The information on the Hessian-function explicitly states that it is not for frequent use on an expensive to evaluate objective function. I tried it anyway, because that is what I need. Most of the time, I get something reasonable (hard to check, but I'll worry about that later), but sometimes I get NaN's. Any idea what the problem might be?

Perhaps some more details are warranted: The function I want to differentiate is a likelihood-function. It is a function of time, and I need the Hessian at the maximum for each timestep. (The maximum is one point associated to the last time period.) The function takes 5 parameters.