statistical bias in get_random_long() with large ranges

[kaganar@…]

Great care is taken within get_random_long() (utils.c) to avoid any statistic biases introduced by limiting the range of genrand_31() via the reminder operator (%), however there seems to be a flaw as exemplified by:

th fdiv(lmath(add,iter(lnum(700),rand(1000000000))),700)

The gist of the code for get_random_long() that handles the actual random number generation seems to be:

x=high-low+1;
n_limit=ULONG_MAX-(ULONG_MAX % x);
do {
  n = genrand_int31();
} while (n >= n_limit);
return low + (n % x);

Presumably ULONG_MAX is 2³²-1, but genrand_int31() only results in numbers from 0 to 2³¹-1, and since an earlier check guarantees n<=2³¹-1, n_limit is always greater than 2³¹-1 and the while-loop condition always fails. Presumably genrand_int32() should be used instead.

--Ian

[Luke]

Calls genrand_int32 instead of genrand_int31.

[raevnos]

Um. Yes. I'm not sure why I made it use 31 bit ints in the first place, considering I pretty much lifted the code from MUX, which uses the MT's 32-bit generator.