DUAL SPACES
• Linear Functionals
• Duals of Some Common Banach Spaces
• Hahn-Banach Theorem (Extension of Linear Functionals)
• The Dual of C[a,b]
• Second Dual Space
• Alignment and Orthogonal Complements
• Minimum Norm Problems
• Hahn-Banach Theorem (Geometric Form)
EL3370 Dual Spaces - 1
LINEAR FUNCTIONALS
Definition In a real vector space !V , !f :V → is a linear functional if for all !x, y ∈V , !α ,β ∈, !f (α x + β y)= α f (x)+ β f ( y)
Examples n n 1. In ! , every linear functional is of the form f (x)= α x , where x = (x ,…,x ) ! ∑k=1 k k ! 1 n 2. In !C[0,1], !f (x)= x(1 2) is a linear functional b 3. In L (a,b), f (x)= y(t)x(t)dt , for any y ∈L (a,b), is a linear functional ! 2 ! ∫a ! 2
EL3370 Dual Spaces - 2
LINEAR FUNCTIONALS (CONT.)
Definition. A linear functional !f in a normed space is bounded if there is an !M > 0 such that ! f (x) ≤ M x for all !x ∈V . The smallest such !M is the norm of !f , ! f
Theorem. Let !f be a linear functional on a normed space !V . The following are equivalent: (1) !f is continuous (2) !f is continuous at !0 (3) !f is bounded
Proof (1)⇒(2): By definition (2)⇒(1): Let x → x , where x ,x ∈V . If f is continuous at 0, then f (x )− f (x) = f (x − x) → f (0) =0, n n n n hence f is continuous at x (2)⇒(3): Let f be continuous at 0. There is a δ >0 such that f (x) ≤1 for all x ∈V such that x ≤δ . Then, by linearity, f (x) ≤(1 δ) x for all x ∈V (3)⇒(2): If f is bounded and x is an arbitrary convergent sequence to 0, then there is an M >0 such n that f (x ) ≤ M x →0. Then, f is continuous at 0 ■ n n
EL3370 Dual Spaces - 3
LINEAR FUNCTIONALS (CONT.)
Example In l (the space of sequences with finitely non-zero entries), with norm equal to the ! 0 maximum of the absolute value of its components, define for x = (x ,…,x ,0,0,…), ! 1 n n f (x)= kx . This linear functional is unbounded ! ∑k=1 k
Linear functionals on !V form a vector space, called algebraic dual of !V , by defining
( f + f )(x)= f (x)+ f (x) 1 2 1 2 ! (λ f )(x)= λ f (x), x ∈V
EL3370 Dual Spaces - 4
LINEAR FUNCTIONALS (CONT.)
Definition The space of bounded linear functionals on a normed space !V is the normed dual of !V , denoted as !V ∗ . The norm of an f ∈V ∗ is f = sup f (x) ! ! x ≤1
Theorem. !V ∗ is a Banach space
Proof (for real normed spaces). We only need to show that V ∗ is complete. Take a Cauchy sequence {x * } in V ∗ . For every x ∈V , {x *(x)} is Cauchy in !, since x *(x)− x * (x) ≤ x * − x * x , so there is an n n n m n m x *(x)∈ such that x∗(x)→ x *(x). The functional x * defined in this way is linear, because ! n x∗(αx +β y)= lim x∗(αx +β y)=αlim x∗(x)+β lim x∗( y)=αx∗(x)+βx∗( y). In addition, for a given ε >0, n n n since {x * } is Cauchy, there is an N ∈ ! such that x∗ − x∗ <ε for all n,m≥ N , or x∗(x)− x∗ (x) <ε x for n n m n m all x ∈ X , hence taking n→∞ gives x∗(x)− x∗ (x) <ε x . Thus, x∗(x) ≤ x∗(x)− x∗ (x) + x∗ (x) < m m m ε + x∗ x , so x∗ is bounded, i.e., x∗ ∈V ∗ . Finally, as x∗(x)− x∗ (x) <ε x , we have x∗ → x∗ . Therefore, ( m ) m m V ∗ is complete ■
EL3370 Dual Spaces - 5
DUALS OF SOME COMMON BANACH SPACES
n • ! (with Euclidean norm): itself!
Take the linear functional
n f (x)= α x , x = (x ,…,x ) ! ∑k=1 k k ! 1 n
n n From Cauchy-Schwarz, f (x) = α x ≤ α 2 x , so every f is bounded, and since ! ∑k=1 k k ∑k=1 k ! n equality is achieved for x = (α ,…,α ), we have f = α 2 ! 1 n ! ∑k=1 k
Conversely, if f ∈(n )∗ , take the standard basis vectors e . Every x = (x ,…,x ) can be ! ! i ! 1 n n n n written as x = x e , so f (x)= x f (e )= α x , with α = f (e ) ! ∑k=1 k k ! ∑k=1 k k ∑k=1 k k ! k k
n Hence the dual space of ! is itself, in the sense that all bounded functionals are of the n n form f (x)= α x , with f = α 2 ! ∑k=1 k k ! ∑k=1 k
EL3370 Dual Spaces - 6
DUALS OF SOME COMMON BANACH SPACES (CONT.)
• l (1≤ p < ∞): l ! ! p ! ! q
The dual of l is l , where 1 p+1 q = 1, in the sense that every bounded linear functional ! p ! q ! ∞ in l can be written as f (x)= α x , with α = {α }∈l , and f = α q ! p ! ∑k=1 k k ! n q !
The dual of l is not l : see bonus slides for proof ∞ 1
• L (a,b) (1≤ p < ∞): L (a,b)! ! p ! ! q
Similarly to l , the dual of L (a,b) is L (a,b), with 1 p+1 q = 1, since every bounded linear ! p ! p ! q ! b functional is of the form f (x)= x(t)y(t)dt , with y ∈L (a,b), and f = y ! ∫a ! q ! q
EL3370 Dual Spaces - 7
DUALS OF SOME COMMON BANACH SPACES (CONT.)
• Dual of a Hilbert space !H
A particular linear functional in !H is !f (x)= (x, y) for !y ∈H . By Cauchy-Schwarz, ! f (x) ≤ x y , and taking !x = y shows that ! f = y . Conversely,
∗ Theorem (Riesz-Fréchet). If !f ∈H , there is a unique !y ∈H such that !f (x)= (x, y) for all !x ∈H , and ! f = y
Proof. If f =0, the theorem is trivial. Otherwise, the set N = {x ∈ H : f (x)=0} is closed (why?) and not ⊥ ⊥ equal to H . Since H = N ⊕ N , take a z ∈ N \{0}, scaled so that f (z)=1. We will show that y is a multiple of z . Given x ∈ H , we have x − f (x)z ∈ N , since f (x − f (x)z)=0. As z ⊥ N , we have 2 (x − f (x)z,z)=0, or (x,z)= f (x)(z,z), hence y = z z . If (x, y )=(x, y ) for all x ∈ H then y = y , 1 2 1 2 which proves the uniqueness of y ■
EL3370 Dual Spaces - 8
HAHN-BANACH THEOREM (EXTENSION OF LINEAR FUNCTIONALS)
Definitions 1. Let !f be a linear functional defined on a subspace !M of a vector space !V . An extension !F of !f (from !M to !N ) is a linear functional defined on N ⊇ M , !N ≠ M , such that !F M = f
2. A real valued function p defined on !V is a sublinear functional if a) p(x + x )≤ p(x )+ p(x ), for all x ,x ∈V ! 1 2 1 2 ! 1 2 b) !p(α x)= αp(x), for all !α ≥ 0 and !x ∈V
Theorem (Hahn-Banach; extension form) Let !V be a real vector space, and !p a sublinear functional on !V . Let !f be a linear functional defined on a subspace M ⊆V satisfying !f (x)≤ p(x) on !M . Then there is an extension !F of !f from !M to !V such that !F(x)≤ p(x) on !V
EL3370 Dual Spaces - 9
HAHN-BANACH THEOREM (EXTENSION OF LINEAR FUNCTIONALS) (CONT.)
Proof Let us define a partial order ≺ on the set E of extensions of f satisfying the conditions of the theorem: say that F ≺ G if G is an extension of F . For every chain C ⊆ E , according to this partial order, define Fˆ as Fˆ(x)= F(x) if x ∈D(F) for some F ∈C . Fˆ is a linear functional with domain ∪ D(F), F∈C and it is an upper bound for the chain C . Thus, by Zorn’s lemma, E has a maximal element F
The linear functional F should have domain equal to V , since otherwise there is an element y ∈V \D(F ). All x ∈ lin{D(F )+ y} are of the form x = m+α y with m ∈D(F ) and α ∈ ! . An extension of F from D(F ) to lin{D(F )+ y} has the form g(x)= F(m)+αg( y), and we can choose g( y) so that g(x)≤ p(x) on lin{D(F )+ y} (see details in the bonus slides), thus contradicting the maximality of F . This contradiction shows that F is the sought extension of f to V ■
Remark In general it is not possible to avoid the use of Zorn’s lemma (or the axiom of choice), since in fact it can be shown that the Hahn-Banach theorem is equivalent to the axiom of choice
EL3370 Dual Spaces - 10
HAHN-BANACH THEOREM (EXTENSION OF LINEAR FUNCTIONALS) (CONT.)
