Linear Analysis

T. W. Körner
November 15, 2007

Small print The syllabus for the course is defined by the Faculty Board Schedules
(which are minimal for lecturing and maximal for examining). Several of the results are
called Exercises. I will do some as part of the lectures but others will be left to the reader.
In general these are simple verifications or recall results from earlier courses. If you find
that you cannot do one of these, consult your supervisor.
I have sketched solutions for supervisors to the exercises in the five example sheets at
the end. These should be available from the departmental secretaries or in tex, ps, pdf
and dvi format by e-mail.
I should very much appreciate being told of any corrections or possible improvements
and might even part with a small reward to the first finder of particular errors. These
notes are written in LATEX 2ε and should be available in tex, ps, pdf and dvi format from
my home page

http://www.dpmms.cam.ac.uk/˜twk/

My e-mail address is twk@dpmms.

Contents
1 Some inequalities 2

2 In finite dimension, norms are equivalent 5

3 Banach spaces 7

4 Continuous linear functions 9

5 Second duals 11

6 Baire category 13

7 Continuous functions 17

8 The Stone–Weierstrass theorem 19

1
9 Ascoli–Arzelà 21

10 Inner product spaces 23

11 Hilbert space 26

12 The dual of Hilbert space 29

13 The spectrum 30

14 Self-adjoint compact operators on Hilbert space 31

15 Using the spectral theorem 34

16 Where next? 37

17 Books 40

18 First example sheet 40

19 Second example sheet 45

20 Third example sheet 50

21 Fourth example sheet 54

22 Supplementary exercises 58

1 Some inequalities
Inequalities lie at the heart of analysis. In this section we prove some in-
equalities which lie at the heart of this course.
We start with some observations from Part 1A.

Exercise 1.1. Suppose f : (0, ∞) → R is twice differentiable with f ′′ (x) < 0

for all x ∈ R. Then if 0 < t < s and 0 < λ < 1, it follows that


λf (t) + (1 − λ)f (s) < f λt + (1 − λ)s .

In other words, if f ′′ (x) < 0, then f is strictly concave. Applying the

result with f = log, we obtain the following result.

2
Lemma 1.2. Suppose p, q are real and positive with
1 1
+ = 1.
p q
Then, if a, b > 0, we have
a b
a1/p b1/q ≤ +
p q
with equality if and only if a = b.
Our result remains true if a, b ≥ 0.
We can now obtain our first version of the Hölder inequality1 . (Here and
elsewhere we will write F to mean either R or C. This reflects the fact that
many theorems of Linear Analysis apply both to real and complex vector
spaces. However, just as in other branches of analysis and algebra, there are
important theorems which apply only to complex or only to real spaces.)
Theorem 1.3. Suppose p, q are real and positive with
1 1
+ = 1.
p q
If aj , bj ∈ F, then
n n
!1/p n
!1/q
X X X
|aj bj | ≤ |aj |p |bj |q
j=1 j=1 j=1

with equality if and only if we can find A, B ∈ R, not both zero, with A|bj |q =
B|aj |p for all 1 ≤ j ≤ n.
I will probably leave out the examination of the case of equality in The-
orems 1.3 and 1.5 but once you have mastered the main proof it a very
instructive exercise to look closely the cases of equality.
If we set p = 2 in Hölder’s inequality we recover the Cauchy-Schwarz
inequality.
Our Hölder inequality is complemented by a ‘reverse Hölder inequality’.
1
Mathematical nomenclature should not be confused with historic truth. Hölder’s paper
makes it clear that he is discussing an inequality previously proved by L. G. Rogers (1862–
1933). Rogers was Professor of Mathematics at Yorkshire College (now the University of
Leeds) and a fine mathematician ‘with little ambition or desire for recognition’. In 1913,
Ramanujan conjectured some remarkable identities which no-one could prove. In 1917,
whilst looking through old journals, Ramanujan came across a 1894 paper of Rogers in
which a more general form of the identities were proved. Rogers’ election to the Royal
Society followed.

3
Theorem 1.4. (i) Suppose p, q are real and positive with
1 1
+ = 1.
p q
If aj ∈ F and
n n
!1/q
X X
|aj bj | ≤ A |bj |q
j=1 j=1

for all choices of bj ∈ F then

n
!1/p
X
|aj |p ≤ A.
j=1

(ii) (Stronger version) Suppose p, q are real and positive with

1 1
+ = 1.
p q
If aj ∈ F and
 n  n
!1/q
X  X
aj b j  ≤ A |bj |q
 

 
j=1 j=1

for all choices of bj ∈ F then

n
!1/p
X
|aj |p ≤ A.
j=1

Putting Hölder’s inequality and the reverse Hölder inequality together,

we get Minkowski’s inequality.
Theorem 1.5. Suppose p > 1. Then
n
!1/p n
!1/p n
!1/p
X X X
|aj |p + |bj |p ≥ |aj + bj |p
j=1 j=1 j=1

with equality only if we can find α, β ∈ F, not both zero, such that βaj = αbj
for all 1 ≤ j ≤ n.
(Exercise 22.2 contains an apparently simpler proof, but the explicit use
of the reverse Hölder inequality in the proof above may help fix it in the
reader’s mind.)
For us, as for Minkowski, this has a clear geometrical interpretation.

4
Theorem 1.6. Suppose p > 1. Then
n
!1/p
X
kakp = |aj |p
j=1

defines a norm on Fn .
We can extend most of our results (but not all) to the cases p = 1 and
p = ∞.
Exercise 1.7. (i) Show that
n
X
kak1 = |aj |
j=1

defines a norm on Fn .
(ii) Show that
kak∞ = max |aj |
1≤j≤n

defines a norm on Fn .
(iii) Show that, if a, b ∈ Fn , then
n
X
|aj ||bj | ≤ kak∞ kbk1 .
j=1

(iv) State and prove the two ‘reverse Hölder’ inequalities corresponding
to (iii).
(v) If ka + bk1 = kak1 + kbk1 , does it follow that we can find α, β ∈ F,
not both zero, such that βaj = αbj for all 1 ≤ j ≤ n? Give reasons.
(vi) If ka + bk∞ = kak∞ + kbk∞ , does it follow that we can find α, β ∈ F,
not both zero, such that βaj = αbj for all 1 ≤ j ≤ n? Give reasons.

2 In finite dimension, norms are equivalent

We have produced a variety of norms on Fn , but the following theorem (which
the reader may have seen in Part 1B) shows that they are all equivalent in a
rather strong sense (called Lipschitz equivalence).
Theorem 2.1. Suppose E is a finite dimensional vector space over F. If
k.k1 and k.k2 are two norms on E, then we can find a constant K > 0 such
that
Kkxk1 ≥ kxk2 ≥ K −1 kxk1
for all x ∈ E.

5
 Note that the proof depends ultimately on the fundamental axiom of
analysis.
However, the unit balls of the different norms k.kp look more and more
different as the dimension of the space increases.

Exercise 2.2. (i) If ∞ > s > r ≥ 1 show, by applying Hölder’s inequality

with aj = xsj , bj = 1 and p = s/r, that, if xj ≥ 0, we have

n
!1/r n
!1/s
−1 −s−1
X X
xrj ≤ xsj nr
j=1 j=1

with equality if and only if x1 = x2 = · · · = xn .

Thus, if we work in Fn ,
−1 −s−1
kxkr ≤ nr kxks

and we cannot improve the inequality.

(ii) If ∞ > s > r ≥ 1, show, by setting t = xj /kxks in the inequality
t ≥ ts [0 ≤ t ≤ 1], or otherwise, that, if xj ≥ 0,
r

n
!1/s n
!1/r
X X
xsj ≤ xrj
j=1 j=1

and identify the cases of equality.

Conclude that, if we work in Fn ,

kxks ≤ kxkr

and we cannot improve the inequality.

(iii) State and prove (the proofs are easy) the corresponding results when
we allow s = ∞.

Once we reach the simplest infinite dimensional space Theorem 2.1 fails.

Definition 2.3. We write c00 for the space of sequences

a = (a1 , a2 , . . . )

with an ∈ F and only finitely many aj non-zero. [Warning: The notation c00
is not universal.]

6
Exercise 2.4. (i) Check that c00 is a vector space.
(ii) Check that
∞
!1/p
X
p
kakp = |aj |
j=1

defines a norm on c00 for ∞ > p ≥ 1.

(iii) Define an appropriate norm k.k∞ and show that is a norm.
(iv) Show that, given any ∞ ≥ s > r ≥ 1 and any K > 0, we can find a
non-zero a ∈ c00 such that
kakr
> K.
kaks

3 Banach spaces
It is generally thought that the algebraic properties of large objects are rather
dull and this is certainly the case for normed spaces. The remedy is to
introduce extra analytic structure, in this case completeness.
Definition 3.1. A Banach space is a complete normed space.
We observe that the norms we have placed on c00 are not complete.
Exercise 3.2. Verify this statement.
Later, in Lemma 6.8, we shall show that c00 cannot be given a complete
norm. Thus c00 is good for nothing except the production of counterexamples.
(But it is very good for this purpose.)
However, the work we have already done enables us to produce some nice
infinite dimensional Banach spaces.
Theorem 3.3. Let 1 ≤ p < ∞. If lp is the collection of sequences
a = (a1 , a2 , . . . )
with an ∈ F and
∞
X
|aj |p convergent,
j=1

then lp is a vector space with the usual coordinatewise definition of addition

and multiplication by scalars. If we set
∞
!1/p
X
kakp = |aj |p ,
j=1

then k.kp is a complete norm on lp .

7
Lemma 3.4. If l∞ is the collection of bounded sequences

a = (a1 , a2 , . . . )

with an ∈ F, then l∞ is a vector space with the usual coordinatewise definition

of addition and multiplication by scalars. If we set

kak∞ = sup |aj |

1≤j

then k.k∞ is a complete norm on l∞ .

At some point, it became fashionable to write lp instead of lp . The reader

must be prepared for both notations.

Exercise 3.5. (i) Show that, if ∞ ≥ s > r ≥ 1, then lr ⊆ ls , lr 6= ls .

(ii) If ∞ ≥ s > r ≥ 1 show that, if x ∈ lr , then x ∈ ls and kxks ≤
kxkr . Give an explicit example of a sequence x(k) ∈ lr with kx(k)ks = 1 but
kx(k)kr → ∞ as k → ∞.

Although we seek to study infinite dimensional Banach spaces by using

geometrical intuition, they differ in important respects from finite dimen-
sional spaces. Here is one example.

Theorem 3.6. The closed unit ball

B̄ = {x ∈ E : kxk ≤ 1}

of an infinite dimensional Banach space (E, k.k) is not compact.

Our proof depends on two results which are of independent interest.

Lemma 3.7. Any finite dimensional subspace of a Banach space is closed.

Theorem 3.8. [Lemma of F. Riesz] If F is a closed subspace of a Banach

space (E, k.k) and F 6= E, then, given any ǫ > 0, we can find an e ∈ E with
kek = 1 such that
ke − f k > 1 − ǫ
for all f ∈ F .

The technique that we use to prove Theorem 3.8 is very useful in a wide
variety of circumstances.

8
4 Continuous linear functions
When we studied linear maps on finite dimensional normed spaces we made
repeated use of the fact that they were continuous.

Exercise 4.1. Suppose (U, k.kU ) and and (V, k.kV ) are finite dimensional
normed spaces. Show that every linear map α : U → V is continuous.

We cannot make this assumption when we study linear maps on infinite

dimensional spaces.

Exercise 4.2. Consider the space c00 of sequences of complex numbers only
finitely many of which are non-zero, equipped with the norm
∞
X
kak1 = |aj |.
j=1

Give an example of a linear map α : c00 → F which is not continuous.

The treatment of continuous linear maps on infinite dimensional vector

spaces is much aided by some simple observations.

Lemma 4.3. Suppose (U, k.kU ) and and (V, k.kV ) are normed spaces and
α : U → V is linear. The following statements are equivalent.
(i) α is continuous everywhere.
(ii) α is continuous at 0.
(iii) There exists a C such that

kαukV ≤ CkukU

for all u ∈ U .

For this reason continuous linear maps are sometimes called bounded
linear maps.
We can define the operator norm and derive its elementary properties in
exactly the same way as in the finite dimension case.

Definition 4.4. Suppose (U, k.kU ) and (V, k.kV ) are normed spaces and α :
U → V is continuous and linear. We define the operator norm kαk of α as
follows.
kαk = sup kαukV .
kukU ≤1

9
Exercise 4.5. (i) Suppose (U, k.kU ) and (V, k.kV ) are normed spaces. The
collection L(U, V ) of continuous linear maps α : U → V forms a vector space
under pointwise addition and multiplication by scalars. The operator norm
is indeed a norm on L(U, V ).
(ii) Suppose (U, k.kU ) is a normed space. Then the identity map ι ∈
L(U, U ) and kιk = 1.
(iii) Suppose (U, k.kU ), (V, k.kV ) and (W, k.kW ) are normed spaces. If
α ∈ L(U, V ) and β ∈ L(V, W ), then βα ∈ L(U, W ) and

kβαk ≤ kβkkαk.

Theorem 4.6. If
(U, k.kU ) and and (V, k.kV ) are Banach spaces, it follows
that L(U, V ), k.k is a Banach space.
(For an improvement see Exercise 18.8.) A particularly important exam-
ple of the space (L(U, V ), k.k) occurs when V = F.
Definition 4.7. If (U, k.kU ) is a normed space over F we say that a con-
tinuous linear map T : U → F is a bounded linear functional. The space
U ′ = L(U, F, k.k) is called the dual space of U .
Theorem 4.8. If ∞ > p ≥ 1 and p−1 +q −1 = 1, then (lp )′ = lq (More exactly,
there is a natural identification of (lp )′ and lq . When we have introduced more
definitions, you may prefer the statement that there is a natural isometric
isomorphism between (lp )′ and lq .)
The following important exercise illustrates both our method of proving
Theorem 4.8 and its limitations.
Exercise 4.9. (i) Show that c, the set of all sequences a such that aj tends
to a limit as j → ∞, is a closed subspace of l∞ with its usual norm. Show
that c0 , the set of all sequences a such that aj → 0 as j → ∞, is a closed
subspace of l∞ with its usual norm.
Because of these results we may consider c and c0 as Banach spaces under
the supremum norm (that is to say the l∞ norm).
(ii) Show that c′0 can be identified in a natural manner with l1 .
(iii) Identify the elements of c′ in a reasonably natural manner.
It is an important fact that the method we use to prove Theorem 4.8 fails
when we try to find the dual of l∞ . In some sense l∞ is too large for us to
deal with. Lemma 4.12 make it clear what is going on.
Definition 4.10. We say that a subset E of a metric space (X, d) is dense
in X, if given any x ∈ X, we can find xn ∈ E with xn → x.

