Supplementary material for Chapter 5

Here are the topics supplementary to Chapter 5. You may wish to skip the starred sections at
a first reading.

Normed vector spaces (NVS), and metrics in Chapter 5 arising from norms (1, 3) page 1
Inner product spaces and proof of Cauchy’s inequality (1, 3) 5
Reason for choice of subscripts d1 , d2 , d∞ (1) 9
Brief mention of lp (3) 12
Sequence spaces: examples including l1 , l2 , l∞ (1, 3) 12
General open sets in R (2) 16

Normed vector spaces (1, 3) This section assumes the reader knows what a vector space
is, and just discusses the normed aspect. Normed vector spaces (NVS) are intermediate in
generality between Euclidean spaces and metric spaces: as we shall see, every NVS gives rise to
a metric space, but not every metric space arises from a NVS. The norm of a vector is a kind of
n
measure of its length: for example
in the Euclidean space R the Euclidean norm of any vector
 n

x = (x1 , x2 , . . . , xn ) is ||x|| =  x2i . This norm has properties which can be abstracted to
i=1
give the following definition.

Definition S.5.1 A normed vector space (NVS) consists of a vector space V over R or C,
together with a function V → R written v → ||v|| and called a norm on V , such that
(N1) for all v ∈ V, ||v||
0; and ||v|| = 0 iff v = 0;
(N2) for any v ∈ V and any scalar α, ||αv|| = |α| ||v||;
(N3) (subadditivity) for all u, v ∈ V, ||u + v||  ||u|| + ||v||.

We immediately relate this to metric spaces.

Proposition S.5.2 Any NVS (V, || ||) gives rise to a metric space (X, d) where, as sets, X = V
and for x, y ∈ V, d(x, y) = ||x − y||.

Proof The following proof is fairly easy.

(M1) For any x, y ∈ X , we know d(x, y)
0 since ||x−y||
0; also, d(x, y) = 0 iff ||x−y|| = 0
iff x − y = 0 by (N1), i.e. iff x = y .

1
(M2) Symmetry of d follows immediately from (N2):
for any x, y ∈ X, d(y, x) = ||y−x|| = ||−(x−y)|| = |−1| ||x−y|| = ||x−y|| = d(x, y).
(M3) For any x, y, z ∈ X, using (N3) we get

d(x, z) = ||x − z|| = ||x − y + y − z||  ||x − y|| + ||y − z|| = d(x, y) + d(y, z). 

We now explore which metrics in Chapter 5 arise this way, and at the same time introduce
examples of NVS. First, here are some norms on Rn . The first is the Euclidean norm which
we havealready seen, sometimes called the l2 norm: if x = (x1 , x2 , . . . , xn ) ∈ Rn , we let

||x||2 = x2i . It is easy to see that this satisfies (N1) and (N2). For (N3) we need Cauchy’s
inequality: for if also y = (y1 , y2 , . . . , yn ) then to prove ||x + y||  ||x|| + ||y|| it is equivalent
to prove the same inequality after squaring both sides, which amounts to proving
  
xi yi  x2i yi2.

This again is equivalent to the same inequality squared, which is Cauchy’s inequality
 2    
xi yi  x2i yi2 .

We prove this in the next section. The metric arising from the l2 norm is the Euclidean metric
d2 of Example 5.4 in the book.

Next we consider the l1 and l∞ norms on Rn . These give rise to the metrics d1 and d∞ of
Example 5.7 in the book. For x ∈ Rn as before, define

n
||x||1 = |xi |, ||x||∞ = max{|x1 |, |x2 |, . . . , |xn |}.
i=1

We now check that these are norms for Rn . It is straightforward to check that (N1) and (N2)
hold for each of them. Let us check (N3) for || ||1 . Given x, y ∈ Rn as before,

||x + y||1 = |x1 + y1 | + |x2 + y2 | + . . . + |xn + yn |  |x1 | + |y1 | + . . . + |xn | + |yn | = ||x||1 + ||y||1.

This completes the proof that || ||1 is a norm for Rn .

Now let us check (N3) for || ||∞ . With x, y ∈ Rn as before, for each i = 1, 2, . . . , n we have
|xi + yi |  |xi | + |yi |  ||x||∞ + ||y||∞, so

||x + y||∞ = max{|x1 + y1 |, |x2 + y2 |, . . . , |xn + yn |}  ||x||∞ + ||y||∞

as required.

2
In the section on sequence spaces we shall look at infinite-dimensional analogues of these norms
on Rn . In the meantime, we consider Examples 5.13, 5.14 and 5.16 in the book. In the first of
these, X is the set of all bounded real-valued functions f : [a, b] → R, and we consider it as a
metric space with the sup metric, so for any f, g ∈ X ,

d(f, g) = sup |f (x) − g(x)|.

x∈[a, b]

This metric arises from a norm on X : first, we can give X the structure of a real vector space
by defining addition and scalar multiplication as follows: for any f, g ∈ X and α ∈ R,

(f + g)(x) = f (x) + g(x); (αf )(x) = αf (x) for all x ∈ [a, b].

