LECTURE SLIDES ON CONVEX OPTIMIZATION AND DUALITY THEORY TATA INSTITUTE FOR FUNDAMENTAL RESEARCH MUMBAI, INDIA JANUARY 2009 PART I

Transcription

1 LECTURE SLIDES ON CONVEX OPTIMIZATION AND DUALITY THEORY TATA INSTITUTE FOR FUNDAMENTAL RESEARCH MUMBAI, INDIA JANUARY 2009 PART I BY DIMITRI P. BERTSEKAS M.I.T.

2 LECTURE 1 NTRODUCTION/BASIC CONVEXITY CONCEPTS LECTURE OUTLINE Convex Optimization Problems Why is Convexity Important in Optimization Multipliers and Lagrangian Duality Min Common/Max Crossing Duality Convex sets and functions Epigraphs Closed convex functions Recognizing convex functions

3 OPTIMIZATION PROBLEMS Generic form: minimize f(x) subject to x C Cost function f : R n R, constraint set C, e.g., C = X { x h 1 (x) = 0,...,h m (x) = 0 } { x g 1 (x) 0,...,g r (x) 0 } Examples of problem classifications: Continuous vs discrete Linear vs nonlinear Deterministic vs stochastic Static vs dynamic Convex programming problems are those for which f is convex and C is convex (they are continuous problems). However, convexity permeates all of optimization, including discrete problems.

4 WHY IS CONVEXITY SO SPECIAL? A convex function has no local minima that are not global A convex set has a nonempty relative interior A convex set is connected and has feasible directions at any point A nonconvex function can be convexified while maintaining the optimality of its global minima The existence of a global minimum of a convex function over a convex set is conveniently characterized in terms of directions of recession A polyhedral convex set is characterized in terms of a finite set of extreme points and extreme directions A real-valued convex function is continuous and has nice differentiability properties Closed convex cones are self-dual with respect to polarity Convex, lower semicontinuous functions are selfdual with respect to conjugacy

5 CONVEXITY AND DUALITY Consider the (primal) problem minimize f(x) s.t. g 1 (x) 0,...,g r (x) 0 We introduce multiplier vectors µ = (µ 1,...,µ r ) 0 and form the Lagrangian function L(x,µ) = f(x)+ r µ j g j (x), x R n, µ R r. j=1 Dual function q(µ) = inf x R n L(x,µ) Dual problem: Maximize q(µ) over µ 0 Motivation: Under favorable circumstances (strong duality) the optimal values of the primal and dual problems are equal, and their optimal solutions are related

6 KEY DUALITY RELATIONS Optimal primal value f = inf f(x) = inf sup L(x, µ) g j (x) 0, j=1,...,r x R n µ 0 Optimal dual value q = sup q(µ) = sup inf L(x,µ) µ 0 µ 0 x R n We always have q f (weak duality - important in discrete optimization problems). Under favorable circumstances (convexity in the primal problem, plus...): We have q = f (strong duality) If µ is optimal dual solution, all optimal primal solutions minimize L(x,µ ) This opens a wealth of analytical and computational possibilities, and insightful interpretations. Note that the equality of sup inf and inf sup is a key issue in minimax theory and game theory.

7 MIN COMMON/MAX CROSSING DUALITY w Min Common Point w Min Common Point w * M _ M w Min Common Point w Min Common Point w * M M M M 0 u Max Crossing Point q * 0 Max Crossing Max Max Crossing Crossing Point q * Point q (a) (a) Point q (b) (b) u Min Common Point w Min Common Point w * Max Crossing Point q w Max Crossing Point q * S _ M M M M 0 u u (c) All of duality theory and all of (convex/concave) minimax theory can be developed/explained in terms of this one figure. This is the novel aspect of the treatment (although the ideas are closely connected to conjugate convex function theory) The machinery of convex analysis is needed to flesh out this figure, and to rule out the exceptional/pathological behavior shown in (c).

8 EXCEPTIONAL BEHAVIOR If convex structure is so favorable, what is the source of exceptional/pathological behavior [like in (c) of the preceding slide]? Answer: Some common operations on convex sets do not preserve some basic properties. Example: A linearly transformed closed convex set need not be closed (contrary to compact and polyhedral sets). x 2 C 1 = { (x 1, x 2 ) x 1 > 0, x 2 > 0, x 1 x 2 1 } C 2 = { (x 1, x 2 ) x 1 = 0 }, x 1 This is a major reason for the analytical difficulties in convex analysis and pathological behavior in convex optimization (and the favorable character of polyhedral sets).

9 COURSE OUTLINE 1) Basic Convexity Concepts (2): Convex sets and functions. Convex and affine hulls. Closure, relative interior, and continuity. 2) More Convexity Concepts (2): Directions of recession. Hyperplanes. Conjugate convex functions. 3) Convex Optimization Concepts (1): Existence of optimal solutions. Partial minimization. Saddle point and minimax theory. 4) Min common/max crossing duality (1): MC/MC duality. Special cases in constrained minimization and minimax. Strong duality theorem. Existence of dual optimal solutions. 5) Duality applications (2): Constrained optimization (Lagrangian, Fenchel, and conic duality). Subdifferential theory and optimality conditions. Minimax theorems. Nonconvex problems and estimates of the duality gap.

