Re: [Numpy-discussion] Random integers

[Tim H. <tim...@ie...>]

Albert Strasheim wrote:
> Hello all
>
> I'm trying to generate random 32-bit integers. None of the following seem to
> do the trick with NumPy 1.0.dev3383:
>
> In [32]: N.random.randint(-2**31, 2**31-1)
> ValueError: low >= high
>
> In [43]: N.random.random_integers(-2**31, 2**31-1)
> OverflowError: long int too large to convert to int
>
> In [45]: N.random.randint(-2**31, 2**31-1)
> ValueError: low >= high
>
> Am I missing something obvious?
>   
I don't think so. This doesn't help you any, but the problem is in 
mtrand.pyx:

        diff = hi - lo - 1
        if diff < 0:
            raise ValueError("low >= high")

The variables diff, hi and lo are all signed c-longs, which means the 
interval can only ever 2**31-1 or you overflow (this is the problem that 
you are seeing). It appears that the underlying rk_interval works on 
unsigned longs, so this is probably fixable with a little care. At the 
moment I don't have time to dig into the ins and outs of this though.

The resulting error distribution is probably imperfect, but you could 
instead use some variation on int(np.random.random() * 2**32 - 2**31).

-tim