The zero vector here is the zero function on [a, b]. This is logically slightly sophisticated,
since ‘vectors’ in X are now whole functions. But the check that all the vector space axioms
hold follows quickly from the analogous properties for R. For example to see that addition (of
functions) is associative, we need to check, for any three functions f, g, h : [a, b] → R, that
(f + g) + h = f + (g + h) as functions. This means (f (x) + g(x)) + h(x) = f (x) + (g(x) + h(x))
for all x ∈ [a, b], which follows immediately from associativity of R. The other vector space
axioms are equally straightforward to check, so we omit them and press on to define the ‘sup
norm’ on X : for any f ∈ X let ||f ||∞ = sup |f (x)|. This exists since f is bounded. Again
x∈[a, b]
(N1) and (N2) are straightforward to check, and we concentrate on (N3). For any f, g ∈ X and
any x ∈ [a, b] we have |f (x) + g(x)|  |f (x)| + |g(x)|  ||f ||∞ + ||g||∞, so

||f + g||∞ = sup |f (x) + g(x)|  ||f ||∞ + ||g||∞,

x∈[a, b]

as required. It is immediate that the metric this norm gives rise to is the sup metric on X .

Now we consider the space X of all continuous functions f : [a, b] → R as in Example 5.14 in
the book. This again has a vector space structure with the definitions of addition and scalar
multiplication as in the above example (in fact, since all continuous real-valued functions on
[a, b] are bounded, we have a vector subspace of the previous vector space). Let us define ‘the
 b
1
L norm’ on X by ||f ||1 = |f (t)|d t. In order to check (N1) in this case, we need Lemma 5.15
a
from the book, which says that a non-negative continuous function whose integral over [a, b] is
zero must be the zero function. This shows that ||f ||1 = 0 implies f = 0. The rest of (N1)
is straightforward to check, as is (N2). For (N3), suppose that f, g are continuous real-valued
functions on [a, b]. Then |f (t) + g(t)|  |f (t)| + |g(t)| for each t ∈ [a, b], and integrating over
[a, b] we get ||f + g||1  ||f ||1 + ||g||1 as required. Again it is clear that the L1 norm gives rise
to the L1 metric of Example 5.14.
3
Recall that in the final example of this trio, Example 5.16, the space X is still the space of
continuous functions f : [a, b] → R, but it now has the L2 (instead of the L1 ) metric. Again
we can see that this metric arises from a norm: for any f ∈ X , define
 b
12
||f ||2 = (f (t))2 .
a

The checks that (N1) and (N2) hold are similar to those for the L1 norm. To see that (N3)
holds we need the Cauchy-Schwarz-Bunyakovsky inequality,
 b  b  b
2
2 2
(f (t)) d t (g(t)) d t
f (t)g(t)d t .
a a a

(This will be proved in the next section.) Given this inequality, the check that (N3) holds here
is entirely analogous to the check that (N3) holds in the case of the norm || ||2 on Rn . Again,
it is clear that the L2 norm gives rise to the L2 metric of Example 5.16.

To complete this section, we show that not every metric arises from a norm. In particular the
discrete metric on a set with more than one point in it does not arise from a norm. For suppose
that X is a set and d is the discrete metric on X . In order for d to arise from a norm, we would
require that X can be given the structure of a real or complex vector space. This already rules
out any finite set with more than one point in it. (A real or complex vector space is infinite
unless it has dimension zero, in which case it contains only one point.) If X can be given the
structure of a real or complex vector space, there is still no norm on X which gives rise to the
discrete metric. For suppose there were such a norm || ||, and let x = 0 in X . Then d(x, 0) = 1,
so we must have ||x|| = d(x, 0) = 1. But then the norm of x/2 would necessarily, by (N2), be
1/2, so ||x/2 − 0|| = 1/2, whereas in the discrete metric d(x/2, 0) = 1.

Another metric which cannot arise from a norm is the p-adic metric of Example 5.11. For
Z does not admit the structure of a real or complex vector space on grounds of cardinality -
it is a countable set, whereas, as we have already noted, any real or complex vector space is
uncountable, unless it it 0-dimensional in which case it contains only one point. Similarly the
word metric on a finitely generated group G in Example 5.12 does not arise from a norm, since
again such a group is countable.

We shall return to NVS in later supplementary material - see the next four sections of S.5, the
last section of S.6 and the first section of S.17.

In the meantime we note that there is a definition of equivalence of norms analogous to what we
call ‘Lipschitz equivalence’ of metrics in Chapter 6. Again, the terminology is not universal.

4
Definition S.5.3 Norms || || and || || on a vector space V are called Lipschitz equivalent if
there exist positive real numbers h, k such that

h||v||  ||v||  k||v|| for all v ∈ V.

It is easy to check that Lipschitz equivalent norms give rise to Lipschitz equivalent metrics.

