Re: [OctDev] control 2.3.53 released

[Lukas R. <luk...@gm...>]

On 03.09.2012, at 12:29, Philip Nienhuis <pr....@hc...> wrote:

> PMFJI
> 
> but, uhm....
> 
> Lukas Reichlin wrote:
>> On 03.09.2012, at 10:39, Rafael Laboissiere<ra...@la...>  wrote:
>> 
>>> * Rafael Laboissiere<ra...@la...>  [2012-08-30 23:47]:
>>> 
>>>> * Carnë Draug<car...@gm...>  [2012-08-27 16:43]:
>>>> 
>>>>> a new release of control package is out, version 2.3.53, by Lukas Reichlin.
>>>> 
>>>> In exercising the tests in inst/@lti/minreal.m, I got the error below.
>>>> Is it normal?
>> Hi Rafael,
>> 
>> You can check whether the observed and expected results are equivalent state-space models (i.e. state-transformation, see command prescale for formulae). This can be done, e.g., by inspection of the Hankel singular values (command hsvd), time response (step, impulse) or frequency response (sigma).
>> If they are the same, there should be nothing to worry about. If you want the same results, use Reference BLAS (and LAPACK) from www.netlib.org instead of ATLAS which you are probably using. The SLICOT authors recommend the use of the reference implementations. Correct results are more important than minor speed advantages of automatically tuned linear algebra software, aren't they? :-)
> 
> ... shouldn't such an interpretation be reflected in the tests then, rather than a plain "assert" comparison that depends on what (allowed!) dependencies happen to be in place.
> Currently the test just seems to yield confusion.
> 
> Philip

Hi Philip

I see your point, but currently, the tests are such that if they pass, one can trust the results. If they fail, it is up to the user to check whether the results in this special case are correct and if he has confidence in his SLICOT dependencies (e.g. ATLAS) in general. I don't know how to do this automatically and how to avoid false positive tests (which would be a very bad thing). Do you have any ideas how to check the results of the failing tests?

Lukas