A note on `p spaces

Ulrik Skre Fjordholm

January 20, 2021

In this note we define the `p spaces (pronounced “ell-pee”, or sometimes

“little ell-pee”, to distinguish it from Lp ) and list some of their properties. The
starting point is real-valued sequences {ai }i∈N in R, which we can think of as
“infinite-dimensional vectors” (a1 , a2 , a3 , . . . ) ∈ R∞ . The space R∞ is too big
to have an interesting structure, so instead we study smaller (but still infinite-
dimensional) subspaces, namely the `p spaces.
Notation. In this note we will think of a sequence {ai }i∈N in K (where K is
one of the fields R or C) as the function
a : N → K, a(i) = ai .
It should be clear that each sequence {ai }i∈N in K gives rise to one and only one
such function a : N → K, and vice versa. We will therefore refer to a function
a : N → K as a sequence (in K). The reason for viewing sequences as functions
becomes apparent when we talk about sequences of sequences of numbers.
We first state the finite-dimensional variant of `p .
Definition 1. Let K = R or K = C and let p ∈ [1, ∞]. For n ∈ N we define the
p-norm of a vector u ∈ Kn as
( P
1/p
n
i=1 |ui |p if p < ∞
kukp =
maxi=1,...,n |ui | if p = ∞.
Definition 2. Let K = R or K = C and let p ∈ [1, ∞]. For a sequence a : N → K
we define ( P
p 1/p


kak`p = i∈N |a(i)| if p < ∞
kak`∞ = supi∈N |a(i)| if p = ∞
and we set
`p (K) = a : N → K : kak`p < ∞ .


Example 3. If a(i) = 1i then i∈N |a(i)|p = i∈N i1p , which is finite only when
P P
p
p > 1. Clearly, |a(i)| 6 1 for all i, so kak`∞ < ∞. Hence,
P a ∈ ` p(R) Pfor all p ∈
1 1
(1, ∞], but not for p = 1. Similarly, if b(i) = √i then i∈N |b(i)| = i∈N ip/2 ,
which is finite only when p > 2.
Remark 4. It is not hard to see that `p (K)P⊂ `∞ (K) for every p ∈ [1, ∞), but
∞
not vice versa. Indeed, if a ∈ `p (K) then i=1 |a(i)|p < ∞, so in particular
p
|a(i)| → 0 as i → ∞, which is equivalent to |a(i)| → 0 as i → ∞. Thus, a is
a sequence converging to 0, so it must be bounded. To see that the converse
inclusion is not true it is enough to observe that a = (1, 1, . . . ) ∈ `∞ (K) but
a∈/ `p (K) for every p < ∞.

1
 In order to show that `p (K) is a normed vector space we first prove Hölder’s
inequality For some p ∈ [1, ∞] we define its conjugate exponent q ∈ [1, ∞] by
(
p
if p < ∞
q = p−1
1 if p = ∞.

Note that
1 1
+ = 1, (1)
p q
and that q is the only element of [1, ∞] satisfying this identity. (Here, we write
1/∞ = 0.) Note also that p = 2 is the only exponent which is its own conjugate.
Theorem 5 (Hölder’s inequality, finite-dimensional version). Let p ∈ [1, ∞]
and let q be its conjugate exponent. For any n ∈ N we have

kuvk1 6 kukp kvkq ∀ u, v ∈ Kn (2)

where uv = (u1 v1 , . . . , un vn ) ∈ Kn .
Proof. If u = 0 or v = 0 then (2) follows immediately, so we may assume
u, v 6= 0. If p = 1 then q = ∞, and
n
X n
X
kuvk1 = |ui vi | 6 max |vj | |ui | = kuk1 kvk∞ .
j=1,...,n
i=1 i=1

The same argument applies to the case p = ∞, q = 1. Last, in the case

p, q ∈ (1, ∞) we first recall Young’s inequality: If s, t > 0 then
sp tq
st 6 + . (3)
p q
We get
n n
kuvk1 1 X X |ui | |vi |
= |ui vi | =
kukp kvkq kukp kvkq i=1 i=1
kuk p kvkq
n
X |ui |p |vi |q
6 p + (by Young’s inequality)
i=1
pkukp qkvkqq
Pn p
Pn
i=1 |ui | |vi |q
= p + i=1 q
pkukp qkvkq
p q
kukp kvkq 1 1
= + = +
pkukpp qkvkqq p q
=1 (by (1)).

Now multiply the above by kukp kvkq to get (2).

