Re: multidimensional optimization

[Brian Gough <bjg at network-theory dot co dot uk>]

Fabrice Rossi writes:

 > use a one dimensional minimization, we can assume that the value of
 > the function at the estimated minimum has been already calculated
 > (in the current sources I think there is no way get the value of
 > the function at the estimated minimum: it has to be
 > recalculated. I'm I right?).

Yes, it has to be calculated separately. The gsl_min_fsolver struct
does not keep track of it.

 >  Right, but there is overlapping with the convergence test. For
 > instance, you can compute the norm of the last move, or something
 > similar, which can be used also for the convergence test.

True. We'll assume that computing the norm is fast compared with
computing the gradient.  The extra flexibility of allowing different
tests will make up for the cost of recomputing it in the tests that
need it.