10 WHAT TO EXPECT FROM THIS COURSE We aim: To develop insight and deep understanding of a fundamental optimization topic To treat rigorously an important branch of applied math, and to provide some appreciation of the research in the field Mathematical level: Prerequisites are linear algebra (preferably abstract) and real analysis (a course in each) Proofs are important... but the rich geometry helps guide the mathematics We will make maximum use of visualization and figures Applications: They are many and pervasive... but don t expect much in this course. The book by Boyd and Vandenberghe describes a lot of practical convex optimization models ( boyd/cvxbook.html) Handouts: Slides, 1st chapter, material in

11 A NOTE ON THESE SLIDES These slides are a teaching aid, not a text Don t expect strict mathematical rigor The statements of theorems are fairly precise, but the proofs are not Many proofs have been omitted or greatly abbreviated Figures are meant to convey and enhance ideas, not to express them precisely The omitted proofs and a much fuller discussion can be found in the Convex Optimization textbook and handouts

12 SOME MATH CONVENTIONS All of our work is done in R n : space of n-tuples x = (x 1,...,x n ) All vectors are assumed column vectors denotes transpose, so we use x to denote a row vector x y is the inner product n i=1 x iy i of vectors x and y x = x x is the (Euclidean) norm of x. We use this norm almost exclusively See the appendix for an overview of the linear algebra and real analysis background that we will use

13 CONVEX SETS αx + (1 α)y, 0 α 1 x y x y x x y y A subset C of R n is called convex if αx + (1 α)y C, x,y C, α [0,1] Operations that preserve convexity Intersection, scalar multiplication, vector sum, closure, interior, linear transformations Cones: Sets C such that λx C for all λ > 0 and x C (not always convex or closed)

14 REAL-VALUED CONVEX FUNCTIONS f(y) f(x) αf(x) af(x) + (1+ (1 - α)f(y) a)f(y) f ( αx f(ax + (1 + (1 - α)y a)y) ) x αx ax + (1- a)y α)y C y Let C be a convex subset of R n. A function f : C R is called convex if f ( αx + (1 α)y ) αf(x) + (1 α)f(y) for all x,y C, and α [0,1]. If f is a convex function, then all its level sets {x C f(x) a} and {x C f(x) < a}, where a is a scalar, are convex.

15 EXTENDED REAL-VALUED FUNCTIONS The epigraph of a function f : X [, ] is the subset of R n+1 given by epi(f) = { (x,w) x X, w R, f(x) w } The effective domain of f is the set dom(f) = { x X f(x) < } We say that f is proper if f(x) < for at least one x X and f(x) > for all x X, and we will call f improper if it is not proper. Note that f is proper if and only if its epigraph is nonempty and does not contain a vertical line. An extended real-valued function f : X [, ] is called lower semicontinuous at a vector x X if f(x) liminf k f(x k ) for every sequence {x k } X with x k x. We say that f is closed if epi(f) is a closed set.

16 CLOSEDNESS AND SEMICONTINUITY Proposition: For a function f : R n [, ], the following are equivalent: (i) {x f(x) a} is closed for every scalar a. (ii) f is lower semicontinuous at all x R n. (iii) f is closed. f(x) epi(f) γ { x f(x) γ } x Note that: If f is lower semicontinuous at all x dom(f), it is not necessarily closed If f is closed, dom(f) is not necessarily closed Proposition: Let f : X [, ] be a function. If dom(f) is closed and f is lower semicontinuous at all x dom(f), then f is closed.

17 XTENDED REAL-VALUED CONVEX FUNCTIONS f(x) Epigraph f(x) Epigraph dom(f) Convex function Convex function x dom(f) Nonconvex function Nonconvex function x x Let C be a convex subset of R n. An extended real-valued function f : C [, ] is called convex if epi(f) is a convex subset of R n+1. If f is proper, this definition is equivalent to f ( αx + (1 α)y ) αf(x) + (1 α)f(y) for all x,y C, and α [0,1]. An improper closed convex function is very peculiar: it takes an infinite value ( or ) at every point.

18 RECOGNIZING CONVEX FUNCTIONS Some important classes of elementary convex functions: Affine functions, positive semidefinite quadratic functions, norm functions, etc. Proposition: Let f i : R n (, ], i I, be given functions (I is an arbitrary index set). (a) The function g : R n (, ] given by g(x) = λ 1 f 1 (x) + + λ m f m (x), λ i > 0 is convex (or closed) if f 1,...,f m are convex (respectively, closed). (b) The function g : R n (, ] given by g(x) = f(ax) where A is an m n matrix is convex (or closed) if f is convex (respectively, closed). (c) The function g : R n (, ] given by g(x) = sup i I f i (x) is convex (or closed) if the f i are convex (respectively, closed).

