54 2.7 CHAPTER 2. WEAK TOPOLOGIES AND REFLEXIVITY The Principle of Local Reflexivity In this section we we proof a result of J. Lindenstrauss and H. Rosenthal [LR] which states that for a Banach space X the finite dimensional subspaces of the bidual X ⇤⇤ are in a certain sense similar to the finite dimensional subspaces of X. Theorem 2.7.1. [LR] [The Principle of Local Reflexivity] Let X be a Banach space and let F ⇢ X ⇤⇤ and G ⇢ X ⇤ be finite dimensional subspaces of X ⇤⇤ and X ⇤ respectively. Then, given " > 0, there is a subspace E of X containing F \ X (we identify X with its image under the canonical embedding) with dim E = dim F and an isomorphism T : F ! E with kT k · kT 1 k 1 + " such that (2.11) (2.12) T (x) = x if x 2 F \ X and hx⇤ , T (x⇤⇤ )i = hx⇤⇤ , x⇤ i if x⇤ 2 G, x⇤⇤ 2 F . We need several Lemmas before we can prove Theorem 2.7.1. The first one is a corollary the Geometric Hahn-Banach Theorem Proposition 2.7.2. (Variation of Geometrical Version of the Theorem of Hahn Banach) Assume that X is a Banach space and C ⇢ X is convex with C 6= ; and let x 2 X \ C (so x could be in the boundary of C). Then there exists an x⇤ 2 X ⇤ so that <hx⇤ , zi < 1 = hx⇤ , xi for all z 2 C 0 , and, if moreover C is absolutely convex (i.e. if ⇢x 2 C for all x 2 C and ⇢ 2 K, with |⇢| 1), then |hx⇤ , zi| < 1 = hx⇤ , xi for all z 2 C 0 . Lemma 2.7.3. Assume T : X ! Y is a bounded linear operator between the Banach spaces X and Y and assume that T (X) is closed. Suppose that for some y 2 Y there is an x⇤⇤ 2 X ⇤⇤ with kx⇤⇤ k < 1, so that T ⇤⇤ (x⇤⇤ ) = y. Then there is an x 2 X, with kxk < 1 so that T (x) = y. Proof. We first show that there is an x 2 X so that T (x) = y. Assume this where not true, then we could find by the Hahn-Banach Theorem (Corollary 1.4.5) an element y ⇤ 2 Y ⇤ so that y ⇤ (z) = 0 for all z 2 T (X) and hy ⇤ , yi = 1 2.7. THE PRINCIPLE OF LOCAL REFLEXIVITY 55 (T (X) is closed). But this yields hT ⇤ (y ⇤ ), xi = hy ⇤ , T (x)i = 0, for all x 2 X, and, thus, T ⇤ (y ⇤ ) = 0. Thus 0 = hx⇤⇤ , T ⇤ (y ⇤ )i = hT ⇤⇤ (x⇤⇤ ), y ⇤ i = hy, y ⇤ i = 1, which is a contradiction. Secondly assume that y 2 T (X) \ T (BX ). Since T is surjective onto its (closed) image Z = T (X) it follows from the Open Mapping Theorem that T (BX ) is open in Z, and we can use the geometric version of the HahnBanach Theorem 2.2.3 , to find z ⇤ 2 Z ⇤ so that hz ⇤ , T (x)i < 1 = hz ⇤ , yi for all x 2 BX . Again by the Theorem of Hahn-Banach (Corollary 1.4.4) we can extend z ⇤ to an element y ⇤ in Y ⇤ . It follows that kT ⇤ (y ⇤ )k = sup hT ⇤ (y ⇤ ), xi = sup hz ⇤ , T (x)i 1, x2BX x2BX and thus, since kx⇤⇤ k < 1, it follows that |hy ⇤ , yi| = |hy ⇤ , T ⇤⇤ (x⇤⇤ )i| = |hx⇤⇤ , T ⇤ (y ⇤ )i| < 1, which is a contradiction. Lemma 2.7.4. Let T : X ! Y be a bounded linear operator between two Banach spaces X and Y with closed range, and assume that F : X ! Y has finite rank. Then T + F also has closed range. Proof. Assume the claim is not true. Put S = T + K and consider the map S : X/N (S) ! Y, x + N (S) ! S(x) which is a well defined linear bounded Operator, which by Proposition 1.3.11 cannot be an isomorphism onto its image. Therefore we can choose a sequence (zn ) in X/N (S) so that we can choose a sequence (xn ) so that lim (T + F )(xn ) = 0 and dist(xn , N (T + F )) n!1 1. Since