Corollary 1 Let !f be a bounded linear functional defined on a subspace !M of a normed space !V . Then there is an extension !F of !f to !V such that ! F = f
Proof. Take p(x)= f x in Hahn-Banach theorem ■
Corollary 2 ∗ Let !x ∈V . Then there is a nonzero !F ∈V on a normed space !V such that !F(x)= F x
∗ Proof. Assume x ≠0 (otherwise, take any F ∈V ). On lin{x}, define the bounded functional ∗ f (αx)=α x , which has norm 1. By Corollary 1, extend f to an F ∈V with norm 1 ■
The Hahn-Banach theorem, particularly as Corollary 1, can be used as an existence theorem for minimization problems
EL3370 Dual Spaces - 11
THE DUAL OF !C[a,b]
Definition !x :[a,b]→ is of bounded variation if there is a !K > 0 such that for every partition of ![a,b] n (i.e., a finite set {t ,…,t } such that a = t A consequence of Hahn-Banach theorem is that the dual of !C[a,b] is (almost) !BV[a,b]: Theorem (Riesz Representation Theorem) ∗ Let !f ∈C[a,b] . Then there is !ν ∈BV[a,b] such that b f (x)= x(t)dν(t), x ∈C[a,b] ! ∫a ∗ and ! f = TV(ν). Conversely, every !ν ∈BV[a,b] defines an !f ∈C[a,b] in this way EL3370 Dual Spaces - 12 THE DUAL OF !C[a,b] (CONT.) Proof b Given f ∈ C[a,b]∗ , we need to find a ν ∈ BV[a,b] such that f (x)= x(t)dν(t) ∫a A natural idea is to set x =u ( s ∈[a,b]), where u (t)=1 for t ∈[a,s] and 0 otherwise, so s s b s f (u )= u (t)dν(t)= dν(t)=ν(s)−ν(a) s ∫a s ∫a and we can set ν(a)=0 and ν(s)= f (u ) s Unfortunately, u ∉ C[a,b]. However, u belongs to the space B of bounded functions on [a,b], and s s C[a,b]⊆ B , so by Hahn-Banach it is possible to obtain an extension F of f from C[a,b] to B , preserving its norm. This will allow us to define ν(s)= F(u )! s b The rest of the proof is to show that: ν ∈ BV[a,b], f (x)= x(t)dν(t), and f =TV(ν) (see bonus ∫a slides for the details) ■ EL3370 Dual Spaces - 13 THE DUAL OF !C[a,b] (CONT.) Riesz Representation Theorem does not provide a unique !ν ∈BV[a,b], since, e.g., !f (x)= x(1/2) can be described by ⎧ 0, 0≤t <1 2 ⎪ v(t)= ⎨ α , t = 1 2 ⎪ 1, 1 2 Definition !NBV[a,b]⊆ BV[a,b], normalized space of functions of bounded variation, consists of those !ν ∈BV[a,b] such that !ν(a)= 0 and are right continuous on !(a,b). Here ! v = TV(v) Then, !NBV[a,b] is the dual of !C[a,b] EL3370 Dual Spaces - 14 SECOND DUAL SPACE ∗ ∗ ∗ ∗ Let !x ∈V . For any !x ∈V , denote !< x,x >:= x (x) (resembling an inner product) ∗ This notation also allows to define a functional !f on !V (for given !x ∈V ): ∗ ∗ ∗ ∗ !f (x )=< x,x >, x ∈V ∗ ∗ ∗ This !f is linear, and since ! f (x ) = < x,x > ≤ x x , so ! f ≤ x , and by Hahn-Banach, ∗ ∗ ∗ ∗ there is an !x ∈V such that !< x,x > = x x , so ! f = x Definition ∗∗ ∗ ∗ ∗ !V = (V ) , second dual of a normed space !V , is the set of bounded linear functionals on !V ∗∗ ∗∗ As we saw, there is a natural mapping !ϕ :V →V given by !ϕ(x)= x , where ∗ ∗∗ ∗ !< x ,x >=< x,x >. This mapping is linear and “norm-preserving”: !ϕ(x) = x . However, ϕ is not always surjective, i.e., there may be an !x∗∗ ∈V ∗∗ not represented by elements of !V Definition. A normed space !V is reflexive if ϕ is surjective. If so, we write !V = V ∗∗ EL3370 Dual Spaces - 15 SECOND DUAL SPACE (CONT.) Examples 1. l and L (1< p < ∞) are reflexive, since (l )∗∗ = (l )∗ = l , where 1 p+1 q = 1 ! p ! p ! ! p q p ! 2. l and L are not reflexive ! 1 ! 1 3. Every Hilbert space is reflexive EL3370 Dual Spaces - 16 ALIGNMENT AND ORTHOGONAL COMPLEMENTS ∗ ∗ ∗ ∗ ∗ For !x ∈V and !x ∈V , !< x,x >≤ x x . In Hilbert spaces, we have equality iff !x is ∗ represented by a nonnegative multiple of !x , i.e., !< x,x >= (x,α x) for some !α ≥ 0. In a normed space !V , we define ∗ ∗ ∗ ∗ Definition. !x ∈V is aligned with !x ∈V if !< x,x >= x x Example Let !x ∈C[a,b], let !Γ := {t ∈[a,b]: x(t) = x } ≠ ∅ b An x∗(x)= x(t)dν(t) is aligned with !x iff ! ∫a ν only varies on Γ so that ν is nondecreasing at !t if !x(t)> 0 and nonincreasing if !x(t)< 0 (exercise!) EL3370 Dual Spaces - 17 ALIGNMENT AND ORTHOGONAL COMPLEMENTS (CONT.) ∗ ∗ ∗ Definition. !x ∈V and !x ∈V are orthogonal if !< x,x > = 0 In Hilbert spaces, this coincides with the original definition of orthogonality ⊥ ∗ ∗ Definition. The orthogonal complement of !S ⊆ V , !S , is the set of all !x ∈V such that ∗ !< x,x > = 0 for all !x ∈S For subsets of !V ∗ , we have a “dual” definition: * ⊥ Definition. The orthogonal complement of !U ⊆ V , ! U , is the set of all !x ∈V such that ∗ ∗ !< x,x > = 0 for all !x ∈U ⊥ ⊥ ∗ ⊥ ∗∗ ⊥ Notice that !U ≠ U (!U ⊆ V ) in general, unless !V is reflexive, since !U ⊆ V , while ! U ⊆ V ⊥ ⊥ Theorem. If !M is a closed subspace of !V , then ! [M ]= M (see proof at the end) EL3370 Dual Spaces - 18 MINIMUM NORM PROBLEMS Goal: Extend the projection theorem to minimum norm problems in normed spaces The situation is more difficult: possibly several optima, given by nonlinear equations, ... Example Let !V = 2 with norm ( y , y ) = max{ y , y }, and let M = {(α ,0): α ∈}. Then ! 1 2 1 2 ! min (2,1)− y = 1, but the optimum is not unique! ! y∈M EL3370 Dual Spaces - 19 MINIMUM NORM PROBLEMS (CONT.) Theorem Let !x ∈V , where !V is a real normed space, and let !d be the distance from !x to a subspace !M . Then, d = inf x − m = max < x,x∗ > m∈M x∗ ≤1 ! x∗∈M⊥ where the max is achieved for some x∗ ∈M ⊥ . Also, if the inf is achieved for some m ∈M , ! ! 0 ! ! 0 then x∗ is aligned with x − m ! 0 ! 0 Proof (≥) Take ε >0, and m ∈ M such that x − m ≤ d + ε . Then, for every x∗ ∈ M ⊥ , x∗ ≤1: ε ε < x,x∗ >= < x −m ,x∗ > ≤ x −m x∗ ≤ d +ε ε ε ∗ ∗ ⊥ ∗ and since ε was arbitrary, we have < x,x >≤ d for all x ∈ M , x ≤1 EL3370 Dual Spaces - 20 MINIMUM NORM PROBLEMS (CONT.) Proof (cont.) (≤) Let N = lin{x +M}. If n ∈ N , then n=αx +m for some m ∈ M , α ∈ ! . Define the functional f on N as f (n)=αd . Then f (n) α d d d f = sup = sup = sup = =1 n∈N n m∈M αx +m m∈M x +m α inf x +m α m∈M By Hahn-Banach, let x∗ be a norm-preserving extension of f to V . Then x∗ =1 and x∗ = f on N . 0 0 0 By construction, x∗ ∈ M ⊥ and < x,x∗ >= d 0 0 Now, let m ∈ M be such that x −m = d , and take x∗ ∈ M ⊥ such that x∗ =1, and < x,x∗ > = d . Then, 0 0 0 0 0 < x −m ,x∗ > = < x,x∗ > = d = x −m x∗ , so x∗ is aligned with x −m ■ 0 0 0 0 0 0 0 EL3370 Dual Spaces - 21 MINIMUM NORM PROBLEMS (CONT.) Corollary (generalized projection theorem) Let !x ∈V and !M be a subspace of !V . Then m ∈M satisfies x − m ≤ x − m for all !m∈M iff ! 0 ! 0 there is an x∗ ∈M ⊥ aligned with x − m ! 0 ! 0 The following is a “dual” version of the previous theorem: Theorem Let !M be a subspace of !V , and !x∗ ∈V ∗ , which is at a distance !d from !M ⊥ . Then d = min x∗ − m∗ = sup < x,x∗ > ∗ ⊥ m ∈M x∈M ! x ≤1 where the min is achieved for m∗ ∈M ⊥ . If the sup is achieved for some x ∈M , then x∗ − m∗ ! ! 