You are on page 1of 57

mlbaker.

org presents

PMATH 453/753
Functional Analysis
Dr. Nicolaas Spronk Fall 2012 (1129) University of Waterloo
www.math.uwaterloo.ca/nspronk
Disclaimer: These notes are provided as-is, and may be incomplete or contain errors.

Contents
1 Normed spaces and Banach spaces

2 Topological spaces and continuous functions

3 `p -spaces

4 Linear operators on normed spaces

5 Separation by hyperplanes

12

6 Consequences of Baire Category Theorem

15

7 Finite-dimensional Banach spaces

20

8 Initial topologies, compactness

22

9 An application of ultrafilters: ultrafilter limits

25

10 Nets

27

11 Extreme points and the Krein-Milman theorem

31

12 Euclidean and Hilbert spaces

34

13 Spectral theory

40

14 Adjoint operators

44

15 Compact operators

46

16 Structure theorem for compact operators

48

17 Operators on Hilbert space

50

Normed spaces and Banach spaces

Let F be R or C. Define

(
|x| =

absolute value of x if x R
modulus of x
if x C.

1.1 Definition. Let X be an F-vector space. A norm on X is a functional k k : X R0 , such that


(i) kxk = 0 if and only if x = 0 (non-degenerate).
(ii) kxk = || kxk, for all F (| |-homogeneity).
(iii) kx + yk kxk + kyk (subadditivity).
We call the pair (X , k k) a normed (vector) space. Often, when no ambiguity arises, we simply say X is a
normed space. If (X , k k) is complete with respect to the metric d(x, y) = kx yk induced by k k (i.e. each
Cauchy sequence in X has a limit in X ) we say that (X , k k) is a Banach space.
1.2 Example. We have the following examples:
(i) (F, | |) is a Banach space1 .
(ii) (Fn , k k2 ) is a Banach space, where
k(x1 , . . . , xn )k2 =

n
X

!1/2
|xi |

i=1

(iii) Let (X, d) be a metric space. Define


Cb (X) = CbF (X) = {f : X F : f is continuous and bounded i.e. kf k = sup |f (x)| < }.
xX

Then (Cb (X), k k ) is a Banach space2 .


(iv) Let 1 p . (Lp [0, 1], k kp ) is a Banach space3 .
(v) On C[0, 1] (note, since ([0, 1], usual metric) is compact, the b for boundedness is redundant), we can put
norm
Z 1
1/p
p
kf kp =
|f |
0

where the integral is a Riemann integral. (C[0, 1], kkp ) is a normed space. It is not complete. Indeed, consider
the sequence of functions

if 0 t 21
0
fn (t) = some affine bit if 12 < t < 12 + n1

1
if 12 + n1 t 1.

This sequence is k kp -Cauchy but converges to no g C[0, 1], i.e. for no such g do we have limn kfn gkp = 0.
1 M147

in the case of R, M247 in the case of C


= R2 .

2 PM351.

3 PM450/354.

(vi) Define
` = {x = (x1 , x2 , . . .) FN : xi = 0 for all but finitely many indices i}.
This is an F-vector space under pointwise operations
(x1 , x2 , . . .) + (y1 , y2 , . . .) = (x1 + y1 , x2 + y2 , . . .),

(x1 , x2 , . . .) = (x1 , x2 , . . .).

We can put many norms on `:


kxk = sup |xi |,

v
u
uX
kxk2 = t
|xi |2 ,

iN

kxk1 =

i=1

|xi |.

i=1

Note that all sums are finite by definition of `. Convince yourself that none of (`, k k ), (`, k k2 ), (`, k k1 )
are complete.

Topological spaces and continuous functions

2.1 Definition. Given a non-empty set X, define its power set, P(X) = {A : A X}. A topology on X is a
subfamily P(X) such that
(i) , X .
(ii) {Ui }iI implies that

Ui (closure under arbitrary unions).

iI

(iii) {U1 , . . . , Un } implies that

n
\

Ui (closure under finite intersections).

i=1

We call the elements of ( -)open sets. The pair (X, ) is called a topological space.
2.2 Example. We have the following examples:
(i) triv = {, X} is called the trivial topology.
(ii) Let (X, ) be a metric space, i.e. : X X R0 satisfies
(i) (x, y) = 0 if and only if x = y (non-degeneracy).
(ii) (x, y) = (y, x) (symmetry).
(iii) (x, z) (x, y) + (y, z) (triangle inequality).
For  > 0, we define the open ball of radius  centered at x by B (x) = {x0 X : (x, x0 ) < }. We let
= {U X : for each x U , there is x > 0 such that Bx (x) U }.
(iii) d = P(X), called the discrete topology. Notice that d = , where is the discrete metric, i.e.
(
1 if x 6= y
(x, y) =
0 if x = y.
(iv) The Sorgenfrey line: Take X = R. Let
= {U R : for each x U , there is x > 0 such that [x, x + x ) U }.
Show there is no metric such that = .

(v) Recall, if 1 , 2 are metrics on X, we say 1 and 2 are equivalent (denoted 1 2 ) if there exist c, C > 0
such that
c1 2 C1 .
If 1 2 , then we have 1 = 2 . However, we may have 1 = 2 without 1 2 . As an example, consider
the metrics on R given by
|s t|
.
1 (s, t) = |s t|,
2 (s, t) =
1 + |s t|
The latter is bounded, whereas the former is unbounded, so there is no way these could be equivalent. However,
they induce the same topology on R.
Assignment #1 is posted; google Nico Spronk, hit PMATH 753 tab under Teaching Fall 2012. I would
advise trying most questions after Fridays lecture.
2.3 Definition. Let (X, ), (Y, ) be topological spaces. A function f : X Y is ( --)continuous at x0 X if
for each neighbourhood4 f (x0 ) V , there is a neighbourhood x0 U such that f (U ) V , i.e. f (x) V
for each x U .
2.4 Remark. Note that in metric topology, V plays the role of  and U plays the role of .
2.5 Definition. Let F = R or C. Let


F
Cb (X, ) = Cb (X, ) = f : X F | f continuous, and bounded: kf k = sup |f (x)| < .
xX

It is straightforward to prove that under pointwise operations, i.e.


(f + g)(x) = f (x) + g(x),

(f )(x) = f (x)

x X

for F, f, g Cb (X, ), this is an F-vector space, and k k is a norm on Cb (X, ).


It is really the completeness of the codomain that drives the following result.
2.6 Theorem. (Cb (X, ), k k ) is a Banach space, i.e. Cb (X, ) is complete with respect to k k .
Proof. Let (fn )
n=1 Cb (X, ) be a Cauchy sequence. Then for all  > 0, there is N N such that for all n, m N
we have kfn fm k < . Then for each fixed x X we observe that |fn (x) fm (x)| kfn fm k hence
(fn (x))
n=1 F is Cauchy. Since (F, | |) is complete, we can define
f (x) = lim fn (x)
n

for all x X. We will show that


sup |fn (x) f (x)| =: kfn f k 0.
xX

Let  > 0. Choose N N such that if n, m N then kfn fm k < . Observe that for x X, we have
|fn (x) f (x)| = lim |fn (x) fm (x)| .
m

It follows that fn f uniformly on X. Also, let us note that f is bounded: we can pick  = 1 and find some N so
that kfN f k < . However, this means that for any x X we have
|f (x)| |f (x) fN (x)| + |fN (x)| 1 + kfN (x)k .
Now we show f is continuous. Fix an arbitrary x0 X,  > 0. Let n0 be such that for n, m n0 we have
kfn fm k < /3. Notice that
kfn f k = lim kfn fm k
m


,
3

for n n0 .

Let x0 U such that


fn0 (U ) B/3 (fn0 (x0 )) F
4 Here, a neighbourhood of x is defined to be an open set containing x. Some authors define it to mean any set containing an open
set containing x.

(we can do this by definition of continuity for fn0 ). Then for x U we have



+ + =
3 3 3
and hence f (U ) B (f (x0 )). Hence f is continuous at x0 . Thus f is continuous at each point in X.
|f (x) f (x0 )| |f (x) fn0 (x)| + |fn0 (x) fn0 (x0 )| + |fn0 (x0 ) f (x0 )| <

2.7 Example. We have:


(i) Let
` = ` (N) = {x = (x1 , x2 , . . .) FN : kxk = sup |xn | < }.
n

Note that ` is an F-vector space under pointwise operations,


x + y = (x1 + y1 , x2 + y2 , . . .),

x = (x1 , x2 , . . .).

The map Cb (N, d ) ` , given by f 7 (f (n))nN is easily seen to be a surjective linear isometry. Hence
these spaces are isometrically isomorphic, hence (` , k k ) is complete. Verify the details.
(ii) Let
c = {x ` : lim xn exists}.
n

We consider the topology on N {} given by S N and S {n, n + 1, . . . , } for S N and n N.


Verify that is indeed a topology on N {}, and that the map Cb (N {}, ) c given by f 7 (f (n))nN
is a bijective linear isometry.
(iii) Let
c 0 = {x ` : lim xn = 0} c .
n

2.8 Proposition. If x0 X, (X, ) a topological space, then the ideal of x0 ,


I(x0 ) = {f Cb (X, ) : f (x0 ) = 0}
is a closed subspace, and hence a Banach space.
Proof. It is easy to show that I(x0 ) is a subspace, and is closed. We note that any closed subset of a complete
metric space is also complete5 .
In terms of c 0 , we simply observe that the map from (ii) above takes I() Cb (N {}, ) onto c 0 .
2.9 Remark. The map N {} {0} { n1 : n N} given by
(
1
if n N
n 7 n
0 if n =
is continuous, with continuous inverse. Note that {0} { n1 : n N} is a closed subspace.

`p -spaces

3.1 Lemma. If I R is an open interval and : I R satisfies 00 (t) > 0 for all t I, then for a < c in I and
0 < s < 1 we have
((1 s)a + sc) < (1 s)(a) + s(c)
(*)
Any function satisfying (*) is called strictly convex.
Proof. By MVT, 0 is strictly increasing. Let b = (1 s)a + sc so that
s=

ba
,
ca

1s=

cb
.
ca

Another application of MVT allows us to find , such that a < < b < < c and
(b) (a)
(c) (b)
= 0 () < 0 () =
.
ba
cb
Thus,
((b) (a)) (c b) < ((c) (b)) (b a) = (b)(c a) < (c)(b a) + (a)(c b)
which reads ((1 s)a + sc) < s(c) + (1 s)(a).
5 PM351.

3.2 Corollary. If p, q > 1 are such that

1
p

1
q

= 1 (such p, q are called conjugate indices) and a, b 0, then


ap
bq
+
p
q

ab
with equality if and only if ap = bq .

Proof. Let (t) = log t for t (0, ), so 00 (t) = t12 > 0. Then
 p

a
bq
1
1
log
+
log(ap ) log(bq ) = log a log b = log(ab).
p
q
p
q
We observe that < holds if ap 6= bq , whereas = holds if ap = bq . Now, apply the strictly decreasing function
x 7 ex to gain the desired result6 .
3.3 Theorem (Hlders Inequality). If a1 , . . . , an , b1 , . . . , bn C then
n

!1/p n
!1/q
n
X

X
X


p
q
ai bi
|ai |
|bi |



i=1

i=1

i=1

where p and q are conjugate indices. Moreover, equality holds if and only if either ai = 0 for all i or there is t 0
such that t|ai |p = |bi |q for all i, and |z| = 1 such that ai bi = z|ai bi | for all i.
P
P
1/p
1/q
Proof. Let A = ( |ai |p ) , B = ( |bi |q ) . Suppose that AB 6= 0. From the corollary above, we have



n
n
n 
(1) 1 X
(2) X
1 X
|ai |p
|bi |q
1
1 1
1 p

a
b

|a
||b
|

+
A +
B q = + = 1.
=

i i
i i
p
q
p
q

AB i=1
AB i=1
pA
qB
pA
qB
p
q
i=1
Note that equality can hold at (1) if and only if |z| = 1 exists as claimed. Also, equality holds at (2) if

p 
q
|ai |
|bi |
=
A
B
for all i (or at least one of A, B = 0), hence t = B q /Ap suffices.
As a corollary, we have the following generalisation of the triangle inequality.
3.4 Theorem (Minkowskis Inequality). If p > 1 and a1 , . . . , an , b1 , . . . , bn C then
!1/p
!1/p
!1/p
n
n
n
X
X
X
|ai + bi |p

|ai |p
+
|bi |p
i=1

i=1

i=1

with equality if and only if either ai = 0 for all i, or there is t 0 such that tbi = ai for all i.
P
Proof. Again, suppose
|ai + bi | =
6 0. Then by the triangle inequality and Hlders inequality,

!1/p
!1/p " n
#1/q
n
n
n
n
X
X
X
X
X
H.I.
p
p1
p
p
(p1)q

|ai + bi |
(|ai | + |bi |) |ai + bi |

|ai |
+
|bi |
|ai + bi |
.
i=1

Note that

(1)

(2)

i=1

i=1

i=1

i=1

1
1
p1
1 1
+ = 1 =
=1 =
= (p 1)q = p.
p q
q
p
p

P
1/q
Divide by ( |ai + bi |p ) , and note that 1
sgn ai = sgn bi , where

1
q

= p1 , so the inequality prevails. Equality at (1) holds if and only if


(

sgn z =

z
|z|

if z 6= 0

if z = 0

or all ai or all bi are zero. Equality at (2) holds if and only if


|ai + bi |(p1)q = |ai + bi |p = t1 |ai |p = t2 |bi |p ,

for some t1 , t2 0.

So t = (t2 /t1 )1/p assuming that not all bi are zero.


6 The

resultPis also a special case of a weighted AM-GM inequality (obtainable from Jensens inequality):
pi > 0 where
pi = 1. See Bollobs Linear Analysis, Chapter 1, Theorem 3.

ai i

ai pi for ai 0,

3.5 Definition. For 1 p < , we define

`p = `p (N) = x = (x1 , x2 , . . .) FN : kxkp =

!1/p
|xi |p

i=1

<

3.6 Proposition. For 1 < p < , `p is a vector space, under pointwise operations, and k kp is a norm on `p .
Proof. If x = (x1 , x2 , . . .) let x(N ) = (x1 , . . . , xN , 0, 0, . . .). Then if x `p ,
!1/p

X
N
(N )
p
kx x kp =
|xi |
0
i=N +1

by the definition of convergence, and kxkp = limN kx(N ) kp . Hence, by Minkowskis Inequality, if x, y `p , then we
have


M.I.
kx + ykp = lim kx(N ) + y (N ) kp lim kx(N ) kp + ky (N ) kp = kxkp + kykp
N

and hence x + y `p with kx + ykp kxkp + kykp . Moreover, it is obvious that if F, and x `p , then
kxkp = ||kxkp < , so x `p .
3.7 Proposition. (`1 , k k1 ) is a normed space.
Proof. Easy exercise.

Linear operators on normed spaces

4.1 Definition. Let X , Y be F-vector spaces. Let


L(X , Y) = {T : X Y | T is linear}.
This is itself an F-vector space under pointwise operations, i.e. for F, S, T L(X , Y) we put
(S + T )x = Sx + T x,

x X .

4.2 Definition. If (X , k k) is a normed space then we define


The closed ball, Bkk (X ) = B(X ) = {x X : kxk 1}.
The open ball/disc, Dkk (X ) = D(X ) = {x X : kxk < 1}.
The sphere, Skk (X ) = S(X ) = {x X : kxk = 1}.
4.3 Proposition. If X , Y are both normed spaces, and T : X Y, then the following are equivalent:
(i) T is continuous on X .
(ii) T is continuous at one point x0 X .
(iii) sup{kT xk : x D(X )} < , i.e. T is bounded.
(iii) sup{kT xk : x B(X )} < .
Proof. (i) (ii): Obvious.
(ii) (iii): Note that
T x0 + D(Y) = {y Y : kT x0 yk < 1}
is an open neighbourhood of T x0 . Hence, by the continuity assumption there is > 0 such that
T x0 + T (D(X )) = T (x0 + D(X )) T x0 + D(Y) = T (D(X )) D(Y) = T (D(X ))
i.e. we have sup{kT xk : x D(X )} < 1/ < .
(iii) (i): If N = sup{kT xk : x D(X )}, then for  > 0, x X we have




1


kT xk = (kxk + ) T
x (kxk + )N


kxk + 
|
{z
}
D(X )

1
D(Y)

and taking  0+ we see kT xk N kxk for x X , which yields


kT x T yk = kT (x y)k N kx yk
so T is Lipschitz, and hence continuous.
(iii) (iii): Obvious.
(iii) (iii): We saw that T satisfying (iii) is Lipschitz and hence (iii) follows.
4.4 Definition. For normed spaces X and Y, let
B(X , Y) = {T L(X , Y) : T is bounded}.
For T B(X , Y), we define its operator norm by
kT k = sup{kT xk : x D(X )} = sup{kT xk : x B(X )} = sup{kT xk : x S(X )}.
We note that kT k is the Lipschitz constant of T .
4.5 Definition. If (X, ) is a topological space, and Y is a normed space, let
CbY (X, ) = {F : X Y : F is -k k continuous and kF k = sup kF (x)k < }.
xX

4.6 Theorem. If Y is a Banach space, then (CbY (X, ), k k ) is also a Banach space.
Proof. Trivial modifications of F-valued case.
4.7 Theorem. If X , Y are normed spaces, then (B(X , Y), k k) is a normed space. If Y is a Banach space, then
(B(X , Y), k k) is a Banach space.
Proof. First, if S, T B(X , Y) then
kS + T k = sup{k(S + T )xk : x D(X )} sup{kSxk : x D(X )} + sup{kT x0 k : x0 D(X )} = kSk + kT k.
|
{z
}
kSx+T xk
kSxk+kT xk

It is easy to show kSk = ||kSk for F. Now, suppose that Y is a Banach space. Let B(X ) have the usual
norm topology from X . Define
: B(X , Y) CbY (B(X )),
T 7 T |B(X )
that is, restricts T to the unit ball. Then is an isometry, i.e. for all T we have
k(T )k = sup{kT xk : x B(X )} = kT k.

It is easy to see that is linear. Hence if (Tn )


n=1 B(X , Y) is Cauchy, then ((Tn ))n=1 is Cauchy as well. Thus

F = lim (Tn )
n

exists, since (CbY (B(X )), k k ) is complete. Now, let T : X Y be given by


(
1
kxkF ( kxk
x) if x 6= 0
T (x) =
0
if x = 0.
Note that kT Tn k = kF (Tn )k 0. It remains to show that T is linear. Let x, x0 X , F, and suppose
that kxk, kx0 k, kx + x0 k =
6 0. Then




1
1
0
0
0
0
0
(x + x ) = kx + x k lim Tn
(x + x )
T (x + x ) = kx + x kF
n
kx + x0 k
kx + x0 k




1
1 0
0
0
0
= lim Tn (x + x ) = lim Tn (x) + lim Tn (x ) = kxk lim Tn
x + kx k lim Tn
x
n
n
n
n
n
kxk
kx0 k




1
1 0
= kxkF
x + kx0 kF
x = T (x) + T (x0 ).
kxk
kx0 k
Thus T is linear.
8

4.8 Definition. For a vector space X , we define the algebraic dual space of X by X 0 = L(X , F); it consists of
the linear functionals on X . If X is a normed space7 , then X = B(X , F) denotes the continuous dual space, i.e.
the subspace of continuous linear functionals on X .
4.9 Proposition. If x `1 , then fx : c 0 F given by
fx (y) =

xi yi

i=1

is a bounded linear functional. Moreover, all bounded linear functionals arise in this way, with kfx k = kxk1 .
4.10 Remark. This identification x 7 fx is a linear isometric isomorphism c 0
= `1 .
Proof. Let x `1 , and y c 0 . We establish that

xi yi

i=1

converges. If m < n, observe


n

n
n
X
X
X
m<n


xi yi
|xi ||yi |
|xi |kyk 0



i=m

i=m

i=m

so the series converges absolutely. Thus



X
X
X


x
y

|x
||y
|

|xi |kyk = kxk1 kyk .



i i
i
i


i=1

i=1

i=1

This shows fx is well-defined. It is easy to check that it is linear. Observe that kfx k kxk1 is (essentially) already
shown. Let
y (N ) = (sgn x1 , . . . , sgn xN , 0, 0, . . .) c 0 ,
and ky (N ) k 1. Then
kfx k sup |fx (y (N ) )| = sup

N
X

|xi | =

N N n=1

|xi | = kxk1 .

n=1

Now suppose that f c 0 . Let en be the sequence with 0 everywhere except for a 1 in the nth position. Define x
by putting xi = f (ei ). We observe, by linearity, that
n
X

|xi | = f (y (n) ) = |f (y (n) )| kf k < .

i=1

Thus,

|xi | = sup

i=1

n
X

|xi | < = x `1 .

nN i=1

We observe that f |` = fx |` , where


` = c 00 = {y = (y1 , . . . , yn , 0, 0, . . .) : y1 , . . . , yn F, n N}.
Note that ` is dense in c 0 , i.e. y = limn (y1 , . . . , yn , 0, 0, . . .) because
n

ky (y1 , y2 , . . . , yn , 0, 0, . . .)k = sup |yi | 0.


in+1

Using the continuity of f and the fact that f = fx on ` (a dense subset), we have
f (y) = lim f ((y1 , . . . , yn , 0, 0, . . .)) = lim fx ((y1 , . . . , yn , 0, 0, . . .)) = fx (y)
n

so that f = fx .
7 More

generally, X is a topological vector space.