19 LECTURE 2 LECTURE OUTLINE Differentiable Convex Functions Convex and Affine Hulls Caratheodory s Theorem Closure, Relative Interior, Continuity

20 DIFFERENTIABLE CONVEX FUNCTIONS f(z) f(x) + (z x) f(x) x z Let C R n be a convex set and let f : R n R be differentiable over R n. (a) The function f is convex over C iff f(z) f(x) + (z x) f(x), x,z C Implies that x minimizes f over C iff f(x ) (x x ) 0, x C (b) If the inequality is strict whenever x z, then f is strictly convex over C, i.e., for all α (0,1) and x,y C, with x y f ( αx + (1 α)y ) < αf(x) + (1 α)f(y)

21 PROOF IDEAS αf(x) + (1 α)f(y) f(y) f(x) f(z) + (x z) f(z) f(z) f(z) + (y z) f(z) x z = αx + (1 α)y (a) y f(z) f(x) + f( x + α(z x) ) f(x) α f(x) + (z x) f(x) x x + α(z x) z (b)

22 TWICE DIFFERENTIABLE CONVEX FUNCTIONS Let C be a convex subset of R n and let f : R n R be twice continuously differentiable over R n. (a) If 2 f(x) is positive semidefinite for all x C, then f is convex over C. (b) If 2 f(x) is positive definite for all x C, then f is strictly convex over C. (c) If C is open and f is convex over C, then 2 f(x) is positive semidefinite for all x C. Proof: (a) By mean value theorem, for x,y C f(y) = f(x)+(y x) f(x)+ 1 2 (y x) 2 f ( x+α(y x) ) (y x) for some α [0,1]. Using the positive semidefiniteness of 2 f, we obtain f(y) f(x) + (y x) f(x), x,y C From the preceding result, f is convex. (b) Similar to (a), we have f(y) > f(x) + (y x) f(x) for all x,y C with x y, and we use the preceding result.

23 CONVEX AND AFFINE HULLS Given a set X R n : A convex combination of elements of X is a vector of the form m i=1 α ix i, where x i X, α i 0, and m i=1 α i = 1. The convex hull of X, denoted conv(x), is the intersection of all convex sets containing X (also the set of all convex combinations from X). The affine hull of X, denoted aff(x), is the intersection of all affine sets containing X (an affine set is a set of the form x + S, where S is a subspace). Note that aff(x) is itself an affine set. A nonnegative combination of elements of X is a vector of the form m i=1 α ix i, where x i X and α i 0 for all i. The cone generated by X, denoted cone(x), is the set of all nonnegative combinations from X: It is a convex cone containing the origin. It need not be closed (even if X is compact). If X is a finite set, cone(x) is closed (nontrivial to show!)

24 CARATHEODORY S THEOREM x 4 x 2 x 1 X x x x 2 cone(x) x 1 x conv(x) x 0 (a) (b) x 3 Let X be a nonempty subset of R n. (a) Every x 0 in cone(x) can be represented as a positive combination of vectors x 1,...,x m from X that are linearly independent (so m n). (b) Every x / X that belongs to conv(x) can be represented as a convex combination of at most n + 1 vectors.

25 PROOF OF CARATHEODORY S THEOREM (a) Let x be a nonzero vector in cone(x), and let m be the smallest integer such that x has the form m i=1 α ix i, where α i > 0 and x i X for all i = 1,...,m. If the vectors x i were linearly dependent, there would exist λ 1,...,λ m, with m λ i x i = 0 i=1 and at least one of the λ i is positive. Consider m (α i γλ i )x i, i=1 where γ is the largest γ such that α i γλ i 0 for all i. This combination provides a representation of x as a positive combination of fewer than m vectors of X a contradiction. Therefore, x 1,...,x m, are linearly independent. (b) Apply part (a) to the subset of R n+1 Y = { (z,1) z X } consider cone(y ), represent (x,1) cone(y ) in terms of at most n + 1 vectors, etc.

26 AN APPLICATION OF CARATHEODORY The convex hull of a compact set is compact. Proof: Let X be compact. We take a sequence in conv(x) and show that it has a convergent subsequence whose limit is in conv(x). By Caratheodory, { a sequence in conv(x) can n+1 } be expressed as, where for all k and i=1 αk i xk i i, αi k 0, xk i X, and n+1 i=1 αk i = 1. Since { (α k 1,...,αn+1 k,xk 1,...,xk n+1 )} is bounded, it has a limit point { (α1,...,α n+1,x 1,...,x n+1 ) }, which must satisfy n+1 i=1 α i = 1, and α i 0, x i X for all i. Thus, the vector n+1 i=1 α ix i, which belongs { to conv(x), is a limit point of the n+1 } sequence, so conv(x) is compact. Q.E.D. i=1 αk i xk i Note the convex hull of a closed set need not be closed.

