Você está na página 1de 18

# REAL ANALYSIS LECTURE NOTES

311

16. Banach Spaces II Theorem 16.1 (Open Mapping Theorem). Let X, Y be Banach spaces, T L(X, Y ). If T is surjective then T is an open mapping, i.e. T (V ) is open in Y for all open subsets V X.
X Y Proof. For all > 0 let B = {x X : kxkX < } X, B = X {y Y : kykY < } Y and E = T (B ) Y. The proof will be carried out by proving the following three assertions. Y (1) There exists > 0 such that B E for all > 0. Y (2) For the same > 0, B E , i.e. we may remove the closure in assertion 1. (3) The last assertion implies T is an open mapping. S En , the Baire category Theorem 15.16 implies there exists 1. Since Y = n1

n such that E n 6= , i.e. there exists y E n and > 0 such that B Y (y, ) E n . Suppose ky 0 k < then y and y + y 0 are in B Y (y, ) E n hence there exists X x0 , x Bn such that kT x0 (y + y 0 )k and kT x yk may be made as small as we please, which we abbreviate as follows kT x0 (y + y 0 )k 0 and kT x yk 0.

## Hence by the triangle inequality, kT (x0 x) y 0 k = kT x0 (y + y 0 ) (T x y)k kT x0 (y + y 0 )k + kT x yk 0

X with x0 x B2n . This shows that y 0 E2n which implies B Y (0, ) E2n . Since the map : Y Y given by (y) = 2n y is a homeomorphism, (E2n ) = E Y Y Y and (B (0, )) = B (0, 2n ), it follows that B E where 2n > 0. Y 2. Let be as in assertion 1., y B and 1 (kyk /, 1). Choose {n } n=2 P Y X (0, ) such that n < 1. Since y B1 E1 = T B1 by assertion 1. n=1 X there exists x1 B1 such that ky T x1 k < 2 . (Notice that ky T x1 k can be Y X made as small as we please.) Similarly, since yT x1 B2 E2 = T B2 there X exists x2 B2 such that ky T x1 T x2 k < 3 . Continuing this way inductively, X there exists xn Bn such that

(16.1) Since
P

ky kxn k <
P

k=1

n X

## T xk k < n+1 for all n N.

P X xn exists and kxk < 1, i.e. x B1 . Passing

n=1

n=1

n < 1, x

n=1

X to the limit in Eq. (16.1) shows, ky T xk = 0 and hence y T (B1 ) = E1 . X Therefore we have shown B E1 . The same scaling argument as above then X shows B E for all > 0. 3. If x V o X and y = T x T V we must show that T V contains a ball B Y (y, ) = T x + B Y for some > 0. Now B Y (y, ) = T x + B Y T V i B Y T V T x = T (V x). Since V x is a neighborhood of 0 X, there exists X Y X > 0 such that B (V x) and hence by assertion 2., B T B T (V x) Y and therefore B (y, ) T V with := .

312

BRUCE K. DRIVER

Corollary 16.2. If X, Y are Banach spaces and T L(X, Y ) is invertible (i.e. a bijective linear transformation) then the inverse map, T 1 , is bounded, i.e. T 1 L(Y, X). (Note that T 1 is automatically linear.) Theorem 16.3 (Closed Graph Theorem). Let X and Y be Banach space T : X Y linear is continuous i T is closed i.e. (T ) X Y is closed. Proof. If T continuous and (xn , T xn ) (x, y) X Y as n then T xn T x = y which implies (x, y) = (x, T x) (T ). Conversely: If T is closed then the following diagram commutes (T ) @ @ 2 @ R @ -Y X T where (x) := (x, T x). The map 2 : X Y X is continuous and 1 |(T ) : (T ) X is continuous bijection which implies 1 |1 ) is bounded by the open mapping Theorem 16.1. (T Hence T = 2 1 |1 ) is bounded, being the composition of bounded operators. (T As an application we have the following proposition. Proposition 16.4. Let H be a Hilbert space. Suppose that T : H H is a linear (not necessarily bounded) map such that there exists T : H H such that Then T is bounded. hT x, Y i = hx, T Y i x, y H.

Proof. It suces to show T is closed. To prove this suppose that xn H such that (xn , T xn ) (x, y) H H. Then for any z H, On the other hand limn hT xn , zi = hy, zi as well and therefore hT x, zi = hy, zi for all z H. This shows that T x = y and proves that T is closed. Here is another example. Example 16.5. Suppose that M L2 ([0, 1], m) is a closed subspace such that each element of M has a representative in C([0, 1]). We will abuse notation and simply write M C([0, 1]). Then (1) There exists A (0, ) such that kf k Akf kL2 for all f M. (2) For all x [0, 1] there exists gx M such that Moreover we have kgx k A. (3) The subspace M is nite dimensional and dim(M) A2 . Proof. 1) I will give a two proofs of part 1. Each proof requires that we rst show that (M, k k ) is a complete space. To prove this it suces to show M is a closed subspace of C([0, 1]). So let {fn } M and f C([0, 1]) such that kfn f k 0 as n . Then kfn fm kL2 kfn fm k 0 as m, n , and since M is closed in L2 ([0, 1]), L2 limn fn = g M. By passing to a f (x) = hf, gx i for all f M. hT xn , zi = hxn , T zi hx, T zi = hT x, zi as n .

