Sounds like it would be nice to have both implementations in the same class
with a "usePivoting(boolean)" method to select which is used. I would
recommend pivoting as the default since it gives fewer surprises and more
diagnostic information to the naive user.
Except for the extra indexing costs, a pivoting implementation should be
about twice as costly as a non-pivoting implementation. If you don't do the
pivoting by indirection, this could be significantly more due to copying.
I haven't had instances where I would have much cared, but it would be nice
to have the option for extra speed if I needed it.
On Wed, Oct 5, 2011 at 10:52 AM, Luc Maisonobe <luc.maison...@free.fr>wrote:
> 4. Implement the best algorithm as "QRDecomposition" in ("linear"), and
>> remove duplicate code ("LevenbergMarquardt" will call the
>> implementation
>> in "linear"; if replacing the LM currently internal version makes some
>> test fail, we should worry and look for the bug either in LM or in the
>> new QR or in the tests).
>>
>
> I'm not sure one implementation only is always feasible. The one I put in
> Levenberg-Marquardt is perhaps more suited to some cases because of
> pivoting, but I'm quite sure it is not as efficient as the classical one
> which is really fast (I always used it as an example in conferences to show
> that Java can be as fast as fortran in some application, because it is as
> fast as lapack with ATLAS blas on matrices up to about 1000 rows and
> columns).