Inner product spaces and Cauchy’s inequality (1, 3) We have explored one ‘boundary’
of the NVS concept by showing that not all metrics arise from norms. We now explore a stronger
condition than existence of a norm on a real or complex vector space, namely the existence of
an inner product structure. The Euclidean norm on Rn arises from an inner product on Rn : for

any x ∈ Rn we have ||x||2 = x, x. We shall see that not all norms arise from inner products.

Recall that an inner product on a real vector space is a generalization of the ‘scalar’ or ‘dot’
product on R3 : x, y = x1 y1 + x2 y2 + x3 y3 for vectors x = (x1 , x2 , x3 ) and y = (y1 , y2 , y3 ).

We abstract the properties of this dot product to make the next definition. In general, an inner
product on a real vector space V is a function V × V → R written (u, v) → u, v satisfying:

(RI1) (Positivity) u, u
0 for any u ∈ V , and u, u = 0 iff u = 0.

(RI2) (Symmetry) u, v = v, u for all u, v ∈ V .
(RI3) (Bilinearity) For all u1 , u2 , v1 , v2 ∈ V and real numbers α, β both the following hold:
αu1 + βu2, v1  = αu1, v1  + βu2, v1  and u1 , αv1 + βv2  = αu1 , v1  + βu1, v2 

The second part of (RI3) could be omitted, since it follows from the first part of (RI3) and
symmetry.

 To some extent it is just introducing another complication to look at the complex case, and
at a first reading it is probably wise to concentrate on the real case, but in analysis one is often
interested in the complex case (for the usual reasons of algebraic completeness) so we include a
mention of the complex case here.

We may define complex inner product spaces by abstracting the properties of the function
C3 × C3 → C given by ((z1 , z2 , z3 ), (w1 , w2 , w3 )) → z1 w1 + z2 w2 + z3 w 3 .

A complex inner product space is a complex vector space V and a function V ×V → C satisfying:
(CI1) u, u
0 for any u ∈ V , and u, u = 0 iff u = 0.
(CI2) u, v = v, u for all u, v ∈ V .
(CI3) αu1 + βu2 , v = αu1, v + βu2, v for all u1 , u2 , v ∈ V and all α, β ∈ C.

5
We note that from (CI2) each u, u is real, so (CI1) makes sense. This time we have been more
economical in the third property; (CI3) and (CI2) give u, αv1 + βv2  = αu, v1  + βu, v2  for
all u, v1 , v2 ∈ V and α, β ∈ C. The combination of this and (CI3) is called ‘sesquilinearity’
rather than bilinearity; it is linear in the first variable but conjugate linear in the second. Some
textbooks interchange the roles of the first and second factors here, modelling complex inner
products on C3 × C3 → C given by ((z1 , z2 , z3 ), (w1 , w2 , w3 )) → z1 w1 + z2 w2 + z3 w3 . 

We want to show that any inner product on a real or complex vector space V gives rise to a
norm on V . For this we shall need a slightly abstract form of Cauchy’s inequality; the more
concrete forms we need elsewhere follow easily from this abstract version. We state it first in
the real case.
Proposition S.5.4 Let (V,  ) be a real inner product space, and for any u ∈ V define ||u||

by ||u|| = u, u. Then for any u, v ∈ V ,

|u, v|  ||u|| ||v||.

Proof For any real number x, the inner product xu+v, xu+v = ||xu+v||2
0. This says that
x2 ||u||2 +2xu, v+||v||2
0 for all x ∈ R. Think of this as a quadratic expression in x. The fact
that is it always non-negative means that the quadratic equation x2 ||u||2 + 2xu, v + ||v||2 = 0
has either no real roots, or else one repeated real root; for otherwise the graph of the quadratic
function would cross the real axis twice and would be negative for some values of x. Hence the
discriminant “b2 − 4ac” of the quadratic does not exceed zero. This says u, v2  ||u||2||v||2 ,
from which the result follows. 

 The version of Cauchy’s inequality in Proposition S.5.4 holds also in the complex case, where

again we define ||u|| = u, u; the proof has an added complication, which we now explain.
We first note that any non-zero complex number w may be written in polar form, w = |w|eiθ for
some (real) θ ∈ [0, 2π). We also note that Cauchy’s inequality is trivially true if u, v = 0, so
we may assume that u, v = 0, and as above we may write it as |u, v|eiθ for some θ ∈ [0, 2π).
For any complex number z , the inner product zu + v, zu + v
0 by (CI1). Now

zu + v, zu + v = |z|2 ||u||2 + zu, v + zv, u + ||v||2 = |z|2 ||u||2 + zu, v + zu, v + ||v||2

= |z|2 ||u||2 + 2 Re(zu, v) + ||v||2.

Let x be any real number x and set z = xe−iθ . Then |z|2 = x2 and Re(zu, v) = x|u, v|,
since u, v = |u, v|eiθ . The upshot is that

x2 ||u||2 + 2x|u, v| + ||v||2
0 for all x ∈ R,

and Cauchy’s inequality follows just as in the case of a real inner product space. 