## REAL ANALYSIS LECTURE NOTES

313

subsequence if necessary we know that g(x) = limn fn (x) = f (x) for m - a.e. x. So f = g M. i)Let i : (M, k k ) (M, k k2 ) be the identity map. Then i is bounded and bijective. By the open mapping theorem, j = i1 is bounded as well. Hence there exists A < such that kf k = kj(f )k A kf k2 for all f M. ii) Let j : (M, k k2 ) (M, k k ) be the identity map. We will shows that j is a closed operator and hence bounded by the closed graph theorem. Suppose that fn M such that fn f in L2 and fn = j(fn ) g in C([0, 1]). Then as in the rst paragraph, we conclude that g = f = j(f ) a.e. showing j is closed. Now nish as in last line of proof i). 2) For x [0, 1], let ex : M C be the evaluation map ex (f ) = f (x). Then |ex (f )| |f (x)| kf k Akf kL2 which shows that ex M . Hence there exists a unique element gx M such that f (x) = ex (f ) = hf, gx i for all f M. Moreover kgx kL2 = kex kM A. 3) Let {fj }n be an L2 orthonormal subset of M. Then j=1 A2 kex k2 = kgx k2 2 M L
n X j=1

|hfj , gx i|2 =

n X j=1

|fj (x)|2

and integrating this equation over x [0, 1] implies that n n XZ 1 X A2 |fj (x)|2 dx = 1=n
j=1 0 j=1

which shows that n A2 . Hence dim(M) A2 . Remark 16.6. Keeping the notation in Example 16.5, G(x, y) = gx (y) for all x, y [0, 1]. Then Z 1 f (x) = ex (f ) = f (y)G(x, y)dy for all f M.
0

The function G is called the reproducing kernel for M. The above example generalizes as follows.

Proposition 16.7. Suppose that (X, M, ) is a nite measure space, p [1, ) and W is a closed subspace of Lp () such that W Lp ()L (). Then dim(W ) < . Proof. With out loss of generality we may assume that (X) = 1. As in Example 16.5, we shows that W is a closed subspace of L () and hence by the open mapping theorem, there exists a constant A < such that kf k A kf kp for all f W. Now if 1 p 2, then kf k A kf kp A kf k2 and if p (2, ), then kf kp kf k2 kf k kf kp kf k2
2/p p 2 p2

kf k

12/p

## or equivalently, 12/p 2/p kf k2 A kf kp

314

BRUCE K. DRIVER

Let / P D := (Q + iQ) and E = cD Ec . Then (E) = 0 and for x E, N 2 2 / n=1 cn fn (x) B |c| for all c D. By continuity it then follows for x E that 2 N X 2 cn fn (x) B 2 |c| for all c CN .
n=1

from which we learn that kf kp A12/p kf k2 and therefore that kf k AA12/p kf k2 so that in any case there exists a constant B < such that kf k B kf k2 . P Let {fn }N be an orthonormal subset of W and f = N cn fn with cn C, n=1 n=1 then 2 N N X X cn fn B 2 |cn |2 B 2 |c|2 n=1 n=1 PN 2 2 N where |c| := n=1 |cn | . For each c C , there is an exception set Ec such that for x Ec , / 2 N X cn fn (x) B 2 |c|2 .
n=1 N

## and therefore that

Taking cn = fn (x) in this inequality implies that N 2 N X X 2 |fn (x)| B 2 |fn (x)|2 for all x E /
n=1 n=1 n=1 N X

## |fn (x)| B 2 for all x E. /

Integrating this equation over x then implies that N B 2 , i.e. dim(W ) B 2 . Theorem 16.8 (Uniform Boundedness Principle). Let X and Y be a normed vector spaces, A L(X, Y ) be a collection of bounded linear operators from X to Y, F = FA = {x X : sup kAxk < } and
AA

(16.2)

R = RA = F c = {x X : sup kAxk = }.
AA AA

(1) If sup kAk < then F = X. (2) If F is not meager, then sup kAk < .
AA

AA AA

## sup kAxk < for all x X then sup kAk < .

AA AA

(4) If X is a Banach space, then sup kAk = i R is residual. In particular if sup kAk = then sup kAxk = for x in a dense subset of X.
AA AA

Proof. 1. If M := sup kAk < , then sup kAxk M kxk < for all x X
AA AA

showing F = X.

