Conve functions 2{2 Conve functions f : R n! R is conve if dom f is conve and y 2 dom f 2 [0 1] + f ( + (1 ; )y) f () + (1 ; )f (y) (1) f is concave i

Similar documents
Lecture 2: Convex functions

Convex functions. Definition. f : R n! R is convex if dom f is a convex set and. f ( x +(1 )y) < f (x)+(1 )f (y) f ( x +(1 )y) apple f (x)+(1 )f (y)

3. Convex functions. basic properties and examples. operations that preserve convexity. the conjugate function. quasiconvex functions

Convex Functions. Daniel P. Palomar. Hong Kong University of Science and Technology (HKUST)

3. Convex functions. basic properties and examples. operations that preserve convexity. the conjugate function. quasiconvex functions

Convex Functions. Wing-Kin (Ken) Ma The Chinese University of Hong Kong (CUHK)

Convex Functions. Pontus Giselsson

Lecture 1: Background on Convex Analysis

A function(al) f is convex if dom f is a convex set, and. f(θx + (1 θ)y) < θf(x) + (1 θ)f(y) f(x) = x 3

Subgradient. Acknowledgement: this slides is based on Prof. Lieven Vandenberghes lecture notes. definition. subgradient calculus

Shiqian Ma, MAT-258A: Numerical Optimization 1. Chapter 4. Subgradient

Economics 205 Exercises

Convex Optimization. 4. Convex Optimization Problems. Prof. Ying Cui. Department of Electrical Engineering Shanghai Jiao Tong University

Lecture 10. ( x domf. Remark 1 The domain of the conjugate function is given by

EE/ACM Applications of Convex Optimization in Signal Processing and Communications Lecture 17

Mathematics 530. Practice Problems. n + 1 }

The proximal mapping

LMI MODELLING 4. CONVEX LMI MODELLING. Didier HENRION. LAAS-CNRS Toulouse, FR Czech Tech Univ Prague, CZ. Universidad de Valladolid, SP March 2009

1. Introduction. mathematical optimization. least-squares and linear programming. convex optimization. example. course goals and topics

1. Introduction. mathematical optimization. least-squares and linear programming. convex optimization. example. course goals and topics

IE 521 Convex Optimization

Optimality Conditions for Nonsmooth Convex Optimization

A Tutorial on Convex Optimization

Optimisation convexe: performance, complexité et applications.

CHAPTER 2: CONVEX SETS AND CONCAVE FUNCTIONS. W. Erwin Diewert January 31, 2008.

Optimization and Optimal Control in Banach Spaces

4. Convex optimization problems

Lecture 8: Basic convex analysis

CSCI : Optimization and Control of Networks. Review on Convex Optimization

Lecture: Convex Optimization Problems

A Brief Review on Convex Optimization

Solutions Chapter 5. The problem of finding the minimum distance from the origin to a line is written as. min 1 2 kxk2. subject to Ax = b.

Convex Optimization Problems. Prof. Daniel P. Palomar

1 Kernel methods & optimization

Nonlinear Programming Models

4. Convex optimization problems

8. Conjugate functions

Convex Functions and Optimization

UNCONSTRAINED OPTIMIZATION PAUL SCHRIMPF OCTOBER 24, 2013

Convex Optimization and Modeling

COURSE ON LMI PART I.2 GEOMETRY OF LMI SETS. Didier HENRION henrion

Linear and non-linear programming

Subgradients. subgradients and quasigradients. subgradient calculus. optimality conditions via subgradients. directional derivatives

Semidefinite Programming Basics and Applications

Convex analysis and profit/cost/support functions

N. L. P. NONLINEAR PROGRAMMING (NLP) deals with optimization models with at least one nonlinear function. NLP. Optimization. Models of following form:

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

HW1 solutions. 1. α Ef(x) β, where Ef(x) is the expected value of f(x), i.e., Ef(x) = n. i=1 p if(a i ). (The function f : R R is given.

Subgradients. subgradients. strong and weak subgradient calculus. optimality conditions via subgradients. directional derivatives

Geometric problems. Chapter Projection on a set. The distance of a point x 0 R n to a closed set C R n, in the norm, is defined as

Linear Algebra: Characteristic Value Problem

Convex Optimization in Communications and Signal Processing

Convex optimization problems. Optimization problem in standard form

subject to (x 2)(x 4) u,

Chapter 2 Convex Analysis

Lecture 7: Weak Duality

Translative Sets and Functions and their Applications to Risk Measure Theory and Nonlinear Separation

EC5555 Economics Masters Refresher Course in Mathematics September 2013

58 Appendix 1 fundamental inconsistent equation (1) can be obtained as a linear combination of the two equations in (2). This clearly implies that the

COM S 578X: Optimization for Machine Learning

Matrix Support Functional and its Applications

Determinant maximization with linear. S. Boyd, L. Vandenberghe, S.-P. Wu. Information Systems Laboratory. Stanford University

L. Vandenberghe EE236C (Spring 2016) 18. Symmetric cones. definition. spectral decomposition. quadratic representation. log-det barrier 18-1

9. Interpretations, Lifting, SOS and Moments

A : k n. Usually k > n otherwise easily the minimum is zero. Analytical solution:

MATH 409 Advanced Calculus I Lecture 7: Monotone sequences. The Bolzano-Weierstrass theorem.