6
We can now prove
Proposition S.5.5 Given a real or complex inner product space (V,  ), we may define a

norm || || for V by the formula ||u|| = u, u.
Proof We prove this in the real case first, to avoid complications. (N1) and (N2) are straight-
forward to check.
(N3) We want to check that ||u + v||  ||u|| + ||v|| for any u, v ∈ V . It is equivalent to prove
this after squaring both sides, so it is enough to prove

u + v, u + v  ||u||2 + 2||u|| ||v|| + ||v||2.

Now u + v, u + v = ||u||2 + 2u, v + ||v||2 , so it is enough to prove u, v  ||u|| ||v||, which
follows from Cauchy’s inequality, Proposition S.5.4. 

 In the complex case, for checking (N2) we note that

αu, αu = ααu, u = |α|2 u, u, so taking square roots ||αu|| = |α| ||u||.

To check (N3) in the complex case, we note that for any complex number w we have Re w  |w|.
When we expand u+v, u+v in the complex case, we get ||u||2 +||v||2 +2Reu, v and as above
Reu, v  |u, v| Also, |u, v|  ||u|| ||v|| by Cauchy’s inequality. So Reu, v  ||u|| ||v||.
The proof of (N3) is completed as in the real case. 

In particular, Proposition S.5.4 applies to Rn with its usual inner product, and this gives the
classical form of Cauchy’s inequality: for vectors u = (a1 , a2 , . . . , an ) and v = (b1 , b2 , . . . , bn )
in Rn we have  2  

n 
n 
n
ai b1  a2i b2i .
i=1 i=1 i=1

This is used in Chapter 5 to prove that the Euclidean d2 is a metric on Rn .

It is worth mentioning a more elementary proof of the classical Cauchy inequality above. We first
check that for real numbers a1 , a2 , . . . , an and b1 , b2 , . . . , bn the following Lagrange identity
holds  2  

n 
n 
n 
ai bi = a2i b2i − (ai bj − aj bi )2 .
i=1 i=1 i=1 1
i<j
n

Since the third summation on the right-hand side is always non-negative, being a sum of squares,
Cauchy’s inequality follows.
Another case in which Proposition S.5.5 applies is Example 5.16: we can define a (real) inner
product on the space C[a, b] of all continuous functions f : [a, b] → R by the formula
 b
f, g = f (t)g(t)d t for all f, g ∈ C[a, b].
a

7
As before, we need Lemma 5.15 of the book to check positivity, but the other checks that this is
an inner product are straightforward. The abstract form of Cauchy’s inequality in Proposition
S.5.4 gives in this case the Cauchy-Schwarz-Bunyakovsky inequality: for any f, g ∈ C[a, b],
 b
2  b  b
2
f (t)g(t)d t  (f (t)) d t (g(t))2 d t.
a a a

The resulting norm is the L2 -norm discussed in the previous section.

 We can give a complex version of the above example by letting X be the set of all continuous
complex-valued functions on [a, b] and defining
 b
f, g = f (t)g(t)d t for all f, g ∈ X.
a
One can check that this gives a complex inner product space and (using Cauchy’s inequality for
a complex inner product space) that this gives rise to a normed complex vector space. 

The norms which come from an inner product have a more geometric feel than the others. This
impression is reinforced by the next result.
Proposition S.5.6 For any inner product space V , the norm arising from the inner product
satisfies the ‘parallelogram equality’: for any u, v ∈ V

||u + v||2 + ||u − v||2 = 2(||u||2 + ||v||2).

Before proving this we indicate in a diagram how it generalizes an elementary geometric prop-
erty in the Euclidean plane, that the sum of the squares of the lengths of the diagonals in a
parallelogram equals the sum of the squares of the lengths of the sides (a fact which can be
deduced from Pythagoras’ theorem).

 
@  
 @  
@ 
  
 @  
@ 
  
@ 
v  u − v@
R
@3u+v



@
  
  @ 
 @
  
  @ 
 @
  
  @ 

 - @
u

Proof of Proposition S.5.6 The proof is the same in the real and complex cases:
||u + v||2 + ||u − v||2 = u + v, u + v + u − v, u − v = 2(u, u + v, v) = 2(||u||2 + ||v||2). 

8
Proposition S.5.6 enables us to see easily that certain norms cannot arise from any inner product.
For example consider the l1 norm on Rn . Let u = (1, 0, 0, . . . , 0), v = (0, 1, 0, 0, . . . , 0). Then

||u + v||21 + ||u − v||21 = 8, but 2(||u||21 + ||v||21) = 4.

Similarly for the l∞ norm on Rn take u = (1, 1, 0, 0, . . . , 0), v = (1, −1, 0, 0, . . . , 0). We get

||u + v||2∞ + ||u − v||2∞ = 8, but 2(||u||2∞ + ||v||2∞) = 4.

(In fact it can be shown that the parallelogram equality is sufficient as well as necessary for a
norm to arise from an inner product, but we do not include the proof.)