27 RELATIVE INTERIOR x is a relative interior point of C, if x is an interior point of C relative to aff(c). ri(c) denotes the relative interior of C, i.e., the set of all relative interior points of C. Line Segment Principle: If C is a convex set, x ri(c) and x cl(c), then all points on the line segment connecting x and x, except possibly x, belong to ri(c). C x α = αx+(1 α)x S x ǫ S α αǫ x

28 ADDITIONAL MAJOR RESULTS Let C be a nonempty convex set. (a) ri(c) is a nonempty convex set, and has the same affine hull as C. (b) x ri(c) if and only if every line segment in C having x as one endpoint can be prolonged beyond x without leaving C. X z 2 0 C z 1 Proof: (a) Assume that 0 C. We choose m linearly independent vectors z 1,...,z m C, where m is the dimension of aff(c), and we let { m m } X = α i z i α i < 1, α i > 0, i = 1,...,m i=1 i=1 Then argue that X ri(c). (b) => is clear by the def. of rel. interior. Reverse: argue by contradiction; take any x ri(c); use prolongation assumption and Line Segment Princ.

29 OPTIMIZATION APPLICATION A concave function f : R n R that attains its minimum over a convex set X at an x ri(x) must be constant over X. x x* X x aff(x) Proof: (By contradiction.) Let x X be such that f(x) > f(x ). Prolong beyond x the line segment x-to-x to a point x X. By concavity of f, we have for some α (0,1) f(x ) αf(x) + (1 α)f(x), and since f(x) > f(x ), we must have f(x ) > f(x) - a contradiction. Q.E.D. Corollary: A linear function can attain a mininum only at the boundary of a convex set.

30 ALCULUS OF RELATIVE INTERIORS: SUMMARY The relative interior of a convex set is equal to the relative interior of its closure. The closure of the relative interior of a convex set is equal to its closure. Relative interior and closure commute with Cartesian product and inverse image under a linear transformation. Relative interior commutes with image under a linear transformation and vector sum, but closure does not. Neither relative interior nor closure commute with set intersection.

31 CLOSURE VS RELATIVE INTERIOR Let C be a nonempty convex set. Then ri(c) and cl(c) are not too different for each other. Proposition: (a) We have cl(c) = cl ( ri(c) ). (b) We have ri(c) = ri ( cl(c) ). (c) Let C be another nonempty convex set. Then the following three conditions are equivalent: (i) C and C have the same rel. interior. (ii) C and C have the same closure. (iii) ri(c) C cl(c). Proof: (a) Since ri(c) C, we have cl ( ri(c) ) cl(c). Conversely, let x cl(c). Let x ri(c). By the Line Segment Principle, we have αx+(1 α)x ri(c) for all α (0,1]. Thus, x is the limit of a sequence that lies in ri(c), so x cl ( ri(c) ). x x C

32 LINEAR TRANSFORMATIONS Let C be a nonempty convex subset of R n and let A be an m n matrix. (a) We have A ri(c) = ri(a C). (b) We have A cl(c) cl(a C). Furthermore, if C is bounded, then A cl(c) = cl(a C). Proof: (a) Intuition: Spheres within C are mapped onto spheres within A C (relative to the affine hull). (b) We have A cl(c) cl(a C), since if a sequence {x k } C converges to some x cl(c) then the sequence {Ax k }, which belongs to A C, converges to Ax, implying that Ax cl(a C). To show the converse, assuming that C is bounded, choose any z cl(a C). Then, there exists a sequence {x k } C such that Ax k z. Since C is bounded, {x k } has a subsequence that converges to some x cl(c), and we must have Ax = z. It follows that z A cl(c). Q.E.D. Note that in general, we may have A int(c) int(a C), A cl(c) cl(a C)

33 INTERSECTIONS AND VECTOR SUMS Let C 1 and C 2 be nonempty convex sets. (a) We have ri(c 1 + C 2 ) = ri(c 1 ) + ri(c 2 ), cl(c 1 ) + cl(c 2 ) cl(c 1 + C 2 ) If one of C 1 and C 2 is bounded, then cl(c 1 ) + cl(c 2 ) = cl(c 1 + C 2 ) (b) If ri(c 1 ) ri(c 2 ) Ø, then ri(c 1 C 2 ) = ri(c 1 ) ri(c 2 ), cl(c 1 C 2 ) = cl(c 1 ) cl(c 2 ) Proof of (a): C 1 + C 2 is the result of the linear transformation (x 1,x 2 ) x 1 + x 2. Counterexample for (b): C 1 = {x x 0}, C 2 = {x x 0}

34 CONTINUITY OF CONVEX FUNCTIONS If f : R n R is convex, then it is continuous. y k e 3 e 2 x k x k+1 0 e 4 zk e 1 Proof: We will show that f is continuous at 0. By convexity, f is bounded within the unit cube by the maximum value of f over the corners of the cube. Consider sequence x k 0 and the sequences y k = x k / x k, z k = x k / x k. Then f(x k ) ( 1 x k ) f(0) + xk f(y k ) f(0) x k x k + 1 f(z k) + 1 x k + 1 f(x k) Since x k 0, f(x k ) f(0). Q.E.D. Extension to continuity over ri(dom(f)).