## REAL ANALYSIS LECTURE NOTES

315

2. For each n N, let En X be the closed sets given by \ En = {x : sup kAxk n} = {x : kAxk n}.
AA AA

Then F = En which is assumed to be non-meager and hence there exists n=1 an n N such that En has non-empty interior. Let Bx () be a ball such that Bx () En . Then for y X with kyk = we know x y Bx () En , so that Ay = Ax A(x y) and hence for any A A, kAyk kAxk + kA(x y)k n + n = 2n. Hence it follows that kAk 2n/ for all A A, i.e. sup kAk 2n/ < .
AA

3. If X is a Banach space, F = X is not meager by the Baire Category Theorem 15.16. So item 3. follows from items 1. and 2 and the fact that F = X i sup kAxk < for all x X.
AA

## 4. Item 3. is equivalent to F is meager i sup kAk = . Since R = F c , R is

AA AA

residual i F is meager, so R is residual i sup kAk = . Example 16.9. Suppose that lim
N X

{cn } n=1

1

n=1

## an cn exists in C for all a

PN Proof. Let fN 1 be given by fN (a) = n=1 an cn and set MN := max {|cn | : n = 1, . . . , N } . Then |fN (a)| MN kak
1

Then c

and by taking a = ek with k such MN = |ck | , we learn that kfN k = MN . Now by assumption, limN fN (a) exists for all a 1 and in particular, sup |fN (a)| < for all a
N 1

## So by the Theorem 16.8, > sup kfN k = sup MN = sup {|cn | : n = 1, 2, 3, . . . } .

N N

16.1. Applications to Fourier Series. Let T = S 1 be the unit circle in S 1 and m denote the normalized arc length measure on T. So if f : T [0, ) is measurable, then Z Z Z 1 f (w)dw := f dm := f (ei )d. 2 T T

316

BRUCE K. DRIVER

Also let n (z) = z n for all n Z. Recall that {n }nZ is an orthonormal basis for L2 (T ). For n N let n n n X X X Z k k hf, n ik (z) = hf, n iz = f (w)w dw z k sn (f, z) :=
k=n

f (w)
T k=n

k=n

where dn () :=

Pn

n X

wk z k

k=n

k=n

dw =

f (w)dn (z w)dw

k=n n X

k =

n+1 n 1

## with the convention that

n X n+1 n n+1 n 1k . |=1 = lim = 2n + 1 = 1 1 1 k=n

Writing = e , we nd Dn () := dn (ei ) = ei(n+1) ein ei(n+1/2) ei(n+1/2) = i 1 e ei/2 ei/2 1 sin(n + 2 ) = . sin 1 2

Recall by Hilbert space theory, L2 (T ) limn sn (f, ) = f for all f L2 (T ). We will now show that the convergence is not pointwise for all f C(T ) L2 (T ). Proposition 16.10. For each z T, there exists a residual set Rz C(T ) such that supn |sn (f, z)| = for all f Rz . Recall that C(T ) is a complete metric space, hence Rz is a dense subset of C(T ). Proof. By symmetry considerations, it suces to take z = 1 T. Let n f := sn (f, 1). Then Z Z Z |f (w)dn (w)| dw kf k |dn (w)| dw |n f | = f (w)dn (w)dw
T T T

showing

kn k

Since C(T ) is dense in L1 (T ), there exists fk C(T, R) such that fk (w) sgndk (w) in L1 . By replacing fk by (fk 1) (1) we may assume that kfk k 1. It now follows that Z |n fk | fk (w)dn (w)dw kn k kfk k T R and passing to the limit as k implies that kn k T |dn (w)| dw. Hence we have shown that Z Z Z sin(n + 1 ) 1 2 dn (ei ) d = 1 (16.3) kn k = |dn (w)| dw = sin 1 d. 2 2
T 2

## REAL ANALYSIS LECTURE NOTES

317

Since sin x =

Since

for all x 0. Since sin x is odd, |sin x| |x| for all x R. Using this in Eq. (16.3) implies that Z Z sin(n + 1 ) 1 2 d = 2 sin(n + 1 ) d . kn k 1 2 0 2 2 Z 2 sin(n + 1 ) d = 0 2 Z
0

cos ydy

|cos y| dy x

n

## Z (n+ 1 ) 2 dy sin(n + 1 ) d = |sin y| as n , 2 y 0 R1 = {f C(T ) : sup |n f | = }

is a residual set. See Rudin Chapter 5 for more details. Lemma 16.11. For f L1 (T ), let f (n) := hf, n i = Z

f (w)wn dw.

