minimization using the Brent algorithm (brent.c)

Brian Gough bjg@network-theory.co.uk
Sat Mar 22 18:12:00 GMT 2003


Z F writes:
 > Sorry for my rudness, but I think that it is not prudent to require 
 > a "test" point inside the interval. (I know that standard libraries
 > require that). The test F(a)>F(x)<F(b) is a protection againts a
 > "stupid" mistake, nothing more. It does not guarantee that there is no
 > point y between a and b that F(y)>F(a) or F(y)>F(b) so the test is
 > useless and only requires an extra computation of F() which might be
 > time consuming.

GSL requires the three points since its goal is to give a guaranteed
bound on a local minimum.  With three points the algorithm cannot
fail, whereas a two-point algorithm can.

It's true that the user needs their own routine to find the initial
guesses.  If someone wants to write one for GSL I will add it.

-- 
Brian Gough

----------------------------------------------------------------------
Network Theory Ltd           Phone: +44 117 3179309 (UK: 0117 3179309)
15 Royal Park                  Fax: +44 117 9048108 (UK: 0117 9048108)
Bristol BS8 3AL                WWW: http://www.network-theory.co.uk/
United Kingdom               Email: bjg@network-theory.co.uk     
----------------------------------------------------------------------



More information about the Gsl-discuss mailing list