Nelder-Mead forms a new simplex when you start it with b_N as the initial point. That simplex is larger than the simplex it ended up with in the first round. The convergence criterion of Nelder-Mead is the standard deviation of the function evaluated at each point of the simplex. Since the initial simplex of the second round is larger, you end up with a larger value of the standard deviation. Btw, you can change the initial simplex, see Optim.jl, and you can also save the last simplex of the first round.
3 Likes