Reason for choice of subscripts d1 , d2 , d∞ (1) The reason is slightly technical. The point
is that for every r ∈ N there is a metric on Rn called dr , associated to the r -norm which we
define by
 1/r

n
||(x1 , x2 , . . . , xn )||r = |xi |r .
i=1

In the cases r = 1 and r = 2 this gives the 1-norm and the 2-norm and hence the metrics d1
and d2 that we have already met. In fact more generally for any real number p with p
1 in
place of r there is a norm, defined by the same formula. In spite of the danger that the notation
p suggests a prime integer, we stick with the traditional terminology of “p-norm”. The check
that || ||p satisfies sub-additivity (for any p
1) involves a generalization of Cauchy’s inequality
called Minkowski’s inequality, proved below, and this is where things become slightly technical.
But before discussing Minkowski’s inequality, here is why the name for ||(x1 , x2 , . . . , xn )||∞ is
reasonable.

Recall from Exercise 4.11 (with a slight change of notation) that given non-negative real numbers
a1 , a2 , . . . , an we have

lim (ar1 + ar2 + . . . + arn )1/r = max{a1 , a2 , . . . , an }.

r→∞

Taking ai = |xi | for each i = 1, 2, . . . , n, we see that in a fairly precise sense, ||x||r → ||x||∞ as
r → ∞. (In fact, the same is true for the p-norm as p → ∞ through real values.)

 We now tackle Minkowski’s inequality (sub-additivity for || ||p when p
1).

We have already proved sub-additivity for p = 1 so let p be a real number with p > 1. We
define q (‘the conjugate exponent’ to p) by the equation 1/p + 1/q = 1. It is easy to check that
then 1/(p − 1) = q − 1, so for positive real numbers s, t we have s = tp−1 iff t = sq−1 .

9
Proposition S.5.7 (Young’s inequality) With p, q as above and any positive real numbers
a, b,
ab  ap /p + bq /q.

Proof We shall show that

 a  b
p−1
ab  t dt+ sq−1 d s = ap /p + bq /q.
0 0

The right-hand equality comes from elementary integration, while the left-hand inequality is
seen in terms of area in the diagram below. The idea is that the first integral is the area shaded
vertically while the second is the area shaded horizontally. The graph of t → tp−1 as t goes from
0 to a along the horizontal axis, matches the graph of s → sq−1 as s goes from 0 to b along
the vertical axis, because s = tp−1 iff t = sq−1 . The left-hand diagram is for the case a  b and
the right-hand one is for b  a.

t6 t6

- -
a s a s

 b
As already noted, in each diagram the horizontally shaded area is tq−1 d t while the vertically
 a 0
p−1
shaded area is s d s. In each case the sum of the shaded areas is at least as big as the
0
area ab of the rectangle with corners (0, 0), (a, 0), (0, b) and (a, b). This gives the left-hand
inequality at the beginning, and hence completes the proof. 

The proof using area is the traditional one. Here is an alternative. Let f (x) = ax − ap /p − xq /q.
Then f  (x) = a − xq−1 which is zero iff x = a1/(q−1) = ap−1 . This is a local maximum since
f  (x) = (1 − q)xq−2 , and f  (ap−1 ) < 0 since 1 − q < 0. Now f (ap−1 ) = ap − ap /p − ap /q = 0.
Also, f (0) < 0, and f (x) → −∞ as x → ∞, since q > 1. Hence f (x)  0 for all x ∈ [0, ∞),
and in particular f (b)  0, which gives Young’s inequality. 
Next we deduce Hölder’s inequality.

10
Proposition S.5.8 For real numbers a1 , a2 , . . . , an and b1 , b2 , . . . , bn and any p, q ∈ [1, ∞)
with 1/p + 1/q = 1,  1/p  1/q

n 
n 
n
|ai bi |  |ai |p |bi |q .
i=1 i=1 i=1

Proof First we reduce the problem by noting that the target inequality is homogeneous in the
sense that if it holds for given vectors a = (a1 , a2 , . . . , an ) and b = (b1 , b2 , . . . , bn ), then it
holds also for the vectors λa and µb where λ and µ are any real numbers, since the move to
these vectors simply introduces a factor of |λµ| on each side. It is therefore enough to assume
that

n 
n 
n
|ai |p = 1 = |bi |q and prove that |ai bi |  1.
i=1 i=1 i=1

(Then in the general case for vectors a = ..., (a1 , a2 , an )
and b = (b1 , b2 , . . . , bn ) we can
 
take λ = 1/( |ai |p )1/p and µ = 1/( |bi |q )1/q , and put ai = λai , bi = µbi , which gives
  q      
|ai |p = 1 = |bi | . Then |ai bi |  1 gives |ai bi |  ( |ai |p )1/p ( |bi |q )1/q as required.)
  q
So suppose that |ai |p = 1 = |bi | . For each i we see that |ai bi |  api /p + bqi /q by Young’s
inequality, Proposition S.5.7. Summing over i we get
 n  n
n  p  q
|ai bi |  ai /p + bi /q = 1/p + 1/q = 1, as required. 
i=1 i=1 i=1

We are now ready to prove Minkowski’s inequality.