0 ! ! ! 0 is aligned with x ! 0 EL3370 Dual Spaces - 22 MINIMUM NORM PROBLEMS (CONT.) Example (control problem, again) Consider the DC motor problem again, where we want to find the current !u:[0,1]→ to drive a motor governed by !θ(t)+θ(t)= u(t) from θ(0)=θ(0)= 0 to θ(1)= 1, θ(1)= 0, so as to minimize max u(t) ! ! ! ! t∈[0,1] We can formulate this problem in a dual space, by forcing u∈(L )∗ = L , so that we seek a ! 1 ∞ u∈(L )∗ = L of minimum norm ! 1 ∞ As in the previous example, the constraints are 1 t−1 t−1 e u(t)dt = < y1 ,u > = 0, y1(t)= e ∫0 1 (1− et−1 )u(t)dt = < y ,u > = 1, y (t)= 1− et−1 !∫0 2 2 EL3370 Dual Spaces - 23 MINIMUM NORM PROBLEMS (CONT.) Example (cont.) Therefore, the optimization problem is L = min u = min u! −u = min u! −u u L u L ⊥ ∈ ∞ !∈ ∞ u!∈M < y1 ,u>=0 < y1 ,u!>=0 < y ,u>=1 < y ,u!>=0 ! 2 2 where !u is any function in L satisfying the constraints: < y ,u > = 0, < y ,u > = 1, and ! ∞ ! 1 ! 2 M = lin{ y , y } ! 1 2 From the theorems: L = min u = sup < x,u > = sup < ay + by ,u > = sup b u L 1 2 ∈ ∞ x M ∈ ay1+by2 ≤1 ay1+by2 ≤1 < y1 ,u>=0 x ≤1 < y ,u>=1 ! 2 1 and the constraint ay + by ≤1, in L , corresponds to (a− b)et−1 + b dt ≤1 ! 1 2 ! 1 !∫0 EL3370 Dual Spaces - 24 MINIMUM NORM PROBLEMS (CONT.) Example (cont.) This means that the optimal value is equal to sup b 1 (a−b)et−1+b dt≤1 !∫0 which is attained and can be solved explicitly (since it is simple and finite dimensional!) t−1 Qualitatively, the optimal !x(t)= (a− b)e + b is aligned with the optimal !u, i.e., !< x,u > = x u L x Since the right side is the maximum of the left side t over all u∈L with u = L, !u has to take only values 0 1 ! ∞ ! !±L, depending on !sgn(x(t)): “Bang-bang” control! u L As !x is the sum of an exponential and a constant, it can change sign at most once, so !u has to change sign at most once EL3370 Dual Spaces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c:-P&)&>=,.5,*)4.! & &H ! &)4B& &:/&)&/(D/,)-.! I<&3;& ! 9&7>.4&! 9& !" # ! ! ! !"#$% ' " ! # " !# $ " "!# " =!" +# " # =! " = # !$ % # = !/8& .6.5=& ! & ;(*;:**/& ! &H 9&! Id& *.77:4E&! & =:.*B/! & & <& 3;& ! ! " # !" # $ % &' !" " ! " " = # !$ % #" = " !# " # = 9&*.7! & d&.6.5=& &!/)7:/;:./& 9&/8& ! &E:6./& ! & ! ! # ! ! " # ! " # $ % &' " !# " # ! " # # = " = # " " + $ "!# "1$1 " " # " !! ! &3;&! 9&B.;:4.&7>.&/(D/,)-.&! ! <&C)P.&)4&! &@:7>& d&7>.49&;85&.6.5=& ! !" # ! " ! " # $ % &' ! ! " $ % " !# $ " # = $ "$ # " = $ = " # " ! " = = " + # ! 9& ! 9&/8& ! &;85&/8?.&! 9& 9& &:/&?)M:?)*! *=&,58,.5<&a8@9&;85& 9& ! " " # # $ ! " # ! "!" # !" ! !" = # !$ % # =% " !" = # !$ % # = *.7& &/(->&7>)7& " $ d&7>.49&!" # $ % & " # $ & '( 9&@>:->&:/&)&>=,.5,*)4.& !" !! # "$ % # # &! # = ! !" ! !" !" = # ! 3;& !" 9& 84.& -)4& /-)*.& ! "& /8& 7>)7& ! " # $ % &'<& 3;& )*/8& ! " # $ % &'9& 7>.4& ! " # $ % $ % &'9&D(7&7>.&/?)**./7&/(D/,)-.&-847):4:4E& &:/& 9&/8& && & !"##$%&'()*&+,)-./&0&2S& HAHN-BANACH THEOREM (GEOMETRIC FORM) By definition, a hyperplane !H is either closed or dense in !V , since H ⊆ H ⊆V , and H is also a linear variety (why?), so either !H = H or !H = V , because H is maximally proper (i.e., it cannot be inside a larger but proper linear variety) Theorem. Let f ≠ 0 be a linear functional on !V . Then, for every !c , !H = {x : f (x)= c} is closed iff !f is continuous −1 Proof. If f is continuous, then H = f ({c}) is closed since {c} is closed for every c ∈! Conversely, as H = {x : f (x)= c} is a hyperplane, there is a y ∈V \H , and every x ∈V fulfills x =α y +h, with α ∈ ! and h∈ H . Then, consider a sequence {x } in V , with x =α y +h (α ∈ !, h ∈ H ), such n n n n n n that x → x =α y +h∈V (α ∈ ! , h∈ H ). Since H is closed, d = inf y −h >0 (from Exercise set 1), so n h∈H x − x = (α −α)y +h −h ≥ α −α d , which implies that α →α , hence f (x )=α f ( y)→α f ( y)= f (x) , n n n n n n n so f is continuous ■ See bonus slides for an example of a linear functional for which H is dense EL3370 Dual Spaces - 27 HAHN-BANACH THEOREM (GEOMETRIC FORM) (CONT.) A hyperplane !H = {x : f (x)= c} generates the half-spaces !{x : f (x)≤ c}, !{x : f (x)≥ c}, !{x : f (x)< c}, !{x : f (x)> c} If !f is continuous, the first !2 are closed and the last two are open Our goal is to establish a geometric Hahn-Banach theorem: Given a convex set K such that intK ≠ ∅, and x ∉intK , ! ! ! 0 x0 there is a closed hyperplane containing x but disjoint from ! 0 K !intK If !K were !B(0,1), the theorem is easy, since Hahn-Banach assures the existence of an x∗ aligned with x : if x ∈intB(0,1), ! 0 ! 0 ! < x,x∗ > ≤ x x ∗ < x x ∗ = < x ,x∗ >, so H = {x : < x,x∗ > = < x ,x∗ >} is enough ! 0 0 0 0 0 0 ! 0 0 0 We thus need to generalize the unit ball !B(0,1) to arbitrary convex sets! EL3370 Dual Spaces - 28 HAHN-BANACH THEOREM (GEOMETRIC FORM) (CONT.) Definition (Minkowski functional) Let !K be a convex set in a normed space !V , and assume !0∈intK . The Minkowski functional p:V → !+ of !K on !V is 0 p(x)= inf r >0: x ∈ rK { } Properties (assuming !0 ∈intK ) (1) !0≤ p(x)< ∞ for all !x ∈V (2) !p(α x)= αp(x) for !α > 0 (3) p(x + x )≤ p(x )+ p(x ) ! 1 2 1 2 (4) !p is continuous (5) !K = {x : p(x)≤1}, !intK = {x : p(x)<1} !p is then a continuous sublinear functional (see bonus slides for proofs of these properties) EL3370 Dual Spaces - 29 HAHN-BANACH THEOREM (GEOMETRIC FORM) (CONT.) Theorem (Mazur’s Theorem, Geometric Hahn-Banach) Let !K be a convex set in a real normed space !V such that !intK ≠ ∅. Let !W be a linear variety such that !W ∩intK = ∅. Then there is a closed hyperplane containing !W but no ∗ ∗ ∗ interior points of !K , i.e., there is an x ∈V \{0} and !c ∈ such that !< w,x > = c for !w ∈W , ∗ ∗ < k,x > ≤ c for k ∈ K , and !< k,x > < c for !k ∈intK Proof. Idea: use the Minkowski functional p in the Hahn-Banach theorem! Assume that 0∈intK . Let M = linW . Then, W is a hyperplane in M which does not contain 0, so there is a linear functional !f on M such that W = {x : f (x)=1} Let p be the Minkowski functional of K . Since W has no interior points of K , f (x)=1≤ p(x) for x ∈W . Every point in M can be written as αx , where α ∈ ! and x ∈W , so if α >0, f (αx)=α ≤ p(αx) (by homogeneity), and if α <0, f (αx)≤0≤ p(αx), so f (x)≤ p(x) in M By Hahn-Banach, there is an extension F of f to V such that F(x)≤ p(x). Let H = {x : F(x)=1}. Since F(x)≤ p(x) on V , and p is continuous, so is F (because −p(−x )≤ F(x )≤ p(x ), so if x → 0, then n n n n F(x )→ 0), and F(x)<1 on intK , while F(x)≤1 on K . Therefore, H is the desired hyperplane ■ n See bonus slides for example of why !intK ≠ ∅ is needed in Mazur’s theorem EL3370 Dual Spaces - 30 HAHN-BANACH THEOREM (GEOMETRIC FORM) (CONT.) Some corollaries 1. (Supporting Hyperplane Theorem). If x ∉intK ≠ ∅, there is a closed hyperplane !H containing !x such that !K lies on one side of !H (If !x ∈∂K , !H is a supporting hyperplane) Proof. A special case of Mazur’s theorem, where W = {x} ■ 2. (Eidelheit Separating Hyperplane Theorem). If K and K are convex, ! 