le Monday 23 August 2004 15:35, krempp@crans.ens-cachan.fr écrivit :
Plus, using a rejection scheme would completely avoid any added bias (contrary to uniform_smallint in which the added bias can reach 1/32, thus unsuitable for high-precision statistics programs)
the bias of uniform_int is definitely a problem. With current uniform_int using small_int, the observed frequencies on random numbers in {0..2} based on mt1993b are about as much biased as I expected (~0.5%) : frequence [0] : 0.335951 frequence [1] : 0.331972 frequence [2] : 0.332077 These are frequencies over 100e6 iterations, and are signficantly consistent (at least 3 digits have already converged). With proposed change, the same test yields : frequence [0] : 0.333284 frequence [1] : 0.333346 frequence [2] : 0.33337 There is now almost 100 times less bias. (and it will converge to 0 as we increase the number of iterations) I attach the patch. I can commit it and the other one I sent about the overflow bug if there's no objection on those patches. -- Samuel