Proposition S.5.9 Given vectors a = (a1 , a2 , . . . , an ) and b = (b1 , b2 , . . . , bn ) in Rn and a

real number p
1,  1/p  n 1/p  1/p

n  
n
||a + b||p  ||a||p + ||b||p, or equivalently |ai + bi |  |ai | + |bi | .
i=1 i=1 i=1

Proof We have already proved this for p = 1 so we assume that p > 1. We use Hölder’s
inequality in a slightly quirky way. Let us write ci for ai + bi . Multiplying the inequality
|ci |  |ai | + |bi | by |ci |p−1 and then summing over i we get
n 
n 
n
p p−1
|ci |  |ai ||ci| + |bi ||ci |p−1.
i=1 i=1 i=1
We apply Hölder’s inequality to each sum on the right-hand side. For the first sum we get
 1/p  1/q  1/p  1/q

n 
n 
n 
n 
n
|ai | |ci |p−1  |ai |p (|ci |p−1 )q = |ai |p (|ci |p
i=1 i=1 i=1 i=1 i=1

since (|ci |p−1)q = |ci |(p−1)q = |ci|p , using (p − 1)q = p. Similarly,

 1/p  1/q

n 
n 
n
|bi | |ci|p−1  |bi |p (|ci |p .
i=1 i=1 i=1

11
Putting these two inequalities together we get
 1/p  n 1/p   n 1/q
 n   n   
|ci |p  |ai |p + |bi |p |ci |p .
 
i=1 i=1 i=1 i=1

Finally we divide by the last factor here, and use 1 − 1/q = 1/p to get the result.  

Brief mention of lp (3) We have already mentioned that for each p ∈ [1, ∞) a norm,
called the lp norm, may be put on Rn : for x = (x1 , x2 , . . . , xn ) ∈ Rn we define
 n 1/p

||x||p = |xi |p .
i=1

The norm properties (N1) and (N2) are easily checked for || ||p , and subadditivity (N3) follows
from Minkowski’s inequality, Proposition S.5.9. 

Sequence spaces (1, 3) By a sequence space we mean a space in which the points are se-
quences, usually of real or complex numbers. (The terminology ‘sequence space’ is not univer-
sal.) Sequence spaces are examples of normed vector spaces which are much studied in modern
analysis. They are special cases of function spaces, since for example a sequence of real numbers
is formally a function S : N → R.

Sequence spaces do have applications - for example Hilbert space is used in quantum mechanics
- but here we give just a few examples of sequence spaces and ideas about them. In the supple-
mentary material for later chapters we shall return to the examples introduced here and certain
subspaces of them.

We look at the sequence spaces usually called l1 , l2 , l∞ . There are real and complex ver-
sions; analysts tend to prefer the complex versions. Here we concentrate on the real versions.
These are the infinite-dimensional analogues of Rn with the metrics d1 , d2 , d∞ (or the norms
|| ||1 , || ||2, || ||∞ ). So l2 is ‘the most geometric’, being the infinite-dimensional analogue of Rn
with the euclidean metric (or norm). It is called Hilbert space after the German mathematician
Hilbert, who is famous for many things, including the list of problems which he put forward in
1900 as the most important ones unsolved at that time - many but not all have subsequently
been solved, and all of them have given rise to interesting mathematics created to tackle them.

However, there’s a big difference between Rn with the different norms mentioned above and
its infinite-dimensional analogues - in finite dimensions these three norms give rise to met-
rics which are all Lipschitz equivalent in the sense of Chapter 6. But the infinite-dimensional

12
analogues are not equivalent in any reasonable sense, as we shall see in the supplementary
material associated with Chapter 6.

l∞ This is the space of bounded sequences of real numbers. It is a vector space, with co-
ordinate-wise addition and scalar multiplication, that is for (xn ), (yn ) ∈ l∞ and c ∈ R,

(xn ) + (yn ) = (xn + yn ), c(xn ) = (cxn ).

To see that this makes sense we need to check that the sum of two bounded sequences is bounded
and so is the scalar product of a bounded sequence by a number. These are easy to prove. The
other vector space axioms follow co-ordinatewise from the corresponding properties for real
numbers. We then define the sup norm: ||x||∞ = supn∈N |xn | where x is the sequence (xn ). Let
us check that this is indeed a norm.
(N1) For any x = (xn ) ∈ l∞ , we have ||x||∞ = supn∈N |xn |
0. Also, for any such x, ||x||∞ = 0
iff x = (0, 0, . . . , 0, . . . , ).
(N2) For any x = (xn ) ∈ l∞ and real number c we have

||cx||∞ = ||(cxn )||∞ = sup |cxn | = |c| sup |xn | = |c|||x||∞ .

n∈N n∈N

(N3) For any x = (xn ), y = (yn ) ∈ l∞ , from the triangle inequality for R and the definition of
|| ||∞ we have
|xn + yn |  |xn | + |yn |  ||x||∞ + ||y ||∞ ,

and since this is true for every n ∈ N, we get

||x + y ||∞ = sup |xn + yn |  ||x||∞ + ||y ||∞ .

n∈N

The metric arising from this norm is the sup metric, d∞ ( x, y ) = supn∈N |xn − yn |.

