[SciPy-Dev] [SciPy-User] ANN: scipy 0.8.0 release candidate 3

Pauli Virtanen pav@iki...
Mon Jul 19 12:18:22 CDT 2010


Sun, 18 Jul 2010 22:22:43 +0800, Ralf Gommers wrote:
[clip]
> >    assert dc[k] < 1e-9, (v[k], x[k], iv(v[k], x[k]), iv(v[k], 
> >                                                         x[k]+0j))
> > AssertionError: (-1.685450377317264, 1.2676128629915242,
> > 4.748141012443594e-05, (4.7481418802688318e-05+0j))
> >
> > relative error <~ 2e-7
>
> That does seem a little high. I guess Pauli has to decide whether this
> qualifies as a bug or not.

Works for me. No idea how to fix this, since it probably is compiler 
specific.

On the other hand, that's pretty close to a zero of Iv, where there is 
cancellation error (since v < 0 we use a reflection identity), so it's 
probably a difficult spot anyway.

Maybe the treshold could be made platform-dependent -- I wouldn't like to 
bump it if it fails only on a single platform/compiler combination.

	Pauli



More information about the SciPy-Dev mailing list