FUNCTIONAL ANALYSIS

PIOTR HAJLASZ

1. Banach and Hilbert spaces

In what follows K will denote R of C.

Definition. A normed space is a pair (X, k · k), where X is a linear space
over K and
k · k : X → [0, ∞)
is a function, called a norm, such that

(1) kx + yk ≤ kxk + kyk for all x, y ∈ X;

(2) kαxk = |α|kxk for all x ∈ X and α ∈ K;
(3) kxk = 0 if and only if x = 0.

Since kx − yk ≤ kx − zk + kz − yk for all x, y, z ∈ X,

d(x, y) = kx − yk
defines a metric in a normed space. In what follows normed paces will always
be regarded as metric spaces with respect to the metric d. A normed space
is called a Banach space if it is complete with respect to the metric d.
Definition. Let X be a linear space over K (=R or C). The inner product
(scalar product) is a function
h·, ·i : X × X → K
such that

(1) hx, xi ≥ 0;
(2) hx, xi = 0 if and only if x = 0;
(3) hαx, yi = αhx, yi;
(4) hx1 + x2 , yi = hx1 , yi + hx2 , yi;
(5) hx, yi = hy, xi,

for all x, x1 , x2 , y ∈ X and all α ∈ K.

As an obvious corollary we obtain
hx, y1 + y2 i = hx, y1 i + hx, y2 i, hx, αyi = αhx, yi ,
Date: February 12, 2009.
1
2 PIOTR HAJLASZ

for all x, y1 , y2 ∈ X and α ∈ K.

For a space with an inner product we define
p
kxk = hx, xi .
Lemma 1.1 (Schwarz inequality). If X is a space with an inner product
h·, ·i, then
|hx, yi| ≤ kxk kyk for all x, y ∈ X .

Proof. We can assume that hx, yi =

6 0. For x, y ∈ X and t ∈ R we have
0 ≤ hx + ty, x + tyi = hx, xi + t[hx, yi + hy, xi] + t2 hy, yi
= hx, xi + 2t re hx, yi + t2 hy, yi .
We obtained a quadratic function of a variable t which is nonnegative and
hence
0 ≥ ∆ = 4(re hx, yi)2 − 4hx, xihy, yi ,
(re hx, yi)2 ≤ hx, xihy, yi .
If |α| = 1, then replacing y by αy we obtain
hx, xihy, yi = hx, xihαy, αyi ≥ (re hx, αyi)2 = (re (αhx, yi))2 .
In particular for α = hx, yi/|hx, yi| we have
!!2
hx, yi
hx, xihy, yi ≥ re hx, yi = |hx, yi|2 .
|hx, yi|
This completes the proof. 2
Corollary 1.2. If X is a space with an inner product h·, ·i, then
p
kxk = hx, yi
is a norm.

Proof. The properties kαxk = |α|kxk and kxk = 0 if and only if x = 0 are
obvious. To prove the last property we need to apply the Schwarz inequality.
kx + yk2 = hx + y, x + yi = hx, xi + 2 re hx, yi + hy, yi
≤ kxk2 + 2kxk kyk + kyk2 = (kxk + kyk)2 .

Definition. A space with an inner product h·, ·i is called a Hilbert space if

it is a Banach space with respect to the norm
p
kxk = hx, xi .
Proposition 1.3 (The Polarization Identity). Let h·, ·i be an inner product
in X.
 FUNCTIONAL ANALYSIS 3

(1) If K = R, then
1
hx, yi = (kx + yk2 − kx − yk2 ) ,
4
for all x, y ∈ X.
(2) If K = C, then
1
1
kx + yk2 − kx − yk2 − i kix + yk2 − kix − yk2


hx, yi =
4 4
for all x, y ∈ X.

Proof is left as an easy exercise.

Theorem 1.4. Let (X, k · k) be a normed space over K (= R or C). Then
there is an inner product h·, ·i such that
p
kxk = hx, xi
if and only if the norm satisfies the Parallelogram Law, i.e.
kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 for all x, y ∈ X .

The Parallelogram Law has a nice geometric interpretation.

Proof. The implication ⇒ follows from a direct computation. To prove the

other implication ⇐ we define the inner product using the Polarization Iden-
tities and we check that it has all the required properties. We leave the details
as an exercise. 2

1.1. Examples. 1. Cn with respect to each of the following norms is a

Banach space
Xn
kxk1 = |xi |,
i=1
kxk∞ = max |xi |,
i=1,2,...,n
n
!1/p
X
kxkp = |xi |p , 1 ≤ p < ∞,
i=1
where x = (x1 , . . . , xn ) ∈ n
C . The Banach space (Cn , k · kp ) is denoted by
`pn .
4 PIOTR HAJLASZ

2. Cn with the inner product

n
X
hx, yi = xi yi
i=1

is a Hilbert space. Note that the corresponding norm is

n
!1/2
p X
2
hx, xi = |xi | = kxk2
i=1

so `2n is a Hilbert space. The metric associated with the norm is

v
u n
uX
d(x, y) = t |xi − yi |2 ,
i=1

i.e. it is the Euclidean metric.

3. The spaces `pn were defined over K = C, but we can also do the same
construction for K = R by replacing Cn by Rn . The resulting space is also
denoted by `pn , but in each situation it will be clear whether we talk about
the real or complex space `pn so there is no danger of a confusion.
4. Consider `p2 over K = R. Then the shape of the unit ball {x : kxk ≤ 1} is

5. A subset of the Euclidean space Rn is called an ellipsoid if it is the image

of the unit ball in Rn under a nondegenerate linear mapping L : Rn → Rn
(i.e. det L 6= 0).
For every ellipsoid E in Rn there is an inner product in Rn such that
E is the unit ball in the associated norm. Indeed, if L : Rn → Rn is an
isomorphism such that E = L(B n (0, 1)), then it suffices to define the inner
product as
hx, yi = (L−1 x) · (L−1 y)
where · stands for the standard inner product in Rn .
More precisely, if L : Rn → Rn is non-degenerate, then according to the
polar decomposition theorem (P. Lax, Linear Algebra, p. 139)
L = RU
 FUNCTIONAL ANALYSIS 5

where U is unitary and R is positive self-adjoint. The mapping R can be

computed explicitly
√
LLT = RU U T RT = R2 , R = LLT .
According to the spectral theorem there is an orthonormal basis v1 , . . . vn in
Rn (with respect to the standard inner product) such that
 
λ1 0
 λ2 
R=
 
. . 
 . 
0 λn
in this basis. That means the mapping L has the following structure. First we
rotate (the mapping U ) and then we apply R which has a simple geometric
meaning of extending the length of vectors v1 , . . . , vn by factors λ1 , . . . , λn .
Now if B n (0, 1) is the unit ball, then U (B n (0, 1)) is also the unit ball, so
L(B n (0, 1)) = R(B n (0, 1)) is an ellipsoid
√ with semi-axes λ1 v1 , . . . , λn vn of
the lengths being eigenvalues of R = LLT . These numbers are called sin-
gular values of L.

Now the ellipsoid E = L(B n (0, 1)) is the unit ball for the inner product
n n n

X X  X ai bi
ai vi , bi vi = .
i=1 i=1 i=1
λ2i
We will see later (Corollary 5.12) any real inner product space space H
of dimension n is isometrically isomorphic to `2n , i.e. Rn with the standard
inner product, so if L : `2n → H is this isometric isomorphism, the unit ball
in H is L(B n (0, 1)), so it is an ellipsoid. Thus we proved.
Theorem 1.5. A convex set in Rn is a unit ball for a norm associated with
an inner product if and only if it is an ellipsoid.

6. `∞ , the space of all bounded (complex, real) sequences x = (an )∞

n=1 with
the norm
kxk∞ = sup |xn |
n
is a Banach space. This is very easy to check.
6 PIOTR HAJLASZ

7. c1 , the space of all (complex, real) convergent sequences with the norm
k · k∞ is a Banach space.
8. c0 , the space of all (complex, real) sequences that converge to zero with
the norm k · k∞ is a Banach space.
9. Note that c0 ⊂ c ⊂ `∞ and both c0 and c are closed linear subspaces of
`∞ with respect to the metric generated by the norm.
Exercise. Prove that `∞ , c and c0 are Banach spaces.
Exercise. Prove that the spaces c and c0 are separable, while `∞ is not.
10. `p , 1 ≤ p < ∞ is the space of all (complex, real) sequences x = (xn )∞
n=1
such that
∞
!1/p
X
p
kxkp = |xn | .
n=1
It follows from the Minkowski inequality for sequences that k · kp is a norm
and that `p is a linear space. We will prove now that `p is a Banach space,
i.e. that it is complete. Let xn = (ani )∞ p
i=1 be a Cauchy sequence in ` , i.e. for
every ε > 0 there is N such that for all n, m > N
∞
!1/p
X
n m p
kxn − xm kp = |ai − ai | < ε.
i=1
Hence for each i the sequence (ani )∞
n=1 is a Cauchy sequence in K (= C or
n
R). Let ai = limn→∞ ai . Fix an integer k. Then for n, m > N we have
k
X
|ani − am p
i | <ε
p

i=1
and passing to the limit as m → ∞ yields
k
X
|ani − ai |p ≤ εp .
i=1
Now taking the limit as k → ∞ we obtain
∞
X
|ani − ai |p ≤ εp ,
i=1
i.e.
kxn − xm kp ≤ ε where x = (ai )∞
i=1 .
p p
This proves that x ∈ ` and xn → x in ` . The proof is complete.
In particular the space `2 is a Hilbert space because its norm is associated
with the inner product
∞
X
hx, yi = xi yi .
i=1
 FUNCTIONAL ANALYSIS 7

Exercise. Prove that `p , 1 ≤ p < ∞ is separable.

11. We will prove that `p for p 6= 2 is not an inner product space. Let
x = (1, 0, 0, . . .), y = (0, 1, 0, 0, . . .). If 1 ≤ p < ∞, then

kx + yk2p + kx − yk2p = 21+2/p , 2kxk2p + 2kyk2p = 4 ,

and thence the Parallelogram Law is violated. The same example can also
be used in the case p = ∞. In the real case this result can also be seen as a
consequence of the fact that the two dimensional section of the unit ball in
`p , p 6= 2, along the space generated by the first two coordinates is not an
ellipse.

12. If X is equipped with a positive measure µ, then for 1 ≤ p < ∞, Lp (µ)

is a Banach space with respect to the norm
Z 1/p
p
kf kp = |f | dµ .
X

Also L∞ (µ) is a Banach space with the norm being the essential supremum
of |f |. For the proofs see the notes from Analysis I.

For p = 2 the space L2 (µ) is a Hilbert space with respect to the inner
product
Z
hf, gi = f g dµ .
X

13. If X = {1, 2, 3, . . .} and µ is the counting measure (i.e. µ(A) = #A),

then Lp (µ) = `p . In particular this gives another proof that `p is a Banach
space. However the proof given above is much more elementary.

14. If X is a compact metric space, then the space of continuous functions

on X with respect to the norm

kf k = sup |f (x)|
x∈X

is a Banach space. This space is denoted by C(X). The resulting metric in

C(X) is the metric of uniform convergence.

15. Let H 2 be the class of all holomorphic functions on the unit disc D =
{z ∈ C : |z| < 1} such that
Z 2π
 1 1/2
kf kH 2 = sup |f (reiθ )|2 dθ < ∞.
0<r<1 2π 0

This space is called Hardy space H 2 . We will prove now that it is a Hilbert
space and we will find an explicit formula for the inner product.
8 PIOTR HAJLASZ

Since every holomorphic function in D can be represented as a Taylor

polynomial f (z) = ∞ n
P
n=0 n , for |z| < 1 we have
a z
Z 2π Z 2π  X∞ ∞
1 iθ 2 1 n inθ
 X 
|f (re )| dθ = an r e am rm eimθ dθ
2π 0 2π 0
n=0 m=0
∞ Z 2π
X 1
= an am rn+m ei(n−m)θ dθ
2π 0
n,m=0 | {z }
1 if n=m, 0 if n 6= m
∞
X
= r2n |an |2 .
n=0
Hence
 ∞ 1/2 ∞
X
X
1/2
kf kH 2 = lim r2n |an |2 = |an |2 .
r→1−
n=0 n=0
This proves that the space H 2 is isometrically isomorphic with `2 and that
the norm in H 2 is associated with the inner product
∞ Z 2π
X 1
hf, gi = an bn = lim f (reiθ )g(reiθ ) dθ ,
r→1− 2π 0
n=0

where f = n=0 an z n , g(z) = ∞

P∞ P n
n=0 bn z .

2. Linear operators

If X and Y are normed spaces, then linear functions L : X → Y will be

called (linear) operators of (linear) transforms. Also we will often write Lx
instead of L(x). We say that a linear operator L : X → Y is bounded if there
is a constant C > 0 such that
kLxk ≤ Ckxk for all x ∈ X.
Theorem 2.1. Let L : X → Y be a linear operator between normed spaces.
Then the following conditions are equivalent:

(1) I is continuous;
(2) L is continuous at 0;
(3) L is bounded.

Proof. The implication (1) ⇒ (2) is obvious. (2) ⇒ (3) For if not
there would exist a sequence xn ∈ X such that kLxn k > nkxn k. Then
kL(xn /(nkxn k))k > 1, but on the other hand kL(xn /(nkxn k))k → 0, be-
cause xn /(nkxn k) → 0 which is an obvious contradiction. (3) ⇒ (1) Let
xn → x. Then Lxn → Lx. Indeed,
kLx − Lxn k = kL(x − xn )k ≤ Ckx − xn k → 0 .
 FUNCTIONAL ANALYSIS 9

This completes the proof. 2

For a linear operator L : X → Y we define its norm by
kLk = sup kLxk .
kxk≤1

Then
kLxk ≤ kLk kxk for all x ∈ X.
Indeed,  x  x
kLxk = L kxk = kxk L ≤ kxk kLk .

kxk kxk
Thus L is bounded if and only if kLk < ∞. Moreover kLk is the smallest
constant C for which the inequality
kLxk ≤ Ckxk for all x ∈ X
is satisfied. kLk is called the operator norm.
The class of bounded operators L : X → Y is denoted by B(X, Y ). We
also write B(X) = B(X, X). Clearly B(X, Y ) has a structure of a linear
space.
Lemma 2.2. B(X, Y ) equipped with the operator norm is a normed space.

Proof is very easy and left as an exercise.

Theorem 2.3. If X is a normed space and Y is a Banach space, then
B(X, Y ) is a Banach space.

Proof. Let {Ln } be a Cauchy sequence in B(X, Y ). Then

(2.1) kLn x − Lm xk ≤ kLn − Lm k kxk .
Since the right hand side converges to 0 as n, m → ∞, we conclude that
{Ln x} is a Cauchy sequence for every x ∈ X and hence {Ln x} has a limit in
Y . We denote it by Lx = limn→∞ Ln x. Because {Ln } is a Cauchy sequence
kLn − Lm k < ε for all sufficiently large n and m and passing to the limit in
(2.1) as n → ∞ yields
(2.2) kLx − Lm xk ≤ εkxk for all sufficiently large m .
This gives
kLxk ≤ kLx − Lm xk + kLm xk ≤ (ε + kLm k)kxk .
Hence L ∈ B(X, Y ) and Lm → L in B(X, Y ) because of (2.2). 2
Definition. By a (continuous linear) functional we mean an arbitrary
bounded operator
L : X → K.
The space X = B(X, K) is called the dual space of X. Elements of X ∗ will
∗

usually be denoted by x∗ and we will write hx∗ , xi instead of x∗ (x).

10 PIOTR HAJLASZ

Corollary 2.4. If X is a normed space,then X ∗ is a Banach space.

Proof. Indeed, K is a Banach space and the result follows from Theo-
rem 2.3 2

Definition. We say that the two normed spaces X and Y are isomorphic
if there is an algebraic isomorphism of linear spaces L ∈ B(X, Y ) (i.e. it
is one-to-one and surjection) such that L−1 ∈ B(Y, X). The mapping L is
called an isomorphism of normed spaces X and Y . If X and Y are Banach
spaces we call it isomorphism of Banach spaces X and Y .

We say that X and Y are isometric if there is an isomorphism L ∈

B(X, Y ) such that kLxk = kxk for all x ∈ X.

The following result immediately follows from the equivalence of continu-

ity and boundedness of an operator.

Proposition 2.5. L ∈ B(X, Y ) is an isomorphism if it is an algebraic

isomorphism of linear spaces and there is C > 0 such that

kLxk ≥ Ckxk for all x ∈ X .

The next result is also very easy and left as an exercise.

Proposition 2.6. If normed spaces X and Y are isomorphic and X is a

Banach space, then Y is also a Banach space.

Exercise. Find two homeomorphic metric spaces X and Y such that X is

complete, while Y is not.

Example. We will construct now an example of an algebraic isomorphism of

normed spaces L ∈ B(X, Y ) such that L−1 6∈ B(Y, X). Let X be the space
of all real sequences x = (a1 , a2 , . . .) with only a finite number nonzero
components, equipped with the norm k · k∞ . Let L : X → X be defined by
a2 a3
L(a1 , a2 , . . .) = (a1 , , , . . .) .
2 3
Clearly, L is a linear isomorphism with the inverse

L−1 (a1 , a2 , a3 , . . .) = (a1 , 2a2 , 3a3 , . . .) .

It is easy to see that L is continuous (because kLxk∞ ≤ kxk∞ ), but L−1 :

X → X is not. Indeed, if xn = (0, . . . , 0, n−1 , 0, . . .) where an = n−1 and all
other ai ’s are equal zero, then kxn k∞ → 0, but kL−1 (xn )k∞ = 1.

Theorem 2.7. Let X be a Banach space. Then isomorphisms form an open

subset in B(X).
 FUNCTIONAL ANALYSIS 11

Proof. If A ∈ B(X), kAk < 1, then I − A is an isomorphism.1 Indeed, the

series of operators
X∞
(2.3) An = I + A + A2 + A3 + . . .
n=0
converge absolutely, because
X∞ ∞
X
kAn k ≤ kAkn < ∞ .
n=0 n=0
Hence it easily follows that the sequence of partial sums of (2.3) is a Cauchy
sequence in B(X). Since B(X) is a Banach space, it converges and thus
(2.3) defines a bounded operator in B(X). Now it is easy to check that this
operator is an inverse of I − A, so I − A is an isomorphism.
If L ∈ B(X) is an isomorphism and A ∈ B(X), kAk < kL−1 k−1 , then
L − A = L(I − L−1 A)
is an isomorphism as a composition of isomorphisms. We proved that a
certain ball in B(X) centered at L consists of isomorphisms. 2
Example. We will prove that the real spaces `12 and `∞
2 are isometric. That
2
is quite surprising because both spaces are R equipped with two different
norms
k(x, y)k1 = |x| + |y|, k(x, y)k∞ = max{|x|, |y|} .
However the mapping
L : `12 → `∞
2 , L(x, y) = (x + y, x − y)
is an isometry, because max{|x + y|, |x − y|} = |x| + |y|. Note that in both
spaces the unit ball is a square and the mapping L maps one square onto
another.
Exercise. Find all isometries between `12 and `∞
2 .

Exercise. Prove that the spaces `13 and `∞

3 are not isometric.

Proposition 2.8. The spaces c0 and c are isomorphic.

Proof. The mapping T : c0 → c,

L(a1 , a2 , a3 , . . .) = (a1 + a2 , a1 + a3 , a1 + a4 , . . .)
is an algebraic isomorphism with the inverse
L−1 (a1 , a2 , a3 , . . .) = (a0 , a1 − a0 , a2 − a0 , . . .) ,
where a0 = limn→∞ an . Since kLxk∞ ≤ 2kxk∞ for x ∈ c0 and kL−1 xk∞ ≤
2kxk∞ for x ∈ c we conclude that L and L−1 are continuous, so L is an
isomorphism of Banach spaces c0 and c. 2
1I stands for identity.
12 PIOTR HAJLASZ

Exercise. Prove that the spaces c0 and c are not isometric.

Definition. We say that two norms k · k1 and k · k2 on a linear space X are
equivalent if there are constants C1 , C2 > 0 such that
C1 kxk1 ≤ kxk2 ≤ C2 kxk1 for all x ∈ X .
Proposition 2.9. The two norms k · k1 and k · k2 on a linear space X are
equivalent if and only if the identity mapping id : (X, k · k1 ) → (X, k · k2 ),
id(x) = x is an isomorphism.

Proof. It easily follows from the fact that continuity of a linear mapping
is equivalent with its boundedness. 2

2.1. Examples of dual spaces.

Theorem 2.10. If s = (si ) ∈ `1 , then
∞
X
∗
(2.4) hx , xi = for x = (xi ) ∈ c0
i=1

defines a bounded functional x∗ ∈ c∗0 with kx∗ k = ksk1 . On the other hand
if x∗ ∈ c∗0 , then there is unique x ∈ `1 such that x∗ can be represented by
(2.4). This proves that the space c∗0 is isometrically isomorphic to `1 .

Proof. Let s ∈ `1 . Then x∗ defined by (2.4) is a bounded linear functional

on c0 . Indeed,
∞
X
|hx∗ , xi| ≤ sup |xi | |si | = ksk1 kxk∞
i i=1
proves that
(2.5) kx∗ k ≤ ksk1 .
Now let x∗ ∈ c∗0 . We will prove that there is s ∈ `1 such that x∗ satisfies
(2.4). It is clear that two different elements of `1 define different functionals
on c0 , so uniqueness is obvious. We will also prove that
(2.6) ksk1 ≤ kx∗ k
which together with (2.5) will give the equality kx∗ k = ksk1 . This will com-
plete the proof.
Let ei = (0, . . . , 0, 1, 0, . . .) ∈ c0 with 1 on ith coordinate and let si =
hx∗ , ei i. Define zi = si /|si | if si 6= 0 and zi = 0 if si = 0. Then
k
X
z k = (z1 , . . . , zk , 0, 0, . . .) = zi ei ∈ c0 , kz k k∞ ≤ 1 .
i=1
 FUNCTIONAL ANALYSIS 13

Hence
k
X k
X
kx∗ k ≥ hx∗ , z k i = zi hx∗ , ei i = |si | .
| {z }
i=1 si i=1

Letting k → ∞ we have
∞
X
ksk1 = |si | ≤ kx∗ k
i=1

`1
which proves that s ∈ along with the estimate (2.6). Now it easily follows
that x∗ satisfies (2.4). Indeed, if x = (xi ) ∈ c0 and
k
X
xk = (x1 , . . . , xk , 0, 0, . . .) = xi ei ,
i=1
Pk P∞
then xk
→ x in c0 and i=1 xi si → i=1 xi si , because s ∈ `1 . Hence passing
to the limit in the equality
k
X
hx∗ , xk i = xi si
i=1
yields (2.4). The proof is complete. 2
Exercise. Prove that the dual space c∗ is isometrically isomorphic to `1 .
Theorem 2.11. If s = (si ) ∈ `∞ , then
∞
X
(2.7) hx∗ , xi = si xi for x = (xi ) ∈ `1
i=1

defines a bounded functional x∗ ∈ (`1 )∗ with kx∗ k = ksk∞ . On the other hand
if x∗ ∈ (`1 )∗ , then there is unique s ∈ `∞ such that x∗ can be represented by
(2.7). This proves that the space (`1 )∗ is isometrically isomorphic to `∞ .

Proof. The proof is pretty similar to the previous one, so we will be short.
If s ∈ `∞ , then it is easily seen that x∗ given by (2.4) defines a functional
on `1 with kx∗ k ≤ ksk∞ . Now let x∗ ∈ (`1 )∗ . It remains to prove that
there is s ∈ `∞ such that x∗ satisfied (2.7) and ksk∞ ≤ kx∗ k (uniqueness is
obvious). Let ei = (0, . . . , 0, 1, 0, . . .) ∈ `1 and si = hx∗ , ei i. Let zi = si /|si |
if si 6= 0 and zi = 0 if si = 0. Put z i = zi ei ∈ `1 , so kz i ki ≤ 1. Then
kx∗ k ≥ hx∗ , z i i = |si |. Now taking supremum over all i yields ksk∞ ≤ kx∗ k
and the result easily follows. 2
Exercise. Prove that for every s = (si ) ∈ `1 , hx∗ , xi = ∞
P
i=1 si xi defines a
bounded functional on `∞ with kx∗ k = ksk1 .
Later we will see that not every functional in (`∞ )∗ can be represented by
an element of `1 , and the above exercise proves only that `1 is isometrically
isomorphic to a closed subspace of (`∞ )∗ .
14 PIOTR HAJLASZ

Theorem 2.12. Let 1 < p < ∞. If s = (si ) ∈ `q , where q = p/(p − 1), then
∞
X
(2.8) hx∗ , xi = si xi for x = (xi ) ∈ `p
i=1

defines a bounded functional x∗ ∈ (`p )∗ with kx∗ k = kskq . On the other hand
if x∗ = (`p )∗ , then there is unique s ∈ `q such that x∗ can be represented by
(2.8). This proves that the space (`p )∗ is isometrically isomorphic to `q .

Proof. Again, the proof is very similar to those presented above, so we will
explain only the step where a tiny difference in the argument occurs. Let
x∗ ∈ (`p )∗ . The crucial point is to show that for si = hx∗ , ei i we have kskq ≤
kx∗ kp . To prove this we take zi = si |si |q−2 and z k = (z1 , . . . , zk , 0, 0, . . .) ∈ `p .
Then
k
X
∗ k ∗ k
(2.9) kx k kz kp ≥ hx , z i = |si |q .
i=1

Since kz k kp = ( ki=1 |si |q )1/p inequality (2.9) yields kx∗ k ≥ ( ki=1 |si |q )1/q
P P
and the claim follows after passing to the limit as k → ∞. 2
The last two results are special cases of the following deep result whose
proof based on the Radon-Nikodym theorem is presented in notes from Anal-
ysis I and also will be proved in Section 5.5.
Theorem 2.13. If µ is a σ-finite measure on X and 1 ≤ p < ∞, 1 < q ≤ ∞,
p−1 + q −1 = 1, then for every function g ∈ Lq (µ),
Z
(2.10) Λf = f g dµ for f ∈ Lp (µ)
X

defines a bounded functional Λ ∈ (Lp (µ))∗ with

kΛk(Lp (µ))∗ = kgkLq (µ) .
Moreover for every functional Λ ∈ (Lp (µ))∗ there is unique g ∈ Lq (µ) such
that Λ can be represented by (2.10). This proves that the space (Lp (µ))∗ is
isometrically isomorphic to Lq (µ).

Definition. Let X be a locally compact metric space and K = R or C. We

define

(1) Cc (X) is the space of all continuous functions u : X → K for which

there is a compact set E ⊂ X such that u vanishes outside E, i.e.
u(x) = 0 for all x ∈ X \ E.
(2) C0 (X) is the space of all continuous functions such that for every
ε > 0 there is a compact set E ⊂ X so that |u(x)| < ε for all
x ∈ X \ E.
 FUNCTIONAL ANALYSIS 15

Exercise. Prove that u ∈ C0 (Rn ) if and only if u is continuous and

lim|x|→∞ u(x) = 0.
Theorem 2.14. Let X be a locally compact metric space. Then C0 (X) is a
Banach space with the norm
kuk = sup |u(x)| .
x∈X

Moreover Cc (X) forms a dense subset in C0 (X).

We leave the proof as an exercise.

If µ is a signed measure, then we have unique Hahn decomposition
µ = µ+ − µ− ,
where µ+ and µ− are positive measures concentrated on disjoint sets. We
define the measure |µ| as
|µ| = µ+ + µ− .
The number |µ|(X) is called total variation of µ.
Example. If µ(E) = E f dµ where f ∈ L1 (µ), then |µ|(E) = E |f | dµ.
R R

Theorem 2.15 (Riesz representation theorem). If X is a locally compact

metric space and Φ ∈ (C0 (X))∗ , then there is unique Borel signed measure
µ of finite total variation such that
Z
Φ(f ) = f dµ .
X

Moreover kΦk = |µ|(X).

Thus the dual space of C0 (X) is isometrically isomorphic to the space of

signed measures of finite total variation.

3. Finitely dimensional spaces.

Theorem 3.1. In a finitely dimensional linear space any two norms are
equivalent. In particular every finitely dimensional normed space is a Banach
space.

Proof. Let e1 , e2 , . . . , en be a Hamel basis of the linear space X. For every

x ∈ X we define a norm
n
X
kxk0 = max |xi | where x = xi ei .
i=1,2,...,n
i=1
16 PIOTR HAJLASZ

Note that the space X with respect to the norm k · k0 is locally compact and
complete. It suffices to prove that every norm in X is equivalent with k · k0 .
Let k · k be an arbitrary norm in X. Then

X n X n n
X
0
kxk = xi ei ≤ |xi |kei k ≤ kxk kei k = Ckxk0 .


i=1 i=1 i=1

Now it remains to prove the opposite inequality kxk0 ≤ Ckxk. By contradic-

tion suppose that for every n we can find xn ∈ X such that kxn k0 ≥ nkxn k.
Let yn = xn /kxn k0 . Then kyn k0 = 1 and kyn k ≤ 1/n → 0, so yn → 0
in the norm k · k. Since the space (X, k · k0 ) is locally compact, there is a
subsequence yni → y convergent with respect to the norm k · k0 . Since the
convergence in the norm k · k0 implies the convergence in the norm k · k
(because kxk ≤ Ckxk0 ) we conclude that y = 0 which is a contradiction,
because kyk0 = limi→∞ kyni k0 = 1. The last part of the theorem is easy,
because the space (X, k · k0 ) is complete. 2
As an application we will prove
Proposition 3.2. For every polynomial P on Rn there is a constant C > 0
such that on every ball B(x, r) in Rn we have
Z
(3.1) sup |P | ≤ C|B(x, r)|−1 |P | .
B(x,r) B(x,r)

Remark. The opposite inequality

Z
−1
|B(x, r)| |f | ≤ sup |f |
B(x,r) B(x,r)

is true for any locally integrable f . It is also clear that for every ball there is
a constant C such that (3.1) is satisfied, however, we want to find a constant
C that will be good for all the balls at the same time.
Proof. Let k be the degree of P . The space P k of polynomials of degree
less than or equal k has finite dimension, so all norms in that space are
equivalent. In particular there is a constant C > 0 such that
Z
(3.2) sup |Q| ≤ C|B(0, 1)|−1 |Q| for all Q ∈ P k .
B(0,1) B(0,1)

Now it remains to observe that a linear change of variables allows us to

rewrite (3.1) in an equivalent form of (3.2) without changing the constant
C, where Q is obtained from P by a linear change of variables. Since degree
of Q equals k and the claim follows. 2
While Theorem 3.1 proves that any two norms in a finitely dimensional
vector space are comparable the theorem does not give an explicit estimate
for the comparison. However the following result does.
 FUNCTIONAL ANALYSIS 17

Theorem 3.3 (F. John). Let (X, k · k) be an n-dimensional real normed

space. Then there is a Hilbert norm k · k0 in X such that
√
kxk ≤ kxk0 ≤ n kxk for all x ∈ X.

Proof. The unit ball with respect to the norm k · k is a convex symmetric
body K (symmetric means that x ∈ K ⇒ −x ∈ K) and any ellipsoid E
centered at 0 is a unit balls for a Hilbert norm k · k0 (we choose the inner
product so that the semi-axes are orthonrmal, see Section 1.1, Example √ 5).
Now it suffices to prove that there is an ellipsoid E such that E ⊂ K ⊂ n E.
Indeed, E ⊂ K means that kxk0 √ ≤ 1 ⇒ kxk ≤ 1 which easily implies√
kxk ≤ kxk0 for any x ∈ X and√K ⊂ n E means that kxk ≤ 1 ⇒ kxk0 ≤ n
which easily implies kxk0 ≤ n kxk for any x ∈ X. Therefore we are left
with the proof of the following result.
Theorem 3.4 (F. John ellipsoid theorem). If K is a closed convex sym-
metric body in an n-dimensional real vector space X, then there is a closed
ellipsoid E centered at 0 such that
√
E ⊂ K ⊂ nE .

Proof. Choose any Euclidean coordinate system in X and define E to

be an ellipsoid centered at 0 contained in K that has maximal volume. A
simple compactness argument shows that such an ellipsoid exists. Now we
choose a new Euclidean coordinate system in which E becomes the unit ball
(this system is obtained by rescaling semi-axes). Note that also in this new
system√ E is an ellipsoid of maximal volume in K. It√remains to prove that
K ⊂ n E, i.e. K is contained in the ball of radius n.
By contradiction suppose
√ that there is an element in K whose distance
to 0 is greater than n. By rotating the coordinate system we can assume
that
√
(t, 0, . . . , 0) ∈ K, t > n .
An elementary geometric argument shows that the tangent cone to E with
the vertex at (t, 0, . . . , 0) touches the unit ball E at points where x1 = 1/t.

Since K is convex and symmetric it follows that the set W consisting of

the ball E and the two tangent cones centered at (±t, 0, . . . , 0) is contained
in K.
18 PIOTR HAJLASZ

The ball E is described by the equation

x21 + y 2 ≤ 1, y = (x22 + . . . + x2n )1/2 .
Now consider the family of ellipsoids Eε for 0 < ε < 1,
(1 − ε)n−1 x21 + y 2 /(1 − ε) ≤ 1 .
Since the x1 semi-axis of Eε has length (1 − ε)−(n−1)/2 and the length of
each of the semi-axes x2 , . . . , xn is (1 − ε)1/2 , the ellipsoids Eε have the
same volume as E. It suffices to prove that for ε sufficiently small Eε is
contained in the interior of W . Indeed, it is then contained in the interior
of K and hence even if we enlarge it slightly it is still contained in K which
will contradict maximality of the volume of E.
Since
1
(1 − ε)n−1 = 1 − (n − 1)ε + O(ε2 ), = 1 + ε + ε2 = . . . = 1 + ε + O(ε2 ) ,
1−ε
points of the ellipsoid Eε satisfy
x21 + y 2 − ε((n − 1)x21 − y 2 ) + O(ε2 ) ≤ 1
which is equivalent to
(3.3) (x21 + y 2 − 1)(1 + ε) ≤ ε(nx21 − 1) + O(ε2 ) .
√
We consider the√two parts of the ellipsoid Eε : |x1 | < (1/t + 1/ √ n)/2 and
|x1 | ≥ (1/t + 1/ n)/2 separately. In the first case |x1 | < (1/t + 1/ n)/2 we
have
 1 1 1   1 1 1 
nx21 − 1 < n + √ + − 1 < n + + − 1 = 0,
4t2 2t n 4n 4n 2n 4n
√
where in the last inequality we used t > n. Hence for ε sufficiently small
the right hand side of (3.3) is negative and thus
x21 + y 2 < 1 ,
√
i.e. points of Eε satisfying |x1 | < (1/t + 1/ n)/2 belong to the interior of
E and hence interior √ of W . Regarding the other part of the ellipsoid Eε
≥ (1/t + 1/ n)/2 note that the part of the ball E with |x1 | ≥
with |x1 | √
(1/t + 1/ n)/2 is a compact set contained in the interior of W , because
 FUNCTIONAL ANALYSIS 19

√
|x1 | ≥ (1/t + 1/ n)/2 > (1/t + 1/t)/2 = 1/t (see the picture describing W ).
Now if ε is sufficiently small the ellipsoid Eε is a tiny
√ perturbation of E and
hence also the part of Eε with |x1 | ≥ (1/t + 1/ n)/2 is contained in the
interior of W . The proof is complete. 2
As an application of Theorem 3.1 we immediately obtain
Corollary 3.5. In every finitely dimensional normed space a set is compact
if and only if it is bounded and closed.

That means finitely dimensional normed spaces are locally compact. We

will prove that this characterizes finitely dimensional normed spaces, i.e no
normed space of infinite dimension is locally compact.
Theorem 3.6 (The Riesz lemma). Let X0 6= X be a closed linear subspace
of a normed space X. Then for every ε > 0 there is y ∈ X such that
kyk = 1 and ky − xk ≥ 1 − ε for all x ∈ X0 .

Proof. Fix y0 ∈ X \ X0 and define % = inf x∈X0 kx − y0 k. Clearly % > 0,

because X0 is closed. Choose η > 0 such that η/(% + η) ≤ ε, and then choose
x0 ∈ X0 such that % ≤ ky0 − x0 k ≤ % + η. We will prove that the vector
y = (y0 − x0 )/ky0 − x0 k has the properties that we need. Obviously kyk = 1.
Moreover for x ∈ X0 we have
1
ky − xk = y0 − x0 − ky0 − x0 kx

ky0 − x0 k
1
= y0 − (x0 + ky0 − x0 kx)

ky0 − x0 k | {z }
∈X0
% η
≥ =1− ≥ 1 − ε.
%+η %+η
The proof is complete. 2
Corollary 3.7. In a normed space of infinite dimension no closed ball is
compact.

Proof. It suffices to prove that the closed unit ball centered at 0 is not
compact. To prove the lack of compactness of this ball it suffices to prove
the existence of a sequence {xi } ⊂ X such that
1
kxi k = 1, kxi − xj k ≥ for i 6= j .
2
We construct the sequence by induction. First we choose x1 with kx1 k = 1
arbitrarily. Now suppose that the elements x1 , . . . , xn have already been
defined. Let X0 = span {x1 , . . . , xn }. Since X is infinitely dimensional, X0 6=
X and hence the Riesz lemma implies that there is xn+1 ∈ X \ X0 such that
1
kxn+1 k = 1 and kxn+1 − xk ≥
2
20 PIOTR HAJLASZ

for every x ∈ X0 . 2
Note that if in the Riesz lemma dim X0 < ∞, by local compactness of X0
we can take in the proof x0 ∈ X0 such that ky0 − x0 k = % and hence the
same proof gives
Corollary 3.8. Let X0 6= X be a closed linear subspace of a normed space
X with dim X0 < ∞. Then there is y ∈ X such that
kyk = 1 and ky − xk ≥ 1 for all x0 ∈ X0 .

In general Corollary 3.8 does not hold if dim X0 = ∞ as the next example
shows, see, however, Theorem 14.11.
Example. Consider the closed linear subspace X of C[0, 1] consisting of
functions vanishing at 0. Let
n Z 1 o
X0 = f ∈ X : f (x) dx = 0 .
0
It is easy to see that X0 is a proper closed linear subspace of X. We will
prove that there is no function f ∈ X such that
(3.4) kf k = 1 and kf − gk ≥ 1 for all g ∈ X0 .
Assume that such a function f ∈ X exists. Since f is continuous, f (0) = 0
and kf k = supx∈[0,1] |f (x)| = 1, we conclude that
Z 1
(3.5) |f (x)| dx < 1 .
0
For every h ∈ X \ X0 we set
R1
f (x) dx
g = f − ch, c = R01
0 h(x) dx
and note that the denominator in nonzero, because h 6∈ X0 . Clearly g ∈ X0
and (3.4) yields
1 ≤ kf − gk = kf − (f − ch)k = |c| khk ,
i.e. Z 1  Z 1 
h(x) dx ≤  f (x) dx sup |h(x)| .
   

0 0 x∈[0,1]

Choosing h(x) = x1/n ∈ X \ X0 gives

n Z 1 
≤ f (x) dx for all n = 1, 2, 3, . . .
 
n+1 0
Now passing to the limit as n → ∞ we obtain
Z 1
|f (x)| dx = 1
0
which contradicts (3.5). 2
 FUNCTIONAL ANALYSIS 21

4. Operations on Banach spaces

4.1. Subspace. If (X, k · k) is a normed space and Y ⊂ X is a linear sub-

space, then (Y, k · k) is a normed space.
Proposition 4.1. A subspace Y of a Banach space X is a Banach space if
and only if it is a closed subspace of X.

This result is obvious.

Every metric space is isometric with a dense subspace of a complete metric
space. The idea of the proof is to add to the space points which are abstract
limits of Cauchy sequences. More precisely, in the space
X 0 = {xn }∞ ∞

n=1 ⊂ X : {xn }n=1 is a Cauchy sequence
we introduce the equivalence relation
{xn } ∼ {yn } if and only if d(xn , yn ) → 0 .
Note that if {xn } and {yn } are convergent sequences, then {xn } ∼ {yn } if
and only if they have the same limit. Thus the relation ∼ identifies those
Cauchy sequences that should have the same limit. We define X̂ = X 0 / ∼
with the metric
d([{xn }], [{yn }]) = lim d(xn , yn ) .
n→∞
The space X can be identified with a subset of X̂ through the embedding
onto constant sequences
X 3 x 7→ {x, x, x, . . .}
After this identification X becomes a dense subset of X̂. It is easy to verify
that X̂ is a complete metric space.
If X is equipped with a structure of a normed space, then also X̂ has a
natural normed space structure and hence it is a Banach space. Thus we
sketched a proof of the following result.
Theorem 4.2. Every normed space is isometrically isomorphic to a dense
linear subspace of a Banach space.

4.2. Quotient space. If X is a linear space and Y ⊂ X is a linear subspace,

then the equivalence relation in X
x∼y if and only if x−y ∈Y
defines the linear quotient space
X/Y = X/ ∼ .
Elements of the space X/Y can be identified with cosets [x] = {x + y : y ∈
Y }.
22 PIOTR HAJLASZ

Definition. By a seminorm on a liner space X we mean any function

| · | : X → [0, ∞) such that

(1) | x + y|| ≤ | x|| + | y|| for x, y ∈ X;

(2) | αx|| = |α| | x|| for all x ∈ X, α ∈ K.

The only difference between the seminorm and the norm is that it can vanish
on nonzero elements. A seminorm vanishes on a linear subspace if X.
If X is a normed space and Y ⊂ X is a linear subspace, then we equip
X/Y with a seminorm
k[x]k = inf kzk = inf kx − yk = dist (x, Y ) .
z∈[x] y∈Y

Theorem 4.3. The quotient space X/Y is a normed space if and only if Y
is a closed subspace of X.

Proof. ⇒. Suppose that k · k is a norm in X/Y , but Y is not closed. Then

there is Y 3 yn → x0 ∈ / Y . Since x0 6∈ Y , [x0 ] 6= 0 in X/Y and hence k[x0 ]k >
0 (because k · k is a norm). On the other hand k[x0 ]k ≤ inf n kx0 − yn k = 0
which is a contradiction.
⇐. If k[x]k = 0, then there is a sequence yn ∈ Y such that kx − yn k → 0
and hence x ∈ Y = Y which yields [x] = 0. Thus k · k is a norm. 2
Lemma 4.4. A normed spaceP X is a Banach space if and only if theP
absolute
convergence of the series ∞
n=1 kxn k < ∞ implies its convergence ∞
n=1 xn
in X.

Proof. ⇒. This implication is obvious, because the absolute convergence

implies that {xn } is a Cauchy sequence.
⇐. Let {an } ⊂ X be a Cauchy sequence. In order to prove that {an } is con-
vergent to an element in X it suffices to prove that it contains a convergent
subsequence (if a Cauchy sequence has a convergent subsequence, then the
entire sequence is convergent). Let {ank } be a subsequence such that
kank − ank+1 k ≤ 2−k .
Then
∞
X
kank − ank+1 k < ∞
k=1
and by our assumption, the series
X∞
(ank+1 − ank )
k=1
is convergent. That means the sequence whose (k − 1)th element is
(an2 − an1 ) + . . . + (ank − ank−1 ) = ank − an1
 FUNCTIONAL ANALYSIS 23

is convergent, and hence the sequence {ank } is convergence as well. 2

Theorem 4.5. If Y is a closed linear subspace of a Banach space X, then
the quotient space X/Y is a Banach space.

Proof. It suffices to prove that

X∞ ∞
X
k[xn ]k < ∞ ⇒ [xn ] is convergent .
n=1 n=1
P∞
There are elements yn ∈ Y P∞ that n=1 kxn +yn k < ∞, so it follows from
such
Lemma 4.4 that the series n=1 (xn + yn ) converges to an element x0 ∈ X,
i.e.
X k
(xn + yn ) − x0 → 0 as k → ∞ .


n=1
Hence
k
 X  Xk


xn − x0 + yn → 0 as k → ∞


n=1 n=1
| {z }
∈Y
which yields
X k h Xk i


[xn ] − [x0 ] = xn − x0 → 0 as k → ∞


n=1 n=1
and thus
∞
X
[xn ] = [x0 ] .
n=1
The proof is complete. 2
Exercise. Let x = (x1 , x2 , x3 , . . .) ∈ `∞ . Prove that in the quotient space
`∞ /c0
k[x]k = lim sup |xn | .
n→∞

Exercise. Let Y be a closed subspace of c consisting of constant sequences.

Prove that the quotient space c/Y is isomorphic to c0 .
Now we will prove the following quite surprising result.
Theorem 4.6. Every separable Banach space X is isometrically isomorphic
to a quotient space `1 /Y where Y is a closed subspace of `1 .

