Re: Octave/C++ matrix Inv() comparison

[CdeMills]

ionone wrote
> thank you for your answer
> 
> i'll tend to believe you and i understand what you meant (and thank you
> for that) that different inverse() functions may give different results
> and it's logical, but what i don't understand is that Octave's inverse
> function makes the DSP works and if Octave uses Lapack, using the same
> funtions on the same numbers should give the same results, no ? Even if
> the numbers aren't equal due to approximation and numbers very close, etc
> it should give the exact same results with the same functions. Maybe i
> didn't tried yet the good functions ? 
> 
> Jeff

It should. But you're on the verge of numerical accuracy. What I suggest is
to start by generating small matrices, compute their inverse, and then
max(max(A * (A^-1) - eye(size(A))) as performance criterion. Do a graph,
abcissa = n = (1:20), ordinate = performance criterion. See where a break
occur into the curve; look at associated condition number.

About the DSP versus computer computations: possible influence factors are
f.i. the way rounding is performed when doing a series of arithmetic ops
like sum (a (i) * b(i)) : does rounding occur after computation of a(i)*b(i)
or after the whole sum is computed ?

Once again, on well-behaved matrix, this won't make a difference; but your
matrices are nearly singular.

Regards

Pascal



--
View this message in context: 
http://octave.1599824.n4.nabble.com/Octave-C-matrix-Inv-comparison-tp4655291p4655537.html
Sent from the Octave - Maintainers mailing list archive at Nabble.com.