A1, Due Monday. Office hours today 11-noon, Friday 2-3:30 pm, or by appointment. nspronk@uwaterloo.ca.
Grader needed for AM/PM 331 (Real Analysis), about 3 hours per week, $600 per term. Talk to me by 3pm
today.
4.11 Definition. A Hamel basis of an F-vector space X is any family B X such that
1. B is linearly independent i.e. for any finite distinct e1 , . . . , en B and 1 , . . . , n F, the equation
n
X

i ei = 0

i=1

occurs if and only if i = 0 for all i.


2. B is spanning i.e. for any x X , there are e1 , . . . , en B and 1 , . . . , n F such that
x=

n
X

i ei .

i=1

If we assume the axiom of choice, every vector space admits a basis.


4.12 Question. Given a normed space X , does there exist a non-zero bounded linear functional? Do there exist
enough bounded functionals, i.e. given a point 0 6= x X , is there f X such that f (x) 6= 0? Given x X , is
there f X such that |f (x)| = kxk, kf k 1?
4.13 Definition. Let X be an F-vector space. A sublinear functional on X is a functional p : X R such that
p(tx) = tp(x) for t R0 (non-negative homogeneity).
p(x + y) p(x) + p(y) for x, y X (subadditivity).
4.14 Theorem (Hahn-Banach Theorem). Let X be an R-vector space, p : X R a sublinear functional,
Y X a subspace, and f Y 0 such that f (y) p(y) for y Y. Then there exists F X 0 such that


F = f
Y

and F (x) p(x) for x X . We will refer to F as a Hahn-Banach extension of f (relative to p).
Proof. Step 1 : Suppose there is x X \ Y such that X = spanR {x, Y}. We observe for y+ , y Y, we have
f (y+ ) + f (y ) = f (y+ + y ) p(y+ + x + y x) p(y+ + x) + p(y x)
which implies that
f (y ) p(y x) p(y+ + x) f (y+ ).
Hence there is c R so that
sup{f (y ) p(y x) : y Y} c inf{p(y+ + x) f (y+ ) : y+ Y}.
0

We then define F X 0 = spanR {x, Y} by


F (x + y) = c + f (y).
Observe F |Y = f . It remains to check that F p on X . Suppose = t > 0. Then
c p( 1t y + x) f ( 1t y) = tc p(y + tx) f (y) = F (tx + y) = tc + f (y) p(y + tx).
If = s, s > 0, then
f ( 1s y) p( 1s y x) c = f (y) p(y sx) cs = F (sx + y) = sc + f (y) p(y sx).
Notice that if dim X < (or if Y is of finite codimension) then we use simple induction to finish.
Step 2 : We call a pair (, M) a p-extension of f if Y M and M is a subspace of X , |Y = f and p on M.
We assign the following partial ordering to the set E of p-extensions of f :
(, M) (, N ) if M N and |M = .
10

This is indeed a partial ordering. Let C E be a chain. We let


[
U=
M
(,M)C

so U is a subspace of X , as (M)(,M)C is a totally ordered collection. Define U 0 by


(x) = (x),

whenever x M, for (, M) C.

f where (, M), (,
f C, with M M,
f say, then (x)
Then is well-defined: if x M and y M,
e M)
e
= (x).
Also, is clearly linear. Note that (, U) is an upper bound for C.
By Zorns Lemma, there exists a maximal element (F, M) for E. If it were the case that M ( X , then there would
0
be x X \ M. Then, by Step 1, we would find spanR {x, M} such that
|M = f , and p on spanR {x, M}.
Then (, spanR {x, M}) E and (F, M) < (, spanR {x, M}), which would violate the fact that (F, M) is maximal.
Thus M = X .
4.15 Lemma. Let X be a C-vector space, hence a R-vector space.
(i) If f XR 0 (i.e. a R-valued linear functional), then
fC (x) = f (x) if (ix)
is C-linear, i.e. fC XC 0 .
(ii) Conversely if g XC 0 and f = Re g, then f XR 0 and moreover fC = g.
(iii) If X is a normed C-space, hence a normed R-space, then f XR if and only if fC XC and kf k = kfC k.
Proof. (i) and (ii) are simple exercises.
(iii): For x X , let z(x) = sgn fC (x), so fC (x) = z(x)|fC (x)|. Then we assume f XR :
|fC (x)| = z(x)fC (x) = fC (z(x)x) = |f (z(x)x)| kf kkz(x)xk = kf kkxk
| {z }
R0

and hence kfC k kf k. However kf k kfC k is clear.


4.16 Corollary (to Hahn-Banach). Let X be a normed space, and Y X a subspace with f Y . Then
there is F X such that


F = f
and
kF k = kf k.
Y

Proof. Let p(x) = kf kkxk, for x X . We observe for y Y that


Re f (y) |f (y)| kf kkyk = p(y).
By the Hahn-Banach theorem, there exists F0 XR such that F0 |Y = Re f and F0 p on X , i.e. F0 (x) kf kkxk
and hence
F0 (x) = F0 (x) kf kk xk = kf kkxk,
so |F0 (x)| kf kkxk = p(x). Thus, kF0 k kf k, but clearly kF0 k kf k. If F = R, let F = F0 , otherwise if F = C
we let F = (F0 )C .
4.17 Corollary. If X is a normed space and x X , then there exists f X such that f (x) = kxk.
Proof. Let Y = Fx. Let f0 : Y F be given by f0 (x) = kxk, so clearly f0 Y 0 . Also
kf0 k = sup{|f (x)| : kxk 1} = sup{||kxk : ||kxk 1} 1.
We apply the last corollary to get f X such that f |Y = f0 and kf k = kf0 k 1.

4.18 Theorem. Let X be a normed space and X = (X ) . The map (x 7 x


) : X X given by evaluation
functionals,
x
(f ) = f (x),
x X , f X
is a linear isometry.
11

Proof. That x
(X ) is clear, and it is clear that (x 7 x
) L(X , (X ) ). Now we have for x X , f X
|
x(f )| = |f (x)| kf kkxk
so that k
xk kxk. However, by the last corollary (really, the Hahn-Banach theorem), given we have that there is
f X , kf k 1, such that
|
x(f )| = |f (x)| = kxk,
it follows that k
xk kxk. Thus k
xk = kxk.
4.19 Remark. We have:
(i) If Y is a linear subspace of a normed space X , then its closure Y X is also a linear subspace. Indeed,
observe that x 7 x and x 7 x + y are uniformly continuous on X , hence on Y.
(ii) If X is a normed space, but not complete, then we may define its completion as
X := Xb X .
Since X is a dual space, it is complete, and any closed subset of a complete space is itself complete. This
is, up to isometric isomorphism, the unique Banach space containing a dense copy of X .

Separation by hyperplanes

5.1 Definition. If X is an F-vector space, then:


A hyperplane (containing 0) is any subspace of the form ker f , 0 6= f X 0 .
An (affine) hyperplane is any subset of the form x0 + ker f , x0 X , 0 6= f X 0 .
A R-hyperplane is any set of the form x0 + ker Re f where x0 X , 0 6= f X 0 .
Note ker Re f ker f .
Our goal is a geometric version of Hahn-Banach theorem: given A, B convex sets with A B = , we want to put
a R-hyperplane between them (we want this to be closed in the normed setting).
5.2 Proposition. Let X be a normed space.
(i) If 0 6= f X , then ker f is closed and nowhere dense.
(ii) If f X 0 \ X , then ker f is dense in X .
Thus a hyperplane in X is either closed or dense in X .
Proof. We have:
(i) If 0 6= f X , then ker f = f 1 ({0}) is closed. Also, any proper closed subspace of a normed space is nowhere
dense. Indeed, if Y X is such a subspace, and there is y0 Y, > 0 such that y0 + D(X ) Y, then
D(X ) 1 (Y y0 ) = Y and span D(X ) = X , so X = Y.
(ii) Suppose ker f is not dense in X . Hence there is x0 X and > 0 such that (x0 + D(X)) ker f = . Thus
0
/ f (x0 + D(X )) = f (x0 ) + f (D(X )),
1
f (x0 )

/ f (D(X )) = f (D(X )) = f (D(X )). Thus




1


kf k f (x0 ) .



Indeed, if there were x D(X ) such that |f (x)| > 1 f (x0 ) , then


f (x0 )


f (x) < 1
hence

and we thus have


f

 f (x )  1
0
x = f (x0 )
f (x)

| {z }
kk<1

contradicting our statement before.


12

5.3 Remark. If X is an infinite-dimensional normed space, then in fact X 0 \ X 6= (assuming the Axiom of
Choice). To see this, note that there exists a Hamel basis {e } for X , and we may assume ke k = 1. Find an
unbounded subset { } F. Let f X 0 be given by

X
X
f
e =

(where 6= 0 only for finitely many ). Then


kf k sup |f (e )| = sup | | =

so f is an unbounded linear functional on X .


5.4 Definition. Let X be an F-vector space.
A nonempty subset of X is convex if for any x, y A, and 0 < s < 1, we have (1 s)x + sy A.
A subset A of X is absorbing about x0 A if for every x X , there is  = (x, x0 , A) > 0 such that for
0 s < , we have x0 + sx A.
5.5 Example. If X is a normed space, any open set U X is absorbing about any of its points.
5.6 Lemma. Let X be an F-vector space, and A X be convex and absorbing about 0. Define p : X R
by p(x) = inf{t 0 : x tA} where tA = {ta : a A}. Then p is a sublinear functional8 , which we call the
Minkowski functional (or gauge functional) of A. Moreover,
(i) {x X : p(x) < 1} A {x X : p(x) 1}.
(ii) If X has norm k k, by which A is a neighbourhood of 0, then there is N > 0 such that p(x) N kxk, x X .
Proof. First, note that since A is absorbing at 0, we find for each x, there is t > 0 such that tx = 0 + tx A,
hence x 1t A, and hence the infimum describing p is over a non-empty set, so p(x) is well-defined. Let us verify
sublinearity.
To see positive homogeneity, note that if s > 0, we have
p(sx) = inf{t 0 : sx tA} = inf{t 0 : x st A} = s inf{ st 0 : x st A} = sp(x).
Clearly p(0) = 0. To see subadditivity, first note if s, t 0, we have for a, b A that


s
t
sa + tb = (s + t)
a+
b (s + t)A
s+t
s+t
|
{z
}
convex combination

since A is convex, and hence sA + tA (s + t)A. On the other hand we always have
(s + t)A = {(s + t)a : a A} {sa + tb : a, b A} = sA + tA
so (s + t)A = sA + tA. Now for x, y X we have
p(x) + p(y) = inf{s 0 : x sA} + inf{t 0 : y tA} = inf{s + t : s, t 0, x sA, y tA}
inf{s + t : s, t 0, x + y sA + tA = (s + t)A}
= inf{r 0 : x + y rA} = p(x + y).
Let us prove the remaining claims:
(i) If p(x) < 1, then because A is absorbing, there is 0 < t < 1 so x tA, i.e. 1t x A, but then by convexity of
A, we have x = (1 t)0 + t 1t x A. Also, if x A, then x 1A so p(x) 1.
(ii) Let > 0 be such that D(X ) A. Then for 0 6= x X ,  > 0,
x (kxk + )D(X ) =
Hence p(x)

kxk+

kxk + 
kxk + 
D(X )
A.

1 kxk as  0. Let N = 1 .

8 If A X is not convex, the proof above still shows, at least, that the Minkowski functional of A is finite-valued, non-negative and
positive-homogeneous, with A {x X : p(x) 1}. Furthermore, if A is both convex and balanced, then p is a seminorm on X .
Converses of this result are also true; see Megginsons An Introduction to Banach Space Theory, Chapter 1, Section 9.

13

5.7 Theorem (Separation Theorem/Geometric Form of the Hahn-Banach Theorem). Suppose X is an F-vector space and A, B X are non-empty convex sets with A B = and A is absorbing about a
point a0 . Then there are f X 0 and R such that
Re f (a) Re f (b),

a A, b B.

Moreover, if X admits a norm k k by which A is a neighbourhood of a0 , then f can be chosen to be continuous.


In addition, if A is open, then we can arrange for the inequality to be strict on one side, i.e.
Re f (a) > Re f (b),

a A, b B.

Proof. Let A B = {a b : a A, b B}. It is straightforward to verify that


(i) A B is absorbing around each a0 b, for b B.
(ii) A B is convex.
(iii) In the case that X is normed and A is a neighbourhood of a0 , then A B is a neighbourhood of each a0 b,
for b B. Furthermore, if A is open, A B is open too.
Let x0 = a0 b for some fixed b in B, and set
C = x0 (A B) = {x0 (a b) : a A, b B}.
Then C is convex, absorbing about 0, and is a neighbourhood of 0, given relevant assumptions. Let p be the
Minkowski functional of C. Since A B = , we have 0
/ A B and hence x0
/ x0 (A B) = C. Thus, by the
lemma above, p(x0 ) 1.
Let f0 : Rx0 R be given by f0 (sx0 ) = sp(x0 ). If s 0,
f0 (sx0 ) = sp(x0 ) = p(sx0 ) p(sx0 ),
and if s < 0,
f0 (sx0 ) = sp(x0 ) < 0 p(sx0 ),
XR0

so f0 (sx0 ) p(sx0 ) on Rx0 . Let f


be any Hahn-Banach extension of f0 to all of X , such that f (x) p(x)
for x X . We will show that f satisfies the statement of the Theorem if F = R. If F = C, we will replace f by fC ,
and be done.
If a A, b B, then note x0 (a b) C so we have
f (x0 (a b)) p(x0 (a b)) 1
by (i) of the Lemma. Therefore,
f (x0 ) f (a) + f (b) 1 = f (x0 ) + f (b) 1 + f (a).
Since f (x0 ) = p(x0 ) 1, we have f (b) f (a). Hence, since a A, b B are arbitrary,
sup{f (b) : b B} inf{f (a) : a A}
for some R.
Now assume X admits a norm by which A is a neighbourhood of a0 , hence C is a neighbourhood of 0. By the
lemma, we see p N k k on X , so for x X
f (x) p(x) N kxk,

f (x) = f (x) p(x) N k xk = N kxk

so |f (x)| N kxk so kf k N , i.e. f is continuous. Moreover, if A is open, we cannot have that f (a) = for any
a A, where is as above. Indeed there is  > 0 so that a x0 A, which means that
f (a x0 ) = f (a) p(x0 )
but p(x0 ) 1 so we certainly cannot have f (a) = .
5.8 Definition. In an F-vector space, a (R)-half-space is any set of the form
for some 0 6= f X 0 , R.

H = {x X : Re f (x) },
14

We observe that if X is normed, H is closed if and only if f X 0 is continuous (if is easy because continuous
maps pull back closed sets to closed sets why is only if true?).
5.9 Definition. In an F-vector space, given a non-empty S X , we let its convex hull be given by
X
co S = {1 x1 + . . . + n xn : i [0, 1],
i = 1, xi S, n N}.
We note that
co S =

{C : C S and C is convex},

i.e. co S is the smallest convex set containing S. If X is normed then we let co S = co S, denote the closed convex
hull. We remark that if C X is convex, then its closure C is convex.
5.10 Theorem (Closed Convex Hull Theorem). If X is a normed space and S X is nonempty, then
\
co S = {H : H is a closed half-space containing S}.
Proof. The collection of all closed half-spaces containing S is a collection of closed convex sets, which proves the
direction (arbitrary intersections of closed/convex sets remain closed/convex). It remains to prove . It will
/ H.
suffice to see that for any x0
/ co S that there is a closed half-space H such that co S H and x0
If x0
/ co S, then there is > 0 such that
(x0 + D(X )) co S = ,
but x0 + D(X ) is open and convex, while co S is convex so by the Separation Theorem, there is f X and R
such that
Re f (c) > Re f (x),
c co S, x x0 + D(X ).
In particular, the closed half-space
H = {x X : Re f (x)}
contains co S, but misses x0 .

Consequences of Baire Category Theorem

We now discuss the Banach-Steinhaus Theorem, the Open Mapping Theorem, and the Closed Graph Theorem.
These are traditionally proved as consequences of the Baire Category Theorem9 .
6.1 Theorem (Baire Category Theorem I). If (X, ) is a complete metric space and {Un }
n=1 is a collection
of dense open sets, then

\
Un is dense in X.
n=1

Proof. PM351.
6.2 Definition. Let (X, ) be a metric space.
A set F X is called nowhere dense if its closure F contains a neighbourhood of none of its points, i.e. if
x0 F , there is no > 0 such that
B (x0 ) = {x X : (x, x0 ) < } F .
In other words, F has empty interior. This is equivalent to saying that X \ F is dense in X.
A set M X is called meager (or 1st category) if
M=

Fn

n=1

where each Fn is nowhere dense, and non-meager (or 2nd category) otherwise.
9 In some cases, one can find non-Baire proofs, e.g. of Banach-Steinhaus via gliding hump. By the way, the Open Mapping Theorem
may be proved via Zabrekos Lemma: every countably subadditive seminorm on a Banach space is continuous (which follows from the
fact that every closed, convex, absorbing-about-0 subset in a Banach space contains a nbhd of 0, which is a special incarnation of Baire).

15

6.3 Theorem (Baire Category Theorem II). If (X, ) is a complete metric space, and 6= U X is
open, then U is non-meager.
Proof. If U was meager, i.e. we could write
U=

Fn

n=1

where each Fn is nowhere dense, then each Un = X \ F n would be open and dense, and hence

Un

n=1

is dense in X, hence meets U . This means we cannot write U as above.


A2 posted online since Monday. By next Mondays lecture, we will have covered all relevant lecture material
however pages 1 and 2 are easily accessible now.
6.4 Theorem (Banach-Steinhaus Theorem/Uniform Boundedness Principle). Let X , Y be
normed spaces. If F B(X , Y) is pointwise bounded on a set U non-meager in X , i.e.
sup{kT xk : T F} < ,

x U,

then we have that F is uniformly bounded on X , i.e.


sup{kT k : T F} = sup{kT xk : T F, x D(X )} < .
Proof. Let
Fn = {x X : kT xk n for all T F} =

T 1 (nB(Y))

T F

where T

(nB(Y)) is closed as T B(X , Y) so Fn , as an intersection of closed sets, is again closed. By assumption


U

Fn = U =

(Fn U ).

n=1

n=1

Since U is non-meager, Baire Category says that at least one Fn U has non-empty interior, so there is x0 X ,
> 0 so that
x0 + D(X ) Fn U Fn .
Now, for x D(X ), we can write x =

1
2

Tx =

[x0 + x (x0 x)], hence

1
[T (x0 + x) T (x0 x)],
2 | {z } | {z }
kkn

so kT xk

n
,

so kT k

n
,

T F

kkn

independent of our choice of T F.

6.5 Corollary. If F B(X , Y) is not uniformly bounded, then the subspace


X0 = {x X : sup kT xk < }
T F

is meager in X .
6.6 Theorem (Open Mapping Theorem/Banach-Schauder Theorem). Let X , Y be Banach spaces
and T B(X , Y). Then if T is surjective, i.e. T (X ) = Y, then T is open, i.e. if U X is open, then T (U ) Y is
open.
6.7 Remark. We shall frequently use the following fact: if 6= A X, x X , 0 6= F, then x + A = x + A.
Indeed, ak a in X if and only if x + ak x + a in X .
6.8 Lemma (Main Lemma10 ). Suppose there is r > 0 that
T (D(X )) rD(Y)
then T (D(X )) rD(Y).
10 See

Bollobs Linear Analysis, Chapter 5, Lemma 5.

16

kzk
1
Proof of lemma. Let z rD(Y), and let > 0 be so that kzk < r(1 ) < r. Set y = 1
z, so kyk = 1
< r. We
1
1
will show that y 1 T (D(X )) i.e. there is x 1 D(X ) so T x = y. Then T ((1 )x) = z and so we will have
shown T (D(X )) rD(Y).

Put A = T (D(X )) rD(Y); it follows from our assumptions that A T (D(X )) rD(Y). We proceed inductively:
y rD(Y) A = y1 A (y + rD(Y))
y y1 + rD(Y) y1 + A = y2 (y1 + A) (y + 2 rD(Y))
y yn + n rD(Y) yn + n A = yn+1 (yn + n A) (y + n+1 rD(Y))
Thus, we found a sequence (yn )
n=1 Y. We note
yn+1 yn + n A
kyn+1 yk < n+1 r, kyn yk < n r = kyn+1 yn k < n+1 r + n r
yn+1 yn + n A = yn+1 yn n A
and A = T (D(X )) rD(Y) which yields
yn+1 yn n rD(Y), i.e. kyn+1 yn k < n r
yn+1 yn n T (D(X )) = T ( n D(X ))
i.e. x X , kxk < n s.t. yn+1 yn = T xn . Moreover, we similarly find x0 D(X ) such that y1 = T x0 . We also
note that yn+1 y + n+1 rD(Y) hence kyn+1 yk < n+1 r implies
lim yn = y.