Proof. Let x1 , x2 , x3 , . . . be a countable and dense subset in the unit sphere

{x ∈ X : kxk = 1}. Let T : `1 → X be defined by
∞
X
T (λ1 , λ2 , . . .) = λn xn .
n=1
24 PIOTR HAJLASZ

Then
∞
X
(4.1) kT λk = k λn xn k ≤ kλk1 ,
n=1
so T is continuous. We will prove that T is a surjection. For every x ∈ X,
every positive integer k and every ε > 0 there is n > k such that

x − kxkxn < ε .
Indeed, it is obvious for x = 0; if x 6= 0 the inequality is equivalent to
x ε
− xn <

kxk kxk

and the existence of xn follows from the density in the unit sphere. Let n1
be such that
ε
kx − λn1 xn1 k < , where λn1 = kxk .
2
Let n2 > n1 be such that
ε ε
k(x − λn1 xn1 ) − λn2 xn2 k < , where λn2 = kx − λn1 xn1 k < .
4 2
Let n3 > n2 be such that
ε ε
k(x−λn1 xn1 −λn2 xn2 )−λn3 xn3 k < , where λn3 = kx−λn1 xn1 −λn2 xn2 k <
8 4
etc. We obtain a sequence λn1 , λn2 , . . . such that
ε
(4.2) λnk+1 = kx − (λn1 xn1 + . . . + λnk xnk )k < k .
2
Let
λx,ε = (λ1 , λ2 , λ3 , . . .)
where we put λi = 0 if i 6∈ {n1 , n2 , n3 , . . .}. Clearly λx,ε ∈ `1 , because
ε ε ε
kλx,ε k1 ≤ kxk + + + + . . . = kxk + ε .
2 4 8
Now continuity of T and (4.2) implies that T (λx,ε ) = x. This proves that T
is a surjection onto X. Let
Y = ker T = {λ ∈ `1 : T (λ) = 0} .
Thus T induces an algebraic isomorphism of linear spaces `1 /Y onto X.
Continuity of T implies that Y is a closed subspace of `1 , so `1 /Y is a
Banach space. Note that if T (λ) = x, then `1 /Y 3 [λ] = {γ : T (γ) = x}.
We will prove that this algebraic isomorphism is actually an isometry. To
this end we have to prove that if T (λ) = x, then k[λ]k = kxk. Let T (λ) = x.
Since T (λx,ε ) = x we have λx,ε ∈ [λ] and hence
k[λ]k ≤ kλx,ε k ≤ kxk + ε
for every ε > 0 and thus
(4.3) k[λ]k ≤ kxk .
 FUNCTIONAL ANALYSIS 25

On the other hand if T (λ) = x, then for every γ ∈ [λ] inequality (4.1) implies
that kxk = kT (γ)k ≤ kγk1 and hence
(4.4) k[λ]k = inf kγk1 ≥ kxk .
γ∈[λ]

The two inequalities (4.3) and (4.4) imply that k[λ]k = kxk. This proves
that T is an isometry of `1 /Y onto X. 2
Definition. For a closed subspace M ⊂ X of a normed space X we define
the annihilator
M ⊥ = {x∗ ∈ X ∗ : hx∗ , yi = 0 for all y ∈ M } .

Clearly M ⊥ is a closed subspace of X ∗ and hence it is a Banach space.2

Theorem 4.7. Let X be a normed space and M ⊂ X a closed subspace.
Then the dual space (X/M )∗ to the quotient space is isometrically isomor-
phic to M ⊥ .

Proof. Let π : X → X/M , π(x) = [x] be the quotient map. We define the
map T : (X/M )∗ → X ∗ by
hT (z ∗ ), xi = hz ∗ , [x]i for z ∗ ∈ (X/M )∗ and x ∈ X.
Since
|hT (z ∗ ), xi| ≤ kz ∗ k k[x]k ≤ kz ∗ k kxk ,
T is bounded and
(4.5) kT (z ∗ )k ≤ kz ∗ k .
We actually have equality in (4.5). To see this we need to prove opposite
inequality. Given ε > 0 let [x] ∈ X/M be such that
k[x]k = 1, |hz ∗ , [x]i| ≥ kz ∗ k − ε .
Since
1 = k[x]k = inf kyk ,
y∈[x]
there is y ∈ [x] such that kyk < 1 + ε and obviously [y] = [x]. Hence
kz ∗ k − ε ≤ |hz ∗ , [x]i| = |hz ∗ , [y]i| = |hT (z ∗ ), yi|
≤ kT (z ∗ )k(1 + ε)
and letting ε → 0 yields kz ∗ k ≤ kT (z ∗ )k which together with (4.5) proves
kT (z ∗ )k = kz ∗ k .
We proved that T is an isometric embedding of (X/M )∗ onto a closed sub-
space of X ∗ . Actually
T ((X/M )∗ ) ⊂ M ⊥ .
2Even if X is only a normed space.
26 PIOTR HAJLASZ

Indeed, for z ∗ ∈ (X/M )∗ , T (z ∗ ) ∈ M ⊥ , because for y ∈ M

hT ∗ (z ∗ ), yi = hz ∗ , [y]i = hz ∗ , 0i = 0 .

It remains to show that T is a surjection onto M ⊥ . Let x∗ ∈ M ⊥ . It is easy

to see that
(4.6) hz ∗ , [y]i = hx∗ , yi for y ∈ X

is a well defined3 and bounded4 functional z ∗ ∈ (X/M )∗ such that T (z ∗ ) =

x∗ , so T is surjective. 2

Theorem 4.8. If L ∈ B(X, Y ) is a bounded mapping between Banach spaces

such that dim(Y /L(X)) < ∞, then L(X) is a closed subspace of Y .

Proof. Let π : Y → Y /L(X) be a quotient mapping.5 Let {z1 , . . . , zn } be

a Hamel basis of Y /L(X) and let y1 , . . . , yn ∈ Y be such that π(yi ) = zi .
The space V = span {y1 , . . . , yn } satisfies
L(X) + V = Y, L(X) ∩ V = {0} .
Note that V is a closed subspace as finitely dimensional, so Y /V is a Banach
space and the quotient mapping
η : Y → Y /V
is bounded. The mapping

L̃ : X/ ker L → Y, L̃([x]) = Lx

is bounded, kL̃k ≤ kLk and it is a bijection onto L̃(X) = L(X). Now the
composition
L̃ η
X/ ker L −→ Y −→ Y /V
is a bijection, so it is an isomorphism of Banach spaces. Hence

L̃ ◦ (η ◦ L̃)−1 : Y /V → L(X)

is an isomorphism6 of normed spaces and thus L(X) is a Banach space by

Proposition 2.6. Therefore L(X) is closed in Y . 2

3If [y ] = [y ], then y − y ∈ M and hence hx∗ , y i = hx∗ , y + (y − y )i = hx∗ , y i

1 2 1 2 1 2 1 2 2
and hence (4.6) is well defined.
4
For ỹ ∈ [y] we have |hz ∗ , [y]i| = |hx∗ , ỹi| ≤ kx∗ kkỹk and taking infimum over ỹ ∈ [y]
yields |hz ∗ , [y]i| ≤ kx∗ k k[y]k which proves boundedness of z ∗ .
5We consider it only as an algebraic mapping. Since we do not know yet that L(X) is
closed, so we do not know that Y /L(X) is a normed space.
6It is an isomorphism, because the inverse is η restricted to L(X). Indeed, η ◦ L̃ ◦ (η ◦
−1
L̃) = id on L(X).
 FUNCTIONAL ANALYSIS 27

4.3. Direct sum.

Definition. If X and Y are normed spaces with the norms k · k1 and
k · k2 respectively, then we define X ⊕ Y as X × Y equipped with the norm
k(x, y)k = kxk1 + kyk2 . The following proposition is obvious.
Proposition 4.9. If X and Y are Banach spaces, then X ⊕ Y is a Banach
space as well.

Example. The space C n [a, b] of real functions whose derivatives of order

up to n are continuous on [a, b] is a Banach space with respect to the norm
kf k = sup |f (x)| + sup |f 0 (x)| + . . . + sup |f (n) (x)| .
x∈[a,b] x∈[a,b] x∈[a,b]

This space is isomorphic to Rn ⊕ C[a, b]. Indeed, the mapping

C n [a, b] 3 f 7→ (f (a), f 0 (a), . . . , f (n−1) (a), f (n) ) ∈ Rn ⊕ C[a, b]
is an isomorphism. This follows from the Taylor formula.
Exercise. Provide a detailed proof of the above fact.
In the case of Hilbert spaces we define a direct sum in a slightly different
way than in the case of Banach spaces.
Definition. If (H1 , h·, ·i1 ), (H2 , h·, ·i2 ) are Hilbert spaces, then their direct
sum is the space H1 ⊕ H2 = H1 × H2 with the inner product
h(x1 , x2 ), (y1 , y2 )i = hx1 , y1 i1 + hx2 , y2 i2 .
It is easy to see that (H1 ⊕ H2 , h·, ·i) is a Hilbert space.
Observe, however, that the corresponding norm is

1/2
k(x1 , x2 )k = kx1 k21 + kx2 k22
which is different, (but equivalent) than the norm defined in the case of a
direct sum of Banach spaces.
We can also define a direct sum of an infinite sequence of Hilbert spaces.
Definition. If (Hi , h·, ·ii ), i = 1, 2, . . . are Hilbert spaces, then the direct
sum is the space
M∞ n ∞
X o
Hi = H1 ⊕ H2 ⊕ . . . = (xi )∞ i=1 : x i ∈ H i , i = 1, 2, . . . , kxi k 2
< ∞
i=1 i=1

equipped with the inner product

∞
X
h(xi ), (yi )i = hxi , yi ii
i=1
28 PIOTR HAJLASZ

and hence
∞
X 1/2
k(xi )k = kxi k2i .
i=1

Theorem 4.10. H1 ⊕ H2 ⊕ . . . is a Hilbert space.

Exercise. Prove the above theorem.

Example. R = `21 is a one dimensional Hilbert space and it is easy to see
that
R ⊕ R ⊕ . . . = `2 .

5. Hilbert spaces

In this section we will develop a basic theory of Hilbert spaces. In partic-

ular we will prove the existence of an orthonormal basis. We will also show
many interesting applications.
Definition. A subset E of a linear space V is convex if
x, y ∈ E, t ∈ [0, 1] ⇒ zt = (1 − t)x + ty ∈ E .

Note that if E is convex, then its translation

E + x = {y + x : y ∈ E}
is also convex.
Definition. Let H be a Hilbert space. We say that the vectors x, y ∈ H
are orthogonal if hx, yi = 0. We denote orthogonal vectors by x ⊥ y.
It is easy to see that if x ⊥ y, then kx + yk2 = kxk2 + kyk2 . This fact is
known as the Pythagorean theorem.
For x ∈ H we define
x⊥ = {y ∈ H : y ⊥ x}
and for a linear subspace M ⊂ H
M ⊥ = {y ∈ H : y ⊥ x for all x ∈ M } .
It is easy to see that x⊥ is a closed subspace of H and hence
\
M⊥ = x⊥
x∈M

is also a closed subspace of H, even if M is not closed.

Theorem 5.1. Every nonempty, convex and closed set E in a Hilbert space
H contains a unique element of smallest norm.
 FUNCTIONAL ANALYSIS 29

Proof. Recall the Parallelogram Law:

kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 for all x, y ∈ H .
Let δ = inf{kxk : x ∈ E}. Applying the Parallelogram Law to x/2 and y/2
we have
1 1 1 x + y 2
kx − yk2 = kxk2 + kyk2 − .

4 2 2 2
Since E is convex and x, y ∈ E we also have (x + y)/2 ∈ E. Accordingly
(5.1) kx − yk2 ≤ 2kxk2 + 2kyk2 − 4δ 2 .
If kxk = kyk = δ, then x = y which implies uniqueness of the element with
the smallest norm. Now let yn ∈ E, kyn k → δ. Then (5.1) yields
kyn − ym k2 ≤ 2kyn k2 + 2kym k2 − 4δ 2 → 0 as n, m → ∞ .
Hence {yn } is a Cauchy sequence in H. Thus yn → x0 ∈ H. Since E is
closed, x0 ∈ E and continuity of the norm yields
kx0 k = lim kyn k = δ.
n→∞

The proof is complete. 2

The next example shows that the property described in the above theorem
is not true in every Banach space. However, it is true in reflexive spaces, see
Theorem 14.10.
Example. For y = (y1 , y2 , . . .) ∈ `∞ define ϕ : `1 → C by
∞
X
ϕ(x) = xn yn where x = (x1 , x2 , . . .).
n=1

Since ϕ ∈ (`1 )∗
W = {x ∈ `1 : ϕ(x) = 1}
is convex and closed. If y = (1, 1, 1, . . .), then
∞
X
W = {x ∈ `1 : xi = 1}
n=1

so for x ∈ W , kxk = ∞
P
n=1 |xn | ≥ 1 and hence every vector x ∈ W with
xn ≥ 0 for all n has the smallest norm equal 1. Thus we have infinitely many
vectors of smallest norm.
On the other hand if y = (1/2, 2/3, 3/4, 4/5, . . .), then
∞
n X n o
W = x ∈ `1 : xn = 1 .
n+1
n=1
30 PIOTR HAJLASZ

In particular en (n + 1)/n ∈ W , where en = (0, . . . , 0, 1, 0, . . .) with 1 on nth

coordinate and ken (n + 1)/nk1 = (n + 1)/n, so dist (0, W ) ≤ 1, but also
dist (0, W ) ≥ 1 because for x ∈ W
∞ ∞ ∞
X X n X n
kxk1 = |xn | > |xn | ≥ xn = 1 .
n+1 n+1
n=1 n=1 n=1
Thus dist (0, W ) = 1. This also proves that there are no vectors in W of
smallest norm because kxk1 > 1 for any x ∈ W . 2
Another example is provided in the next exercise.
Exercise. Let M ⊂ C[0, 1] be a subset consisting of all functions f such
that Z 1/2 Z 1
f (t) dt − f (t) dt = 1 .
0 1/2
Prove that M is a closed convex subset of C[0, 1] that has no element of
minimal norm.
Theorem 5.2. Let M be a closed subspace of Hilbert space H. Then

(a) Every x ∈ H has unique decomposition

x = P x + Qx where P x ∈ M , Qx ∈ M ⊥ .
(b) P x and Qx are nearest points to x in M and M ⊥ respectively.
(c) The mappings P : H → M , Q : H → M ⊥ are linear.
(d) kxk2 = kP xk2 + kQxk2 .

The mappings P and Q are called orthogonal projections of H onto M

and M ⊥ .
Corollary 5.3. If M 6= H is a closed subspace, then there is y ∈ H, y 6= 0,
y ⊥ M.
Corollary 5.4. If M is a closed subspace of a Hilbert space H, then H/M
is isometrically isomorphic to M ⊥ . In particular it is a Hilbert space.
 FUNCTIONAL ANALYSIS 31

Observe that this is a very different situation than in the case of Theo-
rem 4.6.
Proof of Theorem 5.2. (a) Uniqueness is easy. If
x = x0 + y 0 = x00 + y 00 , x0 , x00 ∈ M, y 0 , y 00 ∈ M ⊥ ,
then
M 3 x0 − x00 = y 00 − y 0 ∈ M ⊥ .
Since M ∩ M ⊥ = {0} (because hx, xi = 0 implies x = 0) we conclude that
x0 = x00 and y 0 = y 00 . Thus we are left with the proof of the existence of the
decomposition. The set
x + M = {x + y : y ∈ M }
is convex and closed. Let Qx be the element of the smallest norm in x + M
and let P x = x − Qx. Clearly x = P x + Qx. Since Qx ∈ x + M , it follows
that P x ∈ M . We still need to prove that Qx ∈ M ⊥ . To this end we have to
prove that hQx, yi = 0 for all y ∈ M . We can assume that kyk = 1. Denote
z = Qx. The minimizing property of Qx shows that
hz, zi = kzk2 ≤ kz − αyk2 = hz − αy, z − αyi
for all α ∈ K. Hence
hz, zi ≤ hz, zi + |α|2 hy, yi −αhy, zi − αhz, yi ,
| {z }
1

0 ≤ |α|2 − αhy, zi − αhz, yi .

Taking α = hz, yi we have
0 ≤ −|hz, yi|2 ,
i.e. hz, yi = 0. This proves that z = Qx ∈ M ⊥ .
(b) We know that P x ∈ M . If y ∈ M , then
kx − yk2 = kQx + (P x − y)k2 = kQxk2 + kP x − yk2 .
The last equality follows from Qx ⊥ P x − y. Thus the minimal value of
kx − yk2 is attained when y = P x, so P x is the nearest point to x in M .
Similarly one can prove that Qx is the nearest point to x in M ⊥ .
(c) If we apply (a) to x, y and αx + βy, we obtain
αx + βy = P (αx + βy) + Q(αx + βy) ,
α(P x + Qx) + β(P y + Qy) = P (αx + βy) + Q(αx + βy) ,
αP x + βP y − P (αx + βy) = Q(αx + βy) − αQx − βQy .
| {z } | {z }
∈M ∈M ⊥
Since M ∩ M⊥ = {0} both sides are equal zero which proves linearity both
of P and Q.
(d) This is a direct consequence of the orthogonality of P x and Qx. 2
32 PIOTR HAJLASZ

Example. This will be an illustration for a striking application of Corol-

lary 5.3 in the finitely dimensional case. We will prove that for every integer
m ≥ 0 there is ϕ ∈ C0∞ (B n (0, 1)) such that
Z Z
ϕ(x) dx = 1, ϕ(x)xα dx = 0 for 0 < |α| ≤ m .
Rn Rn
Indeed, let N be the number of multiindices α = (α1 , . . . , αn ) such that
|α| ≤ m. Define a linear map
T : C0∞ (B n (0, 1)) → RN , T (ϕ) = (ξα )|α|≤m
where Z
ξα = ϕ(x)xα dx .
Rn
It suffices to prove that the mapping T is surjective. For if not the image
im T 6= RN is a closed subspace of RN and hence there is a nonzero vector
0 6= (ηα )|α|≤m ⊥ im T , i.e.
X Z Z X
(5.2) 0= ηα ϕ(x)xα dx = ϕ(x) ηα xα dx
|α|≤m Rn Rn |α|≤m

for all ϕ ∈ C0∞ (B n (0, 1)).

In particular if ψ ∈ C0∞ (B n (0, 1)), then
X
ϕ(x) = ψ(x) ηα xα ∈ C0∞ (B n (0, 1))
|α|≤m

and hence (5.2) yields

Z  X 2
0= ψ(x)  ηα xα  dx
 
Rn |α|≤m

C0∞ (B n (0, 1)).

Hence |α|≤m ηα xα = 0 in B n (0, 1) and since
P
for all ψ ∈
it is a polynomial, ηα = 0 for all |α| ≤ m which is a contradiction with
(ηα )|α|≤m 6= 0. 2
Recall that for a function ϕ ∈ C0∞ (Rn ) and ε > 0 we define ϕε (x) =
ε−n ϕ(x/ε).
Exercise. Use the above example to prove that for every integer m > 0
there is ϕ ∈ C0∞ (B n (0, 1)) such that for every ε > 0
ϕε ∗ P = P for all polynomials P of degree ≤ m .

Obviously for every y ∈ H, x 7→ hx, yi is a bounded linear functional on

H. It turns out that every functional on H can be represented in that form.
Theorem 5.5 (Riesz representation theorem). If L is a continuous linear
functional on H, then there is unique y ∈ H such that
Lx = hx, yi for x ∈ H .
 FUNCTIONAL ANALYSIS 33

Proof. If L = 0, then we take y = 0. Otherwise

M = {x : Lx = 0}
is a closed linear subspace with M 6= H. Thus there is z ∈ M ⊥ , kzk = 1
(see Corollary 5.3). For any x ∈ H let
u = (Lx)z − (Lz)x .
We have
Lu = (Lx)(Lz) − (Lz)(Lx) = 0 ,
so u ∈ M and hence hu, zi = 0, i.e.
(Lx) hz, zi −(Lz)hx, zi = 0 ,
| {z }
1

Lx = (Lz)hx, zi = hx, (Lz)z i .

| {z }
y

The uniqueness is easy. If hx, yi = hx, y 0 i

for all x ∈ H, then hx, y − y 0 i = 0
for all x ∈ H and in particular for x = y − y 0
ky − y 0 k2 = hy − y 0 , y − y 0 i = 0, y = y0 .
The proof is complete. 2

5.1. Orthonormal basis. As set {uα }α∈A ⊂ H is called orthonormal if


1 if α = β,
huα , uβ i =
0 if α 6= β.
If {uα }α∈A is orthonormal, then with each x ∈ H we associate Fourier
coefficients defined by
x̂(α) = hx, uα i .
Proposition 5.6. Suppose that {uα }α∈A is an orthonormal set in H and
F is a finite subset of A. Let
MF = span {uα }α∈F .

(a) If ϕ is a complex function on A, equal to zero outside F , then the

vector
X
y= ϕ(α)uα ∈ MF
α∈F

satisfies ŷ(α) = ϕ(α) for α ∈ F and

X
kyk2 = |ϕ(α)|2 .
α∈F
34 PIOTR HAJLASZ

(b) If x ∈ H, and X
sF (x) = x̂(α)uα ,
α∈F
then
(5.3) kx − sF (x)k ≤ kx − sk
for every s ∈ MF with the equality only for s − sF (x). Moreover
X
(5.4) |x̂(α)|2 ≤ kxk2 .
α∈F

Proof. Part (a) is obvious. To prove (b) denote sF = sF (x) and observe
that sˆF (α) = x̂(α) for α ∈ F . That means (x − s + F ) ⊥ MF . In particular
hx − sF , sF − si = 0 and hence
(5.5) kx − sk2 = k(x − sF ) + (sF − s)k2 = kx − sF k2 + ksF − sk2
which implies (5.3). Now (5.5) with s = 0 gives
kxk2 = ks − sF k2 + ksF k2 ≥ ksF k2
which is (5.4). 2
Remark. The part (b) says that sF (x) is the best unique approximation
of x in MF , i.e. (see Theorem 5.2) sF (x) is the orthogonal projection of x
onto MF .
If A is any set, then we define `∞ (A) to be the Banach space of all bounded
functions on A (no measurability condition), and `p (A), 1 ≤ p < ∞ is the
Banach space of p-integrable functions with respect to the counting measure.
Thus ϕ ∈ `p (A), 1 ≤ p < ∞ if ϕ(α) 6= 0 for at most countably many α and
X X
|ϕ(α)|p = |ϕ(α)|p < ∞ .
α∈A α∈A
ϕ(α)6=0

Note that X X
|ϕ(α)|p = sup |ϕ(α)|p
α∈A F α∈F

where the supremum if over all finite subsets F ⊂ A.

If A = Z, then `p (A) = `p . Clearly, the functions ϕ that are equal zero
except on a finite subset of A are dense in `p (A), 1 ≤ p < ∞. It is also
obvious that `2 (A) is a Hilbert space with respect to the inner product
X
hϕ, ψi = ϕ(α)ψ(α) .
α∈A
We will need the following elementary result.
Lemma 5.7. Suppose that

(a) X and Y are metric spaces and X is complete;

 FUNCTIONAL ANALYSIS 35

(b) f : X → Y is continuous;
(c) X has a dense subset X0 on which f is an isometry;
(d) f (X0 ) is dense in Y .

Then f is an isometry of X onto Y .

An important part of the lemma is that f is a surjection. Recall that f

is an isometry if it preserves the distances, i.e.
dY (f (x), f (y)) = dX (x, y) for all x, y ∈ X .
Proof. Clearly f is an isometry on X by continuity. To prove that f is a
surjection let y ∈ Y . Since f (X0 ) is dense in Y , there is a sequence xn ∈ X0
such that dY (y, f (xn )) → 0 as n → ∞. This implies that f (xn ) and hence
xn (by isometry) are Cauchy sequences. Hence xn is convergent, xn → x0
(because X is complete) and thus f (x0 ) = y. 2
Theorem 5.8. Let {uα }α∈A be an orthonormal set in a Hilbert space H,
and let P be the space of all finite linear combinations of the vectors uα , i.e.
P = span {uα }α∈A . Then the Bessel inequality
X
|x̂(α)|2 ≤ kxk2
α∈A

holds for all x ∈ H and x 7→ x̂ is a continuous linear mapping of H onto

`2 (A) whose restriction to the closure P is an isometry of P onto `2 (A).

Proof. The Bessel inequality follows from the fact that

X
|x̂(α)|2 ≤ kxk2
α∈F

for every finite subset F ⊂ A. Let

f : H → `2 (A) , f (x) = x̂ .
Linearity of f is obvious. Since
kf (x) − f (y)k2 = kf (x − y)k2`2 (A) ≤ kx − yk2 ,

f is continuous. Moreover f is an isometry of P onto a dense subset of `2 (A)

consisting of functions in `2 (A) which are equal to zero except on a finite
set. Hence the theorem follows from the lemma. 2
Theorem 5.9. Let {uα }α∈A be an orthonormal set in a Hilbert space H.
Then the following conditions are equivalent.

(a) {uα }α∈A is a maximal orthonormal set in H.

(b) The space P = span {uα }α∈A is dense in H.
36 PIOTR HAJLASZ

(c) The Parseval identity

X
|x̂(α)|2 = kxk2
α∈A
holds for all x ∈ H.
(d) X
x̂(α)ŷ(α) = hx, yi for all x, y ∈ H .
α∈A

⊥
Proof. (a) ⇒ (b). Suppose P is not dense, i.e. P 6= H. Let u ∈ P ,
kuk = 1. Then the set {u} ∪ {uα }α∈A is orthonormal which contradicts
maximality of {uα }α∈A .
(b) ⇒ (c). It follows from the previous theorem
ˆ: P = H → `2 (A)
is an isometry, so the Parseval identity follows.
(c) ⇒ (d) this implication follows from the Polarization identity, Proposi-
tion 1.3.
(d) ⇒ (a). If (a) were false, then there would be 0 6= u ∈ H such that
hu, uα i = 0 for all α ∈ A. Let x = y = u. We have
X
0 < hx, yi = x̂(α) ŷ(α) = 0
|{z} |{z}
α∈A 0 0
which is a contradiction. 2
Definition. Any maximal orthonormal set in H is called an orthonormal
basis.
Therefore the above theorem provides several equivalent conditions for an
orthonormal set to be a basis.
A direct application of the Hausdorff maximality theorem (equivalent to
the axiom of choice) gives
Theorem 5.10. Every orthonormal set is contained in an orthonormal ba-
sis.

Definition. We say that two Hilbert spaces H1 and H2 are isomorphic if

there is a linear bijection Λ : H1 → H2 such that
hΛx, Λyi = hx, yi for all x, y ∈ H .
Corollary 5.11 (Riesz-Fisher). If {uα }α∈A is an orthonormal basis in H,
then the mapping x 7→ x̂ is an isometry of H onto `2 (A).

It is easy to see that orthonormal vectors are linearly independent.

 FUNCTIONAL ANALYSIS 37

Corollary 5.12. If a Hilbert space H has finite dimension n, then H is

isomorphic to `2n . If a Hilbert space H is separable and infinitely dimensional,
then it is isomorphic to `2 .

Proof. If dim H = n, then the maximal orthonormal set consists on n

vectors. If dim H = ∞ and H is separable, then the maximal orthonormal
set is countable, so it can be indexed by integers. 2
Theorem 5.13. If H is a separable Hilbert space and {en }n∈Z is an or-
thonormal basis, then for every x ∈ H
∞
X
x= x̂(n)en
n=−∞

in the sense of convergence in H, i.e.

X
x − x̂(n)en → 0 as k → ∞ .
|n|≤k

P
Proof. As we know kx − |n|≤k x̂(n)en k equals to the distance of x to
the space span{en }|n|≤k (see Theorem 5.6(b)). Since span {en }n∈Z is a dense
subset of H this distance converges to 0 as k → ∞. 2
As a variant of Theorems 5.9 and 5.13 we obtain
Theorem 5.14. Let H1 , H2 , . . . be closed subspaces of a Hilbert space H such
that Hi ⊥ Hj for i 6= j and linear combinations of elements of subspaces
Hi are dense in H. Then H is isometrically isomorphic to the direct sum
of Hilbert spaces H1 ⊕ H2 ⊕ . . .. More precisely, for very x ∈ H there are
unique elements xi ∈ Hi such that
∞
X
x= xi
i=1
in the sense of convergence in H and
∞
X
2
kxk = kxi k2 .
i=1

Observe that if {eji }i is an orthonormal basis in each Hi , then {eji }i,j is

an orthonormal basis in H.
Exercise. Prove Theorem 5.14.
Theorem 5.15. Let µ and ν be positive σ-finite measures on X and Y
respectively. If {ϕi }∞ 2
i=1 is an orthonormal basis in L (µ) and {ψi }i=1 an
∞

orthonormal basis in l2 (ν), then hij (x, y) = ϕi (x)ψj (y), i, j = 1, 2, . . . is an

orthonormal basis in L2 (µ × ν).
38 PIOTR HAJLASZ

Proof. It easily follows from the Fubini theorem that the functions hij
form an orthonormal family in L2 (µ × ν) and it remains to prove that
∞
X
kf kL2 (µ×ν) = |hf, hij i|2 for all f ∈ L2 (µ × ν) ,
i,j=1

see Theorem 5.9(c). If f ∈ L2 (µ × ν), then for ν-a.e. y ∈ Y , x 7→ fy (x) =

f (x, y) ∈ L2 (µ) by the Fubini theorem. Denoting
Z
fy (i) = hfy , ϕi iL2 (µ) =
b f (x, y)ϕi (x) dµ(x)
X
we have
Z ∞
X
2
|f (x, y)| dµ(x) = |fby (i)|2
X i=1
for ν-a.e. y ∈ Y . Note that for i = 1, 2, . . .
y 7→ fby (i) ∈ L2 (ν)
because
Z Z
|fby (i)|2 dν(y) = |hfy , ϕi iL2 (µ) |2 dν(y)
Y Y
Z Z Z
≤ kfy kL2 (µ) kϕi k2L2 (µ) dν(y) =
2
|f (x, y)|2 dµ(x) dν(y) < ∞ .
Y | {z } Y X
1
Hence
Z Z 
kf k2L2 (µ×ν) = 2
|f (x, y)| dµ(x) dν(y)
Y X
∞
Z X ∞ Z
X
= |fby (i)|2 dν(y) = |fby (i)|2 dν(y)
Y i=1 i=1 Y
∞ X
X ∞
= |hfby (i), ψj iL2 (ν) |2
i=1 j=1
X∞ Z Z  2
 
= 
 f (x, y)ϕi (x) dµ(x) ψj (y) dν(y)
i,j=1 Y X

X∞
= |hf, hij i|2 .
i,j=1

the proof is complete. 2

5.2. Gramm-Schmidt orthogonalization. If H is a separable Hilbert

space, then an orthonormal basis can be constructed through the Gramm-
Schmidt orthogonalization that we describe next. Let {f1 , f2 , . . .} be a dense
subset of H. We remove (by induction) the function fn if it is a linear
 FUNCTIONAL ANALYSIS 39

combination of f1 , . . . , fn−1 . Denote the remaining functions by g1 , g2 , . . ..

Now the orthonormal basis can be defined as follows
gn − n−1
P
g1 k=1 hgn , ek iek
e1 = , en = n−1
kg1 k gn − P
k=1 hgn , ek iek

This construction has a clear geometric interpretation that we lave to the
reader.
Example. The Legendre polynomials Ln , n = 0, 1, 2, . . . are defined by
1 dn 2 n


Ln (t) = (t − 1) .
n!2n dtn
Observe that L0 = 1 and Ln is a polynomial of degree n. We will prove that
the functions √
2n + 1
√ Ln (t) n = 0, 1, 2, . . .
2
form an orthonormal basis in L2 [−1, 1] obtained from the functions
1, t, t2 , t3 , . . . through the Gramm-Schmidt orthogonalization. To this end
it suffices to show that

(1) hLn , Lm i = 0 for n 6= m,

(2) hLn , Ln i = 2/(2n + 1) for n = 0, 1, 2, . . .,
(3) hP, Ln i = 0 for all polynomials P of degree less than n,
(4) htn , Ln i > 0 for n = 0, 1, 2, . . .

If n > 1 and P is a polynomial of degree less than n, then

(−1)n 1 dn P (t) 2
Z
hP, Ln i = (t − 1)n dt = 0
n!2n −1 dtn
which proves (3)7. Since Ln is a polynomial of degree n it also proves (1).
Recall that (2n)!! = 2·4·6·. . .·(2n) = n!2n and (2n−1)!! = 1·3·5·. . .·(2n−1).
Since (t2 − 1)n = t2n + . . . we have
dn 1 d2n (2n)!
(t2 − 1)n =


n
Ln (t) = 2n
= (2n − 1)!!
dt (2n)!! dt (2n)!!
and hence
1
dn
Z
1
(t2 − 1)n dt


hLn , Ln i = Ln (t) n
(2n)!! −1 dt
1 1
(−1)n 2 · (2n − 1)!!
Z Z
2 n
= (2n − 1)!! (t − 1) dt = (1 − t2 )n dt
(2n)!! −1 (2n)!! 0
Z π/2
2 · (2n − 1)!! 2
= cos2n+1 u du =
(2n)!! 0 2n + 1
7We have no boundary terms resulting from the integration by parts, because deriva-
tives of (t2 − 1)n of order ≤ n − 1 vanish at ±1.
40 PIOTR HAJLASZ

which proves (2). Finally

Z 1 n Z 1
n 1 n d 2 n n!
(1 − t2 )n dt > 0


ht , Ln i = t (t − 1) dt =
(2n)!! −1 dtn (2n)!! −1
which proves (4). The proof is complete. 2

5.3. Distance to a subspace. We will show now how to compute the

distance of an element of a Hilbert spaces to a finitely dimensional subspace.
Definition. The Gramm determinant of vectors x1 , x2 , . . . , xn in an inner
product space is
G(x1 , x2 , . . . , xn ) = det{hxi , xj i}ni,j=1 .
Lemma 5.16. Let x1 , x2 , . . . , xn be vectors in an inner product space H and
let Hn−1 = span {x1 , x2 , . . . , xn−1 }. Then
G(x1 , . . . , xn ) = G(x1 , . . . , xn−1 ) · dist (xn , Hn−1 )2 .

Proof. Consider the Gramm matrix

 
hx1 , x1 i . . . hx1 , xn−1 i hx1 , xn i
 .. . .. .
.. .. 

 . . 

 hxn−1 , x1 i . . . hxn−1 , xn−1 i hxn−1 , xn i 
hxn , x1 i . . . hxn , xn−1 i hxn , xn i
whose determinant equals G(x1 , . . . , xn ). Let w = λ1 x1 + λ2 x2 = . . . +
λn−1 xn−1 be the orthogonal projection of xn onto Hn−1 . Subtract from the
last row the first (n − 1) rows multiplied by λ1 , . . . , λn−1 respectively. This
operation does not change the determinant and in the last row we have
[hxn − w, x1 i, . . . , hxn − w, xn−1 i, hxn − w, xn i] = [0, . . . , 0, hxn − w, xn i]
because xn − w ⊥ Hn−1 , see Theorem 5.2. Now we subtract from the last
column the first (n − 1) columns multiplied by λ1 , . . . , λn−1 respectively. We
obtain the matrix
 
hx1 , x1 i . . . hx1 , xn−1 i 0
 .. .. .. .. 

 . . . . 

 hxn−1 , x1 i . . . hxn−1 , xn−1 i 0 
0 ... 0 hxn − w, xn i
Hence
G(x1 , . . . xn−1 , xn ) = G(x1 , . . . , xn−1 ) · hxn − w, xn i ,
so the result follows because
hxn − w, xn i = hxn − w, xn − wi + hxn − w, wi = dist (xn , Hn−1 )2 .
| {z }
0
The proof is complete. 2
 FUNCTIONAL ANALYSIS 41

Corollary 5.17. For any vectors x1 , . . . , xn in an inner product space

G(x1 , . . . , xn ) ≥ 0 and G(x1 , . . . , xn ) = 0 if and only if the vectors are
linearly dependent.
Theorem 5.18. Let x1 , . . . , xn be linearly independent vectors in an inner
product space H and let
Hn = span {x1 , . . . , xn } .
Then for any x ∈ H
s
G(x1 , . . . , xn−1 , x)
dist (x, H) = .
G(x1 , . . . , xn−1 )

As another corollary we obtain that if x1 , . . . xn are linearly independent

vectors in an inner product space, then the volume of the parallelpiped
spanned by the vectors x1 , . . . , xn equals
p
V (x1 , . . . , xn ) = G(x1 , . . . , xn ) .

Example. We will prove that if B1 , B2 , . . . , Bk is a family of balls in Rn

and Bi 6= Bj for i 6= j, then
det{|Bi ∩ Bj |}ki,j=1 > 0 .
Indeed, let fi = χBi ∈ L2 (Rn ) for i = 1, 2, . . . The functions f1 , f2 , . . . fk are
linearly independent in L2 (Rn ) and hence
G(f1 , . . . , fk ) = det{|Bi ∩ Bj |}ki,j=1 > 0 .
Observe that the corresponding result is not true is we replace balls by cubes.
Why?
We close the section on Hilbert spaces with two elegant applications.

5.4. Müntz theorem. As a profound application of the formula for the

distance to a subspace we will we will prove
Theorem 5.19 (Müntz). Let 0 < p1 < p2 < . . . , limi→∞ pi = ∞. Then
every continuous function in C[0, 1] can be uniformly approximated by func-
tions of the form
(5.6) λ0 + λ1 tp1 + . . . + λn tpn
if and only if
∞
X 1
(5.7) = ∞.
pi
i=1

Proof. A crucial step in the proof is the following lemma.

42 PIOTR HAJLASZ

Lemma 5.20. Let W = span {tp1 , . . . , tpn } ⊂ L2 [0, 1], where −1/2 < p1 <
. . . < pn . Then for any q ≥ 0
r n
q 1 Y |q − pi |
dist L2 (t , W ) = .
2q + 1 q + pi + 1
i=1

Before we prove the lemma we will show how to complete the proof of the
Müntz theorem. First we will prove an L2 version of the Müntz theorem.
Theorem 5.21. Let −1/2 < p1 < p2 < p3 < . . . , limi→∞ pi = ∞. Then the
functions tp1 , tp2 , tp3 , . . . are linearly dense in L2 [0, 1], i.e.
(5.8) span {tp1 , tp2 , tp3 , . . .} = L2 [0, 1]
if and only if
∞
0 1
X
(5.9) =∞
pi
i=1
P0
where in the sum we omit the term with pi = 0.

Proof. Since polynomials are dense in L2 [0, 1] a necessary and sufficient

condition for (5.8) is that for any integer q ≥ 0 and any ε > 0 there are
coefficients λ1 , λ2 , . . . , λn such that
Z 1 n
q X 2
t − λi tpi  < ε .
0 i=1

If q = pi for some i it is obvious, so we can assume that q 6= pi for any i.

According to Lemma 5.20 the condition is equivalent to
∞
Y |q − pi |
=0
q + pi + 1
i=1
or
Y 1 − q/pi
(5.10) = 0.
pi >q
1 + (1 + q)/pi

Indeed, since pi → ∞ we neglect in the second product only a finite number

of nonzero factors which does not affect divergence to 0.
We will also need the following elementary lemma from Advanced Calcu-
lus.
Q∞
Lemma 5.22. If bi > 0, then the product P∞ n=1 (1 + bn ) converges to a finite
positive limitQif and only if the series n=1 bn converges. If 0 < bn < 1, then
the productP ∞ n=1 (1 − bn ) converges to a finite positive limit if and only if
the series ∞ n=1 bn converges.
 FUNCTIONAL ANALYSIS 43

Proof. Since ex ≥ 1 + x we have

1 + b1 + . . . + bn ≤ (1 + b1 ) . . . (1 + bn ) ≤ eb1 +...+bn
which easily implies the first part of the lemma. If 0 < bn < 1, then
1
(5.11) (1 − b1 ) . . . (1 − bn ) =
(1 + a1 ) . . . (1 + an )
where ai = bi /(1 − bi ). Now the expression onQthe right hand side of (5.11)
converges to a positive limit if and only if ∞ n=1 (1 + an ) converges
P∞ and
due
P∞ to the first part of the lemma it remains
P∞ to observe that n=1 n =
a
n=1 bn /(1 − bn ) converges if and only if n=1 bn converges. 2
It remains to prove that (5.10) is equivalent to (5.9). Assume first that
q > 0. If (5.9) is satisfied, then
Y Y
(1 − q/pi ) = 0, (1 + (1 + q)/pi ) = ∞
pi >q p1 >q

by Lemma 5.22 and hence (5.10) follows. On the other hand if the sum at
(5.9) is finite, then both products
Y Y
(1 − q/pi ), (1 + (1 + q)/pi )
pi >q p1 >q

have finite and positive limits, so (5.10) is not satisfied.

If q = 0, (5.10) reads as
∞
Y 1
=0
1 + pi
i=1

which according to Lemma 5.22 is equivalent to (5.9). The proof of the L2

version of the Müntz theorem is complete. 2
We are ready now to complete the proof of the Müntz theorem in the
continuous case. First observe that the condition (5.7) is necessary. Indeed,
density of the functions (5.6) in C[0, 1] imply density in L2 [0, 1] and hence
necessity of (5.7) follows from Theorem 5.21. It remains to prove that the
condition (5.7) is sufficient.
Since (5.6) contains all constant functions, it remains to prove that for
any integer m ≥ 1, tm can be uniformly approximated by functions of the
form (5.6). Let i0 be such that pi > 1 for i ≥ i0 . Then for any t ∈ [0, 1] we
have
Z 
 Xn   t Xn  
µi mpi−1 tpi  = m  µi τ pi −1 dτ 
m
t − τ m−1 −
  
 0 
i=i0 i=i0
v
uZ 1  n 2
u  m−1 X
(5.12) ≤ m τ − µi τ pi −1  dτ .
t 
0 i=i0
44 PIOTR HAJLASZ

Since ∞
P
i=i0 1/(pi − 1) = ∞ we conclude from Theorem 5.21 that for any
ε > 0 there are numbers µi0 , . . . , µn such that the right hand side of (5.12)
is less than ε and hence for λ0 = λ1 = . . . = λi0 −1 = 0, λi = µi mp−1 i ,
i = i0 , . . . , n we have


sup tm − λ0 + λ1 tp1 + . . . + λn tpn  < ε .

t∈[0,1]

The proof is complete. 2

Proof of Lemma 5.20. According to Theorem 5.18
G(tq , tp1 , . . . , tpn )
(5.13) dist L2 (tq , W )2 = .
G(tp1 , . . . , tpn )
Since
Z 1
1
htα , tβ i = tα+β dt =
0 α+β+1
we have
 n  n
p1 p2 pn 1 1
G(t , t , . . . , t ) = det = det ,
pi + pj + 1 i,j=1 ai + bj i,j=1

where ai = pi , bj = pj + 1 and similarly

 n+1
q p1 pn 1
G(t , t , . . . , t ) = det ,
αi + βj i,j=1

where α1 = q, αi = pi−1 , i = 2, 3, . . . , n + 1, β1 = q + 1, βj = pj−1 + 1,

j = 2, 3, . . . , n + 1.
Thus a crucial step is to compute the Cauchy determinant
 m
1
Dm = det .
ai + bj i,j=1

Observe that computing the determinant using the algebraic definition that
involves permutations and then taking the common denominator gives
Pm (a1 , . . . , am , b1 , . . . , bm )
Dm = Qm .
i,j=1 (ai + bj )

where Pm is a polynomial of degree m2 − m.

Exercise. Prove that if P (x, y) is a polynomial of two variables that van-
ishes when x = y, then P (x, y) = (x − y)Q(x, y), where Q is another poly-
nomial. Generalize this result to polynomials of higher number of variables.
The polynomial Pm vanishes when ai = aj for some i 6= i (because the
matrix has identical two rows, so its determinant equals 0) or when bi = bj
 FUNCTIONAL ANALYSIS 45

for some i 6= j. Therefore

Y Y
Pm = (aj − ai ) (bj − bi ) Q(a1 , . . . , am , b1 , . . . , bm ) .
1≤i<j≤m 1≤i<j≤m
| {z }| {z }
Am Bm

The degree of each of the polynomials Am and Bm equals m(m − 1)/2, so

the degree of the polynomial Am Bm is the same as the degree of Pm and
hence Q is a constant
Pm = γm Am Bm .
We will prove that γm = 1. A direct computation shows that γ1 = 1. Mul-
tiplying the last row of the matrix in the determinant Dm by am , letting
am → ∞ and then letting bm → ∞ we obtain
 1 1 1 
 1 1 . . . a1 +bm−1 a1 +bm 
 a +b 
am Dm = 
 .
. .. .
.. .
.. 
 a. . 
am am

am +b1 . . . am +bm−1
 m 
am +bm
 1 1 
 a +b
 1 1
... a1 +bm−1 0 
 .. .. .. .. 
−→ 
 . . . .  = Dm−1 .
1 1
 am−1 +b1 . . . am−1 +bm−1 0 
 
 1 ... 1 1 
On the other hand the same limiting process gives
Q
−1
am 1≤i<j≤m (aj − ai )(bj − bi )
am γm Dm = Qm
i,j=1 (ai + bj )
Q
1≤i<j≤m−1 (aj − ai )(bj − bi ) −1
−→ Qm−1 = γm−1 Dm−1 ,
i,j=1 (ai + bi )

so γm = γm−1 and by induction γm = 1. We proved that

Q
1≤i<j≤m (aj − ai )(bj − bi )
Dm = Qm .
i,j=1 (ai + bj )

Taking ai = pi , bj = pj + 1 we have
2
Q
p1 pn 1≤i<j≤n (pj − pi )
G(t , . . . , t ) = Qn
i,j=1 (pi + pj + 1)

and similarly
Qn 2
Q 2
q p1 pn i=1 (q − pi ) 1≤i<j≤n (pi − pj )
G(t , t , . . . , t ) = Qn .
+ 1)2 ni,j=1 (pi + pj
Q
(2q + 1) i=1 (q + pi + 1)

The two above formulas together with (5.13) give the result. 2
46 PIOTR HAJLASZ

5.5. Radon-Nikodym-Lebesgue theorem. As an elegant application of

the Riesz representation theorem (Theorem 5.5) we will prove the Radon-
Nikodym-Lebesgue theorem.
Recall that if µ and ν are two positive measures on M, then ν is called
absolutely continuous with respect to ν if
E ∈ M, µ(E) = 0 ⇒ ν(E) = 0
and we write ν  µ. If there are disjoint sets A, B ∈ M such that µ is
concentrated on A and ν is concentrated on B we say that the measures are
mutually singular and we write µ ⊥ ν.
Theorem 5.23 (Radon-Nikodym-Lebesgue). Let µ be a σ-finite measure
on M and ν a finite measure on M. Then

(a) There is unique pair of measures νa , νs on M such that

ν = νa + νs , νa  µ, νs ⊥ µ .
(b) There is a nonnegative function h ∈ L1 (µ) such that
Z
νa (E) = h dµ for all E ∈ M.
E

Proof. Uniqueness of νa and νs is easy and left as an exercise, so we need

to prove the existence νa , νs and h. We will prove all this at the same time.
We need
Lemma 5.24. If µ is a positive σ-finite measure on a σ-algebra M in X,
then there is a function w ∈ L1 (µ) such that 0 < w(x) < 1 for every x ∈ X.

