[SciPy-Dev] Doc error in scipy.optimize.fmin and missing info in ref guide

[David Goldsmith]

On Fri, Mar 19, 2010 at 1:09 PM,  <josef.pktd at gmail.com> wrote:
> On Fri, Mar 19, 2010 at 4:05 PM, David Goldsmith
> <d.l.goldsmith at gmail.com> wrote:
>> On Fri, Mar 19, 2010 at 12:37 PM,  <josef.pktd at gmail.com> wrote:
>>> On Fri, Mar 19, 2010 at 3:16 PM, David Goldsmith
>>> <d.l.goldsmith at gmail.com> wrote:
>>>> On Mon, Mar 15, 2010 at 4:05 PM, Rob Clewley <rob.clewley at gmail.com> wrote:
>>>>> Hi,
>>>>>
>>>>> In both my old version and the one in the svn trunk there is an
>>>>> inconsistency in the docstring of scipy.optimize.fmin.
>>>>>
>>>>>    Other Parameters
>>>>>    ----------------
>>>>>    xtol : float
>>>>>        Relative error in xopt acceptable for convergence.
>>>>>    ftol : number
>>>>>        Relative error in func(xopt) acceptable for convergence.
>>>>>
>>>>> As you see from the following code snippet from the body of the
>>>>> function, the convergence test is in terms of absolute error for both
>>>>> x and f, as the sim array contains x values and fsim the function
>>>>> values.
>>>>>
>>>>>    if (max(numpy.ravel(abs(sim[1:]-sim[0]))) <= xtol \
>>>>>                and max(abs(fsim[0]-fsim[1:])) <= ftol):
>>>>>        break
>>>>
>>>> OK, folks, so what's the _desired_ behavior: absolute or relative
>>>> error (i.e., where's the bug: in the code or in the doc)?
>>>
>>> I think the docs should reflect the actual behavior of the function.
>>> Whether the criteria will be changed and made consistent across
>>> functions is a different question. But it's not a bug, so the docs
>>> should adjust.
>>>
>>> my opinion
>>>
>>> Josef
>>
>> Well, I'm pretty sure we formally adopted the opposite policy: the
>> docs should reflect _desired_ behavior - if the code doesn't implement
>> desired behavior then a bug report is to be filed.  (I thought this
>> was in the Wiki Q&A section, but apparently not, so, UIAM, it's "only"
>> in an email thread somewhere).  Can anyone else confirm and,
>> hopefully, supply a link to the email thread in which this was
>> discussed?
>
> I know of that policy, but there are a lot of mistakes in the docs,
> and I don't think we change the code to match the mistakes in the
> docs.

No, of course not, but the idea is that a "mistake" in the doc is not
a mistake if in fact it documents the desired behavior (and thus the
real mistake is in the code, in which case we've basically agreed that
we _will_ change the code to match the doc; for better or worse, this
means that when "mistakes" are found in the docs - mistakes which in
fact may not be mistakes but only appear to be so because they don't
match code behavior - we will need to address them on a case-by-case
basis, such as is being done in this instance).
>
> But you can also file a ticket for changing and reviewing convergence criteria

OK, yours is a "vote" for "bug in the doc."

> I don't think relative (to what?) error makes much sense if x or f goes to zero.
>
> Josef

Agreed: the most robust behavior it to allow specification of both
types of tolerance.

DG