Now
x=

xn ,

n=0

note that
kxk

kxn k <

n =

n=0

n=0

1
.
1

Using the linearity and continuity of T we have


Tx =

T xn = y1 +

n=0

(yn+1 yn ) = . . . = yN +
|{z}

n=1

by
kT x yk kyN yk +

X
n=N

(yn+1 yn ) = y
{z
}
|
kk< n r

n r 0

n=N

hence kT x yk = 0, as claimed.
Proof of Open Mapping Theorem. It suffices to show that T (D(X )) rD(Y) for some r > 0. Indeed, supposing
this, note that if U X is open and x U then x + D(X ) U for some > 0. Thus, U x D(X ). Hence
T (U x) T (D(X )) rD(Y),
so that
T x + rD(Y) T x + T (U x) = T (U ).
Thus, T (U ) is indeed open. Now, since T (X ) = Y we see that
Y=

nT (D(X ))

n=1

S
S
because nT (D(X )) = T ( nD(X )). Since Y is a Banach space, the Baire Category Theorem tells us that some
nT (D(X )) is not nowhere dense, hence there is y0 Y and > 0 such that
y0 + D(Y) nT (D(X )) = nT (D(X )).
17

Since nT (D(X )) is convex, and symmetric, i.e.


nT (D(X )) = nT (D(X )) = nT (D(X ))
and for y D(Y), we have that y is just the midpoint of two vectors in y0 + D(Y) nT (D(X )), that is,
y =

1
[(y0 + y) (y0 y) ] nT (D(X ))
2 | {z } | {z }
y0 +D(Y)

so that

ny

T (D(X )), which yields that

n D(Y)

y0 +D(Y)

T (D(X )). Hence by the Main Lemma,

n D(Y)

T (D(X )).

6.9 Theorem (Inverse Mapping Theorem). If X , Y are Banach spaces and T B(X , Y) is bijective, then
T 1 B(Y, X ).
Proof. Since T L(X , Y) and bijective, we clearly have that T 1 L(Y, X ). Since T B(X , Y) and surjective,
the Open Mapping Theorem says there is r > 0 such that T (D(X )) rD(Y). Applying T 1 , we see that D(X )
rT 1 (D(Y)) so 1r D(X ) T 1 (D(Y)), i.e. kT 1 k 1r < , so that T 1 B(Y, X ).
6.10 Definition. Let X , Y be normed spaces and X Y denote their direct sum, i.e.
X Y = {(x, y) : x X , y Y}.
If 1 p < , let
k(x, y)kp = (kxkp + kykp )1/p ,

k(x, y)k = max{kxk, kyk}.

It is trivial to check that {k kp : 1 p } is a family of norms on X Y. If X , Y are Banach spaces, then


X p Y = (X Y, k kp )
is also a Banach space for each p.
6.11 Fact. Any norm on X Y such that k(x, 0)k = kxk and k(0, y)k = kyk, is equivalent to k k1 .
6.12 Theorem (Closed Graph Theorem). Let X , Y be Banach spaces, and T L(X , Y). Then T B(X , Y)
if and only if its graph,
(T ) = {(x, T x) : x X } X 1 Y,
is closed.
Proof. () Note that if xn x in X , then T xn T x in Y. Hence if (x, y) (T ), then there is (xn )
n=1 X
such that (x, y) = limn (xn , T xn ). However by the above, limn (xn , T xn ) = (x, T x). Hence (x, y) = (x, T x) (T ),
so (T ) = (T ).
() If (T ) is closed in the Banach space X 1 Y, then (T ) itself is a Banach space (it is trivial to see (T ) is a
subspace). Define 1 : (T ) X by 1 (x, T x) = x. Note that
k1 (x, T x)k = kxk kxk + kT xk = k(x, T x)k = k1 k 1.
Clearly, 1 is linear, so 1 B((T ), X ). Also, it is clear that 1 is a bijection. Thus by the Inverse Mapping
Theorem, 1 1 B(X , (T )). Hence, if x X ,
kT xk kxk + kT xk = k(x, T x)k1 = k1 1 xk1 k1 1 kkxk
and hence kT k k1 1 k < .
6.13 Proposition (Closed Graph Test). Given normed spaces X , Y and T L(X , Y), we have (T )
X 1 Y is closed if and only if whenever xn 0 in X and T xn y in Y, then it must be the case that y = 0.
The point is that we know/assume that limn T xn exists.
Proof. Note that in (T ), (xn , T xn ) (x, z) if and only if (xn x, T (xn x)) = (xn x, T xn T x) (0, z T x).
That is, [(x, z) (T ) if and only if z = T x] occurs if and only if [T xn y if and only if y = T x].
6.14 Definition. Let Y be a normed space and X Y be a subspace. We say that X is (boundedly) complemented if there is P B(Y, Y) = B(Y) such that
Im P = P (Y) = X
and P is idempotent, i.e. P 2 = P P = P .
18

6.15 Remark (obvious projection). In general, if X Y is a subspace, let B be a basis for X . Then
there is B 0 Y so B B 0 is a basis for Y. Hence each y Y admits scalars {ye }eBB 0 such that ye 6= 0 for only
finitely many elements e, and
X
ye e.
y=
eBB 0

Define P : Y X by
Py =

ye e.

eB

Then P 2 = P , and Im P = X . However, it is unlikely that such P is bounded.


Assignment #2 due Wednesday October 10. Office hours Friday 2-3:30, Tuesday 2:30-4. Assignment #1 will
be available for retrieval in the wooden box adjacent to my office door as of about 11am today.
6.16 Theorem. If Y is a Banach space and X Y is a closed subspace, then X is boundedly complemented in Y
if and only if there is a closed subspace Z Y such that X Z = {0} and X + Z = Y.
Proof. () We suppose there is P B(Y) such that P 2 = P and Im P = X . Let Z = ker P . If x X Z, then
x = P x = 0, so indeed X Z = {0}. Also if y Y, we have
y = P y + (I P )y
|{z} | {z }
X

Z, check

so that X + Z = Y. Also Z = P 1 ({0}), so it is closed.


() Let J : X 1 Z Y be given by J(x, z) = x + z. Then J is linear, ker J = 0, i.e. as X + Z = Y.
Also kJ(x, z)k = kx + zk kxk + kzk = k(x, z)k1 , which implies kJk 1. By the Inverse Mapping Theorem,
J 1 B(Y, X 1 Z).
Let i : X , Y be the inclusion map and P : X 1 Z X be given by P (x, y) = x, so kP k 1. Let P = i P J 1 .
Then Im P = X . Also, for y Y,
kP yk = ki P J 1 yk kP J 1 yk kJ 1 yk kJ 1 kkyk
so kP k kJ 1 k < hence P B(Y). Finally, observe that
1
1
1

P2 = i P
| J{z }i P J = i P J = P
identity on X

6.17 Remark (Murray, 1940s). If 1 p (p 6= 2), then `p admits subspaces which are not boundedly
complemented11 .
6.18 Theorem. c 0 is not boundedly complemented in ` .
Proof. We suppose, for sake of contradiction, that P B(` ) exists with Im P = c 0 , P 2 = P . Let F P(N) be
an uncountable collection of infinite sets, such that for distinct sets F1 , F2 F, |F1 F2 | < . For each F F,
let yF = (I P )F . Note that c 0 = ker(I P ), and clearly F
/ c 0 , so yF 6= 0 for any F . Observe for distinct
F1 , . . . , Fm F and 1 , . . . , m F, we have
m
X
i=1

i Fi =

m
X

i Fi \Gi +

kk =

{z

max

i=1,...,m

where Gi =

(j1 + . . . + jk )Fj1 ...Fjk

k=2 1j1 ...jk m

i=1

m
X

|i |

{z

cc0

Fj . Hence
m


m

m
X


X

X





i yFi = (I P )
i Fi kI P k
i Fi \Gi = kI P k max |i |.

i=1,...,m




j6=i

i=1

i=1

i=1

Now, set for n, k N,


Fn,k = {F F : |k (yF )| >

1
n}

11 In fact, it was proved by J. Lindenstrauss and L. Tzafriri in 1971 (On the complemented subspaces problem, Israel J. Math. 9 (1971),
263-269; MR 43 #2474) that a Banach space is isomorphic to a Hilbert space if all of its closed subspaces are boundedly complemented!

19


where k ((xi )
i=1 ) = xk , so k ` with kk k = 1. Now, if F1 , . . . , Fm are distinct elements of Fn,k , then with
i = sgn k (yFi ), we have

m
m
m
X
X
X
m



i k (yFi ) =
|k (yFi )|
i yFi
kI P k



| {z }
n
i=1

i=1

i=1

1/n

so m nkI P k, hence Fn,k is finite. Since each yF 6= 0, for F F, we see that


F=

Fn,k

n=1 k=1

but Fn,k is finite, which contradicts that F is uncountable.

Finite-dimensional Banach spaces

7.1 Lemma. Given a finite-dimensional R-vector space, let {e1 , . . . , ed } be a basis, and let kxk1 =
Pd
x = i=1 xi ei . If B = Bkk1 (X ), then B is compact.

Pd

i=1

|xi | where

Proof. Let us accept the Bolzano-Weierstrass Theorem, any sequence in [1, 1] has a convergent subsequence. Let
(x(n) )
n=1 B. We will show that this has a convergent subsequence.
(nk(1) )
)k(1)=1

(n)

(x1 )
n=1 [1, 1] has a convergent subsequence (x1
(nk(1) )
)k(1)=1

(x2

(nk(2) )
)k(2)=1

[1, 1] has convergent subsequence (x2

.
..
(nk(d1) )
)k(d1)=1

(xd

(nk(d) )
)k(d)=1

[1, 1] has convergent subsequence (xd

Check that (x(nk(d) ) )


k(d)=1 converges in B.
7.2 Theorem. Let X be a finite-dimensional F-vector space. Then any two norms on X are equivalent.
Proof. Given a norm k k on X , we will show that k k k k1 , where k k1 is the R-norm given with respect to
R-basis {e1 , . . . , ed }.
Pd
First, let M = sup kei k. Then for x = i=1 xi ei ,
i=1,...,d

kxk

d
X

kxi ei k =

i=1

d
X

|xi |kei k M kxk1 .

i=1

Thus we have for x, y X ,


|kxk kyk| kx yk M kx yk1
so k k is k k1 -continuous. Now if S = Skk (X ), this is closed in Bkk1 (X ), and hence compact. Thus
m = min kxk
xS

exists, and m > 0. Now if x X \ {0} we have








1

x
m

1
|kxk
{z }
S

hence mkxk1 kxk.


7.3 Corollary. Let (X , k k) be a finite-dimensional normed space.
(i) (Heine-Borel) A subset K X is compact if and only if K is closed and bounded.
20

(ii) (X , k k) is complete, hence a Banach space.


(iii) If Y is any normed space, then L(X , Y) = B(X , Y).
(iii) X 0 = X .
Proof. We have:
(i) () Straightforward exercise from PM 351.
() We have an m > 0 such that k k mk k1 . Hence Bkk (X ) mBkk1 (X ), so for any r > 0, rBkk (X )
rmBkk1 (X ). The map x 7 rmx on X is continuous, so rmBkk1 (X ) is compact. Then K rBkk (X )
rmBkk1 (X ), for large enough r, and is a closed subset, hence compact.

(ii) Given a Cauchy sequence (xn )


n=1 X the set {xn }n=1 is bounded, so {xn }n=1 is compact. A Cauchy
sequence has a unique cluster point, hence limit point.

(iii) Given T L(X , Y), we let for x X


|||x||| = kxk + kT xk.
Then ||| ||| is a norm, so there is M > 0 so that ||| ||| M k k. Then we have for x X that
kT xk kxk + kT xk = |||x||| M kxk.
Hence kT k M < .
(iii) Let Y = F.

Balls in finite-dimensional spaces


If k k is a norm on X , then Bkk (X ) is convex, symmetric (x Bkk (X ) whenever x Bkk (X )) [if F = C then
Bkk (X ) is balanced (zx Bkk (X ) whenever |z| 1 in C and x Bkk (X ))], and absorbing about 0.
If we are given, in an F-vector space, a set B such that B is convex, symmetric (balanced if F = C) and absorbing
about 0, then the Minkowski functional k kB is a norm (exercise).
Given a norm k k on X , we have
k kBkk (X ) = k k.
Now suppose dim X < . We have that if A X is convex and absorbing about 0, then A is a neighbourhood of
0 (w.r.t. any norm we may put on X ). Indeed, since A is absorbing at 0, given a R-basis {e1 , . . . , ed } for X , there
are 1 , . . . , d > 0 such that i ei = 0 i ei A. Let
B = co{i ei : i = 1, . . . , d} A.
Check that B is closed. We note that B is convex, symmetric, and absorbing at 0, and hence
B = BkkB (X ) A
i.e. DkkB (X ) A.
7.4 Proposition. If X is a finite-dimensional subspace of a normed space Y, then X is closed and boundedly
complemented.
Proof. Let {e1 , . . . , ed } be a basis for X . Define f1 , . . . , fd X 0 by

d
X
fi
j ej = i
j=1

(1 , . . . , n F). Since X is finite-dimensional, with the norm inherited from Y, X 0 = X . We find Hahn-Banach
extensions F1 , . . . , Fd for f1 , . . . , fd respectively so each Fi Y . Then define P : Y Y by
Py =

d
X

Fi (y)ei .

i=1

21

Then P is linear, Im P = X , P 2 = P , and


kP k

d
X

k|Fi kkei k < .

i=1

We note that X , with norm from Y, is complete, hence closed. Alternatively, X = ker(I P ), which is closed.
7.5 Theorem. Let X be a normed space. Then B(X ) is compact if and only if X is finite-dimensional.
Proof. () Heine-Borel, above.
7.6 Lemma (Rieszs Lemma). If X is a normed space and Y is a proper closed subspace, then given  > 0,
there is x0 B(X ) such that d(x0 , Y) 1 .
Proof. Given x X \ Y, we define f : Y + Fx F by f (y + x) = for y Y, F. We note that ker f = Y
which is closed so f is continuous. (Propn before Separation Theorem). Thus by Hahn-Banach Theorem, there is
an extension F X of f . Find x0 B(X ) such that |F (x0 )| (1 )kF k. We observe for any y in Y
1
|F (x0 )|
|F (x0 y)| =
1 .
kF k
kF k

kx0 yk

Proof of () of theorem. We show that if dim X 6< , then B(X ) is not compact. We perform an induction; given
0 <  < 1, find
x1 S(X ), i.e. kx1 k = 1.
x2 B(X ) such that d(x2 , Fx1 ) 1 .
x3 B(X ) such that d(x3 , span{x1 , x2 }) 1 .
.
..
xn B(X ) such that d(xn , span{x1 , . . . , xn1 }) 1 .
Note, at each stage, span{x1 , . . . , xn } is finite-dimensional, and hence closed. Moreover, dim X 6< , so these
finite-dimensional subspaces are proper. We note if n > m then
kxn xm k d(xn , span{x1 , . . . , xn1 }) 1 
and we see that (xn )
n=1 B(X ) admits no Cauchy subsequence.

Initial topologies, compactness

8.1 Definition. Let X be a non-empty set, {(X , )}A be a collection of topological spaces, and {f : X
X }A be a family of functions. We define the initial topology = (X, {f }A ) as follows:
U for each x U , there exist U1 1 , . . . , Un n (n N) such that x

n
\

f1
(Ui ) U.
i

i=1

Hence the subsets

n
\

)
f1
(Ui )
i

: n N, each i A, each Ui i

i=1

form a base for , i.e. any U in is the union of such sets. We might call the sets {f1 (U ) : A, U } a
subbase, i.e. finite intersections of such sets form a base.
8.2 Remark. Each f : (X, ) (X , ) is - -continuous (A1Q1). In fact, is the coarsest topology on
X which allows each f to be continuous, i.e. if P(X) is any topology for which each f : X X is
- -continuous, then . We remark that is trivially a topology: it is closed under finite intersections, and
arbitrary unions.
8.3 Example (Metric Topology). Let : X X [0, ) be a metric on X. For each x X, let
x (y) = (x, y), so {x : X R}xX is a family of functions on X (we equip R with its usual topology, of course).
Its not difficult to show that
= (X, {x }xX ).
22

8.4 Example (Product Topology). Let {(X , )}A be a collection of topological spaces. Let
Y
X=
X = {x = (x )A : x X for all A}
A

denote the Cartesian product. For each , denote by : X X the canonical projections, i.e. x = x .
Then we define the product topology on X by
= (X, { }A ).
8.5 Remark. Note the following:
(i) If A = N, then basic open neighbourhoods of any point in X =
n
\

nN

Xn , are given by

p1
i (Ui ) = U1 U2 . . . Un Xn+1 Xn+2 . . .

i=1

(ii) Consider the set RR (functions from R to R) with the product topology . Were effectively selecting finitely
many points t1 , . . . , tn . [DIAGRAM]
8.6 Example (Linear Topologies). Let X be a normed space and Z X . We may write Z = {f : X
F}f Z . The initial topology (X , Z) is called the linear topology on X from Z. Note that (X , Z) kk . In
particular:
The weak topology is defined as w = (X , X ) P(X ).
The weak* topology is defined as w = (X , X ) P(X ).
8.7 Remark. Recall that X X = {
x : x X }, where x
(f ) = f (x). What do subbasic w-open sets of X look
like? Fix f X \ {0}, x0 X . Then
f 1 (f (x0 ) + D) = {x X : f (x) f (x0 ) + D} = {x X : f (x) f (x0 ) D} = {x X : |f (x) f (x0 )| < }
or alternatively, f 1 (f (x0 ) + D) = {x X : |g(x) g(x0 )| < 1} where g = 1 f . Here,
(
(1, 1)
if F = R
D = D(F) =
{z : |z| 1} if F = C.
8.8 Example (Relative Topology). Let (Y, ) a topological space. For X Y , let : X , Y denote the
canonical injection/embedding map. The relative (or relativized) topology on X is denoted
|X = (X, {}).
Note U |X if and only if there exists V such that U = X V .
8.9 Definition. If (X, ) is a topological space, a set K X is -compact if given any open cover of K, i.e. any
family O such that
[
K
U
U O

there is a finite subcollection {U1 , . . . , Un } O such that


K

n
[

Ui .

i=1

We say that (X, ) is a compact space if X itself is ( -)compact.


8.10 Definition. Let (X, ) be a topological space. A set F X is ( -)closed if X \ F . If S X is any set,
we define its -closure by
\

S = S = {F X : F S and F is -closed}.

Since unions of open sets are open, intersections of closed sets are closed. Thus S is the smallest -closed set
containing S.
8.11 Proposition. S = {x X : for every U with x U , we have U S 6= }.
23

Proof. () If for x X, there exists a -neighbourhood U of x such that U S = , then S X \ U so x


/ S.
() If x
/ S then x X \ S so x is in the set described by the RHS.
8.12 Proposition. A topological space (X, ) is compact if and only if forTeach family F P(X)Twith FIP (finite
n
intersection property, i.e. for any finite collection F1 , . . . , Fn F we have i=1 Fi 6= ), we have F F F 6= .
T
Proof. () Suppose (X, ) is compact. If F P(X) is such that F F F = , then {X \ F : F F } is an open
cover of X, and so admits a finite subcover {X \ Fi }ni=1 . But then
X=

n
[

(X \ Fi ) = X \

i=1

n
\

Fi =

i=1

n
\

Fi = =

i=1

n
\
i=1

Fi

n
\

Fi =

i=1

so we dont have FIP.


() If O is a -open cover of X, then F = {X \ U : U O} is a family of closed sets. If O admitted no finite
subcover, then every finite collection {X \ Ui }ni=1 F would have
n
\

(X \ Ui ) 6= .

i=1

However, by assumption, this would mean that


\

(X \ U ) 6= ,

U O

but then

O ( X violating the fact that O is an open cover.

8.13 Definition. Given a set X, an ultrafilter is a collection U P(X) such that:


1. U has the FIP.
2. For any A X, either A U or X \ A U (note that both cannot hold, else FIP is violated).
8.14 Example. If x X, then Ux = {U X : x U } is an ultrafilter, called a principal (or trivial) ultrafilter.
8.15 Lemma (Ultrafilter Lemma). If X 6= and F P(X) has FIP, then an ultrafilter U F exists.
Proof. See Dr. Spronks website for a tidied-up proof.
Tn
(Repair: for G1 , G2 G, G3 such that 6= G3 G1 G2 , F 0 = { i=1 Fi : F1 , . . . , Fn F, n N} then F 0
so 6= ).
Let S
= {G P(X) : F G, G has FIP}. We partially order by inclusion. If is a chain in then let
G = TG G. If G1 G1 , . . . , Gn Gn for some Gi , up to reindexing G1 . . . Gn so G1 , . . . , Gn Gn .
n
Hence i=1 Gi 6= since Gn . Trivially, F G so G and is an upper bound. Hence by Zorns Lemma, a
maximal element U of exists.
For each instance of G1 , . . . , Gn replace it G1 , G2 G then G1 G1 , G2 G2 for some G1 , G2 , up to reindexing,
G1 G2 so G3 G2 such that 6= G3 G1 G2 .
We observe A X and if A U 6= for each U in U then U {A U }U U . Indeed, if U1 , U2 U, then there
is U3 U1 U2 and A U3 6= so A (U1 U2 ) 6= hence U {A U }U A = U by max. of U. If A U = for
at least one U in U, then (X \ A) U , so U {X \ A} so again U {X \ A} = U by max. of U.
8.16 Corollary. Non-principal ultrafilters exist for infinite X.
Proof. Let F0 = {X \ E : E X is finite} so that F0 has FIP. Any ultrafilter U containing F0 is non-principal, i.e.
U 6= Ux for any x X.
Q
8.17 Theorem (Tychonoffs Theorem). If {(X , )}A is a family of compact spaces, then X = A X
is compact (with the product topology ).
Proof. See Dr. Spronks website for a tidied-up proof.
Let F P(X) have FIP. We let U be any ultrafilter containing F. Fix, for the moment, A. We observe that
{p (U )}U U has FIP in X . Indeed, if U1 , . . . , Un U then
!
n
n
\
\
p (Ui ) p
Ui 6=
i=1

i=1

| {z }
6=

24

(p has full domain(?)). Then by an earlier proposition, U U p (U )


6= in X . Thus there is x
T

and hence for any -neighbourhood V of x , V p (U ) 6= for each U U. Then p 1 (V )U 6=


U U p (U )
for any U U.
We
for all and obtain x = (x )A X. If 1 , . . . , n T
A and xi TVi i thenTthe -neighbourhood
Tn do this

1
p
(V
)

U.
Hence
x

U
for
each
U

U
and
thus
x

i
i=1 i
U U U
F F F so
F F F 6= . Hence
(X, ) is compact.