10
Definition 4.11. A metric space (X, d) is separable if we can find a countable
dense subset of X.
Lemma 4.12. The space lp is separable if ∞ > p ≥ 1 but l∞ is not separable.

5 Second duals
Students find it hard to understand the treatment of second duals in 1B
Algebra but the subject actually becomes easier to understand when treated
in a more general setting.
Lemma 5.1. Let U be a normed vector space. Then the map J : U → U ′′
given by
(Ju)(T ) = T u
for u ∈ U and T ∈ U ′ is a well defined continuous linear map.
In 1B Algebra, when U is a finite dimensional space it is easy to show
that J is injective. In the infinite dimensional case we need a further idea.
Definition 5.2. [Warning. This is non-standard.] We say that a dual U ′ of
a normed space U is sufficiently rich if, whenever u ∈ U ,

sup kT uk = kuk.
T ∈U ′ ,kT k≤1

It is easy to check that all the spaces we have met so far are sufficiently
rich.
Exercise 5.3. (i) Show that the duals of the lp spaces are sufficiently rich
[1 ≤ p ≤ ∞]. (Note that, although we cannot fully identify the dual of l∞ we
can, nonetheless, show that it is sufficiently rich.)
(ii) Show that if U is a Banach space with sufficiently rich dual then
L(U, U ) has sufficiently rich dual. (We need this later in the proof of Lemma 13.8
but the reader should not loose any sleep over this.)
The reader will find it easy to check (and should carry out the check)
that any specific space has a sufficiently rich dual.
If we use the axiom of choice, then it can be shown that all duals are
sufficiently rich. Any text that the reader consults will use the axiom of
choice and will therefore omit the condition that the dual is sufficiently rich.
Lemma 5.4. Let U be a normed vector space with a sufficiently rich dual.
Then the map J : U → U ′′ introduced in Lemma 5.1 is isometric (that is to
say kJukU ′′ = kukU ) and so injective.

11
 (Here and everywhere else, we use the operator norm on U ′ and U ′′ .)
If U ′ is sufficiently rich, it is reasonable to use J to give natural identifi-
cation of U with JU and write U ⊆ U ′′ .
Lemma 5.5. If U is a Banach space with sufficiently rich dual then (with
the natural identification) U is a closed subspace of U ′′ .
In 1B Algebra, when U is a finite dimensional space, a dimensional ar-
gument shows that J is surjective (and so, with the natural identification,
U = U ′′ ). If 1 < p < ∞ then the work we have already done shows that
(lp )′′ = lp .
However, the following important example shows that U may be a proper
subspace of its second dual U ′′ .
Exercise 5.6. (Part (i) and most of part (iii) were done in Exercise 4.9.)
Let c0 be the the subset of l∞ consisting of those sequences a such that that
aj → 0 as j → ∞.
(i) Show that c0 is a closed subspace of l∞ and so (c0 , k.k∞ ) is a Banach
space.
(ii) Show that c0 is separable.
(iii) Show that c′0 can be identified in a natural manner with l1 . Show
also that the dual of c0 is sufficiently rich.
(iv) Deduce that c′′0 can identified in a natural manner with l∞ and so the
mapping J : c0 → c′′0 introduced in Lemma 5.1 cannot be surjective.
There are two ways of looking at Banach spaces. One is to study each
space in its own right using the the language and insights of linear analysis
as tools. Each of the spaces l1 , l2 and l∞ could be, and has been, the object
of a lifetime’s study. This is the point of view of the present course or, at
least, the present lecturer2 .
The second way of looking at Banach spaces is to study them as general
structures. In this case we do not study individual spaces but isomorphism
classes or isometric isomorphism classes of Banach spaces.
Definition 5.7. (i) The normed spaces (U, k.kU ) and (V, k.kV ) are isomor-
phic if there exists a vector space isomorphism T : U → V such that both T
and T −1 are continuous.
(ii) The normed spaces (U, k.kU ) and (V, k.kV ) are isometrically isomor-
phic if there exists a vector space isomorphism T : U → V such that
kT ukV = kukU
for all u ∈ U .
2
Who has spent most of his mathematical life studying a particular representation of
1
l .

12
 From this point of view, all we know so far about the lp spaces is that l∞ is
not isomorphic to lp for 1 ≤ p < ∞. (If the reader is interested, Exercise 18.9
shows that l1 is not isomorphic to lp for p 6= 1 and Exercise 22.5 shows that
l2 is not isomorphic to lp for p 6= 2. It has been shown that, in fact, spaces
lp with distinct p are not isomorphic.) We give a very simple example of
Banach space isomorphism in Exercise 6.15.
For the avoidance of doubt, the reader is instructed that, both in the
notes and in examination questions, statements about lp and similar spaces
refer to their concrete realisations and not to their isomorphism classes. If
isomorphism is to be considered, this will be stated explicitly.

6 Baire category
The Baire category is a profound triviality which condenses the folk wisdom
of a generation of ingenious mathematicians into a single statement.

Theorem 6.1. If (X, d) is a complete

T∞ metric space and Uj is an open dense
subset of X for each j ≥ 1, then j=1 Uj 6= ∅.

(The same ideas are used to prove a similar result for compact topological
spaces in Exercise 19.5.)
In some sense,the property of belonging to Uj is stable (since Uj is open,
small perturbations leave us within Uj ) and the property of not belonging
to Uj unstable (since Uj is dense, we can move to Uj by arbitrarily small
perturbations).

Exercise 6.2. Consider the space of n×n complex matrices with the operator
norm. Recall that given any matrix A we can find a non-singular matrix B
such that BAB −1 is upper triangular. By using this result, or otherwise, show
that the set of of matrices with n distinct eigenvalues is open and dense.

For historical reasons Baire’s theorem is associated with some rather un-
helpful nomenclature.

Definition 6.3. Consider

S∞ a metric space (X, d). If Ej is closed with dense
complement and E ⊆ j=1 Ej , then E is said to be of first category.

Baire’s theorem can be restated as follows.

Theorem 6.4. [Baire’s category theorem] If (X, d) is a complete metric

space, then X is not of first category.

The following remark is very useful.

13
Lemma 6.5. Consider a metric space (X, d). The countable union of sets
of the first category is of first category.
The next remark that may already have occurred to the reader.
Exercise 6.6. If (X, d) is a complete metric space and E is a subset of first
category, then X \ E is dense in E.
Recall that a point x in a metric space (X, d) is called isolated if we can
find a δ > 0 such that d(x, y) < δ implies y = x.
Lemma 6.7. A complete metric space without isolated points is uncountable.
Observe that this gives us a new proof that R is uncountable (and so
transcendental numbers exist) which does not depend on establishing decimal
representation.
Lemma 6.8. (i) If E is an infinite dimensional Banach space over F, then
E cannot have a countable spanning set. In other words, we cannot find a
sequence e1 , e2 , . . . in E such that every u ∈ E can be written
N
X
u= λj ej
j=1

for some λj ∈ F and some N ≥ 1.

(ii) The space c00 cannot be given a complete norm.
Exercise 6.9. Consider the Banach space (lp , k.kp ). We know that consid-
ered as a set, lr is a subset of lp whenever p ≥ r ≥ 1. With this convention
r p
S
p>r l is of first category in (l , k.kp ).

Banach and Steinhauss used the Baire category theorem to isolate another
piece of folk wisdom.
Theorem 6.10. [Principle of uniform boundedness] Suppose (U, k.kU )
and (V, k.k) are Banach spaces. If we have a family T of continuous linear
maps T : U → V such that supT ∈T kT ukV < ∞ for each u ∈ U , then
supT ∈T kT k < ∞.
Here is a typical use of the principle. We work on the circle T = R/2πZ
(but if the reader prefers she may work on [−π, π]). To see why this re-
sult may be interestiong recall applied lecturers writing down the following
‘aspirational prose’.
We have gn → δ, that is to say the continuous function gn tends to the delta
function, and so Z Z
gn (t)f (t) dt → gn (t)δ(t) dt = f (0).

14
Exercise 6.11. Suppose that gn : T → R is continuous and
1
Z
gn (t)f (t) dt → f (0)
2π T
as n → ∞ for all continuous functions f : T → R. Then the following must
be true.
(i) There exists a constant K such that
1
Z
|gn (t)| dt ≤ K
2π T
for all n ≥ 1.
(ii) If δ > 0 and f is a continuous function with f (t) = 0 for |t| < δ, then
1
Z
f (t)gn (t) dt → 0
2π T
as n → ∞.
(iii) We have
1
Z
gn (t) dt → 1
2π T
as n → ∞.
[In Exercise 22.7 we establish that these necessary conditions are also suf-
ficient. In Exercise 22.9 we use Exercise 6.11 to establish that the Fourier
series of a continuous function need not converge pointwise to that function.]
We use the Baire category theorem to prove the following series of rather
more subtle results.
Theorem 6.12. [Open mapping theorem] Suppose (U, k.kU ) and (V, k.k)
are Banach spaces. If T ∈ L(U, V ) is surjective, then T maps open sets in
U to open sets in V .
Exercise 6.13. It is easy to see that linearity is essential for results like
these. Give an example of a continuous surjective map f : R → R which is
not open.
We give an example of the use of the open mapping theorem in Exer-
cise 18.11 The reader may recall a very useful ‘open mapping theorem’ in
complex variable theory. The following is an immediate consequence of The-
orem 6.12.
Theorem 6.14. [Inverse mapping theorem]3 Suppose that (U, k.kU ) and
(V, k.kV ) are Banach spaces. If T ∈ L(U, V ) is bijective, then T −1 is contin-
uous (so T is an isomorphism).
3
Called the inversion theorem in the syllabus.

15
 (For a variation on this theme, see Exercise 21.4.)
Here is a simple example of the use of Theorem 6.14.

Exercise 6.15. The space c of sequences with limits and the space c0 of
sequences with limit zero (both equipped with the supremum norm) are Banach
space isomorphic.

We introduce the last of this group of theorems with an exercise.

Exercise 6.16. (i) Let (X, d) be a metric space. If f : X → X is continuous,

then the graph 

x, f (x) : x ∈ X
is closed with respect to the product metric.
(ii) If g : R → R is given by g(x) = x−2 for x 6= 0 and g(0) = 0, then the
graph 

x, g(x) : x ∈ X
is closed in the usual metric but g is not continuous.

Theorem 6.17. [Closed graph theorem] Suppose (U, k.kU ) is Banach

space and T : U → U is a linear function. If the graph


u, T (u) : u ∈ U

is closed with respect to the product norm, then T is continuous.

To see how such a theorem can be used, we recall some definitions and
results from from 1B algebra. The reader can check that they apply without
change in the infinite dimensional case.

Exercise 6.18. If U is a vector space, we say that a linear map P : U → U

is a projection if P 2 = P . Show that, for such a P ,

(I − P )−1 (0) = P (U ) and P −1 (0) = (I − P )(U ).

Show further that every u ∈ U can be written uniquely in the form u = v + w

with v ∈ P (U ) and w ∈ P −1 (0).

Theorem 6.19. Suppose (U, k.kU ) is Banach space and P : U → U is a

projection. Then P is continuous if and only if the kernel P −1 (0) and image
P (U ) are closed.

16
7 Continuous functions
We recall the discussion of continuous functions in the Topological and Metric
Spaces course.
Exercise 7.1. Let (X, τ ) and (Y, σ) be topological spaces. The following two
statements about a function f : X → Y are equivalent.
(i) If U ∈ σ, then f −1 (U ) ∈ τ .
(ii) Given x ∈ X and V ∈ σ with f (x) ∈ V , we can find a W ∈ τ with
x ∈ W and F (W ) ⊆ V .
Any f satisfying the conditions of Exercise 7.1 is called continuous. We
shall be interested in continuous functions F : X → F where (X, τ ) is a
topological space4 and F has its usual topology.
Even if the reader has not seen the next three exercises before, she should
have no difficulty with them.
Exercise 7.2. (i) Let (X, τ ) be a topological space and let fn : X → F be
continuous. Suppose that f : X → F is such that we can find ǫn → 0 with
|fn (x) − f (x)| < ǫn for all x ∈ X.
Show that f is continuous. (In other words, the uniform limit of continuous
functions is continuous.)
(ii) Let C0 (X) be the space of bounded continuous functions f : X → F.
Show that
kf k∞ = sup |f (x)|
x∈X

defines a complete norm on C0 (X).

Note that the completeness of l∞ is a special case where X = N and τ is
the discrete topology.
Exercise 7.3. If (X, τ ) is compact, show that every continuous function
f : X → R is bounded.
Exercise 7.4. Show that, if E is subset of Fn with the usual topology, then
every continuous function f : E → Fn is bounded if and only if E is compact.
These results strongly suggest that we should study the space C(X) =
CF (X) of continuous functions f : X → F with the uniform norm kf k∞ =
supx∈X |f (x)| in the case when X is compact.
However, if we simply demand that X is compact, the space C(X) may
not have much to do with the set X.
4
As usual, we shall sometimes merely refer to X with the topology τ being understood.

17
Exercise 7.5. If X has the indiscrete topology τ = {X, ∅}, then C(X)
consists of the constant functions.
The following simple observation puts us on a profitable path.
Exercise 7.6. If C(X) is such that, given x 6= y, we can find an f ∈ C(X)
with f (x) 6= f (y) (informally, if C(X) separates the points of X), then X is
Hausdorff.
In this section we prove the remarkable fact that the converse also holds
for compact spaces. Thus it is natural to study C(X) when X is compact
and Hausdorff5 .
We need to recall a couple of elementary topological results.
Exercise 7.7. (i) In a compact space, every closed set is compact.
(ii) In a Hausdorff space, singleton sets {a} are closed.
We now start our theorem sequence.
Theorem 7.8. If (X, τ ) is compact and Hausdorff, then, given A and B
non-empty disjoint closed sets, we can find disjoint open sets U and V such
that A ⊆ U and B ⊆ V .
(A space satisfying the conclusions of Theorem 7.8 is called normal. See
Exercise 22.11 for more on this topic.)
Theorem 7.9. [Urysohn’s lemma] If (X, τ ) is compact and Hausdorff,
then, given A and B non-empty disjoint closed sets, we can find an f ∈
CR (X) such that 0 ≤ f (x) ≤ 1 for all x ∈ X and

f (a) = 1 when a ∈ A
f (b) = 0 when b ∈ B.

