How to use fmincon for constrained maximum likelihood?

Hi, I'm trying to solve a constrained minimization problem but I get several error messages. Could you help me in adjusting the code? Thanks!

clear all 
n=2; 
thetatrue=[1 1 2 2 0.2]; 

mu = [0 0]; %mean and variance covariance matrix
sd = [1 thetatrue(2*n+1); thetatrue(2*n+1) 1];

load data     
X=data(:,1:n); 
Y=data(:,n+1:size(data,2));  
B1(:,2)=-X(:,1); 
B1(:,1)=-1;   
B2(:,2)=-X(:,2);
B2(:,1)=-1;

C1=(all(bsxfun(@eq,Y,[0 0]),2)); 
C2=1-C1;   

cdf=@(x) mvncdf( [B1*[x(1);x(3)],  B2*[x(2);x(4)] ] ,mu,[1 x(5); x(5) 1]);             

options=optimset('Algorithm',...
'interior-point','Display','iter','MaxIter',10000,'TolX',10^-30,'TolFun',10^-30);

theta0=thetatrue; 
[theta,fval,exitflag,output]=...
        fmincon(@(x) log_lik(x,cdf,C1,C2),theta0,[],[],[],[],[-Inf; -Inf; -Inf; -Inf; -1], ...
                 [+Inf; +Inf; +Inf; +Inf; 1],[],options);

And

function val=log_lik(theta,cdf,C1,C2)
        g=cdf(theta);
        val=-sum(C1.*log(g)+C2.*log(1-g));

I have also another question: from the theory I know that the set of global minimizers could be non singleton; is there a way to understand it from the results of the optimization procedure?

Answer 1

[Alan Weiss]

I ran exactly the code you gave with your data. fmincon ran to completion:

                                              First-order      Norm of
   Iter F-count            f(x)  Feasibility   optimality         step
      0       6    1.503710e+03    0.000e+00    1.317e+03
  User objective function returned NaN; trying a new point...
      1      14    1.052526e+03    0.000e+00    1.146e+03    1.602e+02
  User objective function returned NaN; trying a new point...
      2      29    2.470200e+02    0.000e+00    1.015e+02    5.340e+00
      3      35    8.972820e+01    0.000e+00    5.300e+01    4.382e+00
      4      41    6.477636e+01    0.000e+00    1.000e+01    2.883e+00
      5      47    6.067017e+01    0.000e+00    1.829e+01    8.142e-01
      6      53    5.640428e+01    0.000e+00    2.277e+01    8.831e-01
      7      59    4.446661e+01    0.000e+00    2.729e+01    2.578e+00
      8      65    3.106304e+01    0.000e+00    2.037e+01    3.292e+00
      9      72    2.835120e+01    0.000e+00    8.090e+00    6.124e-01
     10      78    2.596154e+01    0.000e+00    1.259e+00    1.105e+00
     11      84    2.594904e+01    0.000e+00    2.985e-01    2.495e-02
     12      90    2.594832e+01    0.000e+00    1.000e-01    5.350e-03
     13      96    2.594832e+01    0.000e+00    1.995e-02    6.247e-04
     14     102    2.594832e+01    0.000e+00    1.996e-04    3.078e-05
     15     108    2.594832e+01    0.000e+00    3.789e-06    6.280e-06
     16     114    2.594832e+01    0.000e+00    2.000e-06    2.072e-06
     17     121    2.594832e+01    0.000e+00    1.780e-06    1.765e-07
     18     127    2.594832e+01    0.000e+00    2.175e-06    5.074e-08
     19     135    2.594832e+01    0.000e+00    1.263e-06    6.788e-08
     20     147    2.594832e+01    0.000e+00    1.582e-06    3.423e-08
     21     156    2.594832e+01    0.000e+00    1.780e-06    7.848e-09

Local minimum possible. Constraints satisfied.

fmincon stopped because the size of the current step is less than
the selected value of the step size tolerance and constraints are 
satisfied to within the default value of the constraint tolerance.

So unless you give us the error message you saw, I regard that part of the question as answered.

For the second question: no, you cannot tell if there are other local minima by looking at the result of the first minimization run.

For more details, see Local Vs. Global Optima.

Alan Weiss

MATLAB mathematical toolbox documentation