An application of ultrafilters: ultrafilter limits

Let F0 = {N \ E : E P(N) is finite}. Observe that F0 certainly has the Finite Intersection Property. Let U be
any ultrafilter containing F0 . Define U : P(N) {0, 1} R by
(
1 if A U
U (A) =
0 if A
/ U.
We note that U () = 0. Also, if A, B P(N) and A B = then at most one of A or B is in U, and hence
U (A) + U (B) = U (A B) (using the property that for E P(N), either E U or N \ E U). If we have a
partition
n
G
N=
Ei ,
i=1

then exactly one Ei U and the others are not. It follows that the variation V (U ) = 1. Hence U F A(N); see
A1Q4. By A1Q4, there is LU ` such that LU (S ) = U (S). We note the following facts (the proofs are similar
to A2Q4):
(i) LU (1) = 1, where 1 denotes the all-ones sequence, and kLU k = V (U ) = 1.
(ii) LU |c 0 = 0, LU (x) 0 if xn 0 for all n, and
lim inf xn LU (x) lim sup xn
n

(if each xn R).

(iii) LU is not translation-invariant:


LU (2N ) 6= LU (2N1 )
noting that only one of 2N or 2N 1 is a member of U. However LU (2N1 ) = LU (1 2N ).
9.1 Definition. We call LU ` the ultrafilter limit given by U.
We can construct Banach limits: Cesaro operator: S B(` ), where we define
Sx = (x1 , 21 (x1 + x2 ), 31 (x1 + x2 + x3 ), . . .).
Note that kSk = 1. We note that Sx S(1 x) c0 and hence LU S is a Banach limit.
9.2 Proposition. The cardinality of the set of ultrafilters on N is c.
Proof. Let F be a family of infinite sets, |F| = c, and |F E| < for any F 6= E in F. For F F, define
GF = F0 {F } which has FIP (where F0 consists of the cofinite sets in N, as above). Let UF GF be an ultrafilter.
Since |F E| < for F 6= E in F, we find that UF 6= UE .
9.3 Remark. In fact, the cardinality of ultrafilters on N is 2c .
On A3, we see that one can create product spaces that are compact, but not sequentially compact (i.e. compact
spaces in which sequences do not have convergent subsequences. This will justify our development of nets, to come).
9.4 Definition. A topological space (X, ) is Hausdorff if for x 6= y in X, we have neighbourhoods x Ux
and y Uy such that Ux Uy = .
9.5 Example. We have:
(i) Any metric space is Hausdorff: indeed if is a metric, then (x, y) = 0 if and only if x = y.
(ii) If X is a normed space, then the weak topology (X , X ) is Hausdorff (consequence of the Hahn-Banach
Theorem).
25

(iii) The weak-* topology (X , X ) is Hausdorff (by definition).


(iv) If {(X , )}A is a family of Hausdorff spaces, then their product X =
with the product topology .

X is Hausdorff when equipped

Proof. If x 6= y in X, find in A such that x 6= y . Find neighbourhoods x U , y V and


1
1
U V = , then it is easy to check that p1
(U ) p (V ) = p (U V ) = .
(v) Sierpinski space: ({0, 1}, ), where = {, {1}, {0, 1}}. Notice that this is not Hausdorff.
9.6 Exercise. If (X, ) is a topological space, then C((X, ), ({0, 1}, )) = {U : U }, that is, the continuous
functions are precisely the indicator functions of -open sets.
9.7 Proposition. We have:
(i) If (X, ) is compact, and K X is -closed, then K is -compact.
(ii) If (X, ) is Hausdorff, then any -compact set K X is closed.
9.8 Remark. In the Sierpinski space ({0, 1}, ), {1} is compact but not closed.
Proof. We have:
(i) Let O be an open cover for K. Then O {X \ K} is a cover for X. Since X is compact, that cover
admits a finite subcover for X, which is hence itself a finite subcover for K.
(ii) Fix x X \ K. SFor each y K find neighbourhoods y Uy , x Vy S so that Uy Vy = . We
n
haveTthat K {Uy : y K}Tso there areSy1 , . . . , yn K such
Tn that K i=1 Uyi . Now we have that
n
n
n
x i=1 Vyi and moreover ( i=1 Vyi ) ( i=1 Uyi ) = , so ( i=1 Vyi ) K = . Thus X \ K is open.
9.9 Proposition. We have:
(i) If (X, ) is compact, (Y, ) is a topological space and : X Y is --continuous, then (X) is -compact.
(ii) If (X, ) is compact, (Y, ) is Hausdorff and : X Y is a --continuous bijection, then 1 : Y X is
- -continuous.
Proof. We have:
(i) If O is an open cover of (X), then by A1Q1, {1 (U )}U is a cover of X. Extract a finite subcover,
using compactness of X.
(ii) If K X is closed, then K is -compact, hence (K) is -compact and hence -closed, as is Hausdorff.
Hence we have for -closed K X that (1 )1 (K) = (K) which is closed. Use A1Q1.
Final exam date: TBA.
Next week: Im away. Well have a guest lecturer E. Elgun.
Office hours: F 2:30-3:30, week of Oct. 29, M 2:30-4, T 2:30-4
9.10 Theorem (Alaoglus Theorem). If X is a normed space, then B(X ) is w -compact.
Proof. Let : X , FX be given by (f ) = (f (x))xX (its injectivity is clear). If U F is open, and x X , then
f x
1 (U ) f (x) U (f ) x1 (U ) (here, x : FX F are the coordinate projections), so
(
x1 (U )) = x1 (U ) (X ).
Thus if x1 , . . . , xn X , and U1 , . . . , Un F are open, we see that
!
n
n
n
\
\
\
1

x
i (Ui ) =
(
x1
(U
))
=
x1
(Ui ) (X )
i
i
i
i=1

i=1

{z

i=1

basic set in w

and hence : X (X ) is an open map from w to |(X ) , thus by A1Q1, 1 : (X ) X is -w-continuous.

We prove, now, that (B(X )) is -closed in FX . Suppose g = (g(x))xX (B(X )) . Now for x, y in X , F
and for each n in N we have
1
6= (B(X )) x1 (g(x) + n1 D) y1 (g(y) + n1 D) x+y
(g(x + y) + n1 D)

26

(D = D(F)) so contains (fn ) for some fn B(X ). We note that


|g(x + y) (g(x) + g(y))| |g(x + y) fn (x + y)| + |g(x) fn (x)| + |||g(y) fn (y)|
<

1
1
||
2 + ||
+ +
=
n n
n
n

(0 = fn (x + y) (fn (x) + fn (y))) which holds for all n, so g, i.e. x 7 g(x), is linear. Likewise, we see that for
each x in X we find fn (B(X ))
|g(x) fn (x)| <

1
1
1
= |g(x)| |fn (x)| + kxk +
n
n
n

and it follows that kgk 1. Hence g = (g), g B(X ).


Q
We now observe that (B(X )) xX kxkB FX (B = B(X )). The latter set is compact, by Tychonoffs
theorem; moreover (B(X )) is -closed, and thus itself -compact. Then
B(X ) = 1 ((B(X )))
| {z }
-compact

and hence B(X ) is w -compact, by the continuity of 1 .


9.11 Corollary. Any bounded w -closed subset A X is w -compact.
Proof. The proof could have easily been conducted on rB(X ) for any r > 0, so rB(X ) is w -compact. If A is
closed, and A rB(X ) for some r > 0, then A is a closed subset of a compact set.

10

Nets

These are like sequences in some ways, but unlike sequences in other ways.
10.1 Definition. We make the following definitions.
A directed set is a pair (N, ) where is
a preorder on N (reflexive and transitive, but not necessarily antisymmetric: we allow 0 , 0
without = 0 ).
a cofinal ordering, i.e. 1 , 2 N , there exists 3 such that 1 3 , 2 3 .
Let X 6= . A net is a map ( 7 x ) : N X where (N, ) is a directed set.
If (M, ) is another directed set, a map : M N is directed if
0 in M implies () (0 ) in N .
is cofinal, i.e. for any in N , then there is M such that ().
If (x )N X is a net, and : M N is a directed map, then (x() )M is called a subnet of (x )N .
We often write = () and hence (x )M .
If (x )N X is a net and A X, we say that (x )N is
eventually in A if there is A N such that x A whenever A .
frequently in A if for any in N , there is 0 such that x 0 A.
10.2 Example. We have:
(i) Sequences are nets.
(ii) R, [0, ), [a, b] are directed sets via usual . So any map [a, b] X is a net.
(iii) (Riemann sums) Fix a compact interval [a, b] R. Let
N = {(s0 , s1 , . . . , sn ; t1 , . . . , tn ) : a = s0 < s1 < . . . < sn = b, ti [si1 , si ], i = 1, . . . , n, n N}.

27

We say that (s0 , . . . , sn ; t1 , . . . , tn ) (s00 , . . . , s0m ; t01 , . . . , t0m ) iff {s0 , . . . , sn } {s00 , . . . , s0m }. Inspect that this
is a cofinal preordering. If f : [a, b] F is a function, a Riemann sum is given for = (s0 , . . . , sn ; t1 , . . . , tn )
by
n
X
S (f ) =
f (ti )(si si1 ).
i=1

(iv) Let X 6= . A family of sets F P(X) is filtering if for any F1 , F2 F, there is 6= F3 F such that
F1 F2 F3 . Now let N = {(x, F ) : x F F }. We let (x, F ) (x0 , F 0 ) iff F F 0 . This is a cofinal
preordering. We let (x)(x,F )N X and this is a net.
10.3 Definition. Let X be a non-empty set, (x )N a net on X. We say that (x )N is an ultranet if for any
A X, either (x )N is eventually in A, or is eventually in X \ A.
10.4 Remark. In general, if (x )N is a net in X and A X, then either (x )N is frequently in A or is
eventually in X \ A.
10.5 Proposition. If (x )N is a net in X and A X is such that (x )N is frequently in A then there is a
subnet (x )M which is eventually in A. Moreover, this subnet can be arranged to be an ultranet.
Proof. Let = {F P(X) : A F, F is filtering, and for all F in F, (x )N is frequently in F }. Recall: filtering, F1 , F2 F, then there is 6= F3 such that F3 F1 F2 . Observe that {A} , if F = A {x 0 }N, 0
then {F }N , i.e. if 1 , 2 N , 3 1 , 2 and F3 = F1 F2 and since (x )N is frequently in A, there is
3 such that x A, so F3 6= .
Given F we let
M = {(, F ) : x F F}
0

We preorder M , (, F ) ( , F ) iff 0 in N , F F 0 . [We let ,F = . ()] We remark that this preordering is cofinal: if (1 , F1 ), (2 , F2 ) M , find 6= F3 F1 F2 and 3 1 , 2 such that x3 F3 , so
(3 , F3 ) (1 , F1 ), (2 , F2 ). [] It is obvious that (, F ) 7 is a directed map. Thus (x )(,F )M is a subnet of
(x )N . Moreover, (x )(,F )M is eventually in A. Indeed if (0 , F0 ) M such that F0 A (and F0 F), which
exists by assumption on , then for (, F ) (0 , F0 ) we have x0 F F0 A, so x , (, F ) (0 , F0 ) in M is in
A.
Now, if F , the Ultrafilter Lemma provides an ultrafilter U F. We observe for U U, that U F 6= for each
F in F, hence U (check!). Then (x )(,U )M (where M is as above, U replacing F) is an ultranet (check!).
10.6 Definition. Suppose (X, ) is a topological space, and (x )N is a net in X. If x0 X, we say that
x0 is a limit point for (x )N if each neighbourhood x0 U has that (x )N is eventually in U .
x0 is a cluster point for (x )N if each neighbourhood x0 U has that (x )N is frequently in U .
In the first case we write x0 = -lim x . We may abuse notation and write -lim rather than -lim if is a metric.
N

10.7 Proposition. If (x )N is a net in a topological space (X, ) and x0 in X is a -cluster point, then there is
a subnet (x )M such that x0 = -limM x .
Proof. We let Ox0 = {U : x0 U }. Then Ox0 is a filtering family of sets, and for U in Ox0 , (x )N is
frequently in U . Consider the subnet (x )M of the prior proposition. Check that for U in Ox0 , (x )M is
eventually in U .
10.8 Remark. By A3Q5, f : (X, ) (Y, ) is continuous if and only if it preserves net convergence.
(i) If , P(X) are topologies, then iff id : (X, ) (X, ) is continuous, iff -limN x = x0 whenever
-limN x = x0 .
(ii) In FX , the product topology is the topology of pointwise convergence. If (f (x))xX FX , (f )N FX is
a net, then f0 = -limN f iff for each x in X, f0 (x) = limN f (x) in F (F has metric topology). Check!
Q
(iii) X = A X , each (X , ) a topological space, then a net (x )N converges to x(0) in (product topology)
()
(0)
if and only if -limN x = x .
10.9 Theorem (Metrisation Theorem). Let X be a separable normed space. Then (B(X ), w ) is metrisable.
Proof. Since X is separable, B(X ) contains a dense countable set {xn }
n=1 (from this, we can extract a linearly
0
independent, countable set {x0n }
such
that
span{x
}
is
dense
in
X
. We can use x0n s in place of elements xn
n n=1
n=1
28

in this proof). Let for f, g X


(f, g) =

X
|f (xn ) g(xn )|
2n
n=1

Observe |f (xn ) g(xn )| = |(f g)(xn )| kf gk kxn k kf gk < , so (f, g) kf gk < . It is


| {z }
1

straightforward to see that this is a metric. Let us check:


non-degeneracy: If f 6= g then for at least one xn , f (xn ) 6= g(xn ) so (f, g) > 0.
triangle law: (f, g) (f, h) + (h, g) for f, g, h X which is straightforward.
Now we show that for (f )N , a net in B(X ), w* -limN f = f0 implies -limN f = f0 . We have if
w* -limN f = f0 , then given  > 0, there is n0 such that

X
k=n0 +1


2
< ,
k
2
2

and then for k = 1, . . . , n0 there is k in such that



.
2
If 1 , . . . , n0 , we calculate (f , f0 ) < , and it follows that -limN f = f0 . Hence we see that id :
B(X , w ) (B(X ), ) is continuous. But the latter is Hausdorff, while the first is compact, and hence = w
on B(X ).
|f (xk ) f0 (xk )| <

10.10 Exercise. If X is a normed space and (B(X ), w ) is metrisable, then X is separable.


Note: Closed Convex Hull Theorem implies that if C is a convex set with C X, normed space, then
C is k k-closed C is w-closed.
10.11 Theorem (w -Separation Theorem). Let X be a normed space. Suppose A, B be convex, disjoint in
X of which B is w -open. Then there exists x X and R such that
Re f (x) < Re g(x)

f A, g B.

Proof. By the Separation Theorem, there is F X and R such that


Re F (f ) < Re F (g)

f A, g B.

We know B is w -open, i.e. B (X , X ). So for any f0 B, then there are x1 , . . . , xn X such that
f0 U =

n
\

xbi 1 (f0 (xi ) + D) B.

i=1

Tn

Let Y = i=1 ker xbi , then xbi (Y + f0 ) = xbi (f0 ) = f0 (xi ) f0 (xi ) + D. So Y + f0 U B. For any f Y ,
Re F (f + f0 ) > since f + f0 B. Hence Re F (f ) > Re F (f0 ), which means F |Y = 0 (why?), so Y ker F .
Together with the next lemma will prove the result; the lemma will conclude that F span{c
x1 , . . . , x
cn }.
Tn
0
10.12 Lemma. In an F -vector space X, if f0 , . . . , fn X such that ker f0 i=1 ker fi , then f0 span{f1 , . . . , fn }.
Proof. Let T : X Fn be given by x 7 (f1 (x), . . . , fn (x)). Then
ker T =

n
\

ker fi .

i=1

Let R = Im T Fn is a vector subspace. Define g0 R0 by g0 (T x) = f0 (x). Notice that if T x = T y, then


T (x y) = 0 so f0 (x y) = 0 i.e. f0 (x) = f0 (y). Then g0 (T x) = f0 (x) = f0 (y) = g0 (T y) so g0 is well-defined.
Also, g0 is linear. g0 : Fn R F. Let g be any extension of g0 to Fn , g (Fn )0 . Then there are 1 , . . . , n F
such that
n
X
g(y1 , . . . , yn ) =
i yi .
i=1

Hence,
f0 =

n
X
i=1

if we restrict our attention to R.


29

i fi

10.13 Theorem (w -Closed Convex Hull Theorem). If S X , then


\

cow S = {H : H is a w -closed half-space containing S}.


Proof. Similar to the Closed Convex Hull Theorem.
c )10.14 Theorem (Goldstines Theorem). Let X be a normed space. Then B(X ) is w -dense (i.e. (X , X

dense) in B(X ).
w

Proof. Let A = B(X ) . Since B(X ) is w -closed (by Alaoglu) A B(X ). Let F0 X \ A, noting that
X \ A is w -open. Then there is w -open basic neighbourhood U of F0 , U X \ A but this means U A = .
Such U can be chosen to be convex i.e.
n
\
1
U=
fbi (F0 (fi ) + D)
i=1

for some f1 , . . . , fn X . Hence, by the w -Separation Theorem, there are f X and R such that
Re F (f ) > Re G(f )

F U, G A.

In particular, F0 U , so Re F0 (f ) > . Since B(X ) is symmetric


Re x
(f ) = Re f (x) 0

for some x B(X )

so 0. Moreover, since Re F0 (f ) > , and f 6= 0 so that > 0. Now, f 0 =


Re F (f 0 ) > 1 Re G(f 0 ),

1
f.

Then

F U, G A.

(1)

for x B(X )

(2)

Hence
Re F0 (f 0 ) > 1 Re x
(f 0 )

now (2) implies k Re f 0 k 1, so kf 0 k 1. Hence (1) implies |F0 (f 0 )| > 1, thus kF0 k > 1. Thus F0
/ B(X ).

Therefore B(X ) = A.
10.15 Definition. A normed space X is called reflexive if X = X .
Since dual spaces are always complete, a reflexive normed space is Banach.
10.16 Theorem. Given a Banach space X , the following are equivalent:
1. X is reflexive.
2. B(X ) is weakly compact.
3. (X , X ) = (X , X ) i.e. weak and weak* topologies concide on X .
4. X is reflexive.
X normed is reflexive if X = X . We saw that any reflexive X is Banach.
Proof. Consider the map f : (B(X ), (X , X )) (B(X ), (X , X )) given by x 7 x
. This is a homeomorphism.

X (X ) = X

the space of evaluation functionals.


(i) (ii): If X is reflexive, the range of the map f is (B(X ), (X , X )) which is compact by the Alaoglu
Theorem. Thus the domain B(X ) is also w-compact.
(ii) (i): If B(X ) is w-compact, then the range of the map f is w -compact in X . By Goldstines Theorem, we
get B(X ) = B(X ). Thus X = X .
(i) (iii): Obvious.
(iii) (iv): We have (B(X ), (X , X )) = (B(X ), (X , X )), and the latter is compact by Alaoglu. Thus (ii)
(i) applied to X yields that X is reflexive.
(iv) (i): Suppose X is reflexive. B(X ) is convex and norm-closed in B(X ), hence by Closed Convex Hull
c so (X , X ) = (X , X
c ). Hence B(X )
Theorem, B(X ) is (X , X )-closed. By our assumption X = X
c

is closed with respect to (X , X ). By Goldstine. B(X ) = B(X ) hence X = X i.e. X is reflexive.


30

10.17 Corollary. Any finite-dimensional normed space is reflexive.


Proof. B(X ) is norm-compact (Heine-Borel) implies B(X ) is w-compact.
10.18 Corollary. If Y is a closed subspace of a reflexive X then Y is also reflexive.
Proof. By Hahn-Banach theorem, then Y = X |Y so (Y, Y ) = (Y, X |Y ) = (X , X )|Y . Then B(Y ) = B(X )Y
is norm-closed and convex, hence weakly closed, by Closed Convex Hull Theorem.
B(Y )
| {z }

w-closed

B(X )
| {z }

= B(Y ) is w-compact

w-compact

Therefore, Y is reflexive.