1 ! 2 such that intK ≠ ∅ and K ∩intK = ∅. Then there is closed ! 1 ! 2 1 hyperplane separating K and K (i.e., an x∗ ∈V ∗ such ! 1 ! 2 ! that sup x,x∗ inf x,x∗ ) x∈K < > ≤ x∈K < > ! 1 2 Proof. Let K = K +(−K ). intK ≠ ∅ , and 0∉intK , so apply Mazur with W = {0} ■ 1 2 3. If !K is closed and convex, and !x ∉K , there is a closed halfspace that contains !K but not !x Proof. d = inf x − y > 0, so B(x,d /2) does not intersect K . Apply y∈K Corollary 2 with K = K and K = B(x,d /2) ■ 1 2 4. (Dual formulation of convex sets). A closed convex set is equal to the intersection of all the closed half-spaces that contain it Proof. Follows directly from Corollary 3 ■ EL3370 Dual Spaces - 31 BONUS: DETAILS OF PROOF OF HAHN-BANACH THEOREM To show that !g( y) can be chosen so that !g(x)≤ p(x) on !lin{M + y}, notice that for every m ,m ∈M , f (m )+ f (m )= f (m + m )≤ p(m + m )≤ p(m − y)+ p(m + y), or ! 1 2 ! 1 2 1 2 1 2 ! 1 2 f (m )− p(m − y)≤ p(m + y)− f (m ) ! 1 1 2 2 so sup [ f (m)− p(m− y)]≤ inf [p(m+ y)− f (m)]. Take a !c ∈ between these bounds. ! m∈M m∈M For !x = m+α y , we then define !g(x)= f (m)+αc . If !α > 0, ⎡ ⎛ m⎞ ⎤ ⎡ ⎛ m⎞ ⎛ m ⎞ ⎛ m⎞ ⎤ ⎛ m ⎞ f (m)+αc = α ⎢ f ⎜ ⎟ + c⎥ ≤α ⎢ f ⎜ ⎟ + p⎜ + y⎟ − f ⎜ ⎟ ⎥ = αp⎜ + y⎟ = p(m+α y) ! ⎣ ⎝ α ⎠ ⎦ ⎣ ⎝ α ⎠ ⎝ α ⎠ ⎝ α ⎠ ⎦ ⎝ α ⎠ and if !α = −β < 0, ⎡ ⎛ m⎞ ⎤ ⎡ ⎛ m⎞ ⎛ m ⎞ ⎛ m⎞ ⎤ ⎛ m ⎞ f (m)− βc = β ⎢ f ⎜ ⎟ − c⎥ ≤ β ⎢ f ⎜ ⎟ + p⎜ − y⎟ − f ⎜ ⎟ ⎥ = βp⎜ − y⎟ = p(m− β y) ⎣ ⎝ β ⎠ ⎦ ⎣ ⎝ β ⎠ ⎝ β ⎠ ⎝ β ⎠ ⎦ ⎝ β ⎠ Thus !f (m)+αc ≤ p(m+α y) for all α , and !g is an extension of !f to !lin{M + y} EL3370 Dual Spaces - 32 ⊥ ⊥ BONUS: PROOF THAT ! [M ] = M ⊥ ⊥ It is clear that M ⊆ [M ] Let x ∉ M . On lin{x +M}, define the linear functional f (αx +m)=α for all m ∈ M . Then, f (x +m) 1 f = sup = m∈M x +m inf x +m m∈M ∗ * and since M is closed, f <∞. Thus, by Hahn-Banach, we can extend f to an x ∈V . Since f (x)=0 on ∗ ⊥ ∗ ⊥ ⊥ M , x ∈ M , but also < x,x > =1, so x ∉ [M ] ■ EL3370 Dual Spaces - 33 BONUS: PROOF THAT l IS NOT THE DUAL OF l ! 1 ∞ Let c be the subspace of l consisting of sequences x = {x } such that x →0. We will first show that 0 ∞ n n ∞ c∗ = l . First, let z ∈l and define ϕ :c → ! so that ϕ (x)= z x . ϕ is linear, and bounded, since 0 1 1 z 0 z ∑n=1 n n z ∞ ϕ (x) = z x ≤ z 1 x ∞ , and if x = sgn(z ), then ϕ (x) = z 1 x ∞ , so ϕ = z 1 . Conversely, if z ∑n=1 n n n n z z N f :c → ! is a bounded linear functional, define z = {z } by z = f (e ); then, if v = sgn(z )e , 0 n n n ∑n=1 n n N z = f (v) ≤ f v ∞ = f , so taking N →∞ we conclude that z ∈l . Furthermore, for every x ∈ c , ∑n=1 n 1 0 ∞ N N ∞ f (x)= f x e = f lim x e = lim x f (e )= z x =ϕ (x) (by continuity of f ), so (∑n=1 n n ) ( N→∞ ∑n=1 n n ) N→∞ ∑n=1 n n ∑n=1 n n z there is a one-to-one correspondence between c∗ and l 0 1 Now we show that there are elements of l∗ which are not of the form ϕ with z ∈l . Since c ⊆ l , ∞ z 1 0 ∞ and x ∈l , x =1, does not belong to c , Mazur’s theorem with K = B(x,1/2) and W = c implies the ∞ n 0 0 existence of a f ∈l∗ such that f (c )=0 and f (x)=1; this f cannot be written in the form ϕ , since ∞ 0 z otherwise z =0 and f ≡0, which is not true ■ EL3370 Dual Spaces - 34 BONUS: DETAILS OF DERIVATION OF DUAL OF !C[a,b] (1) ν ∈ BV[a,b]: Let a =t n n ⎛ n ⎞ n v(ti )−v(ti 1 ) = εi[v(ti )−v(ti 1 )]= F⎜ εi[ut −ut ]⎟≤ F εi[ut −ut ] = F = f ∑ − ∑ − ∑ i i−1 ∑ i i−1 i 1 i 1 ⎝ i 1 ⎠ i 1 = = = != ##"##$ =1 so ν ∈ BV[a,b], with TV(v)≤ f b n (2) f (x) x(t)d (t): For x C[a,b], let z x(t )[u u ] B . Then, z x = ν ∈ =∑ i−1 t − t ∈ − B = ∫a i=1 i i−1 max max x(t ) x(t ) 0 as max t t 0, since x is uniformly continuous. Also, continuity i t ≤t≤t i−1 − i → i i − i−1 → i−1 i n b of F yields F(z)→ F(x)= f (x), but F(z)= x(t )[v(t − v(t )]→ x(t)dv(t), by definition of the ∑i=1 i−1 i i−1 ∫a b Riemann-Stieltjes integral, so f (x)= x(t)dv(t) ∫a n (3) f =TV(ν): Since F(z) = x(t )[v(t −v(t )] ≤ x TV(v), since F(z)→ F(x)= f (x), we obtain ∑i=1 i−1 i i−1 f (x) ≤ x TV(v), i.e., f ≤TV(v). Combining this with (1) yields f =TV(v) ■ EL3370 Dual Spaces - 35 BONUS: EXAMPLE OF A DENSE HYPERPLANE It is easy to build examples of closed hyperplanes, since these are the contour surfaces of bounded functionals. Finding dense hyperplanes is a bit harder. Here is an example: Consider the space c ⊆ l of sequences that converge to 0, and let {e } be the standard 0 ∞ n n∈! basis of these spaces. Let v = {1,1 2,1 3,1 4,…}∈c . The set {v ,e ,e ,…} is linearly 0 0 0 1 2 independent (recall that linear independence refers to finite linear combinations); this set can be complemented with a set {b } ⊆ c to obtain a Hamel basis, i.e., a linearly i i∈I 0 independent set {v ,e ,e ,…}∪{b } such that every x ∈c can be written as a finite linear 0 1 2 i i∈I 0 combination of elements of this set (recall the application of Zorn’s Lemma from Lecture 1) Define a functional f :c → ! by 0 ∞ f α v + α e + β b = α ( 0 0 ∑n=1 n n ∑i∈I i i ) 0 Note that f ≠ 0 (because f (v )= 1≠ 0), and since e ∈Ker f for every n∈!, Ker f = 0 n {x ∈c : f (x)= 0} is a dense hyperplane in c . Exercise: prove that f is unbounded 0 0 EL3370 Dual Spaces - 36 BONUS: PROOF OF PROPERTIES OF MINKOWSKI FUNCTIONAL (1) 0≤ p(x)< ∞ for all x ∈V : By definition, p(x)≥ 0. Also, since 0∈intK , there is an δ > 0 such that B(0,δ )⊆ K , so for every ε >0, x ∈( x δ + ε)B(0,δ )⊆ ( x δ + ε)K , hence p(x)≤ x δ + ε for all ε > 0, so p(x)≤ x δ < ∞ (2) p(αx)= αp(x) for α > 0: p(αx)= inf{r >0:αx ∈ rK }= inf{αr! >0: x ∈ r!K }=αp(x) (3) p(x + x )≤ p(x )+ p(x ): Given ε > 0, let p(x )≤ r ≤ p(x )+ ε so x /r ∈K (i = 1,2). Then, 1 2 1 2 i i i i i (x + x )/(r + r )= [r /(r + r )](x /r )+ [r /(r + r )](x /r )∈K , since this is a convex combination of 1 2 1 2 1 1 2 1 1 2 1 2 2 2 x /r and x /r , and K is convex. Hence, p(x + x )≤ r + r ≤ p(x )+ p(x )+2ε , and since ε > 0 was 1 1 2 2 1 2 1 2 1 2 arbitrary, p is sublinear (4) p is continuous: By (1), if x → 0, 0≤ p(x )≤ x δ → 0, so p is continuous at 0. Also, given x, y ∈V , n n n p(x)= p(x − y + y)≤ p(x − y)+ p( y) and p( y)≤ p( y − x)+ p(x), so −p( y − x)≤ p(x)− p( y)≤ p(x − y). Thus, if x → x , −p(x − x)≤ p(x)− p(x )≤ p(x − x ), so p(x )→ p(x) and p is continuous n n n n n (5) K = {x : p(x)≤1}, intK = {x : p(x)<1}: If x ∈K , p(x)≤1, so if x ∈K → x ∈V , by (4), p(x)≤1, thus n K ⊆ {x :p(x)≤1}, and if x ∈intK , there is an ε > 0 such that (1+ ε)x ∈K , so intK ⊆ {x :p(x)<1}. Conversely, if x ∉K , there is an ε ∈(0,1) such that (1−δ)x ∉ K for all 0≤δ ≤ε , hence p(x)>1 and −1 K = {x :p(x)≤1}, and if p(x)<1, x ∈K , so p ((0,1)) is a nbd of x , thus intK = {x :p(x)<1} ■ EL3370 Dual Spaces - 37 8CE"&S%[;Y%!C%[)%E))!