Question

I’m having trouble generating random numbers that do not follow a discrete uniform distribution.

So for example, say I have 5 numbers (to keep it simple), a probability of number k being generated would be k/15. (k = 1 to 5)

My idea is to generate a random number j using rand(), and if this number j is:

1 => then number 1 is generated

2 or 3 => num 2

4 or 5 or 6 => num 3

7 or 8 or 9 or 10 => num 4

11 or 12 or 13 or 14 or 15 => num 5

Now scale this to generating 1-10, 1-100, 1-1000. Does this work the way I intend it to? I’ve constructed a loop that pretty much does this every time a number needs to be generated, I think it’s probably inefficient since it goes up until it finds the j number generated in one of the ranges… What could be a better way to do this?

EDIT: or maybe create an array once with the proper numbers and then pull out with rand() better solution?

Answer 1

Consider that the sum s of integers from 1 to n is s = n * (n + 1) / 2. Solve for n to get n = (± sqrt(8*s + 1) - 1) / 2. We can ignore the negative square root, as we know n is positive. Thus n = (sqrt(8*s + 1) - 1) / 2.

So, plugging in integers for s between 1 and 15:

s  n
01 1.000000
02 1.561553
03 2.000000
04 2.372281
05 2.701562
06 3.000000
07 3.274917
08 3.531129
09 3.772002
10 4.000000
11 4.216991
12 4.424429
13 4.623475
14 4.815073
15 5.000000

If we take the ceiling of each computed n (the smallest integer not less than n), we get this:

s  n
01 1
02 2
03 2
04 3
05 3
06 3
07 4
08 4
09 4
10 4
11 5
12 5
13 5
14 5
15 5

Thus you can go from the uniform distribution to your distribution in constant space and constant time (no iteration and no precomputed tables):

double my_distribution_from_uniform_distribution(double s) {
    return ceil((sqrt(8*s + 1) - 1) / 2);
}

N.B. This relies on sqrt giving an exact result for a perfect square (e.g. returning exactly 7 given exactly 49). This is normally a safe assumption, because IEEE 754 requires exact rounding of square roots.

IEEE 754 doubles can represent all integers from 1 through 2^53 (and many larger integers, but not contiguously after 2^53). So this function should work correctly for all s from 1 to floor((2^53 - 1) / 8) = 1125899906842623.