11

Extreme points and the Krein-Milman theorem

11.1 Definition. Let X be a normed space and C X be a convex set. A face of C is a non-empty convex subset
F C, such that if x, y C, t [0, 1] and (1 t)x + ty F then both x, y F . Furthermore, a face E of C is
called an extreme point if E is a singleton. For example [DIAGRAM].
11.2 Theorem (Krein-Milman Theorem). Let X be a normed space and C X is w -compact and

convex. Then C = cow ext(C).


Last time, we saw that for X a normed space and C X, a face F of C is nonempty, convex, and if (1tx)+ty F
for some x, y C and t [0, 1] then x, y F . This should have read t (0, 1). There exists x F so that for
any y C, t = 0 implies x, y F implies F = C.
If F = {f }, then we called f (or {f }) an extreme point of C.

Proof of Krein-Milman Theorem. C w -closed, convex implies cow ext(C) C. First we will prove that ext(C) 6=
. We take
F = {F C : F is a w -closed face of C}.
C F. Then (F, ) is a partially ordered set. Let C be a chain in F. Define
\
FC =
F
F C

and noting each F is w -closed we have that the intersection above is too. Furthermore, FC is a face: suppose
x, y C, t [0, 1] such that
(1 t)x + ty FC = (1 t)x + ty F for any F C
so indeed x, y F for any F C. So x, y FC . Hence FC F, so it is an upper bound with respect to , for C.
By Zorns lemma, there is a maximal element of F which we call E. E is w -compact, so for x X
Re x
(E) R
| {z }
compact

Hence Re x
(E) has a minimum value, say mx . Let Ex = {f E : Re f (x) = Re x
(f ) = mx } 6= , w -closed. And
Ex is a face (exercise). Hence Ex E but E was minimal so Ex = E. Now if f, g E, then for any x X
Re f (x) = mx = Re g(x) = f = g

Thus E = {f } is a singleton and f ext(C). Finally, we will prove that C cow ext(C). Let f0 X \cow ext(C).
We want to show f0
/ C.

C cow ext C. We proved that ext C 6= . We need to prove C cow ext C. Let f0 X \ cow ext C. Note

that cow ext C is convex, and X \ cow ext C is w -open. By the w -separation theorem, there is x X, R
such that Re f0 (x) < Re f (x) for any f ext C. C is w -compact, so mx = min Re x
(C) exists. Let
Cx = {f C : Re f (x) = mx }
noting that this is w -closed. Also, Cx is a face. By the first part of this proof applied to Cx , we conclude Cx admits
an extreme point, say f . Then f is also an extreme point of C. So Re f (x) = mx . Hence Re f0 (x) < mx thus

f0
/ C. Thus, C = cow ext C.
31

11.3 Corollary. If X is a normed space, and C X is a w-compact convex set, then C = cow ext C.
(X,
X
c )) (X , (X , X
c )) is a homeomorphism onto its range.
Proof. The embedding (X, (X, X )) (X,

C is w-compact, hence C is w -compact. Then by the Krein-Milman theorem,

w
C = cow ext C = co\
ext C.

Hence C = cow ext C.


11.4 Corollary. If X is a normed space, C X is a norm-compact convex set, then C = cokk ext C.
Proof. C is norm-compact and convex, so C is w-compact. Then C = cow ext C = cokk ext C, by the Closed Convex
Hull Theorem.
11.5 Theorem (Minkowskis Inequality). kf + gkp kf kp + kgkp with equality for 1 < p < if and only
if f = g for some 0.
11.6 Proposition. Let X be any of `p , Lp (R) or Lp ([0, 1]). Then ext B(X) = S(X).
Proof. If f D(X), i.e. kf kp < 1, then
(1 kf kp )0 + kf kp

1
f = f = f
/ ext B(X).
kf kp
| {z }
D(X)
/

Thus ext B(X) S(X). Conversely, f S(X) i.e. kf kp = 1, then {f } is a face: let f = (1 t)x + ty for some
x, y B(X) and t [0, 1]. Note
1 = kf kp = k(1 t)x + tykp (1 t) kxkp +t kykp = kxkp = kykp = 1.
| {z } |{z}
1

So, kf kp = (1 t)kxkp + tkykp . By the equality case of Minkowski we know that there is 0 with (1 t)x = ty.
Taking norms, (1 t) kxkp = t kykp . Hence (1 t) = t, so that x = y. Thus f = (1 t)x + tx = x = y. So
|{z}
| {z }
1

x, y {f }, but {f } is a face of B(X). Therefore S(X) = ext B(X).


11.7 Proposition. B(cc0 ) has no extreme points.
c0 ), then limn xn = 0. For some n0 , such that |xn0 | 21 . Let
Proof. If x = (xn )
n=1 B(c
(
(
xn
if n 6= n0
xn if n 6= n0
yn =
zn =
2xn0 if n = n0 .
0
if n = n0 .

c0 ) and 21 (y + z) = x.
Then y = (yn )
n=1 , z = (zn )n=1 B(c

Can c 0 be a dual space? That is, is there a normed space X with X = c 0 ? Assume yes. So X = c 0 . Consider

(cc0 , (cc0 , X)).


B(cc0 ) is convex and w -compact by Alaoglu. By Krein-Milman, B(cc0 ) = cow ext B(cc0 ). Contradiction.
Office hours: 2:30-4pm, today and tomorrow. A3 due Wednesday.
11.8 Example. Consider the unit ball of R2 in the p-norm (2 < p < ).

We now prove that the extreme points of the set P (X) of probability measures on X is just the set of Dirac measures
x ; integration of a function f against such a x merely results in evaluation at x, that is, x (f ) = f (x). It is just
as nontrivial if we choose (X, ) to just be [0, 1] with the usual topology.
11.9 Theorem. For (X, ) a compact Hausdorff space, define
P (X) = { C R (X, ) : kk 1, (1) = 1}.
Then ext P (X) = {x : x X}, where for each x we define x (f ) = f (x).
32

Proof. Since we use symbols f, g for elements of C = C R (X, ) we shall use , , for elements of C . Observe that
1 ({1}) is w -compact (B(C ) is w -compact by Banach-Alaoglu, and 1
1 ({1}) is w -closed).
P (X) = B(C ) 1
Also, P (X) is evidently convex, so it admits extreme points by the Krein-Milman Theorem.
If P (X), we note for 0 f 1 in C, then 0 1 f 1 and hence |(f )| 1 and |1 (f )| = |(1 f )| 1
since kf k , k1 f k 1. Thus 0 (f ) 1. From this it follows that if g 0, then (g) 0.
Now, if g C, let g + = max{g, 0}, g = max{g, 0} so g = g + g and |g| = g + + g . If we fix 0 f 1 in C
and in P (X) we let (g) = (f g), then
0

z }| { z }| {
|(g)| = |(g + g )| = | (f g + ) (f g ) | (f g + ) + (f g ) = (f (g + + g ))
|{z}
|{z}
0

= (f |g|) (f kgk ) = (f )kgk .

(*)

noting that f |g| f kgk . Now, fix ext P (X). Find 0 f 1 in C such that 0 < (f ) < 1 (since kk = 1, this
1
is always possible, exercise). Let (g) = (f
) (f g) for g in C. Then using (*),
|(g)| =

1
1
|(f g)|
(f )kgk = kgk
(f )
(f )

1
1
so kk 1. Also (1) = (f
) (f 1) = 1. Hence P (X). Similarly, if we put (g) = 1(f ) ((1 f )g), then
noting that 1 (f ) = (1 f ), we see that defines an element of P (X). We observe (f ) + (1 (f )) = .
1
In particular = , i.e for g in C, (g) = (f
) (f g) = (g) so (f g) = (f )(g). Since span{f : 0 f 1} = C, it
follows that (f g) = (f )(g) for f, g C. Now, suppose for each x in X Sthere is fx C such that fx (x) 6= 0 but
x Ux , hence X = xX Ux . Since X is compact, there are
fx ker . Let Ux = fx1 (R \ {0})
Sn so Ux is open andP
n
x1 , . . . , xn in X such that X = i=1 Uxi . Thus f := i=1 fx2i > 0 on X so 1/f C. Then

1 = (1) =

( f1 f )

( f1 )(f )

( f1 )

n
X

(fxi )2 = 0.

i=1

This is absurd. Hence there is some x in X such that x (f ) = f (x) = 0 whenever (f ) = 0, ker x ker . However,
by the lemma following w -Separation Theorem, this implies x = c for some c R. However 1 = x (1) = c(1) =
c, so x = . Thus ext P (X) {x }xX .
It remains to show that each x is an extreme point of P (X). If x = (1 t) + t for 0 < t < 1 and , P (X),
then computations similar to but much simpler than (*) show that for f in C,
t|(f )| t(|f |) tx (|f |) = t|f (x)|.
Hence if f ker x , then f ker so ker x ker and, as above, x = . Similarly x = , and x ext P (X).
11.10 Remark. w -closed convex hull of extreme points of a w -closed convex set in the dual space is that set
itself (this is the K-M theorem). Its very tricky to exhibit an example where it is not, a fortiori, the norm-closure.
11.11 Exercise. Prove (only the second equality needs proof) that
(
)

X
X
co ext P (X) = co{x }xX =
i xi : xi X, i 0 for all i and
i = 1 .
i=1

i=1

11.12 Exercise. Let (X, ) = ([0, 1], usual top.). Let m P ([0, 1]) be given by the Riemann integral (i.e. m is
integration against the usual Lebesgue measure, which reduces to Riemann integration because all f are continuous):
Z 1
m(f ) =
f.
0

Then m
/ co{t }t[0,1] (otherwise, we would have covered this in MATH 147).

11.13 Remark. Observe, however, that m cow {t }t[0,1] . We use Riemann sums; let
N = {(s0 , . . . , sn , t1 , . . . , tn ) : 0 = s0 < s1 < . . . < sn = 1, and ti [si1 , si ] for i = 1, . . . , n, and n N}.
Pre-order N by declaring that
= (s0 , . . . , sn , t1 , . . . , tn ) (s00 , . . . , s0n0 , t01 , . . . , t0n0 ) = 0 {s0 , . . . , sn } {s00 , . . . , s0n0 }.
33

If is as above,
S (f ) =

n
n
X
X
(si si1 )f (ti ) =
(si si1 )ti (f ) co{t }t[0,1] .
i=1

i=1

We have (where (s0 , . . . , sn , t1 , . . . , tn ) = N )


m(f ) = lim S (f )
N

i.e.

m = w* -lim
N

n
X
(si si1 )ti .
i=1

11.14 Exercise. In ` = ` R , we have `


= FA(N) (by A1Q4), and
P = { FA(N) : V () 1, (N) = 1}.
Here, ext P = {U : U P(N) is an ultrafilter}, i.e. the extreme points are precisely the ultrafilter limits.
Note that S = T.

12

Euclidean and Hilbert spaces

12.1 Definition. Let X be a vector space. A form [, ] : X X F (F = R or C) is called Hermitian if


(i) [x + x0 , y] = [x, y] + [x0 , y], for x, x0 , y X , F.
(ii) [x, y] = [y, x], for x, y X (symmetry if F = R, skew-symmetry if F = C).
We shall say [, ] is positive if
(iii) [x, x] 0, for x X .
Finally, we shall say [, ] is non-degenerate if
(iv) If 0 6= x X , then [x, y] 6= 0 for some y in X .
We state the following in a little bit more generality than is typical for this, but it often gets used.
12.2 Proposition. Let [, ] be a positive Hermitian form on X , and let p(x) = [x, x]1/2 for x in X . Then for x, y
in X , F, we have
(i) p(x) = ||p(x).
(ii) |[x, y]| p(x)p(y) (Cauchy-Schwarz inequality).
(iii) p(x + y) p(x) + p(y).
Moreover, if [, ] is non-degenerate, then [x, x] > 0 for x 6= 0. Also, in the case of non-degeneracy, assuming y 6= 0,
equality holds at (ii) if and only if x = y for some F, and equality holds at (iii) if and only if x = ty, for some
t 0 in R.
Recall that extreme points of the unit ball in an `p space were all the elements of the unit sphere. We want to have
this fact in storage, in case we want to ask the same question for what well call a Hilbert space. So its not that
interesting, but there are actually some situations where this fact might be useful to know.
Pattersons book shows the following version of the proof (which I believe is a corrupted version of something that
appeared in Terry Taos blog).
Proof. We have:
(i) p(x) = [x, x]1/2 = ([x, x])1/2 = (||2 [x, x])1/2 = ||p(x).
(ii) First, let = sgn[x, y] and note that
[x,y]

z }| {
0 p(x y) = [x y, x y] = [x, x] [y, x] [x, y] + ||[y, y] = p(x)2 2 Re [x, y] + p(y)2
| {z }
2

|[x,y]|

34

and hence |[x, y]|

p(x)2 +p(y)2
.
2

We observe for t > 0,

t2 p(x)2 + t12 p(y)2


.
2
If p(x) = 0, take t ; if p(y) = 0, take t 0+ , to see that |[x, y]| = 0 in either of these cases. If [x, y] 6= 0,
then p(x)p(y) 6= 0, and we substitute t = p(y)/p(x), above to get (ii).
|[x, y]| = |[tx, 1t y]|

(iii) Note
p(x + y)2 = p(x)2 + 2 Re[x, y] + p(y)2
2

p(x) + 2|[x, y]| + p(y)


= (p(x) + p(y))

(as above)
(**)

which gives (iii).


It is immediate from (the proof) of (iii), that if [, ] is non-degenerate then [x, x] > 0 for x 6= 0. Fix y 6= 0. We
observe that if x 6= y for any in F, then = is impossible to obtain at (*). Hence for = to be achieved in (ii),
we require x = y. Likewise, = can achieved at (**) only if x = ty for some t 0. Hence, this is required for =
at (iii).
12.3 Corollary. Let (, ) be a non-degenerate positive Hermitian form on a vector space X , and kxk = (x, x)1/2 .
Then (X , k k) is a normed space.
Proof. Immediate from above.
There are many features of `p spaces that these spaces will share, but not all of them. This is perhaps a very
extreme kind of something called rotundity.
12.4 Definition. We call a non-degenerate positive Hermitian form an inner product and the pair (X , (, )) an
inner product space or Euclidean space. If X is complete with respect to the norm k k induced by the inner
product, we call (X , (, )) a Hilbert space.
12.5 Proposition (Polarisation Identities). Let [, ] be a Hermitian form on X .
In the R case: 4[x, y] = [x + y, x + y] [x y, x y].
P3
In the C case: 4[x, y] = k=0 ik [x + ik y, x + ik y].
Proof. For the real case, just bash it out. For the complex case, observe

P3

k=0

ik = 0,

P3

k=0

i2k = 0.

12.6 Definition. Suppose (, ) is an inner product with associated norm k k. We call a set S X orthogonal if
x 6= y S we have (x, y) = 0; write x y (read as is perpendicular to).
n

n
X 2 X


xi =
kxi k2 .
12.7 Proposition (Pythagoras Law). If {x1 , . . . , xn } X is orthogonal, then


i=1

i=1

12.8 Proposition (Parallelogram Law). If x, y X , then kx + yk + kx yk = 2kxk + 2kyk2 .


12.9 Exercise (tedious exercise). If (X , k k) is a normed space which satisfies parallelogram law, then the
form (, ) given by
3
kx + yk2 kx yk2
1X 4
(x, y) :=
,
(x, y) :=
i kx + ik yk2
4
4
k=0

can be shown to be an inner product which gives the norm.


12.10 Example. We have:
R1
(i) C[0, 1]. (f, g) = 0 f g (Riemann integral). This gives rise to an (incomplete) Euclidean space.
R1
(ii) L2 [0, 1]. (f, g) = 0 f g (Lebesgue integral). Hilbert space (PM 450/354).
P
P
2
N
(iii) `2 = `2 (N) = {(xn )
n=1 F :
n=1 xn yn which converges (absolutely) by
n=1 |xn | < }. We let (x, y) =
Hlders inequality (exercise). We observe kxk2 = (x, x)1/2 , and `2
= `2 , so is complete.
P
(iv) be a set (possibly
uncountable), and define `2 () = {x = (x ) F : |x |2 < }. We define the
P
meaning of as follows: if (a ) , a 0, we can write
X
X
X
a = sup
a = lim
a

F F

35

F F

where F = {F : F is finite} directed by F F 0 iff F F 0 . We note if


with a 0, then for Fn = { : a n1 }, we have

a < for a = (a )

Fn = { : a 6= 0}.

n=1

If

a < , each Fn is necessarily finite, so a = { : a > 0} is countable. If x, y `2 () we write


(x, y) =

x y =

x y

x y

which is a countable series. If (x(n) )


n=1 `2 () is k k2 -Cauchy, then
!

[
(x(n) )
x(n)
n=1 `2
n=1

and the countable union of countable sets is countable, from which it follows that `2 () is complete.
12.11 Proposition. Let (X , (, )) be an Euclidean space. For each x in X , the functional fx : X F given by
fx (y) = (y, x) is in X and kfx k = kxk.
Proof. fx X 0 by the properties of (, ). Also kfx k kxk is a consequence of the Cauchy-Schwarz inequality.
1
1
1
x) = ( kxk
x, x) = kxk
(x, x) = kxk, so kfx k kxk.
Observe if x = 0 then fx = 0; if x 6= 0, then fx ( kxk
12.12 Remark (Notation). Let X be an Euclidean space and 6= S X . We define
S = {y X : (y, x) = 0 for all x in S},
T
pronounced S-perp. Observe that S = xS ker fx , and each of these kernels is a closed subspace, so this is
itself a closed subspace of X .
12.13 Theorem (Complementation Theorem). Let X be an Euclidean space and Y X be a complete
subspace (e.g. X is Hilbert, Y closed; or in general Y is finite dimensional). Then for x in X then there is a unique
decomposition x = xY + xY such that xY Y, xY Y . Moreover, the map P : X X given by P x = xY is
linear, and Im P = Y, P 2 = P and if Y 6= {0}, kP k = 1.
Proof. First, let us find xY . Let d = dist(x, Y) = inf{y Y : kxyk}. Let for each n, yn Y be so kxyn k < d+ n1 .
The parallelogram law for n, m N gives
k(x yn ) (x ym )k2 + k(x yn ) + (x ym )k2 = 2kx yn k2 + 2kx ym k2
so
kym yn k2 = 2kx yn k2 + 2kx ym k2 k2x (yn + ym )k2
= 2kx yn k2 + 2kx ym k2 4kx 21 (yn + ym ) k2
|
{z
}
Y
{z
}
|
< 2(d +

1 2
n)

+ 2(d +

1 2
m)

4d2
2 n,m
+

4d 0 .

So (yn )
n=1 is Cauchy in Y. By completeness, xY = limn yn . Now, let xY = x xY . Observe that kxY k = d.
Suppose there were y in Y such that (xY , y) 6= 0. We let z = (xY , y)y, so z Y, (xY , z) = |(xY , y)|2 > 0.
Then for any  > 0 we have
kx (xY + z )k2 = kxY zk2 = (xY z, xY z) = kxY k2 2 Re(xY , z) + 2 kzk2
| {z }
Y

= d2 [2(xY , z) kzk2 ] < d2 if 0 <  <

2(xY , z)
kzk2

which contradicts the definition of d = dist(x, Y). Hence xY Y . We have x = xY + xY , xY Y, xY Y .