Proof. Since µ is σ-finite it is the union of countably many sets of finite

measure En ∈ M. We may also assume that the sets are pairwise disjoint.
Now the function w(x) = 2−n (1 + µ(En )) if x ∈ En satisfies the claim. 2
Define a new measure dϕ = dν + w dµ. It is a finite measure and for every
measurable nonnegative function f
Z Z Z
f dϕ = f dν + f w dµ .
X X X

The Schwarz inequality for f ∈ L2 (ϕ)

yields
Z  Z
|f | dϕ ≤ kf kL2 (ϕ) ϕ(X)1/2 .
 
 f dν  ≤
 
X X

Accordingly
Z
f 7→ f dν
X
 FUNCTIONAL ANALYSIS 47

is a bounded functional on L2 (ϕ) and hence the Riesz representation theorem

gives the existence of g ∈ L2 (ϕ) such that
Z Z
(5.14) f dν = f g dϕ for all f ∈ L2 (ϕ) .
X X
For f = χE , ϕ(E) > 0 we obtain
Z
ν(E) = g dϕ .
E
Hence Z
1 ν(E)
0≤ g dϕ = ≤ 1.
ϕ(E) E ϕ(E)
Since E can be chosen arbitrarily we conclude that g(x) ∈ [0, 1], ϕ-a.e.
Therefore we may alternate g on a set of ϕ-measure zero so that g(x) ∈ [0, 1]
for all X without affecting (5.14). We can rewrite (5.14) as
Z Z
(5.15) (1 − g)f dν = f gw dµ .
X X
Let
A = {x : 0 ≤ g(x) < 1} B = {x : g(x) = 1}
and define the measures νa and νs by
νa (E) = ν(E ∩ A), νs (E) = ν(E ∩ B)
for E ∈ M. If we take f = χB , (5.15) gives
Z
0= w dµ .
B

Since w(x) > 0 for all x, we conclude that µ(B) = 0 and hence νs ⊥ µ.8
Applying (5.15) to
f = (1 + g + . . . + g n ) χE
we obtain Z Z
(1 − g n+1 ) dν = g(1 + g + . . . + g n )w dµ .
E E
If x ∈ B, then g(x) = 1 and if x ∈ A, then g n+1 (x) decreases monotonically
to zero. Hence letting n → ∞ the left hand side converges to ν(A ∩ E) =
νa (E). On the right hand side the function that we integrate increases to a
measurable function h and the monotone convergence theorem gives
Z
νa (E) = h dµ .
E

Hence νa  µ. Finally if we take E = X we obtain that h ∈ L1 (µ), because

νa (X) < ∞. 2

8Indeed, µ(B) = 0 implies that µ is concentrated on X \ B. Since ν is concentrated

s
on B and B ∩ (X \ B) = ∅ we obtain that µ ⊥ νs .
48 PIOTR HAJLASZ

6. Fourier Series

In this section we will present a basic theory of Fourier series. It is one of

the most important applications of the theory of Hilbert spaces.
We denote the space of continuous, Lp ,. . . etc functions on R with the
period 1 by C(S 1 ), Lp (S 1 ). . . etc. This notation comes from the fact that
functions with period 1 can be identified with functions on the unit circle
through the mapping x 7→ e2πix . The space L2 (S 1 ) is a Hilbert space with
respect to the inner product
Z 1
(6.1) hf, gi = f (x)g(x) dx .
0

The C(S 1 )
and Lp (S 1 )
norms will be denoted by k·k∞ and k·kp respectively,
2 1
but the L (S ) norm will be simply denoted by k · k. As a direct application
of Theorem 5.13 we have.
Theorem 6.1. The functions en (x) = e2πinx , n ∈ Z form an orthonormal
basis in L2 (S 1 ). Hence any f ∈ L2 (S 1 ) can be represented as a series
∞
X
f= fˆ(n) en
n=−∞

where Z 1
fˆ(n) = hf, en i = f (x)e−2πinx dx ,
0
in the sense that the series converges to f in the L2 norm, i.e.
X
f − fˆ(n) en → 0 as k → ∞.

|n|≤k

Moreover the Plancherel identity

∞
X
kf k2 = |fˆ(n)|2
n=−∞

is satisfied for all f ∈ L2 (S 1 ) and the mapping f 7→ {fˆ(n)}n∈Z is an isometry

of Hilbert spaces
∧
: L2 (S 1 ) → `2 (Z) .

Proof. We only need to prove that the family {en }n∈Z is an orthonormal
basis in L2 (S 1 ). It is easy to see that the functions en are orthonormal, so
we are left with the proof that linear combinations of the en ’s are dense in
L2 (S 1 ) (see Theorem 5.9(b)). Since linear combinations of the functions en
are exactly trigonometric polynomials of period 1, they are dense in C(S 1 )
by the Weierstrass theorem, and the claim follows from the density of C(S 1 )
in L2 (S 1 ). 2
 FUNCTIONAL ANALYSIS 49

√ Another orthonormal
√ basis in L2 (S 1 ) is given by the functions 1,
2 cos(2πnx), 2 sin(2πnx), n = 1, 2, 3 . . . and Theorem 6.1 has an ob-
vious counterpart in this case as well. However, it is more convenient to
work with the basis ek and we will always use it in what follows.

As a first step we will derive a useful integral representation for the partial
sum of the Fourier series. Note that the partial sums of the Fourier series
are well defined for f ∈ L1 (S 1 ). We have

X X Z 1
sn (f, x) = fˆ(k)e2πikx = 2πikx
e f (y)e−2πiky dy
|k|≤n |k|≤n 0
Z 1 X
= f (y) e2πik(x−y) dy ,
0 |k|≤n
| {z }
Dn (x−y)

where
X X
Dn (x) = ek (x) = e2πikx .
|k|≤n |k|≤n

To evaluate the sum Dn (x) note that

n
X
Dn (x)eπix = eπi(2k+1)x ,
k=−n

n
X n−1
X
Dn (x)e−πix = eπi(2k−1)x = eπi(2k+1)x
k=−n k=−n−1

and hence

Dn (x) eπix − e−πix = eπi(2n+1)x − e−πi(2n+1)x ,

so
eπi(2n+1)x − e−πi(2n+1)x sin π(2n + 1)x
Dn (x) = πix −πix
= .
e −e sin πx

We have Dn (0) = 2n + 1 and

Z 1 X Z 1
Dn (x) dx = ek (x) dx = 1 .
0 |k|≤n 0
50 PIOTR HAJLASZ

Now
Z 1
sn (f, x) = f (y)Dn (x − y) dy
0
Z 1−x
(y−x=t)
= f (x + t)Dn (−t) dt
−x
Z 1−x
(Dn (−t) = Dn (t))
= f (x + t)Dn (t) dt
−x
t7→f (x+t)Dn (t) Z 1/2
has period 1
= f (x + t)Dn (t) dt .
−1/2

We proved
Proposition 6.2. If f ∈ L1 (S 1 ), then
Z 1/2
sn (f, x) = f (x + y)Dn (y) dy ,
−1/2

where the Dirichlet kernel

sin π(2n + 1)x
Dn (x) =
sin πx
has the properties
Z 1
Dn (0) = 2n + 1, Dn (x) dy = 1 .
0

As a first application we will prove

Theorem 6.3. For any 1 ≤ m < ∞ and f ∈ C m (S 1 ) the partial sum
X
sn = sn (f ) = fˆ(k)ek
|k|≤n

converges uniformly to f as n → ∞. In fact

kf − sn (f )k∞ ≤ C(m)n−m+1/2 kf (m) k .

Proof. For f ∈ C m (S 1 ) we have

Z 1
(m) ∧
(f ) (n) = f (m) (y)e−2πiny dy
0
Z 1
(by parts)
= (2πin)m f (y)e−2πiny dy ,
0
i.e.
(6.2) (f (m) )∧ (n) = (2πin)m fˆ(n) .
 FUNCTIONAL ANALYSIS 51

Hence for n ≤ n0 < ∞ we can estimate

X
ksn − sn0 k∞ ≤ |fˆ(k)|
|k|>n
X
= |(f (m) )∧ (k)|(2πk)−m
|k|>n
(Schwarz)  X   1/2  X 1/2
≤ (f (m) )∧ (k)2 (2πk)−2m .
|k|>n |k|>n

Bessel’s inequality gives

 X   1/2
(f (m) )∧ (k)2 ≤ kf (m) k
|k|>n

and we also have

 X 1/2
(2πk)−2m ≤ C(m)n−m+1/2 ,
|k|>n

because
∞ ∞
n−2m+1
Z
X 1 dx
≤ = .
k 2m n x2m 2m − 1
k=n+1

Accordingly
(6.3) ksn − sn0 k∞ ≤ C(m)n−m+1/2 kf (m) k .
This implies the uniform convergence of the sequence of partial sums sn .
Since sn → f in L2 and f is continuous, sn ⇒ f uniformly. Now letting
n0 → ∞ in (6.3) yields

ksn − f k∞ ≤ C(m)n−m+1/2 kf (m) k .

The proof is complete. 2
The above result allows us to characterize smooth functions on S 1 in
terms of a rapid decay of the Fourier coefficients.
Theorem 6.4. Let f ∈ L2 (S 1 ). Then f ∈ C ∞ (S 1 ) if and only if for every
positive integer m, |n|m fˆ(n) → 0 as |n| → ∞.9

Proof. If f ∈ C ∞ (S 1 ), then according to (6.2)

|nm fˆ(n)| = C(m)|n|−1 (f (m+1) )∧ (n) ≤ C(m)|n|−1 kf (m+1) k1 → 0

 

9More precisely, if f ∈ L2 (S 1 ) and |n|m fˆ(n) → 0 as |n| → ∞ for any m, then there is
f ∈ C ∞ (S 1 ) such that f = f˜ a.e.
˜
52 PIOTR HAJLASZ

as |n| → ∞. To prove the opposite implication observe that for any integer
m ≥ 0, |n|m+2 fˆ(n) → 0 as |n| → 0 and hence |nm fˆ(n)| ≤ C(m)|n|−2 for all
n. Thus the series of term by term derivatives
∞
X ∞
X
fˆ(n)e(m)
n = (2πi)m nm fˆ(n)e2πinx
n=−∞ n=−∞

converges uniformly by the M -test. That, however, implies that the Fourier
series of f defines a C ∞ function
∞
X
(6.4) f˜ = fˆ(n)en
n=−∞

with
∞
X
f˜(m) = fˆ(n)e(m)
n
n=−∞

for any m. Since the Fourier series at (6.4) converges to f in L2 (S 1 ), we

conclude that f = f˜ a.e. 2
The following result is surprisingly difficult and we will not prove it.
Theorem 6.5 (Carleson). If f ∈ L2 (S 1 ), then sn (f ) → f a.e.

In particular if f ∈ C(S 1 ), then sn (f ) → f a.e. and it is natural to

expect the everywhere convergence. Unfortunately it is not always true, see
Section 9.4.
Theorem 6.6 (Fejer). If f ∈ C(S 1 ), then
s0 + s1 + . . . + sn−1
⇒f
n
uniformly on S 1 .

Proof. We have
Z 1/2
s0 + s1 + . . . + sn−1 D0 (y) + . . . + Dn−1 (y)
= f (x + y) dy .
n −1/2 | n
{z }
Fn (y)

The function Fn is called Fejer kernel. Clearly

Z 1/2 n−1 Z
1 X 1/2
Fn = Dk = 1
−1/2 n −1/2 k=0

and hence
Z 1/2
s0 + s1 + . . . + sn−1
−f = (f (x + y) − f (x))Fn (y) dy .
n −1/2
 FUNCTIONAL ANALYSIS 53

It suffices to prove that

Z 1/2
I(x) := |f (x + y) − f (x)||Fn (y)| dy ⇒ 0 as n → 0.
−1/2
Note that
n−1
1 X sin π(2k + 1)x 1  sin nπx 2
Fn (x) = = ≥ 0.
n sin πx n sin πx
k=0
Indeed, by the formula for the sum of the geometric sequence
n−1 n−1
X X (1 − eπxni )(1 + eπxni )
eπ(2k+1)xi = eπxi e2kπxi =
−2i sin πx
k=0 k=0
and hence
n−1 n−1
X X sin2 πxn
sin π(2k + 1)x = im eπ(2k+1)xi = .
sin πx
k=0 k=0
For a small 0 < δ < 1/2 we split the integral
Z Z
I(x) = + := I1 (x) + I2 (x) .
|y|<δ |y|≥δ
We have
Z δ
I1 (x) = |f (x + y) − f (x)|Fn (y) dy
−δ
Z
≤ max max |f (x + y) − f (x)| Fn (y) dy
|x|≤1/2 |y|≤δ |y|<δ
| {z }
<1
≤ max kfy − f k∞ ,
|y|≤δ

where fy (x) = f (x + y). For every ε > 0 we can find 0 < δ < 1/2 such that
I1 (x) < ε/2 for all x by uniform continuity of f . Now
Z Z 1/2 
4 sin nπy 2
I2 (x) = ≤ kf k∞ dy
|y|≥δ n δ sin πy
2
≤ (sin πδ)−2 kf k∞ ,
n
and to given δ we can find n0 so large that I2 (x) < ε/2 for n > n0 and all
x ∈ [0, 1]. The proof is complete 2
Fourier coefficients can be defined for f ∈ L1 (S 1 ) by the same integral
formula. Clearly
∧
: L1 (S 1 ) → `∞ (Z)
because
Z 1 
|fˆ(n)| =  f (x)e−2πinx dx ≤ kf k1 .
 
0
54 PIOTR HAJLASZ

The following result is a variant of the Fejer theorem.

Theorem 6.7. If f ∈ L1 (S 1 ), then
s + s + . . . + s
0 1 n−1
(6.5) − f → 0 as n → ∞.

n

1

In particular the map ∧ : L1 (S 1 ) → `∞ (Z) is one-to-one.

Proof. As in the proof of the Fejer theorem we have

Z 1/2
s0 + s1 + . . . + sn−1
−f = (f (x + y) − f (x))Fn (y) dy .
n −1/2

Thus
s + s + . . . + s Z 1/2
0 1 n−1
− f ≤ kfy − f k1 Fn (y) dy := I .

n

1 −1/2

Recall that by a well known fact from Analysis I, kfy − f k1 → 0 as |y| → 0

for any f ∈ L1 (S 1 ). We write
Z Z
I= + := I1 + I2 .
|y|<δ |y|≥δ

For ε > 0 we choose δ > 0 such that kfy − f k1 < ε/2 whenever |y| < δ.
Hence for |y| < δ we can estimate I1 by
Z
ε ε
I1 ≤ Fn (y) dy < .
2 |y|<δ 2
For the second integral note that kfy − f k1 ≤ 2kf k1 and hence
Z
I2 ≤ 2kf k1 Fn (y) dy
|y|≥δ
Z 1/2
1  sin nπy 2
= 4kf k1 dy
δ n sin πy
2
≤ kf k1 (sin πδ)−2
n
which is less than ε/2 provided n is sufficiently large. Now it is easy to see
that the convergence (6.5) implies that the mapping ∧ : L1 (S 1 ) → `∞ (Z) is
one-to-one. The proof is complete. 2
Recall that c0 is a closed subspace of `∞ (Z) consisting of sequences con-
vergent to 0 at ±∞.
Theorem 6.8 (The Riemann-Lebesgue lemma).
∧
: L1 (S 1 ) → c0 ,
i.e. |fˆ(n)| → 0 as |n| → ∞.
 FUNCTIONAL ANALYSIS 55

First proof. Since eπi = −1 we have

Z 1 Z 1
−2πinx
ˆ
f (n) = f (x)e dx = − f (x)e−2πinx eπi dx
0 0
Z 1 Z 1 
1 1  −2πinx
= − f (x)e−2πin(x− 2n ) dx = − f x+ e dx .
0 0 2n
Hence Z 1
1  1  −2πinx
fˆ(n) = f (x) − f x + e dx ,
2 0 2n
so
1
|fˆ(n)| ≤ kf − f 1 k1 → 0
2 2n

as |n| → ∞. 2
Second proof. If f ∈ C 1 (S 1 ), then
(f 0 )∧ = 2πinfˆ(n)
and hence
fˆ(n) = (2πin)−1 (f 0 )∧ (n) .
Since k(f 0 )∧ k∞ < ∞ it follows that |fˆ(n)| → 0 as |n| → ∞, i.e.
∧
: C 1 (S 1 ) → c0 .
Since10 kfˆk∞ ≤ kf k1 the claim follows from the density of C 1 (S 1 ) in L1 (S 1 ).
2
As mentioned above (cf. Section 9.4), in general, for f ∈ C(S 1 ) the Fourier
series need not converge to f , however, the following results provides suffi-
cient conditions for the convergence.
Theorem 6.9 (Dini’s criterion). If f ∈ L1 (S 1 ) and for some x with |x| ≤
1/2
Z 1/2 
 f (x + y) − f (x) 

 dy < ∞ ,
y

−1/2
then
lim sn (f, x) = f (x) .
n→∞

Theorem 6.10 (Jordan’s criterion). If f is a function of bounded variation

in a neighborhood of x, then
1
lim sn (f, x) = (f (x+ ) + f (x− ))
n→∞ 2
where f (x± ) denote the right and left hand side limits of f at x.

10k · k is the norm in the Banach space c0 .

∞
56 PIOTR HAJLASZ

We will not prove these theorems.

The following result tells us that the rate of convergence of sn (f ) near

x0 ∈ S 1 depends only on properties of f in a neighborhood of x0 .

Theorem 6.11 (Riemann’s localization principle). If f ∈ L1 (S 1 ) vanishes

in a neighborhood of x0 , then sn (f ) ⇒ 0 uniformly in a (perhaps smaller)
neighborhood of x0 .

Remark. If f = g in a neighborhood of x0 , then

sn (f ) − sn (g) = sn (f − g) ⇒ 0

near x0 and hence sn (f ) is uniformly close to sn (g) in a neighborhood of

x0 . In particular if sn (g) ⇒ g near x0 , then also sn (f ) ⇒ f near x0 and
as an application we obtain the following result: if f ∈ L1 (S 1 ) is C 1 in a
neighborhood of x0 , then sn (f ) ⇒ f in a neighborhood of x0 .

Proof. We can assume that x0 = 0. Suppose that f (x) = 0 for |x| ≤ δ.

Then f (x + y) = 0 for |x| ≤ δ/2 and |y| ≤ δ/2. Note that if |x| ≤ δ/2, the
function
f (x + y)
y 7→
sin πy

vanishes for |y| ≤ δ/2 and hence has no singularity at y = 0. Therefore the
function is integrable on S 1 and for |x| ≤ δ/2 we have

Z 1/2
sin(2n + 1)πy
sn (f, x) = f (x + y) dy
−1/2 sin πy
f (x + y) eπiy e2nπiy − e−πiy e−2nπiy
1/2
Z
= dy
−1/2 sin πy 2i
1
(Q+ )∧ (−n) − (Q− )∧ (n) ,


=
2i

where
f (x + y) ±πiy
Q± (y) = e .
sin πy

Now it follows from the Riemann-Lebesgue lemma that sn (f, x) → 0 as

n → ∞. We still need to prove that the convergence is uniform with respect
to x ∈ [−δ/2, δ/2].
 FUNCTIONAL ANALYSIS 57

For x1 , x2 ∈ [−δ/2, δ/2] we have

Z 1/2  sin(2n + 1)πy 
|sn (f, x1 ) − sn (f, x2 )| ≤ |f (x1 + y) − f (x2 + y)|   dy
 
−1/2 sin πy
Z  sin(2n + 1)πy 
= |f (x1 + y) − f (x2 + y)|   dy
 
δ/2≤|y|≤1 sin πy
πδ
−1
≤ sin kfx1 − fx2 k1
2
πδ
−1
= sin kfz − f k1 ,
2
where z = x2 −x1 . This implies that the family {sn (f )}∞ n=1 is equicontinuous
on the interval [−δ/2, δ/2]. Since it converges at every point of the interval
equicontinuity implies uniform convergence. 2

6.1. Convolution. L1 (S 1 ) is an algebra with respect to the convolution

Z 1
f ∗ g(x) = f (x − y)g(y) dy, for f, g ∈ L1 (S 1 ).
0
Although the function y 7→ f (x − y)g(y) need not be integrable for every x
and hence f ∗ g(x) is not defined for every x, we will prove11 that f ∗ g ∈
L1 (S 1 ) and
kf ∗ gk1 ≤ kf k1 kgk1 .
Obviously, if
Z 1
(6.6) |f (x − y)g(y)| dy < ∞ ,
0
then f ∗ g(x) is finite for such x. Fubini’s theorem yields
Z 1Z 1  Z 1 Z 1
|f (x − y)g(y)| dy dx = |f (x − y)| dx |g(y)| dy = kf k1 kgk1 .
0 0 0 0
Hence (6.6) holds for a.e. x. Moreover
Z 1Z 1  Z 1Z 1
kf ∗ gk1 = f (x − u)g(y) dy  dx ≤ |f (x − y)g(y)| dy dx
 

0 0 0 0
= kf k1 kgk1 .
Note that with the convolution notation we have
Z 1
sn (f, x) = f (y)Dn (x − y) dy = Dn ∗ f (x)
0
and similarly
s0 + s1 + . . . sn−1
= Fn ∗ f (x) .
n
Theorem 6.12. For f, g ∈ L1 (S 1 ) we have
11Just like in Analysis I.
58 PIOTR HAJLASZ

(a) f ∗ g(x) = g ∗ f (x)

(b) (f ∗ g) ∗ h(x) = f ∗ (g ∗ h)(x)

We leave the proof as an easy exercise. 2

Theorem 6.13. For f, g ∈ L1 (S 1 )
(f ∗ g)∧ (n) = fˆ(n)ĝ(n)

Proof. We have
Z 1Z 1 
∧
(f ∗ g) (n) = f (x − y)g(y) dy e−2πinx dx
0 0
Z 1Z 1 
= f (x − y)e−2πin(x−y) dy g(y)e−2πiny dy
0 0
= fˆ(n)ĝ(n)
The proof is complete. 2
Now we will show several applications of the Furier series.

6.2. Riemann zeta funciton.

Theorem 6.14.
∞
X 1 π2
ζ(2) = = .
n2 6
n=1

Proof. Let f (x) = x for x ∈ [0, 1). Then f ∈ L2 (S 1 ) and

1

Z 1  2 for n = 0,
fˆ(n) = xe−2πinx dx =
0 1
− 2πin for n 6= 0.


Hence the Plancherel identity yields

Z 1 ∞
1 2 2 ˆ 2 1 X  1 2 1 1 X 1
= x dx = kf k2 = kf k2 = + = + 2
3 0 4 2πn 4 2π n2
n6=0 n=1

and the claim easily follows. 2

6.3. Writinger’s inequality.

Theorem 6.15 (Writinger’s inequality). If f ∈ C 1 ([a, b]), f (a) = f (b) = 0,
then Z b
(b − a)2 b 0 2
Z
2
|f | ≤ |f |
a π2 a
and the constant (b − a)2 /π 2 cannot be replaced by any smaller number.
 FUNCTIONAL ANALYSIS 59

Proof. It is enough to prove the theorem for a = 0, b = 1/2 as the general

case follows from a linear change of variables, i.e. it suffices to prove the
inequality
Z 1/2 Z 1/2
1
2
|f | ≤ 2 |f 0 |2 .
0 4π 0
The function f can be extended from [0, 1/2] to [−1/2, 1/2] as an odd func-
tion in C 1 (S 1 ). By the oddness of f we have
Z 1/2
ˆ
f (0) = f =0
−1/2
and hence the Planchelel identity yields
Z 1/2 ∞
X ∞
X
0 2 0 ∧
|f | = 2
|(f ) (n)| = |2πinfˆ(n)|2
−1/2 n=−∞ n=−∞
X Z 1/2
≥ 4π 2
|fˆ(n)|2 = 4π 2 |f |2 .
n6=0 −1/2

The equality is achieved for the function f (x) = 2i sin 2πx = e2πix − e−2πix .
2

6.4. The isoperimetric problem.

Theorem 6.16 (Isoperimetric theorem). Among all Jordan curves of fixed
length, the one that encloses the largest area is the circle. All other curves
enclose smaller area.

Proof. We can assume that the fixed length of the Jordan curve is 1.
Denoting the enclosed area by A we can write the theorem in the form of
the inequality
1
(6.7) A≤
4π
with the equality if and only if the curve is a circle. We will assume that the
Jordan curve γ(t) = (x(t), y(t)) is of class C 1 (S 1 ), i.e. γ is

closed:
x(0) = x(1), y(0) = y(1);
smooth:
x, y ∈ C 1 (S 1 ), γ(t) 6= 0 for all t;
Jordan:
γ(t1 ) 6= γ(t2 ) for t1 6= t2 , t1 , t2 ∈ [0, 1);
unit length:
Z 1
1/2
`= ẋ(t)2 + ẏ(t)2 dt = 1 .
0
60 PIOTR HAJLASZ

Since every rectifiable curve (i.e. of finite length) admits an arc-length

parametrization (i.e. of speed 1) we can assume that
ẋ(t)2 + ẏ(t)2 = 1 for all t.
As a consequence of the Plancherel theorem we have
Z 1 ∞
X
1 = 2 2
|ẋ(t)| + |ẏ(t)| dt = ˆ
|ẋ(n)|2 ˆ
+ |ẏ(n)|2
0 n=−∞
∞
X
= 4π 2 n2 (|x̂(n)|2 + |ŷ(n)|2 )
n=−∞
and hence
∞
X
n2 (|x̂(n)|2 + |ŷ(n)|2 ) = (4π 2 )−1 .
n=−∞
It easily follow from the Green theorem that
Z Z 1 Z 1
ẏ∈R
A = x dy = x(t)ẏ(t) dt = x(t)ẏ(t) dt
γ 0 0
∞ ∞
A∈R
X X
= ˆ
x̂(n)ẏ(n) = re ˆ
x̂(n)ẏ(n)
n=−∞ n=−∞
X∞
= π n · 2 re x̂(n)iŷ(n) .
n=−∞
Hence
∞
X
n · 2 re x̂(n)iŷ(n) = Aπ −1 .
n=−∞
Therefore
∞
X ∞
X
(4π 2 )−1 − Aπ −1 = n2 (|x̂(n)|2 + |ŷ(n)|2 ) − n · 2 re x̂(n)iŷ(n)
n=−∞ n=−∞
∞
X
= (n2 − |n|)(|x̂(n)|2 + |ŷ(n)|2 )
n=−∞
∞
X
+ |n|(|x̂(n)|2 − sgn (n)2 re x̂(n)iŷ(n) + |ŷ(n)|2 )
n=−∞
X∞ ∞
X
(6.8) = (n2 − |n|)(|x̂(n)|2 + |ŷ(n)|2 ) + |n| |x̂(n) − sgn (n)iŷ(n)|2
n=−∞ n=−∞
because
|a ± b|2 = |a|2 ± 2 re ab + |b|2 .
Since the right hand side of (6.8) is nonnegative we conclude that
1
(4π 2 )−1 − Aπ −1 ≥ 0, so A ≤
4π
 FUNCTIONAL ANALYSIS 61

which is the isoperimetric inequality (6.7). The equality in the isoperimetric

inequality holds if and only if the right hand side of (6.8) equals 0, so
(6.9) x̂(n) = ŷ(n) = 0 for |n| ≥ 2
and
(6.10) |x̂(n) − sgn (n)iŷ(n)|2 = 0 for |n| = 1.
Now (6.9) yields
x(t) = x̂(−1)e−2πit + x̂(0) + x̂(1)e2πit ,
y(t) = ŷ(−1)e−2πit + ŷ(0) + ŷ(1)e2πit .
Hence
(x(t) − x̂(0))2 + (y(t) − ŷ(0))2 =
(x̂(−1)2 + ŷ(−1)2 )e−4πit + 2x̂(−1)x̂(1) + 2ŷ(−1)ŷ(1) + (x̂(1)2 + ŷ(1)2 )e4πit .
| {z } | {z }
0 by (6.10) 0 by (6.10)

Therefore
(x(t) − x̂(0))2 + (y(t) − ŷ(0))2 = const.
The proof is complete. 2

6.5. Equidistribution of arithmetic sequences. For a number 0 < γ <

1 and x ∈ [0, 1) define12
xn = x + nγ − [x + nγ] ∈ [0, 1) .
If we identify [0, 1) with S 1 via the exponential mapping
[0, 1) 3 t 7→ e2πit ∈ S 1 ,
then xn is identified with a point e2πi(x+nγ) on the circle and hence x 7→ xn
is a rotation of S 1 by the angle 2πnγ.
Theorem 6.17 (Weyl). If 0 < γ < 1 is irrational 13 and f ∈ C(S 1 ), then
n−1 Z 1
1X
(6.11) lim f (xk ) = f (x) dx .
n→∞ n 0
k=0

Proof. First we will prove the theorem in the case in which

f (x) = em (x) = e2πimx for some m ∈ Z.
If m = 0, then both sides of (6.11) are equal 1. If m 6= 0, then
Z 1
em (x) dx = 0
0

12[x] stands for the largest integer ≤ x.

13The result is not true if γ is rational (exercise).
62 PIOTR HAJLASZ

and
n−1 n−1
1X 1 X 2πim(x+kγ)
em (xk ) = e
n n
k=0 k=0
1 e2πimx (1 − e2πimγn )
( geom.
sum
)
= → 0 as n → ∞ .
n 1 − e2πimγ
(Note that sice γ is irrational the denominator is nonzero.) Thus (6.11) is
satisfied. Next (6.11) is satisfied by trigonometric polynomials which are
finite sums of the form
X
(6.12) f (x) = am em (x)
|m|≤k

For a general f ∈ C(S 1 )

and ε > 0 we can find a trigonometric polynomial
fε of the form (6.12) such that14
kfε − f k∞ < ε/3 .
Since the function fε satisfies (6.11) for sufficiently large n we have
 1 n−1 Z 1   n−1 Z 1 
 X  1 X
f (x ) − f ≤ f (x ) − fε 

k ε k
n n
  
k=0 0 k=0 0

 1 n−1   1
Z 
 X
+ (fε (xk ) − f (xk ))  +  (fε − f ) 
  
n
| k=0 {z } | 0 {z }
<ε/3
<ε/3
 1 n−1 Z 1 
 X  2ε
≤ fε (xk ) − fε  + < ε.
n 0 3
k=0
The proof is complete. 2
Corollary 6.18. If 0 < γ < 1 is irrational and 0 ≤ a < b ≤ 1, then
#{k < n : a ≤ xk ≤ b}
lim = b − a.
n→∞ n

Proof. Let f ± be continuous functions that approximate the characteristic

function of the interval [a, b] is the following sense15
f − ≤ χ[a,b] ≤ f + ,
Z 1 Z 1
ε ε
b−a− < f− ≤ f+ < b − a + .
2 0 0 2
Since
n−1
#{k < n : a ≤ xk ≤ b} 1X
= χ[a,b] (xk )
n n
k=0

14By the Stone-Weierstrass theorem or simply by the Fejer theorem.

15A picture will help to see that such functions exist.
 FUNCTIONAL ANALYSIS 63

we conclude from the previous theorem that

Z 1
ε
(b − a) − ε < f − (x) dx −
0 2
n−1
large n 1 X
< f − (xk )
n
k=0
#{k < n : a ≤ xk ≤ b}
≤
n
n−1
1 X
≤ f + (xk )
n
k=0
Z 1
large n ε
< f + (x) dx + < (b − a) + ε .
0 2
Hence
 #{k < n : a ≤ x ≤ b} 
k
− (b − a) < ε
 
n

for all sufficiently large n. 2
Exercise. The sequence 2n starts with 2, 4, 8, 16, 32, 64, 128, 256, . . .

(a) Investigate how many times 7 and 8 will appear as a first digit of the
decimal representation of 2n for n ≤ 45.
(b) The digit c appears as a first digit of 2n with the frequency
#{k < n : 2k = c . . .}
lim .
n→∞ n
Prove that this limit exists for any c ∈ {1, 2, . . . , 9} and show that 7
appears more often than 8.

7. Spherical harmonics

The theory of Fourier series describes functions defined on S 1 . It is very

tempting to develop an analog of the theory of Fourier series for functions
defined on the sphere S n−1 in Rn , and this is what we will do now.
√ √
Trigonometric functions 1, 2 cos(2πnt), 2 sin(2πnt), n = 1, 2, . . . form
an orthonormal basis in L2 (S 1 ). Our purpose now is to find an orthonormal
basis in L2 (S n−1 ), the space of square integrable functions on a sphere in
Rn . The trigonometric functions listed above are defined on [0, 1] and they
correspond to functions on S 1 through the parametrization [0, 1] 3 t 7→
e2πit ∈ S 1 . The corresponding inner product in S 1 is
Z
1
hf, gi = f g dσ
2π S 1
64 PIOTR HAJLASZ

where we integrate with respect to the length element on S 1 . The factor

1/(2π) comes from the fact that in the definition (6.1) of the inner product
we integrate over the interval of length 1, but the length of the circle is 2π.
However, until the end of the section we will consider another inner product
in L2 (S 1 )
Z
hf, gi = f g dσ
S1
√
and hence the orthonormal basis in L2 (S 1 ) corresponds to functions 1/ 2π,
π −1/2 cos(2πnt), π −1/2 sin(2πnt), n = 1, 2, . . . through the parametrization
[0, 1] 3 t 7→ e2πt ∈ S 1 .
More precisely if z = e2πit ∈ S 1 , then
1 1 i
√ z n = √ cos 2πnt + √ sin 2πnt .
π π π
Hence the orthonormal basis on S 1 consists of functions (2π)−1/2 , π −1/2 re z n ,
π −1/2 im z n .
The real and imaginary parts of z n are polynomials in two variables (x, y)
that are harmonic. Moreover they are homogeneous of degree n, i.e. they are
of the form
X n
P (x, y) = ak xk y n−k .
k=0
Therefore it should not come as a surprise that we will search for an or-
thonormal basis in L2 (S n−1 ) among homogeneous harmonic polynomials.
We equip the space L2 (S n−1 ) with the inner product
Z
hf, gi = f (x)g(x) dσ(x) .
S n−1

Let Pk be the linear space of homogeneous polynomials P in Rn of degree

k, i.e.
X
P (x) = aα xα
|α|=k

where we use the multiindex notation.

Let Hk be the linear subspace of Pk consisting of harmonic polynomials,
i.e. ∆P = 0. The elements of Hk are called solid spherical harmonics of
degree k.
By Hk we denote the liner space of restrictions of solid spherical har-
monics of degree k to S n−1 , so Hk is a finitely dimensional linear subspace
of L2 (S n−1 ). The elements of Hk are called surface spherical harmonics of
degree k.
 FUNCTIONAL ANALYSIS 65

Lemma 7.1. The finitely dimensional subspaces {Hk }∞ 2

k=0 of L (S
n−1 ) are

mutually orthogonal.

Proof. Let P̃ ∈ Hk and Q̃ ∈ Hl , k 6= l be restictions of P ∈ Hk and

Q ∈ Hl to the unit sphere. Since P (rx) = rk P (x) it easily follows that
∂P d
= |r=1 P (rx) = krk−1 P (x)|r=1 = kP (x) ,
∂ν dr
where ν is the outward normal to S n−1 . Similarly ∂Q/∂ν = lQ(x). Hence
Green’s formula yields
Z Z  ∂P ∂Q 
(k − l)hP̃ , Q̃i = (k − l) P Q dσ = Q −P
S n−1 S n−1 ∂ν ∂ν
Z
= (Q∆P − P ∆Q) dx = 0
|x|≤1
by harmonicity of P and Q. 2
Theorem 7.2. Every polynomial P ∈ Pk can be uniquely represented as
P (x) = P0 (x) + |x|2 P1 (x) + . . . + |x|2l Pl (x)
where Pj ∈ Hk−2j , j = 0, 1, 2, . . . , l.

Proof. It suffices to prove that every P ∈ Pk can be uniquely represented

as
(7.1) P = Q1 + |x|2 Q2 , Q1 ∈ Hk , Q2 ∈ Pk−2 .
The result will follow then from a routine inductive argument. With each
polynomial X
P (x) = aα xα ∈ Pk
|α|=k
we associate a differential operator
∂  X ∂ |α|
P = aα α
∂x ∂x
|α|=k

and we define a new inner product in Pk as follows

∂  X ∂ |α|  X  X
(7.2) hP, Qi = P Q= aα α bβ xβ = aα bα α!
∂x ∂x
|α|=k |β|=k |α|=k

where the last equality follows from the fact that

∂ |α| xβ ∂ |α| xα
= 0 if α 6
= β , = α!
∂xα ∂xα
Note that the right hand side of (7.2) proves that this is a true inner product.
Let
|x|2 Pk−2 = |x|2 P (x) : P ∈ Pk−2 ⊂ Pk .

66 PIOTR HAJLASZ

It suffices to prove that the orthogonal complement of the subspace |x|2 Pk−2
is Hk , i.e.
(7.3) Pk = Hk ⊕ |x|2 Pk−2
because it readily implies unique representation (7.1). To prove (7.3) observe
that P1 ∈ Pk is in the orthogonal complement of |x|2 Pk−2 if and only if for
every P2 ∈ Pk−2
h|x|2 P2 , P1 i = 0 .
We have
 ∂   ∂ 
0 = h|x|2 P2 , P1 i = ∆ P2 P1 = P2 ∆P1 = hP2 , ∆P1 i
∂x ∂x
for every P2 ∈ Pk−2 , where in the right hand side we have the inner product
in Pk−2 . This is, however, true if and only if ∆P1 = 0, i.e. P1 ∈ Hk . 2
Corollary 7.3. The restriction of any polynomial to the unit sphere S n−1
is a finite sum of surface spherical harmonics.

Proof. Any polynomial is a sum of homogeneous polynomials of possibly

different degrees. If P is homogeneous, then the representation from Theo-
rem 7.2 gives
P (x) = P0 (x) + P1 (x) + . . . + Pl (x), for x ∈ S n−1 ,
where Pi are surface spherical harmonics. 2
By the Stone-Weierstrass theorem restrictions of polynomials to S n−1 are
dense in C(S n−1 ) and hence are dense in L2 (S n−1 ). Therefore linear com-
binations of surface spherical harmonics are dense in L2 (S n−1 ), see Corol-
lary 7.3. Now Theorem 5.14 and Lemma 7.1 give
Theorem 7.4. For every f ∈ L2 (S n−1 ) there are unique elements Yk ∈ Hk ,
k = 0, 1, 2, . . . such that
X∞
f= Yk
k=1
in the sense of convergence in L2 (S n−1 ). Moreover
∞
X
2
kf k = kYk k2 .
k=1
Accordingly,
L2 (S n−1 ) = H0 ⊕ H1 ⊕ H2 ⊕ . . .

In particular if {Y1k , . . . , Yakk } is an orthonormal basis in Hk , then {Yik }i,k

is an orthonormal basis in L2 (S n−1 ) consisting of surface spherical harmon-
ics.
On every Riemannian manifold M , in particular on every smooth sub-
manifold of the Euclidean space, one can define a natural Laplace operator
 FUNCTIONAL ANALYSIS 67

∆M called the Lapalce-Beltrami operator. The general construction is quite

involved, but in the case of the sphere it can be easily done as follows.
Definition. If f ∈ C 2 (S n−1 ), then we extend f to a neighborhood of S n−1
as a 0-homogeneous function, i.e. f˜(x) = f (x/|x|) and we define the Laplace-
Beltrami operator on S n−1 (spherical Lapalcean) as follows
∆S f = ∆f˜|S n−1 .

We will prove now that the surface spherical harmonics are eigenfunctions
of the spherical Laplacean ∆S .
Theorem 7.5. If Y ∈ Hk , then
∆S Y (x) = −k(k + n − 2)Y (x) .

Proof. If Y is a restriction to the sphere of P ∈ Hk , then P is k-

homogeneous and hence the 0-homogeneous extension of Y is Ỹ (x) =
|x|−k P (x). Since P is harmonic we have16
∆ |x|−k P (x) = ∆|x|−k P (x) + 2∇|x|−k · ∇P (x) + |x|−k ∆P (x) .



| {z }
0

Because |x|−k = (x21 + . . . + x2n )−k/2 an easy computation shows that

∇|x|−k = −k|x|−(k+2) hx1 , . . . , xn i ,
∆|x|−k = div ∇|x|−k = −k(n − k − 2)|x|−(k+2) .
Hence
∆S Y (x) = ∆ |x|−k P (x) ||x|=1


n
X ∂P
= −k(n − k − 2)P (x) − 2k xi (x)
∂xi
|i=1 {z }
= kP (x)
= −k(k + n − 2)P (x)||x|=1
= −k(k + n − 2)Y (x) .
We used here
PnEuler’s theorem which asserts that a k-homogeneous function
f satisfies i=1 xi ∂f (x)/∂xi = kf (x). The proof is complete. 2
According to this theorem Hk is the eigenspace of the Laplace-Beltrami
operator ∆S corresponding to the eigenvalue −k(k + n − 2). Moreover
eigenspaces corresponding to different eigenvalues are mutually orthogonal.
Therefore there is an orthonormal basis of L2 (S n−1 ) consisting of smooth
eigenfunctions of ∆S (surface spherical harmonics). Analogous result holds
also for the Laplace-Beltrami operator on any smooth compact Riemannian
16We use here ∆(f g) = (∆f )g + 2∇f · ∇g + f ∆g.
68 PIOTR HAJLASZ

manifold, but the proof is to complicated to be presented here17. As we will

see later results of this type are very general: existence of an orthonormal
basis in a Hilbert space consisting of eigenfunctions is a general consequence
of the spectral theorem.
We proved in Analysis II that f ∈ C ∞ (S 1 ) if and only if np fˆ(n) → 0 as
|n| → ∞ for every p < ∞. As a striking application of Theorem 7.5 we will
generalize this result to the case of functions on S n−1 .
Theorem 7.6. Let f ∈ L2 (S n−1 ) has the spherical harmonics expansion
X∞
(7.4) f= Yk , Yk ∈ Hk .
k=0
Then f ∈ C ∞ (S n−1 )
(possibly after correction on a set of measure zero) if
and only if for every N > 0
Z
(7.5) |Yk |2 dσ = O(k −N ) as k → ∞ .
S n−1

Proof. We will need the following lemma.18

Lemma 7.7. If f, g ∈ C 2 (S n−1 ), then
Z Z
f ∆S g dσ = g∆S f dσ .
S n−1 S n−1

Proof. Let f˜ and g̃ be 0-homogeneous extensions of f and g to the annulus

Aε = B n (0, 1 + ε) \ B n (0, 1 − ε) .
Then the classical Green formula gives
∂ f˜
Z Z  ∂g̃
˜ ˜ f˜


f ∆g̃ − g̃∆f dx = − g̃ dσ = 0 ,
Aε ∂Aε ∂ν ∂ν
because f˜ and g̃ are constant along the direction of ν and hence ∂g̃/∂ν =
∂ f˜/∂ν = 0. Finally the fundamental theorem of calculus gives
Z Z
1
f˜∆g̃ − g̃∆f˜ dx = 0



f ∆S g − g∆S f dσ = lim
S n−1 ε→0 ε Aε

and the lemma follows. 2

Suppose f ∈ C ∞ (S n−1 ). We will prove (7.5). If we rewrite (7.4) with
normalized spherical harmonics
X∞
f= ak Yk0 , where Yk0 = Yk /kYk kL2 ,
k=0

17F.W. Warner, Foundations of differentiable manifolds and Lie groups, Chapter 6.

