On Aug 12, 2:45 pm, William Stein <[email protected]> wrote:
> On Thu, Aug 12, 2010 at 12:30 PM, John H Palmieri
>
>
>
>
>
> <[email protected]> wrote:
> > How can I tell if numerical noise is actually noise, or if it is
> > indicative of a bug?  For example, with one or two OS/processor
> > combinations, I get this (from chmm.pyx):
>
> >    sage: m.viterbi([0,1,10,10,1])
> > Expected:
> >    ([0, 0, 1, 1, 0], -9.0604285688230899)
> > Got:
> >    ([0, 0, 1, 1, 0], -9.0604285688230917)
>
> > Can I tell how accurate this is actually supposed to be?  I can
> > certainly just change the doctest to
>
> >    ([0, 0, 1, 1, 0], -9.0604285688230...)
>
> > but if the code is actually supposed to be accurate to a few more
> > decimal places, this is concealing a bug.  Sometimes I can look at the
> > code and easily figure out its supposed accuracy, but more frequently
> > I can't.  So what should be done in cases like this?
>
> I'm the upstream author of 100% of this code, and know precisely what
> it does, which is a sequence of floating point ops and calls to the
> math library, which *of course* can be machine dependent.
>
> Put in dots.

Great, that's very helpful.  (My guess was that it was just a sequence
of floating point ops, as you say, but I wanted to check, to make sure
we were preserving the integrity of Sage's doctests.)

--
John

-- 
To post to this group, send an email to [email protected]
To unsubscribe from this group, send an email to 
[email protected]
For more options, visit this group at http://groups.google.com/group/sage-devel
URL: http://www.sagemath.org

Reply via email to