Since Y Y = {y Y : (y, x) = 0x Y} {y Y : (y, y) = 0} = {0}, so Y Y = {0}. This shows that the
orthogonal decomposition is unique i.e. if x = x0Y + x0Y = xY + xY then xY x0Y = x0Y xY which forces both
pairs to coincide.
36

We first note, by Pythagoras law, kxk2 = kxyk2 + kxY k2 so kxY k kxk. If x, x0 are in Y and F, then
(x + x0 )Y + (x + x0 )Y = xY + xY + (x0Y + x0Y )
= xY + x0Y Y + xY + x0Y
| {z }
|
{z
}
Y

so by uniqueness, we see (x + x0 )Y = xY + x0Y . Thus P : X X , P x = xY is linear. Also kP k 1. If y Y


then P y = y, so kP k 1.
12.14 Remark. In the situation above, we call P = PY the orthogonal projection onto Y.
We also note that X = Y 2 Y .
12.15 Proposition. If H is a Hilbert space and 6= S H then S = (span S) and (S ) = span S.
Proof. We have
x S S

{x}
|{z}

span S {x} .

closed linear space

Now for any closed, hence complete, subspace Y we have Y 2 Y = H = (Y ) 2 Y . We have that Y (Y ) ,
and hence Y = (Y ) . We apply this to Y = span S.
12.16 Theorem (Riesz Representation Theorem). Let H be a Hilbert space, and f H . Then there
is a unique vector x0 H such that f = fx0 , i.e. f (x) = (x, x0 ) for x H, kfx0 k = kx0 k.
Proof. Say f 6= 0. We first note that (ker f ) is one-dimensional. Indeed, if x1 , x2 (ker f ) \ {0} then f (x1 ) 6=
0 6= f (x2 ), and
f (x2 )x1 f (x1 )x2 ker f (ker f ) = {0},
so {x1 , x2 } cannot be linearly independent. Hence (ker f ) = span{x1 }, kx1 k = 1. Let x0 = f (x1 )x1 . Let P = Pker f
be the orthogonal projection onto ker f . Now, if x H,
x = P x + x1 for some F.
|{z}
(IP )x

Thus
(x, x0 ) = (

Px
|{z}

ker f ={x1

+x1 , f (x1 )x1 ) = (x1 , f (x1 )x1 ) = f (x1 ) (x1 , x1 ) = f (x1 ) = f ( |{z}
P x +x1 ) = f (x).
| {z }
1

ker f

So we see indeed that f = fx0 . If (x, x0 ) = (x, y0 ) for all x in H, then with x = x0 y0 , we see (x0 y0 , x0 y0 ) = 0
so x0 = y0 .
Final exam date is Dec. 12, 12:30-3pm (see website). Possible talk topics posted soon.
If F = C, x 7 fx is conjugate linear, i.e. fx+x0 = fx + fx0 (passing scalars at cost of conjugation). Generally, we
where H
denotes the conjugate space. We have H = H
as sets, but scalar multiplication in H
is
write H
=H
governed by the rule x = x where the right-hand side is scalar multiplication in H, and (x, y)H = (y, x)H .
12.17 Remark (on the completion of an Euclidean space). We have:
kk

(i) Let (X , (, )) be an Euclidean space. Let H = X X . We observe that H is a Hilbert space. If F, G H

let (xn )
n , G = limn yn . We observe for n, m in N
n=1 , (yn )n=1 X be so F = limn x
|(xn , yn ) (xm , ym )| |(xn , yn ) (xn , ym )| + |(xn , ym ) (xm , ym )| kxn kkyn ym k + kxn xm kkym k
just using Cauchy-Schwarz and biadditivity of the form giving the inner product. So it is easy to see that
0
((xn , yn ))
n=1 F is Cauchy, and hence converges to something, call it (F, G). We observe if (xn )n=1 ,
0
0
0
(yn )n=1 X , F = limn x
n , G = limn yn , then the same technique as above shows that limn (xn , yn ) =
limn (x0n , yn0 ). Hence (F, G) is well-defined and may be shown to define an inner product on H for which
kF k = (F, F )1/2 .
(ii) Now, if f X , consider f X , so f|H H , so f|H = fG0 . Hence for x in X , f (x) = (
x, G0 ). It follows

H
=
H.
that H = X , i.e. X
H
so
X
H
=
=
=
37

12.18 Theorem (Gram-Schmidt Orthogonalisation). Let {x1 , x2 , x3 , . . .} be a linearly independent sequence in an Euclidean space X . Then there exists an orthogonal sequence {e1 , e2 , . . .} such that span{e1 , . . . , en } =
span{x1 , . . . , xn }.
Proof. Let En = span{x1 , . . . , xn } and Pn denote the orthogonal projection from X to En (complementation theorem). We define
e1 = x1
e 2 = x 2 P1 x 2
e3 = x3 P2 x3
..
.
en = xn Pn1 xn .
We observe, inductively, then en En . Hence span{x1 , . . . , xn } = span{e1 , . . . , en } holds. Moreover, each en
En1 , and it follows that {e1 , . . . , en } is orthogonal, hence {e1 , e2 , . . .} is orthogonal.
12.19 Theorem. We have:
(i) If X is a separable Euclidean space, then X contains an orthogonal sequence, E = {e1 , e2 , . . .} such that
span E = X .
(ii) If H is any Hilbert space, there is an orthogonal set E H such that span E = H.
Proof. We have:

(i) If {xn }
n=1 is a dense subset of X , then span{xn }n=1 = X . Now let

n1 = min{k N : xk 6= 0}
n2 = min{k N : xk
/ span{xn1 }}
..
.
nm = min{k N : xk
/ span{xn1 , . . . , xnm1 }}

we have that span{xnm }


m=1 = span{xn }n=1 = X . We apply Gram-Schmidt to (xkm )m=1 to obtain orthogonal

sequence (em )m=1 . Observing that span{xn1 , . . . , xnm } = span{e1 , . . . , em }, we see that span{em }
m=1 =
span{xnm }
.
m=1

(ii) Let O = {S SH : O
/ S, S is orthogonal}. Partially order O by inclusion. If C O is a chain, it is easy
to verify that SC S is an orthgonal set, not including O. Hence by Zorns Lemma, there is a maximal
orthogonal set E. If we had span E ( H, we let PE denote the orthogonal projection onto span E. Then if
x H \ span E, we would see that 0 6= (I PE )x E, contradicting maximality.
12.20 Definition. A set E in an Euclidean space is orthonormal if for e, e0 in E,
(
1 if e = e0
0
(e, e ) =
0 if e 6= e0 .
12.21 Remark. If {e1 , . . . , en } is an orthonormal set, let E = span{e1 , . . . , en }. If x PX , we have orthogonal
n
decomposition x = PE x + (I PE )x, where (I PE )x E. Moreover PE x E, so PE x = i=1 i ei . Then for each
i,

n
X
(x, ei ) =
i ei + (I PE x), ei
| {z }
i=1

to E

and each ej is orthogonal to ei unless j = i. So we obtain (x, ei ) = (i ei , ei ) = i . Hence PE x =

Pn

i=1 (x, ei )ei .

12.22 Theorem (Abstract Riesz-Fischer Theorem). We have:


(i) If X is an Euclidean space and E X is an orthonormal set, then for x in X
X
|(x, e)|2 kxk2
(Bessels inequality).
eE

(ii) Let dim X P


= . X is complete if and only if for any orthonormal sequence (en )
n=1 in X , and any `2 we

have that i=1 i ei converges in X .


38

Proof. We have:
(i) If F E is finite then
X

|(x, e)|2 = kPF xk2 kxk2

eF

where PF is the orthogonal projection onto span F . But


(
)
X
X
2
2
|(x, e)| = sup
|(x, e)| : F E finite kxk2 .
eE

(ii) () We let xn =

Pn

i=1

eF

i ei . Then for m < n we have



2
n
n
X

X
m,n


kxn xm k =
i ei =
|i |2 0.

P yth.
2

i=m+1

i=m+1

So (xn )
n=1 X is Cauchy, hence converges.
(n)
() Let (x(n) )
}n=1 is separable. Hence, by the
n=1 be a Cauchy sequence in X . Then E = span{x
last theorem, there is an orthogonal (hence we can make it orthonormal) sequence {en }
n=1 such that E =
P
(n)
(n)
2
(n) 2
(n)
span{en }
.
From
(i)
above,
we
have
that
|(x
,
e
)|

kx
k
so
if

=
(x
, ei ), then (n) =
i
n=1
i
i=1
(n)
(i )
i=1 `2 . We observe for n, m in N that

k(n) (m) k22 =

|(x(n) , ei ) (x(m) , ei )|2 kx(n) x(m) k2

i=1
(n)
so (P
)n=1 `2 is Cauchy, and hence converges to some in `2 , as `2
= `2 and is complete. Thus

(n)
x = i=1 i ei X , by assumption. Given  > 0, find n0 such that k k2 < 2 for n n0 . Then, for
P
P
(n)
fixed n n0 , find N such that i=N +1 |i |2 < (/3)2 , i=N +1 |i |2 < (/3)2 . Then

N



X
X
X







(n)
(n)
kx x(n) k (i i )ei +
i ei +
i ei




i=1
i=N +1
i=N +1
m

m

X

X







(n)
(n)
i ei + + = .
k k2 + lim
i ei + lim
m
m
3 3 3

i=N +1
i=N +1
{z
}
|
P
(

m
i=N +1

|i |2 )1/2

12.23 Theorem (Orthonormal Basis Theorem). Let H be a Hilbert space and E H be an orthonormal
set, with E = span E and PE = PE be the orthogonal projection on E. Then for x, y in H we have:
X
X
(i)
(x, e)e = lim
(x, e)e = PE x, where F = {F E : F finite} directed by F F 0 iff F F 0 .
F F

eE

(ii)

eF

|(x, e)| = kPE xk2 (Bessels identity).

eE

(iii) (PE x, PE y) = (PE x, y) = (x, PE y) =

(x, e)(e, y) (Parsevals identity).

eE

Moreover, if for every x, y in H we have any of


X
(i)
(x, e) = x.
eE

(ii)

|(x, e)|2 = kxk2 .

eE

(iii) (x, y) =

(x, e)(e, y).

eE

Then span E = H.

39

Proof. We have:
P
(i) It follows from Bessels inequality (from the theorem above) that PeE |(x, e)|2 Pkxk2 , so Ex = {e E :
(x, e) 6= 0} is countable, hence a sequence, so by Riesz-Fischer x0 = eE (x, e)e = eEx (x, e)e converges in
H. Also, x0 E. Now if e0 E, then
!
X
X
0
0
0
(x, e)(e, e0 )
(x x , e) = (x, e )
(x, e)e, e = (x, e0 )
eEx

eEx

by the continuity of y 7 (y, e0 ), but this is just equal to (x, e0 ) (x, e0 ) = 0. It follows that x x0 E, hence
x x0 E. Also
(x x0 , x0 ) = (x, x0 ) (x0 , x0 )
!
X
= x,
(x, e)e

!
X

eE

(x, e)e,

eE

(x, e )e

e0 E

!
=

(x, e)(x, e)

eE

(x, e) e,

eE

X X
eE

(x, e )e

()

e0 E

eE

|(x, e)|2

(x, e)(x, e0 )(e, e0 ) = 0

e0 E

so x = x0 + x x0 is the unique orthogonal decomposition with x0 E. Hence PE x = x0 .


P
(ii) Let x0 be as above. For finite F E, we let x0 = x0F + x0 x0F where x0F = eF (x0 , e)e is the closest point
in span F to x0 , so this is an orthogonal sum. By Pythagoras,
X
kx0 k2 = kx0F k2 + kx0 x0F k2 =
| (x0 , e) |2 + dist(x0 , span F )2 .
| {z }
eF

If we take limit for F in F we get


X
lim kx0F k2 =
|(x, e)|2 ,
F F

eE

(x,e)

lim dist(x0 , span F ) = dist(x0 , span E) = 0

F F

where by (i), limF F x0F = x0 = PE x.


(iii) Using computations similar to () we compute any of (PE x, PE y), (PE x, y) or (x, PE y) to obtain

eE (x, e)(e, y).

Now, if (i) holds, then H span E H. If (ii) holds then x = PE x + (I PE )x is an orthogonal decomposition
by which kxk2 = kPE xk2 so k(I PE )xk2 = 0, thus PE = I which gives (i). If (iii) holds, let x = y and obtain
(ii).
12.24 Corollary. If H is a Hilbert space and E H is an orthonormal set for which span E = H, then the map
U : H `2 (E) given by
U x = ((x, e))eE
defines a surjective linear isometry such that (U x, U y) = (x, y).
Proof. That (U x, U y)`2 = (x, y)H is Parsevals identity and hence kU xk2 = (U x, U x)`2 = (x, x)H = kxk2 , so U is
an isometry. Surjectivity of U is a consequence of the Riesz-Fischer Theorem.
Talk topics have been posted.

13

Spectral theory

Let X be a complex Banach space and consider B(X ) = B(X , X ). Compose!


13.1 Definition. Let T B(X ).
The spectrum of T is (T ) = { C : I T does not admit a (two-sided) inverse in B(X )}.
The point spectrum of T is p (T ) = { C : ker(I T ) ) {0}}.
40

Of course, p (T ) (T ) and equality holds if dim X < .


13.2 Example. We have the following examples:
(i) For dim X < , and T B(X )
= Mdim X (C), we have
(T ) = { C : det(I T ) = 0}.
(ii) Let S B(`p ) (1 p ) be given by S(x1 , x2 , . . .) = (0, x1 , x2 , . . .), the (unilateral) shift operator.
Note kSxkp = kxkp for all x `p so ker S = {0}, so 0
/ p (S). We claim p (S) = . Let C \ {0}. If
x ker(I S), then
x Sx = 0 = x = (x1 , x2 , . . .) = Sx = (0, x1 , x2 , . . .)
so if x 6= 0, let k = min{n N : xn 6= 0}. Then xk = (Sx)k = xk1 , so xk1 6= 0, contradicting minimality
of k. So ker(I S) = {0}.
Question: Is (T ) always non-empty?
13.3 Theorem (Inversion Theorem). Let X be a Banach space.
(i) If T B(X ) with kT k < 1, then

T k (the Neumann series) converges in B(X ) (note T 0 = I), and

k=0

T k = (I T )1 .

k=0

(ii) If T B(X ) and S G(X ) = {U B(X ) : U 1 B(X )} and kT Sk <


G(X ) is open in B(X ) and S 7 S 1 is continuous.

1
kS 1 k ,

then T G(X ). Moreover,

Proof. Note kST k kSkkT k.


Pn
(i) Let Sn = k=0 T k , and note that for m < n,

n
n

X
X
kT km+1 m


kT kk <
T k
0.
kSn Sm k =


1 kT k
k=m+1

k=m+1

So, (Sn )
n=1 is Cauchy, hence convergent. Now,
(I T )Sn =

n
X

(T k T k+1 ) = I T n+1 I.

k=0

Similarly, Sn (I T ) I. So

k=0

T k = (I T )1 .

(ii) Observe kS 1
T )k kS 1 kkS T k < 1 so T = (S (S PT )) = S(I S 1 (S T )) G(X ) from (i),
P(S

1
with T = k=0 (S 1 (S T ))k S 1 . Therefore, T 1 S 1 = k=1 (S 1 (S T ))k S 1 , so
kT 1 S 1 k

X
k=1

kS 1 kkS T k
k
(kS 1 kkS T k) kS 1 k =
kS 1 k,
{z
}
|
1 kS 1 kkS T k
<1

so limT S kT 1 S 1 k = 0, so S 7 S 1 is continuous. Moreover, for S G(X ), S +


so G(X ) is open.

1
kS 1 k D(B(X ))

13.4 Definition. If X is a complex Banach space and T B(X ), the resolvent set of T is given by
(T ) = { C : (I T )1 B(X )} = C \ (T ).
If B is a complex Banach space and U C is open we call F : U B holomorphic if for each z0 in U ,
F 0 (z0 ) = lim

zz0

F (z) F (z0 )
z z0

exists.
41

G(X )

13.5 Remark. A holomorphic function is continuous on its domain.


13.6 Proposition. Let X be a complex Banach space, T B(X ).
(i) (T ) is open in C.
(ii) The resolvent function, R : (T ) B(X ) given by R(z) = (zI T )1 , is holomorphic.
(iii) (T ) {z C : |z| kT k} and kR(z)k

1
for |z| > kT k.
|z| kT k

Proof. We have:
(i) Let F : C B(X ), F (z) = zI T . Then F is continuous, and (T ) = F 1 (G(X )) is open.
(ii) If z, z0 (T ),
R(z) R(z0 ) = (zI T )1 (z0 I T )1 = (zI T )1 [(z0 I T ) (zI T )](z0 I T )1
= (z0 z)(zI T )1 (z0 I T )1
so that

R(z) R(z0 )
zz0
= (zI T )1 (z0 I T )1
((z0 I T )1 )2 .
z z0

(iii) If |z| > kT k, then k z1 T k < 1 so zI T = z(I z1 T ) is invertible. So (T ) {z C : |z| > kT k}. Also,
1

R(z) = (zI T )
so that

1
=
z

1


1X 1 k
1
=
T
I T
z
z
zk
k=0

1 X 1
1
kR(z)k
kT kk =
k
|z|
|z|
|z|
k=0

1
1

1
|z| kT k

1
.
|z| kT k

13.7 Corollary. (T ) is compact.


Proof. (T ) = C \ (T ), and (T ) is open, so (T ) is closed. By the above, (T ) is bounded. Heine-Borel.
Assign 4, hand-in Friday. Assign 5 will still be posted by Wed.
Office hours this week, T & Th. 2:30-4pm.
13.8 Theorem (Liouvilles Theorem). Let f : C C be holomorphic. If f is bounded, then f is constant.
Proof. Complex analysis (PMATH 352, say).
When you have a finite-dimensional operator, you can come up with some sort of basis and realize it as the matrix.
We know that the spectrum is exactly the set of eigenvalues of said matrix. Most proofs of the Fundamental
Theorem of Algebra are consequences of Liouvilles Theorem in complex analysis.
13.9 Theorem (Liouvilles Theorem for Banach space-valued functions). Let X be a C-Banach
space, and let F : C X be a holomorphic function. If F is bounded, then F is constant.
Proof. Let f X . Then f F : C C is holomorphic, i.e.
f F (z) f F (z0 )
f (F (z) F (z0 ))
lim
= lim
=f
zz0
zz
z z0
z z0
0

F (z) F (z0 )
lim
zz0
z z0

exists for all z0 C, and is bounded if F is bounded:


sup |f F (z)| sup kf kkF (z)k.
zC

zC

Applying Liouvilles Theorem to f F we see for z, z 0 C that f (F (z)) = f (F (z 0 )). Now, applying this, for any
f X and appealing to the Hahn-Banach Theorem (that theres a sufficient quantity of linear functionals to
separate all points), then we see that F (z) = F (z 0 ) for all z, z 0 C.
13.10 Theorem. Let X be a C-Banach space and T B(X ). Then (T ) 6= .

42

Proof. Let R(z) = (zI T )1 denote the resolvent function, for z (T ). Suppose that (T ) = C. We observe
that
1
lim kR(z)k lim
= 0,
|z|
|z| |z| kT k
and that R, being holomorphic, is continuous, i.e. bounded on 2kT kB C. [diagram]. Hence R is bounded on all
of C, and is thus constant. Due to the limit above, this means that R(z) = 0 for all z. This contradicts that each
R(z) is invertible.
13.11 Theorem (Spectral Mapping Theorem). Let X be a C-Banach space. If T B(X ), and p(t) C[t]
(polynomials with complex coefficients) then
(p(T )) = p((T )) = {p() : (T )}.
Proof. We suppose p(t) 6= 0. If z0 C, write
n
Y

p(t) z0 =

(t zk )

k=1

where z1 , . . . , zn is the family of zeros of p(t) z0 (counting multiplicity) [Fundamental Theorem of Algebra]. Then
n
Y

p(T ) z0 I =

(T zk I).

k=1

Thus p(T ) z0 I is non-invertible, if and only if at least one T zk I is non-invertible (i.e. zk (T ) for some k).
Thus z0 (p(T )) if and only if p() z0 = 0 for some (T ).
13.12 Theorem (Spectral Radius Formula). Let X be a C-Banach space. Let T B(X ) and define the
spectral radius of T by r(T ) = max{|| : (T )}. Then
r(T ) = lim kT n k1/n .
n

Proof. Let us first see that limn kT n k1/n exists. Let (n) = log kT n k for n N. Observe that
kT n+m k kT n kkT m k = (n + m) (n) + (m).
Fix, for the moment, m and let n > m, so n = qn m + rn where qn = 0, 1, 2, . . ., rn = 0, 1, . . . , m 1 and then
(n)
(qn m + rn )
(qn m) + (rn )
qn (m) + (rn )
(m) (rn )
=

+
n
n
n
n
m
n
so lim sup
n

(m)
(n)

, hence
n
m
lim sup
n

so

limn (n)
n

(n)
(m)
(n)
inf
lim inf
n
mN m
n
n

exists. Exponentiation shows limn kT n k1/n exists.

Now, by the spectral mapping theorem, we have that r(T n ) = r(T )n . Also, we saw earlier that r(T n ) kT n k.
Thus
r(T ) kT n k1/n = r(T ) lim kT n k1/n .
n

We let f B(X ) and consider f R : (T ) C. Then f R has a Laurent series for all |z| > r(T ). We write
f R(z) =

X
fk
k=0

2k

(note we have omitted the polynomial part but we will prove we didnt need that anyway). We have for |z| > kT k,
R(z) =

X
k=1

43

1
T k,
2k+1

and hence we can compute fk+1 = f (T k ), f0 = 0. Hence, by Banach-Steinhaus, we see that

X
k=0

1
Tk
z k+1

converges for all |z| > r(T ). Now, if |z| < limn kT n k1/n , we can find  > 0 such that
(1 + )|z| < lim kT n k1/n
n

and there is n0 such that (1 + )|z| < kT n k1/n for n n0 . But then (1 + )n <

1
n
|z|n kT k

but, this means

X
1 k
T
zk

k=0

cannot converge, since the nth grows without bound. Thus we have that |z| > r(T ) implies |z| > limn kT n k1/n .
Hence r(T ) limn kT n k1/n .
13.13 Example. Suppose dim X < , where X is a C-vector space. Then L(X ) = B(X )
= Mdim X (C) for any
norm on X . Here, if T B(X ), (T ) = p (T ) = { C : det(I T ) = 0}. Observe r(T ) = max{|| : (T )}
depends only on algebraic structure of T , whereas kT k depends on our choice of norm on X . But, by SRF,
r(T ) = limn kT n k1/n depends only on algebraic structure.