18The same lemma is true for the Laplace-Beltrami operator on any compact Riemann-
ian manifold without boundary, see F.W. Warner, Foundations of differentiable manifolds
and Lie groups.
 FUNCTIONAL ANALYSIS 69

then (7.5) is equivalent to |ak | = O(k −N/2 ) as k → ∞. Applying Lemma 7.7

m times we have
Z Z
∆S m f 2 ≥


m 0

m
L
∆S f · Yk dσ = f ∆S Yk0 dσ
S n−1 S n−1
Z
= (−k(k + n − 2))m f Yk0 dσ ,
n−1
| S {z }
ak
so
|ak | ≤ C(k(k + n − 2))−m = O(k −2m )
and hence (7.5) follows because m can be arbitrarily large.
Suppose now that f ∈ L2 (S n−1 ) has the expansion (7.4) satisfying (7.5).
It remains to prove that the series at (7.4) converges uniformly to a smooth
function on the sphere.

P∞ Recall that if the functions

P∞gi ∈ C ∞ (Ω) are such that for every multiindex
α gi k∞ < ∞, then i=1 Dα gi converges uniformly on Ω, and hence
PkD
i=1
∞
g = i=1 gi ∈ C ∞ (Ω). Thus a good estimate of derivatives of any order of
Yk will imply smoothness of f . Actually if Ỹk = Yk (x/|x|) in a neighborhood
of the sphere, it suffices to prove the estimate19
sup Dα Ỹk (x) ≤ Cα k n/2+|α| kYk kL2 .
 
(7.6)
|x|=1

Since Pk = |x|k Ỹk ∈ Hk is harmonic, the estimate will follow from suitable
estimates for harmonic functions.
Recall that harmonic functions have the mean value property, i.e. if u is
harmonic on Rn , then
Z
u(x) = u dσ
S n−1 (x,r)
for every x ∈ Rn and every r > 0. This easily implies Rthat if ϕ ∈
C0∞ (B n (0, 1)) is radial (i.e. constant on spheres centered at 0), Rn ϕ(x) dx =
1 and ϕε (x) = ε−n ϕ(x/ε), then
u(x) = (ϕε ∗ u)(x)
for every ε > 0 and every x ∈ Rn . Thus
Dα u(x) = (Dα ϕε ∗ u)(x)

19If F is a representation of Y in a local coordinate system (parametrization of a

k k
neighborhood of a point on the sphere), then it follows from the chain rule that on any
compact set K, supK |Dα Fk | ≤ C β≤α sup|x|=1 |Dβ Ỹk | ≤ C(α, K)kn/2+|α| kYk kL2 and
P
P α
hence k supK |D Fk | < ∞ by (7.5). This shows that Fk converges to a smooth function,
P
so does k Yk .
70 PIOTR HAJLASZ

and the Schwarz inequality easily yields

Z 1/2
α −n/2−|α| α
(7.7) |D u(x)| ≤ ε kD ϕkL2 |u|2 .
| {z } B(x,ε)
Cα

Now
Z Z Z 1+ε
2 2
|Pk (x)| dx = |Yk | dσ t2k+n−1 dt ≤ (1+ε)2k+n kYk k2L2 .
|x|≤1+ε S n−1 0

Now for any x ∈ S n−1 inequality (7.7) gives

Z 1/2
−n/2−|α|
α
|D Pk (x)| ≤ Cα ε |Pk |2 ≤ Cα ε−n/2−|α| (1+ε)k+n/2 kYk kL2 .
B(x,ε)

Taking ε = 1/k we see that (1 + ε)k+n/2 is bounded by a constant and hence

sup Dα Pk (x) ≤ Cα0 k n/2+|α| kYk k2L2
 
|x|=1

Since Ỹk (x) = |x|−k Pk (x), the Leibnitz rule implies (7.6). 2

8. Baire category theorem

The Baire category theorem proved below plays an important role in many
areas of mathematics. In this section we will show its applications outside
functional analysis and in the next two sections we will use it in the proofs
of two fundamental theorems in functional analysis, the Banach-Stenihaus
theorem and the Banach open mapping theorem.
Theorem 8.1 (Baire). The intersection of a countable family of open and
dense sets in a complete metric space is a dense set.

Proof. Let V1 , V2 , V3 , . . . be open and dense sets in a complete metric space.

We need to prove that their intersection V1 ∩ V2 ∩ . . . is dense. To this end
it suffices to prove that for every open set W 6= ∅
∞
\
(8.1) W∩ Vi 6= ∅ .
i=1
The density of V1 yields the existence of a ball
B(x1 , r1 ) = B1 ⊂ B1 ⊂ V1 ∩ W, r1 < 1 .
The density of V2 implies the existence of a ball
B(x2 , r2 ) = B2 ⊂ B2 ⊂ B1 ∩ V2 , r2 < 1/2 .
Similarly there is a ball
B(x3 , r3 ) = B3 ⊂ B3 ⊂ B2 ∩ V3 , r3 = 1/3
 FUNCTIONAL ANALYSIS 71

etc. We obtain a decreasing sequence of balls B(xi , ri ). The centers {xi }

form a Cauchy and hence convergent sequence xi → x ∈ X. Since the balls
Bj are closed and xi ∈ Bj for i ≥ j we conclude
∞
\ ∞
\
x∈ Bj ⊂ W ∩ Vj
j=1 j=1

which proves (8.1). 2

Definition. We say that a subset E of a metric space X is nowhere dense
if the closure E has no interior points. Sets which are countable unions of
nowhere dense sets are first category. All other sets are are second category.
Therefore the Baire theorem can be restated as follows.
Theorem 8.2 (Baire). A nonempty complete metric space is second cate-
gory.

As a first striking application we will prove

Proposition 8.3. If f ∈ C ∞ (R) and for every x ∈ R there is a nonnegative
integer n such that f (n) (x) = 0, then f is a polynomial.

The following exercise shows that the result cannot be to easy.

Exercise. Prove that there is a function f ∈ C 1000 (R) which is not a poly-
nomial, but has the property described in the above proposition.
Proof of the proposition. Let Ω ⊂ R be the union of all open intervals
(a, b) ⊂ R such that f |(a,b) is a polynomial. The set Ω is open, so
∞
[
(8.2) Ω= (ai , bi ) ,
i=1
where ai < bi and (ai , bi ) ∩ (bi , bj ) 6= ∅ for i 6= j. Observe that f |(ai ,bi ) is
a polynomial (Why?).20 We want to prove that Ω = R. First we will prove
that Ω = R. To this end it suffices to prove that for any interval [a, b], a < b
we have [a, b] ∩ Ω 6= ∅. Let
En = {x ∈ R : f (n) (x) = 0} .
The sets En ∩ [a, b] are closed and
∞
[
[a, b] = En ∩ [a, b] .
n=0

20It suffices to prove that f is a polynomial on every compact subinterval [c, d] ⊂ (a , b ).

i i
This subinterval has a finite covering by open intervals on which f is a polynomial. Taking
an integer n larger than the maximum of the degrees of these polynomials, we see that
f (n) = 0 on [c, d] and hence f is a polynomial of degree < n on [c, d].
72 PIOTR HAJLASZ

Since [a, b] is complete, it follows from the Baire theorem that for some n
the set En ∩ [a, b] has nonempty interior (in the topology of [a, b]), so there is
(c, d) ⊂ En ∩[a, b] such that f (n) = 0 on (c, d). Accordingly f is a polynomial
on (c, d) and hence
(c, d) ⊂ Ω ∩ [a, b] 6= ∅.
The set X = R \ Ω is closed and hence complete. It remains to prove that
X = ∅. Suppose not. Observe that every point x ∈ X is an accumulation
point of the set, i.e. there is a sequence xi ∈ X, xi 6= x, xi → x. Indeed,
otherwise x would be an isolated point, i.e. there would be two intervals
(8.3) (a, x), (x, b) ⊂ Ω, x 6∈ Ω .
The function f restricted to each of the two intervals is a polynomial, say of
degrees n1 and n2 . If n > max{n1 , n2 }, then f (n) = 0 on (a, x) ∪ (x, b). Since
f (n) is continuous on (a, b), it must be zero on the entire interval and hence
f is a polynomal of degree ≤ n − 1 on (a, b), so (a, b) ⊂ Ω which contradicts
(8.3).
The space X = R \ Ω is complete. Since
∞
[
X= X ∩ En ,
n=1

the second application of the Baire theorem gives that X ∩ En has a

nonempty interior in the topology of X, i.e. there is an interval (a, b) such
that
(8.4) X ∩ (a, b) ⊂ X ∩ En 6= ∅ .
Accordingly f (n) (x) = 0 for all x ∈ X ∩ (a, b). Since for every x ∈ X ∩ (a, b)
there is a sequence xi → x, xi 6= x such that f (n) (xi ) = 0 it follows from the
definition of the derivative that f (n+1) (x) = 0 for every x ∈ X ∩ (a, b), and
by induction f (m) (x) = 0 for all m ≥ n and all x ∈ X ∩ (a, b).

We will prove that f (n) = 0 on (a, b). This will imply that (a, b) ⊂ Ω which
is a contradiction with (8.4). Since f (n) = 0 on X∩(a, b) = (a, b)\Ω it remains
to prove that f (n) = 0 on (a, b)∩Ω. To this end it suffices to prove that for any
interval (ai , bi ) that appears in (8.2) such that (ai , bi )∩(a, b) 6= ∅, f (n) = 0 on
(ai , bi ). Since (a, b) is not contained in (ai , bi ) one of the endpoints belongs
to (a, b), say ai ∈ (a, b). Clearly ai ∈ X ∩ (a, b) and hence f (m) (ai ) = 0 for
all m ≥ n. If f is a polynomial of degree k on (ai , bi ), then f (k) is a nonzero
constant on (ai , bi ), so f (k) (ai ) 6= 0 by continuity of the derivative. Thus
k < n and hence f (n) = 0 on (ai , bi ). 2
Exercise. As the previous exercise shows the theorem is not true if we
only assume that f ∈ C 1000 . Where did we use in the proof the assumption
f ∈ C ∞ (R)?
 FUNCTIONAL ANALYSIS 73

Theorem 8.4. If a complete metric space X has no isolated points, then

every dense Gδ set is uncountable.

Proof. Suppose that a dense Gδ set E is countable, T E = {x1 , x2 , . . .}.

Then there are open and dense sets Vn such that E = ∞ n=1 Vn . Now the
sets Wn = Vn \{x1 ,T
. . . , xn } are still dense (because the space has no isolated
points) and hence ∞ n=1 Wn 6= ∅ by the Baire theorem. On the other hand
∞
\ ∞
\
Wn = Vn \ {x1 , x2 , . . .} = E \ E = ∅
n=1 n=1
which is a contradiction. 2
Theorem 8.5. If f : X → Y is a continuous mapping between metric
spaces, then the set of points where f is continuous is Gδ and the set of
points where f is discontinuous if Fσ .

Proof. We define oscillation of f at a point x ∈ X by

osc f (x) = lim diam (f (B(x, r)) .
r→0+
It is obvious that f is continuous at x if and only if osc f (x) = 0. It is also
easy to see that the sets
Un = {x ∈ X : osc f (x) < 1/n}
are open and hence
∞
\
{x : f is continuous at x} = Un
n=1
is Gδ . The second part of the theorem follows from the fact that the com-
plement of a Gδ set is Fσ . 2
Riemann constructed a function f : [0, 1] → R that is continuous exactly
at irrational points. However as a consequence of the above two results we
have
Corollary 8.6. There is no function f : [0, 1] → R that is continuous
exactly at rational points.

Since the set of irrational points if Gδ , Riemann’s example is a conse-

quence of the following more general result.
Exercise. Let E ⊂ R be a given Gδ set. Prove that there is a function
f : R → R such that the set of points where f is continuous equals E.
The last application of the Baire theorem provided in this section is more
important because the proof is based on the so called Baire category method
which, in many instances, is used to prove the existence of functions (or
other objects) that have strange properties.
74 PIOTR HAJLASZ

Theorem 8.7 (Banach). The set of function in C[0, 1] that are not differ-
entiable at any point of [0, 1] is a dense Gδ subset of C[0, 1]. In particular it
is not empty.

The existence of a nowhere differentiable function has been proved for

the first time by Weierstrass. His construction was explicit. The above re-
sult proves existence of such functions, but does not provide any explicit
example. On the other hand it proves that nowhere differentiability is a
typical property of continuous functions.

Proof of the theorem. Let f ∈ C[0, 1], x ∈ [0, 1] and 0 < r ≤ 1/2. If both
numbers x + r, x − r belong to the interval [0, 1], then we define

n |f (x + r) − f (x)| |f (x − r) − f (x)| o
(8.5) D(f, x, r) = min , .
r r

If only one of the numbers x + r, x − r belongs to [0, 1] we define D(f, x, r)

to be value of that number on the right hand side of (8.5) that exists, for
example if x + r > 1, we set D(f, x, r) = |f (x − r) − f (x)|/r. Let
[ [ 
Gn = f ∈ C[0, 1] : D(f, x, r) ≥ a for all x ∈ [0, 1] .
a>n 0<r<1/n

T∞
It is easy to see that if f ∈ n=1 Gn , then
 
 f (x + r) − f (x) 
lim sup 
  = +∞ for all x ∈ [0, 1]
r→0 r 

and hence f is not differentiable at any point of [0, 1]. Thus it remains to
prove that everyTset Gn is open and dense, because then the Baire theorem
will imply that ∞ 21
n=1 Gn is a dense Gδ subset of C[0, 1]. The fact that the
sets Gn are open is easy and left to the reader. To prove density it suffices
to prove that every g ∈ C[0, 1] can be uniformly approximated by functions
g̃ ∈ C[0, 1] such that

∃0<r<1/n ∀x∈[0,1] D(g̃, x, r) ≥ n + 1 ,

because this implies g̃ ∈ Gn . The construction of the function g̃ approximat-

ing g is explained on a sequence of pictures. First we approximate g with ε
accuracy by a simple function

21Because C[0, 1] is complete.

 FUNCTIONAL ANALYSIS 75

Then we modify it to a piecewise linear function which has flat parts and
parts with the slope larger than n + 1.

Finally we add little teeth to the flat part, so the slope of each tooth is
also larger than n + 1.

Clearly supx∈[0,1] |g(x) − g̃(x)| ≤ ε and if r is very small D(g̃, x, r) > n + 1

for all x ∈ [0, 1]. 2
76 PIOTR HAJLASZ

The proof is based on the following general idea. Given a complete metric
space X, we want to prove that there is an element x ∈ X that has a certain
property P . We find other, simpler to deal with, properties Pn , n = 1, 2, . . .
such that

(a) The sets Gn = {x ∈ X : x satisfies Pn } are open and dense;

(b) x has the property P if x has all the properties Pn .

x∈ ∞
T
Then
T∞ n=1 Gn has the property P and such an x exists, because the
set n=1 Gn is nonempty by the Baire theorem. This is what is called the
Baire category method.
Exercise. We say that a function f ∈ C ∞ (0, 1) is analytic at a ∈ (0, 1) if
there is ε > 0 such that f (x) = ∞ (n) (a)(x − a)n /n! for |x − a| < ε. Use
P
n=0 f
the Baire category method to prove that there is a function f ∈ C ∞ (0, 1)
that is not analytic at any point.

9. Banach-Steinhaus theorem

The following theorem as well as each of the four corollaries that follow
are called Banach-Steinhaus theorem.
Theorem 9.1 (Banach-Steinhaus). Let X be a complete metric space and
let {fi }i∈I be a family of continuous real-valued functions on X. Then exactly
one of the following two conditions is satisfied

(a) There is a nonempty open set U ⊂ X and a constant M > 0 such

that
sup |fi (x)| ≤ M for all x ∈ U ;
i∈I
(b) There is a dense Gδ set E ⊂ X such that
sup |fi (x)| = ∞ for all x ∈ E.
i∈I

Proof. If ϕ(x) = supi∈I |fi (x)|, then the sets

Vn = {x ∈ X : ϕ(x) > n}
are open because of continuity of the functions fi . Suppose that the condition
(a) is not satisfied. Then for every open set U 6= ∅, Vn ∩ UT6= ∅ and hence
the sets Vn are dense. According to the Baire theorem E = n Vn is a dense
Gδ set and obviously points of E satisfy (b). 2
Corollary 9.2. Let X be a complete metric space and let {fi }i∈I be a family
of continuous real-valued functions on X. If the functions in the family are
pointwise bounded, i.e.
sup |fi (x)| < ∞ for every x ∈ X,
i∈I
 FUNCTIONAL ANALYSIS 77

then there is a nonempty open set U ⊂ X and a constant M > 0 such taht
sup |fi (x)| ≤ M for all x ∈ U .
i∈I

The above two results are very important in the case X is a Banach space.
Corollary 9.3. Let X be a Banach space and Y a normed space. Let
{Li }i∈I ⊂ B(X, Y ). Then either
sup kLi k < ∞
i∈I
or there is a dense Gδ set E ⊂ X such that
sup kLi xk = ∞ for all x ∈ E .
i∈I

Proof. The functions fi (x) = kLi xk, i ∈ I are continuous and real-valued.
If the second condition is not satisfied, then it follows from the Banach-
Steinhaus theorem that there is an open ball B(x0 , r0 ) ⊂ X such that
sup |fi (x)| = M < ∞ for all x ∈ B(x0 , r0 ) .
i∈I
For x 6= 0
r0
y = x0 + x ∈ B(x0 , r0 )
2kxk
and hence for all i ∈ I we have
2kxk  r0  4M
kLi xk = Li x0 + x − Li x0 ≤ kxk

r0 2kxk r0
which yields
4M
sup kLi k ≤ < ∞.
i∈I r0
The claim is proved. 2
Corollary 9.4. Let X be a Banach space and Y a normed space. Let
{Li }i∈I ⊂ B(X, Y ). If for every x ∈ X
sup kLi xk < ∞
i∈I
then
sup kLi k < ∞ .
i∈I

Corollary 9.5. Let X be a Banach space and Y a normed space. If

{Ln }∞n=1 ⊂ B(X, Y ) is a pointwise convergent sequence, i.e. for every x ∈ X
the limit
Lx := lim Ln x
n→∞
exists, then L ∈ B(X, Y ). Moreover
sup kLn k < ∞
n
78 PIOTR HAJLASZ

and
kLk ≤ lim inf kLn k .
n→∞

Proof. Clearly L : X → Y is a linear mapping. The existence of the limit

Lx implies that
sup kLn xk < ∞ for every x ∈ X
n
and hence
sup kLn k < ∞
n
by the Banach-Steinhaus theorem. We have


kLxk = lim kLn xk = lim inf kLn xk ≤ lim inf kLn k kxk
n→∞ n→∞ n→∞

and the result follows. 2

Now we will present several applications of the Banach-Steinhaus theo-
rem.

9.1. Multilinear operators. Let X and Y be normed spaces over K = R

or C. We say that
B :X ×Y →K
is a two-linear functional if it is a linear functional with respect to each
variable, i.e.
B(x, ·) : Y → K is linear for every x ∈ X,

B(·, y) : X → K is linear for every y ∈ Y .

Similarly as in the case of linear functionals one can prove
Lemma 9.6. For a two-linear functional B : X × Y → K the following
conditions are equivalent.

(a) The function B is continuous.

(b) The function B is continuous at (0, 0).
(c) There is a constant C > 0 such that
|B(x, y)| ≤ Ckxk kyk for all x ∈ X, y ∈ Y .

It is known that even for a function on R2 continuity with respect to each

variable does not imply continuity. However we have
Theorem 9.7. If X and Y are Banach spaces and B : X × Y → K is a
two-linear functional which is continuous with respect to each variable, then
B is continuous.
 FUNCTIONAL ANALYSIS 79

Proof. It suffices to prove continuity at (0, 0), i.e. the implication

(xn , yn ) → (0, 0) ⇒ B(xn , yn ) → 0 .

Define a family of functionals Tn ∈ Y ∗ by Tn (y) = B(xn , y). Since for every

y ∈ Y , Tn (y) → B(0, y) = 0, we have

sup |Tn (y)| < ∞ for every y ∈ Y .

n

and thus
sup kTn k = M < ∞
n

by the Banach-Steihnaus theorem. Hence |Tn (y)| ≤ M kyk for every n and
all y ∈ Y . In particular

|B(xn , yn )| = |Tn (yn )| ≤ M kyn k → 0

and the result follows. 2

9.2. Landau’s theorem.

Theorem 9.8 (Landau). If the series ∞

P
i=1 ηi ξi converges for every (ξi ) ∈
`p , 1 ≤ p ≤ ∞, then (ηi ) ∈ `q , where 1/p + 1/q = 1.

Proof. We will prove the theorem in the case 1 < p ≤ ∞, but an obvi-
ous modification gives also the proof in the case p = 1. Define a bounded
functional on `p by the formula
n
X
Tn x = ηi ξi , n = 1, 2, 3, . . . ,
i=1

where x = (ξi ) ∈ `p . The functional Tn is fiven by (η1 , . . . , ηn , 0, 0, . . .) ∈ `q

and hence (see Theorem 2.12)
n
X 1/q
kTn k = |ηi |q .
i=1

Thus {Tn }n is a sequence of bounded functionals of `p , such that Tn x is

convergent for every x ∈ `p (by the assumption in the theorem) and hence
the Banach-Steinhaus theorem yields
∞
X 1/q
|ηi |q = sup kTn k < ∞ .
n
i=1

The proof is complete. 2

80 PIOTR HAJLASZ

9.3. Matrix summability methods. If a sequence (an ) converges to a

limit g, then also
a1 + a2 + . . . + an
(9.1) cn = → g.
n
There are, however, divergent sequences (an ) for which the sequence of arith-
metric means (9.1) converges. Thus assigning to the sequence (an ) the limit
of the sequence of arithmetric means (cn ) allows us to extend the notion
of the limit to a larger class of sequences, not necessarily convergent in the
oridinary sense. This particular method is known as the Cesaro summability
method.
We may, however, consider more general methods by taking instead of
arithmetic means other linear combinations of the ai ’s. This leads to the
following
Definition. We say that a sequence (an ) is summable to a generalized limit
g by the matrix summability method A = (ξij )∞ i,j=1 if

(1) the seriesP ∞

P
i=1 ξni ai converges for every n = 1, 2, 3, . . .;
∞
(2) limn→∞ i=1 ξni ai = g .

For example if
1
 
0 0 0 ...
 
1 0 1
1
   
 1 1

0 0 ... 
 
1

   2 2 
A1 =  A2 = 
 
1  
 1 1 1

0 ... 
 
 . .  
3 3 3
 .  



0 ...
then the sequence (an ) is summable to g by the method A1 if an → g and
by the method A2 if
a1 + a2 + . . . + an
→ g,
n
so we can recover both the classical notion of the convergence and the Cesaro
summability method.
Definition. We say that the matrix method A is regular if every convergent
sequence if summable to the same limit by the method A.
Clearly methods A1 and A2 are regular.
Theorem 9.9 (Toeplitz). The matrix summability method A = (ξij )∞
i,j=1 is
regular if and only if the following conditions are satisfied:

(a) supn ∞
P
i=1 |ξni | < ∞,
(b) limn→∞ P ξni = 0 for i = 1, 2, 3, . . .,
(c) limn→∞ ∞ i=1 ξni = 1.
 FUNCTIONAL ANALYSIS 81

Proof. Suppose that the matrix method A is regular. If (an ) =

(0, 0, . . . , 0, 1, 0, . . .) with 1 on ith coordinate, i.e. an = δni , then an → 0
and hence
X∞
lim ξni = lim ξnj aj = 0
n→∞ n→∞
j=1
which is condition (b). If (an ) = (1, 1, 1, . . .), then an → 1 and hence
∞
X ∞
X
lim ξni = lim ξni ai = 1
n→∞ n→∞
i=1 i=1
which is condition (c). The condition (a) is more difficult to prove and it
will actually follow from the Banach-Steinhaus theorem.
On the space of convergent sequences we define functionals22
m
X
Tnm x = ξni ai , where x = (ai ) ∈ c.
i=1
We have
m
X
kTnm k = |ξni | .
i=1
Since the series ∞
P
i=1 ξni ai converges for every x = (ai ) ∈ c, the sequence
(Tnm x)m converges for every x ∈ c and hence it follows from the Banach-
Steinhaus theorem that for every n
∞
X
sup kTnm k = |ξni | < ∞ .
m
i=1
This, however, implies that
∞
X
Tn x = ξni ai
i=1
is a bounded functional on c with
∞
X
kTn k = |ξni | .
i=1
By the assumption, the sequence Tn x converges for every x ∈ c and the
second application of the Banach-Steinhaus theorem gives
∞
X
sup kTn k = sup |ξni | < ∞
n n
i=1
which is the condition (a) of the theorem.
22It is easy to see that every sequence (ξ ) ∈ `1 defines a functional on T ∈ c∗ by
i
Tx = ∞
P P∞
i=1 ξi ai , where x = (ai ) ∈ c and kT k = i=1 |ξi |. This follows from the argument
used in the proof of Theorem 2.10.
82 PIOTR HAJLASZ

It remains to prove that conditions (a), (b), (c) imply that the method A
is regular.
For every x = (an ) ∈ c the series
∞
X
Tn x = ξni ai
i=1

converges and it actually defines a bounded functionals on c with

∞
X
sup kTn k = sup |ξni | < ∞ .
n n
i=1

Note that also

T x = lim an
n→∞
is a bounded functional on c. We have to prove that for every x ∈ c,
Tn x → T x, as n → ∞. Let A0 be a subset of c consisting of the se-
quences (1, 0, 0, . . .), (0, 1, 0, 0, . . .), (0, 0, 1, 0, 0, . . .),. . . and also of the se-
quence (1, 1, 1, 1, . . .). The conditions (b) and (c) readily imply that Tn x →
T x for all x ∈ A0 . By the linearity we also have
(9.2) Tn x → T x for x ∈ X0 = span A0 .
It is easy to see that X0 is a dense subset of c (why?), so we have convergence
on a dense subset of c. Since
sup kTn − T k ≤ kT k + sup kTn k < ∞
n n

it easily follows that

Tn x → T x for all x ∈ c.
Indeed, given x ∈ c and ε > 0 there is x00 ∈ X0 such that
ε
(9.3) kx0 − xk ≤
2 supn kTn − T k
and hence there is n0 such that for n > n0
|(Tn − T )x| ≤ kTn − T k kx − x0 k + |Tn x0 − T x0 | < ε
by (9.2) and (9.3). 2

9.4. Divergent Fourier series. If f ∈ C(S 1 ), then according to the Car-

leson theorem (Theorem 6.5) the sequence of partial sums of the Fourier
series converges to f a.e. This is natural to expect that this sequence actu-
ally converges to f everywhere. Surprisingly, this is not always true. We will
use the Banach-Steinhaus theorem to demonstrate existence of functions in
C(S 1 ) such that the Fourier series diverges on an uncountable and dense
subset of S 1 .
 FUNCTIONAL ANALYSIS 83

Theorem 9.10. There is a dense Gδ set E ⊂ C(S 1 ) such that for each
f ∈ E the set
{x ∈ [0, 1] : sup |sn (f, x)| = +∞}
n

is a dense and uncountable Gδ subset of [0, 1].

Proof. In the first step we will prove existence of continuous functions

with unbounded partial sums of the Fourier series at 0. For each integer n
we define a bounded functional on C(S 1 )

Λn f = sn (f, 0), f ∈ C(S 1 ) .

By Proposition 6.2 we have

Z 1/2
Λn f = f (y)Dn (y) dy ,
−1/2

where
sin π(2n + 1)y
Dn (y) = .
sin πy
Clearly
Z 1/2
(9.4) kΛn k ≤ |Dn (y)| dy .
−1/2

Actually we have equality

Z 1/2
(9.5) kΛn k = |Dn (y)| dy.
−1/2

Indeed, let

1 if Dn (y) ≥ 0,
g(y) =
−1 if Dn (y) < 0.

It is easy to see that there is a sequence fj ∈ C(S 1 ) such that −1 ≤ fj ≤ 1,

fj → g in L1 (S 1 ) and kfj k∞ = 1. Hence
Z 1/2
kΛn k ≥ lim Λn fj = lim fj (y)Dn (y) dy
j→∞ j→∞ −1/2
Z 1/2 Z 1/2
= g(y)Dn (y) dy = |Dn (y)| dy
−1/2 −1/2
84 PIOTR HAJLASZ

which together with (9.4) proves (9.5). We claim that kΛn k → ∞ as n → ∞.

Indeed, | sin πy| ≤ |πy| and hence
Z 1/2
kΛn k = |Dn (y)| dy
−1/2
Z 1/2
2 dy
≥ | sin π(2n + 1)y|
π 0 y
π(2n+1)y=t Z (n+1/2)π
dy/y=dt/t 2 dt
= | sin t|
π 0 t
n Z kπ
2X 1
> | sin t| dt
π kπ (k−1)π
k=1 | {z }
2
n
4 X 1
= → ∞ as n → ∞.
π k
k=1
We proved that
sup kΛn k = ∞.
n
According to the Banach-Steinhaus theorem there is a dense Gδ set E0 ⊂
C(S 1 ) such that
sup |Λn f | = sup |sn (f, 0)| = ∞, for every f ∈ E0 .
n n

In the above reasoning we could replace 0 by any other point x ∈ [0, 1], so
for each x ∈ [0, 1] there is a dense Gδ set Ex ⊂ C(S 1 ) such that
sup |sn (f, x)| = ∞ for all f ∈ Ex .
n

let {x1 , x2 , . . .} ⊂ [0, 1] be a dense subset. Then by the Baire theorem

Now T
E = i Exi is a dense Gδ set in C(S 1 ) and for every f ∈ E
sup |sn (f, xi )| = ∞ for all i = 1, 2, 3, . . .
n
We still need the following lemma that we leave as an exercise.
Lemma 9.11. If f ∈ C(S 1 ), then the set {x ∈ [0, 1] : supn |sn (f, x)| = ∞}
is Gδ .

Accordingly supn |sn (f, x)| = ∞ for x in a dense Gδ set, which is uncount-
able by Theorem 8.4. 2

10. Banach open mapping theorem

Another consequence of the Baire theorem is the following fundamental

result.
 FUNCTIONAL ANALYSIS 85

Theorem 10.1 (Banach open mapping theorem). If X and Y are Banach

spaces and T ∈ B(X, Y ) is a surjection, then T is an open mapping, i.e.
T (U ) ⊂ Y is open whenever U ⊂ X is open.

Proof. Since every ball can be mapped by an affine transformation onto

the unit ball, it suffices to prove that the image of the unit ball contains a
neighborhood of 0, i.e.
B(0, δ) ⊂ T (B(0, 1)) for some δ > 0.
Surjectivity of T gives
∞
[
Y = T (B(0, k))
k=1
and by Baire theorem at least one of the sets T (B(0, k)) has nonempty
interior, i.e.
B(y0 , η) ⊂ T (B(0, k)) .
For y such that kyk < η there exist sequences (x0n ), (x00n ) ⊂ B(0, k) such that
T x0n → y0 , T x00n → y0 + y .
This yields
T (x00n − x0n ) → y, x00n − x0n ∈ B(0, 2k)
and hence
B(0, η) ⊂ T (B(0, 2k)) .
Taking 2δ = η/2k we obtain
B(0, 2δr) ⊂ T (B(0, r)) for every r > 0.
In particular for y ∈ B(0, δ) (r = 1/2) there is x1 ∈ B(0, r) = B(0, 1/2)
such that ky − T x1 k < δ/2. Hence y − T x1 ∈ B(0, δ/2) (r = 1/4) and now
we find x2 ∈ B(0, r) = B(0, 1/4) such that
k(y − T x1 ) − T x2 k < δ/4 .
Hence
y − T (x1 + x2 ) ∈ B(0, δ/4), (r = 1/8) .
Next we find x3 ∈ B(0, 1/8) such that
ky − T (x1 + x2 + x3 )k < δ/8 .
By induction, we construct a sequence (xn ) ⊂ X such that
kxn k < 2−n and ky − T (x1 + . . . + xn )k < 2−n δ .
Since X is a Banach space we conclude that
X∞ X ∞
x= xn ∈ X, kxk ≤ kxn k < 1
n=1 n=1
and
T x = lim T (x1 + . . . + xn ) = y .
n→∞
86 PIOTR HAJLASZ

The proof is complete. 2

Corollary 10.2. If X and Y are Banach spaces and T ∈ B(X, Y ) is an
isomorphism of linear spaces X and Y , then T is an isomorphism of Banach
spaces X and Y , i.e. T −1 ∈ B(Y, X).

Compare this result with an example that follows Proposition 2.6.

Corollary 10.3. Let X be a Banach space with respect to any one of the
two norms k · k1 , k · k2 . If there is a constant C > 0 such that
kxk1 ≤ Ckxk2 for all x ∈ X
then the norms are equivalent and hence there is a constant C 0 > 0 such
that
kxk2 ≤ C 0 kxk1 for all x ∈ X.

Proof. The identity mapping id : (X, k·k2 ) → (X, k·k1 ) is an isomorphism

of Banach spaces by Corollary 10.2. 2
If f : X → Y is a function between Banach spaces, then its graph
Gf = {(x, y) ∈ X × Y : x ∈ X, y = f (x)}
is a subset of the Banach space X ⊕ Y .
Theorem 10.4 (Closed graph theorem). A linear mapping between Banach
spaces T : X → Y is bounded if and only if GT is a closed subset of X ⊕ Y .

Proof. The implication ⇒ is a direct consequence of the definition of

continuity, but the other implication ⇐ is more difficult. GT is a closed
linear subspace of the Banach space X ⊕ Y , so it is a Banach space with
respect to the norm
k(x, T x)k = kxk + kT xk .
The projection on the first component
X ⊕ Y 3 (x, y) 7→ x ∈ X
is a bounded operator. Its restriction to GT is a bounded operator as well.
Since it is an isomorphism of linear spaces GT and X, we conclude from the
Banach open mapping theorem that it is an isomorphism of Banach spaces
GT and X. Hence the inverse mapping
X 3 x 7→ (x, T x) ∈ GT
is bounded, i.e. kxk + kT xk ≤ Ckxk and thus kT xk ≤ (C − 1)kxk, which
proves boundedness of T . 2
According to the closed graph theorem in order to prove boundedness of
a linear mapping between Banach spaces T : X → Y it suffices to prove the
implication
(10.1) xn → x, T xn → y ⇒ y = Tx.
 FUNCTIONAL ANALYSIS 87

Now we will show two applications.

10.1. Symmetric operators are bounded.

Theorem 10.5 (Hellinger-Toeplitz). If a linear mapping of a Hilbert space
T : H → H is symmetric, i.e.
hT x, yi = hx, T yi
for all x, y ∈ H, then it is bounded.

Proof. We need to prove the implication (10.1). For every z ∈ H we have

hy, zi = lim hT xn , zi = lim hxn , T zi
n→∞ n→∞
= hx, T zi = hT x, zi .
Hence hT x − y, zi = 0 for all z ∈ H and taking z = T x − y we obtain
T x − y = 0. 2

10.2. Complemented subspaces.

Theorem 10.6. If M and N are closed subspaces of a Banach space X,
such that
M + N = X and M ∩ N = {0} ,
then the space X is isomorphic to the direct sum M ⊕ N . The isomorphism
is given by
M ⊕ N 3 (m, n) 7→ m + n ∈ X .
Moreover the spaces X/M and X/N are isomorphic to N and M respectively
under the quotient map.

Proof. M and N are Banach spaces as closed subspaces of a Banach space.

Hence M ⊕ N is a Banach space. The mapping
M ⊕ N 3 (m, n) 7→ m + n ∈ X
is one-to-one (because M ∩ N = {0}) and onto (because M + N = X).
Hence it follows from the open mapping theorem that it is an isomorphism
of Banach spaces. Similarly the mappings
M → X/N, m 7→ [m],

N → X/M n 7→ [n]
are one-to-one and onto, hence they are isomorphisms of Banach spaces. 2
Proposition 10.7. If L : X ⊕ Y → Z is an isomorphism of Banach spaces,
then L(X) and L(Y ) are closed subspaces of Z and hence Z = L(X)⊕L(Y ).
88 PIOTR HAJLASZ

Proof. The subspaces L(X) and L(Y ) are closed since

L(X) = L−1 (X × {0}), L(Y ) = L−1 ({0} × Y )
and the result easily follows. 2
Definition. Let M be a closed subspace of a Banach space X. We say that
M is complemented in X if there is a subspace N of X such that
M + N = X, M ∩ N = {0} .
Proposition 10.8. Every closed subspace M of a Hilbert space X is com-
plemented.

Proof. H = M ⊕ M ⊥ . 2
Definition. Let X be a Banach space. The mapping P ∈ B(X) = B(X, X)
is called a projection if P 2 = P , i.e. P (P (x)) = P (x) for all x ∈ X. We denote
the kernel (null space) and the range of the projection by
N (P ) = {x ∈ X : P x = 0}, R(P ) = {P x : x ∈ X} .
Theorem 10.9.

(a) If P ∈ B(X) is a projection, then23

X = R(P ) ⊕ N (P ) .
(b) If X = M ⊕ N is a direct sum of closed subspaces, then there is a
projection P ∈ B(X) such that
M = R(P ) and N = N (P ) .
Corollary 10.10. A closed subspace of a Banach space is complemented if
and only if it is the image of a projections.

Proof of Theorem 10.9. (a) N (P ) is closed as a preimage of 0 of a contin-

uous mapping. Since
R(P ) = N (I − P ) = {x ∈ X : x = P x} ,
also R(P ) is closed. Now it suffices to show that
R(P ) ∩ N (P ) = {0} and R(P ) + N (P ) = X .
If x ∈ R(P ) ∩ N (P ), then x = P x and P x = 0, so x = 0. Moreover every
element x ∈ X can be represented as
P x + (x − P x) .
x = |{z}
| {z }
∈R(P ) ∈N (P )

(b) If P : M ⊕ N → M , P (x, y) = x is the projection onto the first compo-

nent, then M = R(P ) and N = N (P ). The proof is complete. 2
23i.e. the subspaces N (P ) and R(P ) are closed and R(P )+N (P ) = X, R(P )∩N (P ) =
{0}
 FUNCTIONAL ANALYSIS 89

The following theorem is a consequence of the Hahn-Banach theorem and

we will prove it in the next section.
Theorem 10.11. Let M be a closed subspace of a Banach space X.

(a) If dim M < ∞, then M is complemented in X.

(b) If dim(X/M ) < ∞, then M is complemented in X.

Actually it turns out that a property of being a complemented subspace

is quite rare. Namely we have
Theorem 10.12 (Lindenstrauss-Tzafriri). If X is a real Banach space and
every closed subspace of X is complemented, then X is isomorphic to a
Hilbert space.

This is a very difficult theorem and we will not prove it. We will prove,
however
Theorem 10.13 (Phillips). The space c0 is not complemented in `∞ . Equiv-
alently there is no bounded linear projection of `∞ onto c0 .

Proof. We will need the following lemma.

Lemma 10.14. There is an uncountable family {Ai }i∈I of subsets of N such
that

(a) Ai is infinite for every i ∈ I;

(b) Ai ∩ Aj is finite for i 6= j.

Proof. It suffices to prove the lemma with N replaced by Q ∩ (0, 1). For
each irrational number i ∈ (0, 1) \ Q := I let Ai be a sequence of rationals
in (0, 1) convergent to i. It is easy to see that the family {Ai }i∈I has the
desired properties. 2
By contradiction suppose that
`∞ = c0 ⊕ X, X ⊂ `∞ ,
so X ' `∞ /c0 . Consider the family of functionals {en }n∈N on X
hen , xi = xn , x = (xi ) ∈ X ⊂ `∞ .
The family {en }n in X ∗ is total, i.e. it has the property that
hen , xi = 0 for all n ⇒ x = 0.
It suffices to prove that there is no countable total family of functionals in
(`∞ /c0 )∗ .
90 PIOTR HAJLASZ

Let {Ai }i∈I be an uncountable family of subsets of N as in the lemma.

We can identify `∞ with the space of bounded functions f : N → C and for
each i ∈ I we define
fi = χAi ∈ `∞ , [fi ] ∈ `∞ /c0 .
Note
Pm that for any finite collection of functions fi1 , . . . , fim from the family,
k=1 fik > 1 on a finite subset A ⊂ N by the property (b) of the lemma.
Since
Xm
χA fik ∈ c0
k=1
we conclude that
Xm X m m
X
[f ] ≤ f − χ f = 1.

ik ik A ik
∞
k=1 k=1 k=1

By the same argument if bk ∈ C, |bk | ≤ 1, k = 1, 2, . . . , m, then

Xm
bk [fik ] ≤ 1 .


k=1

We will prove now that for every x∗ ∈ (`∞ /c0 )∗ the set
{[fi ] : hx∗ , [fi ]i =
6 0}
is countable. To this end it suffices to prove that for each integer n the set
C(n) = {[fi ] : hx∗ , [fi ]i ≥ 1/n}
is finite. Choose [fi1 ], . . . , [fim ] ∈ C(n) and let

bk = sgnhx∗ , [fik ]i = hx∗ , [fik ]i/|hx∗ , [fik ]i| .

Pm
It follows from the above observation that for x = k=1 bk [fik ] we have
kxk ≤ 1. Hence
kx∗ k ≥ |hx∗ , xi| ≥ m/n, m ≤ nkx∗ k,
so C(n) is finite for every n and actually #C(n) ≤ nkx∗ k. Accordingly, if
{hn }n∈N is a countable subset of (`∞ /c0 )∗ , then
∞
[
{[fi ] : hhn , [fi ]i =
6 0 for some n = 1, 2, 3, . . .} = {[fi ] : hhn , [fi ]i =
6 0}
n=1

is countable and hence there is24 fi such that hhn , [fi ]i = 0 for all n, so the
family {hn }n cannot be total. 2

24Because the family {f }

i i∈I is uncountable.
 FUNCTIONAL ANALYSIS 91

11. Hahn-Banach Theorem

Let X be a linear space over R. We do not require X to be a linear normed

space. The Hahn-Banach theorem says that functionals defined on subspaces
of X can be extended to functionals on X in a way that if they satisfy a
certain inequality on the subspace, then the extension will satisfy the same
inequality.
Definition. A function p : X → R defined on a real linear space X is called
a Banach functional if
p(x + y) ≤ p(x) + p(y), p(tx) = tp(x)
for all x, y ∈ X and t ≥ 0.
Theorem 11.1 (Hahn-Banach). Let p : X → R be a Banach functional on
a linear space X over R and let M be a linear subspace of X. If f : M → R
is a linear functional satisfying
f (x) ≤ p(x) for x ∈ M ,
then there is a linear functional F : X → R being an extension of f , i.e.
f (x) = F (x) for x ∈ M
and such that
−p(−x) ≤ F (x) ≤ p(x) for x ∈ X.

Proof. In the first step we will prove that if f˜ is a functional defined on a

proper subspace M̃ of X and it satisfies f˜(x) ≤ p(x) on that subspace, then
we can extend it to a bigger subspace M̃ ˜ ⊃ M̃ . Actually M̃ ˜ will be obtained
from M̃ by adding one independent vector, so that dim(M̃/ ˜ M̃ ) = 1, and of
˜ ˜ . In the second part
course the extension will satisfy f˜(x) ≤ p(x) for x ∈ M̃
we will use the Hausdorff maximality theorem to conclude the existence of
the maximal extension. Then it will easily follow that the maximal extension
is defined on X and has all required properties.
Suppose that f˜ : M̃ → R satisfies
f˜(x) ≤ p(x), for x ∈ M̃
and that M̃ 6= X. Let x1 ∈ X \ M̃ and define
˜ = {x + tx : x ∈ M̃ , t ∈ R} .
M̃ 1

Then M̃ is a proper subspace of M̃ ˜ . For x, y ∈ M̃ we have

f˜(x) + f˜(y) = f˜(x + y) ≤ p(x + y) ≤ p(x − x1 ) + p(x1 + y) .

Hence
f˜(x) − p(x − x1 ) ≤ p(x1 + y) − f˜(y) .
92 PIOTR HAJLASZ

Let
α = sup f˜(x) − p(x − x1 ) .
x∈M̃

Then
(11.1) f˜(x) − α ≤ p(x − x1 ) and f˜(y) + α ≤ p(x1 + y) .
˜ by the formula
We extend f˜ to M̃
˜
f˜(x + tx1 ) = f˜(x) + tα .
Replacing x and y by x/(−t), t < 0 and x/t, t > 0 in (11.1), after simple
calculations we arrive at
˜ ˜.
f˜(x) ≤ p(x) on M̃
Now we want to apply the Hausdorff maximality theorem. Consider the
family of pairs (M̃ , f˜), where M̃ ⊃ M is a subspace of X and f˜ : M̃ → R is
an extension of f satisfying
f˜(x) ≤ p(x) for x ∈ M̃ .
The family is partially ordered by the relation
˜ f˜˜)
(M̃ , f˜) ≤ (M̃,
˜ and f˜˜ is an extension of f˜. According to the Hausdorff maximality
if M̃ ⊂ M̃
theorem there is a maximal element in the family. Because of the first part
of the proof the maximal element must be a functional defined on all of X.
Denote it by (X, F ). Hence
F (x) = f (x) for x ∈ M , F (x) ≤ p(x) for x ∈ X.
Now it suffices to observe that
F (x) = −F (−x) ≥ −p(−x)
by linearity of F . 2
Theorem 11.2 (Hahn-Banach). Let M be a subspace of a linear space X
over K (= R of C), and let p be a seminorm25 on X. Let f be a linear
functional on M such that
|f (x)| ≤ p(x) for x ∈ M .
Then there is an extension F : X → K of f such that
F (x) = f (x) for x ∈ M ; |F (x)| ≤ p(x) for x ∈ X.