%! " !" %HE%D#\"IW&%=;)CI)D]% ! " = ! !"#$ !" !" !" & !" 3;&!"#$ 9&:7&:/&487&,8//:D*.&:4&E.4.5)*&78&;:4B&)&>=,.5,*)4.& &-847):4:4E& &/(->&7>)7& & (30+2,.<,?-7:4,%*:./&84&84.&/:B.&8;& & & !" = ! " = # # # ! ! # = " = # # # ! ! # = # = " ! ( ! ( $ & & ! " = !( !" ! $# =& "& $ & 34& 9&*.7& "# % % ' ) *+&)4B& "# % % ' ) *+9&)/&/>8@4&:4& ! ! !" !" =# ! !" 7>.& ;:E(5.& D.*8@<& a87.& 7>)7& !"#$ <& O>:*.& ! "#$ 9& 7>.5.& :/& 84*=& 84.& >=,.5,*)4.& -847):4:4E&7>.&,*)4.& &H4)?.*=&:7/.*;G& I9&)4B& &*:./&84&D87>&/:B./&8;&:7& & & " ! & !"##$%&'()*&+,)-./&0V& BONUS: WEAK∗ CONVERGENCE A sequence {x } in a normed space V is said to (strongly) converge to x ∈V if x − x →0 n n Another notion of convergence in V is: {x } converges weakly to x ∈V if < x ,x∗ >→< x,x∗ > n n as n→∞ for all x∗ ∈V ∗ . If x → x strongly, then < x ,x∗ > − < x,x∗ > ≤ x∗ x − x →0, so n n n x → x weakly too n Example. In l , consider the sequence {e } with e = {0,…,0,1,0,…}, with 1 in the n-th ! 2 n n place. For every x∗ =(η ,η ,…), < e ,x * >=η →0 as n→∞, so e →0 weakly. However, 1 2 n n n e −0 =1 for every n, so e does not converge to 0 strongly n n In the dual space V ∗ , we can similarly define strong and weak convergence (since it is a normed space), but an even weaker notion is available: {x ∗ } converges in weak∗ -sense to x∗ ∈V ∗ if < x,x∗ >→< x,x∗ > as n→∞ for all x ∈V n n Weak convergence in V ∗ implies weak ∗ convergence, since in the former one requires that < x ,x∗ >→< x,x∗ > for all x ∈V ∗∗ , where V ∗∗ ⊇V n EL3370 Dual Spaces - 39 BONUS: WEAK∗ CONVERGENCE (CONT.) Example. For V = c ⊂ l , the set of sequences that converge to 0, V ∗ = l and V ∗∗ = l 0 ∞ ! 1 ∞ (why?). In V ∗ = , e →0 weak ∗ , but not weakly, since < x ∗ ,x∗∗ >→/ 0 for x∗∗ =(1,1,…,)∈V ∗∗ !l1 n Weak ∗ convergence corresponds to point-wise convergence for linear functionals, so it has the relative topology on V ∗ of the product topology of !V = !; in particular, the ∏ x∈V weak ∗ topology is Hausdorff but not first-countable in general. Also, Theorem (Banach-Alaoglu). The closed unit ball in V ∗ , K , is weak ∗ -compact Proof. Since K = {x∗ ∈V ∗ : x∗(x) ≤ x for all x ∈V }=V ∗ ∩ [− x , x ]=:V ∗ ∩D, and D is compact by ∏ x∈V Tychonoff, it suffices to show that K is closed. If f ∈ K ∩D, pick x, y ∈V , α,β ∈ ! , and ε >0, and let g ∈ K such that f (x)− g(x) <ε , f ( y)− g( y) <ε and f (αx +β y)− g(αx +β y) <ε . Then, f (αx +β y)−α f (x)−β f ( y) ≤ f (αx +β y)− g(αx +β y) + α g(x)− f (x) + β g( y)− f ( y) <(1+ α + β )ε and f (x) ≤ f (x)− g(x) + g <ε +1, and since ε,x, y,α,β were arbitrary, f is linear and f ≤1, so f ∈ K and K is closed ■ Weak*-compactness is useful for establishing existence of solutions of optimization problems involving a weak∗ -continuous functional on a dual space EL3370 Dual Spaces - 40 BONUS: WEAK∗ CONVERGENCE (CONT.) ∗ ∗ ∗ ∗ The weak* topology of V is the weakest in which functionals f (x )=< x,x > on V , for some x ∈V , are continuous (i.e., every other such topology contains the open sets of the ∗ ∗ weak* topology). This is because for f (x ) = < x,x > <ε to hold, the topology should ∗ ∗ ∗ contain open sets of the form {x ∈V : < x,x > <ε} for all x ∈V , as well as finite intersections of these sets, which generate the weak* topology. Conversely, ∗ ∗ ∗ Theorem. Every weak*-continuous linear functional on V has the form f (x )=< x,x > for some x ∈V Proof. Let f be such a functional. Then, for every ε >0, there is a set U = {x∗ : < x ,x∗ > <δ,…, < x ,x∗ > <δ} where x ,…,x ∈V , such that if x∗ ∈U , f (x∗ ) <ε . Thus, by 1 n 1 n linearity, f (x∗ ) ≤(ε /δ)max { < x ,x∗ > }, so if < x ,x∗ >=0 for all i , then f (x∗ )=0; assume now i∈{1,…,n} i i w.l.o.g. that {x ,…,x } is l.i. Let F(x∗ )=(< x ,x∗ >,…,< x ,x∗ >, f (x∗ )). Since there is a nbd of (0,…,0,1) 1 n 1 n n+1 T ∗ ∗ not intersecting R(F), by Mazur’s theorem there is a λ ∈ ! \{0} such that λ F(x )=0 for all x , i.e., n < λ x ,x∗ >+λ f (x∗ )=0 for all x∗ , and since {x ,…,x } are l.i., λ ≠0, so ∑i=1 i i n+1 1 n n+1 n f (x∗ )=< (−λ /λ )x ,x∗ > ■ ∑i=1 i n+1 i EL3370 Dual Spaces - 41 BONUS: WEAK∗ CONVERGENCE (CONT.) ∗ ∗ Let B,B be the closed unit balls in V ,V , respectively. The following is a converse of the Banach-Alaoglu theorem: ∗ ∗ Theorem (Krein-Smulian). If E ⊆V is convex and such that E ∩(rB ) is weak*-compact for every r >0, then E is weak*-closed Proof Firstly note that E is norm-closed, since if {x∗ } is a sequence in E that converges to x∗ ∈V ∗ , then {x∗ } n n is bounded, i.e., x∗ ≤ M for some M > 0, thus x∗ ∈ E ∩(M B∗ ) for all n, so x∗ ∈ E ∩(M B∗ )⊆ E . This n n ∗ means that if x ∈V \E , there is a ball centered at x which does not intersect E ; thus, by translation and scaling, the theorem will be proven by showing that “if E ∩B∗ = ∅, then there exists an x ∈V such ∗ ∗ ∗ ∗ ∗ that < x,x >≥1 for every x ∈E ”: since {x : < x,x >≥1} is a weak -closed half-space, E is the intersection of closed half-spaces, and hence it is weak ∗ -closed ∗ ∗ Next, given a subset F ⊆ X , define its polar P(F):= {x : < x,x > ≤1 for all x ∈F}; notice that the −1 ∗ intersection of all sets P(F) as F ranges over all finite subsets of r B is exactly rB EL3370 Dual Spaces - 42 BONUS: WEAK∗ CONVERGENCE (CONT.) Proof (cont.) To establish the proposition, let F = {0}, and assume that finite sets F ,…,F have been chosen such 0 0 k−1 that F ⊆ i−1B and P(F )∩!∩P(F )∩E ∩kB∗ = ∅; note that this holds for k = 1. Let Q = i 0 k−1 P(F )∩!∩P(F )∩E ∩(k+1)B∗ ; if P(F)∩Q ≠ ∅ for every finite set F ⊆ k−1B , the weak ∗ -compactness 0 k−1 ∗ of Q implies (via the finite intersection property) that kB ∩Q ≠ ∅ , which contradicts the properties of F ,…,F . Hence, there is a finite set F ⊆ k−1B such that P(F )∩Q = ∅, which yields the sequence {F } 0 k−1 k k k By construction, {F } satisfies E ∩∩∞ P(F )= E ∩P(∪F )= ∅ . Now, arrange the elements of ∪F in a k k=1 k k k sequence {x }; note that x →0. Define T :V ∗ → c (the space of sequences converging to 0) by k k 0 Tx∗ = {< x ,x∗ >}. T(E) is a convex subset of c , and, due to E ∩P(∪F )= ∅ , Tx∗ = sup < x ,x∗ > ≥1 for n 0 k n n every x∗ ∈E . Thus, since c∗ = l (see previous bonus slides), by the separating hyperplane theorem 0 1 ∞ ∞ there is a y ∈l (i.e., y < ∞ ) such that EL3370 Dual Spaces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a87.&7>)7&)&,8*=78,.&:/&7>.