The Karush-Kuhn-Tucker. Nuno Vasconcelos ECE Department, UCSD

Similar documents
Optimization. Nuno Vasconcelos ECE Department, UCSD

Linear discriminants. Nuno Vasconcelos ECE Department, UCSD

Shuai Dong. Isaac Newton. Gottfried Leibniz

Mathematical Economics MEMF e ME. Filomena Garcia. Topic 2 Calculus

OPTIMISATION. Introduction Single Variable Unconstrained Optimisation Multivariable Unconstrained Optimisation Linear Programming

PHYS 705: Classical Mechanics. Calculus of Variations II

ORDINARY DIFFERENTIAL EQUATIONS EULER S METHOD

, rst we solve te PDE's L ad L ad n g g (x) = ; = ; ; ; n () (x) = () Ten, we nd te uncton (x), te lnearzng eedbac and coordnates transormaton are gve

Solutions HW #2. minimize. Ax = b. Give the dual problem, and make the implicit equality constraints explicit. Solution.

Finite Difference Method

COS 521: Advanced Algorithms Game Theory and Linear Programming

On a nonlinear compactness lemma in L p (0, T ; B).

Lagrange Multipliers Kernel Trick

Assortment Optimization under MNL

Maximal Margin Classifier

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

Physics 2A Chapter 3 HW Solutions

MMA and GCMMA two methods for nonlinear optimization

Solutions to exam in SF1811 Optimization, Jan 14, 2015

Which Separator? Spring 1

Complex Variables. Chapter 18 Integration in the Complex Plane. March 12, 2013 Lecturer: Shih-Yuan Chen

A Simple Research of Divisor Graphs

EE 330 Lecture 24. Small Signal Analysis Small Signal Analysis of BJT Amplifier

The Gaussian classifier. Nuno Vasconcelos ECE Department, UCSD

On Pfaff s solution of the Pfaff problem

CHAPTER-5 INFORMATION MEASURE OF FUZZY MATRIX AND FUZZY BINARY RELATION

CENTROID (AĞIRLIK MERKEZİ )

CENTROID (AĞIRLIK MERKEZİ )

5 The Laplace Equation in a convex polygon

Problem Set 4: Sketch of Solutions

CIS526: Machine Learning Lecture 3 (Sept 16, 2003) Linear Regression. Preparation help: Xiaoying Huang. x 1 θ 1 output... θ M x M

Work is the change in energy of a system (neglecting heat transfer). To examine what could

CHAPTER 6 CONSTRAINED OPTIMIZATION 1: K-T CONDITIONS

Parameter estimation class 5

Single Variable Optimization

CHAPTER 7 CONSTRAINED OPTIMIZATION 1: THE KARUSH-KUHN-TUCKER CONDITIONS

14 Lagrange Multipliers

Solutions to Homework 7, Mathematics 1. 1 x. (arccos x) (arccos x) 1

Lecture 2 Solution of Nonlinear Equations ( Root Finding Problems )

Section 15.6 Directional Derivatives and the Gradient Vector

Competitive Experimentation and Private Information

The Finite Element Method: A Short Introduction

Machine Learning. What is a good Decision Boundary? Support Vector Machines

CHAPTER 7 CONSTRAINED OPTIMIZATION 2: SQP AND GRG

Logarithmic functions

Practical Newton s Method

Stanford University CS254: Computational Complexity Notes 7 Luca Trevisan January 29, Notes for Lecture 7

Some modelling aspects for the Matlab implementation of MMA

Maximum Likelihood Estimation (MLE)

Pattern Classification

Not-for-Publication Appendix to Optimal Asymptotic Least Aquares Estimation in a Singular Set-up

Degrees of Freedom. Spherical (ball & socket) 3 (3 rotation) Two-Angle (universal) 2 (2 rotation)

Endogenous timing in a mixed oligopoly consisting of a single public firm and foreign competitors. Abstract

Lecture Notes on Linear Regression

Chapter 3 Differentiation and Integration

Exercises of Fundamentals of Chemical Processes

Convex Optimization. Optimality conditions. (EE227BT: UC Berkeley) Lecture 9 (Optimality; Conic duality) 9/25/14. Laurent El Ghaoui.