25i.e. p(x + y) ≤ p(x) + p(y), p(αx) = |α|p(x) for x, y ∈ X and α ∈ K, see Section 4.2.
 FUNCTIONAL ANALYSIS 93

Proof. Observe that in Theorem 11.1 we required X to be a real space,

but now it can also be a complex one. Actually when K = R, Theorem 11.2
immediately follows from Theorem 11.1, because any seminorm is Banach
functional, so we are left with the case K = C.
Every linear space over C can be regarded as a linear space over R. For
example Cn can be regarded as R2n . Thus if X is a linear space over C we
have two kinds of functionals: C-linear and R-linear. More precisely C-linear
functionals are
Λ : X → C, Λ(ax + by) = aΛx + bΛy a, b ∈ C
and R-linear functionals are
Λ : X → R, Λ(ax + by) = aΛx + bΛy a, b ∈ R .
The functional f : M → C given in the assumptions of the theorem is clearly
C-linear and we need find its C-linear extension. The functional u = re f :
M → R is R-linear. Note that
f (x) = u(x) − iu(ix) .
Now according to Theorem 11.1, the functional u can be extended to an
R-linear functional
U : X → R, U (x) ≤ p(x) for x ∈ X,
and hence
F (x) = U (x) − iU (ix)
is a C-linear extension of f . Since for every x ∈ X there is α ∈ C such that
|α| = 1 and αF x = |F x| we conclude
|F x| = F (αx) = U (αx) ≤ p(αx) = p(x) .
where the second equality follows from the fact that F (αx) ∈ R and U =
re F . 2
Corollary 11.3. If f is a bounded functional defined on a subspace M of a
normed space X, then there is a functional on X, i.e. there is F ∈ X ∗ such
that26
F (x) = f (x) for x ∈ X and kf k = kF k .

Proof. |f (x)| ≤ kf k kxk for x ∈ M . The function p(x) = kf k kxk is a

seminorm on X. Hence there is an extension F such that |F (x)| ≤ kf k kxk
for all x ∈ X. Thus kF k ≤ kf k and the opposite inequality kF k ≥ kf k is
obvious. 2
Corollary 11.4. If X is a normed space and x0 ∈ X, x0 6= 0, then there is
x∗ ∈ X ∗ such that
kx∗ k = 1 and hx∗ , x0 i = kx0 k .
26Here kf k = sup
kxk≤1 |f (x)| and kF k = sup kxk≤1 |F (x)|.
x∈M x∈X
94 PIOTR HAJLASZ

Proof. On a subspace M = {αx0 : α ∈ K} we define a functional f (αx0 ) =

αkx0 k. Clearly kf k = 1 and f (x0 ) = kx0 k. Now it suffices to take x∗ = F ,
where F in a norm preserving functional from Corollary 11.3. 2
The corollary implies that if X 6= {0}, then X ∗ 6= {0}.
Recall that the norm in X ∗ is defined by
kx∗ k = sup hx∗ , xi .
x∈X
kxk≤1

As a corollary from the Hahn-Banach theorem we also have

Corollary 11.5. If X is a normed space and x ∈ X, then
kxk = sup hx∗ , xi .
x∗ ∈X ∗
kx∗ k≤1

Theorem 11.6. Let X0 be a linear subspace of a normed space X and

x1 ∈ X \ X0 . Suppose that the distance of x1 to X0 is positive, i.e.
d = inf{kx1 − x0 k : x0 ∈ X0 } > 0 .
Then there is a functional x∗ ∈ X ∗ such that
1
hx∗ , x1 i = 1, kx∗ k = , hx∗ , x0 i = 0 for x0 ∈ X0 .
d

Proof. Let K = R or C and

X1 = {x0 + αx1 : x0 ∈ X0 , α ∈ K}.
Define a functional on X1 by
hx∗1 , x0 + αx1 i = α .
For α 6= 0 we have
 x 
0
kx0 + αx1 k = |α| x1 − − ≥ |α|d = |hx∗1 , x0 + αx1 i| d ,

α
which yields
1
(11.2) kx∗1 k ≤ .
d
Now let xn0 ∈ X0 be a sequence such that
kxn0 − x1 k → d .
We have
1 = |hx∗1 , xn0 − x1 i| ≤ kx∗1 k kxn0 − x1 k → kx∗1 k d ,
i.e. kx∗1 k ≥ 1/d. This inequality together with (11.2) gives kx∗1 k = 1/d. Now
let x∗ be a norm preserving extension of x∗1 to X. Then kx∗ k = 1/d. We
have
hx∗ , x0 i = hx∗1 , x0 + 0 · x1 i = 0 for x0 ∈ X0 ,
hx∗ , x1 i = hx∗1 , 0 + 1 · x1 i = 1 .
The proof is complete. 2
 FUNCTIONAL ANALYSIS 95

Theorem 11.7. If the dual space X ∗ to a normed space X is separable,

then X is separable too.

Proof. Let {x∗i }∞ ∗

i=1 ⊂ X be a countable and dense subset. For each i we
can find xi ∈ X such that
1
kxi k ≤ 1 and hx∗i , xi i ≥ kx∗i k .
2
Let X0 be a subspace of X that consists of finite linear combinations of the
xi ’s. It suffices to prove that X0 is a dense subset of X (Why?).
Suppose X0 is not dense in X. Then it follows from the previous theorem
that there is a functional 0 6= x∗ ∈ X ∗ such that hx∗ , xi = 0 for all x ∈ X0 .
Let x∗ij be a sequence such that
kx∗ij − x∗ k → 0 as j → ∞.
We have
0 ← kx∗ij − x∗ k = sup |hx∗ij − x∗ , xi| ≥ hx∗ij − x∗ , xij i
kxk≤1
1 1
= hx∗ij , xij i ≥ kx∗ij k → kx∗ k > 0 .
2 2
Contradiction. 2
Example. Every sequence s = (si ) ∈ `1 defines a functional on `∞ by
∞
X
∗
(11.3) hx , xi = si xi , where x = (xi ) ∈ `∞ .
i=1
It is easy to see (see Exercise following Theorem 2.11) that kx∗ k = ksk1 .
That means `1 is isometrically isomorphic to a closed subspace of (`∞ )∗ .
However `1 6= (`∞ )∗ because `1 is separable and `∞ is not.
There is, however, a more direct proof of this fact.
(11.4) hx∗ , xi = lim xi , for x = (xi ) ∈ c
i→∞
is a bounded nonzero functional on c ⊂ `∞ and hence according to the
Hahn-Banach theorem it can be extended to a bounded functional of `∞ . It
is clear that this functional cannot be of the form (11.3). Indeed, the value
of the functional (11.4) does not change if we change value of any finite
number of coordinates of x = (xi ) ∈ c and this is not true for the functional
(11.3). 2
As an application of the Hahn-Banach theorem we will prove Theo-
rem 10.11.
Theorem 11.8. Let M be a closed subspace of a Banach space X.

(a) If dim M < ∞, then M is complemented in X.

96 PIOTR HAJLASZ

(b) If dim(X/M ) < ∞, then M is complemented in X.

Proof. (a) Let e1 , . . . , en be a Hamel basis of M . Then every element

x ∈ M can be represented as
x = α1 (x)e1 + . . . + αn (x)en .
The coefficients α1 (x), . . . αn (x) are linear functionals on M . The space M
is isomorphic as a Banach space to27 Kn and hence every linear functional
on M is continuous. Thus the functions αi are continuous. By the Hahn-
Banach theorem the functionals αi can be extended to bounded functionals
αi∗ ∈ X ∗ . Let
n
\
N= ker αi∗ .
i=1
It is easy to see that
M ∩ N = {0} and M + N = X.
(b) Let π : X → X/M , π(x) = [x] be the standard quotient mapping. Let
e1 , . . . , en be a basis in X/M . Let xi ∈ X be such that π(xi ) = ei and let
N = span {x1 , . . . , xn }. Now it easily follows that
M ∩ N = {0}, M +N =X.
The proof is complete. 2
We close this section with several applications of the Hahn-Banach theo-
rem.

11.1. Banach limits. As we have seen in Section 9.3 matrix summability

methods provide a way to extend the notion of limit to some sequences that
are not necessarily convergent, but is it possible to extend the notion of
limit to all bounded sequences? The following result provides a satisfactory
answer.
Theorem 11.9 (Mazur). To each bounded sequence of real numbers x =
(xn ) ∈ `∞ we can assign a generalized limit LIM n→∞ xn (called Banach
limit) so that

(a) For convergent sequences x = (xn ) ∈ c

LIM xn = lim xn ;
n→∞ n→∞
(b) For all x = (xn ), y = (yn ) ∈ `∞ and a, b ∈ R
LIM (axn + byn ) = a LIM xn + b LIM yn ;
n→∞ n→∞ n→∞
(c) For all x = (xn ) ∈ `∞ and all k ∈ N
LIM xn+k = LIM xn ;
n→∞ n→∞

27Rn or Cn .
 FUNCTIONAL ANALYSIS 97

(d) For all x = (xn ) ∈ `∞

lim inf xn ≤ LIM xn ≤ lim sup xn .
n→∞ n→∞ n→∞

Proof. It is easy to see that

x1 + x2 + . . . + xn
p : `∞ → R, p(x) = lim sup
n→∞ n
∞
is a Banach functional. Let M ⊂ ` be a linear subspace consisting of
sequences for which the limit
x1 + x2 + . . . + xn
`(x) = lim ∈R
n→∞ n
exists. Clearly c ⊂ M and ` : M → R is a linear functional such that
`(x) = p(x) ≤ p(x) for x ∈ M .
According to the Hahn-Banach theorem ` can be extended to a linear func-
tional on `∞ denoted by LIM n→∞ xn such that
(11.5) −p(−x) ≤ LIM xn ≤ p(x) for all x ∈ `∞ .
n→∞
Properties (a) and (b) are obvious. It is easy to see that
p(x) ≤ lim sup xn , lim inf xn = − lim sup(−xn ) ≤ −p(−x) .
n→∞ n→∞ n→∞
The two inequalities combined with (11.5) yield (d). We are left with the
proof of (c). We have
LIM xn+1 − LIM xn = LIM (xn+1 − xn ) .
n→∞ n→∞ n→∞
Since
(x2 − x1 ) + (x3 − x2 ) + . . . + (xn+1 − xn )
p((xn+1 − xn )) = lim sup
n→∞ n
xn+1 − x1
= lim sup = 0,
n→∞ n
and similarly p(xn − xn+1 ) = 0, inequality (11.5) yields
LIM xn+1 = LIM xn
n→∞ n→∞
and (c) follows by induction. 2
This result can be generalized to the class of bounded real-valued func-
tions on [0, ∞). Let B [0, ∞) be the class of all bounded functions f :
[0, ∞) → R (no measurability condition). B [0, ∞) is a Banach space with
respect to the supremum norm28 k · k∞ .
Theorem 11.10 (Banach). In the space B [0, ∞) there is a functional
LIM t→∞ x(t) such that
28In the notation used in Section 5.1, B [0, ∞) = `∞ ([0, ∞)).
98 PIOTR HAJLASZ

(a) If the limit limt→∞ x(t) exists, then

lim x(t) = LIM x(t) ;
t→∞ t→∞

(b)
LIM (ax(t) + by(t)) = a LIM x(t) + b LIM y(t) ;
t→∞ t→∞ t→∞
(c)
LIM x(t + τ ) = LIM x(t) for every τ > 0 ;
t→∞ t→∞
(d)
lim inf x(t) ≤ LIM x(t) ≤ lim sup x(t) .
t→∞ t→∞ t→∞

Note that Mazur’s theorem follows from that of Banach. Indeed, if x =

(xn ) ∈ `∞ , it suffices to apply the Banach theorem to the function
∞
X
x(t) = xn χ[n−1,n)
n=1

and set
LIM xn := LIM x(t) .
n→∞ t→∞

Proof of Theorem 11.10. The subspace M ⊂ B [0, ∞) consisting of func-

tions for which the limit
(11.6) lim x(t)
t→∞

exists is a linear subspace and the limit (11.6) is a linear functional on M .

We want to extend it to an appropriate functional on B [0, ∞) and the main
difficulty is a construction of a suitable Banach functional.
For x ∈ B [0, ∞) and t1 , . . . , tn ≥ 0 we define
n
1 X
β(x; t1 , . . . , tn ) = lim sup x(t + ti )
t→∞ n
i=1

and then

p(x) = inf β(x; t1 , . . . , tn ) : ti ≥ 0, n ∈ N .
We will prove that p(x) is a Banach functional. Clearly p(tx) = tp(x) for
t > 0 and we only need to prove subadditivity p(x + y) ≤ p(x) + p(y).
Let t1 , . . . , tn ≥ 0 and s1 , . . . , sm ≥ 0. Then ti + sj is a collection of nm
numbers. Denote these numbers by u1 , . . . , unm . It it easy to see that the
subadditivity of p follows from the lemma.
Lemma 11.11.
β(x + y; u1 , . . . , unm ) ≤ β(x; t1 , . . . , tn ) + β(y; s1 , . . . , sm ) .
 FUNCTIONAL ANALYSIS 99

Proof. We have
nm m n n m
1 X 1 X1X 1X 1 X
(x+y)(t+uk ) = x(t+ti +sj )+ y(t+ti +sj ) .
nm m n n m
k=1 j=1 i=1 i=1 j=1

Applying lim supt→∞ to both sides yields

m n
1 X 1X
β(x + y; u1 , . . . , unm ) ≤ lim sup x(t + ti + sj )
m t→∞ n
j=1 i=1
| {z }
β(x;t1 ,...,tn ) because t + sj → ∞
n m
1 X 1 X
+ lim sup y(t + ti + sj )
n t→∞ m
i=1 j=1
| {z }
β(y;s1 ,...,sm ) because t + ti → ∞

= β(x; t1 , . . . , tn ) + β(y; s1 , . . . , sm ) .
This completes the proof of the lemma and hence that of the fact that p(x)
is a Banach functional. 2
If x ∈ M ⊂ B(0, ∞), then
lim x(t) = p(x) ≤ p(x)
t→∞
and hence according to the Hahn-Banach theorem the functional x 7→
limt→∞ x(t) extends from M to a linear functional on B [0, ∞) that we
denote by LIM t→∞ x(t) such that
−p(−x) ≤ LIM x(t) ≤ p(x) for all x ∈ B [0, ∞).
t→∞

It is easy to see that p(x) ≤ lim supt→∞ x(t) and hence29

lim inf x(t) ≤ LIM x(t) ≤ lim sup x(t) .
t→∞ t→∞ t→∞
It remains to show that
LIM xτ (t) = LIM x(t)
t→∞ t→∞
where xτ (t) = x(t + τ ). Let y = xτ − x. Then
1 2
p(y) ≤ β(y; 0, τ, 2τ, . . . , (n − 1)τ ) = lim sup (x(t + nτ ) − x(t)) ≤ kxk∞ .
t→∞ n n
Since this inequality holds for any n ∈ N, we have p(y) ≤ 0. Similarly
p(−y) ≤ 0 and hence
0 ≤ −p(−y) ≤ LIM (xτ − x) ≤ p(y) ≤ 0 ,
t→∞
so LIM t→∞ (xτ − x) = 0, LIM t→∞ xτ (t) = LIM t→∞ x(t). The proof is com-
plete. 2
29Because lim inf
t→∞ x(t) = − lim supt→∞ (−x(t)) ≤ −p(−x).
100 PIOTR HAJLASZ

11.2. Finitely additive measures. In this section we will prove the fol-
lowing surprising result.
Theorem 11.12 (Banach). There is a finitely additive measure µ : 2R →
[0, ∞], i.e.
µ(A ∪ B) = µ(A) + µ(B) for A, B ⊂ R, A ∩ B = ∅
defined on all subsets of R and such that

(a) µ is invariant with respect to isometries, i.e.

µ(A + t) = µ(A), µ(−A) = µ(A)
for all A ⊂ R, t ∈ R.
(b)
µ(A) = L1 (A)
for all Lebesgue measurable sets A ⊂ R.

Banach proved that the above result holds also in R2 , i.e. there is a finitely
additive measure
2
µ : 2R → [0, ∞]
defined on all subsets of R2 , invariant under isometries of R2 and equal to
the Lebesgue measure on the class of Lebesgue measurable sets. According
to the Vitali example such measures cannot be, however, countably additive.
Surprisingly, the Banach theorem does not hold in Rn , n ≥ 3. This is
related to an algebraic fact that the group of isometries of Rn , n ≥ 3 contains
a free supgroup of rank 2. Namely Banach and Tarski proved30 that the unit
ball in R3 can be decomposed into a finite number of disjoint sets (later it
was shown that it suffices to take 5 sets)
B 3 (0, 1) = A1 ∪ A2 ∪ A3 ∪ A4 ∪ A5
is a way that there are isometries τ1 , . . . , τ5 of R3 such that
B 3 (0, 1) = τ1 (A1 ) ∪ τ2 (A2 ), B 3 (0, 1) = τ3 (A3 ) ∪ τ4 (A4 ) ∪ τ5 (A5 )
and a similar decomposition is possible for B n (0, 1) for any n ≥ 3.
Clearly the sets Ai cannot be Lebesgue measurable, because we would
obtain a contradiction by comparing volumes. This also shows that there
is no finitely additive measure in Rn , n ≥ 3 invariant under isometries and
equal to the Lebesgue measure on the class of Lebesgue measurable sets.
The Banach theorem will follow from a somewhat stronger result. As in
the case of Fourier series we can identify bounded functions on R with period
1 with bounded functions on S 1 via the exponential mapping t R7→ e2πit . Thus
1
the integral of a function on S 1 corresponds to the integral 0 x(t) dt of a
function x : R → R with period 1.
30See S. Wagon, The Banach-Tarski paradox, Cambridge Univ. Press 1999.
 FUNCTIONAL ANALYSIS 101

The space of all (not necessarily measurable) real-valued functions on R

with period 1 will be denoted by B(S 1 ). Since every bounded function on
R with period 1 is uniquely determined by a bounded function on [0, 1) we
can further identify B(S 1 ) with bounded real-valued functions on [0, 1).
Theorem 11.13
R (Banach). In the space B(S 1 ) there is a linear functional
denoted by x(t) dt such that

(a) Z Z Z


ax(t) + by(t) dt = a x(t) dt + b y(t) dt ;
(b) Z
x(t) dt ≥ 0 if x ≥ 0
(c) Z Z
x(t + τ ) dt = x(t) dt for all τ ∈ R;
(d) Z Z
x(1 − t) dt = x(t) dt ;

(e) For any Lebesgue measurable function x ∈ B(S 1 )

Z Z 1
x(t) dt = x(t) dt ,
0
where on the right hand side we have the integral with respect to the
Lebesgue measure.

Before we prove this theorem we show how to conclude Theorem 11.12.

Proof of Theorem 11.12. Theorem 11.13 defines a generalized integral of
a bounded function on [0, 1). For A ⊂ R and k ∈ Z let
xk = χ(A−k)∩[0,1) .
R
Thus xk (t) dt corresponds to a generalized measure of the set A ∩ [k, k + 1)
and it is natural to define
X∞ Z
µ(A) = xk (t) dt .
k=−∞
It is easy to check now that µ satisfies the claim of Theorem 11.12. 2
Proof of Theorem 11.13. For x ∈ B(S 1 ) and t1 , . . . , tn ≥ 0 we define
n
1 X
β(x; t1 , . . . , tn ) = sup x(t + ti ) : t ∈ R
n
i=1
and then 
p(x) = inf β(x; t1 , . . . , tn ) : ti ≥ 0, n ∈ N.
102 PIOTR HAJLASZ

As in the proof of Theorem 11.10 one can show that p(x) is a Banach func-
tional such that
(11.7) p(xτ − x) ≤ 0, p(x − xτ ) ≤ 0 .
For a Lebesgue measurable function x ∈ B(S 1 ) we define a functional
Z 1
∗
hx1 , xi = x(t) dt .
0
It is easy to see that
hx∗1 , xi ≤ p(x)
and hence x∗1 can be extended to a functional x∗ on B(S 1 ) such that
−p(−x) ≤ hx∗ , xi ≤ p(x) for all x ∈ B(S 1 ).
Finally we define Z
1
x(t) dt = hx∗ , x + x− i ,
2
where x− (t) = x(1−t). The properties (a), (d) and (e) are obvious. Property
(b) follows from the observation that if x ≥ 0, then p(−x) ≤ 0 and hence
0 ≤ −p(−x) ≤ hx∗ , xi, 0 ≤ −p(−x− ) ≤ hx∗ , x− i .
Finally, property (c) follows from the inequality (11.7). 2

11.3. Runge’s theorem. Let Ω = B 2 (0, 2) \ B 2 (0, 1) be an annulus. The

function f (z) = z −1 is holomorphic in Ω, but it cannot be uniformly ap-
proximated by polynomials on on compact subsets of Ω. Indeed, if γ is a
positively oriented circle inside Ω centered at 0, then
Z
z −1 dz = 2πi
γ

and the corresponding integral for any complex polynomial equals 0. Note
that the domain Ω is not simply connected. Hence there is no complex
version of the Weierstrass approximation theorem. However, we have
Theorem 11.14 (Runge). If Ω ⊂ C is simply connected, then every holo-
morphic function on Ω can be uniformly approximated on compact subsets
of Ω by complex polynomials.

Proof. Since Ω is simply connected, every compact subset of Ω is contained

in a simply connected compact set K ⊂ Ω, so the complement C \ K is
connected. Let γ be a positively oriented Jordan curve in Ω such that K is
in the interior of γ. If f ∈ H(Ω), then
Z
1 f (ξ)
f (z) = dξ, z ∈ K .
2πi γ ξ − z
 FUNCTIONAL ANALYSIS 103

The integral on the right hand side can be uniformly approximated by Rie-
mann sums
Z m
1 f (ξ) 1 X f (ξk )
dξ ≈ ∆ξk
2πi γ ξ − z 2πi ξk − z
k=1
which is a linear combination of functions
1
z 7→ ∈ C(K) .
ξk − z
Thus it remains to prove that every function
1
(11.8) z 7→ ∈ C(K), ξ0 6∈ K
ξ0 − z
can be uniformly approximated on K by complex polynomials. Let M ⊂
C(K) be the closure of the subspace of complex polynomials. Suppose that
the function (11.8) does not belong to M . According to Theorem 11.6 there
is a functional x∗ ∈ C(K)∗ which vanishes on M and is nonzero on the
function (11.8), i.e.
hx∗ , p(z)i = 0, p any complex polynomial,

1
(11.9) hx∗ , i=
6 0.
ξ0 − z
The function
1
(11.10) ξ 7→ hx∗ , i
ξ−z
is complex differentiable in C \ K, so it is holomorphic in C \ K. If |ξ| > R =
maxz∈K |z|, then
∞
1 1 1 X zn
= =
ξ−z ξ 1 − z/ξ ξ n+1
n=0
and hence the function
1
z 7→
ξ−z
can be uniformly approximated in C(K) by complex polynomials, so
1
hx∗ , i=0 if |ξ| > R.
ξ−z
Since the function (11.10) is homomorphic in C \ K, C \ K is connected and
it vanishes for ξ sufficiently large
1
hx∗ , i=0 for all ξ ∈ C \ K
ξ−z
which contradicts (11.9). 2
104 PIOTR HAJLASZ

11.4. Separation of convex sets. Let X be a linear space over K.31

Definition. We say that a convex set W ⊂ X is absorbing if for each x ∈ X
there is ε > 0 such that the whole interval {tx : t ∈ [0, ε]} ⊂ W is contained
in W .32
We say that W is absorbing at a ∈ W if W − a is absorbing. Equivalently
if for every x ∈ X there is ε > 0 such that {a + tx : t ∈ [0, ε]} ⊂ W .
We say that W is balanced if αx ∈ W for all |α| ≤ 1 whenever x ∈ W .
Clearly 0 ∈ W if W is absorbing or balanced. If K = R, then W is
balanced if it is central symmetric, i.e. symmetric with respect to 0 and if
K = C, then W is balanced if the set contains the whole disc {αx : |α| ≤ 1}
in the complex plane generated by x, span {x}.
The above definitions are are motivated by the following obvious result:
Proposition 11.15. If p is a seminorm on X, then the unit ball W = {x :
p(x) < 1} is convex, balanced and absorbing at any point of W .

In Theorem 11.17 we will show that the conditions from the proposition
characterize the unit ball for a seminorm.
Definition. For each convex and absorbing set W ⊂ X the Minkowski
functional is n x o
µW (x) = inf s > 0 : ∈W .
s
Theorem 11.16. Let W ⊂ X be convex and absorbing. The Minkowski
functional has the following properties

(a) µW (x) ≥ 0;
(b) µW (x + y) ≤ µW (x) + µW (y);
(c) µW (tx) = tµW (x) for all t ≥ 0;
(d) µW (x) = 0 if and only if {tx : t ≥ 0} ⊂ W .
(e) If x ∈ W , then µW (x) ≤ 1;
(f) If µW (x) < 1, then x ∈ W
(g) If µW (x) > 1, then x 6∈ W .

Proof. All the properties but (b) are obvious. To prove (b) let x, y ∈ X
and s1 , s2 > 0 be such that x/s1 , y/s2 ∈ W . It follows from the convexity of
W that
x+y s1 x s2 y
= + ∈W
s1 + s2 s1 + s2 s1 s1 + s2 s2
Hence µW (x + y) ≤ s1 + s2 and the claim follows upon taking the infimum.
2
31As always K = R or K = C.
32Equivalently a convex set W ⊂ X is absorbing if X = S∞ nW .
n=1
 FUNCTIONAL ANALYSIS 105

Note that the properties (b) and (c) show that the Minkowski functional
is a Banach functional.
The following result provides a geometric characterization of sets that are
unit balls for norms and seminorms.
Theorem 11.17. If W ⊂ X is convex, balanced and absorbing at any point
of W , then there is a unique seminorm p such that W = {x ∈ W : p(x) < 1}.
Moreover p is a norm if and only if for each x 6= 0 there is t > 0 such that
tx 6∈ W .

Proof. Let p = µW . First we will prove that p is a seminorm. The inequal-

ity p(x + y) ≤ p(x) + p(y) follows from the previous result. Let now x ∈ X,
α ∈ K. If α = 0, then p(αx) = |α|p(x), so we can assume α 6= 0. If |α| = 1,
then for s > 0, x/s ∈ W if and only if αx/s ∈ W , because W is balanced
and hence
(11.11) p(αx) = p(x) for |α| = 1 .
If α 6= 0 is arbitrary, then
α 
p(αx) = p |α|x = p(|α|x) = |α|p(x)
|α|
by (11.11) and Theorem 11.16(c). This proves that p is a seminorm. Accord-
ing to Theorem 11.16(e,f)
(11.12) {x : p(x) < 1} ⊂ W ⊂ {x : p(x) ≤ 1} .
Let x ∈ W . Since W is absorbing at x there is t > 0 such that x+tx = y ∈ W
and thus x = (1 + t)−1 y. Since y ∈ W , p(y) ≤ 1, so
p(x) = (1 + t)−1 p(y) ≤ (1 + t)−1 < 1 .
This inequality together with (11.12) proves that
W = {x : p(x) < 1} .
Uniqueness is easy, because if q is another seminorm such that W = {x :
q(x) < 1}, then it is easy to show that
n x o
q(x) = inf s > 0 : ∈ W = p(x) .
s
Now it follows from condition (d) of Theorem 11.16 that p(x) is a norm, i.e.
p(x) 6= 0 for x 6= 0 is and only if for each x 6= 0 there is t > 0 such that
tx 6∈ W . 2
Now we will use the fact that the Minkowski functional is a Banach func-
tional and we will prove results about separation of convex sets known also
as geometric Hahn-Banach theorems.
Observe that the Hahn-Banach theorem in a form involving a Banach
functional applies only to real linear spaces. However, if X is a complex
106 PIOTR HAJLASZ

linear space, the real part of a functional is R-linear and we can still apply
the Hahn-Banach theorem (cf. Theorem 11.2).
Theorem 11.18. Let W1 , W2 be disjoint convex subsets of a normed space
X and assume that W1 is open. Then there is a functional x∗ ∈ X ∗ and
c ∈ R such that
re hx∗ , xi < c ≤ re hx∗ , yi
for all x ∈ W1 and y ∈ W2 .

Proof. It suffices to prove the theorem in the case of real normed spaces.
Indeed, if X is a complex normed space, then it can still be regarded as a
real normed space. If we can find a bounded R-linear functional x̃∗ such that
hx̃∗ , xi < c ≤ hx̃∗ , yi
for all x ∈ W1 and y ∈ W2 , then
hx∗ , xi = hx̃∗ , xi − ihx̃∗ , ixi
is a bounded C-linear functional that satisfies re hx∗ , xi = hx̃∗ , xi.
Henceforth we will assume that X is a real normed space. Fix x1 ∈ W1
and x2 ∈ W2 . Then the set
W = W1 − W2 + x2 − x1
| {z }
x0

is open, convex and 0 ∈ W . Since W is open it is absorbing and hence we

can define the Minkowski functional µW for W .
Define a functional x∗1 on the one dimensional space {αx0 : α ∈ R} by
hx∗1 , αx0 i = α .
Since the sets W1 and W2 are disjoint, x0 = x2 − x1 6∈ W . Thus µW (x0 ) ≥ 1
and hence it easily follows that
hx∗1 , αx0 i ≤ µW (αx0 )
for all α ∈ R. Now the fact that µW is a Banach functional allows us to
apply the Hahn-Banach theorem according to which there is a functional x∗
being an extension of x∗1 such that
hx∗ , xi ≤ µW (x) for all x ∈ X.
The functional x∗ is bounded because W is open. Indeed, if B(0, ε) ⊂ W ,
then for kxk < ε, hx∗ , xi ≤ µW (x) ≤ 1, so kx∗ k ≤ ε−1 .
For any x ∈ W1 , y ∈ W2 , x − y + x0 ∈ W and thus
µW (x − y + x0 ) < 1 ,
because W is open. Using hx∗ , x0 i = 1 we obtain
hx∗ , x − y + x0 i = hx∗ , xi − hx∗ , yi + 1 ≤ µW (x − y + x0 ) < 1 ,
 FUNCTIONAL ANALYSIS 107

so
(11.13) hx∗ , xi < hx∗ , yi for all x ∈ W1 , y ∈ W2 .
The sets x∗ (W1 ) and x∗ (W2 ) are intervals in R as convex sets. Moreover
x∗ (W1 ) is open, since W1 is open.33 Therefore (11.13) yields the existence of
c ∈ R such that x∗ (W1 ) ⊂ (−∞, c), x∗ (W2 ) ⊂ [c, ∞). 2
Theorem 11.19. Let W1 , W2 be disjoint convex subsets of a normed space
X. Assume that W1 is compact and W2 is closed. Then there is x∗ ∈ X ∗ ,
c ∈ R and ε > 0 such that
re hx∗ , xi ≤ c − ε < c ≤ re hx∗ , yi
for all x ∈ W1 and y ∈ W2 .

Proof. The distance between W1 and W2 is positive

dist (W1 , W2 ) = inf{d(x, y) : x ∈ W1 , y ∈ W2 } > 0 ,
because W1 is compact and W2 is closed. Hence there is δ > 0 such that
W1δ = W1 + B(0, δ) = {x + y : x ∈ W1 , kyk < δ}
is disjoint with W2 . Clearly W1δ is open and convex so we can apply previous
result to the pair of convex sets W1δ , W2 . Therefore
re hx∗ , xi < c ≤ re hx∗ , yi
for all x ∈ W1δ , y ∈ W2 and some c ∈ R. Since W1 is compact, the set
{re hx∗ , xi : x ∈ W1 }
is a compact subset of (−∞, c), so it is contained in (−∞, c − ε] for some
ε > 0. 2
Corollary 11.20. Let W be a closed convex subset of a real normed space
X. Then W is the intersection of all closed half-spaces {x : hx∗ , xi ≥ c},
x∗ ∈ X ∗ , c ∈ R that contain W .

Proof. For x 6∈ W we set W1 = {x}, W2 = W and apply Theorem 11.19.

2

11.5. Convex hull.

Definition. Let X be a vector space and E ⊂ X. The convex hull co(E) of
E is the intersection of all convex sets that contain E. In other words co(E)
is the smallest convex set that contains E.
If X is a normed space and E ⊂ X, then it is easy to see that the closure
co(E) is convex. This set is called the closed convex hull of E and it is obvious
that co(E) is the smallest closed convex set that contains E.
33x∗ (B(x, r)) is an open interval centered at hx∗ , xi.
108 PIOTR HAJLASZ

Theorem 11.21 (Mazur). If X is a Banach space and K ⊂ X is compact,

then co(K) is compact.

Proof. A metric space is compact if and only if it is complete and totally

bounded.34 Clearly co(K) is complete as a closed subset of a complete metric
space, and it remains to prove that it is totally bounded. Since in any metric
space the closure of a totally bounded set is totally bounded it suffices to
prove that co(K) is totally bounded. Fix ε > 0. Since K is compact, there
is a finite covering of K by balls of radius ε/2.
K ⊂ B(x1 , ε/2) ∪ . . . ∪ B(xn , ε/2) = {x1 , . . . , xn } + B(0, ε/2)
⊂ co({x1 , . . . , xn }) + B(0, ε/2) .
The set on the right hand side is convex, so
co(K) ⊂ co({x1 , . . . , xn }) + B(0, ε/2) .
Note that the set co({x1 , . . . , xn }) is totally bounded as a bounded set in a
finitely dimensional space span {x1 , . . . , xn } and hence it has a finite covering
by balls of radius ε/2
co({x1 , . . . , xn }) ⊂ {y1 , . . . , ym } + B(0, ε/2)
and hence
co(K) ⊂ {y1 , . . . , ym } + B(0, ε/2) + B(0, ε/2) = B(y1 , ε) ∪ . . . ∪ B(ym , ε)
which proves that co(K) is totally bounded. 2

12. Banach space valued integration

The existence of vector valued integrals is described in the next theorem.

Theorem 12.1. Let µ be a probability Borel measure on a compact metric
space E and let f : E → X be a continuous function with values in a Banach
space X. Then there is a unique element y ∈ X such that
Z
(12.1) hx∗ , yi = hx∗ , f i dµ for every x∗ ∈ X ∗ .
E
Moreover y ∈ co(f (E)).

Observe that z 7→ hx∗ , f (z)i is a continuous scalar valued function on E,

so the integrals on the right hand side of (12.1) are well defined.
The vector y from the above theorem is denoted by
Z
y= f dµ
E

34A metric space is totally bounded if for every ε > 0 there is a finite covering of the
space by balls of radius ε.
 FUNCTIONAL ANALYSIS 109

and hence we can rewrite (12.1) as

D Z E Z
∗
(12.2) x , f dµ = hx∗ , f i dµ .
E E
Moreover
Z
(12.3) f dµ ∈ co(f (E)) .
E
If µ is a finite Borel measure, then up to a constant factor it is a probability
measure and hence we can define the integrals of vector valued functions
for any finite measures. However, if µ is not a probability measure, then, in
general, the property (12.3) will be lost.
Proof. Uniqueness of y is obvious, because if hx∗ , y1 i = hx∗ , y2 i for all
x∗ ∈ X ∗ , then y1 = y2 , so we are left with the proof of the existence.
It follows from Mazur’s theorem that the set K = co(f (E)) is compact.
We have to find y ∈ K such that (12.1) is satisfied. For each finite subset
F ⊂ X ∗ let
n Z o
KF = y ∈ K : hx , yi = hx∗ , f i dµ for all x∗ ∈ F .
∗
E
KF is a closed subset of K, so it is compact. We have to prove that the
intersection of sets KF over all finite subsets F ⊂ X ∗ is nonempty. To this
end it suffices to prove that each of the sets KF in nonempty, because the
intersection of a finite number of sets KF is also a set of the same type and
the rest follows from the well known lemma.
Lemma 12.2. If {Ki }i∈I is a collection of compact subsets of a metric
T the intersection of every finite subcollection of {Ki }i∈I is
space Z such that
nonempty, then i∈I Ki is nonempty.
T
T Proof. Suppose i∈I S Ki = ∅. Let Gi = Z \ Ki . Fix j ∈ I. Then Kj ∩
i∈I K i = ∅, so K j ⊂ i∈I Gi . This is an open covering of a compact set Kj ,
so there is a finite subcovering Kj ⊂ Gi1 ∪. . .∪Gin , i.e. Kj ∩Ki1 ∩. . .∩Kin = ∅
which is a contradiction with our assumptions. 2
Let F = {x∗1 , . . . , x∗n } ⊂ X ∗ . We have to prove that KF 6= ∅, i.e. that
there is y ∈ K = co(f (E)) such that
Z
(12.4) hx∗i , yi = hx∗i , f i dµ for i = 1, 2, . . . , n.
E
Let
L : X → Kn , L(x) = hx∗1 , xi, . . . , hx∗n , xi

and let W = L(f (E)). Since L is linear the set L(K) is compact, convex
and it contains W , so co(W ) ⊂ L(K). It suffices to prove that
Z Z 
z = (z1 , . . . , zn ) := hx∗1 , f i dµ, . . . , hx∗n , f i dµ ∈ co(W ) .
E E
110 PIOTR HAJLASZ

Indeed, since co(W ) ⊂ L(K) it will imply that z = L(y) for some y ∈ K
which is (12.4).
Suppose z 6∈ co(W ). The sets {z} and co(W ) are convex, compact and
disjoint, so according to Theorem 11.19 there is a functional in (Kn )∗ that
separates the two sets, i.e. there is (c1 , . . . , cn ) ∈ Kn such that
n
X n
X
re ci ti < re ci zi for all t = (t1 , . . . , tn ) ∈ co(W ).
i=1 i=1
If t = L(f (s)), s ∈ E, we have
n
X n
X
re ci hx∗i , f (s)i < re ci zi .
i=1 i=1
Since µ is a probability measure, integration of the inequality over E gives
X n Z Xn
re ci hx∗i , f i dµ < re ci zi
i=1 E i=1
| {z }
zi
which is an obvious contradiction. The proof of the theorem is complete. 2
Theorem 12.3. Under the assumptions of Theorem 12.1
Z Z
f dµ ≤ kf k dµ .


E E
R
Proof. Let y = E f dµ. According to the Hahn-Banach theorem there is
x∗ ∈ X ∗ such that kx∗ k = 1 and hx∗ , yi = kyk. Hence
Z Z Z
f dµ = hx∗ , yi = hx∗ , f i dµ ≤ kf k dµ .


E E E
The proof is complete. 2

12.1. Banach space valued holomorphic functions.

Definition. Let X be a complex banach space and Ω ⊂ C an open set. We
say that f : Ω → X is (strongly) holomorphic if the limit
f (w) − f (z)
f 0 (z) = lim
w−z w→z
exists in the topology of X for every z ∈ Ω.
Holomorphic functions are continuous and hence we can integrate them
along curves in Ω. If γ is a positively oriented Jordan curve in Ω such that
the interior of the curve ∆γ is contained in Ω, then for every z ∈ ∆γ we
have the Cauchy formula
Z
1 f (ξ)
(12.5) f (z) = dξ .
2πi γ ξ − z
 FUNCTIONAL ANALYSIS 111

Indeed, if x∗ ∈ X ∗ , then Ω 3 z 7→ hx∗ , f (z)i ∈ C is a holomorphic function

and hence
hx∗ , f (ξ)i
Z Z
∗ 1 D
∗ 1 f (ξ) E
hx , f (z)i = dξ = x , dξ
2πi γ ξ − z 2πi γ ξ − z
by (12.2) and hence the Cauchy formula (12.5) follows.
By the same argument we prove that
Z
(12.6) f (z) dz = 0 .
γ

This easily implies that a substantial part of the theory of homomorphic

functions can be generalized to the vectorial case. In particular one can easily
show that holomorphic functions in a domain Ω ⊂ C can be represented
as a power series in any disc contained in Ω and a similar result hold for
Laurent expansions. That implies that holomorphic functions are infinitely
differentiable. the following result is an easy exercise.
Proposition 12.4. Let X be a complex Banach space, Y be a closed sub-
space and Ω ⊂ C be an open connected set. If f : Ω → X is holomorphic
and on a certain open subset of Ω it takes values into Y , then it takes values
into Y on all of Ω.

13. Reflexive spaces

If X is a Banach space, then X ∗∗ = (X ∗ )∗ is the dual to the dual space of

X, called the second dual space. Observe that every element x ∈ X defines
a bounded functional on X ∗ by the formula
x∗ 7→ hx∗ , xi .
Denote this functional by κ(x) ∈ X ∗∗ , i.e.
hκ(x), x∗ i = hx∗ , xi .
Observe that
kκ(x)k = sup hx∗ , xi = kxk
x∗ ∈X ∗
kx∗ k≤1

by Corollary 11.5. Therefore the canonical embedding

κ : X → X ∗∗
is an isometrical isomorphism between X and a closed subspace of X ∗∗ .
Definition. We say that a Banach space X is reflexive if κ(X) = X ∗∗ .
Warning. James constructed a nonreflexive space X such that X is iso-
metrically isomorphic to X ∗∗ .
112 PIOTR HAJLASZ

Theorem 13.1. Let µ be a σ-finite measure and 1 < p, q < ∞, p−1 + q −1 =

1. Then the space Lp (µ) is reflexive.

Proof.35 The space (Lp (µ))∗ is isometrically isomorphic to Lq (µ) by The-

orem 2.13. If f ∈ Lp (µ), then κ(f ) defines a functional on (Lp )∗ = Lq by
Z
hκ(f ), gi = gf dµ g ∈ Lq (µ) .
X
Since every functional on (Lp )∗
= Lq is of that form36 we obtain that
p p ∗∗
κ(L (µ)) = (L (µ)) and hence Lp (µ) is reflexive. 2
The canonical embedding κ : `1 → (`1 )∗∗ = (`∞ )∗ gives an isometric
isomorphism between `1 and a closed subspace of (`∞ )∗ , but as we know κ
is not surjective, so `1 is not reflexive.
Theorem 13.2. Any Hilbert space is reflexive.

Proof. According to the Riesz representation theorem (Theorem 5.5) for

every x∗ ∈ H ∗ there is a unique element T (x∗ ) ∈ H such that
hx∗ , xi = h x, T (x∗ )ii for x ∈ H ,
where on the left hand side we have evaluation of the functional and inner
product on the right hand side.37 Moreover kx∗ kH ∗ = kT (x∗ )kH .
If x ∈ H, then κ(x) is a functional on H ∗ defined by
(13.1) hκ(x), x∗ i = hx∗ , xi = h x, T (x∗ )ii
and we want to show that every functional on H ∗ is of that form. The
mapping T : H ∗ → H is bounded one-to-one and onto, however, if H is a
complex Hilbert space, T is not quite linear, because
T (x∗1 + x∗2 ) = T (x∗1 ) + T (x∗2 ) but T (αx∗ ) = αT (x∗ ) .
The space H ∗ is a Hilbert space with the inner product
h x∗1 , x∗2i := h T (x∗2 ), T (x∗1 )ii .
Now if x∗∗ ∈ H ∗∗ , then it follows from the Riesz representation theorem
that there is z ∗ ∈ H ∗ such that
hx∗∗ , x∗ i = h x∗ , z ∗i = h T (z ∗ ), T (x∗ )ii ,
and hence
hx∗∗ , x∗ i = hκ(T (z ∗ )), x∗ i
by (13.1). The proof is complete. 2
35We can briefly summarize the proof as follows: (Lp )∗ = Lq , (Lp )∗∗ = (Lq )∗ = Lp , so
the second dual to Lp is Lp itself and hence Lp is reflexive.
36Again by Theorem 2.13.
37We use thick notation for the inner product to distinguish it from the notation for
the functional.
 FUNCTIONAL ANALYSIS 113

Theorem 13.3. A closed subspace of a reflexive space is reflexive.

