How to use fminunc providing gradient and hessian?

Hi, could you help me in making this code (attached) working? Thanks!

Answer 1

[Matt J]

You need to return the gradient and Hessian

function [val,g,H]=log_lik(theta,ix)

Also, these tolerances look unrealistically small

'TolX',10^-30,'TolFun',10^-30

Answer 2

[Paul]

What Matt said and the way you want to calculate the gradient and hessian is not gonna work. since val is just a single value, the function jacobian is not gonna work.

[MRC]

In fact it does not work. What do you mean by single value?

[Matt J]

What Matt said...

It definitely isn't going to work if the function doesn't return g and H. Whether g and H are correctly computed is another matter, but it doesn't invalidate my advice.

[Matt J]

What do you mean by single value?

The jacobian function is from the Symbolic Math Toolbox. But you are doing non-symbolic math here.

[MRC]

Are you suggesting that I cannot use the jacobian function inside fminunc? Or, how should I change my code in order to use it?

[Paul]

Maybe you misunderstood me Matt. I meant what you said is right and that the way he was calculating the jacobian and hessian wasn't gonna work. Cris, jacobian works on symbolic expressions. You could try using fminunc without the gradient and hessian. Or else derive the gradient and hessian expressions yourself by hand or using matlab. These should depend on the values in theta.

[Matt J]

@Paul: Ah! I get it now.

@Cris: If you are setting GradObj and Hessian to 'on' because of Walter's remark here, I think it's too soon. I think you should try with the default gradient and Hessian calculation first and then add your own once you're getting decent optimization results and you just need to make small improvements to speed and accuracy.