The Karush-Kuhn-Tucker conditions

Similar documents
Lecture 18: Optimization Programming

I.3. LMI DUALITY. Didier HENRION EECI Graduate School on Control Supélec - Spring 2010

Duality. Lagrange dual problem weak and strong duality optimality conditions perturbation and sensitivity analysis generalized inequalities

Optimization Problems with Constraints - introduction to theory, numerical Methods and applications

IE 5531: Engineering Optimization I

Convex Optimization Boyd & Vandenberghe. 5. Duality

More on Lagrange multipliers

Optimality, Duality, Complementarity for Constrained Optimization

4TE3/6TE3. Algorithms for. Continuous Optimization

5. Duality. Lagrangian

The Karush-Kuhn-Tucker (KKT) conditions

Numerical Optimization

Lecture: Duality.

Optimality Conditions for Constrained Optimization

TMA 4180 Optimeringsteori KARUSH-KUHN-TUCKER THEOREM

Constrained Optimization

Convex Optimization M2

In view of (31), the second of these is equal to the identity I on E m, while this, in view of (30), implies that the first can be written

Lecture 3. Optimization Problems and Iterative Algorithms

Karush-Kuhn-Tucker Conditions. Lecturer: Ryan Tibshirani Convex Optimization /36-725

Lecture 2: Linear SVM in the Dual

On the Method of Lagrange Multipliers

Optimization over a polyhedron

Lectures 9 and 10: Constrained optimization problems and their optimality conditions

ICS-E4030 Kernel Methods in Machine Learning

Generalization to inequality constrained problem. Maximize

Constrained optimization

Nonlinear Optimization

Constrained Optimization and Lagrangian Duality

Solution Methods. Richard Lusby. Department of Management Engineering Technical University of Denmark

Nonlinear Optimization: What s important?

Convex Optimization & Lagrange Duality

Convex Optimization. Dani Yogatama. School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA. February 12, 2014

Lecture: Duality of LP, SOCP and SDP

Machine Learning. Support Vector Machines. Manfred Huber

CE 191: Civil & Environmental Engineering Systems Analysis. LEC 17 : Final Review

CS-E4830 Kernel Methods in Machine Learning

Lecture 6: Conic Optimization September 8

EE/AA 578, Univ of Washington, Fall Duality

Support Vector Machines: Maximum Margin Classifiers

SVM and Kernel machine

Structural and Multidisciplinary Optimization. P. Duysinx and P. Tossings

SECTION C: CONTINUOUS OPTIMISATION LECTURE 9: FIRST ORDER OPTIMALITY CONDITIONS FOR CONSTRAINED NONLINEAR PROGRAMMING

Introduction to Optimization Techniques. Nonlinear Optimization in Function Spaces

Math 5311 Constrained Optimization Notes

ISM206 Lecture Optimization of Nonlinear Objective with Linear Constraints

Mathematical Economics. Lecture Notes (in extracts)

Introduction to Machine Learning Lecture 7. Mehryar Mohri Courant Institute and Google Research

Chap 2. Optimality conditions

Optimality Conditions

Convex Optimization and Modeling

Lagrange Duality. Daniel P. Palomar. Hong Kong University of Science and Technology (HKUST)

CONSTRAINED NONLINEAR PROGRAMMING

The Degree of Central Curve in Quadratic Programming

5 Handling Constraints

Optimization. Yuh-Jye Lee. March 28, Data Science and Machine Intelligence Lab National Chiao Tung University 1 / 40

UNDERGROUND LECTURE NOTES 1: Optimality Conditions for Constrained Optimization Problems

Constrained optimization: direct methods (cont.)

Support Vector Machines for Regression

Nonlinear Programming, Elastic Mode, SQP, MPEC, MPCC, complementarity

ON LICQ AND THE UNIQUENESS OF LAGRANGE MULTIPLIERS

Linear Support Vector Machine. Classification. Linear SVM. Huiping Cao. Huiping Cao, Slide 1/26

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation

Mechanical Systems II. Method of Lagrange Multipliers

Motivation. Lecture 2 Topics from Optimization and Duality. network utility maximization (NUM) problem:

Two hours. To be provided by Examinations Office: Mathematical Formula Tables. THE UNIVERSITY OF MANCHESTER. xx xxxx 2017 xx:xx xx.