Exercise 7.7 now tells us that C(X) separates points whenever X is com-
pact and Hausdorff. It is, perhaps, worth remarking that Urysohn’s lemma
has a much simpler proof if τ is derived from a metric.
The following simple remark comes in useful in our proof of Urysohn’s
lemma.
Exercise 7.10. Let (X, τ ) be a topological space and let R have its usual

topology. A function f :
X → R is continuous if and only if f −1 (−∞, a)
is open and f −1 (−∞, a] is closed for all a ∈ R.
5
This result so impressed a retired French general that he proposed using the word
‘compact’ to mean ‘compact and Hausdorff’. The innovation was not popular but the
reader should be aware of this possible source of confusion.

18
 In fact we can prove an apparently stronger result than Urysohn’s lemma.
Theorem 7.11. [Tietze’s extension theorem] If Y is closed subset of a
compact Hausdorff space (X, τ ), then, given any f ∈ CR (Y ) (where Y has
the subspace topology), we can find an F ∈ CR (X) such that F (y) = f (y) for
all y ∈ Y .
To see that Tietze’s extension theorem is non-trivial consider the following
example.
Exercise 7.12. Consider the closed interval X = [−4, 4] with the usual
topology and the open interval Y = (0, 1). Show that if f : Y → R is defined
by f (y) = sin(1/y) then f is continuous but there does not exist an F ∈ C(X)
such that F (y) = f (y) for all y ∈ Y .
We strengthen Theorem 7.11 in two steps.
Corollary 7.13. If Y is closed subset of a compact Hausdorff space (X, τ ),
then, given any f ∈ CF (Y ), we can find an F ∈ CF (X) such that F (y) = f (y)
for all y ∈ Y .
Corollary 7.14. If Y is closed subset of a compact Hausdorff space (X, τ ),
then, given any f ∈ CF (Y ), we can find an F ∈ CF (X) such that F (y) = f (y)
for all y ∈ Y and kF k∞ = kf k∞ .

8 The Stone–Weierstrass theorem

Unless the reader has lead a very sheltered life she will have done the following
important exercise many times before. (If not, she should do it at once.)
Exercise 8.1. [Cauchy’s example] Let E(x) = exp(−1/x2 ) for x 6= 0 and
E(0) = 0.
(i) Show that E is infinitely differentiable on R \ {0} with
E (n) (x) = Pn (1/x)E(x)
for some polynomial Pn .
(ii) Show that E is infinitely differentiable everywhere with E (n) (0) = 0
for all n.
(iii) Use the fact that a power series is infinitely P
differentiable term by
term to show that we cannot find aj ∈ R with E(x) = ∞ j
j=−∞ aj x .

(Exercise 19.7, which uses the Baire category theorem from a later section,
provides an even stronger result.) Weierstrass must, therefore, have been
delighted to prove the following result.

19
Theorem 8.2. The set of real polynomials is uniformly dense in CR ([a, b]).
In other words, given any continuous real function f : [a, b] → R and any
ǫ > 0, we can find a real polynomial with
|P (t) − f (t)| < ǫ
for all t ∈ [a, b].
When Stone was asked to contribute an article to the first issue of the
American Mathematical Monthly, he produced the following far reaching ex-
tension of Weierstrass’s theorem.
Theorem 8.3. [The Stone–Weierstrass theorem] Consider a compact
Hausdorff space X. Suppose that A is a subspace of CR (X) with the following
properties.
(i) If f, g ∈ A then f × g ∈ A.
(ii) 1 ∈ A.
(iii) If x, y ∈ X then we can find an f ∈ A such that f (x) 6= f (y).
Then A is dense in (CR , k.k∞ ).
(If A is a subspace of C(X) satisfying (i), we sometimes say that A is a
subalgebra of C(X).)
Our proof of the Stone–Weierstrass theorem makes use of the following
fact.
Lemma 8.4. We can find aj ∈ R such that
∞
X
(1 − x)1/2 = aj x j
j=0

for all real x with |x| < 1.

Our version of the Stone–Weierstrass theorem deals with real valued func-
tions. The following example shows that it will not apply in the complex case
without modification.
Example 8.5. We work in the complex plane C. Let
D̄ = {z ∈ C : |z| ≤ 1} and D = {z ∈ C : |z| < 1}.
We write A(D̄) for the set of f ∈ C(D̄) such that f is analytic on D. Then
A(D̄) is a subspace of CC (D̄) with the following properties.
(i) If f, g ∈ A(D̄) then f × g ∈ A.
(ii) 1 ∈ A(D̄).
(iii) If z, w ∈ D̄ then we can find an f ∈ A(D̄) such that f (z) 6= f (w).
However, A(D̄) is not uniformly dense in C(D̄).

20
 Instead we produce the following variation.

Theorem 8.6. [The complex Stone–Weierstrass theorem] Consider a

compact Hausdorff space X. Suppose that A is a subspace of CC (X) with the
following properties.
(i) If f, g ∈ A, then f × g ∈ A.
(ii) 1 ∈ A.
(iii) If x, y ∈ X, then we can find an f ∈ A such that f (x) 6= f (y).
(iv) If f ∈ A, then its complex conjugate f ∗ ∈ A.
Then A is dense in (CC (X), k.k∞ ).

The following exercise gives a typical application and clears up matters

left vague in the 1B methods course.

Exercise 8.7. We work on the circle T = R/2πZ. If f : T → C is continuous

we write
1
Z
fˆ(n) = f (t) exp(−int) dt.
2π T
(i) The collection of trigonometric polynomials nj=−n aj exp(ijt) is uni-
P
formly dense in CC (T).
(ii) (Uniqueness of Fourier series.) If f, g ∈ CC (T) and fˆ(n) = ĝ(n) for
all n ∈ Z, then f = g.
(iii) If f ∈ CC (T) and ∞
P ˆ
n=−∞ |f (n)| converges, then

∞
fˆ(n) exp(int)
X
f (t) =
n=−∞

for all t ∈ T.

Exercise 19.3 gives another example of Stone–Weierstrass in action.

9 Ascoli–Arzelà
It is frequently possible to show that a problem can be solved ‘apart from an
error which can be made as small as we like’. Under these circumstances an
appeal to compactness, if available, will often show that the problem has an
exact solution.
The Ascoli–Arzelà theorem enables us to characterise the compact subsets
of C(X) when X is a compact metric space.

21
Definition 9.1. Let (X, τ ) be a topological space and (Y, ρ) a metric space.
We say that a collection F of functions f : X → Y is equicontinuous at x if
given ǫ > 0 we can find a U ∈ τ with x ∈ U such that
y ∈ U implies ρ(f (x), f (y)) < ǫ for all f ∈ F.
If F is equicontinuous at all points of X we say that F is equicontinuous.
Exercise 9.2. If (X, d) and (Y, ρ) are metric spaces, write out the definition
of equicontinuity in ǫ, δ form.
Theorem 9.3. [Ascoli–Arzelà] Let (X, τ ) be a compact Hausdorff space.
Then a subset F of C(X) is compact under the the uniform norm if and only
if both the following conditions hold.
(i) F is closed and bounded in the uniform norm.
(ii) F is equicontinuous.
We shall prove the Ascoli–Arzelà theorem by a direct attack. A cleaner
proof depending on results from the Metric and Topological course is given
in Exercise 19.11 but the basic ideas of the two proofs are the same.
A typical example of the use of these ideas appears in the proof of the
following nice result.
Theorem 9.4. If η > 0 and f : [x0 − η, x0 + η] × [y0 − η, y0 + η] → R is
continuous, then we can find a δ with η ≥ δ > 0 and a differentiable function
φ : (x0 − δ, x0 + δ) → R
such that φ(x0 ) = y0 and
φ′ (t) = f (t, φ(t))
for all t ∈ (x0 − δ, x0 + δ).
In Part 1B we used the contraction mapping theorem (another idea from
‘abstract analysis’) to prove the following theorem.
Theorem 9.5. If η > 0, K > 0 and f : [x0 − η, x0 + η] × [y0 − η, y0 + η] → R
satisfies the Lipschitz condition.
|f (x, y) − f (x′ , y)| ≤ K|x − x′ |
for all x, x′ ∈ [x0 − η, x0 + η] and all y ∈ [y0 − η, y0 + η], then we can find a
δ with η ≥ δ > 0 and a unique differentiable function
φ : (x0 − δ, x0 + δ) → R
such that φ(x0 ) = y0 and
φ′ (t) = f (t, φ(t))
for all t ∈ (x0 − δ, x0 + δ).

22
Exercise 9.6. By using the mean value theorem, establish that if

f : [x0 − η, x0 + η] × [y0 − η, y0 + η] → R

has continuous first partial derivative ∂f (x, y)/∂x, then it satisfies a Lipschitz
condition.

Our new theorem establishes existence under much more general condi-
tions that those of Theorem 9.5, but the solution need not be unique.

Exercise 9.7. The differential equation x′ (t) = 3x(t)2/3 has more than one
solution with x0 = 0.

It is helpful, when considering the form of our proof for Theorem 9.4, to
observe that if there are different solutions of the equations then a series of
‘approximate solutions’ may switch between approximating one solution and
another.
Even in the Lipschitz case we cannot hope to prove more than the exis-
tence of local solutions since no global solution may exist.

Exercise 9.8. Find all the solutions of x′ (t) = (1 + x(t)2 ). Observe that
there is no solution which is valid over an interval of length greater than π.

10 Inner product spaces

Since the reader’s first arrival in Cambridge she has been bombarded with
inner product spaces. In this section we recall some of the results she already
knows. She should check where appropriate that the results hold in infinite
dimensional spaces.

Definition 10.1. Let V be a vector space over C. Suppose that exists a map
p : V 2 → C such that, writing hu, vi = p(u, v) we have
(i) hλ1 u1 + λ2 u2 , vi = λ1 hu1 , vi + λ2 hu2 , vi for all λ1 , λ2 ∈ C, u1 , u2 , v ∈
V.
(ii) hu, vi = hv, ui∗ for all u, v ∈ V .
(iii) hu, ui ≥ 0 for all u ∈ V .
(iv) hu, ui = 0 implies u = 0.
Then we say that (V, p) is an inner product space. We call p an inner
product.

A similar definition applies with C replaced by R except that the complex

conjugation in condition (ii) is superfluous.

23
Exercise 10.2. (i) (Cauchy–Schwarz) If V is an inner product space then

|hu, vi|2 ≤ hu, uihv, vi

with equality if and only if u and v are linearly dependent.

(ii) If V is an inner product space then

kuk22 = hu, ui, kuk2 ≥ 0

defines a norm on V .
(iii) (Parallelogram law) With the notation of (ii)

ku + vk22 + ku − vk22 = 2 kuk22 + kvk22 ).

We derived the norm from the inner product but the process can be
reversed and we can recover the inner product from the norm.

Exercise 10.3. [The polarisation identity] With the notation and as-
sumptions of Exercise 10.2,

4hu, vi = ku + vk22 − ku − vk22 + i(ku + ivk22 − ku − ivk22 )

for all u v ∈ V .

(For an interesting sidelight see Exercise 22.12.)

Definition 10.4. Let V be an inner product space.

(i) If u, v ∈ V and hu, vi = 0 we say that u and v are orthogonal and
write u ⊥ v.
(ii) A collection E of vectors is said to be orthonormal if, whenever e, f ∈
E (
0 if e 6= f ,
he, f i =
1 if e = f .

We have the following extensions of Pythagoras’s theorem.

Exercise 10.5. Consider an inner product space V . Suppose e1 , e2 , . . . , en

are orthonormal vectors in V and f ∈ V . Then
n
X n
X
kf − λj ej k22 ≥ kf k22 − |hf, ej i|2
j=1 j=1

with equality if and only if λj = hf, ej i.

24
Theorem 10.6. [Bessel’s inequality] Consider an inner product space V .
Suppose e1 , e2 , . . . is an orthonormal sequence of vectors in V and f ∈ V .
Then
∞
X
|hf, ej i|2 ≤ kf k22
j=1

with equality if and only if

N
X
f − hf, ej iej → 0


j=1 2
as n → ∞.
We illustrate these familiar general results with a familiar special case.
Note that Exercise 10.7 (iii) resolves a problem left open by the 1B mathe-
matical methods course.
Exercise 10.7. We work on T = R/2πZ.
(i) Show that if f ∈ CR (T), f (t) ≥ 0 for all t and
1
Z
f (t) dt = 0,
2π T
then f (t) = 0 for all t.
(ii) Show that the formula
1
Z
hf, gi = f (t)g(t)∗ dt = 0
2π T

defines an inner product on CC (T). From now on we consider CC (T) with

this inner product.
(iii) Show that, if we write ej (t) = exp ijt, then the ej are orthonor-
mal. By using the fact that the trigonometric polynomials are dense in
(CC (T), k.k∞ ) (Exercise 8.7), show that the trigonometric polynomials are
dense in (CC (T), k.k2 ). Hence show that
N

X
f − hf, ej iej → 0


j=−M 2

as M, N → ∞.
(iv) (Parseval’s formula) Use (iii) to show that, if we write fˆ(j) = hf, ej i,
then ∞
1
Z
ˆ
X
2
|f (n)| = |f (t)|2 dt
j=−∞
2π T

25
for all f ∈ CC (T). Show also that
∞
1
Z
fˆ(n)ĝ(n)∗ =
X
f (t)g(t)∗ dt.
j=−∞
2π T

However, we note the following important fact.

Exercise 10.8. (C(T), k.k2 ) is not complete.

This result needs careful proof. We need to show, not that Cauchy se-
quence does not converge to the obvious answer, but that it does not converge
to any continuous function. The next exercise illustrates this remark.

Exercise 10.9. Write

(
1 − 2n |t| for |t| ≤ 2−n ,
∆n (t) =
0 otherwise.

Show that,if we define fn ∈ C(T) by

n
X
fn = n∆n (t − 2πr/n),
j=1

then kf k2 → 0.

11 Hilbert space
The work of this section depends on the following key result.

Theorem 11.1. Let V be an infinite inner product space. The following

statements are equivalent.
(i) V is separable.
(ii) There exists an orthonormal sequence ej such that

Xn
f − hf, ej iej → 0


j=1 2

as n → ∞ for all f ∈ V .

Our proof calls on an old friend, the Gramm–Schmidt orthogonalisation

process.

