Karesh-Kuhn-Tucker Optimality Criteria

Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

Optimality Criteria

Big question: How do we know that we have found the optimum for min f(x)?
Answer: Test the solution for the necessary and sufficient conditions

Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

Optimality Conditions Unconstrained Case

Let x* be the point that we think is the minimum for f(x) Necessary condition (for optimality): f(x*) = 0 A point that satisfies the necessary condition is a stationary point
It can be a minimum, maximum, or saddle point

How do we know that we have a minimum? Answer: Sufficiency Condition: The sufficient conditions for x* to be a strict local minimum are: f(x*) = 0 Georgia Institute of Technology 2f(x*) is positive definite
Optimization in Engineering Design

Systems Realization Laboratory

Constrained Case KKT Conditions To proof a claim of optimality in constrained minimization (or maximization), we have to check the found point with respect to the (Karesh) Kuhn Tucker conditions. Kuhn and Tucker extended the Lagrangian theory to include the general classical single-objective nonlinear programming problem:
minimize Subject to f(x) gj(x) 0 for j = 1, 2, ..., J hk(x) = 0 for k = 1, 2, ..., K x = (x1, x2, ..., xN)
Georgia Institute of Technology Systems Realization Laboratory

Optimization in Engineering Design

Interior versus Exterior Solutions

Interior: If no constraints are active and (thus) the solution lies at the interior of the feasible space, then the necessary condition for optimality is same as for unconstrained case: f(x*) = 0 Exterior: If solution lies at the exterior, then the condition f(x*) = 0 does not apply because some constraints will block movement to this minimum.
Some constraints will (thus) be active.

We cannot get any more improvement (in this case) if for x* there does not exist a vector d that is both a descent direction and a feasible direction.
In other words: the possible feasible directions do not intersect the possible descent directions at all. See Figure 5.2
Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

Mathematical Form
A vector d that is both descending and feasible cannot exist if -f = mi (gi) (with mi 0) for all active constraints iI.
See page 152-153 This can be rewritten as 0 = f + mi (gi) This condition is correct IF feasibility is defined as g(x) 0. If feasibility is defined as g(x) 0, then this becomes -f = mi (-gi)

Again, this only applies for the active constraints. Usually the inactive constraints are included as well, but the condition mj gj = 0 (with mj 0) is added for all inactive constraints jJ.
This is referred to as the complimentary slackness condition. Note that this condition is equivalent to stating that mj = 0 for inactive constraints

Note that I+J = m, the total number of (inequality) constraints. Georgia Institute of Technology
Optimization in Engineering Design

Systems Realization Laboratory

Necessary KKT Conditions

For the problem: Min f(x) s.t. g(x) 0 (n variables, m constraints) The necessary conditions are:
f(x) + mi gi(x) = 0 (optimality) gi(x) 0 for i = 1, 2, ..., m (feasibility) mi gi(x) = 0 for i = 1, 2, ..., m (complementary slackness condition) mi 0 for i = 1, 2, ..., m (non-negativity)

Note that the first condition gives n equations.

Georgia Institute of Technology Systems Realization Laboratory

Optimization in Engineering Design

Necessary KKT Conditions (General Case)

For general case (n variables, M Inequalities, L equalities):
Min f(x) s.t. gi(x) 0 for i = 1, 2, ..., M hj(x) = 0 for J = 1, 2, ..., L

In all this, the assumption is that gj(x*) for j belonging to active constraints and hk(x*) for k = 1, ...,K are linearly independent
This is referred to as constraint qualification

The necessary conditions are:

f(x) + mi gi(x) + lj hj(x) = 0 (optimality) gi(x) 0 for i = 1, 2, ..., M (feasibility) hj(x) = 0 for j = 1, 2, ..., L (feasibility) mi gi(x) = 0 for i = 1, 2, ..., M (complementary slackness condition) mi 0 for i = 1, 2, ..., M (non-negativity) (Note: lj is unrestricted in sign)
Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

Necessary KKT Conditions (if g(x)0)

If the definition of feasibility changes, the optimality and feasibility conditions change.

The necessary conditions become:

f(x) - mi gi(x) + lj hj(x) = 0 (optimality) gi(x) 0 for i = 1, 2, ..., M (feasibility) hj(x) = 0 for j = 1, 2, ..., L (feasibility) mi gi(x) = 0 for i = 1, 2, ..., M (complementary slackness condition) mi 0 for i = 1, 2, ..., M (non-negativity)

Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

Restating the Optimization Problem Kuhn Tucker Optimization Problem: Find vectors x(Nx1), m(1xM) and l (1xK) that satisfy:
f(x) + mi gi(x) + lj hj(x) = 0 (optimality) gi(x) 0 for i = 1, 2, ..., M (feasibility) hj(x) = 0 for j = 1, 2, ..., L (feasibility) mi gi(x) = 0 for i = 1, 2, ..., M (complementary slackness condition) mi 0 for i = 1, 2, ..., M (non-negativity)

If x* is an optimal solution to NLP, then there exists a (m*, l*) such that (x*, m*, l*) solves the KuhnTucker problem. Above equations not only give the necessary conditions for optimality, but also provide a way of finding the optimal point.
Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

10

Limitations Necessity theorem helps identify points that are not optimal. A point is not optimal if it does not satisfy the KuhnTucker conditions. On the other hand, not all points that satisfy the KuhnTucker conditions are optimal points. The KuhnTucker sufficiency theorem gives conditions under which a point becomes an optimal solution to a single-objective NLP.

Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

11

Sufficiency Condition
Sufficient conditions that a point x* is a strict local minimum of the classical single objective NLP problem, where f, gj, and hk are twice differentiable functions are that 1) The necessary KKT conditions are met. 2) The Hessian matrix 2L(x*) = 2f(x*) + mi2gi(x*) + lj2hj(x*) is positive definite on a subspace of Rn as defined by the condition: yT 2L(x*) y 0 is met for every vector y(1xN) satisfying: gj(x*)y = 0 for j belonging to I1 = { j | gj(x*) = 0, uj* > 0} (active constraints) hk(x*)y = 0 for k = 1, ..., K y0

Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

12

KKT Sufficiency Theorem (Special Case) Consider the classical single objective NLP problem.
minimize Subject to f(x) gj(x) 0 for j = 1, 2, ..., J hk(x) = 0 for k = 1, 2, ..., K x = (x1, x2, ..., xN)

Let the objective function f(x) be convex, the inequality constraints gj(x) be all convex functions for j = 1, ..., J, and the equality constraints hk(x) for k = 1, ..., K be linear. If this is true, then the necessary KKT conditions are also sufficient. Therefore, in this case, if there exists a solution x* that satisfies the KKT necessary conditions, then x* is an optimal solution to the NLP problem. In fact, it is a global optimum.
Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

13

Limitations .

Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

14

Closing Remarks Kuhn-Tucker Conditions are an extension of Lagrangian function and method. They provide powerful means to verify solutions But there are limitations
Sufficiency conditions are difficult to verify. Practical problems do not have required nice properties. For example, You will have a problems if you do not know the explicit constraint equations (e.g., in FEM).

If you have a multi-objective (lexicographic) formulation, then I would suggest testing each priority level separately.

Optimization in Engineering Design

Georgia Institute of Technology Systems Realization Laboratory

15