Check of the gradient of the negative log-likelihood

To verify the the supplied gradient of the negative log-likelihood calculated by using Gaussian filtering (differentiated Uuscented Kalman Filter) I used the DerivativeCheck from Matlab by:

Opt =optimset('DerivativeCheck','on','GradObj','on','Display','iter',,'Algorithm','sqp','FinDiffType','central');
problem = createOptimProblem('fmincon','objective',@neg_logLik,'x0',th0,'lb',LB,'ub',UB,'options',Opt);

I get the following message:

Objective function derivatives:
Maximum relative discrepancy between derivatives = 0.9
Caution: user-supplied and central finite-difference derivatives do
not match within 1e-006 relative tolerance.
To displaying full derivativecheck output, I used the following code:

To displaying full 'DerivativeCheck' output, I used the following code:

[x,fval,exitflag,output,lambda,grad]=fmincon(...)

From https://www.mathworks.com/matlabcentral/answers/46854-displaying-full-derivativecheck-output.

I did this with MaxIter = 0 and GradObj = 'off'. the output grad will be fmincon's idea of the gradient of −L(θ) calculated by finite difference at the initial point and I get the following results:

My supplied derivatives results:

-120.5610
  42.5878
 -28.1507
   9.7675
and finite difference results:

-12.0561
   4.2588
  -2.8151
   0.9767

My results and the results of the finite difference look the same but my results multiplied by 10. What's wrong?

Answer 1

[Matt J]

My results and the results of the finite difference look the same but my results multiplied by 10. What's wrong?

Isn't it obvious? Your gradient computation is off by a factor of 10.