Theorem 6 (Hölder’s inequality, infinite-dimensional version). Let p ∈ [1, ∞]
and let q be its conjugate exponent. Then

kabk`1 6 kak`p kbk`q ∀ a ∈ `p (K), b ∈ `q (K), (4)

where ab is the sequence (ab)(i) = a(i)b(i).

2
Proof. For an arbitrary n ∈ N, apply (2) to the vectors u = (a(1), . . . , a(n)) and
v = (b(1), . . . , b(n)) to get
n
X
|a(i)b(i)| 6 kukp kvkq .
i=1

p 1/p p 1/p
Pn
P∞

From the fact that kukp = i=1 |a(i)| 6 i=1 |a(i)| = kak`p , and
likewise for v, we get
Xn
|a(i)b(i)| 6 kak`p kbk`q .
i=1

Taking the limit n → ∞ now yields (4).

Remark 7. Hölder’s inequality is useful in many applications. For instance,
Hölder’s inequality implies that if a sequence a lies in both `p1 and `p2 for
some p1 , p2 ∈ [1, ∞], then it also lies in every space “in between”. (Such a
result is sometimes called an interpolation result.) Indeed, assume that, say,
p1 < p2 < ∞, and let p ∈ (p1 , p2 ). Then there is some α ∈ (0, 1) such that p =
αp1 + (1 − α)p2 . Apply Hölder’s inequality with exponent α1 (whose conjugate
1
exponent is 1−α ) to get
∞
X ∞
X
|a(i)|p = |a(i)|αp1 |a(i)|(1−α)p2
i=1 i=1
∞
!1/(1/α) ∞
!1/(1/(1−α))
X
1/α X
(1−α)p2 1/(1−α)
|a(i)|αp1


6 |a(i)|
i=1 i=1
∞
!α ∞
!1−α
X X
p1
= |a(i)| |a(i)|p2
i=1 i=1
(1−α)p2
= kakαp
`p1 kak`p2
1
< ∞.

The case p2 = ∞ is easier:

∞
X ∞
X ∞
X
|a(i)|p = |a(i)|p1 |a(i)|p−p1 6 kak`p−p
∞
1
|a(i)|p1 < ∞.
i=1 i=1 i=1

Remark 8. Combining Remarks 4 and 7, we see that if a ∈ `p (K) then also

a ∈ `r (K) for all r ∈ [p, ∞]. (Exercise: Show by example that a ∈ `p (K) does
not necessarily imply a ∈ `r (K) for r < p.)
In order to show that k·k`p is a norm, we first show that its finite-dimensional
version k · kp is a norm.
Theorem 9. For every p ∈ [1, ∞] and n ∈ N, the function k · kp is a norm on
Kn .
Proof. It is clear that kukp > 0 for all u 6= 0, and that kukp = 0 implies
p 1/p
Pn

u = 0. Let u ∈ Kn and α ∈ K. If p < ∞ then kαukp = i=1 |αui | =
p 1/p
Pn

|α| i=1 |ui | = |α|kukp . For p = ∞ we have kαuk∞ = maxi=1,...,n |αui | =
|α| maxi=1,...,n |ui | = |α|kuk∞ .

3
 Last, we show the triangle inequality. If u, v ∈ Kn and p = 1 then
n
X n
X
ku + vk1 = |ui + vi | 6 |ui | + |vi | = kuk1 + kvk1 .
i=1 i=1

If p = ∞ then


ku + vk∞ = max |ui + vi | 6 max |ui | + |vi |
i=1,...,n i=1,...,n

6 max |ui | + max |vi | = kuk∞ + kvk∞ .

i=1,...,n i=1,...,n

Finally, if p ∈ (1, ∞), then

n
X n
X
ku + vkp = |ui + vi |p = |ui + vi | · |ui + vi |p−1
i=1 i=1
n
X n
X
p−1
6 |ui | · |ui + vi | + |vi | · |ui + vi |p−1
i=1 i=1
n
!1/p n
!1/q
X X
p (p−1)q
6 |ui | |ui + vi |
i=1 i=1
n
!1/p n
!1/q
X X
p (p−1)q
+ |vi | |ui + vi |
i=1 i=1
p
where we have applied Hölder’s inequality (2). Since q = p−1 we get (p−1)q = p.
(p−1)q 1/q
Pn

Divide both sides by i=1 |ui + vi | to get
n
!1−1/q
X
|ui + vi |p 6 kukp + kvkp .
i=1

