dual2nd grad -nan issue

[Manushanker]

HI Allan,

I am computing a Hessian using the dual2nd forward type. The computed Jacobian and Hessian are all nan. When I debugged the code I found that for certain dual2nd variable it has grad (under val) to be -nan(ind) and val and grad under grad is also -nan(ind). Can you please let me know what is the reason for nan? If need I can share you my code.

Thanks,

Manushanker

[Manushanker]

I used the function 'isnan' to identify the dual2nd variables with val.grad, grad.val, grad.grad equal to nan and forced their value to be zero, then the library computes the correct Jacobian and Hessian values. I still would like to know why the val.grad becomes nan even though the value of the variable is not nan.

[allanleal]

Note f(0) is not always an issue. For example, f(x) = exp(x). f(0) = 1, fx(0) = 1.

[Manushanker]

My function is not a single function, but a set of functions. For a certain part of the function the val.grad value becomes nan, which makes all the dependent variables also to be val.grad nan. The code is bit involved than simple sqrt(x) or exp(x), I will used your suggestion to narrow down the issue. Thanks a lot!

[allanleal]

Yes, that's clear. Somewhere inside your more complicated function, with multiple variables, you have a mathematical operation at a certain value for which derivative does not exist. Check this to get a runtime exception so that you can debug and pinpoint where the error is happening: https://stackoverflow.com/a/3616017/418875

[Manushanker]

Thanks for the stackoverflow link. I will keep you updated!

[Manushanker]

The reason for nan was the value of x in sqrt(x) was zero, which caused the derivative to be nan. Thanks for your time and help!