Exercises. Exercises. Basic terminology and optimality conditions. 4.2 Consider the optimization problem

Unconstrained minimization: assumptions

Exercise 1. Let f be a nonnegative measurable function. Show that. where ϕ is taken over all simple functions with ϕ f. k 1.

Definition of convex function in a vector space

15. Conic optimization

Examples of Dual Spaces from Measure Theory

Constrained Optimization and Lagrangian Duality

Convex Optimization. Convex Analysis - Functions

Winter Lecture 10. Convexity and Concavity

The Erwin Schrodinger International Boltzmanngasse 9. Institute for Mathematical Physics A-1090 Wien, Austria

Handout 2: Elements of Convex Analysis

Probability and Random Processes

1. Sets A set is any collection of elements. Examples: - the set of even numbers between zero and the set of colors on the national flag.

3.1 Convexity notions for functions and basic properties

Springer-Verlag Berlin Heidelberg

ON THE ARITHMETIC-GEOMETRIC MEAN INEQUALITY AND ITS RELATIONSHIP TO LINEAR PROGRAMMING, BAHMAN KALANTARI

Math 172 Problem Set 5 Solutions

1. Introduction. mathematical optimization. least-squares and linear programming. convex optimization. example. course goals and topics

1. Introduction. mathematical optimization. least-squares and linear programming. convex optimization. example. course goals and topics

1. Introduction. mathematical optimization. least-squares and linear programming. convex optimization. example. course goals and topics

A notion of Total Dual Integrality for Convex, Semidefinite and Extended Formulations

3 (Due ). Let A X consist of points (x, y) such that either x or y is a rational number. Is A measurable? What is its Lebesgue measure?

C.7. Numerical series. Pag. 147 Proof of the converging criteria for series. Theorem 5.29 (Comparison test) Let a k and b k be positive-term series

M17 MAT25-21 HOMEWORK 6

Static Problem Set 2 Solutions

A : k n. Usually k > n otherwise easily the minimum is zero. Analytical solution:

Entrance Exam, Real Analysis September 1, 2017 Solve exactly 6 out of the 8 problems

6.2 Fubini s Theorem. (µ ν)(c) = f C (x) dµ(x). (6.2) Proof. Note that (X Y, A B, µ ν) must be σ-finite as well, so that.

Optimization Methods. Lecture 18: Optimality Conditions and. Gradient Methods. for Unconstrained Optimization

Bellman function approach to the sharp constants in uniform convexity

Lecture: Duality of LP, SOCP and SDP

Duality Uses and Correspondences. Ryan Tibshirani Convex Optimization

In English, this means that if we travel on a straight line between any two points in C, then we never leave C.

Chapter 13. Convex and Concave. Josef Leydold Mathematical Methods WS 2018/19 13 Convex and Concave 1 / 44

Transcription:

ELEC-E5420 Conve functions 2{1 Lecture 2 Conve functions conve functions, epigraph simple eamples, elementary properties more eamples, more properties Jensen's inequality quasiconve, quasiconcave functions log-conve and log-concave functions K-conveity

Conve functions 2{2 Conve functions f : R n! R is conve if dom f is conve and y 2 dom f 2 [0 1] + f ( + (1 ; )y) f () + (1 ; )f (y) (1) f is concave if ;f is conve conve concave neither `Modern' denition: f : R n! R [ f+1g (but not identically +1) f is conve if (1) holds as an inequality in R [ f+1g

Conve functions 2{3 Epigraph & sublevel sets The epigraph of the function f is epi f = f( t) j 2 dom f f () t g : f () epi f f conve function, epi f conve set The (-)sublevel set of f is C() = f 2 dom f j f () g : f conve ) sublevel sets are conve (converse false)

Conve functions 2{4 Dierentiable conve functions f dierentiable and conve () 8 0 : f () f ( 0 ) + rf ( 0 ) T ( ; 0 ) f () 0 f ( 0 )+rf ( 0 ) T ( ; 0 ) Interpretation 1st order Taylor appr. is a global lower bound on f supporting hyperplane to epi f: ( t) 2 epi f =) f () 2 6 rf ( 0) 4 ;1 3 7 5 T 2 6 64 ; 0 t ; f ( 0 ) 3 7 5 0 2 4 rf ( 0) ;1 3 5 f twice dierentiable and conve () r 2 f () 0

