Re: Sparse matrix problem

[Michael Goffioul]

> With g++ 4.1.2 on my system, it prints (nan,nan).  That seems like a
> bug to me, but maybe it is standard conforming (and if it is, wtf?!?)?
>
> If it is a bug, then the bug should be fixed in the compiler, not
> Octave.
>   
[snip]
> so maybe we need a better than grammar school implementation here.
>   

The problem also exists in MSVC (altough (1+i)/0 gives good result, 
(1+i)/(0+0i) does not).
In the end, the fundamental problem is in the complex division 
implementation:

   x/y is converted into x*y'/|y|^2
   if y = 0+0i
   x*y' => 0+0i (whatever x)
   |y|^2 => 0
   hence x/y => NaN+NaNi (whatever x)

However, the complex division is only valid if y is non-zero. In the 
MSVC complex header,
one can see that the result of x/y when y = 0+0i is explicitely set to 
NaN+NaNi, it's even
not the result of a computation like 0/0 (like in the gcc headers).

Michael.