There are interesting subspaces of l∞ such as the one often labelled c, the space of all convergent
sequences (of real numbers). A convergent sequence is bounded, so c ⊆ l∞ , and the sum of two
convergent sequences is convergent as is the scalar product of a convergent sequence by a constant
number. So c is a vector subspace of l∞ and we take it with the sup norm.


∞
l2 2
The (real) Hilbert space l consists of all sequences (xn ) of real numbers such that x2n

n=1
∞

converges. The norm is defined by ||(xn )||2 =  x2n . It is actually an inner product space,
n=1
1 ∞
unlike l and l , and it has geometric notions such as orthonormal bases and the like. Recall
that this is ‘the infinite-dimensional generalization of Rn ’. We now prove some of these claims.

13
We first show that l2 is a real vector space under coordinate-wise addition and scalar multipli-
cation. It will be enough to show that if (xi ), (yi) ∈ l2 and a, b ∈ R then (axi + byi ) ∈ l2 . For
any n ∈ N we have

n 
n 
n 
n
2 2
(axi + byi ) = a x2i +b 2
yi2 + 2ab xi yi .
i=1 i=1 i=1 i=1

By Cauchy’s inequality in Rn ,
 1/2  1/2
 n 
n 
n
xi yi  x2i yi2 .
i=1 i=1 i=1

Hence  1/2  n 1/2


n 
n 
n 
n 
(axi + byi )2  a2 x2i + b2 yi2 + 2|ab| x2i yi2
i=1 i=1 i=1 i=1 i=1
  n 1/2  1/2 2
  
n 
= |a| x2i + |b| yi2
 
i=1 i=1
 ∞ 1/2  1/2 2
  
∞ 
 |a| x2i + |b| yi2 .
 
i=1 i=1


n
This shows that the sequence of partial sums (axi + byi )2 is bounded above, and since it
i=1
is also monotonic increasing, it converges, so the sequence (axi + byi ) is in l2 .


∞

The metric defined by this norm is d2 ((xn ), (yn )) =  (xn − yn )2 .
i=1

We can define a real inner product on l2 by the formula


∞
(xi ), (yi ) = xi yi for any (xi ), (yi ) ∈ l2 .
i=1

We first need to check that the sum on the right converges. But it is absolutely convergent, since
for any n ∈ N (using Cauchy’s inequality),
 1/2  1/2  1/2  1/2

n 
n 
n 
∞ 
∞
|xi yi |  x2i yi2  x2i yi2 ,
i=1 i=1 i=1 i=1 i=1

which says that the monotonic increasing sequence of partial sums ( ni=1 |xi yi |) is bounded
above; hence it converges.
We now check that   defined above is a real inner product on l2 . Positivity is clear -
∞
x, x = x2i = 0 iff x is the zero vector in l2 .
i=1

14
Symmetry is immediate, and bilinearity follows from (axi +byi )zi = axi zi +byi zi for real numbers,
since we can sum this for i going from 1 to ∞, to get for any vectors x, y, z ∈ l2 and any real
numbers a, b that
ax + by , z  = ax, z  + by, z .
Hence l2 is a real inner product space.

From Proposition S.5.5 we know that the formula || x ||2 = x, x gives a norm on l2 .

l1 This has to do with absolute convergence. We take l1 to be the set of all real sequences (xi )

∞
such that |xi | converges. We have to check that this is a vector space under coordinate-wise
1=1
sum and scalar product. So suppose that x = (xi ) and y = (yi ) are in l1 and a, b ∈ R. For
each i ∈ N we have |axi + byi |  |a||xi | + |b||yi|. So for each n ∈ N,


n 
n 
n 
∞ 
∞
|axi + byi |  |a| |xi | + |b| |yi|  |a| |xi | + |b| |yi|.
i=1 i=1 i=1 i=1 i=1
∞
As before this shows that i=1 |axi + byi | converges, so ax + by ∈ l1 , and this shows that the
latter is a vector space.

∞
1
We define a norm on l by ||x||1 = |xn |. The properties (N1) and (N2) are easily checked.
n=1
For (N3), suppose that x = (xi ) and y = (yi ) are in l1 . For each i ∈ N, |xi + yi |  |xi | + |yi|.
Summing this from 1 to n, we get

n 
n 
n 
∞ 
∞
|xi + yi |  |xi | + |yi |  |xi | + |yi | = ||x||1 + ||y||1 .
i=1 i=1 i=1 i=1 i=1

Since this holds for all n ∈ N it follows that ||x + y||1  ||x||1 + ||y||1 as required.
∞
The metric arising from this norm is d1 ((xn ), (yn )) = |xn − yn |.
i=1

Just as in the finite-dimensional case, we can show that || ||1 , || ||∞ do not arise from inner
products on l1 , l∞ . Also as in the finite-dimensional case, there are complex analogues of l1 , l2
and l∞ . Their story is very similar to the real case, give or take a conjugate or two.

