The answer is that you cannot ensure that nonlinear constraints are satisfied at every step. Instead, your code must accept points that are infeasible and handle them appropriately. You can put try-catch statements in your code to handle errors that result from infeasible points, or do something else. Your code usually should return NaN for infeasible points. Or it can return an arbitrary large value, such as 1e6, which presumably is too large to be a minimum.
Alan Weiss
MATLAB mathematical toolbox documentation
