“…Roughly speaking, we will use a projected gradient direction as a descent direction to calculate the next iterate when the current projected Newton direction is not descent. As will be shown in Section 2, this modification will not only ensure the global convergence of the projected Newton-Krylov methods in [1] but also inherit the computational advantages of these methods, such as the capacity of solving extreme large-scale problems mentioned earlier, the matrix-free operation, and preconditioning technique, and so on, see, for example, [7]. In other words, we extend the theory of Newton-Krylov methods in [2,8] fully to the projected case.…”