Image classification. Given the bag-of-features representations of images from different classes, how do we learn a model for distinguishing i them?

Lecture 10 Support Vector Machines. Oct

Quantum Mechanics I - Session 4

Lecture 10 Support Vector Machines II

Kernel Methods and SVMs Extension

Exponentials and Logarithms Review Part 2: Exponentials

Machine Learning. Support Vector Machines. Eric Xing , Fall Lecture 9, October 6, 2015

Affine transformations and convexity

2.8 The Derivative as a Function

Bayesian decision theory. Nuno Vasconcelos ECE Department, UCSD

ESCUELA LATINOAMERICANA DE COOPERACIÓN Y DESARROLLO Especialización en Cooperación Internacional para el Desarrollo

Physics 2A Chapters 6 - Work & Energy Fall 2017

Chapter Eight. Review and Summary. Two methods in solid mechanics ---- vectorial methods and energy methods or variational methods

Summary with Examples for Root finding Methods -Bisection -Newton Raphson -Secant

What is LP? LP is an optimization technique that allocates limited resources among competing activities in the best possible manner.

( ) 2 ( ) ( ) Problem Set 4 Suggested Solutions. Problem 1

: Numerical Analysis Topic 2: Solution of Nonlinear Equations Lectures 5-11:

Lecture 26 Finite Differences and Boundary Value Problems

Polynomial Barrier Method for Solving Linear Programming Problems

Pre-Calculus Summer Assignment

PHYS 1101 Practice problem set 12, Chapter 32: 21, 22, 24, 57, 61, 83 Chapter 33: 7, 12, 32, 38, 44, 49, 76

SELECTED SOLUTIONS, SECTION (Weak duality) Prove that the primal and dual values p and d defined by equations (4.3.2) and (4.3.3) satisfy p d.

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

General Tips on How to Do Well in Physics Exams. 1. Establish a good habit in keeping track of your steps. For example, when you use the equation

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

Mean Field / Variational Approximations

One Dimensional Axial Deformations

CHAPTER 4d. ROOTS OF EQUATIONS

SIMPLE LINEAR REGRESSION

TR/95 February Splines G. H. BEHFOROOZ* & N. PAPAMICHAEL

COMP4630: λ-calculus

1. Fundamentals 1.1 Probability Theory Sample Space and Probability Random Variables Limit Theories

The Prncpal Component Transform The Prncpal Component Transform s also called Karhunen-Loeve Transform (KLT, Hotellng Transform, oregenvector Transfor

Continuity and Differentiability Worksheet

( ) [ ( k) ( k) ( x) ( ) ( ) ( ) [ ] ξ [ ] [ ] [ ] ( )( ) i ( ) ( )( ) 2! ( ) = ( ) 3 Interpolation. Polynomial Approximation.

Maximum Likelihood Estimation

DIFFERENTIAL SCHEMES

Neural networks. Nuno Vasconcelos ECE Department, UCSD

Classical Mechanics Virtual Work & d Alembert s Principle

Spring Force and Power

Numerical Differentiation

f(x,y) = (4(x 2 4)x,2y) = 0 H(x,y) =

Math 324 Advanced Financial Mathematics Spring 2008 Final Exam Solutions May 2, 2008

Transcription:

e Karus-Kun-ucker condtons and dualt Nuno Vasconcelos ECE Department, UCSD

Optmzaton goal: nd mamum or mnmum o a uncton Denton: gven unctons, g, 1,...,k and, 1,...m dened on some doman Ω R n mn w, w Ω subect to g w w,, or compactness we wrte gw nstead o g w,. Smlarl w we derved necessar and sucent or local optmalt n te absence o constrants equalt constrants onl

Mnma condtons unconstraned let w be contnuousl derentable w s a local mnmum o w and onl w s a local mnmum o w and onl as zero gradent at w w and te Hessan o at w s postve dente w n t d d w d R were d d w d R, 1 n M L 3 1 1 n n L

Mama condtons unconstraned let w be contnuousl derentable w s a local mamum o w and onl w s a local mamum o w and onl as zero gradent at w w and te Hessan o at w s negatve dente w n t d d w d R were d d w d R, 1 n M L 4 1 1 n n L

