[ADOL-C] Question about NaNs in jacobian

[Gil Serrancolí Masferrer]

Dear all,


We are using ADOLC, and we observed an issue we cannot solve so far.


You can see a piece of code attached to this message. In function F, we evaluate a certain function. It contains tanh, pow, sqrt, etc., all continuous functions at all points we evaluate the function. The trace is computed (in F_adolc) at the first time that this function is evaluated.


When evaluating the jacobian of this function at certain points (0,0), (0.1), the jacobian seems reasonable (I compared it to the one obtained using finite differences). However, when evaluating the jacobian at (1,0), the jacobian reports NaNs. Then, we need to compute again the trace (run F_adolc) to avoid NaNs.

Do you know what is the reason of this issue? It seems the trace should be valid for the entire domain, so why do we get NaNs when we do not create the trace again?


Thank you very much for your attention.


Best regards,


Gil Serrancolí
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/adol-c/attachments/20171117/f7fe9edd/attachment.html>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: TestNaNsJacobian_onlyADOLC.cpp
URL: <http://list.coin-or.org/pipermail/adol-c/attachments/20171117/f7fe9edd/attachment.ksh>