&-846.M&>(**&8;&:7/&6.57:-./&H85& .M75.?.&,8:47/I& !"##$%&'()*&+,)-./&0&KK& BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) The Krein-Milman theorem is one of the cornerstones of functional analysis, as it extends to very general vector spaces some of the intuitive results of Euclidean geometry shown in the previous slide. To state it, we first need some definitions: Definitions - Topological vector space: a vector space V over F with a Hausdorff topology such that the addition +:V 2 →V and scalar multiplication ⋅:F ×V →V operations are continuous. Every normed space induces a topological vector space - Locally convex topological (LCT) vector space: a real topological vector space V such that for every x ∈V and every nbd U of x , there is a convex nbd W ⊆U of x - V ∗ (dual of a topological vector space V ): set of continuous linear functionals on V - Extreme point of a convex set K : a point x ∈ K such that if x = λ y +(1−λ)z with y,z ∈ K and λ ∈[0,1], then either λ =0 or λ =1 - conv(X) (convex hull of a subset X of a real vector space): set of all convex combinations (i.e., expressions α x +!+α x where α ,…,α ≥0, α =1 and n ∈ !) of points in X 1 1 n n 1 n ∑i i EL3370 Dual Spaces - 45 BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) In locally convex topological vector spaces, linear functionals can separate points: Theorem (separation of points) ∗ If V is an LCT vector space, V separates points in V (i.e., if x, y ∈V , x ≠ y , there is an ∗ f ∈V such that f (x)≠ f ( y)) Proof. Since V is an LCT vector space, all properties of Minkowski functional of a convex set hold. In particular, if U is a convex set and 0∈ int U , then p is finite, because scalar multiplication is continuous, so for every z ∈V there is an α ∈ ! such that αz ∈U Therefore, Mazur’s theorem applies to LCT vector spaces, as well as its 4 corollaries. Corollary 3, in particular implies this result, by taking K = { y} ■ Corollary (separation of a convex set and a point) ∗ If K is a closed, convex set in an LCT vector space V , and y ∈V \K , then there is an f ∈V such that f (x)≤ c for all x ∈ K , and f ( y)> c EL3370 Dual Spaces - 46 BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) Theorem (Krein-Milman) Let V be an LCT vector space, and K ⊆V a convex, compact and non-empty set. Then: (i) K has at least one extreme point (ii) K is the closure of the convex hull of its extreme points Proof. (i) Call a set E ⊆ K extreme if it is convex, compact, non-empty and such that if x ∈ E is a convex combination of y,z ∈ K , then both y,z ∈ E . Since K is extreme, the collection F of all extreme subsets of K is not empty. Partially order {E } by set inclusion; we will show that every subchain {E } j j of F has a lower bound in F , ∩E . ∩E is clearly convex, compact (as it is the intersection of compact j j sets), and if x ∈∩E is a convex combination of y,z ∈ K , then for every j , x ∈ E , so both y,z ∈ E , j j j and hence y,z ∈∩E . It remains to show that ∩E is non-empty: if ∩E = ∅ , then ∪E C = K , but since K j j j j C C C is compact, there is a finite subcollection {E j } of {E j } whose union is K ; however, since {E j } is k k C totally ordered, there is a largest extreme set E j = K , i.e., E j = ∅, which is a contradiction k k By Zorn’s lemma, K has a minimal extreme subset E . If E had at least two points, by the theorem on ∗ separation of points, there is an f ∈V that separates them EL3370 Dual Spaces - 47 BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) Proof (cont.) Since E is compact and f is continuous and non-constant, it achieves its maximum, say, f , on some max proper subset M of E , which is closed (and hence compact) due to the continuity of f . M is convex, and if x ∈ M satisfies x = λ y +(1−λ)z for some y,z ∈ K and λ ∈[0,1], then both y,z ∈ E and furthermore f = f (x)= λ f ( y)+(1−λ)f (z), so f ( y)= f (z)= f and y,z ∈ M ; thus, M is a smaller max max extreme subset of K than E , a contradiction, so E has only one point, which is an extreme point of K (ii) Let K be the set of all extreme points of K , and C = conv(K )⊆ K . If x ∉ C , by the corollary on the e e e e separation of a convex set and a point, there is an f ∈V ∗ such that f ( y)≤ c for all y ∈ C and f (x)> c . e As before, f achieves its maximum over K , f , on some closed non-empty set E ⊆ K . If y ∈ E max satisfies y = λz+(1−λ)w for some z,w ∈ K and λ ∈[0,1], then f = f ( y)= λ f (z)+(1−λ)f (w), so max f (z)= f (w)= f and z,w ∈ E , hence E is extreme. By part (i), E should contain at least one extreme max point, p, which belongs to K ⊆ C , hence f = f (p)≤ c . Therefore, since f (x)> c ≥ f , x ∉ K . Thus, e e max max C ⊇ K , which implies that C = K ■ e e EL3370 Dual Spaces - 48 BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) Example: c is not the dual space of a Banach space 0 By Banach-Alaoglu, the closed unit ball of a dual space is weak∗ -compact, as well as convex and non-empty, hence by Krein-Milman, it should contain at least one extreme point However, the closed unit ball of c does not have extreme points: indeed, let x ∈c , x ≤1. 0 0 By definition, x = {x }, with x → 0, so there is an N ∈! such that x <1/2 for all n ≥ N . n n n Let y1 , y2 ∈c be such that y1 = y2 = x for n ≤ N and y1 = x +2−n , y2 = x −2−n for n > N . 0 n n n n n n n 1 2 1 2 1 2 Then, y ≤1, y ≤1, x = (1/2)( y + y ), but y ≠ x , y ≠ x , so x is not an extreme point EL3370 Dual Spaces - 49 BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) To apply Krein-Milman, the following lemma is often useful: Lemma. If V is an LCT vector space, K ⊆ V a compact and convex set, and F ⊆ K a compact set such that conv(F)= K , then the extreme points of K are contained in F Proof. Let x be an extreme point of K not in F . Since F is compact, there is a nbd U of 0 such that 0 (x +U )∩F = ∅ , and a convex nbd U of 0 such that U −U ⊆ U , so (p+U)∩(F +U)= ∅, hence p ∉F +U . 0 0 The family { y +U} is an open cover of F , so by compactness, { y +U} is a finite subcover. Let y∈F i i=1,…,n Q = conv([ y +U]∩K )⊆ y +U ; note that Q is closed, and hence a compact subset of K , therefore i i i i K = conv(Q ∪!∪Q )= conv(Q ∪!∪Q ) 1 n 1 n (This result is proven by induction on n: for n = 2, the mapping ϕ :[0,1]×Q ×Q → conv(Q ∪Q ) given 1 2 1 2 by ϕ(α , y , y )= α y +(1−α )y is continuous, and [0,1]×Q ×Q is compact, hence ϕ([0,1]×Q ×Q )= 1 2 1 2 1 2 1 2 conv(Q ∪Q ) is compact, so conv(Q ∪Q )= conv(Q ∪Q )) 1 2 1 2 1 2 n Since x ∈K , x = α x for x ∈Q , α ≥ 0, α +!