Constraned optmzaton wt equalt constrants onl eorem: consder te problem arg mn subect to were te constrant gradents are lnearl ndependent. en s a soluton and onl tere ets a unque vector λ, suc tat m λ 1 + λ + m 1 λ, s.t. 5

Alternatve ormulaton state te condtons troug te Lagrangan m t t b tl tt, 1 L m + λ λ te teorem can be compactl wrtten as, L λ,, L L L st L, λ λ λ λ λ te entres o λ are reerred to as Lagrange multplers,, L s.t. λ 6

Geometrc nterpretaton dervatve o along d s lm α w + αd α w d w d. w.cos d, w ts means tatt greatest ncrease wen d no ncrease wen d snce tere s no ncrease wen d s tangent to so-contour k te gradent s perpendcular to te tangent o te so-contour allows geometrc nterpretaton o te Lagrangan g condtons no ncrease 7

Lagrangan optmzaton geometrc nterpretaton: snce s a so-contour o, s perpendcular to te so-contour sas tat span{ }.e. to tangent space o te constrant surace ntutve drecton o largest ncrease o s to constrant surace te gradent s zero along te constrant no wa to gve an nntesmal gradent step, wtout endng up volatng t t s mpossble to ncrease and stll sats te constrant span{} tg plane 8

Eample consder te problem mn 1 + subect to 1 + 1 1 to te so-contours o 1 + k 1 1 1 1 + 1 + 1 9 1 + -1

Eample consder te problem mn 1 + subect to 1 + to te so-contour o 1 + - 1 + 1 1 1 1 1 + 1 + -1 1

Eample recall tat dervatve along d s w + +α d w lm d. w.cos d, w α α crtcal pont crtcal pont - movng along te tangent s descent as long as cos tg, < -.e. π/ < angle,tg < 3π/ - can alwas nd suc d unless tg - crtcal pont wen - to nd wc tpe we need nd order as beore 11

Alternatve vew consder te tangent space to te so-contour ts s te subspace o rst order easble varatons { } V, space o or wc + satses te constrant up to rst order appromaton V easble varatons 1

Feasble varatons multplng our rst Lagrangan condton b + λ t ollows tat m 1, V ts s a generalzaton o n unconstraned case mples tat V and tereore note: Hessan constrant onl dened or n V makes sense: we cannot move anwere else, does not reall matter wat Hessan s outsde V 13

Inequalt constrants wat appens wen we ntroduce nequaltes? arg mn subect to, g we start b denng te set o actve nequalt constrants { g } A or eample 1, 1 +, -5 1 5, -5 5 14

Actve nequalt constrants we ave a mnmum at,-5 1 5 <, - 1-5 <, and 5 < arenactve - 5 s actve -5 note tat a local mnmum or ts problem would stll be a local mnmum we removed te nnactve constrants nnactve constrants do not do antng actve constrants are equaltes nnactve actve 15

Constraned optmzaton ence, te problem arg mn subect to, g s equvalent to arg mn subect to, g, A ts s a problem wt equalt constrants, tere must be a λ and µ, A, suc tat + m 1 λ + µ g A wc does not cange we assgn a zero Lagrange g multpler to te nnactve constrants 16

Constraned optmzaton lettng µ, A, + m 1 r λ + µ g 1 tere s one nal constrant t wc s µ, or all due to te ollowng pcture g as to pont nward oterwse we would ave a mamum o g as to pont outward oterwse g would ncrease nward,.e. g would be non-negatve g nsde wen we put togeter all tese constrants we obtan te amed Karus-Kun-ucker Kun KK condtons 17

e KK condtons eorem: or te problem, arg mn g to subect s a local mnmum and onl tere est λ and µ suc tat 1 1 g r m + + µ λ,, v A r m µ µ { },,, 1 1 A g V were V g v r m + + and µ λ 18 { },, A g V were and

Geometrc nterpretaton Let s orget te equalt constrants or now later we wll see tat te do not cange muc consder te problem arg mn subect to g rom te KK condtons, te soluton satses µ [ L, µ ], µ, A wt L, µ r + 1 µ g 19

Geometrc nterpretaton wc s equvalent to [ ] g [ L, µ ] mn + µ L mn µ wt we tus ave µ, and µ, A + µ g L + µ g L or w z - b w z - b were 1 b L, w, z µ g plane n z-space normal w, bas b L s n al-space ponted to b w

