejections as far as you wish.