05868750

j

05868750 Andreani, Roberto Mart{\'\i}nez, J.M. Svaiter, B.F. A new sequential optimality condition for constrained optimization and algorithmic consequences. SIAM J. Optim. 20, No. 6, 3533-3554 (2010). 2010 Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA EN nonlinear programming optimality condition approximate KKT conditions constraint qualification stopping criteria

doi:10.1137/090777189

The authors provide improved Karush-Kuhn-Tucker conditions (KKT) in limit form for the nonlinear optimization problem $$\text{Minimize}\quad f(x)\text{ subject to }h(x)= 0,\ g(x)\le 0,\tag P$$ where $f:\bbfR^n\to \bbfR$, $h: \bbfR^n\to \bbfR^m$, $g: \bbfR^n\to \bbfR^p$ have continuous first derivatives. In detail, a feasible point $x^*$ fulfills the ``complementary approximate Karush-Kuhn-Tucker conditions'' (CAKKT) if there exist sequences $\{x^k\}\subset\bbfR^n$, $\{\lambda^k\}\subset\bbfR^m$ and $\{\mu^k\}\subset \bbfR^p_+$ such that $$\lim_{k\to \infty} s^k= x^*,$$ $$\lim_{k\to \infty}\Vert\nabla f(x^k)+ \nabla h(x^k)\lambda^k+\nabla gf(x^k)\mu^k\Vert= 0,$$ $$\align \lim_{k\to \infty}\mu^k_i g_i(x^k)= 0\quad &\text{for all }i= 1,\dots, p,\\ \lim_{k\to\infty} \lambda^k_i h_i(x^k)= 0\quad &\text{for all }i= 1,\dots, m.\endalign$$ Obviously, CAKKT are satisfied if the classical KKT hold. In the main theorems of the paper it is proved that 1. a minimizer of (P) fulfills CAKKT -- without constraint qualification, 2. if a constraint qualification is satisfied then CAKKT are sufficient for KKT, 3. CAKKT are stronger than other known approximate KKT, 4. in the convex-affine case CAKKT are sufficient for minimality. It is pointed out that optimization methods (e.g. the augmented Lagrangian method under suitable assumptions) which produces CAKKT sequences are more efficient. J\"org Thierfelder (Ilmenau)