Finally we show that l1 ⊆ l2 ⊆ l∞ , but both inclusions are strict.

If (xn ) ∈ l2 then {x2n : n ∈ N} is bounded and hence {|xn | : n ∈ N} is bounded, so l2 ⊆ l∞ .
The reverse inclusion does not hold, since if x = (1, 1, . . . , 1 . . .), then x ∈ l∞ but x ∈ l2 .

Next l1 ⊆ l2 , since if |xn | converges then |xn | is bounded, so there exists K ∈ R such that
 2
|xn |  K for all n ∈ N, so x2n  K|xn |, and by comparison xn converges. The reverse
inclusion does not hold, for if x = (1, 1/2, 1/3, . . . , 1/n, . . .), then by elementary theory of
series, x ∈ l2 but x ∈ l1 .
15
General open sets in R (2) We have seen in Example 5.33 that an ‘open interval’ (a, b) in
R is open in R according to Definition 5.32. In this section we ask what a general open set in
R looks like. The language of Chapter 12 can be used to give a more streamlined account, as
we shall mention in S.12.

Proposition S.5.10 A subset X of R is open in R iff X is the disjoint union of a countable

collection of open intervals.

Proof Since a single open interval is open in R, any union of open intervals is open in R.
Conversely, suppose that X ⊆ R is open in R. First we define an equivalence relation on X :
if x, y ∈ X then x ∼ y iff for some a, b with a < b we have x, y ∈ (a, b) ⊆ X . This is an
equivalence relation since:
Reflexivity follows from the definition of X being open in R: given any x ∈ X , there exists
ε > 0 such that (x − ε, x + ε) ⊆ X . So x ∼ x.
Symmetry follows from the definition of ∼.
Transitivity: suppose that x, y, z ∈ X and that x ∼ y and y ∼ z . Then there are intervals
(a, b), (c, d) such that x, y ∈ (a, b) ⊆ X and y, z ∈ (c, d) ⊆ X . From a < y < b and
c < y < d we get a < d and c < b. I claim that (a, b) ∪ (c, d) is an open interval (u, v) with
x, z ∈ (u, v) ⊆ X . This is one of these ‘obvious’ facts which really needs a proof. First, if either
(a, b) ⊆ (c, d) or (c, d) ⊆ (a, b) then the conclusion really is obvious. Suppose neither holds.
Then
either c < a or d > b (otherwise (c, d) ⊆ (a, b)) and also
either a < c or b > d (otherwise (a, b) ⊆ (c, d)).
So we must have either
(i) c < a and b > d or
(ii) a < c and d > b.
In case (i), (a, b) ∪ (c, d) = (c, b). To see that the left-hand side is contained in the right-hand
side, note that c < a implies (a, b) ⊆ (c, b) and b > d implies (c, d) ⊆ (c, b). Conversely if
t ∈ (c, b) then either c < t  a < d or a < t < b, so t ∈ (a, b) ∪ (c, d).
In case (ii), (a, b) ∪ (c, d) = (a, d). The proof is entirely similar to the proof in (i).

Thus ∼ is an equivalence relation. I claim that the corresponding equivalence classes are (dis-
joint) open intervals. For suppose that E is an equivalence class of X under ∼, and let x ∈ E .
Consider the set S of y ∈ R such that (x, y) ⊆ X . Then S is non-empty since X is open so
(x, x + ε) ⊆ (x − ε, x + ε) ⊆ X for some ε > 0.

16
Case (1) If S is not bounded above, then for any y > x there is some z > y with (x, z) ⊆ X ,
so y ∈ E . Thus (x, ∞) ⊆ E .
Case (2) If S is bounded above let b be its least upper bound. Then whenever x < z < b there
is some y ∈ S with z < y , so (x, y) ⊆ X and in particular z ∈ X . So (x, b) ⊆ X . Thus
(x, b) ⊆ E . Moreover, in this case there is no y > b such that (x, y) ⊆ X since that would
contradict the fact that b is the sup of S . So there is no y > b with y ∈ E .

By an entirely similar proof we can show that either (−∞, x) ⊆ E or for some a < x we have
(a, x) ⊆ E and there is no y < a such that (y, x) ⊆ E .

The upshot is that E is one of the following: (−∞, ∞), (−∞, b), (a, ∞), (a, b). Thus E is an
open interval as claimed.

It is of the nature of equivalence classes that distinct ones are disjoint. To see that there are
only countably many such equivalence classes, choose a rational nunber in each. This gives an
injection from the set of equivalence classes into Q and shows that the former is countable (see
Corollary S.2.7). This completes the proof of Proposition S.5.10. 

17