Re: Test failure for fftfilt.m

[Jordi Gutiérrez Hermoso]

On 3 April 2013 00:51, Ed Meyer <address@hidden> wrote:
>
>
>
> On Mon, Mar 18, 2013 at 8:35 AM, Jordi Gutiérrez Hermoso
> <address@hidden> wrote:
>>
>> On 18 March 2013 01:05, Ed Meyer <address@hidden> wrote:

>> > I could put a patch up but for some reason other patches I've put
>> > up don't get applied so there is something I don't understand
>> > about the process.
>>
>> Sorry, something might have just gotten lost or forgotten in the
>> way. Which patches of yours have gone unapplied?
>>
> Besides the patch for 37297 there is one for 34461 and 34634. The
> first two mainly deal with test tolerances. 34634 adds the ability
> to read a mat file with undocumented data types for names which is
> readable by ML.

Ok, I'll look into those now.

> The test tolerance issue has come up several times in the short time
> I've been on the mail list; I believe that instead of simply
> increasing the tolerances, tests should be written to reflect the
> theoretical error bounds, which always involve some measure of the
> data or solution. As I've said before, don't take my word for it,
> see e.g. Golub & VanLoan or most any numerical methods text like
> Forsythe & Moler :-)

Oh, I completely agree, but it's very difficult to know what the
actual theoretical tolerance should be. It would require some careful
analysis of each problem plus knowledge of the algorithms that are
being used to compute each test. Frequently these algorithms differ
since they're implemented in libraries external to Octave; indeed, it
is often when these external libraries change that we see differing
tolerances.

- Jordi G. H.