Then f c0 and the map f L1 (T ) c0 is a one to one bounded linear transformation into but not onto c0 . P 2 Proof. By Bessels inequality, f (n) < for all f L2 (T ) and in nZ particular lim|n| f (n) = 0. Given f L1 (T ) and g L2 (T ) we have Z [f (w) g(w)] wn dw kf gk f (n) g (n) = 1
T

and hence

linear transformation from L1 (T ) to c0 . R To see that is injective, suppose f = f 0, then T f (w)p(w, w)dw = 0 for all polynomials p in w and w. By the Stone - Wierestrass and the dominated convergence theorem, this implies that Z f (w)g(w)dw = 0
T

for all g L2 (T ). Since L2 (T ) is dense in L1 (T ), it follows that lim supn f (n) = 1 0 for all f L , i.e. f c0 . Since f (n) kf k , we have f kf k showing that f := f is a bounded
1 c0 1

n n

318

BRUCE K. DRIVER

## (16.4). Therefore ran() 6= c0 .

If were surjective, the open mapping theorem would imply that 1 : c0 L (T ) is bounded. In particular this implies there exists C < such that (16.4) kf kL1 C f for all f L1 (T ). c0 Taking f = dn , we nd dn = 1 while limn kdn kL1 = contradicting Eq.
1 c0

16.2. Hahn Banach Theorem. Our next goal is to show that continuous dual X of a Banach space X is always large. This will be the content of the Hahn Banach Theorem 16.15 below.

Proposition 16.12. Let X be a complex vector space over C. If f X and u = Ref XR then (16.5)
Conversely if u and f is dened by Eq. (16.5), then f X and kukXR = kf kX . More generally if p is a semi-norm on X, then

XR

## f (x) = u(x) iu(ix).

|f | p i u p.

Proof. Let v(x) = Im f (x), then v(ix) = Im f (ix) = Im(if (x)) = Ref (x) = u(x). Therefore
Conversely for u XR let f (x) = u(x) iu(ix). Then

## f (x) = u(x) + iv(x) = u(x) + iu(ix) = u(x) iu(ix).

while

f ((a + ib)x) = u(ax + ibx) iu(iax bx) = au(x) + bu(ix) i(au(ix) bu(x))

(a + ib)f (x) = au(x) + bu(ix) + i(bu(x) au(ix)). So f is complex linear. Because |u(x)| = |Ref (x)| |f (x)|, it follows that kuk kf k. For x X choose S 1 C such that |f (x)| = f (x) so Since x X is arbitrary, this shows that kf k kuk so kf k = kuk.34
34

kxk=1 kxk=1

## kf k2 = sup |f (x)|2 = sup (|u(x)|2 + |u(ix)|2 ).

Supppose that M =

## Replacing x0 by x0 if necessary, we may assume that x0 , 0=

kxk=1

sup |u(x)| and this supremum is attained at x0 X with kx0 k = 1. u(x0 ) = M. Since u has a maximum at

since

d | |1 dt 0

+ it| =

d | dt 0

0

## REAL ANALYSIS LECTURE NOTES

319

For the last assertion, it is clear that |f | p implies that u |u| |f | p. Conversely if u p and x X, choose S 1 C such that |f (x)| = f (x). Then holds for all x X. |f (x)| = f (x) = f (x) = u(x) p(x) = p(x)

Denition 16.13 (Minkowski functional). p : X R is a Minkowski functional if (1) p(x + y) p(x) + p(y) for all x, y X and (2) p(cx) = cp(x) for all c 0 and x X. Example 16.14. Suppose that X = R and Notice that if x 0, then p(x) = x/2 and if x 0 then p(x) = x, i.e. x/2 if x 0 p(x) = |x| if x 0. p(x) = inf { 0 : x [1, 2] = [, 2]} .

From this formula it is clear that p(cx) = cp(x) for all c 0 but not for c < 0. Moreover, p satises the triangle inequality, indeed if p(x) = and p(y) = , then x [1, 2] and y [1, 2] so that which shows that p(x + y) + = p(x) + p(y). To check the last set inclusion let a, b [1, 2], then a + b = ( + ) a+ b ( + ) [1, 2] + + since [1, 2] is a convex set and
+

## x + y [1, 2] + [1, 2] ( + ) [1, 2]

= 1.

TODO: Add in the relationship to convex sets and separation theorems, see Reed and Simon Vol. 1. for example. Theorem 16.15 (Hahn-Banach). Let X be a real vector space, M X be a subspace f : M R be a linear functional such that f p on M . Then there exists a linear functional F : X R such that F |M = f and F p. Proof. Step (1) We show for all x X \ M there exists and extension F to M Rx with the desired properties. If F exists and = F (x), then for all y M and R we must have f (y)+ = F (y+x) p(y+x) i.e. p(y+x)f (y). Equivalently put we must nd R such that p(y + x) f (y) for all y M and > 0 p(z x) f (z) for all z M and > 0. So if R is going to exist, we have to prove, for all y, z M and , > 0 that p(y + x) f (y) f (z) p(z x) or equivalently (16.6) f (z + y) p(y + x) + p(z x) = p(y + x) + p(z x).

