[Ipopt] derivative-check errors

Andreas Waechter andreasw at watson.ibm.com
Thu Sep 10 20:55:52 EDT 2009

Hi Sourav,

> Many thanks for those clarifications. I have a small question.
> Is quasi-Newton the default option. i.e., if I do not provide a hessian,
> IPOPT automatically calculates the hessian by quasi-Newton approximation?

No, you have to set the option explicitly if you do not provide the 


> regards,
> Sourav
>> Hi Sourav,
>> The third column in the derivative checker output is the RELATIVE error,
>> where the denominator is the larger of the two values - this is why it is
>> one.
>> If the values in the second column are insanely large, it probably means
>> that there is something wrong with the evaluation of the first
>> derivatives, so that the finite difference approximation for second
>> derivatives is giving incorrect numbers.
>> Since you have memory errors, I suggest you verify your code with a memory
>> checker first, such as valgrind on Linux, before trying to fix anything
>> else.
>> I'm not sure if I understood what you wrote about not giving structure or
>> values in eval_h.  You need to provide both if you are using the 2nd order
>> derivative checker or want Ipopt to use your second derivatives during the
>> optimization (i.e., if you are not using the quasi-Newton approximation).
>> Regards
>> Andreas

More information about the Ipopt mailing list