Re: multidimensional root-finding problem :-(

[Fabrice Rossi <rossi at ufrmd dot dauphine dot fr>]

Peter S. Christopher wrote:

> Hi there,
>
> 	I've had experience with the Multi-Min package. My best guess is
> that some of the routines use ONLY fdf and some of them use f, df and fdf.

That's right. As said before by Ivo, the purpose of fdf is to save some 
calculation. For instance, when you deal with neural networks, you can 
save a factor 2 if you compute f and df together rather than f and then 
(in another call) df. BUT the basic contract between a function provider 
and a minimizer is that the three functions work. That's why I suggest 
to start programming f and df, and provide a dummy version of fdf that 
calls both f and df. Then, if you need faster calculation, you can 
provide an efficient version of fdf. If I recall correctly (I wrote the 
first version of multimin), f is called in the line search (i.e., the 
one dimensional part of multimin), whereas fdf is used for the high 
level part. I think I did not use df alone, but the code as evolved.

This is only valid for multimin, but my understanding of multi-root is 
that it basically works in the same way. So Slaven I suggest you 
separate f and df, and call those functions for fdf...

Fabrice