320

BRUCE K. DRIVER

But f (z + y) = f (z + x) + f (z x) p(z + x) + p(z x) which shows that Eq. (16.6) is true and by working backwards, there exist an R such that f (y) + p(y + x). Therefore F (y + x) := f (y) + is the desired extension. Step (2) Let us now write F : X R to mean F is dened on a linear subspace D(F ) X and F : D(F ) R is linear. For F, G : X R we will say F < G if D(F ) D(G) and F = G|D(F ) , that is G is an extension of F. Let F = {F : X R : M D(F ), F p on D(F )}. Then (F, <) is a partially ordered set. If F is a chain (i.e. a linearly ordered S subset of F) then has an upper bound G F dened by D(G) = D(F )
F

and G(x) = F (x) for x D(F ). Then it is easily checked that D(G) is a linear subspace, G F, and F < G for all F . We may now apply Zorns Lemma (see Theorem B.7) to conclude there exists a maximal element F F. Necessarily, D(F ) = X for otherwise we could extend F by step (1), violating the maximality of F. Thus F is the desired extension of f. Corollary 16.16. Suppose that X is a complex vector space, p : X [0, ) is a semi-norm, M X is a linear subspace, and f : M C is linear functional such that |f (x)| p(x) for all x M. Then there exists F X 0 (X 0 is the algebraic dual of X) such that F |M = f and |F | p.

Proof. Let u = Ref then u p on M and hence by Theorem 16.15, there exists 0 U XR such that U |M = u and U p on M . Dene F (x) = U (x) iU (ix) then as in Proposition 16.12, F = f on M and |F | p. Theorem 16.17. Let X be a normed space M X be a closed subspace and x X \ M . Then there exists f X such that kf k = 1, f (x) = = d(x, M ) and f = 0 on M . Proof. Dene f : M Cx C by f (m + x) for all m M and C. Notice that km + xk = ||kx + m/k || and hence |f (m + x)| = || km + xk which shows kf k 1. In fact, since |f (m + x)| = = inf kx + mk, kf k = 1. By
mM

Hahn-Banach theorem there exists F X such that F |M Cx = f and |F (x)| kxk for all x X, i.e. kF k 1. Since 1 = kf k kF k 1 we see kF k = kf k.

Corollary 16.18. The linear map x X x X where x(f ) = f (x) for all x X is an isometry. (This isometry need not be surjective.) Proof. Since |(f )| = |f (x)| kf kX kxkX for all f X , it follows that x kkX kxkX . Now applying Theorem 16.17 with M = {0} , there exists f X x such that kf k = 1 and |(f )| = f (x) = kxk , which shows that kkX kxkX . x x This shows that x X x X is an isometry. Since isometries are necessarily injective, we are done.

## REAL ANALYSIS LECTURE NOTES

321

Denition 16.19. A Banach space X is reexive if the map x X x X is surjective. Example 16.20. Every Hilbert space H is reexive. This is a consequence of the Riesz Theorem, Proposition 10.15. Example 16.21. Suppose that is a nite measure on a measurable space (X, M), then Lp (X, M, ) is reexive for all p (1, ), see Theorem 13.14. Example 16.22 (Following Riesz and Nagy, p. 214). The Banach space X := C([0, 1]) is not reexive. To prove this recall that X may be identied with complex measures on [0, 1] which may be identied with right continuous functions of bounded variation (F ) on [0, 1], namely Z Z 1 F F (f X f dF = f dF ).
[0,1] 0

Dene X by

() =

x[0,1]

({x}) =

x[0,1]

(F (x) F (x)) ,

so () R the sum of the atoms of . Suppose there existed an f X such that is () = [0,1] f d for all X . Choosing = x for some x (0, 1) would then imply that Z f (x) = f x = (x ) = 1
[0,1]

showing f would have to be the constant function,1, which clearly can not work.

Example 16.23. The Banach space X := L1 (R, m) is not reexive. As we have seen in Theorem 13.14, X L (R, m). The argument in Example 13.15 shows = (L (R, m)) L1 (R, m). 16.3. Weak and Strong Topologies. Denition 16.24. Let X and Y be be a normed vector spaces and L(X, Y ) the normed space of bounded linear transformations from X to Y. (1) The weak topology on X is the topology generated by X , i.e. sets of the form N = n {x X : |fi (x) fi (x0 )| < } i=1 where fi X and > 0 form a neighborhood base for the weak topology on X at x0 . (2) The weak- topology on X is the topology generated by X, i.e. where xi X and > 0 forms a neighborhood base for the weak topology on X at f X . (3) The strong operator topology on L(X, Y ) is the smallest topology such that T L(X, Y ) T x Y is continuous for all x X. (4) The weak operator topology on L(X, Y ) is the smallest topology such that T L(X, Y ) f (T x) C is continuous for all x X and f Y . N n {g X : |f (xi ) g(xi )| < } i=1

Theorem 16.25 (Alaoglus Theorem). If X is a normed space the unit ball in X is weak - compact.