Proof. Let Y be a closed subspace of a reflexive space X. We shall prove

that Y is reflexive.
Every functional x∗ ∈ X ∗ defines a functional on Y as a restriction to Y ,
σ : X ∗ → Y ∗, hσ(x∗ ), yi = hx∗ , yi .
Clearly
kσ(x∗ )k ≤ kx∗ k, so σ(x∗ ) ∈ Y ∗
Fix y0∗∗ ∈ Y ∗∗ . We have to prove that there is x0 ∈ Y such that
(13.2) hy0∗∗ , y ∗ i = hy ∗ , x0 i for y ∗ ∈ Y ∗ .
Let τ (y0∗∗ ) ∈ X ∗∗ be defined by38
hτ (y0∗∗ ), x∗ i = hy0∗∗ , σ(x∗ )i .
Since
|hτ (y0∗∗ ), x∗ i| ≤ ky0∗∗ k kx∗ k
we have τ (y0∗∗ ) ∈ X ∗∗ . The canonical embedding κ : X → X ∗∗ is an isomor-
phism39 and hence it is invertible κ−1 : X ∗∗ → X. Thus
x0 = κ−1 (τ (y0∗∗ )) ∈ X ,
but we will prove that x0 ∈ Y . Note that
hτ (y0∗∗ ), x∗ i = hκ(x0 ), x∗ i = hx∗ , x0 i .
Suppose x0 6∈ Y . According to Theorem 11.6 there is x∗ ∈ X ∗ such that
hx∗ , x0 i =
6 0, hx∗ , yi = 0 for all y ∈ Y
and hence σ(x∗ ) = 0, so
0 = hy0∗∗ , σ(x∗ )i = hτ (y0∗∗ ), x∗ i = hx∗ , x0 i =
6 0
which is a contradiction. We proved that
x0 = κ−1 (τ (y0∗∗ )) ∈ Y .
For y ∗ ∈ Y ∗ let x∗ ∈ X ∗ be an extension of y ∗ , so y ∗ = σ(x∗ ). We have
x ∈Y
hy0∗∗ , y ∗ i = hy0∗∗ , σ(x∗ )i = hτ (y0∗∗ ), x∗ i = hx∗ , x0 i 0= hy ∗ , x0 i
which proves (13.2). 2
Theorem 13.4. A Banach space isomorphic to a reflexive space is reflexive.

38Roughly speaking Y ⊂ X, X ∗ ⊂ Y ∗ , Y ∗∗ ⊂ X ∗∗ , so every element in Y ∗∗ can be

regarded as an element in X ∗∗ .
39Because X is reflexive.
114 PIOTR HAJLASZ

Proof. Let T : X → Y be an isomorphism of Banach spaces and let X be

reflexive. Let y0∗∗ ∈ Y ∗∗ . We have to prove that there is y0 ∈ Y such that
(13.3) hy0∗∗ , y ∗ i = hy ∗ , y0 i for y ∗ ∈ Y ∗ .
The mapping T ∗ : Y ∗ → X ∗ defined by hT ∗ y ∗ , xi = hy ∗ , T xi is an isomor-
phism and hence it is invertible (T ∗ )−1 : X ∗ → Y ∗ . Let x∗∗
0 ∈X
∗∗ be defined

by
hx∗∗ ∗ ∗∗ ∗ −1 ∗
0 , x i = hy0 , (T ) x i .
Since X is reflexive, there is x0 ∈ X such that
hx∗∗ ∗ ∗ ∗
0 , x i = hx , x0 i for x ∈ X
∗

and hence
hy0∗∗ , y ∗ i = hy0∗∗ , (T ∗ )−1 (T ∗ y ∗ )i = hx∗∗ ∗ ∗
0 ,T y i
= hT ∗ y ∗ , x0 i = hy ∗ , T x0 i .
Thus (13.3) holds with y0 = T x0 ∈ Y . 2
Corollary 13.5. Finitely dimensional Banach spaces are reflexive.
Theorem 13.6. X is reflexive if and only if X ∗ is reflexive.

Proof. Suppose that X is reflexive. In order to prove reflexivity of X ∗ we

have to show that for every
x∗∗∗
0 ∈ X ∗∗∗ = (X ∗ )∗∗ = (X ∗∗ )∗
there is x∗0 ∈ X ∗ such that
hx∗∗∗ ∗∗ ∗∗ ∗
0 , x i = hx , x0 i for x
∗∗ ∈ X ∗∗ .

Since X is reflexive, κ : X → X ∗∗ is an isomorphism and hence if is invert-

ible. Note that the composition
κ x∗∗∗
X −→ X ∗∗ −→
0
K
defines x∗0 = x∗∗∗ ∗
0 ◦ κ ∈ X and hence
hx∗∗∗ ∗∗ ∗∗∗ −1 ∗∗ ∗∗∗ −1 ∗∗
0 , x i = hx0 , κ(κ (x ))i = hx0 ◦ κ, κ (x )i
= hx∗0 , κ−1 (x∗∗ )i = hκ(κ−1 (x∗∗ )), x∗0 i
= hx∗∗ , x∗0 i ,
where in the second to last inequality we employed the definition of κ ac-
cording to which hκ(x), x∗ i = hx∗ , xi. This proves reflexivity of X ∗ .
Now if X ∗ is reflexive, then by what we already proved X ∗∗ is reflexive.
The canonical embedding κ : X → X ∗∗ gives an isomorphism between X
and a closed subspace of the reflexive space X ∗∗ . Hence reflexivity of X
follows from Theorem 13.3 and Theorem 13.4. 2
Theorem 13.7. If X is reflexive and M ⊂ X is a closed subspace, then the
quotient space X/M is reflexive.
 FUNCTIONAL ANALYSIS 115

Proof. The space X ∗ is reflexive by Theorem 13.6. The space (X/M )∗

is isomorphic to the annihilator M ⊥ by Theorem 4.7. Since M ⊥ is a closed
subspace of X ∗ it is reflexive by Theorem 13.3 and thus Theorem 13.4 implies
reflexivity of (X/M )∗ . This and Theorem 13.6 yields reflexivity of X/M . 2
Corollary 13.8. If T : X → Y is a bounded linear surjection of a reflexive
space X onto a Banach space Y , then Y is reflexive.

Proof. The space Y is isomorphic to X/ ker T and hence it is reflexive as

a consequence of Theorem 13.7 and Theorem 13.4. 2

14. Weak convergence

Definition. Let X be a normed space. We say that a sequence {xn }∞

n=1 ⊂ X
converges weakly to x0 ∈ X if for every functional x∗ ∈ X ∗
hx∗ , xn i → hx∗ , x0 i as n → ∞.
We denote weak convergence by xn * x0 .
Clearly if xn → x0 in norm, then xn * x0 .
Theorem 14.1. Let X be a normed space. If xn * x0 in X, then
sup kxn k < ∞
n
and
(14.1) kx0 k ≤ lim inf kxn k .
n→∞

Proof. The canonical embedding defines κ(xn ) ∈ X ∗∗ by

hκ(xn ), x∗ i = hx∗ , κ(xn )i .
Observe that X ∗ is a Banach space, even if X is only a normed space. Since
hκ(x0 ), x∗ i = hx∗ , x0 i = lim hx∗ , xn i = lim hκ(xn ), x∗ i ,
n→∞ n→∞
the sequence κ(xn ) of functionals on X∗
is pointwise convergent to κ(x0 ).
Therefore the Banach-Steinhaus theorem (Corollary 9.5) yields
sup kxn k = sup kκ(xn )k < ∞
n n
and
kx0 k = kκ(x0 )k ≤ lim inf kκ(xn )k = lim inf kxn k .
n→∞ n→∞
The proof is complete. 2
Note that we can prove (14.1) as a direct consequence of the Hahn-Banach
theorem. Indeed, let x∗ ∈ X ∗ be such that kx∗ k = 1, hx∗ , x0 i = kx0 k. Then
kx0 k = hx∗ , x0 i = lim hx∗ , xn i ≤ lim inf kxn k .
n→∞ n→∞
116 PIOTR HAJLASZ

In the previous proof instead of the Hahn-Banach theorem we employed the

Banach-Steinhaus theorem, but the Hahn-Banach theorem was still there –
it was needed for the proof of the equality kxk = kκ(x)k.
Weak convergence is, in general, much weaker than the convergence in
norm. For example if
xn = (xni ) = (0, 0, . . . , 0, 1, 0, . . .) ∈ `p , 1 < p < ∞,

i.e. xni = δni , then kxn − xm kp = 21/p and hence no subsequence of xn can
be convergent in the norm. However, xn is weakly convergence to 0. Indeed,
according to Theorem 2.12 for every x∗ ∈ (`p )∗ there is s = (si ) ∈ `q such
that
∞
X
hx∗ , xn i = si xni = sn
i=1

and clearly hx∗ , x

n i = sn → 0 as n → ∞. Note also that the same sequence
is not weakly convergent in `1 . Indeed, if we take s = (1, −1, 1, −1, 1 . . .) ∈
`∞ = (`1 )∗ , then hx∗ , xn i = ±1 is not convergent. Actually in `1 weak
convergence is equivalent with the convergence in norm.
Theorem 14.2 (Schur). In the space `1 a sequence is weakly convergent if
and only if it is convergent in norm.

Proof. Suppose that xn * x0 weakly in `1 . Then yn = xn − x0 * 0 and

we have to prove that kyn k1 → 0. By contradiction suppose that kyn k1 does
not converge to 0. Then there is ε > 0 and a subsequence (still denoted by
yn ) such that kyn k1 ≥ ε. Denote yn = (ynk ). Note that for every integer p
p
X
(14.2) |ynk | → 0 as n → ∞
k=1

because of weak convergence to 0. Hence for each p we can find n so large

that that the sum at (14.2) is less than ε/4. On the other handPthe sum of the
series is greater or equal to ε, so we can find q > p such that qk=p+1 |ynk | >
3ε/4 and ∞
P
k=q |ynk | < ε/4. Hence using induction we can find two sequences
of integers {ni } and {pi } such that
pi pi+1 ∞
X ε X 3ε X ε
|yni k | < , |yni k | > , |yni k | < .
4 4 4
k=1 k=pi +1 k=pi+1

Now we need to define a functional on which we will evaluate the sequence.

For k ≤ p1 we set sk = 0 and for pi < k ≤ pi+1 , i = 1, 2, . . . we define
( y
ni k

sk = sgn (yni k ) = |yni k | if yni k 6= 0,

0 if yni k = 0.
 FUNCTIONAL ANALYSIS 117

Then x∗ = (sk ) ∈ `∞ = (`1 )∗ is a bounded functional. We have

pi+1
∗
X ε ε ε
hx , yni i ≥ |yni k | − − >
4 4 4
k=pi +1

so hx∗ , yni i does not converge to 0. Contradiction. 2

Weak convergence in the space of continuous functions is descried in the
following result.
Theorem 14.3. Let X be a compact metric space. Then a sequence fn ∈
C(X) converges weakly in C(X) to f ∈ C(X) if and only if there is M > 0
such that

(a)
|fn (x)| ≤ M for all x ∈ X and n = 1, 2, 3 . . .

and the sequence converges pointwise

(b)
fn (x) → f (x) for all x ∈ X.

Proof. Suppose fn * f . Then condition (a) follows from Theorem 14.1.

Since for every x ∈ X the value at the point x, i.e. g 7→ g(x) is a bounded
functional on C(X), the pointwise convergence (b) follows from the definition
of weak convergence.
Now suppose that both conditions (a) and (b) are satisfied. According to
the Riesz representation theorem for every functional Φ ∈ C(X)∗ there is a
signed Borel measure µ of finite total variation such that
Z
hΦ, gi = g dµ
X

and hence hΦ, fn i → hΦ, f i by the dominated convergence theorem. 2

While weak convergence is weaker than the convergence in norm, the
following result shows that a sequence of convex combinations of a weakly
convergent sequence converges in norm.
Theorem 14.4 (Mazur’s lemma). Let X be a normed space and let xn * x0
weakly in X. Then un → x0 in norm for some sequence un of the form
N (n)
X
un = ank xk
k=n
PN (n)
where ank ≥ 0, k=n ank = 1.
118 PIOTR HAJLASZ

Proof. Let Wn be the closure of the set of all convex combinations of

elements of the sequence {xn , xn+1 , xn+2 . . .}. It remains to prove that x0 ∈
PN (n)
Wn , because it will imply existence of coefficients ank ≥ 0, k=n ank = 1
such that
NX(n) 1
ank xk − x0 < .

n

k=n

Suppose x0 6∈ Wn . The set Wn is convex and closed and the set W 0 = {x0 }
is convex and compact. According to Theorem 11.19 there is x∗ ∈ X ∗ , c ∈ R
and ε > 0 such that
re hx∗ , x0 i ≤ c − ε < c ≤ re hx∗ , xi
for all x ∈ Wn . In particular
re hx∗ , x0 i ≤ c − ε < c ≤ re hx∗ , xi i for i = n, n + 1, . . .
which contradicts weak convergence xi * x0 . 2
Definition. Let X be a normed space. A sequence of functionals {x∗n } ⊂ X ∗
converges weakly-∗ to x∗0 ∈ X ∗ if for every x ∈ X
hx∗n , xi → hx∗0 , xi as n → ∞.
∗
We denote weak-∗ convergence by x∗n * x∗0 .
∗
Theorem 14.5. If x∗n * x∗0 , then
sup kx∗n k < ∞ and kx∗0 k ≤ lim inf kx∗n k .
n n→∞

Proof. It is an immediate consequence of Corollary 9.5. 2

The following result is a special case of the Banach-Alaoglu theorem.
Theorem 14.6 (Banach-Alaoglu, separable case). Let X be a separable
normed space. Then every bounded sequence in X ∗ has a weakly-∗ convergent
subsequence.

Proof. Let {x∗n }n ⊂ X ∗ be a bounded sequence and let {x1 , x2 , . . .} ⊂ X

be a countable and dense set. For each i, hx∗n , xi i is a bounded sequence of
scalars so it has a convergent subsequence. Using the diagonal method we
find a subsequence {x∗nk }k such that for each i = 1, 2, 3, . . . the sequence
hx∗nk , xi i is convergent. Thus {x∗nk }k is a bounded sequence of functionals
that converges on a dense subset of X and it easily follows40 that hx∗nk , xi
converges for every x ∈ X. 2

40Compare with an argument that follows (9.2).

 FUNCTIONAL ANALYSIS 119

Corollary 14.7. Let X be a locally compact metric space and let {µn }n be
a sequence of finite positive Borel measures on X such that
sup µn (X) < ∞ .
n
Then there is a subsequence µnk and a finite Borel measure µ on X such
that
µ(X) ≤ lim inf µnk (X)
k→∞
and for every f ∈ C0 (X)
Z Z
f dµnk → f dµ .
X X

Proof. The measures µn define bounded functionals in C0 (X)∗

Z
hΦn , f i = f dµ
X
with
sup kΦn k = sup µn (X) < ∞ ,
n n
see Theorem 2.15. Since the space C(X) is separable41, there is a weakly-∗
∗
convergent subsequence Φnk * Φ. Clearly µ is represented by a signed Borel
measure of finite total variation and
Z Z
(14.3) f dµnk → f dµ for all f ∈ C0 (X).
X X
Since the measures µn are positive it easily follows from (14.3) that µ is
positive and Theorem 14.5 yields
µ(X) = kΦk ≤ lim inf kΦnk k = lim inf µnk (X) .
k→∞ k→∞
The proof is complete. 2
In the situation described in the above result we say that the measures
∗
µnk converge weakly-∗ to the measure µ, µnk * µ.42
If X is a reflexive space, then X is isometric to the dual space to X ∗ ,
so the weak convergence is equivalent with the weak-∗ convergence. Note
also that for reflexive spaces, X is separable if and only if X ∗ is separable
(Theorem 11.7) and hence Theorem 14.6 shows that a bounded sequence in
a separable reflexive space has a weakly convergent subsequence. However,
this result is also true without assuming separability.
Theorem 14.8. A bounded sequence in a reflexive space has a weakly con-
vergent subsequence.
41Why?
42Sometimes in the literature this convergence is called weak convergence of measures
µnk * µ, but we prefer to call it weak-∗ convergence to be consistent with the language
of functional analysis.
120 PIOTR HAJLASZ

Proof. Let {xn }n ⊂ X be a bounded sequence in a reflexive space X. Let

X0 = span {x1 , x2 , . . .} .
Then X0 is a closed separable subspace of X, so X0 is reflexive. Since the
dual space to X0∗ is separable as isometric to X0 , we conclude that X0∗ is
separable (Theorem 11.7). Let {x∗i }i ⊂ X0∗ be a countable and dense subset.
By the diagonal argument there is a subsequence {xnk }k such that for every
i = 1, 2, 3, . . .
hx∗i , xnk i is convergent as k → ∞.
The density of {x∗i }i in X0∗ easily implies that43
hx∗ , xnk i is convergent for every x∗ ∈ X0∗ .
Observe that every functional x∗ defines a functional in X0∗ as a restriction
to X0 , so
hx∗ , xnk i is convergent for every x∗ ∈ X ∗ .
If κ : X → X ∗∗ is the canonical embedding, then the limit
lim hκ(xnk ), x∗ i = lim hx∗ , xnk i := hx∗∗ , x∗ i
k→∞ k→∞
defines an element x∗∗
∈ X ∗∗ .
Indeed, linearity of x∗∗ is obvious and bound-
edness follows from the fact that
sup kκ(xnk )k = sup kxnk k < ∞ .
k k

Now it is clear that xnk * x = κ−1 (x∗∗ ). Indeed, for x∗ ∈ X ∗

hx∗ , xnk i → hx∗∗ , x∗ i = hx∗ , xi .
The proof is complete. 2
Corollary 14.9. Let mu be a σ-finite measure and 1 < p < ∞. Then every
bounded sequence {fn }n ⊂ Lp (µ) has a weakly convergence subsequence, i.e.
there is a subsequence {fnk }k and f ∈ Lp (µ) such that for every g ∈ Lq (µ),
p−1 + q −1 = 1 Z Z
fnk g dµ → f g dµ as k → ∞.
X X

The following result generalizes Theorem 5.1 to any reflexive space, see
also an example that follows Theorem 5.1.
Theorem 14.10. Every nonempty, convex and closed set E in a reflexive
space X contains an element of smallest norm.

Proof. Let {xn } ⊂ E be a sequence such that kxn k → inf x∈E kxk. Hence
the sequence {xn } is bounded, so it has a weakly convergent subsequence

43Verify the Cauchy condition.

 FUNCTIONAL ANALYSIS 121

xn * x. By Mazur’s lemma a sequence of convex combinations of xnk con-

verge to x in norm. Since convex combinations belong to E we conclude that
x ∈ E. Now Theorem 14.1 yields
kxk ≤ lim inf kxnk k = inf kxk
k→∞ x∈E

and the claim follows. 2

In the case of reflexive spaces the Riesz lemma (Theorem 3.6) has the
following stronger version.44
Theorem 14.11. Let X0 6= X be a closed linear subspace of a reflexive
space X. Then there is y ∈ X such that
kyk = 1 and ky − xk ≥ 1 for all x ∈ X0 .

Proof. Fix y0 ∈ X \ X0 and let M = X0 − y0 . Since M is convex and

closed, Theorem 14.10 yields the existence of z0 ∈ M such that
kz0 k = inf kzk .
z∈M

Hence z0 = x0 − y0 , x0 ∈ X0 satisfies
kx0 − y0 k = inf kx − y0 k
x∈X0

and the proof of the Riesz lemma shows that the vector y = (y0 − x0 )/ky0 −
x0 k satisfies the claim. 2

14.1. Direct methods in the calculus of variations. In this section we

will show an application of weak convergence to an abstract approach to
existence of minimizers of variational problems.
Let I : X → R be a function defined on a normed space X. Henceforth
I will be called functional even if it is not linear. Actually in all interesting
cases it will not be linear.
The problem is to find reasonable conditions that will guarantee existence
of x̃ ∈ X such that
(14.4) I(x̃) = inf I(x)
x∈X

An element x̃ ∈ X satisfying (14.4) is called minimizer of I and a problem

of finding a minimizer is called a variational problem.
Definition. We say that a functional I is sequentially weakly lower semi-
continuous (swlsc) if for every sequence xn * x weakly convergent in X,
I(x) ≤ lim inf I(xn ) .
n→∞

44Compare with Corollary 3.8 and an example that follows it.

122 PIOTR HAJLASZ

We say that I is coercive if

kxn k =⇒ I(xn ) → ∞ .
Theorem 14.12. If X is a reflexive Banach space and I : X → R is swlsc
and coercive, then there is x̃ ∈ X such that
I(x̃) = inf I(x) .
x∈X

Proof. Let xn ∈ X be a sequence such that

I(xn ) → inf I(x) .
x∈X
Such a sequence is called a minimizing sequence. Coercivity of I implies that
xn is bounded in X. Since the space is reflexive, xn has a weakly convergent
subsequence
xnk * x̃ in X
and the sequential weak lower semicontinuity yields
I(x̃) ≤ lim inf I(xnk ) = inf I(x) .
k→∞ x∈X
The proof is complete. 2
In general the swlsc condition is very difficult to check, because it does
not follow from continuity of I.
Definition. A functional I : X → R is called convex if
I(tx + (1 − t)y) ≤ tI(x) + (1 − t)I(y) for x, y ∈ X, t ∈ [0, 1]
and strictly convex if
I(tx + (1 − t)y) < tI(x) + (1 − t)I(y) for x, y ∈ X, x 6= y, t ∈ (0, 1).
We say that I is lower semicontinuous if
xn → x =⇒ lim inf I(xn ) .
n→∞

The lower semicontinuity is much weaker than swlsc, because we require

convergence of xn to x in norm. In particular if I is continuous, then it is
lower semicontinuous.
Theorem 14.13. If X is a normed space and I : X → R is convex and
lower semicontinuous, then I is swlsc.

Proof. We have to prove that

xn * x =⇒ I(x) ≤ lim inf I(xn ) .
n→∞
45
We can assume that I(xn ) has a limit. Denote the limit by g,
lim I(xn ) = g .
n→∞

45Otherwise we choose a subsequence x

nk such that I(xnk ) → lim inf n→∞ I(xn ).
 FUNCTIONAL ANALYSIS 123

It follows from Mazur’s lemma that for some sequence

N (n) N (n)
X X
un = ank xk , ank ≥ 0, ank = 1
k=n k=n
of convex combinations of xk , un → x in norm. Now lower semicontinuity
and coercivity of I yields
 NX
(n) 
I(x) ≤ lim inf I(un ) = lim inf I ank xk
n→∞ n→∞
k=n
N (n)
X
≤ lim inf ank I(xk ) = g .
n→∞
k=n
The proof is complete. 2
As a corollary we obtain a result of fundamental importance in the convex
calculus of variations.
Theorem 14.14 (Mazur-Schauder). If I : X → R is a convex, lower semi-
continuous and coercive functional defined on a reflexive Banach space X,
then I attains minimum in X, i.e. there is x̃ ∈ X such that
I(x̃) = inf I(x) .
x∈X
If in addition I is strictly convex, then the minimizer x̃ is unique.

15. Weak topology

The following result which is interesting on its own will be useful later.
Proposition 15.1. Let s be be the space of all (real or complex) sequences
x = (xk )∞
k=1 . The space s with the metric
∞
X |xk − yk |
d(x, y) = 2−k
1 + |xk − yk |
k=1
is a complete metric space. Moreover for each sequence of positive numbers,
r1 , r2 , . . . > 0 the set
K = {x ∈ s : |xk | ≤ rk for k = 1, 2, . . .}
is compact in (s, d).

Proof. The fact that d is a metric directly follows from an elementary

inequality
a+b 1 b
≤ + a, b ≥ 0.
1+a+b 1+a 1+b
Observe that d is the metric of convergence on each coordinate, i.e.
xn = (xnk )∞ ∞
k=1 → x = (xk )k=1
124 PIOTR HAJLASZ

if and only if xnk → xk for each k = 1, 2, 3 . . . If xn = (xnk )∞

k=1 is a Cauchy
sequence, them for each k, xnk is a Cauchy sequence of numbers, so it is
convergent. That means xn converges on each of its coordinates and thus it
converges in the metric d. this proves that (s, d) is complete.
If xk ∈ K, for n = 1, 2, 3 . . ., then for each n, the sequence xnk is bounded,
so it has a convergent subsequence. Using the diagonal argument we can find
a subsequence xni that converges on each of its coordinates, so it converges
in the metric d. Thus K is compact. 2
As an application we have
Theorem 15.2. If X is a separable Banach space, then there is a metric ρ
on X ∗ such that for every bounded set E ⊂ X ∗ and {x∗n } ⊂ E, x∗ ∈ E,
∗
x∗n * x∗ if and only if ρ(x∗n , x∗ ) → 0.

That means for each bounded set E ⊂ X ∗ , the metric ρE being the
restriction of ρ to E is such that the convergence in ρE is equivalent to the
weak-∗ convergence in E, i.e. the weak-∗ convergence in bounded subsets of
X ∗ is metrizable. This is true in bounded sets only as the weak-∗ convergence
in X ∗ , dim X = ∞ is not metrizable (see Theorem 15.4).
The following is a version of the separable Banach-Alaoglu theorem (The-
orem 14.6).
Corollary 15.3. If X is a separable Banach space, then every closed ball
B = {x∗ ∈ X ∗ : kx∗ k ≤ r}
is compact in the metric ρB .

Proof. If x∗n ∈ B, then we can find a weakly-∗ convergent subsequence

∗
x∗ni *x∗ by Theorem 14.6. It follows then from Theorem 14.5 that x∗ ∈ B
and hence ρB (x∗ni , x∗ ) → 0. 2
Proof of Theorem 15.2. Let {x1 , x2 , . . .} ⊂ B(0, 1) ⊂ X be a dense subset
of the unit ball in X. Consider the mapping
Φ : X ∗ → s, Φ(x∗ ) = (hx∗ , x1 i, hx∗ , x2 i, . . .) .
The mapping is one-to-one. Indeed, if hx∗ , xi i = hy ∗ , xi i for all i, then hx∗ −
y ∗ , xi = 0 for all x ∈ X, kxk ≤ 1 by the density argument and hence x∗ = y ∗ .
Thus
ρ(x∗ , y ∗ ) = d(Φ(x∗ ), Φ(y ∗ ))
is a metric in X ∗ . It remains to prove that for every bounded set E ⊂ X ∗ ,
{x∗n } ⊂ E and x∗ ∈ E,
∗
x∗n * x∗ if and only if d(Φ(x∗n ), Φ(x∗ )) → 0.
 FUNCTIONAL ANALYSIS 125

∗
If x∗n * x∗ , then hx∗n , xi i → hx∗ , xi i for every i, so every coordinate
of Φ(x∗n ) converges to the corresponding coordinate of Φ(x∗ ) and hence
d(Φ(x∗n ), Φ(x∗ )) → 0.

Conversely, if d(Φ(x∗n ), Φ(x∗ )) → 0, then coordinates of Φ(x∗n ) converge

to coordinates of Φ(x∗ ), i.e.

(15.1) hx∗n , xi i → hx∗ , xi i

and since the sequence x∗n is bounded

(15.2) hx∗n , xi → hx∗ , xi

∗
for every x with kxk ≤ 1 and hence for every x ∈ X, i.e. x∗n * x∗ . 2

Note that the metric ρ is defined by an explicit formula

∞
X |hx∗ − y ∗ , xk i|
ρ(x∗ , y ∗ ) = 2−k ,
1 + |hx∗ − y ∗ , xk i|
k=1

where {x1 , x2 , . . .} ⊂ B(0, 1) is a dense subset.

The assumption that E is bounded was employed only once in the proof
of the implication from (15.1) to (15.2) and the boundedness was a crucial
assumption here as the following result shows.

Theorem 15.4. Let X be a separable Banach space, dim X = ∞. Then

∗
there is no metric d in X ∗ such that x∗n * x∗ if and only if d(x∗n , x∗ ) → 0.

Proof. Let {x1 , x2 , . . .} ⊂ X be a dense subset and let

Xn = span {x1 , x2 , . . . , xn }

By the Hahn-Banach theorem there is a functional x∗n ∈ X ∗ such that kx∗n k =

∗
1 and hx∗n , xi = 0 for x ∈ Xn . Note that x∗n * 0. Indeed, if x ∈ X, then for
every ε > 0 there is m such that

kx − xm k < ε .

Hence for n > m we have

|hx∗n , xi| = |hx∗n , x − xm i| < ε .

∗
Similarly for every k > 0, kx∗n * 0 as n → ∞. Suppose now that there is a
metric d as in the statement on the theorem. Then d(kx∗n , 0) → 0 as n → ∞.
∗
Hence for each k we can find n(k) such that d(kx∗n(k) , 0) < 1/k, so kx∗n(k) * 0
as k → ∞. However, kkx∗n(k) k = k and we arrive to a contradiction with the
fact that the weakly-∗ convergent sequence is bounded, Theorem 14.5. 2
126 PIOTR HAJLASZ

15.1. Topological spaces. In addition to the norm topology Banach spaces

are equipped with so called weak topology and dual spaces with weak-∗
topology. These topologies do not come from metric, so we have to introduce
topological spaces.
Definition. A topological space is a set X with a family T of subsets of X
called open sets that satisfy the following properties

(a) ∅, X ∈ T ;
(b) T is closed under finite intersections, i.e. if U, V ∈ T , then US∩V ∈ T ;
(c) T is closed under arbitrary unions, i.e. if {Ui }i∈I ⊂ T , then i∈I Ui ∈
T.

The family T is called topology in X.

Let X0 ⊂ X be a subset of a topological space. Then
T0 = {U ∩ X0 : U ∈ T }
is the induced topology in X0 . Obviously (X0 , T0 ) is a topological space.
We say that E ⊂ X is closed is X \ E is open, i.e. X \ E ∈ T .
If A ⊂ X, the closure of A denoted by A is the intersection of all closed
sets that contain A, i.e. it is the smallest closed set that contain A. Clearly,
if x ∈ A then for every open set U such that x ∈ U we have U ∩ A 6= ∅.
A topological space (X, T ) is called Hausdorff if for every x, y ∈ X, x 6= y
there are opens sets U, V ∈ T such that x ∈ U , y ∈ V , U ∩ V = ∅.
Example. If (X, d) is a metric space, then the family of all open sets is a
Hausdorff topology in X.
Definition. Let (X, T ), (Y, F) be two topological spaces. We say that a
mapping f : X → Y is continuous if preimages of open sets are open, i.e.
f −1 (U ) ∈ T whenever U ∈ F.
We say that a topology T1 is weaker than a topology T2 if T1 ⊂ T2 . In this
situation T2 is called a stronger topology. Clearly if a mapping is continuous
with respect to T1 , then it is also continuous with respect to the stronger
topology T2 .
A family B ⊂ T is called a base if every U ∈ T is a union of elements of
B.
Example. If (X, d) is a metric space, then the family of all open balls is a
base for the topology generated by the metric.
Definition. We say that a topological space is compact if every open cov-
ering has a finite subcovering. A subset of a topological space is compact if
it is compact with respect to the induced topology.
 FUNCTIONAL ANALYSIS 127

Proposition 15.5.

(a) Let E be a closed subset of a compact space. Then E is compact with

respect to the induced topology.
(b) Let E be a compact subset of a Hausdorff topological space. Then E
is closed.46
(c) A continuous image of a compact set is compact.
(d) Let f : X → Y be a continuous one-to-one mapping of a compact
space into a Hausdorff space. Then f −1 : f (X) → X is continuous.

Proof. We leave the proofs of (a)-(c) as an exercise and we will only prove
property (d). We need to show that for every open set U ⊂ X, (f −1 )−1 (U ) =
f (U ) is an open subset of f (X). X \ U is closed, so it is compact by (a).
Hence f (X \ U ) is compact by (c). Since f (X) is Hausdorff, f (X \ U ) is
closed and hence f (U ) = f (X) \ f (X \ U ) is open. 2
Definition. Let K be a family of functions from a set Y into a topological
space (X, T ). The K-weak topology in Y is the weakest topology on Y for
which all the functions in the family K are continuous.
The K-weak topology is constructed as follows. Observe that the sets
B = {f1−1 (U1 ) ∩ . . . ∩ fn−1 (Un ) : fi ∈ K, Ui ∈ T , i = 1, 2, . . . , n}
must be open and that the family of all possible unions of the sets from B
is a topology, so it is the K-weak topology and B is a base.
Let (X1 , T1 ), . . . , (Xn , Tn ) be topological spaces. In the Cartesian product
n
Y
Xi = X1 × . . . × Xn
i=1

we consider open rectangles

U1 × . . . × Un , Ui ∈ Ti , i = 1, 2, . . . , n.
Qn
Then the family of all unions of open rectangles defines
Qna topology in i=1 Xi
which is called the product topology and denoted by i=1 Ti = T1 × . . . × Tn .
Clearly open rectangles form a base for this topology.
Note that each projection
πi : X 1 × . . . × X n → X i , πi (x1 , . . . , xn ) = xi
is continuous with respect to the product topology and that the product
topology is the weakest one which makes all the projections continuous.

46The claim is not true if the space is not Hausdorff. Indeed, {∅, X} is a topology in
X and every subset is compact, but only ∅ and X are closed.
128 PIOTR HAJLASZ

If (X1 , d1 ), . . . , (Xn , dn ) are metric spaces, thenQit is easy to see that the
topology induced by the metric d in the product ni=1 Xi
n
X
d((x1 , . . . , xn ), (y1 , . . . , yn )) = di (xi , yi )
i=1

coincides with the product topology.

One can also define a product topology in a product of an arbitrary (pos-
sibly uncountable) number of topological spaces.
Definition. Let (Xi , Ti )i∈I be an arbitrary family of topological spaces.
The Cartesian product
Y
Xi = {(xi )i∈I : xi ∈ Xi }
i∈I

is equipped with the product topology which is the weakest topology for
which each of the projection
Y
πj : Xi → Xj , πj ((xi )i∈I ) = xj
i∈I

is continuous.
Let i1 , . . . , in ∈ I and Ui1 ∈ Ti1 , . . . , Uin ∈ Tin be chosen arbitrarily. It is
easy to see that the sets
{(xi )i∈I : xi1 ∈ Ui1 , . . . , xin ∈ Uin } = πi−1
1
(Ui1 ) ∩ . . . ∩ πi−1
n
(Uin )
form a base for the product topology.
Q
It is easy to see that if the spaces {Xi }i∈I are Hausdorff, then i∈I Xi is
Hausdorff.
Theorem 15.6 (Tychonov). Let {Xi }i∈I Q be an arbitrary collection of com-
pact topological spaces. Then the product i∈I Xi is compact.

We will not prove it.

15.2. Weak topology in Banach spaces.

Definition. Let X be a Banach space. The weak topology in X is the weak-
est topology with respect to which all functionals x∗ ∈ X ∗ are continuous.
Let X be a normed space. The weak-∗ topology is the weakest topology in
X ∗ with respect to which all functions of the form x∗ 7→ hx∗ , xi for x ∈ X
are continuous.
Exercise. Prove that weak and weak-∗ topologies are Hausdorff.
 FUNCTIONAL ANALYSIS 129

Theorem 15.7 (Banach-Alaoglu). Let X be a normed space. The closed

unit ball in X ∗ , i.e.
B = {x∗ ∈ X ∗ : kx∗ k ≤ 1}
is compact in the weak-∗ topology.

Proof. For each x ∈ X let

Bx = {λ ∈ C : |λ| ≤ kxk} .
Each set Bx is compact, so is the product
Y
K= Bx .
x∈X

Elements of K can be identified with functions f : X → C such that |f (x)| ≤

kxk for all x ∈ X. Observe that the unit ball B in X ∗ is a subset of K.
Actually, X ∗ ∩ K = B. It remains to prove that the topology in B induced
from K coincides with the weak-∗ topology and that B is a closed subset of
K.
The sets
n
\
U= {f : X → C : |f (x)| ≤ kxk, f (xi ) ∈ Ui }
i=1
form a base of the topology in K and the sets
n
\
V= {x∗ ∈ X ∗ : hx∗ , xi i ∈ Ui }
i=1
form a base of the weak-∗ topology in X ∗ . Since
U ∩B =V ∩B
it easily implies that the weak-∗ topology and the topology induced from K
coincide on B.
It remains to prove that B is a closed subset of K. Let f0 be in the closure
of B in the topology of K. We have to prove that f0 ∈ B.
Fix x, y ∈ X, α, β ∈ C. The set
Vε = {f ∈ K :
|f (x) − f0 (x)| < ε, |f (y) − f0 (y)| < ε, |f (αx + βy) − f0 (αx + βy)| < ε}
is open and contains f0 . Hence Vε ∩ B 6= ∅ by the definition of the closure.
Thus there is x∗ ∈ Vε ∩ B, and hence it satisfies
|hx∗ , xi − f0 (x)| < ε, |hx∗ , yi − f0 (y)|, ε
| αhx∗ , xi + βhx∗ , yi −f0 (αx + βy)| < ε .
| {z }
hx∗ ,αx+βyi
130 PIOTR HAJLASZ

The above inequalities yield

|f0 (αx + βy) − αf0 (x) − βf0 (y)| < (1 + |α| + |β|) ε .
Since ε > 0 can be arbitrarily small we have
f0 (αx + βy) = αf0 (x) + βf0 (y) .
Moreover |f0 (x)| ≤ kxk as an element of K which implies f0 ∈ B. 2
The following result shows that in the case of separable Banach spaces,
Corollary 15.3 is equivalent to the Banach-Alaoglu theorem.
Theorem 15.8. Let X be a separable Banach space. Then the weak-∗ topol-
ogy in the unit ball B in X ∗ coincides with the topology induced by the metric
ρB .

Proof. Let
Φ : X ∗ → s, Φ(x∗ ) = (hx∗ , x1 i, hx∗ , x2 i, . . .)
be a mapping defined in the proof of Theorem 15.2. The mapping is one-to-
one. It remains to prove that Φ is continuous. Indeed, since s is Hausdorff
as a metric space and B is compact in the weak-∗ topology, it will follow
from Proposition 15.5(d) that
Φ|B : B → s
is a homeomorphism onto the image and hence the weak-∗ topology in
B will coincide with the topology generated by the metric ρ(x∗ , y ∗ ) =
d(Φ(x∗ ), Φ(y ∗ )).

Fix r > 0 and let w = (wi ) ∈ s. Let N be such that ∞ −k < r/2.
P
k=N +1 2
Then the set
A(w, r, N ) = {t ∈ s : |ti − wi | < r/2, i = 1, 2, . . . , N }
is open, contained in B(w, r) and w ∈ A(w, r, N ). Hence the sets A(w, r, N )
form a base in s. Thus it remains to prove that Φ−1 (A(w, r, N )) is open in
the weak-∗ topology. By the definition of the weak-∗ topology the functions
x∗ 7→ hx∗ , xi i are continuous, so the sets
{x∗ ∈ X ∗ : |hx∗ , xi i − wi | < r/2}
are open as preimages of B(wi , r/2) ⊂ C and hence
N
\
Φ−1 (A(w, r, N )) = {x∗ ∈ X ∗ : |hx∗ , xi i − wi | < r/2}
i=1

is open. 2
 FUNCTIONAL ANALYSIS 131

16. Compact operators

Definition. Let A ∈ B(X, Y ) be a bounded operator between Banach

spaces. We say that A is compact if it maps bounded sets onto relatively
compact sets, i.e. if for every bounded set E ⊂ X, A(E) ⊂ Y is compact.
The class of compact operators will be denoted by K(X, Y ) with K(X) =
K(X, X).
Equivalently A ∈ B(X, Y ) is compact if for every bounded sequence
{xn }n ⊂ X, {Axn }n ⊂ Y has a convergent subsequence.
It is also easy to see that A is compact if and only if
A(B(0, 1)) ⊂ Y is compact.
For a bounded mapping A ∈ B(X, Y ) we define47
R(A) = A(X), N (A) = ker A .
Proposition 16.1. If dim R(A) < ∞, then A is compact.

Since the closed unit ball in an infinitely dimensional space is not compact
(Corollary 3.7) we have
Proposition 16.2. If dim X = ∞, then the identity mapping id : X → X
is not compact.
Theorem 16.3. If An ∈ B(X, Y ) is a sequence of compact operators be-
tween Banach spaces and An → A in norm, then A is compact.

Proof. Let {xn }n ⊂ X be a bounded sequence, say kxn k ≤ M for all n.

We need to show that Axn ∈ Y has a convergent subsequence. Since each
sequence {Ai xn }n has a convergent subsequence, by diagonal argument we
find a subsequence {xnk }k such that for each i = 1, 2, 3 . . .
Ai xnk is convergent as k → ∞.
Given ε > 0, let i be such that
kA − Ai k < ε/3M .
The Cauchy condition gives existence of k0 such that for k, l ≥ k0
kAi xnk − Ai xnl k < ε/3
and hence
kAxnk − Axnl k ≤ kAi xnk − Ai xnl k + kAxnk − Ai xnk k + kAxnk − Ai xnl k
ε
≤ + kA − Ai kM + kA − Ai kM < ε .
3
Thus the sequence {Axnk }k satisfies the Cauchy condition and hence it is
convergent, because Y is a Banach space. 2
47R stands for range and N for null space.
132 PIOTR HAJLASZ

Theorem 16.4. Let A ∈ B(X, Y ) be a compact operator between Banach

spaces. If xn * x weakly in X, then Axn → Ax in norm.

Proof. First we will prove that Axn * Ax weakly in Y . If y ∗ ∈ Y ∗ , then

x∗ = y ∗ ◦ A, i.e. hx∗ , xi = hy ∗ , Axi is a bounded functional in X ∗ and hence
weak convergence xn * x yields
hy ∗ , Axn i = hx∗ , xn i → hx∗ , xi = hy ∗ , Axi .
Suppose that Axn does not converge to Ax in norm. Then there is a subse-
quence Axnk and ε > 0 such that
(16.1) kAxnk − Axk ≥ ε .
The sequence xn is bounded (as weakly convergent) and since A is compact,
there is a convergent subsequence Axnkl → y. Since Axnkl * Ax weakly
we easily conclude that y = Ax, so Axnkl → Ax in norm which contradicts
(16.1). 2
Theorem 16.5. Let X, Y, Z, V be Banach spaces. If A ∈ B(X, Y ) is com-
pact and B ∈ B(Y, Z), C ∈ B(V, X) are bounded operators, then the opera-
tors BA ∈ B(X, Z), AC ∈ B(V, Y ) are compact.

Proof. Easy exercise. 2

It follows from Proposition 16.1 and Theorem 16.3 that if an operator A ∈
B(X, Y ) can be approximated in norm by operators with finitely dimensional
range, then A is compact. In the case of operators into Hilbert spaces this
property characterizes compact operators.
Theorem 16.6. Let A ∈ B(X, H) be a bounded operator between a Banach
space and a Hilbert space. Then A is compact if and only if there is a sequence
of operators An ∈ B(X, H) such that dim R(An ) < ∞ and An → A in norm.

Proof. We only need to prove the implication from left to right. We can
assume that dim R(A) = ∞ as otherwise the claim is obvious. Since A is
compact, R(A) is a union of countably many relatively compact sets, so
R(A) is separable. Let {ϕi }∞
i=1 be an orthonormal basis in R(A) and let

Pn : R(A) → span {ϕ1 , ϕ2 , . . . , ϕn }

be the orthogonal projection. It remains to prove that An = Pn A → A in
norm. If not, there is ε > 0 and a subsequence (still denoted by An ) such
that
kAn − Ak ≥ ε .
Hence there is a sequence xn ∈ X, kxn k = 1 such that
ε
(16.2) k(An − A)xn k ≥ .
2
 FUNCTIONAL ANALYSIS 133

Since A is compact, Ank xnk → y ∈ R(A) for some subsequence. We have

(A − Ank )xnk = (I − Pnk )Axnk = (I − Pnk )y + (I − Pnk )(Ank − y) → 0
as k → ∞ which contradicts (16.2). 2
Recall that for a bounded operator A ∈ B(X, Y ) we define the adjoint
operator A∗ ∈ B(Y ∗ , X ∗ ) by
hA∗ y ∗ , xi = hy ∗ , Axi .
Theorem 16.7. Let Ω ⊂ Rn be open and K ∈ L2 (Ω × Ω). Then the integral
operator Z
Kf = K(x, y)f (y) dy , x∈Ω
Ω
defines a compact operator K : L2 (Ω) → L2 (Ω).

Proof. The operator K : L2 (Ω) → L2 (Ω) is well defined, because

Z Z 2
2
kKf k2 =  K(x, y)f (y) dy  dx
 
ZΩ  ZΩ Z 
2
≤ |K(x, y)| dy |f (y)|2 dy dx
Ω Ω Ω
= kKk2L2 (Ω×Ω) kf k22 ,
i.e.
(16.3) kKf k2 ≤ kKk2 kf k2 .
Let {ϕi }∞ 2
i=1 be an orthonormal basis in L (Ω). According to Theorem 5.15
∞
the functions {ϕi (x)ϕj (y)}i,j=1 form an orthonormal basis in L2 (Ω × Ω). We
can write
∞
X
K(x, y) = aij ϕi (x)ϕj (y)
i,j=1
where the series converges to K in the norm of L2 (Ω × Ω), i.e. the functions
m
X
Km (x, y) = aij ϕi (x)ϕj (y)
i,j=1

converge to K in L2 (Ω× Ω). Now it follows from inequality (16.3) that

the operators Km ∈ B(L2 (Ω), L2 (Ω)) converge to K ∈ B(L2 (Ω), L2 (Ω)) in
norm. Since the range of each operator Km is finitely dimensional
Xm Z
Km f (x) = aij ϕi (x) f (y)ϕj (y) dy ∈ span {ϕ1 , . . . , ϕm }
i,j=1 Ω

the claim follows from Theorem 16.6. 2

Theorem 16.8 (Schauder). An operator A ∈ B(X, Y ) between Banach
spaces is compact if and only if A∗ ∈ B(Y ∗ , X ∗ ) is compact.
134 PIOTR HAJLASZ

Proof. ⇒. Let A ∈ B(X, Y ) be compact. In order to prove compactness

of A∗ we need to show that for every bounded sequence {yn∗ }n ⊂ Y ∗ , say
supn kyn∗ k ≤ M , {A∗ yn∗ }n ⊂ X ∗ has a convergent subsequence.