14

Adjoint operators

14.1 Definition. Let X , Y be vector spaces and T L(X , Y). Define T L(Y 0 , X 0 ) by (T (f ))(x) = f (T x), so
T f = f T = f T . Linearity is clear. This is the adjoint of T .
14.2 Proposition (on the adjoint). Let X , Y, Z be normed spaces, and T, S B(X , Y), R B(Y, Z).
(i) T B(Y , X ), kT k = kT k.
(ii) If F, (S + T ) = S + T .
(iii) If T = (T ) B(X , Y ), then T x
= Tcx for all x X .
(iv) RS B(X , Z) implies (RS) = S R B(Z , X ).
Proof. We have:
(i) and (iii) If f Y ,
kT f k = sup{|T f (x)| : x B(X )} = sup{|f (T x)| : x B(X )} kf kkT k
so kT k kT k. If x X , f Y ,
T x
(f ) = x
(T f ) = T f (x) = f (T x) = Tcx(f ).
So kT k = kT |X k kT k kT k kT k. So kT k = kT k.
(ii) Obvious.
(iv) If f Z , (RS) f = f R S = S (f R) = S R f .
14.3 Definition. Let X be a normed space, Y X a subspace, and Z X a subspace.
\
The annihilator of Y in X is Y a = {f X : f|Y = 0} =
ker y (w -closed subspace of X ).
| {z }
yY w -closed

The pre-annihilator of Z on X is Za = {x X : x
|Z = 0} =

ker f (closed subspace of X ).

f Z

14.4 Remark. We have:


a

(i) Y a = Y and (Y a )a = Y (A2Q5a).


(ii) Za = (Z

)a and (Za )a = Z

(prove as in A2Q5a, using w -closed convex hull theorem).


44

14.5 Example. Recall `1


= ` , so for x `1 , y ` , let
hx, yi = fy (x) =

xi yi

i=1

denote the dual pairing. Now, c 0 ` and with respect


to the duality above, (cc0 )a = {0} `1 since c 0
= `1 .
w
a
a
Hence, ((cc0 )a ) = {0} = ` which captures that c 0 = ` .
14.6 Theorem (Kernel-Annihilator Theorem). Let X , Y be normed spaces, and T B(X , Y). Then
ker T = Im(T )a and ker(T ) = (Im T )a .
Proof. We have
ker T = {x X : T x = 0} = {x X : T g(x) = g(T x) = 0 g Y } = Im(T )a .
ker(T ) = {g Y : T g = 0} = {g Y : g(T x) = T g(x) = 0 x X } = (Im T )a .
14.7 Corollary. We have:
(i) ker(T ) = {0} if and only if Im T = Y.
(ii) ker T = {0} if and only if Im T

= X .

Proof. Apply preceding remark to Kernel-Annihilator.


14.8 Theorem. Let X , Y be Banach spaces, T B(X , Y). The following are equivalent:
(i) T is invertible; i.e. there is an inverse T 1 B(Y, X ).
(ii) T is invertible.
(iii) Im T is dense in Y and T is bounded below: inf{kT xk : x S(X )} > 0.
(iv) Both T and T are bounded below.
14.9 Remark. If T is bounded below, m = inf{kT xk : x S(X )}, we have kT xk mkxk for all x X .
Proof. (i) (ii): T T 1 = IX = T 1 T , so by an earlier proposition, (T 1 ) T = IX = T (T 1 ) so (T 1 ) =
(T )1 .
(ii) (iii): By the last corollary, (Im T )a = ker(T ) = {0}, so Im T = Y. Also, if x X \ {0}, choose f X ,
kf k = 1, f (x) = kxk, by Hahn-Banach. Then
kxk = f (x) = T (T )1 f (x) = (T )1 f (T x) k(T )1 f kkT xk k(T )1 kkT xk
so

1
k(T )1 k kxk

kT xk, so T is bounded below.

(iii) (i): If T is bounded below, then ker T = {0}. We also have that if T is bounded below, then Im T is closed.
Let (yn )
n=1 Im T with y = limn yn in Y. Write yn = T xn for some xn in X , and we observe
kxm xk k

1
kT (xm xk )k = kym yk k
m

and hence (xn )


n=1 is Cauchy, hence x = limn xn exists in X . We have y = limn yn = limn T xn = T x.
We note Im T = Y by assumptions. Hence T 1 exists. Finally, for y Y, kyk = kT T 1 yk mkT 1 yk so
1
1
kT 1 yk m
kyk, i.e. kT 1 k m
.
(i), (ii) (iv): (i) (iii) hence T is bounded below. (ii) (iii) (stated for T ), so T is bounded below.
(iv) (iii): T bounded below implies T is bounded below. Also, T bounded below implies ker(T ) = {0} implies
Im T = Y by Corollary to Kernel-Annihilator Theorem.
14.10 Remark. Reasons why T B(X ) may not be invertible:
ker T ) {0}.
Im T , Im T ( X .
T fails to be bounded below.
14.11 Corollary. If X is a C-Banach space, T B(X ), then (T ) = (T ).
45

Proof. Given in C, we have by (i) (ii) above that


IX T is invertible IX T is invertible.
14.12 Remark. We have:
Point spectrum: p (T ) = { C : ker(I T ) ) {0}}.
Compression spectrum: com (T ) = { C : Im(I T ) ( X }.
Approximate point spectrum: ap (T ) = { C : I T is not bounded below}.
If T B(X ) is not bounded below, then there is xn S(X ) such that kT xn k <

1
n,

i.e. limn T xn = 0. Hence


n

ap (T ) = { C : there is (xn )
0}.
n=1 S(X ) s.t. (I T )xn
{z
}
|
n
0
T xn xn
14.13 Example. Fix 1 p < , S B(`p ), S(x1 , x2 , . . .) = (0, x1 , x2 , . . .). We
showed earlier that p (S) = .
P
Lets compute S B(`q ), p1 + 1q = 1. Write for x `p , y `q , hx, yi = fy (x) = i=1 xi yi . We have
hx, S yi = fy (Sx) = hSx, yi =

xi1 yi =

i=2

xi yi+1

i=1

and we find that S (y1 , y2 , . . .) = (y2 , y3 , y4 , . . .) (back-shift).


Observe that if || < 1 then x() = (1, , 2 , 3 , . . .) `p and S x() = (, 2 , 3 , . . .) = (1, , 2 , . . .) = x() .
Hence p (S ) D. Moreover, kS k = kSk = 1, so (S ) B. Conclusion is that (S ) = B, hence (S) = B.
Check that D com (S) (use Corollary to K-A Theorem).
1
14.14 Lemma. If (Tn )
k < ,
n=1 G(X ) (X a Banach space) and T = limn Tn exists and supnN k(Tn )
then T G(X ).

Proof. Let M = supnN k(Tn )1 k. For sufficiently large n,


kTn T k <

1
1

,
M
k(Tn )1 k

which implies T G(X ) by the Inversion Theorem.


14.15 Proposition. If T B(X ) (X a C-Banach space), and if (T ) denotes the (topological) boundary of (T ),
then (T ) ap (T ).
Proof. If (T ), then there is (n )
n=1 (T ) = C \ (T ) such that = limn n . We observe k(T
n
n I) (T I)k = | n | 0 while T I
/ G(X ) (i.e. (T ) (T ), as (T ) is closed), so
supnN k(T n I)1 k = , by the Lemma above. By dropping to subsequence, assume limn k(T n I)1 k = .
Fix, for each n, kxn k = 1 in X so that
n = k(T n I)1 xn k > k(T n I)1 k
Let yn =

1
n

(T n I)

xn , so kyn k = 1. Compute

(T I)yn = (T n I)yn + (n )yn =

15

1
n
, so n .
n

1
n
xn + (n )yn 0.
n

Compact operators

15.1 Definition. Let X , Y be Banach spaces. An operator K L(X , Y) is compact if K(B(X )) Y is compact.
In particular, any compact operator is a fortiori bounded.
15.2 Remark. We observe for K B(X , Y) that the following are equivalent:
46

1. K is compact.

2. For every sequence (Kxn )


n=1 K(B(X )) there is a converging subsequence (Kxnk )k=1 .

3. K(B(X )) is totally bounded, i.e. given  > 0, there are x1 , . . . , xn B(X ) such that
K(B(X ))

n
[

(Kxi + D(Y)) .

i=1

15.3 Proposition. Let X , Y, Z be Banach spaces and K(X , Y) = {K B(X , Y) : K is compact}. Then
(i) K(X , Y) is a closed subspace of B(X , Y).
(ii) If S B(Y, Z) or T B(Z, X ) then for K K(X , Y) then SK, KT are both compact.
Proof. We have:

(i) Suppose K, L K(X , Y), F. If (xn )


n=1 B(X ), then ((Kxn , Lxn ))n=1 K(B(X )) L(B(X )), a

sequence in a compact space, admits a converging subsequence ((Kxnk , Lxnk ))k=1 and hence ((K+L)xnk )
k=1
converges in K(B(X )) Y.

Now, suppose (Kn )


n=1 K(X , Y) and K = limn Kn . Given  > 0 we let n0 be so that kK Kn k <
/3 for all n n0 , and fix such an n n0 . Since Kn (B(X )) is totally bounded, there is an (/3)-net
{Kn x1 , . . . , Kn xm } for this set. We observe for x B(X ) that there is j = 1, . . . , m such that kKn xKn xj k <
/3, and hence
kKx Kxj k kKx Kn xk + kKn x + Kn xj k + kKn xj Kxj k <




+ + = ,
3 3 3

so {Kx1 , . . . , Kxm } is an -net for K(B(X )).


(ii) We observe
compact

z
}|
{
SK(B(X )) = S(K(B(X)) S(K(B(X )))
| {z }
compact

and hence SK(B(X )) is compact. On the other hand


KT (B(Z)) K(kT kB(X )) = kT kK(B(X ))
has compact closure.
15.4 Example. We have:
(i) Let for Banach spaces X , Y
F(X , Y) = {F B(X , Y) : Im F is finite-dimensional}.
These are the so-called finite rank operators. By the Heine-Borel theorem, F(X , Y) K(X , Y).
(ii) F(X , Y)

kk

K(X , Y), from the proposition above.


kk

15.5 Remark. Often, F(X , Y) = K(X , Y) if X , Y have the approximation property [Grothendieck]. It
turns out the approximation property characterizes this equality. You can devise separable Banach spaces for
which this equality fails to be true.
(iii) Let I = [0, 1]. Let k C(I 2 ) (the kernel). Define K : C(I) C(I) by
Z 1
Kf (x) =
k(x, y)f (y) dy.
0

Weve claimed that C(I) is the codomain, which isnt completely obvious, so lets check that. Observe
Z 1
|Kf (x) Kf (x0 )|
|k(x, y) k(x0 , y)| |f (y)| dy sup |k(x, y) k(x0 , y)| kf k
yI

and k, being continuous on a compact space I 2 , is uniformly continuous, so Kf C(I). Its obvious that
R1
K is linear, and for all x I, |Kf (x)| 0 |k(x, y)| |f (y)| dy kkk kf k , so kKf k kkk kf k , so
kKk kkk .
47

These are the so-called integral operators. We will approximate K by finite-rank operators, and hence
show that K is compact. Let A = span{ : C(I)}, (x, y) = (x)(y). We observe that A is an
algebra of functions, which is point separating and conjugate closed (if F = C). Hence by Stone-Weierstrass
kk

Theorem, A
= C(I 2 ). Hence if k C(I 2 ) then there is (kn )
n=1 A such that limn kk kn k = 0.
R1
n
Hence if Kn f = 0 kn (, y)f (y) dy, then kK Kn k kk kn k 0. If
m(n)

kn =

n,i n,i ,

i=1

then Im Kn span{n,1 , . . . , n,m(n) } i.e.

R1
0

R1
n,i ()n,i (y)f (y) dy = [ 0 n,i f ] n,i .

15.6 Remark. If k, K are as above, then by Hlders inequality for 1 < p < shows that kKf k
kkk kf kp , so K K(Lp (I), C(I)). The injection map J : C(I) Lp (I) is linear and contractive (kJk 1).
Hence JK K(Lp (I)).
15.7 Theorem. If X , Y are Banach spaces and K B(X , Y) then K K(X , Y) if and only if K K(Y , X ).
Proof. We will observe that K is w -norm continuous on bounded sets. Indeed, suppose (f )N M B(Y )
(M > 0) with f0 = limN f . Then f0 M B(Y ) (Alaoglu). Given  > 0, let {Kx1 , . . . , Kxn } be an (/2M )-net

and for each N ,
of K(B(X )). Then for x B(X ), there is xj so kKx Kxj k < 2M
|f (Kx) f0 (Kx)| |f (Kx) f (Kxj )| + |f (Kxj ) f0 (Kxj )| + |f0 (Kxj ) f0 (Kx)|
kf kkKx Kxj k + |f (Kxj ) f0 (Kxj )| + kf0 kkKxj Kxk
<  + |f (Kxj ) f0 (Kxj )|.
Hence
kK f K f0 k = sup |K f (x) K f0 (x)| = sup |f (Kx) f0 (Kx)|
xB(X )

xB(X )
N

 + max |f (Kxj ) f0 (Kxj )| 


j=1,...,n

and hence limN kK f K f0 k = 0, so A3P5 tells us that K : M B(Y ) X is w -norm continuous.


Now, B(Y ) is w -compact, by Alaoglus Theorem, so K (B(Y )) is k k-compact in X .
The usual Arzel-Ascoli proof will be typed and posted on the website.
Proof, continued. () We showed that on bounded sets in Y , K is w -norm continuous.
\ )) = K (B(X ))
() We have, from above, that if K is compact, then K is compact. Hence we see that K(B(X

K (B(X )); the first equality here is from the first proposition on adjoint operators. Also, the last space is compact.
\ )) is compact (
Hence K(B(X ))
= K(B(X
= denotes isometry).
15.8 Corollary. If X , Y are reflexive spaces, then for K B(X , Y) we have K K(X , Y) if and only if K is
w-norm continuous on bounded sets.
Proof. On reflexive spaces, w = w coincide. Also, K = (K ) . Appeal to the proof above.

16

Structure theorem for compact operators

16.1 Lemma (Key Lemma). Let X be a Banach space, and K K(X ). Suppose that there exist:
a sequence Y0 ( Y1 ( Y2 ( . . ., each Yn a closed subspace, and
scalars (n )
n=1 such that (n I K)Yn Yn1 .
Then limn n = 0.
Proof (notes 16.1). Suppose not. Then by dropping to a subsequence if necessary, we may suppose |n |  > 0 for
each n, for some  > 0. By the Riesz Lemma (used to prove B(X ) is not compact if dim X 6< ) there exists for each

48

i some xi Yi such that dist(xi , Yi1 ) > 21 , and kxi k 1. Let yi = (K i I)xi Yi1 , so Kxi = i xi + yi Yi .
If j < i, we have
kKxi Kxj k = ki xi + yi Kxj k = |i | kxi +

1

(yi Kxj )k |i | dist(xi , Yi1 ) .
i | {z }
2
Yi1

Hence (Kxi )
i=1 has no converging subsequence, which means that K is not compact.
16.2 Remark. We have:
(i) Let C be a compact set. If the interior 6= then || = c. Indeed, let z0 . For each z S
(i.e. |z| = 1) let rz = max{r > 0 : z0 + rz } > 0. Then {z0 + rz z} , so |S| ||. By the
Cantor-Bernstein-Schroeder Theorem, we have that || = c. Thus, if || 0 then = and hence
= is countable.
(ii) If T L(X ) (X a C-vector space) and {1 , . . . , n } is a collection of distinct eigenvalues of T and xi
ker(i I T ) \ {0} then {x1 , . . . , xn } is linearly independent.
Proof by induction. Let n = 2. If x1 = 2 x2 then 1 2 x2 = 1 x1 = T x1 = T (2 x2 ) = 2 2 x2 , so 2 (1
2 )x2 = 0 which contradicts that x2 6= 0. Hence having proved this for n 1, suppose x1 = 2 x2 + . . . + n xn .
Then 1 (2 x2 + . . . + n xn ) = 1 x1 = T x1 = T (2 x2 + . . . + n xn ) = 2 2 x2 + . . . + n n xn , so 2 (1
2 )x2 + . . . + n (1 n ) = 0, which contradicts the linear independence.
16.3 Theorem (Structure Theorem for Compact Operators). Let X be a C-Banach space, K
K(X ).
(i) If C\{0}, each ker[(I K)n ] is finite-dimensional and there is n0 so that ker[(I K)n ] = ker[(I K)n0 ]
for n n0 .
(ii) If dim X 6< , (K) = p (K) {0} and p (K) = {1 , 2 , . . .} where limn n = 0 if p (K) is infinite.
Proof. We have:
(i) Consider first the sequence ker(I K) ker(I K)2 . . .. If n0 , as suggested, does not exist, then let
n1 = 1 and nk+1 = min{n N : ker(I K)nk ( ker(I K)n }. Let Yk = ker(I K)nk , and Y0 = {0}, so
Y0 ( Y1 ( Y2 ( . . .. We have that (I K)Yj Yj1 , so by the Key Lemma we would obtain limn = 0,
but is constant which is absurd. Hence n0 exists, as claimed.
Now suppose dim ker(I K)n 6< for some n. Pick the nk , above, for which dim Yk1 < and dim Yk 6< .
Let {xi }
i=1 Yk \ Yk1 be linearly independent. Let V0 = Yk1 and Vn = span{V0 , x1 , . . . , xn }. We observe
that (I K)Vn V0 Vn1 . As well, V0 ( V1 ( V2 ( . . ., and hence by the Key Lemma we have limn = 0,
which is still absurd.
(ii) We first prove that ap (K) p (K) {0}. If ap (K) \ {0}, then there is (xn )
n=1 S(X ) such that
n

k(I K)xn k 0. Since K is compact, (Kxn )n=1 K(B(X )) admits a subsequence (Kxnj )
j=1 which
j

converges to y. Hence kxnj yk k(I K)xnj k + kKxnj yk 0, so K 1 y = limj Kxnj = y


so Ky = y. Hence p (K). Suppose {n }
n=1 (K) is a sequence of distinct elements. We recall
that (K) ap (K) by an earlier Proposition, and moreover, ap (K) p (K) {0}, from above. Consider
spaces Y0 = {0}, Yn = span{ker(i I K) : i = 1, . . . , n}. By linear independence of non-zero eigenvectors
associated to distinct eigenvalues, we have that
Y0 ( Y1 ( Y2 ( . . .
We observe that (n I K)Yn Yn1 . Hence, by the Key Lemma, limn n = 0. Hence we conclude that

[
1
(K)
(K) \ D {0}
{z n }
n=1 |
finite

so that (K) is countable. Hence we have that (K) itself is countable (remark, last class). We thus have
(K) = (K) = ap (K) p (K) {0}. Thus, either |p (K)| < , in which case 0 p (K) since each
non-zero eigenvalue has finite-dimensional eigenspace, that is, dim ker[(I K)n ] < , by (i) above. And, if
|p (K)| <
6 , then 0 is the only cluster point of p (K). Because (K) is compact, 0 (K).

49

16.4 Aside. |p (K)| < . p (K) = {1 , . . . , n }. Let Vk = ker[(k I K)nk ] where nk is the maximal exponent
as in (i). dim Vk < . Then dim span{v1 , . . . , vn } < .
16.5 Example. It is possible for 0 6= K K(X ) with empty point spectrum.
Let K B(`p ), 1 p < , be given by K(x1 , x2 , . . .) = (0, x1 , 21 x2 , 13 x3 , . . .) (weighted shift).
K(x1 , x2 , . . .) = (0, x1 , 12 x2 , 31 x3 , . . .)
1
1
K 2 (x1 , x2 , . . .) = (0, 0, 12 x1 , 32
x2 , 43
x3 , . . .)
..
.
1
2
32
k!
K n (x1 , x2 , . . .) = (0, 0, . . . , 0, n!
x1 , (n+1)!
x2 , (n+2)!
x3 , . . . , (n+k)!
xk , . . .).

