Re: [OctDev] Re: sparse tests

[Paul K. <pki...@us...>]

On Dec 12, 2003, at 10:56 PM, Andy Adler wrote:

> On Thu, 11 Dec 2003, Paul Kienzle wrote:
>
>> The sparse tests at line 240 is failing since sparse(eye(3)).^x does 
>> not
>> fill in the zeros with 1 for x==0 or inf for x<0.  This is 
>> bug-compatible
>> with matlab.  Shall we correct it?
>
> This is bug compatible with Matlab. I remember asking on this list
> about this (mis) feature when I was developing it.
>
> Advantages of the matlab way:
>  - faster (ignore everything corresponding to a zero)
>  - compatible
>
> Advantages of the correct way:
>  - correct ;-)
>
> I agree with John that introducing "feature" variables is a bad
> idea. I say we leave it and put a _big_bad_warning_ somewhere.

It would be a little more convincing if other operators didn't
convert to full as necessary, like + and /.  I haven't checked
functions like cos/sin though.

BTW, this came up in the context of minimizing the warnings
produced by the sparse tests.  I've got it down to almost
nothing now.

- Paul