Let S = {x ∈ H : kxk ≤ 1}. The set K = A(S) ⊂ Y is compact. The

family of functions
fn : K → K, fn (y) = hyn∗ , yi
is bounded and equicontinuous. Boundedness is easy and equicontinuity fol-
lows from the estimate
|fn (y1 ) − fn (y2 )| ≤ M ky1 − y2 k .
According to the Arzela-Ascoli theorem, there is a uniformly convergent
subsequence {fnk }. Hence also the functions
gk : S → K, gk (x) = fn (Ax) = hA∗ yn∗ k , xi
converge uniformly on S.48 The limit
hx∗ , xi := lim hA∗ yn∗ k , xi
k→∞
exists for every x ∈ X and it defines a bounded functional x∗ ∈ X ∗ . Moreover
kA∗ yn∗ k − x∗ k = sup |gk (x) − hx∗ , xi| → 0 as k → ∞.
x∈S

This completes the proof of compactness of A∗ .

⇐. Suppose now that A∗ ∈ B(Y ∗ , X ∗ ) is compact. Let {xn }n ⊂ X be a
bounded sequence. We need to prove that {Axn }n ⊂ Y has a convergent
subsequence. By the first part of the proof A∗∗ ∈ B(X ∗∗ , Y ∗∗ ) is compact
and hence A∗∗ (κX (xn )) has a convergent subsequence A∗∗ (κX (xnk )), where
κX : X → X ∗∗ is the canonical embedding. Since
κY (Axnk ) = A∗∗ (κX (xnk ))
where κY : Y → Y ∗∗ is the canonical embedding and hence an isometric
embedding, the sequence {Axnk }k satisfies the Cauchy condition, so it is
convergent. 2

16.1. Fredholm operators. Fredholm operators play an important role

in the theory of integral and elliptic equations. We will describe later an
application to the Fredholm theory of integral equations. Another applica-
tion appears in the famous Atiyah-Singer index theorem which deals with
the index of a special Fredholm operator defined as an elliptic operator on
manifolds. The Atiyah-Singer theorem plays a fundamental role in algebraic
topology and in contemporary physics.
48Although the family g is bounded and equicontinuous, we could not apply the
k
Arzela-Ascoli theorem directly to gk , because S is not compact.
 FUNCTIONAL ANALYSIS 135

Definition. A bounded operator A ∈ B(X, Y ) between Banach spaces is

called Fredholm operator if
dim N (A) < ∞, dim(Y /R(A)) < ∞.
Note that according to Theorem 4.8 R(A) is a closed subspace on Y .
The space of Fredholm operators will be denoted by Fred (X, Y ). The
quantity
ind A = dim N (A) − dim(Y /R(A))
is called the Fredholm index of A.
Theorem 16.9. Fred (X, Y ) is an open subset of B(X, Y ). The Fredholm
index ind : Fred (X, Y ) → Z is continuous and hence constant on connected
components of Fred (X, Y ).

Proof. Let A ∈ Fred (X, Y ). By Theorem 11.8 the subspaces N (A) ⊂ X

and R(A) ⊂ Y are complemented, so there are closed subspaces X1 ⊂ X
and Y1 ⊂ Y such that
X = N (A) ⊕ X1 , Y = R(A) ⊕ Y1 .
For any operator L ∈ B(X, Y ) we define
L̃ : X1 ⊕ Y1 → Y, L̃(x1 , y1 ) = Lx1 + y1 .
It is easy to see that
à : X1 ⊕ Y1 → Y
is a bijection and hence it is an isomorphism. If B ∈ B(X, Y ), then kà −
B̃k ≤ kA − Bk and since isomorphisms form an open subset in the space
of all bounded transformations (Theorem 2.7) it follows that if kA − Bk
is sufficiently small, then also B̃ is an isomorphism. In particular N (B) ∩
X1 = {0}, so dim N (B) ≤ dim N (A) < ∞, because the quotient map X →
X/X1 ' N (A) restricted to N (B) is one-to-one. Moreover for any y ∈ Y
there is x1 ∈ X1 ⊂ X and y1 ∈ Y such that
Bx1 + y1 = B̃(x1 , y1 ) = y
and hence
R(B) + Y1 = Y
which gives
dim(Y /R(B)) ≤ dim Y1 = dim(Y /R(A)) < ∞.
This proves that B is a Fredholm operator and it remains to show that
ind B = ind A.
Since N (B) ⊕ X1 is a closed subspace of X of finite codimension, there is
a finitely dimensional subspace Z ⊂ X such that
(16.4) X = N (B) ⊕ Z ⊕ X1 ,
136 PIOTR HAJLASZ

so
dim N (B) ⊕ Z = dim(X/X1 ) = dim N (A) ,

(16.5) dim N (B) = dim N (A) − dim Z.

The mapping
B̃ : X1 ⊕ Y1 → Y
is an isomorphism, so Y = B(X1 ) ⊕ Y1 (see Proposition 10.7) and hence
dim Y1 = dim(Y /B(X1 )). On the other hand (16.4) shows that B restricted
to Z ⊕ X1 gives an isomorphism onto R(B), so
R(B) = B(Z) ⊕ B(X1 )
and in particular dim B(Z) = dim Z. Hence
dim(Y /R(B)) = dim(Y /B(X1 )) − dim B(Z) = dim Y1 − dim Z
= dim(Y /R(A)) − dim Z
which together with (16.5) immediately implies that ind A = ind B. 2
The above result shows in particular that if [0, 1] 3 t → A(t) is a continu-
ous one parameter family of Fredholm operators, then ind A(0) = ind A(1).
Theorem 16.10. Let K ∈ B(X) be a compact operator. Then I + K ∈
Fred (X) and ind (I + K) = 0.

Proof. First we will prove that dim N (I + K) < ∞. To this end it suffices
to prove that the closed unit ball in N (I + K) is compact, see Theorem 3.7.
Let xn ∈ N (I + K), kxn k ≤ 1. Since K is compact, after passing to a
subsequence we may assume that Kxn → y converges. Since xn + Kxn = 0,
xn → −y.
Now we will prove that R(I + K) is closed. The space N (I + K) is finitely
dimensional, so it is complemented
(16.6) X = N (I + K) ⊕ V
for some closed subspace V ⊂ X. The mapping I + K restricted to V is a
bijection onto R(I + K) and to prove that R(I + K) is closed it suffices to
show that the inverse mapping is continuous

−1
(I + K)|V : R(I + K) → V .
We need to prove continuity at 0 only. Suppose the mapping is not contin-
uous at 0. Then there is a sequence xn ∈ V such that (I + K)xn → 0, but
xn 6→ 0. Without loss of generality we may assume that kxn k ≥ ε. Then the
sequence yn = xn /kxn k satisfies kyn k = 1, yn + Kyn → 0. By compactness
of K, after passing to a subsequence we have Kyn → y, yn → −y. Hence
kyk = 1, y ∈ V , y ∈ N (I + K) which is a contradiction with (16.6).
 FUNCTIONAL ANALYSIS 137

In order to prove that I + K is a Fredholm operator we still need to show

that the space R(I + K) has a finite codimension. Suppose the codimension
is infinite. Then we can find a sequence of closed subspaces
R(I + K) = H0 ⊂ H1 ⊂ H2 ⊂ . . .
such that dim(Hn+1 /Hn ) = 1. Applying the Riesz lemma (Theorem 3.6),
we can find xn ∈ Hn , kxn k = 1 such that
1
kxn − yk ≥ for all y ∈ Hn−1 .
2
For k < n we have
1
kKxn − Kxk k = k(xn + Kxn ) − (xk + Kxk ) + xk −xn k ≥ .
| {z } | {z } |{z} 2
∈H0 ∈H0 ∈Hk

This shows that the sequence {Kxn } cannot have a convergent subsequence,
which contradicts compactness of K. Hence R(I +K) has finite codimension.
We proves that I + K ∈ Fred (X). Now [0, 1] 3 t 7→ I + tK is a continuous
family of Fredholm operators and from the previous result we have ind (I +
K) = ind I = 0. 2
Definition. Let A, B ∈ B(X, Y ). We say that A is congruent to B modulo
compact operators if A − B is compact and we write
A≡B mod K(X, Y ) .
It is easy to see that this is an equivalence relation. Moreover if
A≡B mod K(X, Y ), A1 ≡ B1 mod K(Y, Z) ,
then
A1 A ≡ B1 B mod K(X, Z) .
Indeed
A1 A − B1 B = A1 (A − B) + (A1 − B1 )B
is compact by Theorem 16.5.
Definition. We say that A ∈ B(X, Y ) is invertible modulo compact opera-
tors if there is A1 ∈ B(Y, X) such that
AA1 ≡ IY mod K(Y ), A1 A ≡ IX mod K(X) ,
i.e. A1 A = IX + K1 and AA1 = IY + K2 for some K1 ∈ K(X), K2 ∈ K(Y ).
We call A1 inverse of A modulo compact operators.
Theorem 16.11. Let A ∈ B(X, Y ) be a bounded operator between Banach
spaces. Then A is Fredholm if and only if it is invertible modulo compact
operators.

Proof. Let A ∈ Fred (X, Y ). We can write

X = N (A) ⊕ X1 , Y = R(A) ⊕ Y1 .
138 PIOTR HAJLASZ

Note that A restricted to X1 is an isomorphism from X1 onto R(A), so it is

invertible. Consider the following composition of operators
π (A|X1 )−1 ι
Y = R(A) ⊕ Y1 −→ R(A) −→ X1 −→ X ,
where π is the projection onto the first component and ι is the inclusion.
Denote this composition by
A1 = ι ◦ (A|X1 )−1 ◦ π : Y → X .
It is easy to check that IY − AA1 is the projection onto Y1 and IX − A1 A
is the projection onto N (A). Both mappings are compact as mappings with
finitely dimensional range.49
Conversely, suppose that A ∈ B(X, Y ) is invertible modulo compact op-
erators, i.e. there is A1 ∈ B(Y, X), K1 ∈ K(X), K2 ∈ K(Y ) such that
A1 A = IX + K1 , AA1 = IY + K2 .
Hence N (A) ⊂ N (A1 A) = N (IX + K1 ) shows that dim N (A) < ∞, because
IX + K1 ∈ Fred (X) and R(IY + K2 ) = R(AA1 ) ⊂ R(A) shows that R(A)
has finite codimension, so A ∈ Fred (X, Y ). 2
Corollary 16.12. The composition of Fredholm operators is Fredholm. If A
is Fredholm and K is compact, then A + K is Fredholm and ind(A + K) =
ind A.

Proof. Let A ∈ Fred (X, Y ), B ∈ Fred (Y, Z) and A1 , B1 be the inverse

operators modulo compact ones, i.e.
A1 A = IX + K1 , AA1 = IY + K2 ,
B1 B = IY + K3 , BB1 = IZ + K4 .
Then
A1 B1 BA = IX + K1 + A1 K3 A , BAA1 B1 = IZ + K4 + BK2 B1 ,
| {z } | {z }
compact compact
so A1 B1 is the inverse of BA modulo compact operators and hence BA ∈
Fred (X, Z).
A similar argument can be used to show that A + K is Fredholm. Finally,
[0, 1] 3 t 7→ A + tK is a continuous family of Fredholm operators, so ind A =
ind (A + K). 2
Theorem 16.13. If A ∈ Fred (X, Y ), B ∈ Fred (Y, Z), then
ind (BA) = ind B + ind A .

We will not prove it.50

49Hence we can always select an inverse of A modulo compact operators with finitely
dimensional range.
50For a proof, see S. Lang, Real and Functional Analysis. Third edition, p. 423.
 FUNCTIONAL ANALYSIS 139

16.2. Spectrum of compact operators. In this section we will assume

that X is a complex Banach space.
Definition. For a bounded operator T ∈ B(X) the spectrum σ(T ) is defined
as
σ(T ) = {λ ∈ C : T − λI is not invertible} .
If λ ∈ σ(T ), then T − λI is not surjective or it is not ont-to-one. In the latter
case λ is called eigenvalue of T and there is 0 6= x ∈ X such that
T x = λx .
Each such vector is called eigenvector of T .
Since T − λI = λ(λ−1 T − I) is invertible for |λ| > kT k, by Theorem 2.7,
the spectrum σ(T ) is a bounded set satisfying |λ| ≤ kT k for λ ∈ σ(T ).
Theorem 16.14. If T ∈ B(X) is a bounded operator, then eigenvectors
corresponding to distinct eigenvalues are linearly independent.

Proof. We argue by induction. Suppose that any collection of (n − 1)

eigenvectors corresponding to distinct eigenvalues is linearly independent.
Suppose that T xi = λi xi , xi 6= 0, i = 1, 2, . . . , n and λi 6= λj for i 6= j.
We need to prove that the vectors x1 , . . . , xn are linearly independent. Let
c1 , . . . , cn be scalars such that
(16.7) c1 x1 + . . . + cn xn = 0 .
Applying the operator T we have
c1 λ1 x1 + . . . + cn λn xn = 0 .
At least one of the eigenvalues is nonzero, say λ1 6= 0. Dividing the second
equation by λ1 and subtracting from the first one gives
c2 (1 − λ2 /λ1 )x1 + . . . + cn (1 − λn /λ1 )xn = 0 .
| {z } | {z }
6=0 6=0

Since by the assumption the vectors x2 , . . . , xn are linearly independent,

c2 = . . . = cn = 0 and hence (16.7) gives c1 = 0 which proves linear
independence of x1 , . . . , xn . 2
Theorem 16.15. If A ∈ B(X) is compact, then every nonzero element in
the spectrum 0 6= λ ∈ σ(A) is an eigenvalue. If dim X = ∞, then 0 ∈ σ(A).

Proof. If 0 6= λ ∈ σ(A), then A − λI is a noninvertible Fredholm operator.

Since ind (A − λI) = 0, N (A − λI) 6= 0 and hence there is 0 6= x ∈ X such
that Ax = λx. If dim X = ∞, then A is not surjective by Corollary 3.7 and
hence A = A − 0 · I is not invertible, so 0 ∈ σ(A). 2
140 PIOTR HAJLASZ

Theorem 16.16. If A ∈ B(X) is compact, then eigenvalues from a finite or

a countable set. If there are infinitely many eigenvalues, then we can order
them as λ1 , λ2 , λ3 , . . . in such a way that
|λ1 | ≥ |λ2 | ≥ |λ3 | ≥ . . . lim λn = 0.
n→∞

Proof.51 It suffices to prove that for any r > 0 the number of eigenvalues
satisfying |λ| ≥ r is finite. If this is not true, then there are distinct eigenval-
ues λi , |λi | ≥ r, i = 1, 2, 3, . . . and corresponding eigenvectors xi , kxi k = 1.
Let
Hn = span {x1 , . . . , xn } .
By the Riesz lemma (Theorem 3.6) we can find wn ∈ Hn such that
1
kwn k = 1, kwn − yk ≥ for any y ∈ Hn−1 .
2
We have wn = an xn +yn−1 , yn−1 ∈ Hn−1 Thus for k < n, Awk ∈ Hk ⊂ Hn−1
and hence
kAwn − Awk k = kan λn xn + Ayn−1 − Awk k
r
= |λn |k an xn + yn−1 −(yn−1 − λ−1
n (Ayn−1 − Awk ))k ≥ .
| {z } | {z } 2
wn ∈Hn−1

Therefore the sequence {Awn } has no convergent subsequence which con-

tradicts compactness of A. 2

16.3. The Fredholm-Riesz-Schauder theory. Our aim is to apply the

above theory to the following problem. Given a compact operator A ∈ B(X)
in a Banach space X and a parameter 0 6= λ ∈ C we want to solve the
equation
(16.8) Ax − λx = y where y ∈ X is given and x ∈ X is unknown.
Theorem 16.17. If λ 6= 0 is not an eigenvalue of A, then for every y ∈ X,
(16.8) has a unique solution.

Proof. By Theorem 16.15, λ 6∈ σ(A), so A − λI is invertible and hence

(16.8) has the unique solution
x = (A − λI)−1 y .
The proof is complete. 2
Note that by Theorem 16.16 we have at most countably many eigenvalues
that form a discrete set whenever we are away from 0. Thus for most of the
values of λ, (16.8) can be uniquely solved for any y ∈ X. The interesting
case is, however, when λ is an eigenvalue of A.
51The proof is very similar to the proof that if K is compact, then R(I + K) has finite
codimension, see the proof of Theorem 16.10.
 FUNCTIONAL ANALYSIS 141

Along with the equation (16.8) we consider the adjoint equation

(16.9) A∗ x∗ − λx∗ = y ∗ .

The first part of the following definition already appears in Section 4.2.
Definition. Let X be a normed space. For linear subspaces M ⊂ X, N ⊂
X ∗ we define annihilators
M ⊥ = {x∗ ∈ X ∗ : hx∗ , xi = 0 for all x ∈ M } ,
⊥
N = {x ∈ X : hx∗ , xi = 0 for all x∗ ∈ N } .
Clearly M ⊥ and ⊥N are closed subspaces of X ∗ and X respectively.
The following result provides a complete description of solvability of the
equations (16.8) and (16.9).
Theorem 16.18 (Riesz-Schauder). If X is a Banach space, A ∈ B(X) is
compact and λ 6= 0, then
(16.10) dim N (A − λI) = dim N (A∗ − λI) < ∞ .
Moreover
(16.11) R(A − λI) =⊥N (A∗ − λI), R(A∗ − λI) = N (A − λI)⊥ .

Before we prove the theorem we will see how it applies to the equations
(16.8) and (16.9). The following description is called the Fredholm alterna-
tive.
It follows from (16.10) that 0 6= λ ∈ C is not an eigenvalue of A if an only
if it is not an eigenvalue of A∗ and in this case both equations
Ax − λx = y, A∗ x∗ − λx∗ = y ∗
have unique solutions for all y ∈ X and y ∗ ∈ X ∗ .
Also 0 6= λ ∈ C is an eigenvalue of A if and only if it is an eigenvalue of
A∗ and the dimensions corresponding of eigenspaces for A and A∗ are finite
and equal. Moreover the equation
Ax − λx = y
has a solution if and only if y ∈⊥N (A∗ − λI), i.e. if
hy ∗ , yi = 0 whenever A∗ y ∗ = λy ∗ .
Note that since dim(A∗ − λI) < ∞, this is a finite number of conditions to
check. Similarly the equation
A∗ x∗ − λx∗ = y ∗
has a solution if and only if y ∗ ∈ N (A − λI)⊥ , i.e.
hy ∗ , yi = 0 whenever Ay = λy.
142 PIOTR HAJLASZ

The above theory was first considered by Fredholm in the setting of inte-
gral equations. Let Ω ⊂ Rn be an open set and let K ∈ L2 (Ω × Ω). As we
know the operator K : L2 (Ω) → L2 (Ω),
Z
Kf (x) = K(x, y)f (y) dy
Ω

is compact, see Theorem 16.7. We consider the Fredholm integral equation

Z
(16.12) f (x) − µ K(x, y)f (y) dy = g(x) ,
Ω

where 0 6= µ ∈ C and g ∈ L2 (Ω) are given.

Theorem 16.19 (Fredholm).

(a) The equation (16.12) has a solution for every g ∈ L2 (Ω) if and only
if the only solution to
Z
(16.13) f (x) − µ K(x, y)f (y) dy = 0
Ω

is f = 0.
(b) The equation (16.13) has nonzero solutions, if and only if there are
nonzero solutions to
Z
(16.14) f (y) − µ K(x, y)f (x) dx = 0
Ω

The spaces of solutions to (16.13) and (16.14) are finitely dimen-

sional and have the same dimension.
(c) If the equation (16.13) has nonzero solutions, then (16.12) has solu-
tions if and only if
Z
g(x)f (x) dx
Ω

for every f ∈ L2 (Ω) that solves (16.14).

(d) The set of numbers µ for which (16.13) has nonzero solutions is
at most countable. If this set is infinite, then the numbers form an
infinite sequence µn , |µn | → ∞.

Proof. The equation (16.12) can be written as

Kf − λf = −µ−1 g, λ = µ−1 .
It has solutions for all g ∈ L2 (Ω) if and only if R(K − λI) = L2 (Ω) which
us equivalent to N (K − λI) = {0}, so (a) follows.
The equation (16.13) with µ 6= 0 has nonzero solutions if and only if
λ = µ−1 is an eigenvalue of K. Thus (d) follows from Theorem 16.16.
 FUNCTIONAL ANALYSIS 143

The dual space to L2 (Ω) is isometrically isomorphic to L2 (Ω) and the

isomorphism is given by52
Z
2 2 ∗
L (Ω) 3 f 7→ Λf ∈ (L (Ω)) , hΛf , gi = g(x)f (x) dx .
Ω
We have
Z Z 
∗
hK Λf , gi = hΛf , Kgi = K(x, y)g(y) dy f (x) dx
Ω Ω
Z Z 
= g(y) K(x, y)f (x) dx dy ,
Ω Ω
so Z
∗
K Λf = ΛK̃f , K̃f (y) = K(x, y)f (x) dx .
Ω
Hence the adjoint operator to K − λI is K̃ − λI and the equation adjoint
to (16.13) reads as
Z
(16.15) f (y) − µ K(x, y)f (x) dx = 0 .
Ω

Now (16.12) has a solution if and only if g ∈⊥ N (K̃ − λI) =⊥ N (K ∗ − λI),

i.e.
hΛf , gi = 0 whenever K ∗ Λf = λΛf ,
i.e. Z
g(x)f (x) dx = 0 whenever f solves (16.15).
Ω
Since f solves (16.15) if and only if f solves (16.14), the part (c) follows. 2
We prepare for the proof of Theorem 16.18 now.
Theorem 16.20. Let X and Y be normed spaces and T ∈ B(X, Y ). Then
N (T ∗ ) = R(T )⊥ and N (T ) =⊥R(T ∗ ) .

Proof. It follows from an obvious sequence of identifications. 2

However, we will need the following less obvious results.
Theorem 16.21. Let X and Y be normed spaces and T ∈ B(X, Y ). Then
R(T ) =⊥N (T ∗ ) .

Recall that
⊥
N (T ∗ ) = {y ∈ Y : hy ∗ , yi = 0 whenever T ∗ y ∗ = 0} .

52We refer to Theorem 2.13 rather than to Theorem 5.5.

144 PIOTR HAJLASZ

First we will prove that ⊥N (T ∗ ) ⊂ R(T ). If not, then there is y0 ∈⊥N (T ∗ ) \

R(T ). The Hahn-Banach theorem (Theorem 11.6) yields the existence of
y ∗ ∈ Y ∗ such that
hy ∗ , y0 i =
6 0 and hy ∗ , yi = 0 for all y ∈ R(T ).
The second condition implies that for all x ∈ X
hT ∗ y ∗ , xi = hy ∗ , T xi = 0 ,
so T ∗ y ∗ = 0 and hence hy ∗ , y0 i =
6 0 implies that y0 6∈⊥N (T ∗ ) which is a
contradiction.
It remains to prove that R(T ) ⊂⊥N (T ∗ ). Let y ∈ R(T ). Then there is a
sequence xn ∈ X such that T xn → y. If T ∗ y ∗ = 0, then
hy ∗ , yi = lim hy ∗ , T xn i = lim hT ∗ y ∗ , xn i = 0
n→∞ n→∞

which proves y ∈⊥N (T ∗ ). 2

Theorem 16.22. Let X and Y be Banach spaces and T ∈ B(X, Y ). If
R(T ) is closed, then
R(T ∗ ) = N (T )⊥ .

Proof. First we prove that

R(T ∗ ) ⊂ N (T )⊥ = {x∗ ∈ X ∗ : hx∗ , xi = 0 whenever T x = 0} .
If x∗ ∈ R(T ∗ ), then x∗ = T ∗ y ∗ for some y ∗ ∈ Y ∗ . Then for any x ∈ X
satisfying T x = 0 we have
hx∗ , xi = hT ∗ y ∗ , xi = hy ∗ , T xi = 0
so x∗ ∈ N (T )⊥ .
The proof of N (T )⊥ ⊂ R(T ∗ ) is more difficult.
R(T ) is a Banach space as a closed subspace of Y . The mapping
T̃ : X/ ker T → R(T ), T̃ ([x]) = T x
is a well defined bounded bijection. Hence it is an isomorphism. Thus the
inverse mapping
T̃ −1 (T x) = [x]
is also bounded. That easily implies that there is C > 0 such that for every
y ∈ R(T ) we can find x ∈ X such that T x = y and kxk ≤ Ckyk.
Let x∗ ∈ N (T )⊥ . We have to find y ∗ ∈ Y ∗ such that T ∗ y ∗ = x∗ . Since x∗
vanishes on all x ∈ X such that T x = 0 it easily follows that
hΛ, T xi = hx∗ , xi
 FUNCTIONAL ANALYSIS 145

is a well defined linear functional on R(T ). This functional is bounded.

Indeed, if y ∈ R(T ), then there is x ∈ X such that T x = y, kxk ≤ ckyk and
hence
|hΛ, yi| = |hΛ, T xi| = |hx∗ , xi| ≤ Ckx∗ kkyk .
By the Hahn-Banach theorem Λ can be extended to y ∗ ∈ Y ∗ . Now for any
x ∈ X we have
hT ∗ y ∗ , xi = hy ∗ , T xi = hΛ, T xi = hx∗ , xi
and hence T ∗ y ∗ = x∗ . 2
Proof of Theorem 16.18. If T = A − λI, then R(T ) = R(A − λI) is closed
and hence
R(A − λI) = R(T ) =⊥N (T ∗ ) =⊥N (A∗ − λI)
by Theorem 16.21. The second equality in (16.11) follows from Theo-
rem 16.22 since
R(A∗ − λI) = R(T ∗ ) = N (T )⊥ = N (A − λI)⊥ .
We are left with the proof of (16.10). By the Schauder theorem (Theo-
rem 16.8) A∗ ∈ B(X ∗ ) is compact, so both operators A − λI and A∗ − λI
are Fredholm and hence the null spaces are finitely dimensional. We need to
prove that the dimensions are equal.
We start with a general observation. If x1 , . . . , xn are linearly independent
elements in a normed space X, then it follows from the Hahn-Banach the-
orem (Theorem 11.6) that there are functionals x∗1 , . . . , x∗n ∈ X ∗ such that
hx∗i , xj i = δij . The converse result is also true.53
Lemma 16.23. If x∗1 , . . . , x∗n are linearly independent elements in X ∗ , then
there exist points x1 , . . . , xn ∈ X such that hx∗i , xj i = δij .

Proof. Let Nj = ker x∗j and

Mj = N1 ∩ . . . ∩ Nj−1 ∩ Nj+1 ∩ . . . ∩ Nn .
It suffices to prove that Mj \ Nj 6= ∅ for j = 1, 2, . . . , n. We will prove that
M1 \N1 6= ∅ as the argument for j 6= 1 is the same. Suppose by contradiction
that M1 ⊂ N1 , so
(16.16) hx∗1 , xi = 0 whenever hx∗j , xi = 0 for j = 2, . . . , n.
Consider the linear mapping
A : X → Cn−1 , Ax = [hx∗2 , xi, . . . , hx∗n , xi]
and define on A(X) a linear functional Λ
hΛ, Axi = hx∗1 , xi .
53It holds in any linear space, not only in a normed space as the proof is based on
linear algebra only.
146 PIOTR HAJLASZ

Note that this functional is well defined, because if Ax = Ax̃, then hx1 , xi =
hx∗1 , x̃i by (16.16). The functional Λ can be extended linearly to a functional54
on Cn−1 , so it is of the form
n
X
hΛ, [y2 , . . . , yn ]i = aj yj
j=2

for some scalars aj ∈ C. This, however gives for any x ∈ X

n
DX E
hx∗1 , xi = aj x∗j , x
j=2

which contradicts linear independence of functionals. 2

Now we can complete the proof of the theorem. Let x1 , . . . , xn be a Hamel
basis of N (A − λI) and y1∗ , . . . , ym
∗ be a Hamel basis of N (A∗ − λI).

As we observed above there are elements x∗1 , . . . , x∗n ∈ X ∗ and y1 , . . . , ym ∈

X such that
hx∗i , xj i = δij , hyi∗ , yj i = δij .
Suppose that n < m. Consider the operator
n
X
Bx = Ax + hx∗i , xiyi .
i=1

This is a sum of a compact operator and an operator with a finitely dimen-

sional range, so B is compact. We claim that N (B − λI) = {0}. Indeed, if
x ∈ N (B − λI), then Bx = λx, i.e.
n
X
λx − Ax = hx∗i , xiyi .
i=1

Since yj∗ ∈ N (A∗ − λI) for j = 1, 2, . . . , n we have

0 = hλyj∗ − A∗ yj∗ , xi = hyj∗ , λx − Axi = hx∗j , xi ,
so λx − Ax = 0, x ∈ N (A − λI). Thus x = ni=1 ai xi , but 0 = hx∗j , xi = aj
P

and hence x = 0. We prove that N (B − λI) = {0}. Since B − λI is Fredholm

of index 0 we conclude that R(B − λI) = X. In particular there is x ∈ X
such that
Bx − λx = yn+1 ,
but this yields
D n
X E
∗ ∗ ∗
1 = hyn+1 , yn+1 i = hyn+1 , Bx − λxi = yn+1 , Ax − λx + hx∗i , xiyi = 0
i=1

54Linear algebra.
 FUNCTIONAL ANALYSIS 147

∗
since yn+1 ∈ N (A∗ − λI) and hyn+1
∗ , y i = 0 for i = 1, 2, . . . , n. The contra-
i
diction proves that n ≥ m, i.e.
dim N (A − λI) ≥ dim N (A∗ − λI) .
Applying this result to A∗ in place of A we have
dim N (A∗ − λI) ≥ dim N (A∗∗ − λI) .
If x ∈ N (A − λI), then it easily follows that κ(x) ∈ N (A∗∗ − λI) and hence
dim N (A∗∗ − λI) ≥ dim N (A − λI) ,
so the above inequalities give (16.10).

16.4. Spectral theorem.

Definition. Let H be a Hilbert space. We say that a linear operator T :
H → H is self-adjoint if
hT x, yi = hx, T yi for all x, y ∈ H.
It immediately follows from the Hellinger-Toeplitz theorem (Theorem 10.5)
that self-adjoint operators are bounded.
Theorem 16.24. If T ∈ B(H) is self-adjoint, then

(a) hAx, xi is real for all x ∈ H,

(b) Eigenvalues of T are real,
(c) Eigenspaces of T corresponding to distinct eigenvalues are orthogo-
nal.

Proof. Since hAx, xi = hx, Axi = hAx, xi, (a) follows. If T x = λx, x 6= 0,
then
λhx, xi = hT x, xi = hx, T xi = λhx, xi
so λ ∈ R which is (b). If λ1 6= λ2 are eigenvalues and x1 , x2 6= 0 are corre-
sponding eigenvectors, then using the fact that λ2 ∈ R we have
λ1 hx1 , x2 i = hT x1 , x2 i = hx1 , T x2 i = λ2 hx1 , x2 i ,
so hx1 , x2 i = 0 which is (c). 2
The spectrum of a bounded operator is bounded. more precisely if λ ∈
σ(T ), then |λ| ≤ kT k (see a remark proceeding Theorem 16.14).
Theorem 16.25. If A ∈ B(H) is compact and self-adjoint, then kAk or
−kAk is an eigenvalue of A.

Proof. Let xn ∈ H be a sequence such that kxn k = 1, kAxn k → kAk. After

passing to a subsequence we may assume that Axn → y ∈ H, so kyk = kAk
and A2 xn → Ay. We have
Schwarz
kAyk = lim kA2 xn k ≥ lim hA2 xn , xn i = lim hAxn , Axn i = kAk2 .
n→∞ n→∞ n→∞
148 PIOTR HAJLASZ

Hence
kA2 yk kyk ≥ hA2 y, yi = kAyk2 ≥ kAk4 = kAk2 kyk2 ≥ kA2 yk kyk ,
so
hA2 y, yi = kA2 yk kyk.
This is possible only if the vectors A2 y and y are parallel, A2 y = cy. We
have
hA2 y, yi kAk4
c= = = kAk2 .
hy, yi kAk2
Let x = y + kAk−1 Ay. If x = 0, then Ay = −kAk y and hence −kAk is an
eigenvalue of A. If x 6= 0, then
Ax = Ay + kAk−1 A2 y = Ay + kAk−1 kAk2 y = Ay + kAk y = kAk x ,
so kAk is an eigenvalue. 2
Let A ∈ B(H) be compact and self-adjoint. According to Theorem 16.16
all nonzero eigenvalues form a finite or infinite sequence {λi }N
i=1 (N is finite
or N = ∞). If N = ∞ and we order the eigenvalues in such a way that
|λ1 | ≥ |λ2 | ≥ |λ3 | ≥ . . ., then
lim λi = 0.
i→∞
Let
Eλi = {x ∈ H : Ax = λi x} .
Eλi is the eigenspace corresponding to the eigenvalue λi . Since Eλi = N (A−
λi I),
dim Eλi < ∞.
Moreover Theorem 16.24 yields
Eλi ⊥ Eλj for i 6= j.
Theorem 16.26 (Spectral theorem). Let A ∈ B(H) be compact and self-
adjoint. Then
N
M
H = ker A ⊕ Eλi .
i=1

Proof. Let
N
M
Y = Eλi .
i=1
Then H = Y ⊕ Y ⊥ and it remains to show that Y ⊥ = ker A. It is easy to
see that A(Y ) ⊂ Y and hence for every x ∈ Y ⊥ and y ∈ Y we have
hAx, yi = hx, Ayi = 0 ,
i.e. Ax ∈ Y ⊥ . Hence A|Y ⊥ : Y ⊥ → Y ⊥ is a compact self-adjoint operator.
According to Theorem 16.25 at least one of the numbers ±kA|Y ⊥ k is its
eigenvalue. It cannot be a nonzero eigenvalue, because all eigenvectors with
 FUNCTIONAL ANALYSIS 149

nonzero eigenvalues are contained in Y . Thus kA|Y ⊥ k = 0, A|Y ⊥ = 0, i.e.

Y ⊥ = ker A. 2
Note that ker A = E0 is the eigenspace corresponding to the zero eigen-
value and hence the spectral theorem says that the Hilbert space is the direct
sum of eigenspaces.
Example. Let {λi }∞
i=1 ⊂ C be any sequence such that λi → 0

Exercise. Prove that the operator A : `2 → `2 ,

A(x1 , x2 , x3 , . . .) = (λ1 x1 , λ2 x2 , λ3 x3 , . . .)
is compact.
Hence we can find a compact operator with a prescribed set of eigenvalues.
If λi 6= 0 for all i, then ker A = {0}, so compact operators can be one-to-
one even in the infinitely dimensional spaces. If λi ∈ R for all i, then A is
compact self-adjoint.
Exercise. Let K ∈ L2 (Ω×Ω). Prove that the integral operator K : L2 (Ω) →
L2 (Ω),
Z
Kf (x) = K(x, y)f (y) dy
Ω
is self-adjoint if and only if

K(x, y) = K(y, x) a.e. in Ω × Ω.

16.5. Sobolev spaces and the eigenfunctions od the Laplace oprta-

tor. In Section 7 we proved that the eigenfunctions of the Laplace operator
on the sphere55 form an orthonormal basis in L2 (S n−1 ) (Theorem 7.1 and
Theorem 7.5). In this section we will prove s similar result: suitable eigen-
functions of the Laplace operator form an orthonormal basis in L2 (Ω), where
Ω ⊂ Rn is any bounded open set. This will be a consequence of the spectral
theorem from the previous section. To prove, or even to state the result,
we need to introduce Sobolev spaces which is one of the main tools in the
theory of partial differential equations and calculus of variations.
Definition. Let Ω ⊂ Rn be an open set, u, v ∈ L1loc (Ω) and let α be a
miltiindex. We say that Dα u = v in the weak sense if
Z Z
vϕ = (−1)|α| uDα ϕ for all ϕ ∈ C0∞ (Ω).
Ω Ω

55Laplace-Beltrami.
150 PIOTR HAJLASZ

The weak derivative (if it exists) is unique. Indeed, if Dα u = v1 , Dα u = v2

weakly, then Z
(v1 − v2 )ϕ = 0 for all ϕ ∈ C0∞ (Ω)
Ω
and hence v1 = v2 a.e.
If u ∈ C ∞ (Ω), then the classical partial derivative Dα u equals to the weak
one due to the integration by parts formula.
Definition. Let 1 ≤ p ≤ ∞ and let m ≥ 1 be an integer. The Sobolev
space W m,p (Ω) is the space of all functions u ∈ Lp (Ω) such that the weak
derivatives Dα u exist for all |α| ≤ m and belong to Lp (Ω). The Sobolev
space is equipped with the norm
X
kukm,p = kDα ukp .
|α|≤m

Theorem 16.27. W m,p (Ω) is a Banach space.

Proof. If {uk } ⊂ W m,p (Ω) is a Cauchy sequence, then for every α, |α| ≤ m,
{Dα uk } is a Cauchy sequence in Lp (Ω), so Dα uk converges to some function
uα ∈ Lp (Ω) (we will writeu instead of u0 ). Since
Z Z Z Z
α α |α| α |α|
uD ϕ ← uk D ϕ = (−1) D uk ϕ → (−1) uα ϕ
Ω Ω Ω Ω
we conclude that α
D u = uα weakly, so u ∈ W m,p (Ω) and uk → u in the
norm of W m,p (Ω). 2
Theorem 16.28 (Meyers-Serrin). For 1 ≤ p < ∞ smooth functions are
dense in W m,p (Ω), i.e. for every u ∈ W m,p (Ω) there exist a sequence uk ∈
C ∞ (Ω), kuk km,p < ∞ such that ku − uk km,p → 0 as k → ∞.

We will not prove it. The space of functions u ∈ C ∞ (Ω) with kukm,p <
inf ty forms a normed space with respect to the Sobolev norm and the
above theorem shows that the Sobolev space can be equivalently defined as
its completion.
Theorem 16.29. For 1 < p < ∞ the Sobolev space W m,p (Ω) us reflexive.

Proof. Let N be the number of multiindices |α| ≤ m. The space

Lp (Ω) ⊕ . . . ⊕ Lp (Ω) = Lp (Ω)N
| {z }
N times
is reflexive, so is any of its closed subspaces. The mapping
T : W m,p (Ω) → Lp (Ω)N , T (u) = (Dα u)|α|≤N
is an isometric embedding, so its image R(T ) is a closed subspace of Lp (Ω)N .
Hence W m,p (Ω) is reflexive as isomorphic to a reflexive space. 2
 FUNCTIONAL ANALYSIS 151

The notion of weak derivative can be generalized as follows.

Definition. If P = |α|≤m aα Dα is a differential operator, with smooth
P

coefficients and u, v ∈ L1loc (Ω), then we say that P u = v in Ω in the weak

sense if
Z Z X
vϕ = u (−1)|α| Dα (aα ϕ) for all ϕ ∈ C0∞ (Ω).
Ω Ω |α|≤m

= 0 for all ϕ ∈ C0∞ (Ω).

R
In particular ∆u = 0 in weak sense if Ω u∆ϕ

As before P u is unique (if it exists) and coincides with the classical dif-
ferential operator applied to u is u is smooth.
Definition. For 1 ≤ p < ∞ and m ≥ 1 we define W0m,p (Ω) as the closure
of C0∞ (Ω) in the Sobolev norm.
Theorem 16.30 (Poincaré lemma). If Ω ⊂ Rn is bounded, 1 ≤ p < ∞ and
m ≥ 1, then there is a constant C = C(m, p, Ω) such that
kDα ukp for all u ∈ W0m,p (Ω).
X
kukm,p ≤ C
|α|=m

Proof. By the density argument it suffices to prove the result for u ∈

C0∞ (Ω) and it actually suffices to prove the inequality
n
∂u
X
kuk ≤ C
∂xi p

i=1
as the general case follows by iteration of this inequality.
Let M > 0 be such that Ω ⊂ [−M, M ]n . Then for every x ∈ [−M, M ]n
Z x1 Z M 
∂u  ∂u 

u(x) = (t, x2 , . . . , xn ) dt ≤  dx .
−M ∂x1 −M ∂x1


Applying the Hölder inequality we get

Z M  ∂u p
p p−1 p−1
|u(x)| ≤ 2 M  dt
 
∂x

−M 1

and the result follows by integration with respect to x ∈ Ω. 2

The result shows that if Ω is abounded set, then
X
k∇m ukp := kDα ukp
|α|=m

is an norm on W0m,pequivalent to k · km,p . Note that it is not a norm in

m,p
W (Ω) because it vanishes on constant functions. In particular this shows
that nonzero constant functions do not belong to W0m,p (Ω).
152 PIOTR HAJLASZ

Theorem 16.31 (Rellich-Kondrachov). If Ω ⊂ Rn is bounded and 1 ≤ p <

∞, then the embedding, i.e. the inclusion
W01,p (Ω) ⊂ Lp (Ω)
is a compact operator.

We will not prove this result.

Consider the following Dirichlet problem. Given a bounded open set Ω ⊂
Rn , find f ∈ W01,2 (Ω) such that
(16.17) −∆f = g weakly in Ω,
i.e. Z Z
∇f · ∇ϕ = gϕ for all ϕ ∈ C0∞ (Ω).
Ω Ω
Since C0∞ (Ω) is dense in W01,2 (Ω) this condition is equivalent to
Z Z
(16.18) ∇f · ∇h = gh for all h ∈ W01,2 (Ω).
Ω Ω

Theorem 16.32. For every g ∈ L2 (Ω) there is a unique solution f ∈

W01,2 (Ω) of the Dirichlet problem (16.17).

Proof. Since f 7→ k∇f k2 is an equivalent norm on W01,2 (Ω),

Z
[f, h] = ∇f · ∇h
Ω
R
is an equivalent inner product. Since h 7→ Ω gh is a bounded linear func-
tional on W01,2 (Ω) it follows from the Riesz representation theorem (Theo-
rem 5.5) that there is unique f ∈ W01,2 (Ω) such that
Z
[f, h] = gh for all h ∈ W01,2 (Ω) ,
Ω
i.e. Z Z
∇f · ∇h = gh for all h ∈ W01,2 (Ω)
Ω Ω
which means f is a solution to (16.17). 2

For each g ∈ L2 (Ω) denote by T g = f ∈ W01,2 (Ω) the unique solution to

(16.17). The operator
T : L2 (Ω) → W01,2 (Ω)
is linear. It is also bounded, because (16.18) h = f and the Poincaré in-
equality give
Z Z
2
|∇f | = gf ≤ kgk2 kf k2 ≤ Ckgk2 k∇f k2 ,
Ω Ω
so
kT gk1,2 = kf k1,2 ≈ k∇f k2 ≤ ckgk2 .
 FUNCTIONAL ANALYSIS 153

We can regard T as an operator from L2 (Ω) into L2 (Ω)

T : L2 (Ω) → W01,2 (Ω) ⊂ L2 (Ω)
and the Rellich-Kondrachv theorem implies that T is compact. Observe also
that T is one-to-one.
Theorem 16.33. T : L2 (Ω) → L2 (Ω) is compact and self-adjoint.

Proof. We are left with the proof that T is self-adjoint. Let g1 , g2 ∈ L2 (Ω)
and f1 = T g1 , f2 = T g2 . We have to prove that
hT g1 , g2 i = hg1 , T g2 i ,
i.e.
Z Z
(16.19) f1 g2 = g1 f2 .
Ω Ω
Since f1 = T g1 is a weak solution to −∆f1 = g1 we have
Z Z
(16.20) ∇f1 · ∇h = g1 h for h ∈ W01,2 (Ω)
Ω Ω
and similarly −∆f2 = g2 yields
Z Z
(16.21) ∇f2 · ∇h = g2 h for h ∈ W01,2 (Ω).
Ω Ω
The two equalities give
Z Z Z Z
(16.21) (16.20)
f1 g2 = ∇f2 · ∇f1 = ∇f2 ·∇f1 = g1 f2
Ω |{z} Ω |{z} Ω |{z} Ω |{z}
h ∇h ∇h̃ h̃
which is (16.19). 2
Since the operator T : L2 → L2 is compact and self-adjoint we can apply
the spectral theorem.
Definition. Let Ω ⊂ Rn be open and bounded. We say that λ is an eigen-
value of −∆ in W01,2 (Ω) if there is 0 6= g ∈ W01,2 (Ω) such that
−∆g = λg weakly in Ω.
Lemma 16.34. If Ω ⊂ Rn is bounded, the eigenvalues of −∆ in W01,2 (Ω)
are strictly positive.

Proof. Suppose that −∆g = λg weakly in Ω, where 0 6= g ∈ W01,2 (Ω), i.e.