Note

1
n!

k!
(n+k)!

for k in N. We can compute


kK n xkp

1
1
kxkp = kK n k
n!
n!

hence the spectral radius formula tells us that


1
= 0.
r(K) = lim kK n k1/n lim
n
n
n
n!
Use an estimate like n! (n/2)n/2 . Think back to the theory of converging series the root test is the primary
theoretical test. Theres a really nice exercise that shows that if the ratio of the n + 1st term over the nth term
has really nice behaviour, then so do the roots, so the ratio test follows from the root test. Hence (K) = {0}. So
either you believe the proof we gave in terms of compact operators on infinite-dimensional Banach spaces, or you
believe the Liouville theorem argument that the spectrum is non-empty. Either way it has to contain 0. We note
for x 6= 0 that
!1/p

X
1
p
kKxkp =
|xn |
np
n=1
and necessarily at least one |xn |p is not zero, so this norm is not zero, which tells us that ker K = {0}. This gives
us a little bit of a warning: even as pretty as this theorem, the first part of it might essentially have no content;
there might be no eigenvalues to deal with. Because we were only able to get finite dimensionality of eigenspaces
with non-zero eigenvalues, we see that the zero part of the spectrum can contain all the craziness of the operator.
16.6 Remark. There are quite a few ways to manufacture operators that contain only 0 in the spectrum. Let
I = [0, 1], let K be a continuous function on K C(I 2 ) such that k(x, y) = 0 when x y. For example, let
(
0
if x y
k(x, y) =
x y if x > y
R1
R1
If we let K be the integral operator coming from this kernel, Kf (x) = 0 k(x, y)f (y) dy = x k(x, y)f (y) dy. Now
we can write down a formula for what happens when we iterate this operator. We can check
kK n f kp

kkkn
kf kp
n!

and hence r(K) = {0}. If k is continuous partially differentiable in the 1st coordinate, then one can show that
p (K) = .
16.7 Remark. If Kn (x1 , x2 , . . .) = (0, x, 21 x2 , . . . , n1 xn , 0, 0, . . .). Kn F(`p ) K(`p ). Check
k(K Kn )xkp

17

1
kxkp = lim kKn Kk = 0.
n
n+1

Operators on Hilbert space

17.1 Definition. Let H, L be C-Hilbert space. Given T B(H, L). We define B(L, H) by (x, T y) := (T x, y) for
all x H, y L. Indeed, for fixed y L, x 7 (T x, y) is a bounded linear functional so by Riesz Representation
Theorem there exists T y H such that
(T x, y) = (x, T y).
Check that y 7 T y is linear.
50

Cost: (T ) = T for C.
Advantage: If L = H we can now compose T S in H.
Office hours this week. Thursday 2:30-4. Friday 2-3:30. Typo 5(b) p (Mf ) 6= .
Alternate exam time: if you wish to take advantage, send an email including all other exam times.
Recall: for H a Hilbert space, T B(H), we define T B(H) by
(T x, y) = (x, T y).
von Neumann preferred Hilbert spaces to model quantum mechanics. It plays host to non-commutative geometry;
I am involved with non-commutative harmonic analysis (a more general kind of harmonic analysis). The notion of
the adjoint is just mildly different than it is in the Banach space setting. If we have an operator on a Banach space,
then the adjoint will be defined its dual space. Hilbert spaces are essentially self-dual (we say essentially, because
the small cost of conjugation, in the complex realm).
17.2 Proposition. If H, K, L are Hilbert spaces and S, T B(H, K) and R B(K, L) then
(i) T B(K, H) and kT k = kT k.
(ii) (T + S) = T + S (meaningful only if F = C, but generally for spectral theory this is the domain in
which we wish to stay).
(iii) (RT ) = T R .
(iv) T = T (in Banach spaces, we couldnt quite say anything this nice; compare to the statement T x
= Tcx).
(v) If K = H, then kT T k = kT k2 (C -identity; makes a profound theory no study of Banach algebras is
complete without a substudy of C -algebras).
This is not very hard to prove. We will give very small hints at the details of C -algebra theory.
Proof. (i), (iii) and (iv) are pretty much just like the Banach space case. Also, (ii) is very straightforward (also
similar to the Banach space case).
(v): First, we have
kT T k kT kkT k = kT k2 .
Hence, it is really required to show kT T k kT k2 . We have
kT k2 = sup{kT xk2 : x B(H)} = sup{(T x, T x) : x B(H)} = sup{(T T x, x) : x B(H)}.
By Cauchy-Schwarz,
sup{kT T xkkxk : x B(H)} = kT T k.
This will be the driving force for the theorem on the spectral theorem for compact Hermitian operators on Hilbert
space.
We mention two theorems whose proofs are just like the Banach space case.
17.3 Theorem (Kernel-Annihilator Theorem). If H, K are Hilbert spaces, T B(H, K), then
ker T Im(T ).
17.4 Remark. In Hilbert spaces, if Y is a subspace of H, then Y = Y a = Ya (due to reflexivity).
17.5 Theorem (Schauder). If H, L are Hilbert spaces, and K B(H, L) then the following are equivalent:
(i) K K(H, L) K K(L, H).
(ii) K is w -norm continuous on bounded sets.
There is only a modest difference, because K is only slightly differently described as K from the Banach space
setting.
Lets talk about some particular classes of operators on a Hilbert space.
17.6 Definition. Let H be a Hilbert space, and T B(H). We say that
51

(i) T is positive if (T x, x) 0.
(ii) T is Hermitian (or self-adjoint) if T = T .
(iii) T is normal if T T = T T .
17.7 Proposition. If H is a Hilbert space, T B(H), then
(i) T is Hermitian if and only if [x, y] = (T x, y) is a Hermitian form.
(ii) T is positive implies T is Hermitian.
(iii) T T is positive.
(iv) If T is Hermitian, then T 2 is positive.
Proof. (i) ()
[y, x] = (T y, x) = (y, T x) = (y, T x) = (T x, y) = [x, y].
() First, observe that if [x, y] = (T x, y) defines a Hermitian form, then [x, x] = [x, x] and hence [x, x] R. Thus
we have the polarisation identity
(T x, y) = [x, y] =

k=0

k=0

1X k
1X k
1X k
i [x + ik y, x + ik y] =
i (T (x + ik y), x + ik y) =
i (x + ik y, T (x + ik y)) = (x, T y).
{z
} 4
|
4
4
R

k=0

Remark: In finite dimensions, represent T with respect to an orthonormal basis {e1 , . . . , ed } = , and we have
[T ] = [tij ] and [T ] = [T ] = [tji ] and tij = (T ej , ei ).
(ii): If T is positive, then [x, y] = (T x, y) is a positive form, i.e. [x, x] 0 for x H. Hence this form is Hermitian,
by polarisation identity.
(iii): (T T x, x) = (T x, T x) = kT xk2 0.
(iv): T = T implies that T T = T 2 .
17.8 Proposition. If T B(H) with H a Hilbert space, then there exist unique Hermitian operators Re T, Im T
such that T = Re T + i Im T .
1
Proof. Let Re T = 12 (T + T ) and Im T = 2i
(T T ). Check that (Re T ) = Re T and (Im T ) = Im T (recall

(T ) = T ). Now suppose T = T1 + iT2 where Tk = Tk for k = 1, 2. Then

0 = T T = (T1 Re T ) + i(T2 Im T )
where T1 Re T , T2 Im T are each Hermitian. Thus, adding
0 = 0 + 0 = 2(T1 Re T ) = T1 = Re T,
and subtracting,
0 = 0 0 = 2i(T2 Im T ) = T2 = Im T.
Now we talk about why Hermitian and normal operators are absolutely as nice as an operator can be.
17.9 Corollary. If T B(H), T is normal if and only if Re T and Im T commute.
Proof. () Obvious.
() If Re T Im T = Im T Re T , then T = Re T + i Im T and Re T i Im T necessarily commute too.
17.10 Proposition. If H is a Hilbert space and T B(H), then T is normal if and only if kT xk = kT xk for all x
in H.
Proof. () kT xk2 = (T x, T x) = (T T x, x) = (T T x, x) = (T x, T x) = kT xk2 .
() 0 = kT xk2 kT xk2 = (T T x, x)(T T x, x) = ([T T T T ]x, x). Since T T T T is Hermitian, polarisation
identity shows that ([T T T T ]x, y) = 0 for all x, y H. It follows that T T T T = 0, by the Hahn-Banach
Theorem and Riesz Representation Theorem.
If we take a normal operator, its spectral radius is as large as can be: its always the norm of the operator. This is a
very magical property, and its exactly the sort of engine thats going to drive the spectral theory for both Hermitian
and normal operators. Well see this over the next two lectures (proof is just a nice mixture of the Spectral Radius
Formula, and the C -identity).
52

Office hours Th 2:30-4, F 2:00-3:30. A5 due Monday. Please make copies for study purposes. Final exam
questions posted soon.
17.11 Proposition. If H is a Hilbert space and T B(H) is normal (i.e. T T = T T ) then kT n k = kT kn .
Proof. Observe that kT k2 = kT T k by the C*-identity. Let H = T T so H = H and hence kHk2 = kH Hk =
kH 2 k. Thus for k in N:
k
k1
k1
k
kHk2 = kHk22
= kH 2 k2
= . . . = kH 2 k
Thus if n N, 2k n,
k

kHkn kHk2

= kHk2 = kH 2 k = kH n H 2

k kH n kkH 2

k kH n kkHk2

If H 6= 0, i.e. kHk =
6 0, we see that kHkn kH n k hence kH n k = kHkn . Thus
kT kn kT kn = (kT k2 )n = kT T kn = kHkn = kH n k = k(T T )n k = k(T )n T n k
where normality of T is used in the last equality.
k(T T )n k kT kn kT n k = kT kn kT n k.
Hence as above kT kn kT n k, so kT kn = kT n k.
17.12 Corollary. If T B(H) is normal, then r(T ) = kT k.
Proof. Recall Spectral Radius Formula: r(T ) = limn kT n k1/n .
This is a special property. For exame K K(`2 ) given by
Kx = (0, x1 , 21 x2 , 13 x3 , . . .)
has that kKk = 1 while r(K) = 0, i.e. (K) = {0}. Recall p (K) = .
17.13 Proposition. Let H be a Hilbert space and U B(H). Then the following are equivalent:
(i) (U x, U y) = (x, y) for x, y in H.
(ii) kU xk = kxk for x in H (i.e. U is an isometry).
(iii) U U = I.
Furthermore, if any of (i) (iii) holds, then the following are equivalent (we are not suggesting the following are
equivalent to those above):
(iv) U is surjective.
(v) U U = I (i.e. U is normal).
Recall if U U = I = U U , then U is a unitary.
Proof. (i) (ii) and (iii) (i): easy.
(ii) (iii): If kU xk2 = kxk2 , then
((U U I)x, x) = (U x, U x) (x, x) = 0
and by applying polarisation identity (U U I is Hermitian), we see that U U I = 0.
(iv) (v) (with (iii) holding).
(iii) U is injective. If U is also surjective and bounded below: kU xk kxk, by (ii), means that U 1 B(H)
exists. U 1 = IU 1 = U U U 1 = U (by (iv)).
(v) (iv): U U = I = U U (i.e. by (iii)) then U = U 1 so U is bijective.
In finite dimensions, U U = I implies U is invertible and hence U = U 1 . In infinite dimensions, this may not be
true:
U B(`2 ),
U (x1 , x2 , . . .) = (0, x1 , x2 , x3 , . . .)

53

Calculate:
U (x1 , x2 , . . .) = (x2 , x3 , . . .)
U U (x1 , x2 , . . .) = (0, x2 , x3 , . . .)

so kU U k = 1, (U U ) = U U = U U , (U U )2 = U |U{z
U} U = U U . One can calculate, moreover, that U U = I
I

or kU xk2 = kxk2 (so this is in fact an isometry, as advertised).


17.14 Lemma. Let H be a Hilbert space and H B(H) be Hermitian.
(i) p (H) R (comment: in fact (H) R).
(ii) 6= in p (H) then ker(I H) ker(I H).
(iii) If p (H), and P = Pker(IH) is the orthogonal projection and T H = HT for T B(H) then T P = P T .
Proof. We have:
(i) The form [x, y] = (Hx, y) is a Hermitian form, so [x, x] = [x, x], so [x, x] R. If x ker(I H) \ {0} for
some p (H) then (x, x) = (x, x) = (Hx, x) = (x, Hx) = (x, x) = (x, x) so = so R.
(ii) If x ker(I H), y ker(I H) then
(x, y) = (x, y) = (Hx, y) = (x, Hy) = (x, y) = (x, y)
( R). So ( )(x, y) = 0 so (x, y) = 0.
(iii) If , P are as above, then for x ker(I H),
HT x = T Hx = T (x) = T x
so T x ker(I H). Hence T (ker(I H)) ker(I H), which means that
i.e. T |{z}
P x ker(I H),

T P = P T P,

ker(IH)

so T P x = P T P x. Also, T H = HT implies (T H) = (HT ) hence HT = T H and as above T P = P T P .


Hence
P T = (T P ) = (P T P ) = P T P = T P.
LAST TIME:
17.15 Proposition. N B(H) normal if and only if r(N ) = kN k.
17.16 Lemma. H B(H) Hermitian
(i) p (H) R
(ii) 6= in p (H) implies ker(I H) ker(I H).
(iii) If T H = HT , T B(H), then for P = Pker(IH) (l projection), T P = P T .
17.17 Theorem (Spectral Theorem for Compact Hermitian Operators). Let H K(H) be a
hermitian operator (H a hilbert space) and let {1 , 2 , . . .} be a list of (H) \ {0} = p (H) \ {0}. Then
M
H = `2
ker(n I H) 2 ker H
n=1,2,3,...

P
i.e. if x H then x has a unique decomposition x = n=1,2,... xn + x0 where xn ker(n I H), x0 ker H, with
P
xn xm for n 6= m, i.e. kxk2 = n=1,2,... kxn k2 + kx0 k2 . Thus, letting Pn = Pker(n IH) (orthogonal projection),
we have
X
H=
n Pn
n=1,2,...

where this sum converges in norm if {1 , 2 , . . .} is infinite.

54

Proof. Let E = span{ker(n I H) : n = 1, 2, . . .}. Then part (ii) of the above lemma tells us that
M
E = `2
ker(n I H)
n=1,2,...

We note that E contains all eigenspaces corresponding to nonzero eigenvalues of H. Now, let H0 = E and P0 = PH0
(orthogonal projection).
We observe, first, that HH0 H0 . Indeed ,if x H0 , and y E then (Hx, y) = (x, Hy) = 0 since HE E
so Hy E and x E by assumption. That is Hx H0 whenever x H0 . Thus H = HP0 = P0 HP0 (by
proof of part (iii) of lemma above) is a hermitian operator: H0 = (P0 HP0 ) = P0 HP0 = P0 HP0 . Since H0 is
compact and since r(H0 ) = kH0 k, there exists (H0 ) such that || = kH0 k. Further, if H0 6= 0, i.e. kH0 k =
6 0,
then p (H0 ). If 6= 0 and x ker(I H0 ) \ {0} then x E because x ker(0I H) = ker H and
E ker H0 = ker P0 HP0 ker P0 = E. Thus for x as above, P0 x = x. Thus Hx = HP0 x = H0 x = x. But this
contradicts the fact that E contains all eigenspaces
of H corresponding to nonzero eigenvalues. Thus H0 = 0 and
L
H0 = ker H. Hence H = E 2 E = `2 n=1,2,... ker(n I H) 2 ker H.
To see the operator form of this, let
H1 =

n Pn .

We observe that if {1 , 2 , . . .} is infinite then limn n = 0 by the structure theorem for general compact
operators and

2


2
! 2

N
N

X




X
X






n Pn x
n Pn x = sup
n Pn = sup H1
H1




xB(H)
xB(H)
n=1

n=1

n=N +1

which by Pythagoras formula and the orthonormal basis theorem is


sup

|n |2 kPn xk2

xB(H) n=N +1

sup |n |2 0.

CHECK nN +1

So
H1 = lim

N
X

n Pn

n=1

in norm. We observe that the structure of H, above, tells us that H H1 = 0, so H = H1 .


17.18 Corollary. Let H = H K(H) (with H a Hilbert space) and T B(H) be such that T H = HT . Then
X
T =
Pn T Pn + P0 T P0
n=1,2,...

where P1 , P2 , . . . are from above. The sense of convergence above is pointwise (strong operator topology):
N
X

T x = lim

Pn T Pn x

n=1

(if {1 , 2 , . . .} is infinite).
Proof. Part (iii) of the Lemma shows that T Pn = Pn T Pn for n = 0, 1, 2, . . . Moreover
M
H = `2
Im Pn 2 Im(P0 )
n=1,2,...

means exactly that


I=

Pn + P0

n=1,2,...

in the pointwise sense described above, i.e. Ix =


result.

n=1,2,...

55

Pn x + P0 x. Then T x = T Ix, and we obtain the desired

17.19 Corollary (Simultaneous Diagonalisation Theorem). If H1 , H2 K(H) are both Hermitian and
H2 H1 = H1 H2 then there are finite rank orthogonal projections {Q1 , Q2 , . . .} K(H) such that Im Qn Im Qm
for n 6= m and for any scalars 1 , 2 C we can write
X
1 H1 + 2 H2 =
n (1 , 2 )Qn
n=1,2,...

where each n (1 , 2 ) is a C-valued bilinear form and limn n (1 , 2 ) = 0, so the series converges in norm.
P
Proof. First, write H1 = n=1,2,... n Pn as in the spectral theorem. We have, by the Corollary above, that
H2 =

Pn H2 Pn + P0 H2 P0

n=1,2,...

where P0 = Pker H1 . We have that each Pn H2 Pn is compact (in fact finite rank if n 0). So in fact each Pn H2 Pn
can be written
X
Pn H 2 Pn =
i,n Qi,n
i=1,2,...

where
X

Qi,n = Pn

i=1,2,...

(in the pointwise sense if infinite i.e. n = 0) by the structure of Pn H2 Pn . We let {Q1 , Q2 , . . .} be an enumeration
of {Qi,n : i = 1, 2, . . . ; n = 1, 2, . . .} and let n (1 , 2 ) = 1 m + 2 i,m where Qn = Qi,m .
17.20 Corollary (Spectral Theorem for Compact Normal Operators). If N K(H) is normal,
then there is a sequence of finite rank projections {Q1 , Q2 , . . .} such that
X
N=
n Qn
n=1,2,...

for scalars {1 , 2 , . . .}, with limn n = 0.


Proof. Let H1 = Re N = 12 (N + N ) and let H2 = Im N =
Re N Im N = Im N Re N if and only if N is normal.

1
2i (N

N ) and we apply the Corollary above. Recall,

17.21 Theorem (Spectral Theorem for Compact Normal Operators). If N K(H) is normal,
then there exists a sequence (Qn )n=1,2,... of finite rank orthogonal projections with Im Qn Im Qm for m 6= n, and
a sequence (n )n=1,2,... C \ {0}, and limn n = 0 if infinite, such that
X
N=
n Qn
n=1,2,...

where the series converges in norm, if infinite.


17.22 Remark. By combining projections associated to n = m (i.e. taking Qn + Qm in place of Qn , Qm ) we
may assume that (n )n=1,2,... = p (N ) \ {0}.
17.23 Corollary. Given an infinite dimensional separable H, and N K(H), normal, there is a unitary U : H `2
such that U N U = Ma where Ma is the multiplication operator by a.
Proof. Let for each projection Qn , let {en1 , . . . , en,m(n) } be an orthonormal basis for the finite-dimensional space
Im Qn . Also, let {e01 , e02 , . . .} (possibly empty) be an orthonormal basis for ker N . Recall that ker N Im Qn =

ker(
Pn I N ). Let (ei )i=1 be an enumeration of {enj : n = 0, 1, 2, . . . ; j = 1, . . . , m(n); j= 1, 2, . . . if n = 0}. Let
U i=1 xi ei = (xi )i=1 `2 . Check that U is unitary. Check that U N U = Ma , a = (ai )i=1 , ai = nj according to
ei = enj .
17.24 Remark. As usual, let C((N )) denote the continuous C-valued functions on (N ). Define N : C((N ))
B(H) to be the unique continuous linear operator such that
N (p) = p(N, N )
if p(z, z) is a polynomial in z, z. Check that kN (p)k = supz(N ) |p(z, z)|. By Stone-Weierstrass, such polynomials
are dense in C((N )). The operator N is called a functional calculus. Often we write f (N ) = N (f ). Note that
for f, g C((N )), f g(N ) = f (N )g(N ), and f (N ) = f (N ) .

56

17.25 Proposition. If T K(H) is positive, then there is a unique positive operator S K(H) such that S 2 = T .
We write S = T 1/2 .
Proof. Recall, T positive means that (T x, x) 0 for all x H. Hence (T ) [0, ). Indeed, if (T ) \ {0},
then p (T ) and we have, for x ker(I T ) \ {0} that
(x, x) = (x, x) = (T x, x) 0
so > 0. By the Spectral Theorem (Hermitian),
X

T =

n Pn .

n=1,2,...

P
Let S = n=1,2,... n Pn . Suppose S1 K(H) and S1 0, S12 = T . Then S1 S12 = S12 S1 , so S1 T = T S1 . Hence by
an earlier corollary,
X
S1 =
Pn S1 Pn + P0 S1 P0
n=1,2,...

where P0 = Pker T . Observe


(Pn S1 Pn x, x) = (S1 Pn x, Pn x) 0.
Hence we again use Spectral Theorem to write
X
n,i Qn,i ,
Pn S1 Pn =

where

spani=1,2,... Im(Qn,i ) Im Pn

i=1,2,...

Then
X
n=1,2,...

n Pn = T = S12 =

(Pn S1 Pn )2 =

n=1,2,...

n=1,2,... m=1,2,...

2n,i Qn,i
| {z }

Q2n,i =Qn,i

where (*) holds since Im Pn Im Pm for n 6= m. This forces 2n,1 = n , which means S1 = S, above.
17.26 Theorem (Polar decomposition). Let K K(H). Then let |K| = (K K)1/2 . There is a unique
operator U B(H) such that
U |K| = K,
ker U = ker K
and kU xk = kxk whenever x (ker K) . Here U is called a partial isometry.
Proof. First, for x H
k|K|xk2 = (|K|x, |K|x) = (|K|2 x, x) = (K Kx, x) = (Kx, Kx) = kKxk2 .
Hence we can define V0 : Im |K| H by V0 |K|x = Kx, so kV0 yk = kyk, i.e. V0 is an isometry. Let V : Im |K| H
be the unique continuous extension of V0 . As usual, V is an isometry and is linear. By the Kernel-Annihilator
Theorem, we know that Im |K| = (ker |K|) and we note from above that ker K = ker |K|. Let U : H H be
given by V P where P = P(ker K) . We observe that if y Im |K| = (ker K) then kU yk = kV P yk = kV yk since
P y = y, hence kU yk = kyk. Also, ker U ker P = ker K so ker U = ker K.

57

You might also like