322

BRUCE K. DRIVER

topology. If f B := {f X : kf k 1}, |f (x)| kf k kxk kxk which implies that f (x) Dx for all x X, i.e. B . The topology on B inherited from the weak topology on X is the same as that relative topology coming from the product topology on . So to nish the proof it suces to show B is a closed subset of the compact space . To prove this let x (f ) = f (x) be the projection maps. Then B = {f : f is linear} = {f : f (x + cy) f (x) cf (y) = 0 for all x, y X and c C} \ \ = {f : f (x + cy) f (x) cf (y) = 0}
x,yX cC

Proof. For all x X let Dx = {z C : |z| kxk}. Then Dx C is a Q compact set and so by Tychonos Theorem Dx is compact in the product
xX

x,yX cC

(x+cy x cy )

({0})

which is closed because (x+cy x cy ) : C is continuous. Theorem 16.26 (Alaoglus Theorem for separable spaces). Suppose that X is a separable Banach space, C := {f X : kf k 1} is the closed unit ball in X and {xn } is an countable dense subset of C := {x X : kxk 1} . Then n=1 (16.7) (f, g) :=
X 1 |f (xn ) g(xn )| 2n n=1

denes a metric on C which is compatible with the weak topology on C , C := (w )C = {V C : V w } . Moreover (C , ) is a compact metric space. Proof. The routine check that is a metric is left to the reader. Let be the topology on C induced by . For any g X and n N, the map f X (f (xn ) g(xn )) C is w continuous and since the sum in Eq. (16.7) is uniformly convergent for f C , it follows that f (f, g) is C continuous. This implies the open balls relative to are contained in C and therefore C . We now wish to prove C . Since C is the topology generated by {|C : x C} , it suces to show x is continuous for all x C. But given x C x there exists a subsequence yk := xnk of {xn } such that such that x = limk yk . n=1 Since
f C

## x sup |(f ) yk (f )| = sup |f (x yk )| kx yk k 0 as k ,

f C

yk x uniformly on C and using yk is continuous for all k (as is easily checked) we learn x is also continuous. Hence C = (|C : x X) . x The compactness assertion follows from Theorem 16.25. The compactness assertion may also be veried directly using: 1) sequential compactness is equivalent to compactness for metric spaces and 2) a Cantors diagonalization argument as in the proof of Theorem 10.35. 16.4. Supplement: Quotient spaces, adjoints, and more reexivity. Denition 16.27. Let X and Y be Banach spaces and A : X Y be a linear operator. The transpose of A is the linear operator A : Y X dened by

## REAL ANALYSIS LECTURE NOTES

323

A f (x) = f (Ax) for f Y and x X. The null space of A is the subspace nul(A) := {x X : Ax = 0} X. For M X and N X let N := {x X : f (x) = 0 for all f N }. c Proposition 16.28 (Basic Properties). (1) kAk = A and A x = Ax for all x X. (2) M 0 and N are always closed subspace of X and X respectively. (3) M 0 = M . N 0 with equality when X is reexive. (4) N (5) nul(A) = ran(A ) and nul(A ) = ran(A)0 . Moreover, ran(A) = nul(A ) and if X is reexive, then ran(A ) = nul(A)0 . c (6) X is reexive i X is reexive. More generally X = X X 0 . Proof. (1) kAk = sup kAxk = sup sup |f (Ax)|
kxk=1 kxk=1 kf k=1 kf k=1 kxk=1 kf k=1

M 0 := {f X : f |M = 0} and

(2) This is an easy consequence of the assumed continuity o all linear functionals involved. (3) If x M, then f (x) = 0 for all f M 0 so that x M 0 . Therefore M M 0 . If x M , then there exists f X such that f |M = 0 while / f (x) 6= 0, i.e. f M 0 yet f (x) 6= 0. This shows x M 0 / and we have 0 shown M M. 0 0 (4) It is again simple to show N N and therefore N N . Moreover, as above if f N there exists X such that |N = 0 while (f ) 6= 0. / If X is reexive, = x for some x X and since g(x) = (g) = 0 for all g N , we have x N . On the other hand, f (x) = (f ) 6= 0 so 0 0 f N / . Thus again N N . (5) nul(A) = {x X : Ax = 0} = {x X : f (Ax) = 0 f X } = x X : A f (x) = 0 f X = x X : g(x) = 0 g ran(A ) = ran(A ) . = {f Y : f (Ax) = 0 x X} = f Y : f |ran(A) = 0 = ran(A)0 .

## = sup sup A f (x) = sup A f = A .

nul(A ) = f Y : A f = 0 = f Y : (A f )(x) = 0 x X x (6) Let X and dene f X by f (x) = () for all x X and set 0 := f . For x X (so x X ) we have x x x 0 () = () f () = f (x) x(f ) = f (x) f (x) = 0.