26
Exercise 11.2. Suppose V is an inner product space. If e1 , e2 , . . . , en are
orthonormalPand f ∈ V then either
(i) f = Pnj hf, ej iej and f ∈ span{e1 , e2 , . . . , en }, or
(ii) f 6= nj hf, ej iej in which case f ∈
/ span{e1 , e2 , . . . , en }. In this case,
setting
Xn
u=f− hf, ej iej
j

and en+1 = kuk−1

2 u, we have e1 , e2 , . . . , en+1 orthonormal and

span{e1 , e2 , . . . , en+1 } = span{e1 , e2 , . . . , f }.

From now on, if

n
X
f − fj → 0,


j=1 2
we feel free to write ∞
X
f= fj .
j=1

Exercise 11.3. (Uniqueness) Let V be an infinite inner product space. If

we have an orthonormal sequence ej , then, if λj ∈ F,
∞
X
λj ej = 0
j=1

implies λj = 0 for all j.

[Note this is result of analysis and not of algebra since it involves limits.]

Definition 11.4. If U is an inner product space, we say that an orthonormal

sequence ej in U is a basis6 (or more exactly an orthonormal basis) for U if
∞
X
x= hx, ej iej
j=1

for all x ∈ U .

We immediately obtain the following remarkable result.

6
NB This is not an algebraic basis.

27
Theorem 11.5. [Riesz–Fisher] All separable complete infinite dimensional
inner product spaces are inner product isomorphic. More precisely, if U and
V are separable complete infinite dimensional inner product spaces with inner
products pU and pV , then there exists a linear map T : U → V such that

pV (T x, T y) = pU (x, y)

for all x, y ∈ U . We note that T is automatically an isometric Banach space

isomorphism.
Since all separable complete infinite dimensional inner product spaces are
isomorphic, we simply talk about the Hilbert7 space H. Sometimes people
talk about complete inner product spaces which are not separable and are
then careful to talk about ‘non-separable Hilbert spaces’ but the study of
such large spaces has not yet been very profitable. (If you want to see such
a space, consult Exercise 22.17.)
Our arguments also give the following results more or less for free.
Exercise 11.6. Consider l2 . If a, b ∈ l2 then ∞ ∗
P
j=1 aj bj is absolutely con-
vergent. Further
X∞
ha, bi = aj b∗j
j=1

defines an inner product which induces the norm k.k2 . With this inner prod-
uct, l2 is (inner product isomorphic to) Hilbert space.
Exercise 11.7. Let U be a separable infinite dimensional inner product space.
Then there exists an inner product preserving linear map J : U → H of U
into the Hilbert space H such that J(U ) is dense in H.
If the reader knows about such things, she will be able to restate Ex-
ercise 11.7 as the observation that the completion of a separable infinite
dimensional inner product space is (inner product isomorphic to) Hilbert
space.
Lemma 11.8. If U is an inner product space, with basis ej , then U is com-
plete if and only if
∞
X
xj ej
j=1
P∞
converges whenever j=1 |xj |2 converges.
7
Hilbert developed the theory H in a non-abstract way for particular purposes. There
is a, no doubt apocryphal, story of his asking ‘What is this Hilbert space which the young
people are talking about?’.

28
12 The dual of Hilbert space
We already know that Hilbert space is isometrically isomorphic to l2 and we
know that l2 has dual space isometrically isomorphic to itself. Thus the dual
space of Hilbert space is isometrically isomorphic to itself.
However, Hilbert space is the infinite dimensional space in which our
geometrical intuition has freest play and it is instructive to follow a geometric
path to a closely related result. Not only does this avoid the inelegant use of
specific bases, but it provides additional insight into the structure of Hilbert
space8 ..
Theorem 12.1. Let U be a complete inner product space. If F is closed
subspace and a ∈ U , then we can find a unique f0 ∈ F such that
ka − f0 k2 ≤ ka − f k2
for all f ∈ F .
(See also Exercises 20.4 and 20.5.)
Lemma 12.2. With the hypotheses and notation of Theorem 12.1, f0 ∈ F
is the unique element of F such that a − f0 is orthogonal to every element of
F.
We immediately deduce the following pleasing result.
Theorem 12.3. [Riesz representation] If U is a complete inner product
space and T ∈ U ′ (that is to say, T : U → F is a continuous linear map),
then there is a unique w ∈ U with
T u = hu, wi
for all u ∈ U .
Exercise 12.4. If U is a complete inner product space and we define by
J(v)u = hu, vi
for all u, v ∈ U , then J(v) ∈ U ′ for all v ∈ U and J has the following
properties.
(i) J(λ1 v1 + λ2 v2 ) = λ∗1 J(v1 ) + λ∗2 J(v2 ) for all λ1 , λ2 ∈ C and all v1 , v2 ∈
V . (We say that J is anti-linear.)
(ii) kJ(v)k = kvk2 for all v ∈ V .
(iii) J is surjective.
8
Since we do not use bases, our results will also apply to non-seperable Hilbert spaces
but the reader may ignore this.

29
 Thus (using the polarisation identity of Exercise 10.3) J : U → U ′ is an
inner product anti-isomorphism and U ′ is naturally anti-isomorphic to U .
(If the reader is interested, but only if she is interested, she may glance at
Exercise 22.15.)
Theorem 12.1 and Lemma 12.2 also give us information on orthogonal
complements which will be used later.
Lemma 12.5. If F is a closed subspace of a Hilbert space H, then
F ⊥ = {g ∈ H : hg, f i = 0 for all f ∈ F }
is a closed subspace of H. Every u ∈ H can be written in one and only one
way as
u=f +g
with f ∈ F and g ∈ F ⊥ .

13 The spectrum
When we studied linear maps α : Cn → Cn , we were particularly interested
in those λ ∈ C such that α − λι was not invertible. This interest carries
over to infinite dimensional spaces9 . The elementary theory is no harder
for general Banach spaces than for Hilbert spaces10 so we shall work in the
general context.
Definition 13.1. If U is a Banach space over C and T : U → U is a
continuous linear map, we define the spectrum σ(T ) of T by
σ(T ) = {λ ∈ C : T − λI not invertible}.
/ σ(T ), then (T − λI)−1 is
The inverse mapping theorem shows that, if λ ∈
a continuous linear map. The structure of the spectrum can be exceedingly
intricate but some useful general results can be obtained by applying the
following simple ‘master theorem’.
Theorem 13.2. If U is a Banach space
P∞ over C and T : U → U is a contin-
j
uous linear map with kT k < 1, then j=0 T converges in the uniform norm
and I − T is invertible with
∞
X
−1
(I − T ) = T j.
j=0

9
Even in the finite dimensional case, the study of such things in real vector spaces
turned out to be less interesting, so we shall stick to complex Banach spaces.
10
However is only true for the elementary theory.

30
Lemma 13.3. If U is a Banach space over C and T : U → U is a continuous
linear map, then σ(T ) is bounded.
Lemma 13.4. If U is a Banach space over C and T : U → U is a continuous
linear map, then σ(T ) is closed.
Definition 13.5. If U is a Banach space over C and T : U → U is a
continuous linear map, we say that λ is an eigenvalue of T if
ker(T − λI) 6= {0}.
If u is a non-zero element of ker(T − λI), we call u an eigenvector with
associated eigenvalue λ.
Exercise 13.6. With the notation just introduced, every eigenvalue of T lies
in σ(T ).
Example 13.7. (i) If K is a non-empty closed bounded set in C, then we
can find a continuous linear map T : l2 → l2 with σ(T ) = K.
(ii) We can find a continuous linear map T : l2 → l2 such that σ(T ) = {0}
but 0 is not an eigenvalue.
Recall that every Banach space we have studied has a sufficiently rich
dual in the sense of Definition 5.2.
Lemma 13.8. If U is a Banach space over C with sufficiently rich dual and
T : U → U is a continuous linear map, then σ(T ) is non-empty.
If we were prepared to develop complex analysis for L(U, U ) valued func-
tions from scratch, we could replace Lemma 13.8 by a stronger and simpler
result.
Lemma 13.9. If U is a Banach space over C and T : U → U is a continuous
linear map, then σ(T ) is non-empty.

14 Self-adjoint compact operators on Hilbert

space
In the previous section we developed the elementary theory of the spectrum
for general Banach spaces. From now on we are only interested in Hilbert
space.
The reader will recall the very pretty theory of diagonalisation for self-
adjoint (that is to say, Hermitian) maps α : V → V on finite dimensional
inner product spaces. We conclude this course by developing a parallel theory
for Hilbert space. We need two definitions of which only the second is really
new.

31
Definition 14.1. Let H be a Hilbert space. A continuous linear map T :
H → H is called self-adjoint (or Hermitian) if

hT x, yi = hx, T yi

for all x, y ∈ H.

Exercise 14.2. The eigenvalues of a self-adjoint continuous linear map are

real.

Definition 14.3. Let H be a Hilbert space. A continuous linear map T :

H → H is called compact if Cl(T (B)), the closure of the image under T of
the unit ball B = {x : kxk ≤ 1}, is compact.

Exercise 14.4. Let H be a Hilbert space. Show that a continuous linear map
T : H → H is compact, if given any xn ∈ H with kxn k ≤ 1, we can find
n(j) → ∞ and a y ∈ H such that

kT xn(j) − yk2 → 0.

Exercise 21.7 gives some insight into what the compact operators11 look
like.
We state the theorem which we wish to prove.

Theorem 14.5. [The spectral theorem] Let H be a Hilbert space. If

T : H → H is a continuous linear compact self-adjoint map, we can find an
orthonormal basis en of eigenvectors whose associated eigenvalues λn are real
and satisfy the condition λn → 0 as n → ∞.

Exercise 14.6. Show that the following is an equivalent statement of Theo-

rem 14.5.
Let H be a Hilbert space. If T : H → H is a continuous linear compact
self-adjoint map we can find an orthonormal basis en and a sequence λn of
real numbers with λn → 0 such that
∞
X
Tu = λj hu, ej iej
j=1

for all u ∈ H.

We give yet another equivalent form in Exercise 21.9.

11
A continuous linear map T : H → H is called an operator.

32
Exercise 14.7. Let H be a Hilbert space. with orthonormal basis en . If λn
is a sequence of real numbers with λn → 0, then the equation
X∞
Tu = λj hu, ej iej
j=1

defines a continuous linear compact self-adjoint map T : H → H.

The proof of Theorem 14.5 parallels its finite dimensional analogue but
additional work is required. For the rest of the section we work in a Hilbert
space H and ‘T is an operator’ will mean that T : H → H is a continuous
linear map.
Lemma 14.8. If T is a self-adjoint operator, then
sup |hx, T xi| = kT k.
kxk2 =1

Lemma 14.9. If T is a compact self-adjoint operator, then at least one of

kT k or −kT k is an eigenvalue.
The next result recalls 1B Linear Algebra.
Exercise 14.10. If T is a self-adjoint operator and e is an eigenvector for
T , then, writing
e⊥ = {f ∈ H : he, f i = 0},
we know that T (e⊥ ) ⊆ e⊥ . The map T |e⊥ : e⊥ → e⊥ is self-adjoint and, if T
is compact, so is T |e⊥ .
Lemma 14.11. If T is a compact operator then, given any ǫ > 0, T has only
finitely many orthonormal eigenvectors with associated eigenvalues having
absolute values greater than ǫ.
Putting these these results together we obtain the spectral theorem for
compact self-adjoint operators.
Exercise 14.12. Suppose that T is a compact self-adjoint (ie Hermitian)
operator. Consider the following properties which T may or may not have.
(A) T −1 (0) = {0}.
(B) T −1 (0) has dimension r for some r ≥ 1.
(C) T −1 (0) has infinite dimension.
(a) T has infinitely many eigenvalues.
(b) T has s eigenvalues for some s ≥ 1.
Which of the pairs (X,y) can be true of T and which cannot? Give reasons
or examples.
This completes the course, but I have added two extra sections. The first
is an extended exercise on the use of the spectral theorem which is strongly
recommended to the reader.

33
15 Using the spectral theorem
In mathematical methods you studied Sturm–Liouville equations
 
d ′
p(t)y (t) + q(t)y(t) = f (t)
dt
on an interval [a, b] subject to conditions

A1 y(a) + A2 y ′ (a) = 0, B1 y(b) + B2 y ′ (b) = 0

with (A1 , A2 ) 6= (0, 0), (B1 , B2 ) 6= (0, 0), p continuously differentiable, f , q

continuous and p(t) > 0 for all t ∈ [a, b]. You showed that it is generally
possible to find a continuous Green’s function G : [a, b]2 → R with G(s, t) =
G(t, s) such that
Z b
y(t) = G(s, t)f (s) ds
a
solves the given problem.
We shall not into the details here. (They are in [3] §19 and in [5].) The
next exercise gives a particular case.
Exercise 15.1. (i) If G : R2 → R is differentiable and g(t) = G(t, t), write
down g ′ (t).
(ii) By using the fundamental theorem of the calculus and differentiation
under the integral show, that under conditions on F that you should specify,
Z t Z t
d ∂F
F (s, t) ds = F (t, t) + (s, t) ds.
dt a a ∂t

(iii) Show that, if

(
(1 − s)t if 1 ≥ t ≥ s ≥ 0,
G(s, t) =
s(1 − t) if 1 ≥ s > t ≥ 0,

then, if f : [0, 1] → R is continuous,

Z 1
y(t) = f (s)G(s, t) ds
0

defines a twice differentiable function with y(0) = y(1) = 0 and

y ′′ (t) = f (t)

for t ∈ [0, 1].

34
 We now investigate the equation
Z b
y(t) = G(s, t)f (s) ds
a

using the methods of linear analysis.

Exercise 15.2. Suppose that G : [a, b]2 → R is continuous. Show that, if

f : [a, b] → C is continuous, then Lf : [a, b] → C given by
Z b
Lf (t) = G(s, t)f (s) ds
a

is continuous.

Exercise 15.3. (This is a reprise of parts Exercises 10.7 and 10.8.) Show
that, if we set Z b
hf, gi = f (t)g(t)∗ dt,
a

we obtain C([a, b]) as an inner product space. Show that C([a, b]) is an infinite
dimensional separable inner product space but is not complete.

Exercise 15.4. We consider C([a, b]) both with the uniform norm k.k∞ and
the inner product derived norm k.k2 . We shall use the Cauchy–Schwarz in-
equality for integrals repeatedly.
(i) Show that

L : (C([a, b]), k.k2 ) → (C([a, b]), k.k∞ )

is a continuous linear map.