MVE165/MMG631 Linear and integer optimization with applications Lecture 13 Overview of nonlinear programming. Ann-Brith Strömberg

Algorithms for constrained local optimization

Support vector machines

2.3 Linear Programming

Convex Optimization and SVM

MATH Dr. Pedro V squez UPRM. P. V squez (UPRM) Conferencia 1/ 17

Constrained Optimization Theory

Date: July 5, Contents

Constraint qualifications for nonlinear programming

Introduction to Machine Learning Prof. Sudeshna Sarkar Department of Computer Science and Engineering Indian Institute of Technology, Kharagpur

Convex Optimization Overview (cnt d)

4. Algebra and Duality

Introduction to Nonlinear Stochastic Programming

Scientific Computing: Optimization

5.5 Quadratic programming

6.254 : Game Theory with Engineering Applications Lecture 7: Supermodular Games

Pattern Classification, and Quadratic Problems

Support Vector Machine (SVM) and Kernel Methods

Example: feasibility. Interpretation as formal proof. Example: linear inequalities and Farkas lemma

1. f(β) 0 (that is, β is a feasible point for the constraints)

Support Vector Machines and Kernel Methods

Numerical optimization

AM 205: lecture 19. Last time: Conditions for optimality Today: Newton s method for optimization, survey of optimization methods

Convex Optimization Lecture 6: KKT Conditions, and applications

Course on Model Predictive Control Part II Linear MPC design

Support Vector Machine (SVM) & Kernel CE-717: Machine Learning Sharif University of Technology. M. Soleymani Fall 2012

Nonlinear Programming

Numerical optimization. Numerical optimization. Longest Shortest where Maximal Minimal. Fastest. Largest. Optimization problems

Interior Point Methods for Convex Quadratic and Convex Nonlinear Programming

Optimization. A first course on mathematics for economists

Lecture 7: Convex Optimizations

AM 205: lecture 19. Last time: Conditions for optimality, Newton s method for optimization Today: survey of optimization methods

Machine Learning Support Vector Machines. Prof. Matteo Matteucci

Cheng Soon Ong & Christian Walder. Canberra February June 2018

Transcription:

Chapter 6 The Karush-Kuhn-Tucker conditions 6.1 Introduction In this chapter we derive the first order necessary condition known as Karush-Kuhn-Tucker (KKT) conditions. To this aim we introduce the alternative theorems. 6.2 Alternative theorems and the Farkas Lemma We present in this section alternative theorems which allows to formulate the not existence of solution as the existence of an alternative system. Given two linear systems (I) e (II), an alternative theorem consists in proving that: System (I) has a solution if and only if system (II) does not have a solution. Among alternative theorems, one of the mst important that weare going to use to prove optimality conditions, is the Farkas Lemma. Theorem 6.3 (Farkas Lemma) Let B a p n matrix and g R n. System Bd 0 g T d < 0 (I) does not have a solution d R n if and only if system B T u = g u 0 (II) has a solution. u R p. 43

44 CHAPTER 6. THE KARUSH-KUHN-TUCKER CONDITIONS 6.4 The Karush-Kuhn-Tucker conditions Consider the problem min f (x) In section 5 we derive the first order optimality conditions (P POL) If x is a local minimizer of problem (P-POL) then there exists NO solution d R n of the system A I(x )d 0, f (x ) T d < 0. (6.1) where A I(x ) is the I(x ) n submatrix of A defined as A I(x ) = (a T i ) i I(x ) and I(x ) = {i : a T i x = b i }. Identifying the system (6.1) with system (I) of Farkas Lemma, we get There exists NO solution d R n to the linear system (6.1) if and only if there exists a solution of the system A T I(x ) u = f (x ), (6.2) u 0. where u R I(x ). Using this result we can state the optimality conditions for problem (P-POL) known as Karush-Kuhn-Tucker conditions (KKT). Theorem 6.5 (Karush-Kuhn-Tucker conditions) If x is a local minimizer of problem (P-POL). Then a multiplier λ R m exists that such that (i) f (x ) A T λ = 0, (iii) λ i (b i a T i x ) = 0 per i = 1,...,m. (iv) Ax b, Proof. Since x is a local minimizer, by definition it is feasible hence (iv) holds. Further there exists no solution to the system (6.1). Identifying the system (6.1) with system (I) of Farkas Lemma, we get that there exists NO solution che esiste to system(6.2). Let u i 0, i I(x ) a solution of the preceding system and define λ R m as: λ i = { u i peri I(x ) 0 peri / I(x ). (6.3)