35 CLOSURES OF FUNCTIONS The closure of a function f : X [, ] is the function clf : R n [, ] with epi(clf) = cl ( epi(f) ) The convex closure of f is the function člf with epi(čl f) = cl( conv ( epi(f) )) Proposition: For any f : X [, ] inf x X f(x) = inf x Rn(clf)(x) = inf x R n(čl f)(x). Also, any vector that attains the infimum of f over X also attains the infimum of clf and čl f. Proposition: For any f : X [, ]: (a) cl f (člf) is the greatest closed (closed convex, resp.) function majorized by f. (b) If f is convex, then clf is convex, and it is proper if and only if f is proper. Also, (clf)(x) = f(x), x ri ( dom(f) ), and if x ri ( dom(f) ) and y dom(clf), (clf)(y) = lim α 0 f ( y + α(x y) ).

36 LECTURE 3 Recession cones LECTURE OUTLINE Directions of recession of convex functions Nonemptiness of closed set intersections Linear and Quadratic Programming Preservation of closure under linear transformation

37 RECESSION CONE OF A CONVEX SET Given a nonempty convex set C, a vector d is a direction of recession if starting at any x in C and going indefinitely along d, we never cross the relative boundary of C to points outside C: x + αd C, x C, α 0 Recession Cone R C C d x + αd 0 x Recession cone of C (denoted by R C ): The set of all directions of recession. R C is a cone containing the origin.

38 RECESSION CONE THEOREM Let C be a nonempty closed convex set. (a) The recession cone R C is a closed convex cone. (b) A vector d belongs to R C if and only if there exists a vector x C such that x + αd C for all α 0. (c) R C contains a nonzero direction if and only if C is unbounded. (d) The recession cones of C and ri(c) are equal. (e) If D is another closed convex set such that C D Ø, we have R C D = R C R D More generally, for any collection of closed convex sets C i, i I, where I is an arbitrary index set and i I C i is nonempty, we have R i I C i = i I R Ci

39 PROOF OF PART (B) C z 3 z 2 z 1 = x + d x x + d 1 x + d 2 x + d 3 x + d x Let d 0 be such that there exists a vector x C with x + αd C for all α 0. We fix x C and α > 0, and we show that x + αd C. By scaling d, it is enough to show that x + d C. Let z k = x + kd for k = 1,2,..., and d k = (z k x) d / z k x. We have d k d = z k x z k x d d + x x z k x, z k x z k x 1, x x z k x 0, so d k d and x + d k x + d. Use the convexity and closedness of C to conclude that x + d C.

40 LINEALITY SPACE The lineality space of a convex set C, denoted by L C, is the subspace of vectors d such that d R C and d R C : L C = R C ( R C ) If d L C, the entire line defined by d is contained in C, starting at any point of C. Decomposition of a Convex Set: Let C be a nonempty convex subset of R n. Then, C = L C + (C L C ). True also if L C is replaced by a subset S L C. S C z x C S 0 d S

41 DIRECTIONS OF RECESSION OF A FUNCTION Some basic geometric observations: The horizontal directions in the recession cone of the epigraph of a convex function f are directions along which the level sets are unbounded. Along these directions the level sets { x f(x) γ } are unbounded and f is monotonically nondecreasing. These are the directions of recession of f. epi(f) γ Slice {(x,γ) f(x) γ} 0 Recession Cone of f Level Set Vγ = {x f(x) γ}

42 RECESSION CONE OF LEVEL SETS Proposition: Let f : R n (, ] be a closed proper convex function and consider the level sets V γ = { x f(x) γ }, where γ is a scalar. Then: (a) All the nonempty level sets V γ have the same recession cone, given by R Vγ = { d (d,0) R epi(f) } (b) If one nonempty level set V γ is compact, then all nonempty level sets are compact. Proof: For each fixed γ for which V γ is nonempty, { (x,γ) x Vγ } = epi(f) { (x,γ) x R n } The recession cone of the set on the left is { (d,0) d R Vγ }. The recession cone of the set on the right is the intersection of R epi (f) and the recession cone of { (x,γ) x R n}. Thus we have { (d,0) d RVγ } = { (d,0) (d,0) Repi(f) }, from which the result follows.

43 RECESSION CONE OF A CONVEX FUNCTION For a closed proper convex function f : R n (, ], the (common) recession cone of the nonempty level sets V γ = { x f(x) γ }, γ R, is the recession cone of f, and is denoted by R f. Recession Cone R f 0 Level Sets of f Terminology: d R f : a direction of recession of f. L f = R f ( R f ): the lineality space of f. d L f : a direction of constancy of f. Example: For the pos. semidefinite quadratic f(x) = x Qx + a x + b, the recession cone and constancy space are R f = {d Qd = 0, a d 0}, L f = {d Qd = 0, a d = 0}

44 RECESSION FUNCTION Function r f : R n (, ] whose epigraph is R epi(f) : the recession function of f. Characterizes the recession cone: R f = { d r f (d) 0 }, L f = { d r f (d) = r f ( d) = 0 } Can be shown that r f (d) = sup α>0 f(x + αd) f(x) α = lim α f(x + αd) f(x) α Thus r f (d) is the asymptotic slope of f in the direction d. In fact, r f (d) = lim α f(x + αd) d, x,d R n if f is differentiable. Calculus of recession functions: r f1 + +f m (d) = r f1 (d) + + r fm (d) r supi I f i (d) = sup i I r fi (d)

45 DESCENT BEHAVIOR OF A CONVEX FUNCTION f(x + ay) f(x + αd) f(x + ay) f(x + αd) f(x) f(x) r f (d) = 0 f(x) f(x) r f (d) < 0 α a α a (a) (b) f(x + ay) f(x + αd) f(x + ay) f(x + αd) f(x) f(x) r f (d) = 0 f(x) f(x) r f (d) = 0 αa α a (c) (d) f(x + ay) f(x + αd) f(x + ay) f(x + αd) f(x) f(x) r f (d) > 0 f(x) r f (d) > 0 f(x) α a α a (e) (f) y is a direction of recession in (a)-(d). This behavior is independent of the starting point x, as long as x dom(f).