(ii) Show that the collection of Lf such that f ∈ C([a, b]) and kf k2 ≤ 1
is equicontinuous.
(iii) Show (Exercise 19.3 is relevant) that, if G(s, t) = G(t, s) for all
t, s ∈ [a, b], then
hLf, gi = hf, Lgi
for all f, g ∈ (C([a, b]).

We know that C([a, b]) is not a complete inner product space, so we

cannot apply the spectral theorem directly. However, Exercise 11.7 tells us
that there exists an inner product preserving linear map J : C([a, b]) → H
of U into the Hilbert space H such that J(C([a, b])) is dense in H.

35
Exercise 15.5. The results of this exercise are not hard but the reader should
not sleep walk through them.
(i) Show that, if u ∈ H, un ∈ C([a, b]) and kJun − uk2 → 0, then Lun
converges uniformly in C([a, b]) to a continuous function g say. Show that if
vn ∈ C([a, b]) and kJvn − uk2 → 0, then
kLvn − gk∞ → 0.
Thus we can write L̃u = g.
(ii) Show that L̃ is a well defined function L̃ : H → C([a, b]).
(iii) Show that
L̃ : H → (C([a, b]), k.k∞ )
is a continuous linear map.
(iv) Show that the collection of L̃f such that f ∈ H and kf k2 ≤ 1 is
equicontinuous.
Exercise 15.6. We now define L̆ = J L̃.
(i) Show that
L̆ : H → H.
is a continuous linear map.
(ii) Show that L̆ is compact.
(iii) From now on we suppose G(s, t) = G(t, s) for all s, t ∈ [a, b]. Show
that L̆ is self-adjoint.
(iv) Deduce that we can find an orthonormal basis wn and a sequence λn
of real numbers with λn → 0 such that
∞
X
L̆u = λj hu, wj iwj
j=1

for all u ∈ H.
The result of the previous exercise tells us something about L̆, which is an
operator on H, and we are interested in L, which is an operator on C([a, b]).
However, this is soon remedied.
Exercise 15.7. (i) If λj 6= 0, use the fact that λj wj = L̆wj to show that
wj = Jej for some ej ∈ C([a, b]).
(ii) Conclude that, if G : [a, b]2 → R is continuous and G(s, t) = G(t, s),
then either we can find an orthonormal sequence vj in C([a, b]) and a sequence
ζj of non-zero real numbers with ζj → 0 having the property
Z
b XN
f (s)G(s, .) ds − ζj hf, vj ivj → 0


a
j=1 2

36
as N → ∞, or we can find a finite orthonormal collection vj in C([a, b]) and
ζj non-zero real numbers with
Z b X
f (s)G(s, t) ds = ζj hf, vj ivj (t).
a

(iii) Show that hf, vj i = 0 for all j implies f = 0 whenever f ∈ C([a, b])
if and only if Z b
f (t)G(s, t) dt = 0
a

for all s implies f = 0 whenever f ∈ C([a, b]).

Exercise 15.8. Briefly identify the ‘eigenfunctions’ vj associated with non-
zero eigenvalues in the case of Exercise 15.1.

16 Where next?
In this section which will neither be examined nor lectured, I look at the
different ways in which the ideas of this course can be developed.
Measure theory Measure theory interacts with linear analysis in many ways.
(1) We have seen that C([a, b]) with the usual inner product can be iden-
tified with a dense subset of of a complete inner product space. It is a
surprising fact that we can realise this complete inner product space as a
space of functions L2 ([a, b]) on [a, b] by using Lebesgue integration.
In much the same way, it can be shown that, if we write
Z b 1/p
p
kf kp = |f (x)| dx ,
a

then k.kp is a norm on C([a, b]) (see Exercise 18.1) and the completion gives
rise in a natural manner to a space of Lp ([a, b]) on [a, b]. These spaces are
natural subjects for linear analysis.
(2) Although we have studied the space C([a, b]) with the the uniform
norm, we did not try to identify its dual (for some members of the dual
see Exercise 18.7). It is not hard to show that that the dual space can be
identified with the space of Borel measures.
(3) The theory of compact self-adjoint operators that we have developed
on Hilbert space corresponds to the theory of Fourier sums
∞
fˆ(j) exp(ijt).
X
f (t) ∼
j=−∞

37
If we are to get something like the theory of Fourier transforms with the
putative inversion formula
Z ∞
1
f (t) ∼ fˆ(x) exp(ixt) dx
2π −∞
we need to extend our notions of integration. (In fact, we only need to
extend the ideas of Riemann integration, but, nonetheless, the extension is
quite subtle.)
Examples Spaces like lp and C([a, b]), k.k∞ ) are good examples of Banach
spaces to start with because they have a great deal of structure. For the
same reason, they are inadequate if we wish to understand what a general
Banach space might look like.
As a result of seventy years of hard work (including that of our own
Professor Gowers) we know that the good behaviour of lp , C([a, b]), k.k∞ )
and similar spaces is not typical of Banach spaces in general. The study of
Banach spaces (like the study of most general mathematical objects) requires
a plentiful stock of examples.
The axiom of choice The reader will be aware of a principle called the axiom
of choice 12 . This asserts that, given a non-empty collection A of non-empty
sets, we can find a function
[
f :A→ A
A∈A

such that f (A) ∈ A. (That is to say, f chooses an element f (A) from each
A ∈ A.) Mathematical logicians have shown that, if the ordinary axioms for
set theory are consistent, then they remain consistent if we add the axiom of
choice but that the axiom of choice is not implied by the ordinary axioms.
It turns out that the general study of Banach spaces takes a more elegant
form if we assume the axiom of choice and, for this reason, it is customary
to assume it. Here are some consequences of this assumption.
Theorem 16.1. Assuming the axiom of choice, every vector space V over F
has an algebraic basis (that is to say a subset E such that any v ∈ V may be
written uniquely as a finite sum
X
v= λg g
g∈G

with G a finite subset of E and λg ∈ F).

12
For historical reasons this axiom has acquired an air of glamour and mystery which it
it hardly deserves.

38
 Using this theorem, it is easy to prove the following result which reinforces
the lesson of Exercise 4.2.

Theorem 16.2. Assuming the axiom of choice, if U is any infinite dimen-

sional normed vector space over F, then there exists a linear map α : U → F
which is not continuous.

We can also prove the following supplement to Exercise 2.4.

Theorem 16.3. Assuming the axiom of choice, we can find an infinite di-
mensional vector space U and two complete norms k.kA k.kB on U such that

kukA kukB
sup = sup = ∞.
u6=0 kukB u6=0 kukA

(For example we can set up an algebraic isomorphism between l2 and l∞ .)

The axiom of choice also enables us to prove a beautiful result of Hahn–
Banach. We shall not discuss this but here are some of its consequence. The
first result sheds light on the paragraph following Theorem 4.8.

Lemma 16.4. We work in l∞ and define en ∈ l∞ by

(
1 if j = n,
enj =
0 otherwise.

Assuming the axiom of choice, there there exists a non-zero continuous linear
functional T : l∞ → C such that

T en = 0

for all n.

The second consequence was already stated in the discussion of Defini-

tion 5.2.

Theorem 16.5. Assuming the axiom of choice, every Banach space has a
sufficiently rich dual.

The strengths and weaknesses of linear analysis using the axiom of choice
are well illustrated by Lemma 16.4. On the one hand, it asserts the existence
of an object T without giving any clue as to what it looks like. On the other
hand, if we did not know the result of Lemma 16.4, we could waste an awful
lot of time trying to show that no such object exists.

39
17 Books
There are many excellent introductions to linear analysis. The book of Bol-
lobás [1] has the advantage of being based on this course and a subsequent
Part III course. I think that [3] and [2] are nice and reasonably simple.
If you wish to learn more about Hilbert space then [5] is an excellent
introduction and, if you simply want to learn more analysis in a non-exam
driven way, then Rudin’s Real and Complex Analysis is a masterpiece.

References
[1] B. Bollobás, Linear Analysis, CUP, 1991.

[2] C. Gofman and G. Pedrick, A First Course in Functional Analysis, Pren-

tice Hall, 1965. (This is now reissued by AMS, Chelsea)

[3] J. D. Pryce Basic Methods of Linear Functional Analysis, Hutchinson,

1973. (Out of print but should be in college libraries.)

[4] W.Rudin Real and Complex Analysis, McGraw Hill, 2nd Ed, 1974.

[5] N. Young An Introduction to Hilbert Space, CUP, 1988.

18 First example sheet

Students who are unsure of their ground should check that they can do the
exercises in the main text. Strong students should at least glance at the
supplementary example sheet. The order of the exercises roughly follows the
order of the lectures.

Exercise 18.1. In this exercise, ∞ > p > 1 and p−1 + q −1 = 1. We work

with the space C([a, b]) of continuous functions on [a, b].
(i) Prove Hölder’s integral inequality
Z b Z b 1/p Z b 1/q
p q
|f (t)g(t)| dt ≤ |f (t)| dt |g(t)| dt
a a a

for all f, g ∈ C([a, b]).

(ii) State and prove an appropriate reverse form of Hölder’s integral in-
equality.

40
 (iii) Show that
Z b 1/p
p
kf kp = |f (t)| dt
a

defines a norm on C([a, b]).

(iv) Show that (C([a, b]), k.kp ) is not complete. (We shall consider the
particular case p = 2 in Exercise 10.8.)
(v) By applying Hölder’s integral inequality with g = 1, p = v/u, or
otherwise, show that
−1 −v −1 )
kF ku ≤ (b − a)(u kF kv

when ∞ > v > u > 1.

(vi) Show that, if ∞ > v > u > 1, then, given any K > 0, we can find
an f ∈ C([a, b]) such that
kf kv > Kkf ku .
[Note that the inequalities in (v) and (vi) run in the opposite way to the lp
case.]
(vii) [Optional extra] Show that, if ∞ > v > u > 1, then given K > 0, we
can find continuous functions f and g which are zero outside some interval
such that
Z ∞ 1/u Z ∞ 1/v
u v
|f (x)| dx >K |f (x)| dx ,
−∞ −∞
Z ∞ 1/v Z ∞ 1/u
v u
|g(x)| dx >K |g(x)| dx .
−∞ −∞

Exercise 18.2. Suppose 1 > p > 0.

(a) Find (x1 , x2 ), (y1 , y2 ) ∈ R2 such that

1/p
(x1 + y1 )p + (x2 + y2 )p ) > (xp1 + xp2 )1/p + (y1p + y2p )1/p .

(b) Show, by considering the behaviour of 1 + tp − (1 + t)p , or otherwise,

that, if a, b ≥ 0, then
ap + bp ≥ (a + b)p .
(c) Show that, if we write lp for the space of complex sequences a with
∞
X
|aj |p < ∞,
j=1

then lp can be made into a vector space in the standard way.

41
 Show that, if we set
∞
X
d(a, b) = |aj − bj |p ,
j=1

then d is a complete metric on lp .

Exercise 18.3. Show that we can find a constant An such that

sup |p′ (t)| ≤ An sup |p(t)|

t∈[0,1] t∈[0,1]

for every real polynomial of degree n or less.

Exercise 18.4. Let E and F be normed spaces. Let A be a dense subset of

E, and let Tn : E → F be a continuous linear map for each n ≥ 1. Show
that if
(a) there exists a K with kTn k ≤ K for all n, and
(b) Tn (a) → 0 for all a ∈ A,
then Tn (e) → 0 for all e ∈ E.
Is the result true if condition (a) is dropped? Give a proof or a counterex-
ample.
If (a) and (b) hold, does it follow that kTn k → 0 as n → ∞? Give a proof
or a counterexample.

Exercise 18.5. (A useful fact.) LetP(V, k.k) be a normed space. P

Show that
∞
it is a Banach space if and only if j=1 xj converges whenever ∞ j=1 kxj k
converges.
In the special case when V = C and kzk = |z| deduce that absolute
convergence implies convergence.

Exercise 18.6. (i) Consider C([0, 1]) with the uniform norm. Show that

E = {f ∈ C([0, 1]) : f (0) = 0}

is a closed subspace of C([0, 1]) and explain why this means that E is a
Banach space under the uniform norm.
Show that  Z 1 
F = f ∈E : f (t) dt = 0
0

is a closed subspace of E. Show that there does not exist a g ∈ E such that
kgk∞ = 1 and
kg − f k∞ ≥ 1

42
for all f ∈ F .
Thus Theorem 3.8 cannot be improved in general.
(ii) Show, however, that, if F is a subspace of a finite dimensional normed
space space (E, k.k) and F 6= E, then we can find an e ∈ E with kek = 1
such that
ke − f k ≥ 1.
for all f ∈ F .

Exercise 18.7. In this question we work with real valued continuous func-
tions although similar results hold for the complex valued case.
(i) Consider the space (C([a, b]), k.k∞ ). Show that, if s ∈ [a, b] and

δs (g) = g(s),

then δs ∈ C([a, b])′ . What is kδs k and why? Can you find a g ∈ C([a, b])
with kgk∞ = 1 and δs (g) = kδs k? Give reasons. Show that C([a, b]) has a
sufficiently rich dual in the sense of Definition 5.2.
(ii) Consider the space (C([a, b]), k.k∞ ). If F ∈ C([a, b]) set
Z b
TF (g) = F (t)g(t) dt.
a

Show that TF ∈ C([a, b])′ . What is kTF k and why? Can you always find a
g ∈ C([a, b]) with kgk∞ = 1 and TF (g) = kTF k? Give reasons.
(iii) Consider the space (C([a, b]), k.k1 ) where, as usual,
Z b
kgk1 = |g(t)| dt.
a

If δs and TF are defined as before, show that δs is not continuous, but TF is.
(iv) [Optional extra] Continuing with the ideas of (iii), find kTF k and
prove your answer.

Exercise 18.8. (i) Show that if (U, k.kU ) is a normed space and (V, k.kV ) is
a Banach space, then (L(U, V ), k.k) is a Banach space.
(ii) Consider c00 (the space of sequences with all but finitely many terms
zero) with the norm
X∞
kak∗ = |aj |
j=1

and the space l1 with its usual norm. Let L(l1 , C00 ) be defined as in Theo-
rem 4.6.

