# Multiple constraint feasibility

Hello dear friends,

I’m trying to solve the following feasibility problem in CVX as:

cvx_begin SDP quiet

cvx_precision high
variable X( N , N ) hermitian

maximize 0

subject to

X           >= 0 ;
trace(X*A) == 1;
trace(X*B) == x_min*trace(X*C);

cvx_end


where \bf{A}, \bf{B} and \bf{C} are Hermitian matrices containing constant values. The problem is solved by exhaustive or bisection search over the parameter x_{min}.

When I run it, it returns infeasible for the “cvx_status” string. On the other hand, a random Gaussian complex vector as \bf{x} can be generated and normalized to satisfies the constraint “trace(X*A) == 1”. Furthermore, the corresponding parameter x_{min} can be found. Then, if we solve the above problem using the found x_{min}, it becomes infeasible. This happens while we just found a feasible solution for the problem by the same x_{min}.

If we omit the constraint “trace(X*A) == 1” from the problem, it becomes solved. But according to above paragraph, it seems that there is a problem with CVX.

Best regards,
Ashkan

It seems that the source of the problem are low order matrices containing elements from the order 10^{-15}. If you multiply the matrices by 10^{15} and scale them up, the problems are not infeasible anymore and get solved.

It seems that either MATLAB or CVX has difficulty handling very small numbers.

Oh, this is most definitely the case. At some point, the solvers must decide whether a number is “close to zero” or “truly zero”.

Dear Michael, thanks for the insight.