Compute gradient and hessian with the python package numdifftools when parameters are bounded

Asked Jun 13 '16 at 12:29

Active Aug 03 '16 at 21:47

Viewed 833 times

I have minimized a log likelihood function using scipy.minimize and I want to compute the standard-errors associated to the parameters.

In cases in which parameters are not bounded, using numdifftools (nd.Gradient and nd.Hessian) works perfectly. I have issues when the parameters are bounded because nd.Gradient tries to compute the function with values of parameters out of bounds where the function is not defined.

My guess is that I should use the options relative to step (step_max, step_nom, etc.) but I do not understand how they work. Can anyone help me understand how to use them?

Thanks

edited Aug 03 '16 at 21:47

Amir

10,600
9
48
75

asked Jun 13 '16 at 12:29

Ariane

Did you get any answer ? I am going to try something and keep you informed ***if you want***. – keepAlive Jun 21 '17 at 00:32
@Kanak I did not get an answer. I managed to do it by luck playing with step_max and method. – Ariane Jun 26 '17 at 13:46

Compute gradient and hessian with the python package numdifftools when parameters are bounded

0 Answers0