Why does fminunc give me the message "User objective function returned NaN; trying a new point..."?

Hi all, I have to solve a minimization problem using fminunc and many starting values but the code stucks somewhere saying that "User objective function returned NaN; trying a new point...". How can I fix this problem? Sometimes it is not even able to solve one optimization.

Here the code:

1) Generation of data

clear all

T=500;  
n=2; 
alpha=[1 1]; %vector 1xn alpha1 alpha2
beta=[2 2]; %vector 1xn beta1 beta2
rho=0.2; %corr(u1, u2)
thetatrue=[alpha beta rho]; 
mu=zeros(1,n);  %mean vector 1xn
sigma=[1 rho; rho 1]; %variance covariance matrix nxn
optionsunc=optimset('Display','iter','MaxIter',100000000,'TolX',10^-30,'TolFun',10^-30,   'MaxFunEvals', 10000000); %options for fminunc

X=[unidrnd(2,T,1)-1 unidrnd(4,T,1)-2]; %matrix Txn
profit1= repmat(alpha,T,1)+[X(:,1).*beta(1) X(:,2).*beta(2)]; %vector Txn 
u = mvnrnd(mu,sigma,T); %vector Txn
profit=profit1+u; %matrix Txn           
prof=(profit>0)==zeros(T,2);                          
sumequil=sum(prof,2);                           
W1=(sumequil==n*ones(T,1)); 
W2=1-W1; 
data=[X W1 W2];

2) Setting up the minimization problem

A1(:,2)=-X(:,1); 
A1(:,1)=-1;   
A2(:,2)=-X(:,2); 
A2(:,1)=-1;
rhogrid=[(-0.9:0.1:0.1) (0.11:0.01:0.19) (0.21:0.01:0.29) (0.3:0.1:0.9)];
L_nstarho=zeros(size(rhogrid,2),1); %allocate space for the max log-lik value at each rho

for j=1:size(rhogrid,2) %for each value of rho
theta0grid=[]; %initiate theta0grid 
while size(theta0grid,1)==0
% Grid of starting values alpha1, alpha2, beta1, beta2 for each rho
alpha1=normrnd(thetatrue(1),1,1,2);   %random draws from N(alpha1true,1) 
alpha2=normrnd(thetatrue(2),1,1,2);   %random draws from N(alpha2true,1)
beta1=normrnd(thetatrue(3),1,1,2);    %random draws from N(beta1true,1)
beta2=normrnd(thetatrue(4),1,1,2);    %random draws from N(beta2true,1) 
CC = {alpha1,alpha2,beta1,beta2};
b = cellfun(@numel,CC);
n1 = cumsum(b);
a = fullfact(b);
idx = bsxfun(@plus,a,[0, n1(1:end-1)]);
v = cat(2,CC{:});
theta0grid = v(idx);  % grid of starting values 
cdfun=@(x) mvncdf( [A1*[x(1);x(3)],  A2*[x(2);x(4)] ]...
                                        ,[0 0],[1 rhogrid(j); rhogrid(j) 1]); %F(-alpha1-beta1*x1, -alpha2-beta2*x2; rhogrid(j))
% Check feasibility of starting values                  
drop=false(size(theta0grid,1),1); 
for i=1:size(theta0grid,1)
%P=cdfun(theta0grid(i,:));
fval=log_lik(theta0grid(i,:),cdfun,W1,W2);
drop(i)=((isnan(fval)) | fval==0);
end
theta0grid(drop,:)=[]; 
end

l_nstarho=zeros(size(theta0grid,1),1);  %allocate the space for listing max log-lik at each starting value (given rhogrid(j))

for t=1:size(theta0grid,1) %for each starting value 
theta0=theta0grid(t,:);  
[theta,fval,exitflag,output]=fminunc(@(x) log_lik(x,cdfun,W1,W2),theta0,optionsunc);
if exitflag>0   %if convergence
l_nstarho(t)=-fval;
else   
l_nstarho(t)=nan;  
end
end
L_nstarho(j)=max(l_nstarho);  %pick the max among the values obtained for each starting value
end

Answer 1

[Walter Roberson]

At the command prompt, command

dbstop if naninf

and run. The program will stop in the debugger when the program generates infinity or nan.

NaN can be 0/0 or 0 * infinity