46 THE ROLE OF CLOSED SET INTERSECTIONS A fundamental question: Given a sequence of nonempty closed sets {C k } in R n with C k+1 C k for all k, when is k=0 C k nonempty? Set intersection theorems are significant in at least three major contexts, which we will discuss in what follows: 1. Does a function f : R n (, ] attain a minimum over a set X? This is true iff the intersection of the nonempty level sets { x X f(x) γ k } is nonempty. 2. If C is closed and A is a matrix, is AC closed? Special case: If C 1 and C 2 are closed, is C 1 + C 2 closed? 3. If F(x,z) is closed, is f(x) = inf z F(x,z) closed? (Critical question in duality theory.) Can be addressed by using the relation P ( epi(f) ) epi(f) cl (P ( epi(f) )) where P( ) is projection on the space of (x,w).

47 ASYMPTOTIC DIRECTIONS Given nested sequence {C k } of closed convex sets, {x k } is an asymptotic sequence if x k C k, x k 0, k = 0,1,... x k, x k x k d d where d is a nonzero common direction of recession of the sets C k. As a special case we define asymptotic sequence of a closed convex set C (use C k C). Every unbounded {x k } with x k C k has an asymptotic subsequence. {x k } is called retractive if for some k, we have x k d C k, k k. x 1 x 2 x 3 x 0 x 4 x 5 Asymptotic Sequence 0 d Asymptotic Direction

48 RETRACTIVE SEQUENCES A nested sequence {C k } of closed convex sets is retractive if all its asymptotic sequences are retractive. Intersections and Cartesian products of retractive set sequences are retractive. A closed halfspace (viewed as a sequence with identical components) is retractive. A polyhedral set is retractive. Also the vector sum of a convex compact set and a retractive convex set is retractive. Nonpolyhedral cones and level sets of quadratic functions need not be retractive. Intersection k=0 C k Intersection Intersection k=0 C k Intersection SC 1 CS 0 d x 3 x 3 S 2 C 2 SC 1 x 2 C S 0 x 2 d x x S 1 C x 2 0 x 0 x 0 0 (a) Retractive (a) Retractive Set Sequence (b) Nonretractive (b) Nonretractive Set Sequence 1 1

49 SET INTERSECTION THEOREM I Proposition: If {C k } is retractive, then k=0 C k is nonempty. Key proof ideas: (a) The intersection k=0 C k is empty iff the sequence {x k } of minimum norm vectors of C k is unbounded (so a subsequence is asymptotic). (b) An asymptotic sequence {x k } of minimum norm vectors cannot be retractive, because such a sequence eventually gets closer to 0 when shifted opposite to the asymptotic direction. x 1 x 2 x 3 x 0 x 4 x 5 Asymptotic Sequence 0 d Asymptotic Direction

50 SET INTERSECTION THEOREM II Proposition: Let {C k } be a nested sequence of nonempty closed convex sets, and X be a retractive set such that all the sets C k = X C k are nonempty. Assume that where R X R L, R = k=0 R C k, L = k=0 L C k Then {C k } is retractive and k=0 C k is nonempty. Special case: X = R n, R = L. Proof: The set of common directions of recession of C k is R X R. For any asymptotic sequence {x k } corresponding to d R X R: (1) x k d C k (because d L) (2) x k d X (because X is retractive) So {C k } is retractive.

51 NEED TO ASSUME THAT X IS RETRACTIVE X X C k C k+1 C k C k+1 Consider k=0 C k, with C k = X C k The condition R X R L holds In the figure on the left, X is polyhedral. In the figure on the right, X is nonpolyhedral and nonretrative, and k=0 C k = Ø

52 LINEAR AND QUADRATIC PROGRAMMING Theorem: Let f(x) = x Qx+c x, X = {x a jx+b j 0, j = 1,..., r}, where Q is symmetric positive semidefinite. If the minimal value of f over X is finite, there exists a minimum of f over X. Proof: (Outline) Write with Set of Minima = X {x x Qx + c x γ k } γ k f = inf x X f(x). Verify the condition R X R L of the preceding set intersection theorem, where R and L are the sets of common recession and lineality directions of the sets Q.E.D. {x x Qx + c x γ k }

53 CLOSURE UNDER LINEAR TRANSFORMATIONS Let C be a nonempty closed convex, and let A be a matrix with nullspace N(A). (a) AC is closed if R C N(A) L C. (b) A(X C) is closed if X is a retractive set and R X R C N(A) L C, Proof: (Outline) Let {y k } AC with y k y. We prove k=0 C k Ø, where C k = C N k, and N k = {x Ax W k }, W k = { z z y y k y } N k x C k C y y k+1 y k AC Special Case: AX is closed if X is polyhedral.

