Through analysis, we have decided that we want to take values which are close enough to 1.0 to map back to 0.0. We came to this conclusion from noting that:

[0.0, 1.0)

when scaled to the integer set:

[0, 4)

there is some value close enough to 1.0 that when multiplying by 4, gets truncated to 4. Therefore:

[0,1-eps] => 0 [1,2-eps] => 1 [2,3-eps] => 2 [3,4-eps] => 3

So 0 gets uneven distribution if we simply clamp. The actual clamp utilized in this file is to map values out of range back to 0 to restore uniform distribution.

Also, for clamping floats when asking for a distribution from [0.0,g) we have determined that for values of g < 0.5, then rand*g=g, which is not the desired result. As above, we clamp to 0 to restore uniform distribution.

Generate a float from [0, RAND_MAX).

Use the boost random number generators if you want a stateful random numbers. If you want more random numbers, use the c-functions since they will generate faster/better randomness across the process.

I tested some of the boost random engines, and picked a good double generator and a good integer generator. I also took some timings for them on linux using gcc 3.3.5. The harness also did some other fairly trivial operations to try to limit compiler optimizations, so these numbers are only good for relative comparisons.

usec/inter algorithm 0.21 boost::minstd_rand0 0.039 boost:lagged_fibonacci19937 0.036 boost:lagged_fibonacci607 0.44 boost::hellekalek1995 0.44 boost::ecuyer1988 0.042 boost::rand48 0.043 boost::mt11213b 0.028 stdlib random() 0.05 stdlib lrand48() 0.034 stdlib rand() 0.020 the old & lame LLRand