Similarly,

324

BRUCE K. DRIVER

c c This shows 0 X 0 and we have shown X = X + X 0 . If X X 0 , for some f X and 0 = f () = x(f ) = f (x) for all x X, x then = f c i.e. f = 0 so = 0. Therefore X = X X 0 as claimed. If X is = X and so X 0 = {0} showing X = X , i.e. X c reexive, then X 0 = {0} and is reexive. Conversely if X is reexive we conclude that X therefore X = {0} = X 0 = X, so that X is reexive.
J f X f X J f X x is the identity map since J f (x) = f (Jx) = f () = x(f ) = f (x) for all

## Alternative proof. Notice that f = J , where J : X X is given by Jx = x, and the composition

x X. Thus it follows that X X is invertible i J is its inverse which can happen i nul(J ) = {0} . But as above nul(J ) = ran(J)0 which will be zero i ran(J) = X and since J is an isometry this is equivalent to saying ran(J) = X . So we have again shown X is reexive i X is reexive. Theorem 16.29. Let X be a Banach space, M X be a proper closed subspace, X/M the quotient space, : X X/M the projection map (x) = x + M for x X and dene the quotient norm on X/M by k(x)kX/M = kx + M kX/M = inf kx + mkX .
mM

## Then (1) (2) (3) (4)

kkX/M is a norm on X/M. The projection map : X X/M has norm 1, kk = 1. (X/M, kkX/M ) is a Banach space. If Y is another normed space and T : X Y is a bounded linear transformation such that M nul(T ), then there exists a unique linear transformation S : X/M Y such that T = S and moreover kT k = kSk .
n

## x + M = 0 X/M. If c C\ {0} , x X, then

mM

Proof. 1) Clearly kx + M k 0 and if kx + M k = 0, then there exists mn M such that kx + mn k 0 as n , i.e. x = lim mn M = M. Since x M, kcx + M k = inf kcx + mk = |c| inf kx + m/ck = |c| kx + M k
mM

because m/c runs through M as m runs through M. Let x1 , x2 X and m1 , m2 M then Taking innums over m1 , m2 M then implies kx1 + x2 + M k kx1 + x2 + m1 + m2 k kx1 + m1 k + kx2 + m2 k. kx1 + x2 + M k kx1 + M k + kx2 + M k.

and we have completed the proof the (X/M, k k) is a normed space. 2) Since k(x)k = inf mM kx + mk kxk for all x X, kk 1. To see kk = 1, let x X \ M so that (x) 6= 0. Given (0, 1), there exists m M such that kx + mk 1 k(x)k .

## REAL ANALYSIS LECTURE NOTES

325

Therefore, k(x + m)k k(x)k kx + mk = = kx + mk kx + mk kx + mk which shows kk . Since (0, 1) is arbitrary we conclude that k(x)k = 1. P 3) Let (xn ) X/M be a sequence such that k(xn )k < . As above there P exists mn M such that k(xn )k 1 kxn + mn k and hence kxn + mn k 2 P P 2 k(xn )k < . Since X is complete, x := (xn +mn ) exists in X and therefore
n=1

by the continuity of ,

(x) =

n=1

showing X/M is complete. 4) The existence of S is guaranteed by the factor theorem from linear algebra. Moreover kSk = kT k because and kT k = kS k kSk kk = kSk

(xn + mn ) =

n=1

(xn )

kSk = sup

## kS((x))k kT xk = sup k(x)k xM / xM k(x)k / kT xk kT xk sup = sup = kT k . x6=0 kxk xM kxk /

Theorem 16.30. Let X be a Banach space. Then (1) Identifying X with X X , the weak topology on X induces the weak topology on X. More explicitly, the map x X x X is a homeo with the relative morphism when X is equipped with its weak topology and X topology coming from the weak- topology on X . (2) X X is dense in the weak- topology on X . (3) Letting C and C be the closed unit balls in X and X respectively, then C := { C : x C} is dense in C in the weak topology on X . . x (4) X is reexive i C is weakly compact. Proof. (1) The weak topology on X is generated by o n f : f X = { X (f ) : f X } . So the induced topology on X is generated by {x X x X x(f ) = f (x) : f X } = X

(16.8)

and so the induced topology on X is precisely the weak topology. (2) A basic weak - neighborhood of a point X is of the form for some {fk }k=1 X and > 0. be given. We must now nd x X such that x N , or equivalently so that |(fk ) (fk )| = |fk (x) (fk )| < x for k = 1, 2, . . . , n. N := n { X : |(fk ) (fk )| < } k=1
n

(16.9)