Z Z
∇g · ∇h = λgh for h ∈ W01,2 (Ω).
Ω Ω
Taking h = g we have
Z Z
0< |∇g|2 = λ |g|2
Ω Ω
and hence λ > 0. 2
154 PIOTR HAJLASZ

Note that µ = 0 is not an eigenvalue of T since T is one-to-one. It follows

immediately from the definition of the operator T that µ 6= 0 is an eigenvalue
of T if and only if λ = µ−1 is an eigenvalue of −∆ on W01,2 (Ω). Moreover
the corresponding eigenspaces are the same.
Hence the lemma yields that the eigenvalues of T are strictly positive.
According to the spectral theorem
N
M
L2 (Ω) = ker
| {zT} ⊕ Eµi .
{0} i=1

Since dim Eµi < ∞ we conclude that N = ∞ and hence

∞
M
2
L (Ω) = E µi , µ1 > µ2 > µ3 > . . . , lim µi = 0 .
i→∞
i=1

As we are already observed the eigenspaces Eµi for T are equal to the
eigenspaces Eµ−1 for −∆, so we have
i

Theorem 16.35. If Ω ⊂ Rn is open and bounded, then:

(a) The eigenvalues of the Laplace operator −∆ : W01,2 (Ω) → L2 (Ω) are
positive and form an increasing sequence
λ1 < λ2 < λ3 < . . . , lim λi = ∞ .
i→∞

(b) The dimensions of the corresponding eigenspaces are finite,

lim Eλi < ∞, i = 1, 2, 3 . . .
(c)
∞
M
L2 (Ω) = Eλi
i=1

and hence we may choose an orthonormal basis of L2 (Ω) consisting

of eigenfunctions of −∆.

Let us state an important deep regularity result.

1,2
Theorem 16.36. If Ω ⊂ Rn is open, g ∈ C ∞ (Ω × R) and f ∈ Wloc (Ω) is a
weak solution to
(16.22) −∆f = g(x, f ) in Ω,
i.e. Z Z
∇f · ∇ϕ = f (x, f )ϕ for ϕ ∈ C0∞ (Ω) ,
Ω Ω
then f ∈ C ∞ (Ω).
 FUNCTIONAL ANALYSIS 155

We will not prove this result.

If f is an eigenfunction of −∆, then −∆f = λf in Ω, so f is a solution to
(16.22) with g(x, ξ) = λξ and hence f ∈ C ∞ (Ω). That means eigenfunctions
of −∆ in W01,2 (Ω) are smooth functions in the classical sense.
Example. If n = 1 and Ω = (0, 1), then −∆ = −d2 /dx2 and it is easy to
see that the functions
√
wn (x) = 2 sin(nπx), n = 1, 2, 3, . . .
are eigenfunctions of −∆ and the corresponding eigenvalues are (nπ)2 . The
functions wn form an orthonormal basis in L2 (0, 1). The corresponding ex-
pansion in this basis is the (sinusoidal) Fourier series.
Theorem 16.35 generalizes ro the case of any compact Riemannian mani-
fold M : the eigenfunctions of the Laplace-Beltrami operator are smooth and
they form an orthonormal basis in L2 (M ).56

17. Banach Algebras

Definition. A complex algebra is a complex linear space with a multiplica-

tion such that

(a) x(yz) = x(yz);

(b) x(y + z) = xy + xz, (x + y)z = xz + yz;
(c) α(xy) = (αx)y = x(αy)

for all x, y, z ∈ A and α ∈ C.

If in addition A is a Banach space,
(17.1) kxyk ≤ kxkkyk for all x, y ∈ A
and if there is a unit element e ∈ A such that kek = 1 and
ex = xe = x for all x ∈ A,
then A is called a Banach algebra.
The unit element is uniquely determined. Indeed, if e, e0 are unit elements,
then e = ee0 = e0 . Moreover the multiplication is continuous, i.e. if xn → x,
yn → y, then xn yn → xy. It follows from xn yn − xy = (xn − x)yn + x(yn − y)
and (17.1).
A complex algebra is commutative if
xy = yx for all x, y ∈ A,
but there are many interesting examples of noncommutative algebras.
56For a proof see F.W. Warner, Foundations of differentiable manifolds and Lie groups,
Chapter 6.
156 PIOTR HAJLASZ

Examples. 1. If X is a compact metric space, C(X), the space of complex

valued continuous functions with the supremum norm is a Banach algebra.
2. If U ⊂ C is the unit disc, then the space of functions that are continuous
on U and holomorphic in U is a Banach algebra. It is called the disc algebra.
3. If X is a Banach space, then the space of bounded operators B(X) is a
Banach algebra. If dim X > 1, then B(X) is noncommutative.
4. L1 (Rn ) with the convolution
Z
f ∗ g(x) = f (x − y)g(y) dy
Rn
is a commutative Banach algebra except for the fact that there is no unit
element.
In the definition of the Banach algebra we require that kxyk ≤ kxkkyk,
however we may think of a continuous multiplication that does not sat-
isfy this inequality, but it is not a problem, because we can always find an
equivalent norm such that the inequality is granted.
Theorem 17.1. Assume that A is a Banach space and a complex algebra
with unit element e 6= 0 such that the multiplication is left and right contin-
uous. Then there is an equivalent norm with respect to which A is a Banach
algebra.

Proof. For each x ∈ A we define Mx ∈ B(A) by Mx (z) = xz. It is a

bounded operator since the multiplication is right continuous. The collection
of all such operators
à = {Mx : x ∈ A} ⊂ B(A)
is a complex algebra with the unit element Me = I. Moreover kMe k =
kIk = 1, kMx My k ≤ kMx kkMy k. It is a Banach algebra, because à is a
closed subspace of B(X). Indeed, if Mxi → T ∈ B(A) in norm of B(A),
then Mxi (e) → T (e) and Mxi (z) → T (z) for all z ∈ A as i → ∞ Hence
T (z) ← Mxi (z) = Mxi (e)z → T (e)z as i → ∞
by left-continuity of the multiplication, so T (z) = T (e)z, T = MT (e) ∈ Ã.

The mapping Φ : A → Ã, Φ(x) = Mx is an isomorphism of complex

algebras, i.e. it is an isomorphism of linear spaces and Φ(e) = Me , Φ(xy) =
Φ(x)Φ(y). Moreover, the inverse mapping is bounded, because for x ∈ A
kxk = kMx ek ≤ kMx kkek .
Thus by the open mapping theorem, Φ is an isomorphism of Banach spaces.
Hence kxk0 = kMx k is an equivalent norm on A which turns it into a Banach
algebra. 2
 FUNCTIONAL ANALYSIS 157

Observe that a closed subalgebra of B(X) that contains I is also a Banach

algebra and as a corollary from the above proof we get
Corollary 17.2. Every Banach algebra A is isomorphic to a closed subal-
gebra of B(A).

Definition. A nonzero linear functional φ : A → C on a complex algebra

is called a complex homomorphism if
φ(xy) = φ(x)φ(y) for all x, y ∈ A.
An element x ∈ A is called invertible if there is another element x−1 ∈ A
such that
x−1 x = xx−1 = e.
It is easy to see that the inverse element x−1 , if exists, is uniquely defined.
Proposition 17.3. If φ is a complex homomorphism on a complex algebra
with unit e, then φ(e) = 1 and φ(x) 6= 0 for every invertible x ∈ A.

Proof. Since φ is nonzero, there is y ∈ A such that φ(y) 6= 0 and hence

φ(y) = φ(ey) = φ(e)φ(y), so φ(e) = 1. Moreover, if x ∈ A is invertible, then
φ(x−1 )φ(x) = φ(x−1 x) = φ(e) = 1, so φ(x) 6= 0. 2
Theorem 17.4. Suppose A is a Banach algebra and x ∈ A, kxk < 1. Then

(a) e − x is invertible.
(b)
kxk2
k(e − x)−1 − e − xk ≤ .
1 − kxk
(c) |φ(x)| < 1 for every complex homomorphism φ on A.

Proof. Proof of the part (a) is very similar to that of Theorem 2.7. The
series
X∞
xn = e + x + x2 + x3 + . . .
n=0
converges absolutely, because kxn k ≤ kxkn and kxk < 1. Hence it converges
to an element in A. It is easy to see that this element is an inverse of e − x.
Since
kxk2
k(e − x)−1 − e − xk = kx2 + x3 + . . . k ≤
1 − kxk
the part (b) follows. Finally if λ ∈ C, |λ| ≥ 1, then e − λ−1 x is invertible by
(a) and thus
1 − λ−1 φ(x) = φ(e − λ−1 x) 6= 0
by Proposition 17.3. Hence φ(x) 6= λ for any such λ, so |φ(x)| < 1. 2
Corollary 17.5. Every complex homomorphism on a Banach algebra is
continuous.
158 PIOTR HAJLASZ

Proof. It is a direct consequence of part (c) of the above theorem. 2

It turns out that the property of complex homomorphisms described in
Proposition 17.3 characterizes complex homomorphisms in the case of Ba-
nach algebras.
Theorem 17.6 (Gleason-Kahane-Żelazko). A linear functional57 φ on a
Banach algebra A is a complex homomorphism if and only if φ(e) = 1 and
φ(x) 6= 0 for every noninvertible x ∈ A.

Proof. The implication from left to right is contained in Proposition 17.3,

so we are left with the proof of the implication from right to left.
Let N = ker φ. If x, y ∈ A, then x − φ(x)e, y − φ(y)e ∈ N and hence
(17.2) x = a + φ(x)e, y = b + φ(y)e
for some a, b ∈ N . Applying φ to the product of the two equations we get
(17.3) φ(xy) = φ(ab) + φ(x)φ(y) .
Thus it remains to prove the implications
(17.4) If a, b ∈ N , then ab ∈ N .
Consider a seemingly weaker statement
(17.5) If a ∈ N , then a2 ∈ N .
Lemma 17.7. Property (17.5) implies property (17.4).

Proof. Suppose (17.5) is true. then (17.3) for x = y gives

φ(x2 ) = φ(x)2 for all x ∈ A.
Replacement of x by x + y gives
φ(xy + yx) = 2φ(x)φ(y) for x, y ∈ A.
Thus
(17.6) If x ∈ N , y ∈ A, then xy + yx ∈ N .
The next identity is easy to check
(xy − yx)2 + (xy + yx)2 = 2[x(yxy) + (yxy)x] .
if x ∈ N , then the right hand side belongs to N by (17.6) and since (xy −
yx)2 ∈ N by (17.6) and (17.5), we conclude that (xy + yx)2 ∈ N and hence
xy + yx ∈ N by another application of (17.5). Thus also (xy + yx) + (xy −
yx) = 2xy ∈ N . We proved that if x ∈ N and y ∈ A, then xy ∈ N which
implies (17.4). 2

57We do not assume continuity of φ.

 FUNCTIONAL ANALYSIS 159

Thus we are left with the proof of the property (17.5). The above argu-
ments were purely algebraic and would work in any complex algebra with
unit, however, the proof of (17.5) is analytic.
By the assumptions, there are no invertible elements in N , so ke − xk ≥ 1
for every x ∈ N by Theorem 17.4. Hence
kλe − xk ≥ |λ| = φ(λe − x) for x ∈ N .
Since every element in A is of the form λe−x, x ∈ N (see (17.2) we conclude
that φ is a continuous linear functional of norm 1.
Lemma 17.8. Suppose f is an entire function on C such that f (0) = 1,
f 0 (0) = 0 and
0 < |f (λ)| ≤ e|λ| for λ ∈ C.
Then f (λ) = 1 for all λ ∈ C.

Proof. Since f has no zeroes, there is another entire function g such that
f = eg .
Clearly g(0) = g 0 (0) = 0 and
(17.7) re g(λ) ≤ |λ|.
If |λ| ≤ r, then re g(λ) ≤ r and hence
|re g(λ)| ≤ |2r − re g(λ)| ,
so
|g(λ)|2 = (re g(λ))2 +(im g(λ))2 ≤ (2r −re g(λ))2 +(im g(λ))2 = |2r −g(λ)|2 .
Thus
(17.8) |g(λ)| ≤ |2r − g(λ)| for |λ| ≤ r.
Consider the function
r2 g(λ)
(17.9) hr (λ) = .
λ2 (2r − g(λ))
This fuinction is holomorphic in the disc {λ : |λ| < 2r} by (17.7) and
|hr (λ)| ≤ 1 if |λ| = r by (17.8). Hence the maximum principle gives
(17.10) |hr (λ)| ≤ 1 for |λ| ≤ r.
if we fix λ and let r → ∞ (17.9) and (17.10) gives g(λ) = 0. Hence f (λ) =
e0 = 1. 2
Now we are ready to complete the proof of (17.5). Let a ∈ N . We can
assume that kak = 1. Since φ has norm 1, |φ(an )| ≤ kan k ≤ kakn = 1 and
hence the function
∞
X φ(an ) n
f (λ) = λ , λ∈C
n!
n=0
160 PIOTR HAJLASZ

is entire. Moreover f (0) = φ(e) = 1, f 0 (0) = φ(a) = 0. If we prove that

f (λ) 6= 0 for every λ ∈ C, then the lemma will imply that f 00 (0) = 0, i.e.
φ(a2 ) = 0 which is (17.5).
The series
∞
X an
E(λ) = λn
n!
n=0
converges in norms for every λ ∈ C and hence continuity of φ shows that
f (λ) = φ(E(λ)) for λ ∈ C.
The functional equation E(λ + µ) = E(λ)E(µ) can be proves in the same
way as in the scalar case. In particular E(−λ)E(λ) = E(0) = e, so E(λ)
is invertible for every λ and hence f (λ) = φ(E(λ)) 6= 0, because by the
assumption φ is nonzero on invertible elements. 2
Definition. Let A be a Banach algebra. By G(A) we denote the set os all
invertible elements in A. It is easy to see that G(A) is a group with respect
to the algebra multiplication.
For x ∈ A the spectrum of x is the set
σ(x) = {λ ∈ C : λe − x is not invertible}
and C \ σ(x) is called the resolvent set of x.
As we shall see σ(x) is always nonempty and we define the spectral radius
of x as
ρ(x) = sup{|λ| : λ ∈ σ(x)}.
It easily follows from Theorem 17.4(a) that σ(x) is bounded, i.e. |λ
leqkxk for λ ∈ σ(x), so ρ(x) ≤ kxk.
Lemma 17.9. If x ∈ G(A) and h ∈ A, khk ≤ 21 kx−1 k−1 , then x + h ∈ G(A)
and
(17.11) k(x + h)−1 − x−1 + x−1 hx−1 k ≤ 2kx−1 k3 khk2 .

Proof. x + h = x(e + x−1 h), kx−1 hk < 1/2, so invertibility of x + h follows

from Theorem 17.4(a), so x + h ∈ G(A) and (x + h)−1 = (e + x−1 h)−1 x−1 .
Hence
(x + h)−1 − x−1 − x−1 hx−1 = [(e + x−1 h)−1 − e − x−1 h]x−1
and the inequality (17.11) follows from Theorem 17.4(b). 2
Theorem 17.10. If A is a Banach algebra, then G(A) is an open set and
the mapping x 7→ x−1 is a homeomorphism of G(A) onto G(A).

Proof. The lemma implies that G(A) is open. It also implies that
k(x + h)−1 − x−1 k ≤ 2kx−1 k3 khk2 + kx−1 k2 khk
 FUNCTIONAL ANALYSIS 161

which yields continuity of x 7→ x−1 . Since x 7→ x−1 maps G(A) onto G(A)
and it is its inverse, it follows that x 7→ x−1 is a homeomorphism of G(A)
onto itself. 2
The following lemma is often useful.
Lemma 17.11. let {cn }∞
n=1 be a sequence of nonnegative numbers such that

(17.12) cm+n ≤ cm cn
for all positive integers m, n. Then the limit
lim c1/n
n→∞ n
1/n
exists and equals inf n≥1 cn .

Proof. If cm = 0 for some m, then cn = 0 for n ≥ m and the lemma

follows, so we may assume that cn > 0 for all n. We put c0 = 1. Fix m. Any
integer n can be represented as
n = q(n)m + r(n), 0 ≤ r(n) < m.
Thus (17.12) implies that
1/n
c1/n
n ≤ cq(n)/n
m cr(n) .
Clearly q(n)/n → 1/m as n → ∞ and since cr(n) attains only a finite number
1/n
of positive values, cr(n) → 1 as n → ∞. Hence

lim sup c1/n

n ≤ c1/m
m for any integer m.
n→∞
Thus
lim sup c1/n
n ≤ inf c1/m
m ≤ lim inf c1/n
n
n→∞ m n→∞
and the lemma follows. 2
Let x be an element of a Banach algebra and cn = kxn k, then cn+m ≤ cn cm
and hence the lemma shows that
(17.13) lim kxn k1/n = inf kxn k1/n .
n→∞ n≥1

Theorem 17.12. If A is a Banach algebra and x ∈ A, then

(a) The spectrum σ(x) is compact and nonempty.

(b) The spectral radius satisfies
(17.14) ρ(x) = lim kxn k1/n = inf kxn k1/n .
n→∞ n≥1

Proof. Although the second equality in (17.14) follows from (17.13) we

will not use this fact and we will conclude (17.14) from different arguments.
162 PIOTR HAJLASZ

As we know σ(x) is a bounded set. To prove that σ(x) is compact define

g : C → A by g(λ) = λe − x. Clearly g is continuous and
Ω := C \ σ(x) = g −1 (G(A)) .
The set Ω is open as a preimage of an open set and hence σ(x) is compact
as bounded and closed. Now we define f : Ω → G(A) by
f (λ) = (λe − x)−1 .
We will prove that f (λ) is a homomorphic A-valued function. If we replace
x and h by λe − x and (µ − λ)e in Lemma 17.9, then for µ sufficiently close
to λ inequality (17.11) reads as
kf (µ) − f (λ) − (µ − λ)f 2 (λ)k ≤ 2kf (λ)k3 |µ − λ|2
and hence
f (µ) − f (λ)
lim = −f 2 (λ) ,
µ→λ µ−λ
so f is holomorphic.
If |λ| > kxk, then it is easy to see that
∞
X
−1
(17.15) f (λ) = (λe − x) = λ−n−1 xn .
n=0
The series converges uniformly on any circle Γr centered at 0 of radius
r > kxk and hence Theorem 12.3 allows us to integrate it term by term
Z ∞ Z
1 X
n 1
(17.16) f (λ) dλ = x λ−n−1 dλ = e.
2πi Γr 2πi Γr
n=0
Suppose that σ(x) = ∅, i.e. Ω = C. Then f is an entire function and the
Cauchy formula (12.6) implies that the left hand side of (17.16) equals zero
which is a contradiction. Thus σ(x) 6= ∅ which completes the proof of the
first part of the theorem.
The same term by term integration as in (17.16) gives
Z
1
(17.17) λn f (λ) = xn , r > kxk, n = 0, 1, 2, . . .
2πi Γr
Since λn f (λ) is holomorphic for all λ > ρ(x), the Cauchy theorem (12.6)
implies that (17.17) holds for all r > ρ(x) and thus
kxn k ≤ rn+1 sup kf (reiθ )k, r > ρ(x).
θ∈[0,1]

Hence
lim sup kxn k1/n ≤ r for all r > ρ(x),
n→∞
i.e.
(17.18) lim sup kxn k1/n ≤ ρ(x) .
n→∞
 FUNCTIONAL ANALYSIS 163

Lemma 17.13. If ab and ba are invertible elements in a complex algebra,

then a and b are invertible.

Proof. We will prove that a is invertible as the proof for b is the same.
Since ba(ba)−1 b = bab(ab)−1 we have (ba)−1 b = b(ab)−1 and thus
a[b(ab)−1 ] = e, [b(ab)−1 ]a = (ba)−1 ba = e
so a−1 = b(ab)−1 . 2
Lemma 17.14. If λ ∈ σ(x), then λn ∈ σ(xn ).

Proof. We have
λn e − xn = (λe − x)(λn−1 e + . . . + xn−1 ) = (λn−1 e + . . . + xn−1 )(λe − x) .
If λn 6∈ σ(xn ), then according to the previous lemma λe − x is invertible, so
λ 6∈ σ(x). 2
If λ ∈ σ(x), then λn ∈ σ(xn ) and hence |λ|n = |λn | ≤ kxn k, so ρ(x) ≤
kxn k1/nfor any n = 1, 2, . . ., i.e.
ρ(x) ≤ inf kxn k1/n
n≥1

which together with (17.18) implies (17.14). 2

Theorem 17.15 (Gelfand-Mazur). If A is a Banach algebra in which every
nonzero element is invertible, then A is isometrically isomorphic to C.

Proof. let x ∈ A and λ ∈ σ(x) (σ(x) is nonempty). Hence λe − x = 0, so

x = λe. Thus σ(x) consists of exactly one point. Denote this point by λ(x).
The mapping x 7→ λ(x) is an isometric isomorphism of A onto C. 2

18. Commutative Banach algebras

Definition. A subset J of a commutative complex algebra is called an ideal

if J is a linear subspace of A and xy ∈ J whenever x ∈ A and y ∈ J. J is
a proper ideal if J 6= A and it is a maximal ideal if it is proper and is not
contained in any larger ideal.
The ideals are interesting, because they are kernels of homomorphisms of
commutative Banach algebras and closed ideals are particularly interesting.
Indeed, if : φ : A → B is a homomorphism of commutative Banach algebras
(i.e. φ(xy) = φ(x)φ(y))), then ker φ is an ideal. Moreover the ideal is closed
if φ is continuous.
Conversely every closed ideal J in a Banach algebra A it is a kernel of
a continuous homomorphism. Indeed, let π : A → A/J be the quotient
mapping. A/J is a Banach space, but as we shall see it is also a Banach
algebra and π is a homomorphism with the kernel J.
164 PIOTR HAJLASZ

The multiplication in A/J is defined by

[x][y] = [xy] (i.e. π(x)π(y) = π(xy)).
This multiplication is well defined because if [x] = [x0 ], [y] = [y 0 ], then
x0 − x, y 0 − y ∈ J,
x0 y 0 − xy = (x0 − x)y 0 + x(y 0 − y) ∈ J
and hence [x0 y 0 ] = [xy]. Thus A/J is a complex algebra with unit [e] 6= 0 and
π is a homomorphism with ker π = J. π is continuous since kπ(x)k ≤ kxk by
the definition of the norm in A/J. It remains to prove that A/J is a Banach
algebra, i.e.
k[e]k = 1, k[x][y]k ≤ k[x]kk[y]k.
Let x, y ∈ A. Then by the definition of the quotient norm for every ε > 0
there are x̃, ỹ ∈ J such that
kx + x̃k ≤ k[x]k + ε, ky + ỹk ≤ k[y]k + ε .
Since
xy = (x + x̃)(y + ỹ) − (xỹ + x̃y + x̃ỹ)
| {z }
∈J
we have
k[xy]k ≤ k(x + x̃)(y + ỹ)k ≤ kx + x̃kky + ỹk ≤ (k[x]k + ε)(k[y]k + ε) ,
and thus
(18.1) k[xy]k ≤ k[x]kk[y]k .
Clearly [e] 6= 0 and hence (18.1) gives k[e]k ≤ k[e]k2 , k[e]k ≥ 1, but on the
other hand k[e]k ≤ kek = 1, so k[e]k = 1.
Proposition 18.1.

(a) No proper ideal of A contains an invertible element.

(b) If J is an ideal in a commutative Banach algebra, then its closure J
is also an ideal.

The proof is left to the reader as a simple exercise.

Theorem 18.2.

(a) If A is a commutative complex algebra with unit, then every proper

ideal is contained in a maximal ideal.
(b) If A is a commutative Banach algebra, then every maximal ideal is
closed.

Proof. (a) Let J ⊂ A be a proper ideal. Consider the family of all proper
ideals that contain J. The family is partially ordered by inclusion. According
to the Hausdorff maximality theorem there is a maximal totally ordered
subfamily. The union of this subfamily is also an ideal. It is proper since the
 FUNCTIONAL ANALYSIS 165

unit element does not belong to any of the ideals in the family. Hence the
union is a maximal ideal.
(b) Let M be a maximal ideal. Then M is an ideal. It is proper since M
has no invertible elements and invertible elements form an open set. Hence
M = M by maximality of M . 2
Theorem 18.3. Let A be a commutative Banach algebra, and let ∆ be the
set of all complex homomorphisms of A.

(a) If φ ∈ ∆, the kernel of φ is a maximal ideal of A.

(b) Every maximal ideal of A is the kernel of some φ ∈ ∆.
(c) An element x ∈ A is invertible if and only if φ(x) 6= 0 for all φ ∈ ∆.
(d) An element x ∈ A is invertible if and only if x does not belong to
any proper ideal of A.
(e) λ ∈ σ(x) if and only if φ(x) = λ for some φ ∈ ∆.

Proof. (a) If φ ∈ ∆, then ker φ is a proper ideal. It is maximal because it

has codimension 1.
(b) Let M be a maximal ideal. Hence M is closed, A/M is a Banach
algebra and π : A → A/M is a homomorphism with ker π = M . It remains
to prove that A/M is isomorphic to C. Any nonzero element in A/M is of
the form π(x) for some x ∈ A \ M . Let
J = {ax + y : a ∈ A, y ∈ M } .
Clearly J is an ideal that contains M as a proper subspace. Hence J = A by
the maximality of M . In particular ax + y = e for some a ∈ A, y ∈ M . Thus
π(e) = π(ax + y) = π(a)π(x), so π(x) ∈ A/M is invertible and according to
the Gelfand-Mazur theorem A/M is isometrically isomorphic to C.
(c) If x is invertible and φ ∈ ∆, then 1 = φ(e) = φ(x−1 )φ(x), so φ(x) 6= 0.
If x is not invertible, then {ax : a ∈ A} is an ideal that does not contain e,
and hence it is contained in a maximal ideal M . Now according to (b) there
is φ ∈ ∆ with ker φ = M , so φ(x) = 0.
(d) If x is invertible, then x is not contained in any proper ideal by
Proposition 18.1(a). The converse implication was proved in (c).
(e) It follows from (c) applied to λe − x in place of x. 2
Theorem 18.4. Suppose f1 , f2 , . . . , fn ∈ A(U ) (disc algebra) are such that
(18.2) |f1 |2 + . . . + |fn |2 > 0 on U ,
i.e. the functions f1 , . . . , fn do not have a common zero in U . Then there
are functions φ1 , . . . , φn ∈ A(U ) such that
f1 (z)φ1 (z) + . . . + fn (z)φn (z) = 1 for all x ∈ U .
166 PIOTR HAJLASZ

Proof. A(U ) is a commutative Banach algebra with pointwise multiplica-

tion and the supremum norm.
J = {f1 φ1 + . . . + fn φn : φ1 , . . . , φn ∈ A(U )}
is an ideal in A(U ) and it suffices to prove that J = A(U ). If J 6= A(U ),
then J is contained in a maximal ideal and hence there is a complex homo-
morphism φ ∈ ∆ such that φ vanishes on J. Let g(z) = z ∈ A(U ), kgk = 1.
Since complex homomorphisms have norm 1 (Theorem 17.4(c)), φ(g) = w
for some kwk ≤ 1. Now from a multiplicative property of φ it easily follows
that for every complex polynomial P , φ(P ) = P (w). The density of poly-
nomials in A(U )58 and continuity of φ imply that φ(f ) = f (w) for every
f ∈ A(U ). Since f vanishes on J, 0 = φ(fi ) = fi (w) for i = 1, 2, . . . , n which
contradicts (18.2). 2
Theorem 18.5 (Wiener’s lemma). Let f : Rn → C be the sum of absolutely
convergent multidimensional Fourier series59
X X
(18.3) f (x) = am e2πim·x , |am | < ∞ .
m∈Zn m∈Z
If f (x) 6= 0 for all x ∈ n
R , then
1 X X
= bm e2πim·x , |bm | < ∞ .
f (x) n
m∈Z m∈Zn

Proof. One can easily check that the functions of the form (18.3) form
a Banach algebra A with respect to the pointwise multiplication and the
supremum norm. For each x ∈ Rn ,
(18.4) A 3 f 7→ f (x)
is a complex homomorphism. If we prove that all complex homomorphisms
are of that form, it will follow that the function f satisfies φ(f ) 6= 0, for all
φ ∈ ∆ and hence f is invertible by Theorem 18.3(c) which is what we want
to prove.
For k = 1, 2, . . . , n put gk (x) = e2πixk , where xk is kth coordinate of
x ∈ Rn . Clearly gk , 1/gk ∈ A and both functions have norm 1. If φ ∈ ∆,
then  1    1 
|φ(gk )| ≤ 1 and   = φ  ≤ 1.
   
φ(gk ) gk
Hence there is y ∈ Rn such that
(18.5) φ(gk ) = e2πiyk = gk (y), k = 1, 2, . . . , n.
Every trigonometric polynomial P is a linear combination of products of the
functions gk and 1/gk and since φ is a complex homomorphism (18.5) implies
that φ(P ) = P (y). Since φ is continuous and trigonometric polynomials are
58Prove it.
59m · x = m x + . . . + m x .
1 1 n n
 FUNCTIONAL ANALYSIS 167

dense in A, we get φ(f ) = f (y) for all f ∈ A and hence φ is of the form
(18.4). 2
Definition. Let ∆ be the set of all complex homomorphisms of a commu-
tative Banach algebra A. The formula
x̂(φ) = φ(x) for φ ∈ ∆
assigns to each x ∈ A a function x̂ : ∆ → C called the Gelfand transform of
x.
Let  be the space of all functions x̂ on ∆. The Gelfand topology on ∆ is
the weakest topology for which all the functions x̂ are continuous. Obviously
 ⊂ C(∆).
Since there is a one-to-one correspondence between ∆ and the maximal
ideals in A, ∆ equipped with the Gelfand topology is called the maximal
ideal space of A.
Theorem 18.6. If A is a commutative Banach algebra, then

(a) ∆ is a compact Hausdorff space;

(b) For each x ∈ A, the range of x̂ is σ(x) and hence
kx̂k∞ = ρ(x) ≤ kxk.

Proof. (a) Let A∗ be the dual Banach space and B the closed unit ball in
A∗ .By the Banach-Alaoglu theorem, B is compact in the weak-∗ topology. It
is Hausdorff, because the weak-∗ topology is Hausdorff. Clearly ∆ ⊂ B and
the Gelfand topology is the restriction of the weak-∗ topology to ∆. Thus it
remains to show that ∆ is a closed subset of B in the weak-∗ topology.60
Let φ0 belongs to the closure of ∆ in the weak-∗ topology. We have to
prove that
φ0 (xy) = φ0 (x)φ0 (y), φ0 (e) = 1 .
Fix x, y ∈ A and ε > 0. The set
V = {x∗ ∈ A∗ : |hx∗ , zi i − φ0 (zi )| < ε, i = 1, 2, 3, 4}
where z1 = e, z2 = x, z3 = y, z4 = xy is open in the weak-∗ topology and
φ0 ∈ V. It follows from the definition of the closure that there is φ ∈ ∆ ∩ V,
so |φ(zi ) − φ0 (zi )| < ε for i = 1, 2, 3, 4. In particular
|1 − φ0 (e)| = |φ(e) − φ0 (e)| < ε
gives φ0 (e) = 1. Moreover
φ0 (xy) − φ0 (x)φ0 (y) = (φ0 (xy) − φ(xy)) + (φ(x)φ(y) − φ0 (x)φ0 (y))
= (φ0 (xy) − φ(xy)) + φ(x)(φ(y) − φ0 (y)) + φ0 (y)(φ(x) − φ0 (x))
60The proof of this part is very similar to a corresponding argument in the proof of the
Banach-Alaoglu theorem.
168 PIOTR HAJLASZ

gives
|φ0 (xy) − φ0 (x)φ0 (y)| ≤ (1 + |φ(x)| + |φ0 (y)|) ε
and hence
φ0 (xy) = φ0 (x)φ0 (y) .
(b) λ is in the range of x̂ if x̂(φ) = φ(x) = λ for some φ ∈ ∆ which is
equivalent to λ ∈ σ(x) by Theorem 18.3(e). 2

18.1. C ∗ -algebras.
Definition. Let A be a complex algebra (not necessarily commutative). By
an involution on A we mean a map x 7→ x∗ of A onto itself such that
(x + y)∗ = x∗ + y ∗ , (αx)∗ = αx∗ ,
(xy)∗ = y ∗ x∗ , x∗∗ = x.
If e is a unit element, then one easily verifies that e∗ = e and if x is invertible,
then (x−1 )∗ = (x∗ )−1 .
A Banach algebra A is called a C ∗ -algebra if it is an algebra with involution
that satisfies
kx∗ xk = kxk2 for all x ∈ A .
An element x ∈ A satisfying x = x∗ is called hermitian or self-adjoint.
Note that kxk2 = kx∗ xk ≤ kxkkx∗ k implies kxk ≤ kx∗ k. Hence also
kx∗ k
≤ kx∗∗ k = kxk. Thus
(18.6) kx∗ k = kxk.

Example. If X is a compact Hausdorff space, then C(X) is a commuta-

tive C ∗ -algebra with the involution f 7→ f . The Gelfand-Najmark theorem
(Theorem 18.9) states that every commutative C ∗ -algebra is isometrically
isomorphic to C(X) for some X.
Example. Let H be a Hilbert space. For A ∈ B(H) we define the adjoint
operator A∗ ∈ B(H) to the the unique operator such that
hAx, yi = hx, A∗ yi for all x, y ∈ H.
Clearly
(A + B)∗ = A∗ + B ∗ , (αA)∗ = αA∗
(AB)∗ = B ∗ A∗ , A∗∗ = A .
Moreover
kA∗ Ak = kAk2 .
To prove the last equality observe that
kA∗ Ak = sup hA∗ Ax, yi .
kxk≤1
kyk≤1
 FUNCTIONAL ANALYSIS 169

Since
|hA∗ Ax, yi| = |hAx, Ayi| ≤ kAk2 kxkkyk
we conclude that kA∗ Ak ≤ kAk2 . On the other hand
kAxk2 = hAx, Axi = hA∗ Ax, xi ≤ kA∗ Akkxk2
implies kAk2 ≤ kA∗ Ak. Thus kA∗ Ak = kAk2 .
Therefore B(H) with the adjoint operator as involution is a C ∗ -algebra.
As a direct consequence of (18.6) we have
kA∗ k = kAk.
Theorem 18.7. Let A be a C ∗ -algebra. If x ∈ A is self-adjoint, then σ(x) ⊂
R and ρ(x) = kxk.

Proof. Since kx2 k = kx∗ xk = kxk2 it easily follows that

n n n −n
kx2 k = kxk2 , ρ(x) = lim kx2 k2 = kxk .
n→∞
Let λ ∈ σ(x). Then for any t ∈ R, λ + it ∈ σ(x + ite). We have
|λ + it|2 ≤ kx + itek2 = k(x + ite)(x + ite)∗ k = k(x + ite)(x − ite)k
= kx2 + t2 ek ≤ kx2 k + t2 .
If λ = a + bi, then the inequality reads as
a2 + b2 + 2bt ≤ kx2 k
which is possible only if b = 0, so λ ∈ R. 2
Lemma 18.8. Let A be a commutative C ∗ -algebra. If φ ∈ ∆ is a complex
homomorphism, then
φ(x∗ ) = φ(x) for all x ∈ A.

Proof. If x = x∗ , then φ(x∗ ) = φ(x) ∈ σ(x) ⊂ R by Theorem 18.3(e) and

Theorem 18.7. If x is arbitrary, then
1 1
x = u + iv, where u = (x + x∗ ), v = (x − x∗ ) .
2 2i
Since u = u∗ , v = v ∗ we have
φ(x∗ ) = φ(u) − iφ(v) = φ(u) + iφ(v) = φ(x) .
The proof is complete. 2
Theorem 18.9 (Gelfand-Najmark). Suppose A is a commutative C ∗ -
algebra, with maximal ideal space ∆. Then the Gelfand transform is an
isometric isomorphism of A onto C(∆) such that
(18.7) (x∗ )ˆ = x̂ for all x ∈ A.
In particular x is self-adjoint if and only if x̂ is a real valued function, i.e.
σ(x) ⊂ R.
170 PIOTR HAJLASZ

Proof. Equality (18.7) means that

φ(x∗ ) = φ(x) for all φ ∈ ∆.
and hence it follows from Lemma 18.8.
Let  ⊂ C(∆) be the space of all functions on ∆ of the form x̂.  is self-
adjoint, i.e. it is closed under the complex conjugation, it separates points
and it does not vanish at any point φ ∈ ∆. Hence  is dense in C(∆) by
the Stone-Weierstrass theorem.
Let x ∈ A. Then y = x∗ x is self-adjoint. For φ ∈ ∆ we have
ŷ(φ) = φ(x∗ x) = φ(x∗ )φ(x) = |φ(x)|2 = |x̂(φ)|2
and hence
kŷk∞ = kx̂k2∞ .
Since y is self-adjoint, Theorem 18.6(b) and Theorem 18.7 yield
kx̂k2∞ = kŷk∞ = ρ(y) = kyk = kx∗ xk = kxk2 ,
i.e. kx̂k∞ = kxk. Thus A 3 x 7→ x̂ ∈  is an isometry. Hence  is complete
and thus closed in C(∆). Since  is dense,  = C(∆). 2
Let A be a C ∗ -algebra and x ∈ A. We say that x is normal is x∗ x = xx∗ .
Note that self-adjoint elements are normal. Let B be the closure
P of the space
of all polynomials in x and x∗ , i.e. polynomials of the form i,j aij xi (x∗ )j .
Clearly B is a commutative C ∗ -subalgebra of B generated by x and poly-
nomials in x and x∗ are dense in B.
Denote the spectrum of x with respect to A and B by σA (x) and σB (x).
Clearly σA (x) ⊂ σB (x).
Theorem 18.10. If A is a C ∗ -algebra, x ∈ A is normal and B is defined
as above, then
σA (x) = σB (x).

Proof. We have to prove that if y = λe−x is invertible in A, then y −1 ∈ B.

yy ∗
is self-adjoint and invertible in A. Hence σA (yy ∗ ) ⊂ R is compact, so
Ω = C \ σA (yy ∗ ) is connected. The function
Ω 3 λ 7→ (λe − yy ∗ )−1 ∈ A
is holomorphic and on the open set C \ σB (yy ∗ ) ⊂ Ω it takes values into B.
Hence it follows from Proposition 12.4 that it takes values into B on all of Ω.
Since yy ∗ is invertible in A, 0 ∈ Ω and hence (yy ∗ )−1 = −(0 · e − yy ∗ )−1 ∈ B.
Since yy ∗ = y ∗ y is invertible in B, Lemma 17.13 implies that y is invertible
in B. 2
Theorem 18.11 (Spectral mapping theorem). Let B be a commutative C ∗ -
algebra and x ∈ B be such that polynomials in x and x∗ are dense in B.
Then the mapping x̂ : ∆ → σ(x) is a homeomorphism. Hence C(σ(x)) 3
 FUNCTIONAL ANALYSIS 171

f →
7 f ◦ x̂ ∈ C(∆) is an isometric isomorphism of algebras. Thus for every
f ∈ C(σ(x)) there is unique element Ψf ∈ B such that
(Ψf )ˆ = f ◦ x̂
and hence Ψ defines an isometric isomorphism of C(σ(x)) onto B that sat-
isfies
(18.8) Ψf = (Ψf )∗ for all f ∈ C(σ(x)).
Moreover, if f (λ) = λ on σ(x), then Ψf = x.

Proof. The mapping x̂ : ∆ → σ(x) is a continuous61 surjection by The-

orem 18.6(b). In order to prove that x̂ is a homeomorphism it suffices to
prove that x̂ is one-to-one.62 Suppose φ1 , φ2 ∈ ∆ and x̂(φ1 ) = x̂(φ2 ), i.e.
φ1 (x) = φ2 (x). Hence φ1 (x∗ ) = φ2 (x∗ ) by Lemma 18.8. Since φ1 and φ2 are
homomorphisms it follows that φ1 (P (x, x∗ )) = φ2 (P (x, x∗ )) for any polyno-
mial P . Now the density of such polynomials in B and continuity of φ1 , φ2
imply that φ1 (y) = φ2 (y) for all y ∈ B, i.e. φ1 = φ2 . Thus x̂ : ∆ → σ(x) is
one-to-one and hence homeomorphism. Equality (18.8) follows from (18.7).
If f (λ) = λ, then (Ψf )ˆ = f ◦ x̂ = x̂, so Ψf = x. 2
Corollary 18.12. Let A be a C ∗ -algebra and let x ∈ A be normal. Then

(a) kxk = ρ(x).

(b) x is self-adjoint if and only if σ(x) ⊂ R.

Proof. Let B be a commutative C ∗ -algebra defined as above. Recall that

σA (x) = σB (x) (= σ(x)) and hence ρA (x) = ρB (x) (= ρ(x)).
(a) Let f ∈ C(σ(x)), f (λ) = λ. Then
ρ(x) = kf k∞ = kΨf k = kxk.
(b) follows directly from the last statement in the Gelfand-Najmark theorem.
2
In the situation described in Theorem 18.11 we write
(18.9) Ψf = f (x) .
P i j
If P (λ) = i,j aij λ λ is a polynomial, then it is easy to see that P (x)
defined by (18.9) coincides with63
X
P (x) = aij xi (x∗ )j
i,j

Note that such polynomials P (λ) are dense in C(σ(x)). Theorem 18.11 allows
us to define not only polynomial functions of x, but any continuous functions
61By the definition of the Gelfand topology.
62Because ∆ is compact.
63Prove it.
172 PIOTR HAJLASZ

of x for f ∈ C(σ(x)). Since Ψ : C(σ(x)) → B is an isometric isomorphism

of algebras we have that
(f + g)(x) = f (x) + g(x), (f g)(x) = f (x)g(x), f (x) = (f (x))∗
kf (x)k = kf k∞ .
Moreover
σ(f (x)) = f (σ(x)).
Indeed, according to Theorem 18.3(e), for any y ∈ B, σ(y) = ŷ(∆), so
σ(f (x)) = (f (x))ˆ(∆) = (Ψf )ˆ(∆) = (f ◦ x̂)(∆)
= f (x̂(∆)) = f (σ(x)).
Corollary 18.13. Let A be a C ∗ -algebra and let x ∈ A be self-adjoint with
σ(x) ⊂ [0, ∞). Then there is another self-adjoint element y ∈ A such that
y 2 = x.

Proof. Let B be the closure of polynomials in x and x∗ . Then B is com-

mutative C ∗ -algebra. We will apply Theorem 18.11 to the algebra B. We
have σB (x) = σA (x) ⊂ [0, ∞) by Theorem 18.10. The function f (t) = t1/2
is continuous on σ(x) and hence y = f (x) is well defined. We have
y 2 = f (x)f (x) = (f · f )(x) = x
Moreover
y ∗ = (f (x))∗ = f (x) = f (x) = y,
so y is self-adjoint. 2

18.2. Applications to the spectral theory. Let H be a Hilbert space

and T ∈ B(H) be a normal operator, i.e. T T ∗ = T ∗ T . With each polynomial
X j
(18.10) P (λ) = aij λi λ
i,j
we associate an operator
X
(18.11) P (T ) = aij T i (T ∗ )j
i,j

Let A ⊂ B(H) be the closure of the space of all operators of the form
(18.11). Clearly A is a commutative C ∗ -subalgebra of B(H) generated by
T . Note that polynomials (18.10) are dense in C(σ(T )). As a consequence
of Theorem 18.11 we have
Theorem 18.14 (Spectral mapping theorem). The mapping P 7→ P (T )
uniquely extends to the isometric isomorphism of algebras
C(σ(T )) 3 f 7→ f (T ) ∈ A.
Moreover f (T ) = f (T )∗ and σ(f (T )) = f (σ(T )).

Also Corollary 18.13 can be stated now as

 FUNCTIONAL ANALYSIS 173

Theorem 18.15. If T ∈ B(H) is self-adjoint and σ(T ) ⊂ [0, ∞), then there
is a self-adjoint operator S such that S 2 = T .

It is very useful to identify the class of operators mentioned in Theo-

rem 18.15.
Theorem 18.16. Suppose T ∈ B(H). Then the following conditions are
equivalent:

(a) hT x, xi ≥ 0 for all x ∈ H.

(b) T = T ∗ and σ(T ) ⊂ [0, ∞).

Operators satisfying condition (a) are called positive.

Proof. Suppose (a). Then
hT x, xi = hx, T xi = hx, T xi = hT ∗ x, xi,
so h(T −T ∗ )x, xi = 0 for all x ∈ H and it easily follows from the polarization
identity that T − T ∗ = 0. Thus T is self-adjoint and hence σ(T ) ⊂ R by
Theorem 18.7. In order to prove that σ(T ) ⊂ [0, ∞) it suffices to prove that
for any λ > 0, T + λI is invertible. The inequality at (a) yields
λkxk2 = hλx, xi ≤ h(T + λI)x, xi ≤ k(T + λI)xkkxk,
so
(18.12) k(T + λI)xk ≥ λkxk.
This inequality easily implies that the operator T + λI is one-to-one and
that the range R(T + λI) is closed. It remains to prove that the operator is
surjective, i.e. R(T + λI)⊥ = {0}. If y ∈ R(T + λI)⊥ , the for any x ∈ H,
0 = h(T + λI)x, yi = hx, (T + λI)yi
and hence (T + λI)y = 0, so y = 0 by (18.12).
Suppose now (b). Let S be a self-adjoint operator such that S 2 = T . We
have
hT x, xi = hS 2 x, xi = hSx, Sxi ≥ 0
and hence (a) follows. 2

Department of Mathematics, University of Pittsburgh, 301 Thackeray

Hall, Pittsburgh, PA 15260, USA, [email protected]