zhenglei-gao
commented
10 years ago -
Some constraints are imposed because the function is not defined outside the region. This would mean that the gradient and hessian are not defined in the usual sense at the boundary; the "left" and "right" derivatives should exist and be the same in the limit.
-
If your solution is actually on the boundary, you might consider some substitution which removes the constraint, and consider the hessian on the reduced, unconstrained, space.
-
the hessian coming from the optimization is built up as an approximation using information from the steps of the optimization.
Shall we use a bordered hessian??