Another flaw with generate_canonical<float> is that if you generate, say, 2^40 of the things, you should expect to get quite a few values smaller than 2^-32; bu

Another flaw with generate_canonical<float> is that if you generate, say, 2^40 of the things, you should expect to get quite a few values smaller than 2^-32; but with minstd_rand, or mt19937 I believe you will not ever see values smaller than 2^-32 or so. A similar problem exists with double, though it takes a much longer-running problem to prove it (ain't nobody got time to generate 2^61 random numbers)