the sequence (F (xn ) : n 2 N) is a bounded sequence in a finite dimensional space, we can, after passing to a subsequence, assume that (F (xn ) : n 2 N) converges to some y 2 Y and, hence, lim T (xn ) = n!1 y. 56 CHAPTER 2. WEAK TOPOLOGIES AND REFLEXIVITY Since T has closed range there is an x 2 X, so that T (x) = y. Using again the equivalences in Proposition 1.3.11 and the fact that T (xn ) ! y = T (x), if n % 1, it follows for some constant C > 0 that xn , N (T )) lim CkT (x lim dist(x n!1 n!1 xn )k = 0, and, thus, y F (x) 2 F (N (T )), F (x) = lim F (xn ) n!1 so we can write y F (x) as F (x) = F (u), where u 2 N (T ). y Thus lim dist(xn lim kF (xn ) n!1 u, N (T )) = 0 and x n!1 F (x) F (u)k = 0. F |N (T ) has also closed range, Proposition 1.3.11 yields (C being some positive constant) lim sup dist(xn x u, N (F )\N (T )) lim sup CkF (xn ) F (x) F (u)k = 0. n!1 n!1 Since T (x+u) = y = F (x+u) (by choice of u), and thus (T +F )(x+u) = 0 which means that x + u 2 N (T + F ). Therefore lim sup dist(xn , N (T + F )) = lim sup dist(xn x u, N (T + F )) lim sup dist(xn x u, N (T ) \ N (F )) = 0. n!1 n!1 n!1 But this contradicts our assumption on the sequence (xn ). Lemma 2.7.5. Let X be a Banach space, A = (ai,j )im,jn an m be n matrix and B = (bi,j )ip,jn a p by n matrix, and that B has only real entries (even if K = C). ⇤⇤ Suppose that y1 , . . . ym 2 X, y1⇤ , . . . yp⇤ 2 X ⇤ , ⇠1 , . . . ⇠p 2 R, and x⇤⇤ 1 , . . . xn 2 BX ⇤⇤ satisfy the following equations: (2.13) n X ai,j x⇤⇤ j = yi , for all i = 1, 2 . . . m, and j=1 (2.14) D yi⇤ , n X j=1 E bi,j x⇤⇤ = ⇠i , for all i = 1, 2 . . . p. j 2.7. THE PRINCIPLE OF LOCAL REFLEXIVITY 57 Then there are vectors x1 , . . . xn 2 BX satisfying: (2.15) n X ai,j xj = yi , for all i = 1, 2 . . . m, and j=1 (2.16) D yi⇤ , n X j=1 E bi,j xj = ⇠i , for all i = 1, 2 . . . p. Proof. Recall from Linear Algebra that we can write the matrix A as a product A = U P V , where U and V are invertible and P is of the form ✓ ◆ Ir 0 P = , 0 0 where r is the rank of A and Ir the identity on Kr . For a general s by t matrix C = (ci,j )is,jt consider the operator TC : `t1 (X) ! `s1 (X), (x1 , x2 , . . . xt ) 7! t ⇣X j=1 ⌘ ci,j xj : i = 1, 2 . . . m . If s = t and if C is invertible then TC is an isomorphism. Also if C (1) and C (2) are two matrices so that the number of columns of C (1) is equal to the number of rows of C (2) one easily computes that TC (1) C (2) = TC (1) TC (2) . Secondly it is clear that TP is a closed operator (P defined as above), since TP is simply the projection onto the first r coordinates in `n1 (X). It follows therefore that TA = TU TP TV is an operator with closed range. Secondly define the operator SA : `n1 (X) ! `m 1 (X) (x1 , . . . xn ) 7! `p1 , TA (x1 , . . . xn ), ⇣D yi⇤ , n X bi,j xj j=1 E⌘p i=1 ! . SA can be written as the sum of SA and a finite rank operator and has therefore also closed range by Lemma 2.7.4. ⇤⇤ is the operator Since the second adjoint of SA ⇤⇤ ⇤⇤ SA : `n1 (X ⇤⇤ ) ! `m 1 (X ) ⇤⇤ (x⇤⇤ 1 , . . . xn ) 7! `p1 , ⇤⇤ TA⇤⇤ (x⇤⇤ 1 , . . . xn ), ⇣D yi⇤ , n X j=1 bi,j x⇤⇤ j E⌘p i=1 ! 58 CHAPTER 2. WEAK TOPOLOGIES AND REFLEXIVITY with TA⇤⇤ : `n1 (X ⇤⇤ ) ! ⇤⇤ `m 1 (X ), ⇤⇤ ⇤⇤ (x⇤⇤ 1 , x2 , . . . xn ) 7! t ⇣X j=1 ⌘ ai,j x⇤⇤ j : i = 1, 2 . . . m , our claim follows from Lemma 2.7.3. Lemma 2.7.6. Let E be a finite dimensional space and (xi )N i=1 is an "-net of SE for some 0 < " < /3. If T : E ! E is a linear map so that (1 Then ") kT (xj )k (1 + "), for all j = 1, 2 . . . N . 