Re: [eigen] numerical differentiation

[ Thread Index | Date Index | More lists.tuxfamily.org/eigen Archives ]


2009/9/28 Rohit Garg <rpg.314@xxxxxxxxx>:
> Trevor does have some good points here.

Just to be clear: I never said he didn't ;)

I just tend to avoid as long as possible dealing with the specifics of
all proposals that appear on this list: otherwise I get sucked into
studying other people's stuff all the time ;)

Now I had a look at Thomas' code and that seems like pretty reasonable
stuff to me as a minimalistic/naive/lightweight implementation, and
then, Rohit's suggestion of allowing to use backends seems very good
to me! Without allowing backends, indeed I'm afraid that this isn't
going to be very useful: doing robust numeric differentiation is
probably much harder than that (and we probably don't want to enter
into that in Eigen).

Thomas --> one remark: calling resize() on fixed size matrices,
passing the right size, is a NOP, it's perfectly OK and we do that all
the time.

> A better idea for what is
> being attempted would be perhaps to provide backends, like it was done
> for FFT. Though in that case, we have a basic FFT code in eigen's core
> as well.
>
> Instead of throwing the kitchen sink into eigen, let's provide
> external interfaces to top notch libraries that exist for that
> purpose.

Sounds very reasonable to me!

But again, even so, it's useful to provide a minimalistic
implementation built-in, and Thomas' code is that.

> This way, we can expose the need of users and avoid
> maintaining the code as well. If a feature gets used often, even if it
> is unsupported, it will end up core developer's brain cycles.

Yes, it's true. Thanks for caring about these aspects, it's
appreciated ;) It will end in our brain cycles, if only when we wonder
whether to move that to Eigen/.

Benoit



Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/