END
Unfortunately, to date, there is no concrete convergence property that has been proved of
the original Nelder-Mead algorithm. The algorithm might even converge to a non-stationary
point of the objective function (see Mickinnon[7] for an example). However, in general, it
has been tested to provide rapid reduction in function values and successful implementations
of the algorithm usually terminate with bounded level sets that contain possible minimum
points. Recently, there are several attempts to modify the Nelder-Mead algorithm to come
up with convergent variants. Among these: the fortified-descent simplical search method (
Tseng [12]) and a multidimensional search algorithm (Torczon [11]) are two of the most
successful ones. See Kelley [5] for a Matlab implementation of the multidimensional search
algorithm of Torczon.