6.4. THE KARUSH-KUHN-TUCKER CONDITIONS 45 By definition λ 0 and (ii) holds. Furthe (iii) derives immediately by the defintion (6.3) of λ. Indeed we have { λi (b i a T i x u i ) = (b i a T i x ) = 0 peri I(x ) 0(b i a T i x ) = 0 peri / I(x ). Finally,(6.2) can be written as f (x ) = and (i) holds. f (x ) = u i a i so that i I(x ) u i a i + 0 a i = A T λ i I(x ) i I(x ) The vector λ R m is called generalized Lagrange multiplier or KKT multiplier associated to constraints. Condition (iii) is called complementarity condition and states that at optimality the product λi (b i a i T x ) vanishes so that if the constraints is inactive the corresponding multiplier is zero. Since λ 0 and Ax b 0 condition (iii)can be equivalently written as λ T (b Ax ) = m i=1 λ i (b i a T i x ) = 0. We can express conditions above using the Lagrangian function. The Lagrangian function associated to problem (P-POL) is L(x,λ) = f (x) + λ T (b Ax) Condition (ii) states that in (x,λ ) the gradient x L(x,λ) = f (x) A T λ vanishes, so that the pair (x,λ ) ia a stationary point of the Lagrangian function. (Karush-Kuhn-Tucker conditions) Let x is a local minimizer of problem (P-POL). Then a multiplier λ R m exists that such that (i) x L(x,λ ) = 0 (stationary), (iii) λ T (b Ax ) = 0 (complementarity), (iv) Ax b, (feasibility).

46 CHAPTER 6. THE KARUSH-KUHN-TUCKER CONDITIONS Candidates to be minimizer of problem (P-POL) can be found by solving the KKT conditions. We consider the following definition. A point x is called KKT point or stationary point of problem (P-POL) if there exists λ R m such that. (i) A x b, (ii) L( x, λ) = 0, (iii) λ 0, (iv) λ T (b A x) = 0. under onvexity assumtpion on f we get Theorem 6.6 (Necessary and sufficient conditions for global optimzality) Let f be a continuously differentiable convex function in R n. A point x is a global minimizer of problem (P-POL) if and only if a multiplier λ R m exists that such that (i) Ax b, (ii) f (x ) A T λ = 0, (iv) λ T (b Ax ) = 0. If f is strictly convex and the conditions (i)-(iv) hold, then x is the unique global minimizer of f on S. We derive now the KKT conditions for a more general problem. min f (x) Dx = h where D is a p n matrix, A is a q n matrix, h R p, b R q. (P GEN) The Lagrangian function associated to (P-GEN) is with λ R q, µ R p. L(x,λ,µ) = f (x) + λ T (b Ax) + µ T (Dx h),

6.4. THE KARUSH-KUHN-TUCKER CONDITIONS 47 The KKT conditions for problem (P-GEN) are satisfied in (x,λ, µ ) R n R q R p when it holds: (i) Ax b, Dx = h (feasibility), (ii) x L(x,λ, µ ) = f (x ) A T λ + D T µ = 0 (stationarity), (iv) λ T (b Ax ) = 0 (complementarity). Theorem 6.7 (Karush-Kuhn-Tucker Conditions for (P-GEN)) Let x is a local minimizer of problem (P-GEN). Then multipliers λ R q, µ R p exist such that: (i) x L(x,λ, µ ) = 0, (iii) λ i (b i a T i x ) = 0 per i = 1,...,q. Consider the special case of Quadratic Programming (QP) problems min 1 2 xt Qx + c T x, (PQ) where Q is a n n symetric and positive semidefinite matrix, c R n. Theorem 6.8 (Optimality conditions for QP) Let Q be a n n symetric and positive semidefinite matrix. A point x is a global minimizer of problem (PQ) if and only if a multiplier λ R m exists that such that: (i) Ax b, (ii) Qx + c A T λ = 0, (iv) λ T (b Ax ) = 0. Further if Q is positive definite and 8i)-(iv) hold, then x is the unique global solution of problem (PQ).