When does relaxation reduce the minimum cost of an optimal control problem?