Matlab 2020a outputs wrong fval for quadprog

Running the following code:

k = optimvar('k',2);
c = optimvar('c',2);
objec = (k-2)'*(k-2)+(c-1)'*(c-1);
prob = optimproblem('Objective',objec);
problem = prob2struct(prob);
[x,fval] = quadprog(problem);
sol.c = x(1:2);
sol.k = x(3:4);
% Matlab documention claims:
% [x,fval] = quadprog(___), for any input variables, also returns fval, the value of the objective function at x:
%       fval = 0.5*x'*H*x + f'*x
fval 
.5*x'*problem.H*x+problem.f'*x 
% the real objective function value at x is
evaluate(objec,sol) 
x'*problem.H*x+problem.f'*x 
% it seems the 0.5 multiplication is incorrect and this is a bug

yields the folloing output:

fval =

-10

ans =

-10

ans =

0

ans =

0

Is this a known bug with quadprog in Matlab 2020a?

Answer 1

[Alan Weiss]

Your problem has an added constant term that quadprog does not take into account, though solve does. Did you try calling

[sol fval] = solve(prob)

sol =

struct with fields:

c: [2×1 double]

k: [2×1 double]

fval =

0

To compute things the way you want, try this, where fval is the result of your quadprog optimization:

fval + problem.f0
% and
.5*x'*problem.H*x+problem.f'*x + problem.f0

Alan Weiss

MATLAB mathematical toolbox documentation