43
 If we set
Tn (a) = (a1 , 2−1 a2 , 3−1 a3 , . . . , n−1 an , 0, 0, . . . ),
show that Tn ∈ L(l1 , c00 ). Show that the Tn form a Cauchy sequence in
L(l1 , c00 ) with no limit point.
Thus Theorem 4.6 may fail if (V, k.k) is not complete.
Exercise 18.9. If T : U → V is an isomorphism between the Banach spaces
U and V (that is to say, a linear bijection such that T and T −1 are con-
tinuous), show that the map T ′ : V ′ → U ′ between the dual spaces given
by
T ′ (v ′ )u = v ′ (T u)
for all v ′ ∈ V ′ and u ∈ U is a well defined isomorphism between V ′ and U ′ .
(Observe, that, on general grounds, the verification must consist of routine
and rather easy steps.)
Deduce that l1 cannot be isomorphic to lp for any p > 1.
Exercise 18.10. Suppose that X, Y and Z are Banach spaces. Suppose that
F : X × Y → Z is linear and continuous in each variable separately, that is
to say that, if y is fixed,
F (., y)) : X → Z
is a continuous linear map and, if x is fixed,
F (x, .) : Y → Z
is a continuous linear map. Show, by using the principle of uniform bound-
edness, that there exists an M such that
kF (x, y)kZ ≤ M kxkX kykY
for all x ∈ X, y ∈ Y . Deduce that F is continuous.
Exercise 18.11. Suppose that U is a vector space with two complete norms
k.kA and k.kB . By applying the open mapping theorem to an appropriate
linear map, show that if there exists a K such that
KkukA ≥ kukB
for all u ∈ U , then there exists a K ′ such that
K ′ kukB ≥ kukA
for all u ∈ U . Thus comparable complete norms are equivalent.
[We could also use the inverse mapping theorem but this comes to much the
same thing.]

44
Exercise 18.12. (i) (Dini’s theorem) Let (X, d) be a compact metric space.
Suppose fn : X → R is a sequence of continuous functions such that, for
each fixed x ∈ X, fn (x) is a decreasing sequence with fn (x) → 0 as n → ∞.
By considering
Bn = {x : fn (x) < ǫ}
for any fixed ǫ > 0 show that fn → 0 uniformly on X.
(ii) Show, by means of an example, that the condition (X, d) compact
cannot be dropped. Show, by means of an example, that the condition fn de-
creasing cannot be dropped. Show, by means of an example, that the condition
fn continuous cannot be dropped.
(iii) Set p0 = 0 and pn+1 (x) = 12 x2 + pn (x) − 12 pn (x)2 . Explain why pn is
a polynomial. Show that

pn (x) ≤ pn+1 (x) ≤ |x|

and all n ≥ 0 for all x ∈ [0, 1]. Hence deduce that pn (x) → |x| as n → ∞
for all x ∈ [0, 1]. Now use Dini’s theorem to show that the convergence is
uniform.
Explain how to use this result as a replacement for Lemma 8.4 in the
proof of the Stone–Weierstrass theorem.

19 Second example sheet

Students who are unsure of their ground should check that they can do the
exercises in the main text. Strong students should at least glance at the
supplementary example sheet. The order of the exercises roughly follows the
order of the lectures.
Exercise 19.1. (i) Here is a typical use of the Stone–Weierstrass theorem.
If f ∈ C[0, 1], we say that f has nth moment
Z 1
En (f ) = f (t)tn dt.
0

Show that, if all the moments of f vanish, then

Z 1
f (t)P (t) dt = 0
0

for all polynomials. Use the Stone–Weierstrass theorem to deduce that

Z 1
f (t)g(t) dt = 0
0

45
for all g ∈ C[0, 1]. Deduce that f = 0.
(ii) (Optional) Let ω = exp(iπ/4). Show that
Z ∞
y n e−ωy dy = n!ω −n−1
0

and deduce that

Z ∞
y 4n+3 exp(−2−1/2 y) sin(2−1/2 y) dy = 0.
0

By making the substitution x = y 4 /4, show that

Z ∞
xn exp(−x1/4 ) sin(x1/4 ) dy = 0
0

for all n although x 7→ exp(−x1/4 ) sin(x1/4 ) is a well behaved non-zero con-

tinuous function. Why does the argument of part (i) fail?
[Both parts have obvious relevance to the question of what we can say about
a random variable X from knowledge of its moments.]

Exercise 19.2. (The Riemann–Lebesgue lemma) (i) The Riemann–Lebesgue

lemma tells us that, if f ∈ C(T), then fˆ(n) → 0 as |n| → ∞. There are
many ways of proving this but you are asked to prove it by finding a dense
subalgebra of C(T) for which the result is true ‘for obvious reasons’ and then
using a density argument to extend the result to all of C(T).
(ii) (Optional) Suppose that φ(n)P > 0 and φ(n) → 0. Show that we can
∞
find 0 < n1 < n2 < . . . such that j=1 φ(nj ) converges. By considering
P ∞
j=1 φ(n j ) cos n j x, or otherwise, show that there exists an f ∈ C(T) such
that φ(n)−1 fˆ(n) 9 0. Thus part (i) cannot be improved.

Exercise 19.3. In this question you may use results about Riemann integra-
tion in one dimension but not in higher dimensions.
(i) Suppose that f : [0, 1]2 → R is continuous. Explain why the function
g : [0, 1] → R defined by
Z 1
g(x) = f (x, y) dy
0

is continuous and so Z 1 Z 1 
f (x, y) dy dx
0 0
exists.

46
 (ii) Show that, if u, v : [0, 1] → R are continuous,
Z 1 Z 1  Z 1 Z 1 
u(x)v(y) dy dx = u(x)v(y) dx dy.
0 0 0 0

(iii) Use the theorem of Stone–Weierstrass to show that the collection of

functions h of the form
n
X
h(x, y) = uj (x)vj (y)
j=1

with uj , vj : [0, 1] → R continuous, is uniformly dense in C([0, 1]2 ).

(iv) Deduce that
Z 1 Z 1  Z 1 Z 1 
f (x, y) dy dx = f (x, y) dx dy
0 0 0 0

for all f ∈ C([0, 1]2 ).

Exercise 19.4. (i) Show that we cannot find a sequence Pn of polynomials

such that Pn (x) → exp(x) uniformly on [0, ∞) as n → ∞. Show that we can-
not find a sequence Qn of polynomials such that Qn (x) → exp(−x) uniformly
on [0, ∞) as n → ∞.
Why do these results not contradict the Stone–Weierstrass theorem?
Pn(ii) Let j
A0 be the subset of CR ([0, 1]) consisting of polynomials of the form
j=1 aj x . Show that A0 is not uniformly dense in CR ([0, 1]).Why does this
not contradict the Stone–Weierstrass theorem? Identify the uniform closure
of A0 and prove your statement.
(iii)
PnLet A12j be the subset of CR ([0, 1]) consisting of polynomials of the
form j=0 aj x . Show that A1 is uniformly dense in CR ([0, 1]).
(iii)
P Let A2 be the subset of CR ([−1, 1]) consisting of polynomials of the
form nj=0 aj x2j . Show that A2 is not uniformly dense in CR ([−1, 1]).Why
does this not contradict the Stone–Weierstrass theorem? Identify the uniform
closure of A2 and prove your statement.

Exercise 19.5. (i) Show that a topological space (X, τ ) is compact if and
only if it satisfies the following condition:–
If F is a non-empty collection of closed sets such that, whenever

F1 , F2 , . . . , Fn ∈ F,

we have nj=1 Fj =
T T
6 ∅, then F ∈F F 6= ∅.

47
 (ii) By quoting the appropriate theorem or (if you need the practice, re-
proving the result which is used in obtaining that theorem) show that if (X, τ )
is compact and Hausdorff, then, given V ∈ τ and x ∈ V , we can find a closed
set F and an open set U with

x ∈ U ⊆ F ⊆ V.

(iii) (Baire’s theorem for compact Hausdorff spaces.) If (X, τ ) is a com-

pact Hausdorff topological space and Uj is an open subset of X with the
property that Uj ∩ V 6= ∅ for all non-empty open V [j ≥ 1], show that
T ∞
j=1 Uj 6= ∅.

Exercise 19.6. (i) If A is a set of first category in a complete metric space

(X, d) without isolated points show that X \ A is uncountable.
(ii) If A is a set of first category in a complete metric space (X, d) without
isolated points, show, by considering the subspace

Cl{x : d(x, y) < δ},

or otherwise, that
{x : d(x, y) ≤ δ} \ A
is uncountable for all y ∈ X and δ > 0.
[Be careful. A subspace may have isolated points even if the original space
does not.]
Exercise 19.7. Consider the space CR∞ ([0, 1]) of infinitely differentiable real
functions on [0, 1] (with appropriate conventions about left and right deriva-
tives at end points).
(i) Show that, if we define
∞
X
d(f, g) = min(2−j , kf (j) − g (j) k∞ ),
j=0

then d is a complete metric on CR∞ ([0, 1]).

(ii) Suppose q is a point of [0, 1]. Show that

Em (q) = {f ∈ CR∞ ([0, 1]) : |f (j) (q)| ≤ m × j! × j j for all j ≥ 0}

is closed in the d metric and its complement

S∞ is dense.
(iii) Deduce that we can find an F ∈/ m=1 Em (q). Show that

|F (j) (q)||y|j
lim sup =∞
j→∞ j!

48
for all y 6= 0. Deduce that the Taylor expansion of F around q diverges except
at q (ie has zero radius of convergence).
(iv) If F is as in (iii), explain why F cannot have a power series expansion
valid in any open interval containing q.
(v) Up to now we have kept q fixed. Extend our argument to show the
existence of an infinitely differentiable function G : [0, 1] → R which does
have a power series expansion valid in any open interval.
Exercise 19.8. The object of this exercise is to prove the following form of
the Tietze extension theorem.
If (X, τ ) is a topological space and Y is a compact subset of X such that,
given any x, y ∈ Y with x 6= y, we can find a g ∈ CR (X) with g(x) 6= g(y)
then, given any f ∈ CR (Y ), we can find an F ∈ CR (X) such that F (y) = f (y)
for all y ∈ Y .
(i) Prove Tietze’s extension theorem in the form given in Theorem 7.11
using Urysohn’s lemma and the result just stated.
(ii) From now on, we assume that (X, τ ) and Y satisfy the hypotheses of
our theorem. Let us write A for the collection of f ∈ CR (Y ) such that we
can find an F ∈ CR (X) with F (y) = f (y) for all y ∈ Y . Use the Stone–
Weierstrass theorem to show that A is uniformly dense in CR (Y ).
(iii) Show that, if f ∈ A, we can find f˜ ∈ CR (X) with f˜(y) = f (y) for all
y ∈ Y and kf˜k∞ = kf k∞ . Use this result to show that A is uniformly closed
in CR (Y ). Deduce that A = CR (Y ) and our theorem holds.
Exercise 19.9. We work in C and write

∂D = {z ∈ C : |z| = 1}.

Suppose K is a compact subset of C and φ : K → ∂D is continuous. By

considering functions of the form Φ(z)/|Φ(z)|, or otherwise, show that there
is an open set Ω ⊇ K and a continuous function φ̃ : Ω → ∂D.
Explain, giving an explicit K and φ and using ideas from complex analysis
or topology (you are not asked for proofs), why we cannot always take Ω = C.
Exercise 19.10. We work with continuous functions on [0, 1].
(i) Show that, if fn → f , uniformly then the set F = {fn : n ≥ 1} is
equicontinuous.
(ii) Show that, if F = {fn : n ≥ 1} is equicontinuous and fn (q) → f (q)
as n → ∞ at each rational point q ∈ [0, 1], then fn → f uniformly.
Exercise 19.11. Recall the definition of total boundedness from the met-
ric and topological spaces course. Let (X, τ ) be a compact Hausdorff space.
Suppose that F is a subset of C(X).

49
 (i) Explain why F is complete under the natural metric d (that is the
restriction of the metric induced by k.k∞ ) if and only if F is closed in C(X).
Explain why this means that F is compact if and only if F is closed and
totally bounded.
(ii) Show that F is totally bounded if and only if it is uniformly bounded
and equicontinuous.
(iii) Deduce the Ascoli–Arzelà theorem (Theorem 9.3).
Exercise 19.12. Find all continuously differentiable functions x : R → R
with
x′ (t) = 3x(t)2/3 for all t
subject to the condition x(0) = 0 and show (formally or informally according
to taste) that you have found them all.
[Note that x 7→ 2x2/3 is differentiable (and so satisfies a Lipschitz condition)
for x 6= 0.]

20 Third example sheet

Students who are unsure of their ground should check that they can do the
exercises in the main text. Strong students should at least glance at the
supplementary example sheet. The order of the exercises roughly follows the
order of the lectures.
Exercise 20.1. We work with the set A(D) of functions analytic on
D = {z ∈ C : |z| < 1}.
(i) Let 1 > R > r. By considering Cauchy’s formula with the contour
z = Reiθ , or otherwise, show that
FR (M ) = {f ∈ A(D) : |f (z)| ≤ M for all |z| ≤ R}
is equicontinuous at every point z with |z| ≤ r.
(ii) If fn ∈ F and |fn (z)| ≤ M for all |z| ≤ R, show that we can find a
continuous function
F : {z ∈ C : |z| ≤ r} → C
and n(j) → ∞ such that fn(j) (z) → F (z) uniformly for |z| ≤ r. Explain why
F is analytic on Dr = {z : |z| < r}.
(iii) Show that the formula
∞
X
d(f, g) = 2−j min(1, sup |f (z) − g(z)|)
j=1 |z|≤1−2−j

50
gives a complete metric on A(D). Show that the set

F(M ) = {f ∈ A(D) : |f (z)| ≤ M for all |z| < 1}

is compact in this metric.

[Simple developments of these ideas produce the powerful method of normal
families in complex variable theory.]

Exercise 20.2. Consider the space of continuous functions C([0, 1]) with the
uniform norm. Suppose that E is a closed subspace consisting of continuously
differentiable functions.
(i) By applying the closed graph theorem, or otherwise, show that the
differentiation map D : E → C([0, 1]) (given by Df = f ′ ) is continuous.
(ii) Deduce, using the Ascoli–Arzelà theorem, or otherwise that the unit
ball in E is compact.
(iii) Deduce, quoting the appropriate theorem, that E is finite dimen-
sional.

Exercise 20.3. Suppose K is a compact Hausdorff space and C(K) is the

countable union of equicontinuous sets. Show that C(K) is the countable
union of bounded and closed (in the uniform norm) equicontinuous sets. De-
duce, using three of the ‘big theorems’ of the course, or otherwise, that C(K)
is finite dimensional and so K is finite.