54 NEED TO ASSUME THAT X IS RETRACTIVE C C N(A) N(A) X X A(X A(X C) C) A(X A(X C) C) Consider closedness of A(X C) In both examples the condition is satisfied. R X R C N(A) L C However, in the example on the right, X is not retractive, and the set A(X C) is not closed.

55 LECTURE 4 LECTURE OUTLINE Hyperplane separation Proper separation Nonvertical hyperplanes Convex conjugate functions Conjugacy theorem Examples

56 HYPERPLANES a Positive Halfspace {x a x b} Hyperplane {x a x = b} = {x a x = a x} Negative Halfspace {x a x b} x A hyperplane is a set of the form {x a x = b}, where a is nonzero vector in R n and b is a scalar. We say that two sets C 1 and C 2 are separated by a hyperplane H = {x a x = b} if each lies in a different closed halfspace associated with H, i.e., either a x 1 b a x 2, x 1 C 1, x 2 C 2, or a x 2 b a x 1, x 1 C 1, x 2 C 2 If x belongs to the closure of a set C, a hyperplane that separates C and the singleton set {x} is said be supporting C at x.

57 VISUALIZATION Separating and supporting hyperplanes: a a C 1 C 2 C x (a) (b) A separating {x a x = b} that is disjoint from C 1 and C 2 is called strictly separating: a x 1 < b < a x 2, x 1 C 1, x 2 C 2 C 1 C 2 C 1 C 2 x 1 x x 2 a (a) (b)

58 SUPPORTING HYPERPLANE THEOREM Let C be convex and let x be a vector that is not an interior point of C. Then, there exists a hyperplane that passes through x and contains C in one of its closed halfspaces. C ˆx 0 a 0 x 0 a x ˆx 2 ˆx 3 ˆx 1 a 3 a 1 a 2 x 1 x 3 x 2 Proof: Take a sequence {x k } that does not belong to cl(c) and converges to x. Let ˆx k be the projection of x k on cl(c). We have for all x cl(c) a k x a k x k, x cl(c), k = 0,1,..., where a k = (ˆx k x k )/ ˆx k x k. Let a be a limit point of {a k }, and take limit as k. Q.E.D.

59 SEPARATING HYPERPLANE THEOREM Let C 1 and C 2 be two nonempty convex subsets of R n. If C 1 and C 2 are disjoint, there exists a hyperplane that separates them, i.e., there exists a vector a 0 such that a x 1 a x 2, x 1 C 1, x 2 C 2. Proof: Consider the convex set C 1 C 2 = {x 2 x 1 x 1 C 1, x 2 C 2 } Since C 1 and C 2 are disjoint, the origin does not belong to C 1 C 2, so by the Supporting Hyperplane Theorem, there exists a vector a 0 such that 0 a x, x C 1 C 2, which is equivalent to the desired relation. Q.E.D.

60 STRICT SEPARATION THEOREM Strict Separation Theorem: Let C 1 and C 2 be two disjoint nonempty convex sets. If C 1 is closed, and C 2 is compact, there exists a hyperplane that strictly separates them. C 1 C 2 C 1 C 2 x 1 x x 2 a (a) (b) Proof: (Outline) Consider the set C 1 C 2. Since C 1 is closed and C 2 is compact, C 1 C 2 is closed. Since C 1 C 2 = Ø, 0 / C 1 C 2. Let x 1 x 2 be the projection of 0 onto C 1 C 2. The strictly separating hyperplane is constructed as in (b). Note: Any conditions that guarantee closedness of C 1 C 2 guarantee existence of a strictly separating hyperplane. However, there may exist a strictly separating hyperplane without C 1 C 2 being closed.

61 ADDITIONAL THEOREMS Fundamental Characterization: The closure of the convex hull of a set C R n is the intersection of the closed halfspaces that contain C. (Proof uses the strict separation theorem.) We say that a hyperplane properly separates C 1 and C 2 if it separates C 1 and C 2 and does not fully contain both C 1 and C 2. C 1 C 2 C 1 C 2 a a a C 1 C 2 (a) (b) (c) Proper Separation Theorem: Let C 1 and C 2 be two nonempty convex subsets of R n. There exists a hyperplane that properly separates C 1 and C 2 if and only if ri(c 1 ) ri(c 2 ) = Ø

62 PROPER POLYHEDRAL SEPARATION Recall that two convex sets C and P such that ri(c) ri(p) = Ø can be properly separated, i.e., by a hyperplane that does not contain both C and P. If P is polyhedral and the slightly stronger condition ri(c) P = Ø holds, then the properly separating hyperplane can be chosen so that it does not contain the nonpolyhedral set C while it may contain P. a a P C P C P a C Separating hyperplane Separating hyperplane Separating hyperplane On the left, the separating hyperplane can be chosen so that it does not contain C. On the right where P is not polyhedral, this is not possible.

