On Jun 10, 4:01 am, Mark Dickinson <dicki...@gmail.com> wrote: > On Jun 10, 7:25 am, John Yeung <gallium.arsen...@gmail.com> wrote: > > > > > > > On Jun 10, 1:52 am, Steven D'Aprano > > > <ste...@remove.this.cybersource.com.au> wrote: > > > On Tue, 09 Jun 2009 22:21:26 -0700, John Yeung wrote: > > > > Therefore, to me the most up-to-date docs (which say > > > > that uniform(a, b) returns a float in the closed > > > > interval [a, b]) is closer to correct than before, > > > > but still fails to point out the full subtlety of > > > > the behavior. > > > > Which is? > > > That uniform(a, b) will return a random float in the semi-open > > interval [a, b) for certain values of a and b; and in the closed > > interval [a, b] for other values of a and b. (Swap a and b if a > b.) > > > To me, the fact that you sometimes get a semi-open interval and > > sometimes a closed interval is worth noting in the docs. > > Do you want to submit a doc patch? > > For practical purposes, I think you'd be hard-pressed to find a > statistical > test that could reliably distinguish between a large sample of values > from > random.uniform(a, b) and a sample from a 'perfect' uniform > distribution on > the closed interval [a, b]. It's true that there are values of a and > b > such that random.uniform(a, b) can never produce b.
So, the 2.6.2 documentation is STILL wrong. Before it implied it was ALWAYS a semi-open interval, and now it says it's ALWAYS a closed interval. But neither is correct. I think a doc patch is definitely warranted. > But for given a and b, > not too close together, there may be many other values that can't be > produced as well, so it hardly seems worth pointing out one particular > value that can never be produced. > > Example: on a typical system there are almost 2**62 floats in the > range [0, 1); the vast majority of these can never be produced by > random.random(), which can only ever return one of 2**53 distinct > values > (it essentially just returns a value of the form n/2**53 with n > an integer in the range [0, 2**53)). So random.uniform(0, 1) will > miss lots of possible values. > > On the other hand, it's easy to find examples of a and b such that > random.uniform(a, b) has a significant chance of producing b. > For example, take a = 10**16, b = 10**16 + 4, then there's about a 1 > in 4 chance of getting b. Or for a more extreme example, simply > take a = b. Hence the doc change. > > Mark -- http://mail.python.org/mailman/listinfo/python-list