+α = 1. But x is an extreme point, so x = x for ∑k=1 k k k k k 1 n k some k , which implies that x ∈Q ⊆ y +U ⊆ F +U , a contradiction ■ k k EL3370 Dual Spaces - 50 BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) In finite dimensions, the Krein-Milman theorem can be made more explicit: Theorem (Carathéodory) N Let X ⊆ ! . Every point x ∈ conv(X) can be written as a convex combination of at most N +1 points from X Proof. Let x ∈ conv(X). Then, x = λ x +!+λ x , where λ ,…,λ >0, λ +!+λ =1 and x ,…,x ∈ X . If 1 1 n n 1 n 1 n 1 n n> N +1, the points x − x ,…,x − x are linearly dependent, i.e., µ (x − x )+!+µ (x − x )=0 for some 2 1 n 1 2 2 1 n n 1 µ ,…,µ ∈ ! , not all zero. Defining µ = −µ −!−µ , we have that µ x +!+µ x =0 and µ +!+µ =0. 2 n 1 2 n 1 1 n n 1 n Since not all µ ’s are zero, there is at least one µ >0. Then, x = λ x +!+λ x −α(µ x +!+µ x )= i 1 1 n n 1 1 n n (λ −αµ )x +!+(λ −αµ )x . Pick α as the minimum of λ /µ , over the indices j for which µ >0; 1 1 1 n n n j j j note that α >0 and λ −αµ ≥0 for all j , but λ −αµ =0 for at least one index. Thus, j j j j x =(λ −αµ )x +!+(λ −αµ )x , where each coefficient is non-negative, their sum is one, and one of 1 1 1 n n n them is zero. Applying this procedure iteratively one can write x as a convex combination of at most N +1 points ■ EL3370 Dual Spaces - 51 BONUS: KREIN-MILMAN AND CARATHÉODORY THEOREMS (CONT.) Application: Optimal multisine spectrum In experimental design, one searches for an input cumulative spectrum Φopt (a non- decreasing function such that Φ(0)=0) such that the information matrix I of a parameter F d π θ ∈ ! is maximized in some sense, under a power constraint Φ(π )= dΦ(ω)=1. Now, ∫0 π I (Φ)≈ F(ω)dΦ(ω)∈ !d×d , F(ω): symmetric positive semi-definite matrix F ∫0 Note that, under the power constraint, I (Φ) lies in the convex hull of {F(ω):ω ∈[0,π ]}, F which has dimension d(d +1)/2 (since F(ω) is a symmetric matrix). Thus, by Carathéodory’s Theorem, I (Φopt ) is a convex combination of at most n= d(d +1)/2+1 F matrices F(ω ),…,F(ω ), i.e., for some λ ,…,λ ≥0, 1 d(d+1)/2+1 1 n opt π ⎡ ⎤ IF (Φ )= λ1F(ω1 )+!+λnF(ωn )= F(ω)⎣λ1δ(ω −ω1 )+!+λnδ(ω −ωn )⎦dω ∫0 where δ is the Dirac distribution. Therefore Φopt can be replaced by a “multisine” spectrum with frequencies at ω ,…,ω to obtain the same maximal information matrix! 1 n EL3370 Dual Spaces - 52 BONUS: POSITIVE LINEAR FUNCTIONALS Let X be a topological space. A linear functional F on C(X) is positive, denoted F ≥0, if F( f )≥0 whenever f (x)≥0 for all x ∈ X . Positive linear functionals are called Borel measures on X , and those of unit norm are called Borel probability measures on X Lemma (Cauchy-Schwarz inequality for positive linear functionals) 2 2 2 If F :C(X)→ ! is a positive linear functional, and f , g ∈ C(X), then [F( fg)] ≤ F( f )F(g ) 2 2 2 2 Proof. Since F[(λx + y) ]= λ F(x )+2λF(xy)+F( y )≥0 for all λ ∈ !, the discriminant 2 2 2 4[F( fg)] −F( f )F(g ) is non-positive ■ Lemma (characterization of positive linear functionals) A linear functional F :C(X)→ ! is positive iff F(1)= F , where 1: x ∈ C(X)!1 Proof. (⇒) If F ≥0, then F(1)≤ F , since 1 =1, and, by Cauchy-Schwarz, for every f ∈ C(X) of unit 2 2 2 2 2 2 norm, [F( f )] =[F(1 f )] ≤ F(1 )F( f )≤ F(1) F , since f ≤ f ≤1; taking the supremum over all f of 2 unit norm, F ≤ F(1) F , so in conclusion F(1)= F (⇐) If f ∈ C(X) satisfies f (x)≥0 for all x , and f ≤1, then also 1(x)− f (x)≥0, and 1− f ≤1, so F −F( f )= F(1)−F( f )= F(1− f )≤ F , so F( f )≥0. By linearity, F is positive ■ EL3370 Dual Spaces - 53 BONUS: POSITIVE LINEAR FUNCTIONALS (CONT.) Lemma (Jordan decomposition for linear functionals) Every l ∈ C(X)∗ can be written as l = l −l , where l ,l ∈ C(X)∗ are positive, and l = l + l + − + − + − ∗ ∗ Proof. Let S = {l ∈ C(X) :l(1)= l ≤1} and U = {l ∈ C(X) : l ≤1}; by Banach-Alaoglu, S and U are weak ∗ -compact. We will first show that U equals K = conv(S∪(−S)). If l ∈ K , then l =αl −(1−α)l , 1 2 with l ,l ∈ S and α ∈[0,1] (since S is convex, every convex combination in S∪(−S) can be reduced to 1 2 this form); hence l ≤α l +(1−α ) l ≤1, so K ⊆ U . Furthermore, K is weak ∗ -compact: indeed, define 1 2 ϕ :S × S ×[0,1]→U as ϕ(l ,l ,α)= αl −(1−α)l ; this map is continuous and S × S ×[0,1] is the product of 1 2 1 2 compact sets, hence it is compact, and K =R(ϕ) is compact. Suppose now that l ∈U \K . By the ∗ corollary on the separation of a convex set and a point, there is a weak -continuous functional f on ∗ C(X) such that f (l)> c and f (m)≤ c for all m ∈ K . This f is of the form f (m)= m(x) for some x ∈ C(X), and since K is symmetric (i.e., m ∈ K iff −m ∈ K ), m(x) ≤ c for all m ∈ K ˆ ∗ ˆ ˆ Let t ∈ X such that x(t) = max{ x(τ ) :τ ∈ X}, and define m ∈ C(X) as m(x)= x(t), hence m≥0 and mˆ =1, so x = mˆ (x) = sup m(x) . Thus, x ≤ c but l(x)> c , contradicting that l ≤1; hence K =U m∈S Now, take an l ∈ C(X)∗ ; assume w.l.o.g. that l =1. As l ∈U = conv(S∪(−S)), l =αl −(1−α)l for some 1 2 l ,l ∈ S and α ∈[0,1]. Define the positive functionals l =αl and l =(1−α)l , and note that l + l = 1 2 + 1 − 2 + − α l +(1−α) l ≤1= l = l −l ≤ l + l ■ 1 2 + − + − EL3370 Dual Spaces - 54 BONUS: POSITIVE LINEAR FUNCTIONALS (CONT.) Lemma (Extreme points of Borel probability measures) The extreme points of the set of Borel probability measures on X , Δ(X), are the delta measures δ for x ∈X , where δ ( f )= f (x) for all f ∈C(X) x x Proof (Barvinok, 2002, and Dunford&Schwartz, 1958) Let x ∈ X . Assume that δ =αl +(1−α)l , where l ,l ∈ Δ(X) and α ∈(0,1). If f ∈ C(X) satisfies f (x)=1 x 1 2 1 2 and f ( y)≤1 for all y ∈ X , then l ( f )= l (1−(1− f ))= l (1)−l (1− f )≤1 for i =1,2, but δ ( f )= f (x)=1, so i i i i x l ( f )=1. Hence, l ,l agree with δ on every function with its maximum at x . Now, take any f ∈ C(X), i 1 2 x and write it as f = f − f , where f ( y)= min{ f (x), f ( y)} and f ( y)= min{0, f (x)− f ( y)}. Then, f , f 1 2 1 2 1 2 attain their maxima at x , so l ( f )= l ( f )−l ( f )= f (x)− f (x)= f (x)= δ ( f ), i.e., δ is extreme i i 1 i 2 1 2 x x Now, let A = {δ : x ∈X}⊆ Δ(X). Since Δ(X) is convex and weak ∗ -closed, conv(A)⊆ Δ(X) (the left side is x ∗ ∗ the weak -closure). If l ∈C(X) \conv(A), by the corollary on the separation between a convex set and a point, there is an f ∈C(X) such that l( f )> c but δ ( f )= f (x)≤ c for all x ∈X , so l ∉Δ(X) . Therefore, x ∗ ∗ ∗ conv(A)= Δ(X). On the other hand, A is weak -closed in C(X) (why?), and hence weak -compact by Banach-Alaoglu, so the lemma on slide 50 implies that all extreme points of Δ(X) are in A ■ Corollary. The extreme points of U = {l ∈ C(X)∗ : l ≤1} are δ and −δ for x ∈X x x EL3370 Dual Spaces - 55 BONUS: CONVEX FUNCTIONS AND JENSEN’S INEQUALITY Definitions Let X be a normed