Geometrc Interpretaton rom b we ave L, 1 w, z µ g snce µ, w s alwas n te rst quadrant snce rst coordnate s 1, w s never parallel to g as ts can be vsualzed n z-space space as also, two cases: 1 g w z - b L te -ntercept s,l, s te mnmum o L w admssble planes w R g,, g R r-1 1

Geometrc Interpretaton b L, 1 w, µ case g< z g te constrants are nnactve µ w 1, plane s orzontal w z - b L te -ntercept s,l, s te mnmum o L.e. ts s alwas te case n general m o actve and nnactve but beavor s ts w g, g< R, g R r-1

In summar L mn wt µ [ g ] [ L, µ ] mn + µ s equvalent to, and µ, A w z - b 1 b L, w w, z z - b µ g can be vsualzed as g R g< w R easble g,, w g,, g R r-1 g< g R r-1 3

Dualt [ g ] [ L, µ ] mn + L mn µ wt µ, and µ, A does not appear terrbl dcult once I know µ ow I do nd ts value? consder ts uncton or an µ [ L, µ ] mn[ g ] q µ mn + µ wt µ ts s equvalent to w z - b b q µ, w w z - b 1, z µ g te pcture s te same wt L replaced b qµ 4

Dualt notng tat perplane w,b stll as to support easble set and we stll ave µ> ts leads to w g R g< w easble w R g,, w,,qµ g R r-1 g< g R r-1,qµ 5

Dualt note tat qµ L we keep ncreasng qµ we wll get qµ L we cannot go beond L would move to g > ts s eactl te denton o te dual problem ma q µ µ note: [ L, µ ] mn[ g ] q µ mn + µ wt µ qµ ma go to - or some µ, wc means tat tere s no Lagrange multpler plane would be vertcal. ts s avoded b ntroducng te constrant { µ µ > } µ D q D q 6

Dualt we tereore ave a two step recpe to nd te optmal soluton 1.or an µ, solve [ L, µ ] mn[ µ g ] q µ mn +.ten solve ma q µ µ, µ Dq one o te reasons w ts s nterestng s tat te second problem turns out to be qute manageable 7

Dualt eorem: D q s a conve set and q s concave on D q Proo: or an, µ, µ and α [,1] L + 1 α µ + αµ + 1 α µ, αµ g + αµ g + 1 α µ g [ µ g ] 1 α [ µ g ] + + + α α L, µ + 1 α L, µ and takng mn on bot sdes mn L we ave, αµ + 1 α µ mn[ αl, µ + 1 α L, µ ] α mn L, µ + 1 α mn L, µ αµ + 1 α µ αq µ + 1 α q µ q 8

Dualt we ave αµ + 1 α µ αq µ + 1 α q µ q rom wc two conclusons ollow µ D q and µ D q qµ>-, qµ>- αµ+ 1-αµ D q Hence D q s sconve b denton o concavt mples tat q s concave over D q ts s a ver appealng result, snce conve optmzaton problems are among te easest to solve note tat te dual s alwas concave, rrespectve o te prmal problem te ollowng result onl proves wat we alread ave nerred rom te grapcal nterpretaton 9

Dualt eorem: weak dualt t s alwas true tat Proo: q or an µ, and wt g, q µ mn L z, µ + µ g z snce µ g. Hence q ma q µ µ and snce ts olds or an q mn, g 3

Dualt gap q we sa tat tere s no dualt gap. Oterwse tere s a dualt gap. te DG constrans te estence o Lagrange multplers eorem: tere s no dualt gap, te set o Lagrange multplers s te set o optmal dual solutons tere s a dualt gap, tere are no Lagrange multplers l Proo: b denton, µ µ s a Lagrange multpler and onl q µ q wc, rom te prevous teorem, olds and onl q,.e. tere s no dualt gap 31

Dualt gap note tat tere are stuatons n wc te dual problem as a soluton, but or wc tere s no Lagrange g multpler. or eample: ts s a vald dual problem owever te constrant µ g s not satsed and R, q g < g R r-1 n summar, dualt s nterestng onl wen tere s no dualt gap,q 3

33