326

BRUCE K. DRIVER

In fact we will show there exists x X such that (fk ) = fk (x) for k = 1, 2, . . . , n. To prove this stronger assertion we may, by discardn ing some of the fk s if necessary, assume that {fk }k=1 is a linearly inn dependent set. Since the {fk }k=1 are linearly independent, the map x X (f1 (x), . . . , fn (x)) Cn is surjective (why) and hence there exists x X such that (16.10) (f1 (x), . . . , fn (x)) = T x = ( (f1 ) , . . . , (fn )) as desired. (3) Let C X and N be the weak - open neighborhood of as in Eq. (16.8). Working as before, given > 0, we need to nd x C such that Eq. (16.9). It will be left to the reader to verify that it suces again to assume {fk }n is a linearly independent set. (Hint: Suppose that k=1 {f1 , . . . , fm } were a maximal linearly dependent subset of {fk }n , then k=1 each fk with k > m may be written as a linear combination {f1 , . . . , fm } .) As in the proof of item 2., there exists x X such that Eq. (16.10) holds. The problem is that x may not be in C. To remedy this, let N := n nul(fk ) = nul(T ), : X X/N Cn be the projection map and fk = k=1 for k = 1, 2, . . . , n. Then we have (X/N ) be chosen so that fk = fk produced x X such that ( (f1 ) , . . . , (fn )) = (f1 (x), . . . , fn (x)) = (f1 ((x)), . . . , fn ((x))). Since f1 , . . . , fn is a basis for (X/N ) we nd Pn Pn | i=1 i (fi )| i=1 i fi ((x)) Pn Pn k(x)k = sup = sup Cn \{0} Cn \{0} k i=1 i fi k i=1 i fi P P |( n i fi )| k n i fi k i=1 Pn Pi=1 = sup kk sup = 1. n Cn \{0} k Cn \{0} k i=1 i fi k i=1 i fi k

which can be arbitrarily small (i.e. less than ) by choosing suciently close to 1. (4) Let C := { : x C} C X . If X is reexive, C = C is weak x - compact and hence by item 1., C is weakly compact in X. Conversely if C is weakly compact, then C C is weak compact being the continuous image of a continuous map. Since the weak topology on X is Hausdor, it follows that C is weak closed and so by item 3, weak = C. So if X , / kk C = C, i.e. there exists C = C x C such that x = / kk . This shows = (kk x) and therefore X = X .

Hence we have shown k(x)k 1 and therefore for any > 1 there exists y = x + n X such that kyk < and ( (f1 ) , . . . , (fn )) = (f1 (y), . . . , fn (y)). Hence |(fi ) fi (y/)| fi (y) 1 fi (y) (1 1 ) |fi (y)|

16.5. Exercises.

## REAL ANALYSIS LECTURE NOTES

327

16.5.1. More Examples of Banach Spaces. Exercise 16.1. Let (X, M) be a measurable space and M (X) denote the space of complex measures on (X, M) and for M (X) let kk |k(X). Show (M (X), kk) is a Banach space. (Move to Section 14.) Exercise 16.2. Folland 5.9. Exercise 16.3. Folland 5.10. Exercise 16.4. Folland 5.11. 16.5.2. Hahn-Banach Theorem Problems. Exercise 16.5. Folland 5.17. Exercise 16.6. Folland 5.18. Exercise 16.7. Folland 5.19. Exercise 16.8. Folland 5.20. Exercise 16.9. Folland 5.21. Exercise 16.10. Let X be a Banach space such that X is separable. Show X is separable as well. (Folland 5.25.) Hint: use the greedy algorithm, i.e. suppose D X \ {0} is a countable dense subset of X , for D choose x X such that kx k = 1 and | (x )| 1 k k. 2 Exercise 16.11. Folland 5.26. 16.5.3. Baire Category Result Problems. Exercise 16.12. Folland 5.29. Exercise 16.13. Folland 5.30. Exercise 16.14. Folland 5.31. Exercise 16.15. Folland 5.32. Exercise 16.16. Folland 5.33. Exercise 16.17. Folland 5.34. Exercise 16.18. Folland 5.35. Exercise 16.19. Folland 5.36. Exercise 16.20. Folland 5.37. Exercise 16.21. Folland 5.38. Exercise 16.22. Folland 5.39. Exercise 16.23. Folland 5.40. Exercise 16.24. Folland 5.41.

328

BRUCE K. DRIVER

16.5.4. Weak Topology and Convergence Problems. Exercise 16.25. Folland 5.47. Denition 16.31. A sequence {xn }n=1 X is weakly Cauchy if for all V w such that 0 V, xn xm V for all m, n suciently large. Similarly a sequence {fn } X is weak Cauchy if for all V w such that 0 V, fn fm V n=1 for all m, n suciently large. Remark 16.32. These conditions are equivalent to {f (xn )}n=1 being Cauchy for all f X and {fn (x)}n=1 being Cauchy for all x X respectively. Exercise 16.26. Folland 5.48. Exercise 16.27. Folland 5.49. Exercise 16.28. land 5.50. Exercise 16.29. Show every weakly compact subset X is norm bounded and every weak compact subset of X is norm bounded. Exercise 16.30. Folland 5.51. Exercise 16.31. Folland 5.53.