63 NONVERTICAL HYPERPLANES A hyperplane in R n+1 with normal (µ,β) is nonvertical if β 0. It intersects the (n+1)st axis at ξ = (µ/β) u+w, where (u,w) is any vector on the hyperplane. (µ, β) w (µ, 0) (u, w) µ u + w β Nonvertical Hyperplane Vertical Hyperplane 0 u A nonvertical hyperplane that contains the epigraph of a function in its upper halfspace, provides lower bounds to the function values. The epigraph of a proper convex function does not contain a vertical line, so it appears plausible that it is contained in the upper halfspace of some nonvertical hyperplane.

64 NONVERTICAL HYPERPLANE THEOREM Let C be a nonempty convex subset of R n+1 that contains no vertical lines. Then: (a) C is contained in a closed halfspace of a nonvertical hyperplane, i.e., there exist µ R n, β R with β 0, and γ R such that µ u + βw γ for all (u,w) C. (b) If (u,w) / cl(c), there exists a nonvertical hyperplane strictly separating (u,w) and C. Proof: Note that cl(c) contains no vert. line [since C contains no vert. line, ri(c) contains no vert. line, and ri(c) and cl(c) have the same recession cone]. So we just consider the case: C closed. (a) C is the intersection of the closed halfspaces containing C. If all these corresponded to vertical hyperplanes, C would contain a vertical line. (b) There is a hyperplane strictly separating (u, w) and C. If it is nonvertical, we are done, so assume it is vertical. Add to this vertical hyperplane a small ǫ-multiple of a nonvertical hyperplane containing C in one of its halfspaces as per (a).

65 CONJUGATE CONVEX FUNCTIONS Consider a function f and its epigraph Nonvertical hyperplanes supporting epi(f) Crossing points of vertical axis f (y) = sup x R n { x y f(x) }, y R n. f(x) ( y, 1) 0 Slope = y x inf x R n{f(x) x y} = f (y), For any f : R n [, ], its conjugate convex function is defined by f (y) = sup x R n { x y f(x) }, y R n

66 EXAMPLES f (y) = sup x R n { x y f(x) }, y R n f(x) = αx β Slope = α { β if y = α f (y) = if y α β 0 β/α x 0 α y { f(x) = x f (y) = { 0 if y 1 if y > 1 0 x y f(x) = (c/2)x 2 f (y) = (1/2c)y 2 0 x 0 y

67 CONJUGATE OF CONJUGATE From the definition f (y) = sup x R n { x y f(x) }, y R n, note that h is convex and closed. Reason: epi(f ) is the intersection of the epigraphs of the linear functions of y as x ranges over R n. x y f(x) Consider the conjugate of the conjugate: f (x) = sup y R n { y x f (y) }, x R n. f is convex and closed. Important fact/conjugacy theorem: If f is closed proper convex, then f = f.

68 CONJUGACY THEOREM - VISUALIZATION f (y) = sup x R n { x y f(x) }, f (x) = sup y R n { y x f (y) }, y R n x R n If f is closed convex proper, then f = f. f(x) ( y, 1) f (x) = sup y R n { y x f (y) } Slope = y x { y x f (y) } 0 inf x R n{f(x) x y} = f (y),

69 CONJUGACY THEOREM Let f : R n (, ] be a function, let člf be its convex closure, let f be its convex conjugate, and consider the conjugate of f, f (x) = sup y R n { y x f (y) }, x R n (a) We have f(x) f (x), x R n (b) If f is convex, then properness of any one of f, f, and f implies properness of the other two. (c) If f is closed proper and convex, then f(x) = f (x), x R n (d) If člf(x) > for all x Rn, then člf(x) = f (x), x R n

70 A COUNTEREXAMPLE A counterexample (with closed convex but improper f) showing the need to assume properness in order for f = f : f(x) = { if x > 0, if x 0. We have f (y) =, y R n, f (x) =, x R n. But čl f = f, so čl f f.

71 A FEW EXAMPLES l p and l q norm conjugacy, where 1 p + 1 q = 1 f(x) = 1 p n x i p, f (y) = 1 q i=1 n y i q i=1 Conjugate of a strictly convex quadratic f(x) = 1 2 x Qx + a x + b, f (y) = 1 2 (y a) Q 1 (y a) b. Conjugate of a function obtained by invertible linear transformation/translation of a function p f(x) = p ( A(x c) ) + a x + b, f (y) = q ( (A ) 1 (y a) ) + c y + d, where q is the conjugate of p and d = (c a + b).

72 SUPPORT FUNCTIONS Conjugate of indicator function δ X of set X σ X (y) = sup x X y x is called the support function of X. epi(σ X ) is a closed convex cone. The sets X, cl(x), conv(x), and cl ( conv(x) ) all have the same support function (by the conjugacy theorem). To determine σ X (y) for a given vector y, we project the set X on the line determined by y, we find ˆx, the extreme point of projection in the direction y, and we scale by setting σ X (y) = ˆx y X y ˆx 0 σ X (y)/ y