Conve functions 2{5 Simple eamples linear and ane functions: f () = a T + b conve quadratic functions: f () = T P+ 2q T + r with P = P T 0 any norm Eamples on R is conve on R + for 1, 0 concave for 0 1 log is concave, log is conve on R + e is conve jj, ma(0 ), ma(0 ;) are conve Z ;1 e;t2 dt is concave log

Conve functions 2{6 Elementary properties a function is conve i it is conve on all lines: f conve () f ( 0 + th) conve in t for all 0 h positive multiple of conve function is conve: f conve 0 =) f conve sum of conve functions is conve: f 1 f 2 conve =) f 1 + f 2 conve etends to innite sums, integrals: g( y) conve in =) Z g( y)dy conve

Conve functions 2{7 pointwise maimum: f 1 f 2 conve =) maff 1 () f 2 ()g conve (corresponds to intersection of epigraphs) f 2 () epi maff 1 f 2 g f 1 () pointwise supremum: f conve =) sup f conve 2A ane transformation of domain f conve ) f (A + b) conve

Conve functions 2{8 More eamples piecewise-linear functions: f () = mafa T i i + b ig is conve in (epi f is polyhedron) ma distance to any set, sup s2s k ; sk, is conve in f () = [1] + [2] + [3] is conve on R n ( [i] is the ith largest j ) f () = 0 B @ Y i 1 C i A 1=n is concave on R n + f () = m X i=1 log(b i ; a T i );1 is conve on P = f j a T i < b i i = 1 ::: mg least-squares cost as functions of weights, is concave in w f (w) = inf X i w i (a T i ; b i) 2

Conve functions 2{9 Conve functions of matrices Tr X is linear in X more generally, Tr A T X = X i j A ij X ij = vec(a) T vec(x) log det X ;1 is conve on X = X T 0 Proof: let i be the eigenvalues of X ;1=2 0 HX ;1=2 0 f (t) = log det(x 0 + th) ;1 = log det X ;1 0 + log det(i + tx ;1=2 0 HX ;1=2 0 ) ;1 = log det X ;1 0 ; X i is a conve function of t log(1 + t i ) (det X) 1=n is concave on X = X T 0, X 2 R nn ma (X) is conve on X = X T Proof: ma (X) = sup y T Xy kyk=1 kxk = ma (X T 1=2 X) is conve on R nm Proof: kxk = sup u T Xv kuk=1 kvk=1

Conve functions 2 {10 Minimizing over some variables If h( y) is conve in and y, then is conve in f () = inf y h( y) corresponds to projection of epigraph, ( y t)! ( t) h( y) f () y

Conve functions 2 {11 Eample. If S R n is conve then (min) distance to S, is conve in dist( S) = inf k ; sk s2s Eample. If g() is conve, then is conve in y. f (y) = inffg() j A = yg Proof: nd B, C s.t. f j A = yg = fby + Cz j z 2 R k g so f (y) = inf z g(by + Cz) `Modern' proof: f (y) = inf z g() + h(a ; y) where is conve h(z) = 8 >< >: 0 if z = 0 +1 otherwise

Conve functions 2 {12 Composition one-dimensional case f () = h(g()) is conve if g conve h conve, nondecreasing g concave h conve, nonincreasing Eamples f () = ep g() is conve if g is conve f () = 1=g() is conve if g is concave, positive f () = g() p, p 1, is conve if g() conve, positive f 1,..., f n conve, then f () = ; X i conve on f j f i () < 0 i = 1 ::: ng log(;f i ()) is Proof: (dierentiable functions, 2 R) f 00 = h 00 (g 0 ) 2 + g 00 h 0

Conve functions 2 {13 Composition k-dimensional case f () = h(g 1 () ::: g k ()) with h : R k! R, g i : R n! R is conve if h conve, nondecreasing in each arg. g i conve h conve, nonincreasing in each arg. g i concave etc. Eamples f () = ma i g i () is conve if each g i is f () = log X i ep g i () is conve if each g i is Proof: (dierentiable functions, n = 1) f 00 = rh T 2 6 4 g 00 1. g 00 k 3 2 g 0 3 1 7 + 5 6. 4 g 0 7 5 k T r 2 h 2 6 4 g 0 3 1. g 0 7 5 k