space, and f : X → ! - f is convex if f (λx +(1−λ)y)≤ λ f (x)+(1−λ)f ( y) for all x, y ∈ X and λ ∈[0,1] - epi f = {(x, y)∈ X ×!: y ≥ f (x)}: epigraph of f : X → ! A function is convex iff its epigraph is a convex set, so it can be represented as the intersection of closed half-spaces that contain it. Now, for all a,b ∈ ! and x∗ ∈ X ∗, a ≤< x,x∗ >+by whenever y ≥ f (x) ∗ ⇔ a ≤< x,x >+bf (x) ∗ If H = {(x, y)∈ X ×!:a ≤< x,x >+by} contains epi f , ∗ then b cannot be zero; otherwise f would be undefined when a >< x,x >. Thus, every ∗ closed half-space containing epi f has the form H = {(x, y)∈X × !:c+ < x,x >≤ y}, and f (x)= sup c+ < x,x∗ > (c ,x∗ ): c+ EL3370 Dual Spaces - 56 BONUS: CONVEX FUNCTIONS AND JENSEN’S INEQUALITY (CONT.) Jensen’s Inequality Let X be a set, and E a Borel probability measure on a set of functions f : X → !. Also, let g: X → ! be such that E(g) is well-defined, and let f :! → ! be a convex function. Then E( f ! g)≥ f (E(g)) if the left hand side is well-defined Proof. Take a,b ∈ ! such that f ( y)≥ ay +b for all y ∈ ! . Then f (g(x))≥ ag(x)+b for all , so applying E yields E{ f ! g}≥ aE{g}+b. Taking the supremum over all a,b ∈ ! such that f ( y)≥ ay +b for every y ∈ ! finally gives E{ f ! g}≥ sup aE{g}+b = f (E{g}) ■ (a,b): a+by≤ f ( y ) for all y∈X Remark. The Borel probability measure E can be, e.g., the Riemann/Lebesgue integral, or the mathematical expectation operator with respect to a probability measure EL3370 Dual Spaces - 57 BONUS: STONE-WEIERSTRASS THEOREM Weierstrass’ Theorem, on the approximation of continuous functions on a real interval by polynomials, was considerably generalized by M.H. Stone in two directions: - it applies to continuous functions on an arbitrary compact Hausdorff space, and - instead of polynomials, it allows more general algebras of functions We will provide a functional-analytic proof of this result, based on the following lemma: Lemma (spanning criterion) ∗ Let V be a normed space, x ∈V , and Y ⊆V . Then, x ∈ clin Y iff every f ∈V that vanishes on Y also vanishes on x (i.e., if f ( y)=0 for all y ∈Y , then f (x)=0) Proof (⇒) If f ( y)=0 for all y ∈Y , then by linearity f ( y)=0 for all y ∈ lin Y , and by continuity f (x)=0 since x ∈ clin Y (⇐) Assume x ∉ clin Y , and define the functional f on lin{clin Y + x} by f ( y +αx)=α for all y ∈ clin Y and α . Let d = inf x − y , which is >0 since clin Y is closed. Then, y +αx = α α −1 y + x ≥ d α , so y∈clin Y −1 ∗ f ≤ d , and Corollary 1 of Hahn-Banach provides an extension F ∈V of f that vanishes on Y but not on x ■ EL3370 Dual Spaces - 58 BONUS: STONE-WEIERSTRASS THEOREM (CONT.) Theorem (Stone-Weierstrass) Let X be a compact Hausdorff space, and let E be a subalgebra of C(X), i.e., a linear subspace of C(X) such that if f , g ∈ E , also fg ∈ E . In addition, assume that 1 ∈ E , and that E separates points of X , i.e., for every pair x, y ∈ X , x ≠ y , there is an f ∈ E such that f (x)≠ f ( y). Then, E is dense in C(X) Proof (adapted from version due to Louis de Branges) ∗ By the spanning criterion, E is dense in C(X) iff for every l ∈ C(X) , l(x)=0 for all x ∈ E implies that ∗ l =0. Assume the contrary, and let U be the set of l ∈ C(X) of norm ≤1 that vanish on E ; by Banach- Alaoglu, U is weak ∗ -compact. By Krein-Milman, U has an extreme point, say, l , which should have unit norm (why?) ∗ Let g ∈ E be such that 0< g(t)<1 for all t ∈ X . Since E is an algebra, the functionals lʹ,lʹʹ ∈ C(X) defined by lʹ( f )= l(gf ) and lʹʹ( f )= l((1− g)f ) also vanish in E . Write l as l = l −l , where l ,l ≥ 0, and + − + − also lʹ and lʹʹ as lʹ( f )= l(gf )= l (gf )−l (gf )=: lʹ( f )−lʹ( f ) and lʹʹ( f )= l ((1− g)f )− l ((1− g)f )=: + − + − + − l′′( f )− l′′( f ), where l′ ,l′ ,l′′,l′′≥ 0, and note that l = l + l = l (1)+l (1)= lʹ(1)+lʹʹ(1)+lʹ(1)+lʹʹ(1)= + − + − + − + − + − + + − − lʹ + lʹʹ + lʹ + lʹʹ , while l = lʹ+lʹʹ ≤ lʹ + lʹʹ ≤ lʹ + lʹʹ + lʹ + lʹʹ , so combining these inequalities we see + + − − + + − − that they should be all equalities, and in particular l = lʹ + lʹʹ EL3370 Dual Spaces - 59 BONUS: STONE-WEIERSTRASS THEOREM (CONT.) Proof (cont.) Let a = lʹ ≥ min g(x) l >0 and similarly b = lʹʹ >0, so a+b = l =1. Then, l = a(lʹ/a)+b(lʹʹ/b) shows t∈X that l is a convex combination of lʹ/a and lʹʹ/b, both of which have unit norm and vanish in E . Since l is an extreme point, we have that l = lʹ/a = lʹʹ/b, i.e., l ([1− g/a]f )=0 for all f ∈ C(X), or ± l (g) l ( f )= l (gf ), where l ( f )= l ( f )+ l ( f ) + − Using l (g) l ( f )= l (gf ), we will now show that all functions in the nullspace of l in E (which is a closed hyperplane) have a common zero. Indeed, every g ∈Ker l ∩E should have a zero in X , because −1 −1 −1 otherwise g ∈C(X), so taking f = g gives l (g) l (g )= l (1)> 0, which contradicts the assumption that l (g)= 0. Now, if g ,…, g ∈Ker l ∩E , then they share a common zero, since otherwise 1 n g = g2 +!+ g2 ∈E is strictly positive, while l (g)= l (g2 )+!+ l (g2 )=[ l (g )]2 +!+[ l (g )]2 = 0, which 1 n 1 n 1 n is a contradiction. Finally, for every g ∈E , let X = {x ∈X : g(x)= 0}; X is a closed set in X , and every g g finite intersection X X is non-empty, so by the compactness of X , X , i.e., all g ∩!∩ g ∩ g Ker l g ≠ ∅ 1 n ∈ functions in Ker l ∩E share a common zero EL3370 Dual Spaces - 60 BONUS: STONE-WEIERSTRASS THEOREM (CONT.) Proof (cont.) ∗ The kernel of any h∈E for which h( fg)= h( f )h(g) (called a homomorphism, such as l ) satisfies another important property: it is an ideal in E , i.e., it is a linear subspace of E such that if f ∈E and g ∈Ker h, then fg ∈Ker h (since h( fg)= h( f )h(g)= 0); furthermore, it is a maximal ideal, that is, a proper subset of E which is not contained in a larger proper ideal of E (since Ker h has co-dimension 1, so adding an extra dimension would yield E ) Other homomorphisms on E include δ for every x ∈X . Since E separates points on X , x E Ker δ = {g ∈E : g(x)= 0} ≠ Ker δ for x ≠ y (indeed: pick a function g ∈E such that g(x)= 0 and x E y E g( y)= 1; this g belongs to Ker δ but not to Ker δ ) x E y E Therefore, Ker l ∩E ⊆ Ker δ for some x ∈X corresponding to the common zero of the functions in x E Ker l ∩E , but since Ker l is a maximal ideal in E , we have that Ker l ∩E = Ker δ , i.e., x E l ( f )= αδ ( f )= α f (x) for all f ∈E , with α = 1 due to the positivity of l x Since δ is an extreme point of the unit ball of C(X)∗ , δ = l = l + l implies that l = α δ , so l = cδ . On x x + − ± ± x x the other hand, 1∈E , so cδ (1)= l(1)= 0, thus c 0 and l = 0, which contradicts the assumption that x = l = 1, hence E is dense in C(X) ■ EL3370 Dual Spaces - 61