Hi.
On Sun, 13 Oct 2013 00:41:00 -0700, Ted Dunning wrote:
Commons math decided not to support sparse matrices due to the
difficulty
of getting them just right.
The problem is that 0 * Inf = NaN, not 0. This means that for
floating
point math, 0 * x = x is not a true identity.
The problem really arises when the 0 in question could be either
stored as
a 0 or not. Not storing zeros is, of course, the hallmark and
calling card
of sparse matrices because if you assume that the zero multiplication
identity holds, you can avoid lots of arithmetic.
Unfortunately, if the other matrix or vector has Inf in it,
multiplying by
a sparse matrix where the corresponding element is missing gives a 0
instead of NaN. Moreover, the dot product gives a value instead of
NaN.
This is very hard to fix and still retain the desirable speed
properties of
sparse matrices. You could imagine that you have a bitmap of all of
the
Inf values in every matrix, but keeping that bitmap up to date can
absolutely kill performance since you have the potential for a branch
in
critical inner loops.
Most sparse matrices just punt and say they are sorry in a very
non-IEEE-754 sort of way.
How do they do that, practically?
Commons Math has decided instead to remove sparse matrices entirely.
IIRC, this decision was taken by default, because nobody was identified
as a user of CM's sparse matrices implementation. We thus concluded
that
it was harmless to drop this part of the code because it needed work
(cf.
issues clearly explained above) and we are lacking human resources to
perform it (i.e. we need developers interested to fix and maintain this
part of the code). If such a person would come forward, the decision
could probably be reverted.
Regards,
Gilles
On Thu, Oct 10, 2013 at 6:34 AM, Daniel Marbach <[email protected]>
wrote:
Hi,
Does Apache Commons support sparse matrices? I understand that the
interface SparseRealMatrix is deprecated...
Will there be a replacement or has the project decided to stop
supporting
sparse matrices altogether?
Thanks,
Daniel
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
