[SciPy-dev] bug in scipy.stats.norm.cdf ?

[John Hunter]

On 3/4/07, eric jones <eric at enthought.com> wrote:
> It looks like it works with latest install of Python -- Enthought
> Edition on WinXP.  This is python 2.4.3.
>
> In [81]: from scipy.special import _cephes as c
>


Yep, it's an isnan bug.  If I add a dummy function "erfjdh" to ndtr.c
and _cephesmodule.c

double erfjdh(double x)
{
  double y, z;
  if (isnan(x)) {
    mtherr("erf", DOMAIN);
    return -2.0;
  }
  return 2.0;
}


and then run my test code

In [1]: import scipy

In [2]: import scipy.special._cephes as c

In [3]: x = scipy.arange(-3, -2, 0.1)

In [4]: c.erfjdh(x)
Out[4]: array([ 2.,  2.,  2.,  2.,  2., -2.,  2.,  2.,  2.,  2.])


But apparently the internal isnan testing in ndtr.c is different than
the toplevel testing because

In [5]: scipy.isnan(x)
Out[5]: array([False, False, False, False, False, False, False, False,
False, False], dtype=bool)

Still digging...