Exercise 20.4. (i) Let U be a complete inner product space. If F is closed

convex subset of U and a ∈ U , show that we can find a unique f0 ∈ F such
that
ka − f0 k2 ≤ ka − f k2
for all f ∈ F .
[A set F is convex if, whenever x, y ∈ F and 1 ≥ λ ≥ 0, it follows that
λx + (1 − λ)y ∈ F .]
(ii) Give an example of a closed convex subset F of l∞ and a point a ∈ l∞
with the property that the equation

ka − gk∞ = inf ka − f k∞
f ∈F

has more than one solution with g ∈ F .

Exercise 20.5. We work in l2 . Let F be the collection of x ∈ l2 such that

all but finitely many xj are zero and let

a = (1, 1/2, 1/3, . . . ).

51
Show that F is a subspace of l2 , but, given f ∈ F , we can always find a g ∈ F
such that
ka − f k2 > ka − gk2 .
Why does this not contradict Theorem 12.1?

Exercise 20.6. (i) Use Theorem 12.1 and Lemma 12.2 to show that, if U is
complete separable inner product space and F is closed subspace of U , then
F , considered as a space in its own right (inheriting the inner product from
U ), is itself a complete separable inner product space.
(ii) Suppose that U is an inner product space and F is closed subspace
of U . Suppose further that F considered as a space in its own right has an
orthonormal basis e1 , e2 , . . . . Use the arguments
P∞which2 we used to establish
Bessel’s inequality to show that, if a ∈ U and j=1 |xj | < ∞, then
∞ ∞
X X
xj ej − a ≥ ha, ej iej − a



j=1 2 j=1 2

with equality only if xj = ha, ej i. Deduce the conclusions of Theorem 12.1

and Lemma 12.2 in this case.

Exercise 20.7. (i) Consider the inner product space c00 of sequences with
only finitely many non-zero terms and inner product
∞
X
ha, bi = aj b∗j .
j=1

Show that the formula

∞
X
Tu = j −1 uj
j=1

defines a continuous linear map T : c00 → C but there does not exist a w ∈ c00
such that
T u = hu, wi.
Why does this not contradict the Riesz representation theorem (Theorem 12.3)?
(ii) Find a closed subspace F of c00 and an a ∈ c00 such that, given f ∈ F ,
we can always find a g ∈ F such that

ka − f k2 > ka − gk2 .

Why does this not contradict Theorem 12.1?

52
Exercise 20.8. Suppose that U is an inner product space and e1 , e2 , . . . are
orthonormal. Show that
hf, en i → 0
as n → ∞.
Use this result to give another proof of the Riemann–Lebesgue lemma (see
Exercise 19.2).

Exercise 20.9. Suppose that e1 , e2 , . . . is an orthonormal basis for an inner

product space E. Show reasonably carefully that
∞
X
hx, yi = hx, ej ihy, ej i∗ .
j=1

Exercise 20.10. Consider a Hilbert space H with a closed subspace E and T :

E → C a continuous linear functional. Show that there exists a continuous
linear functional T̃ : H → C such that T̃ |E = T and kT̃ k = kT k.
[You can prove this using bases but well brought up people do not use bases
unless they have to. (However, if they have to, they instantly forget their
upbringing.)]

Exercise 20.11. Let E be an inner product space. We say that a non-zero

continuous linear map P : E → E is a projection if P 2 = P . We say that P
is an orthogonal projection if P (U ) ⊥ P −1 (0).
(i) Give a example of an orthogonal projection Q : l2 → l2 such that both
Q(U ) and Q−1 (0) are infinite dimensional.
(ii) Give a example of a projection R : C2 → C2 which is not an orthog-
onal projection.
(iii) If E is an inner product space and P : E → E is a projection, show
that P is an orthogonal projection if and only if kP k = 1.

Exercise 20.12. Suppose that H is a Hilbert space. If C is a subset (NB no

algebraic structure is implied) of H we write

C ⊥ = {g ∈ H : hg, f i = 0 for all f ∈ C}.

(i) Show that C ⊥ is a closed subspace of H.

(ii) Show that C ⊥⊥ = C if and only if C is a closed subspace of H.
(iii) Show that C ⊥⊥ = Cl span C.

53
21 Fourth example sheet
Students who are unsure of their ground should check that they can do the
exercises in the main text. Strong students should at least glance at the
supplementary example sheet. The order of the exercises roughly follows the
order of the lectures.
Exercise 21.1. Let e1 , e2 , . . . be an orthonormal basis for the Hilbert space
H. Write

un = e2n−1
vn = e2n−1 + n−1 e2n

and let U be the closed subspace generated by the un (ie the smallest closed
subspace containing the un ) and V be the closed subspace generated by the
vn .
(i) Show that U ∩ V = {0}.
(ii) Show that every x ∈ U + V has a unique expression as x = P x + Qx
with P x ∈ U and Qx = V . Show that P : U + V → U is linear and satisfies
P 2 = P (so P is a projection) but that P is not continuous.
(iii) Show that U + V is dense in H but U + V 6= H.
Exercise 21.2. Let H be Hilbert space. Suppose that T : H → H is linear
and self-adjoint (that is to say hT x, yi = hx, T yi for all x, y ∈ H). Use the
the principle of uniform boundedness to show that T is continuous.
Exercise 21.3. Let U be a normed space and T : U → U a continuous linear
map. Explain why
ρ(T ) = lim inf kT n k1/n
n→∞

is well defined.
By making the observation that

kT N k+r k ≤ kT N kk kT kr ,

or otherwise, show that, in fact

kT n k1/n → ρ(T )

as n → ∞. (We call ρ(T ) the spectral radius of T .)

Now suppose that U is a Banach space. Show that λI − T is invertible
(that is to say, has continuous inverse) for all |λ| > ρ(T ).
By considering maps of the form

(x1 , x2 , x3 , . . . ) 7→ (0, c1 x1 , c2 x2 , . . . )

54
with appropriate cj , show that we can find T : l2 → l2 such that T n is injective
(so, in particular, T n 6= 0) for all n ≥ 1 but ρ(T ) = 0.

Exercise 21.4. Let U and V be Banach spaces. Suppose T : U → V is

a continuous injective map such that T (U ) is dense in V . Show that the
following three statements are equivalent.
(i) There exists a c > 0 with the property that

kT ukV ≥ ckukU

for all u ∈ U .
(ii) T is surjective.
(iii) T is invertible (that is to say, has continuous inverse).

Exercise 21.5. (i) Consider the the shift map S : l2 → l2 given by

S(a1 , a2 , . . . ) = (0, a1 , a2 , . . . ).

Show that S is a well defined continuous linear map and find its adjoint.
(ii) Let U be a finite dimensional space and T : U → U a linear map.
Explain why λI − T fails to be invertible if and only if there exists a u 6= 0
such that T u = λu.
(iii) Give an example of a continuous linear map R1 : l2 → l2 which
is surjective but not bijective. Give an example of a continuous linear map
R2 : l2 → l2 which is injective but not bijective.
(iv) By using Exercise 21.4, or otherwise, show that, if U is a Banach
space and T : U → U is a continuous map, then λ ∈ σ(T ) if and only if at
least one of the following conditions hold.
(A) λ is an eigenvalue of T , that is to say, there exists a u 6= 0 such that
T u = λu.
(B) λ is not an eigenvalue of T but λ is an approximate eigenvalue of T ,
that is to say, there exist uj ∈ U with kuj k = 1 such that kT uj − λuj k → 0.
(C) (λI − T )(U ) is not dense in V .
(v) Give an example of a continuous linear map T : l2 → l2 such that
σ(T ) = {0} but 0 is not an eigenvalue of T .
[The moral of this question is that the spectrum is much more a complicated
than at first appears. We find an interesting spectrum in Exercise 22.16.]

Exercise 21.6. (i) Let U be a Banach space and T : U → U a continuous

linear map. Suppose that λ is in the frontier of σ(T ) (that is to say λ lies in
σ(T ) and in the closure of the complement of σ(T )). By examining possibility
(C) in Exercise 21.5, show that λ is an approximate eigenvalue (this includes

55
the possibility that λ is an eigenvalue). In other words, the frontier of the
spectrum is composed of approximate eigenvalues.
(ii) Now suppose T : H → H is a continuous Hermitian (that is to say,
self-adjoint) linear map. Show that the approximate eigenvalues of T are
real. Deduce from (i) that the spectrum of T consists of real numbers.
Exercise 21.7. (This quite long, though instructive. Those who want a
shorter, thought still non-trivial, question should ignore parts (i) and (iii),
and do that part of (ii) which says that the limit of finite rank operators is
compact.)
Consider Hilbert space H. Let us write B(H) for the space of continuous
linear maps T : H → H.
(i) Show that the following statements about a bounded sequence xn are
equivalent.
(A) Every subsequence of the sequence has a convergent subsequence.
(B) Given any ǫ > 0 we can find a finite dimensional subspace E such
that, if P is the orthogonal projection onto E, we have k(I − P )xn k2 < ǫ.
(ii) We say that an S ∈ B(H) is of finite rank if the image space S(H)
is finite dimensional. Show that T ∈ B(H) is compact if and only if we can
find finite rank Sn such that kSn − T k → 0 as n → ∞.
[It may be helpful to recall that H is separable]
(iii) Show that the collection of compact operators is a closed nowhere
dense (ie having dense complement) subset of B(H).
Exercise 21.8. (i) We work on the Hilbert space l2 . Show that
S(a1 , a2 , a3 , . . . ) = (a2 , a3 , a4 . . . )
is a continuous linear map from l2 to l2 and find its adjoint map S ∗ . Find
kSk and kS ∗ k.
(ii) We use the notation of Exercise 21.7. Show that the collection of
self-adjoint operators is a closed nowhere dense subset of B(H).
Exercise 21.9. Check that the following statement is equivalent to the spec-
tral theorem for compact self-adjoint operators (Theorem 14.5). Let H be a
Hilbert space. If T : H → H is a continuous linear compact self-adjoint map
we can find a finite or infinite sequence P1 , P2 , . . . of orthogonal projections
(that is to say continuous linear maps Pj : H → H with Pj−1 ({0}) ⊥ Pj (H))
which are mutually orthogonal (that is to say Pj (H) ⊥ Pk (H) for j 6= k) and
have finite dimensional non-trivial images (that is to say 1 ≤ dim Pj (H) <
∞) together with distinct non-zero real λj , with λn → 0 as n → ∞ if the
sequence is infinite, such that
X
T = λj Pj .
j

56
[Although this is just a simple rewrite of Theorem 14.5 it provides a better
jumping off point for generalisation.]
Exercise 21.10. Suppose T : H → H is a compact self-adjoint operator on
a Hilbert space H. Let λ be a non-zero real number and let y ∈ H. Consider
the two equations
T x = λx, (1)
T x = λx + y. (2)
Prove the Fredholm alternative which states that either (1) has no non-zero
solutions and (2) has a unique solution or the space of solutions Eλ of (1)
contains non-zero vectors and (2) has solutions if and only if y ⊥ Eλ . In the
second case, if z is a solution of (2), the set of solutions is
{z + w : w ∈ Eλ }.
[Hint: If this was a 1B mathematical methods question you would consider it
rather easy.]
Exercise 21.11. Work through Section 15.
Exercise 21.12. In this exercise we work in a Hilbert space H.
(i) Suppose that T and S are commuting compact self-adjoint continuous
maps from H to H. If we write
Eλ = {x ∈ H : T x = λx}
show that S(Eλ ) ⊆ Eλ . By repeated use of Theorem 14.5, or otherwise, show
that we can find an orthonormal basis f1 , f2 , f3 , . . . such that each fj is an
eigenvector of both T and S.
(ii) A continuous linear map R : H → H is called normal if RR∗ = R∗ R.
By considering T = (R + R∗ )/2 and S = (R − R∗ )/(2i), or otherwise show
that, if R a compact normal continuous linear map from H to H, we can
find an orthonormal basis fj of H and λj ∈ C with λj → 0 such that
∞
X
Ru = λj hu, fj i
j=1

for all u ∈ U .
(iii) Show conversely that if fj is any orthonormal basis of of H and λj
any sequence in C with λj → 0 then the formula
∞
X
Ru = λj hu, fj i
j=1

for all u ∈ U defines a continuous linear map R : H → H which is compact

and normal.

57
22 Supplementary exercises
These exercises are not intended to be harder but are less closely linked to
the immediate needs of the course. They do, however, provide background
and anyone intending to do more analysis should at least glance at them.
Exercise 22.1. (Revision of 1A) We say that a function f : [a, b] → R is
strictly concave if, whenever a ≤ t < s ≤ b and 0 < λ < 1, it follows that
λf (t) + (1 − λ)f (s) < f λt + (1 − λ)s).
(i) Show that, if g : [a, b] → R is twice differentiable with g ′′ (t) < 0 for
all t ∈ (a, b), then g is strictly concave.
(ii) Give an example of a strictly concave function g : [−1, 1] → R which
is not differentiable at 0.
(iii) Suppose that f : [a, b] → R is strictly concave. By using induction,
or otherwise, prove Jensen’s inequality which states that if x1 , x2 , . . . , xn are
distinct
P points of [a, b] and λ1 , λ2 , . . . , λn are strictly positive real numbers
with nj=1 λj = 1, then
n
! n
X X
f λj xj > λj f (xj ).
j=1 j=1

Deduce that if x1 , x2 , . . . , xn are

Pnpoints of [a, b] and λ1 , λ2 , . . . , λn are
strictly positive real numbers with j=1 λj = 1, then
n
! n
X X
f λj xj = λj f (xj )
j=1 j=1

if and only if x1 = x2 = · · · = xn .
(iv) Use Jensen’s inequality to show that, if aj > 0, then
a1 + a2 + · · · + a n
(a1 a2 . . . an )1/n ≤ .
n
(This is Cauchy’s arithmetic-geometric inequality.) What are the conditions
for equality?
(v) Suppose that p > 1 and let g(x) = (1 + x1/p )p . Show that g is a
concave function.
Suppose that a1 , a2 , . . . , an > 0, nj=1 apj = 1 and b1 , b2 , . . . , bn > 0. By
P
applying Jensen’s inequality with xk = bpk /apk and λk chosen appropriately,
prove Minkowski’s inequality.
n
!1/p n
!1/p n
!1/p
X X X
(aj + bj )p ≤ apj + bpj
j=1 j=1 j=1

58
and obtainPthe conditions for equality. Why does the result fllow for general
values of nj=1 apj ?

