Re: multimin question

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

Fabrice Rossi writes:
 > 1) it is not able to find a x>e (e is the machine precision) such
 > that f(x)<f(0) 2) it has x such that x>e, f(x)<f(0) but it is not
 > able to find a y>x such that f(y)>f(x).
 >  The second case is quite strange. The first case is very common
 > and simply shows that the descent direction does not allow
 > decrease. I considered this to be a bug, so I programmed bracketing
 > to return GSL_FAILURE. But it should return something else. Then in
 > the multimin algorithm, we will react on the result. If the
 > bracketing fails for reason 1 and we have used the gradient as
 > descent direction, we might decide that the algorithm as converged.
 >  Comments?

If you need to introduce an extra error code to distinguish various
cases, and it can be done in a natural way, then feel free to do that.
It is part of the library philosophy to give every distinct type of
error its own error code.