Re: Differing results on right division

[Damian Philipp]

Hello Octave-Gurus,

I still haven't solved this, but some digging has revealed new things.
Find a link to the updated demo application at the end [1].

For one, it occured to me that since I was building Octave from source
anyway, I could insert some tracing code to see what was going on. This
then showed me, that I was using the wrong solver (LSSolve where Octave
uses FSolve). So I updated both my program and the demo application [1]
to use FSolve for full matrices. Sadly, the problem persists.

While updating the demo application [1] I noticed something else: I
print everything using 18 decimal places (which is more than machine
accuracy allows for, but it helps spotting differing entries in the
matrices). When saving a matrix in octave to a file (save matrix.oct
matrix), the output is truncated to mostly 15 or 16 decimal places
(machine accuracy?). When loading the file (load matrix.oct) and
comparing the loaded matrix to the computed matrix, many entries are
different. In my case, no entry differs by more than 4*eps(entry).

These tests were run using ATLAS 3.8.4, Lapack 3.3.1, IT++ 4.2, all
compiled on Debian Lenny, gcc 4.3.2, Intel Xeon.

So the question remains - why do I get different results from a right
division when I use the C++ code copied from octave compared to
performing the division in the octave interpreter?

Regards,
Damian Philipp

[1] Find the updated demo application at
https://rapidshare.com/files/1298191737/accuracy-demo-2.tar.gz

-- 
 University of Stuttgart, IPVS    phone: +49 711 685-88339
 Distributed Systems Group        www.ipvs.uni-stuttgart.de
 Universitaetsstrasse 38          address@hidden
 D-70569 Stuttgart, Germany