Conve functions 2 {14 Jensen's inequality f : R n! R conve two points 2 [0 1] + f ( 1 + (1 ; ) 2 ) f ( 1 ) + (1 ; )f ( 2 ) more than two points f ( X i i 0 X i = 1 + i i i ) X i i f ( i ) continuous version most general form: p() 0 Z p()d = 1 + f (Z p()d) Z f ()p()d f (E ) E f () Interpretation: (zero mean) randomization, dithering increases average value of a conve function

Conve functions 2 {15 Applications Many (some people claim most) inequalities can be derived from Jensen's inequality Eample. Arithmetic-geometric mean inequality a b 0 ) p ab (a + b)=2 Proof. f () = log is concave on R + : 0 1 B (log a + log b) log a @ + b 1 C A 2 2

Conve functions 2 {16 Quasiconve functions f : C! R, C a conve set, is quasiconve if every sublevel set S = f j f () g is conve. y S can have `locally at' regions f () f is quasiconcave if ;f is quasiconve, i.e., superlevel sets f j f () g are conve. A function which is both quasiconve and quasiconcave is called quasilinear. f conve (concave) ) f quasiconve (quasiconcave)

Conve functions 2 {17 Eamples f () = r jj is quasiconve on R f () = log is quasilinear on R + linear fractional function, f () = at + b c T + d is quasilinear on the halfspace c T + d > 0 k ; ak k ; bk f j k ; ak k ; bkg f () = is quasiconve on the halfspace f (a) = degree(a 0 + a 1 t + + a k t k ) on R k+1

Conve functions 2 {18 Properties f is quasiconve if and only if it is quasiconve on lines, i.e., f ( 0 + th) quasiconve in t for all 0 h. modied Jensen's inequality: f : C! R quasiconve if and only if y 2 C 2 [0 1] + f ( + (1 ; )y) maff () f(y)g f () y

Conve functions 2 {19 for f dierentiable, f quasiconve if and only if for all, y f (y) f () ) (y ; ) T rf () 0 S 1 rf () S 2 S 3 1 < 2 < 3 positive multiples f quasiconve 0 =) f quasiconve pointwise maimum f 1 f 2 quasiconve =) maff 1 f 2 g quasiconve (etends to supremum over arbitrary set) ane transformation of domain f quasiconve =) f (A + b) quasiconve

Conve functions 2 {20 projective transformation of domain 0 B f quasiconve =) f A @ + b c T + d on c T + d > 0 1 C A quasiconve composition with monotone increasing function f quasiconve g monotone increasing =) g(f ()) quasiconve sums of quasiconve functions are not quasiconve in general f quasiconve in, y =) g() = inf y quasiconve in f ( y)

Conve functions 2 {21 Nested sets characterization f quasiconve ) sublevel sets S conve, nested, i.e., 1 2 ) S 1 S 2 converse: if T is a nested family of conve sets, then is quasiconve. f () = inff j 2 T g Engineering interpretation: T are specs, tighter for smaller

Conve functions 2 {25 Log-concave functions f : R n! R + is log-concave (log-conve) if log f is concave (conve) Log-conve ) conve concave ) log-concave `Modern' denition allows log-concave f to take on value zero, so log f takes on value ;1 Eamples normal density, f () = e ;(1=2)(; 0) T ;1 (; 0 ) erfc, f () = p 2 Z 1 e;t2 dt indicator function of conve set C: I C () = 8 >< >: 1 2 C 0 62 C

Conve functions 2 {26 Properties sum of log-concave functions not always log-concave (but sum of log-conve functions is log-conve) products (immediate) f g log-concave =) fg log-concave integrals f ( y) log-concave in y =) Z f ( y)dy log-concave convolutions f g log-concave =) Z f ( ; y)g(y)dy log-concave (immediate from the properties above)

Conve functions 2 {27 Log-concave probability densities Many common probability density functions are log-concave. Eamples normal ( 0) f () = 1 r(2) n det e;1 2 (;) T ;1 (;) eponential( i > 0) on R n + f () = 0 B @ n Y i=1 i 1 C A e ;( 1 1 ++ n n ) uniform distribution on conve (bounded) set C f () = 8 >< >: 1= 2 C 0 62 C where is Lebesgue measure of C (i.e., length, area, volume...)

Conve functions 2 {31 K-conveity conve cone K R m induces generalized inequality K f : R n! R m is K-conve if 0 1 =) f ( + (1 ; )y) K f () + (1 ; )f (y) Eample. K is PSD cone (called matri conveity) let's show that f (X) = X 2 is K-conve on fxjx = X T g, i.e., for 2 [0 1], (X + (1 ; )Y ) 2 X 2 + (1 ; )Y 2 (1) for any u 2 R m, u T X 2 u = kxuk 2 is a (quadratic) conve fct of X, so u T (X + (1 ; )Y ) 2 u u T X 2 u + (1 ; )u T Y 2 u which implies (1)