Exercise 22.2. Obtain Minkowski’s inequality by applying Hölder’s inequal-

ity to the observation
n
X n
X n
X
p p−1
|xj + yj | ≤ |xj ||xj + yj | + |yj ||xj + yj |p−1 .
j=1 j=1 j=1

Is this really a different proof to the one given in the lectures using the reverse
Hölder inequality?

Exercise 22.3. The results of Exercise 2.2 depend on clever inequalities13

but there are other ways of arriving at the results.
≥ s > r ≥ 1. Investigate maxima and minima of nj=1 xsj subject
P
Let ∞ P
to xj ≥ 0, nj=1 xrj = 1 using the calculus of variations. (Unless we take care,
which you are not asked to do, the results will not be rigorous but, once we
know what is happening, it is much easier to prove that it happens by some
other technique.)

Exercise 22.4. If V is a vector space over F, we say that E is an algebraic

basis (that is to say a basis in the sense of 1B algebra) if every v ∈ V can be
written uniquely as a finite sum
n
X
v= λj ej
j=1

with e1 , e2 , . . . , en distinct elements of E and λj ∈ F. The collection V ∗ of

linear maps α : V → F is called the algebraic dual (that is to say, the dual
space in the sense of 1B algebra). The proofs of 1B algebra show that U ∗ can
be given the structure of a vector space.
Let c00 be the vector space of complex sequences with only a finite number
of non-zero terms. Explain why c00 has a countable basis. Identify c∗00 in a
natural manner with the space CN of all complex sequences. Show that c∗00
does not have a countable basis. (The argument is not difficult but you should
not sleep walk through it.)
Although this question deals only with one space the reader should require
little convincing that, if we only deal with algebraic duals, the algebraic dual
of an infinite dimensional space will be very much bigger than the the space
(and the dual of the dual will be even bigger).
13
To the writer all inequalities seem clever.

59
Exercise 22.5. (i) Prove the parallelogram law

ka + bk22 + ka − bk22 = 2(kak2 + bk22 )

for all a, b ∈ l2 .
(ii) Use induction to show that, for each n, we can find ζjk (n) = ±1 such
that 2
2n X
2n 2n
X X
ζjk (n)a(k) = 2n ka(k)k22



j=1 k=1 2 k=1

for all a(k) ∈ l2 .

(iii) If (U, k.k) is isomorphic to (l2 , k.k2 ) explain why there is a constant
K independent of n such that
2n X
2n 2 2n 2n X
2n 2
X X X
K ζjk (n)u(k) ≥ 2n ku(k)k2U ≥ K −1 ζjk (n)u(k)



j=1 k=1 U j=1 j=1 k=1 U

for all u(k) ∈ U .

(iv) Show that l2 is not isomorphic to lp when p 6= 2.

Exercise 22.6. Consider the space of n × n complex matrices with the oper-
ator norm. Prove the Cayley–Hamilton theorem by using the fact proved in
Exercise 6.2 that the set of of matrices with n distinct eigenvalues is dense.
Exercise 22.7. Suppose that gn : T → R is continuous and satisfies the
conditions set out in Exercise 6.11 as follows.
(i) There exists a constant K such that

1
Z
|gn (t)| dt ≤ K
2π T
for all n ≥ 1.
(ii) If δ > 0 and f is a continuous function with f (t) = 0 for |t| < δ, then

1
Z
f (t)gn (t) dt → 0
2π T
as n → ∞.
(iii) We have
1
Z
gn (t) dt → 1
2π T
as n → ∞.

60
 If f : T → R is continuous and δ > 0 observe that
1 1
Z Z
gn (t)f (t) dt − f (0) = gn (t)(f (t) − f (0)) dt
2π T 2π |t|<δ
 Z 
1 1
Z
+ gn (t)(f (t) − f (0)) dt + gn (t) dt − 1 × f (0),
2π |t|≥δ 2π T

and, by estimating the three terms separately, show that

1
Z
gn (t)f (t) dt → f (0)
2π T
as n → ∞.
Show that condition (ii) is implied by
(ii)′ If δ > 0 then gn (t) → 0 uniformly for t ∈
/ (−δ, δ).
Show also, by considering the Riemann–Lebesgue lemma or otherwise, that
condition (ii) does not imply (ii)′ .

Exercise 22.8. Let f : T → C be continuous. We write

1
Z
ˆ
f (n) = f (t) exp(−int) dt.
2π T

(i) Show that N |n|

P
n=−N r exp(int) converges uniformly, to Pr (t) say, for
t ∈ T as N → ∞ for each fixed r with 0 < r < 1. Deduce that
∞
1 1
Z Z
r fˆ(n) =
X
|n|
f (t)Pr (−t) dt = f (t)Pr (t) dt.
n=−∞
2π T 2π T

Show that
1 − r2
Pr (t) = .
1 − 2r cos t + r2
(ii) Show that
(A) Pr (t)
Z ≥ 0 for all t.
1
(B) Pr (t) dt = 1.
2π T
(C) Pr (t) → 0 uniformly as r → 1− for t ∈/ (−δ, δ).
(iii) Deduce, by the arguments of Exercise 22.7, or otherwise, that
∞
rn fˆ(n) → f (0)
X

n=−∞

61
as r → 1−. By considering the Fourier coefficients of fa given by fa (t) =
f (t − a), or otherwise, show that
∞
rn fˆ(n) exp(int) → f (t)
X

n=−∞

as r → 1−.

Exercise 22.9. Let f : T → C be continuous. We write

1
Z
ˆ
f (n) = f (t) exp(−int) dt.
2π T

(i) We write Dn (t) = nj=−n exp(ijt). Show that, if t 6= 0,

P

sin (n + 21 )t


Dn (t) = .
sin 2t

What is the value of Dn (0)?

(ii) Show that
n
1 1
Z Z
fˆ(j) =
X
f (t)Dn (−t) dt = f (t)Dn (t) dt.
j=−n
2π T 2π T

(iii) Show that

A(n + 12 )
| sin (n + 12 )t |


|Dn (t)| ≥
r+1
for all rπ/(n + 21 ) ≤ |t| ≤ (r + 1)π/(n + 12 ) and 1 ≤ r ≤ n − 1 for some A > 0
(to be chosen to fit your convenience) independent of n.
(iv) Deduce that
1
Z
|Dn (t)| dt ≥ B log n
2π T
for all n ≥ N and some B > 0 (where both N and B are chosen to fit your
convenience).
(v) Use Exercise 6.11 to show that there exists a continuous function f
whose Fourier sum does not converge to f at 0.

Exercise 22.10. Let (X, τ ) be a topological space. Write x ∼ y if f (x) =

f (y) whenever f ∈ C(X) the space of real valued continuous functions on X.
(i) Show that ∼ is an equivalence relation.

62
 (ii) Show that, if we give X/ ∼ the quotient topology τ / ∼, then the
function J : C(X) → C(X/ ∼) given by J(f )[x] = f (x) (where [x] denotes
the equivalence class of x) is a well defined linear isometry.
(iii) Show that (X/ ∼, τ / ∼) is Hausdorff.
(iv) If (X, τ ) is compact, show that (X/ ∼, τ / ∼) is also compact.
Exercise 22.11. In this exercise we discuss various separation axioms for
topological spaces (X, τ ).
A topological space (X, τ ) is said to be T0 if, given a, b ∈ X with a 6= b,
at least one of these statements is true. (α) There exists a U ∈ τ such that
a ∈ U and b ∈ / U . (β) There exists a V ∈ τ such that a ∈ / V and b ∈ V . A
topological space (X, τ ) is called T1 if given a, b ∈ X with a 6= b, at both of
these statement are true.
(i) Show that (X, τ ) is T1 if and only if every singleton set {x} is closed.
(ii) Consider N. If the elements of σ0 are ∅ and sets of the form {m ∈
N : m ≥ n} show that σ1 is a topology and (N, σ1 ) is T0 but not T1 .
(iii) Consider N. If the elements of U ∈ σ1 if and only if A = ∅ or N \ A
is finite show that σ1 is a topology and (N, σ1 ) is T1 but not Hausdorff. (In
this classification, Hausdorff spaces are called T2 .)
(iv) Show that every T1 normal space is Hausdorff but give an example
of a normal space which is not Hausdorff. (Recall that a topological space is
said to be normal if, given A and B non-empty disjoint closed sets, we can
find disjoint open sets U and V such that A ⊆ U and B ⊆ V .)
[There are very small and simple examples.]
(v) Carry out us much of the proof of Urysohn’s lemma and the Tietze
extension theorem as you need14 to show that Urysohn’s lemma holds with
the hypothesis ‘compact and Hausdorff ’ replaced by ‘normal’ and the Tietze
extension theorem holds in the following form:-
If Y is closed subset of a normal topological space (X, τ ), then, given any
bounded continuous real valued function f on Y (where Y has the subspace
topology), we can find a bounded continuous real valued function F on X
such that F (y) = f (y) for all y ∈ Y .
Exercise 22.12. We work with a real vector space to simplify the algebra
but similar results can be obtained in the complex case.
(i) Consider a real inner product space. Obtain the formulae

4hu, vi = ku + vk22 − ku − vk22

and
ku + vk22 + ku − vk22 = 2 kuk22 + kvk22 .

14
Either do the whole thing as a revision exercise or just cast your eyes over your notes.

63
 (ii) Suppose that V is a real normed vector space such that

ku + vk2 + ku − vk2 = 2 kuk2 + kvk2 .

Show that if we set

p(u, v) = (ku + vk2 − ku − vk2 )/4

then
p(u + w, v) = p(u, v) + p(w, v).
[It may or may not be helpful to consider the parallelepiped associated with
u, v and w.]
(iii) Continuing with the hypotheses and notation of (ii), show that

p(nu, v) = np(u, v)

for n a strictly positive integer. Deduce that p(λu, v) = λp(u, v) when λ is

rational. Show that p is an inner product and k.k is the norm associated with
p.
Exercise 22.13. In some of the exercises we used metrics rather than norms.
This exercise shows why the
Pneed may arise. We work in C ∞ ([0, 1]).
∞
(i) Show that d(f, g) = j=0 2 min(1, kf (j) − g (j) k∞ ) is a metric. Show
−j

(j)
that fn → f (j) uniformly on [0, 1] for each j if and only if d(fn , f ) → 0.
Show that d is a complete metric.
(ii) The object of the rest of this question is to show that there is no norm
with this property. To this end, suppose that k.k is a norm on C ∞ ([0, 1]).
(j)
such that kfn − f k → 0 implies fn → f (j) uniformly on [0, 1] for each j.
Show, by reductio ad absurdum, or otherwise, that there must exist ǫj > 0
with the property that kf (j) k∞ ≥ 1 implies kf k ≥ ǫj .
(iii) Show that we can find a sequence of fn ∈ C ∞ ([0, 1]) such that

kfn(j) k∞ ≤ 2−n for 0 ≤ j ≤ n − 1

kfn(n) k∞ ≥ 2n ǫ−1
n .

(j)
Show that fn → f (j) uniformly on [0, 1] for each j but kfn k → ∞.
Exercise 22.14. Suppose that we have a sequence of metrics dj on a space
X.
(i) Show that
∞
X
2−j min 1, dj (x, y)


d(x, y) =
j=1

64
is a metric on X such that the identity map ι : (X, d) → (X, dj ) is continuous
for each j.
(ii) Show that if ρ is a metric on X such that the identity map ι : (X, ρ) →
(X, dj ) is continuous for each j then ι : (X, ρ) → (X, d) is continuous.
Exercise 22.15. (This easy exercise discusses anti-isomorphism15 .) If (U, +, ., C)
is a vector space show that, if we write

λ • u = λ∗ u,

then (U, +, •, C) is a vector space. Show that (U, +, ., C) and (U, +, •, C) are
algebraically anti-isomorphic.
Exercise 22.16. (i) (This is about difference equations in the sense of the
1A differential equations course.) If λ ∈ C, λ 6= 0, find the general solution
of the difference equation

λun − un+1 = 0 [n ∈ Z].

(ii) Find the general solution of the the difference equation.

λun − un+1 = 0 [n ∈ Z, n 6= 0].

(Note that there are two arbitrary constants.)

(iii) Find the general solution of the the difference equation.
(
0 if n 6= 0
λun − un+1 =
1 if n = 0.

(iv) From now on we work in l2 (Z) the vector space of two sided square
summable sequences of complex numbers

a = (. . . , a−2 , a−1 , a0 , a1 , a2 , . . . )

with norm !1/2

∞
X
kak2 = |aj |2 .
j=−∞

Show that, if 0 < |λ| < 1, there is exactly one u ∈ l2 (Z) such that
(
0 if n 6= 0
λun − un+1 =
1 if n = 0,
15
And shows that there is not much to discuss

65
and that the same is true if |λ| > 1.
(iv) Show that the shift operator S defined by Sa = b with bj = aj+1 is a
well defined continuous linear map from l2 (Z) to itself. Show that, if |λ| =
6 1,
then λ ∈/ σ(S).
(v) Show that 1 is not an eigenvalue of S. By considering v with
(
N −1/2 if 1 ≤ n ≤ N
vn =
0 otherwise,
show that 1 is an approximate eigenvalue of S.
(vi) Show that
σ(S) = {λ ∈ C : |λ| = 1},
that all the points of σ(S) are approximate eigenvalues and that none of them
are eigenvalues.
Exercise 22.17. The object of this exercise is to exhibit a complete inner
product space which is not separable.
(i) Let X be a non-empty set and a : X → R a function such that a(x) ≥ 0
for all x ∈ X and such that there exists a K with
X
a(x) ≤ K
x∈F

whenever F is a finite subset of X. Show that

Y = {x ∈ X : a(x) 6= 0}
is countable.
Using the fact that the sum of a sequence of positive numbers is unaffected
by the order of summation, this means that we can define
X X
a(x) = a(x).
x∈X x∈Y

Suppose b : X →P R a function such thatP

b(x) ≥ 0 for all x ∈ X. In what
follows we say that x∈X b(x) converges if x∈F b(x) ≤ K whenever F is a
finite subset of X and we write
X X
b(x) = b(x).
x∈X b(x)6=0

2
P (ii) If X2 is a non-empty set and 2a : X → C, we say that a ∈ l (X) if
x∈X |a(x)| converges. Show that l (X) is a vector space. Show further
that X
ha, bi = a(x)b(x)∗
x∈X

66
gives a well defined inner product on l2 (X).
(iii) Show that l2 (X) is separable if and only if X is countable.
(iv) Show that l2 (X) is complete.

67