1 3" 1+" kxk kT (x)k kxk, for all x 2 E, 1 " 1 " and thus kT k · kT 1 k (1 + ")2 . (1 ")(1 3") Proof. Let x 2 X. W.l.o.g. we can assume that kxk = 1. Pick j N so that kx xj k ". Then kT (x)k kT (x) T (xj )k + kT (xj )k "kT k + 1 + ", and thus 1+" , 1 " which implies the second inequality of our claim. We also have kT k kT (x)k kT (xj )k kT (x xj )k 1 " " 1+" 1 = 1 " 2" + "2 " 1 " "2 = 1 1 which proves our claim. We are now ready to proof Theorem 2.7.1. Proof of Theorem 2.7.1. Let F ⇢ X ⇤⇤ and G ⇢ X ⇤ be finite dimensional )2 subspaces, and let 0 < " < 1. Choose > 0, so that (1 (1+ )(1 3 ) < ", and a N ⇤⇤ N -net (x⇤⇤ j )j=1 of SF . It can be shown that (xj )j=1 generates all of F , but we can also simply assume that without loss of generality, since we can add a basis of F . Let N X N S : R ! F, (⇠1 , ⇠2 . . . ⇠N ) 7! ⇠j x⇤⇤ j , j=1 3" , 2" 2.7. THE PRINCIPLE OF LOCAL REFLEXIVITY 59 and note that S is surjective. Put H = S 1 (F \ X), and let (a(i) : i = 1, 2 . . . m) be a basis of H, write a(i) as a(i) = (ai,1 , ai,2 , . . . ai,N ), and define A to be the m by N matrix A = (ai,j )im,jN . For i = 1, 2 . . . m put yi = S(a(i) ) = N X j=1 ai,j x⇤⇤ j 2 F \ X, ⇤ choose x⇤1 , x⇤2 , . . . , x⇤N 2 SX ⇤ so that hx⇤⇤ , and pick basis j , xj i > 1 ⇤ ⇤ ⇤ {g1 , g2 , . . . g` } of G. Consider the following system of equations in N unknowns z1⇤⇤ , z2⇤⇤ , ⇤⇤ in X ⇤⇤ : . . . , zN N X ai,j zj⇤⇤ = yi for i = 1, 2 . . . m j=1 hzj⇤⇤ , x⇤j i hzj⇤⇤ , gk i ⇤ = hx⇤⇤ j , xj i for j = 1, 2 . . . N and ⇤ = hx⇤⇤ j , gk i for j = 1, 2 . . . N and k = 1, 2 . . . `. By construction zj⇤⇤ = x⇤⇤ j , j = 1, 2 . . . N , is a solution to these equations. Since kx⇤⇤ k = 1 < 1 + , for j = 1, 2 . . . N , we can use Lemma 2.7.5 and j find x1 , x2 , . . . xN 2 X, with kxj k = 1 < 1 + , for j = 1, 2 . . . N , which solve above equations. Define N X N S1 : R ! X, (⇠1 , ⇠2 . . . ⇠N ) 7! ⇠j xj . j=1 We claim that the null space of S isPcontained in the space of S1 . Indeed Pnull N ⇤⇤ if we assumed that ⇠ 2 KN , and N ⇠ x , but x j j j=1 j=1 j 6= 0, then there is an i 2 {1, 2, . . . N } so that D x⇤i , N X j=1 E x⇤⇤ 6 0, = j but since hx⇤ ⇤j , x⇤i i = hxj , x⇤i i this is a contradiction. It follows therefore that we can find a linear map T : F ! X so that S1 = T S. Denoting the standard basis of RN by (ei )iN we deduce that xi = S1 (ei ) = T S(ei ) = T (x⇤⇤ j ), and thus 1 + > kxj k = kT (x⇤⇤ j )k ⇤ |hx⇤j , xj i| = hx⇤⇤ j , xj i| > 1 . 60 CHAPTER 2. WEAK TOPOLOGIES AND REFLEXIVITY By Lemma 2.7.6 and the choice of it follows therefore that kT k · kT 1 k 1 + ". P (i) Note that for ⇠ 2 H = S 1 (F \ X), say ⇠ = m i=1 i a , we compute S1 (⇠) = m X i=1 i S1 (a (i) )= m X i i=1 N X j=1 ai,j xj = m X i=1 i N X j=1 ai,j x⇤⇤ j = m X i S(a (i) i=1 We deduce therefore for x 2 F \ X, that T (x) = x. Finally from the third part of the system of equations it follows, that ⇤ ⇤ ⇤ hx⇤ , T (x⇤⇤ j )i = hx , xj i = hx , xj i, for all j = 1, 2 . . . N and x 2 G , and, thus (since the x⇤⇤ j generate all of F ), that hx⇤ , T (x⇤⇤ )i = hx⇤⇤ , x⇤ i, for all x⇤⇤ 2 F and x⇤ 2 G. Exercises 1. Prove Proposition 2.7.2. ) = S(⇠).