Since 1 − 1/q = 1/p, the left-hand side equals ku + vkp , so we are done.
Theorem 10. (`p (K), k · k`p ) is a Banach space (a complete normed vector
space) for every p ∈ [1, ∞].
Proof. The proof consist of three parts: `p is a vector space, k · k`p is a norm,
and this space is complete.
Claim: k · k`p is a norm. It is clear that kuk`p > 0 for all u 6= 0, and
that kuk`p = 0 implies u = 0. Let a ∈ `p (K) and α ∈ K. If p < ∞ then
p 1/p p 1/p
P∞
P∞

kαak`p = i=1 |αa(i)| = |α| i=1 |a(i)| = |α|kak`p . For p = ∞ we
have kαak`∞ = supi∈N |αa(i)| = |α| supi∈N |a(i)| = |α|kak`∞ . Last, we show the
triangle inequality. If a, b ∈ `p (K) and p = ∞ then
ka+bk`∞ = sup |a(i)+b(i)| 6 sup |a(i)|+|b(i)| 6 sup |a(i)|+sup |b(i)| = kak`∞ +kbk`∞ .
i∈N i∈N i∈N i∈N

Finally, if p ∈ (1, ∞), let n ∈ N be an arbitrary integer and define u, v ∈ Kn by

u = (a(1), . . . , a(n)) and v = (b(1), . . . , b(n)). Then
n
!1/p
X
|a(i) + b(i)|p = ku + vkp 6 kukp + kvkp 6 kak`p + kbk`p .
i=1

4
Taking the limit n → ∞ yields ka + bk`p on the left-hand side.
Claim: `p (K) is a vector space. Most of the axioms follow immediately; we
only show that `p (K) is closed under addition and multiplication by scalars.
Indeed, from the fact that k · k`p is a norm on `p (K), we find that kαak`p =
|α|kak`p < ∞ whenever α ∈ K and a ∈ `p (K), implying that also αa ∈ `p (K).
If a, b ∈ `p (K) then ka + bk p

`p 6 kak`p + kbk`p < ∞, so also a + b ∈ ` (K).
p
Claim: ` (K), k · k`p is complete. Let {an }n∈N be a Cauchy sequence in
`p (K). Then for every ε > 0 there is some N ∈ N such that kan − am k`p < ε
when n, m > N , so in particular,

|an (i) − am (i)| 6 kan − am k`p < ε ∀ i ∈ N.

It follows that for each i ∈ N, the sequence {an (i)}n∈N is a Cauchy sequence in
K. Since K is complete, {an (i)}n∈N is convergent, converging to some a(i) ∈ K.
We claim that the sequence a = (a(1), a(2), . . . ) lies in `p (K) and that {an }n∈N
converges to a. We split the proof into the cases p = ∞ and p < ∞.
p = ∞: Let ε > 0 and let N be as above. Then |a(i)−an (i)| = limm→∞ |am (i)−
an (i)| 6 ε for all n > N and i ∈ N, so

|a(i)| 6 |a(i) − an (i)| + |an (i)| 6 ε + kan k`p .

Since this holds for every i ∈ N we get kak`∞ 6 ε + kan k`p < ∞, so
a ∈ `∞ (K). Moreover, ka − an k`∞ = supi∈N |a(i) − an (i)| 6 ε, so we
conclude that an → a as n → ∞.
p < ∞: Let ε > 0 and let N be as above. For every n, I ∈ N we have

I
!1/p I
!1/p
X X
|a(i) − an (i)|p = lim |am (i) − an (i)|p 6 ε,
m→∞
i=1 i=1
| {z }
6kan −am k`p <ε

so ka − an k`p 6 ε. Hence, ka − an k`p → 0 as n → ∞. Last, from the

inverse triangle inequality,

kak`p 6 kan k`p + ka − an k`p 6 kan k`p + ε < ∞,

so a ∈ `p (K).

We complete this note by showing that `p has a (Schauder) basis whenever

p < ∞.
Proposition 11. `p (K) is infinite-dimensional for all p ∈ [1, ∞]. If en ∈ `p (K)
is given by
en = (0, . . . , 0, 1, 0, . . . )
(the 1 occuring in the nth position), then {en }n∈N is a Schauder basis for `p (K)
for every p ∈ [1, ∞), but not for `∞ (K).

5
Proof. The set {en }n∈N is infinite and linearly independent, so `p (K) is infinite-
< ∞ and a ∈ `p (K), let αi = a(i) for each i ∈ N. Then the
dimensional. If p P
n
partial sum sn = i=1 αi ei = (a(1), . . . , a(n), 0, . . . ) satisfies

∞
!1/p
X
p
ka − sn k`p = |a(i)| .
i=n+1
P∞ p
From the fact that i=1 |a(i)|P∞ < ∞, the above sum must converge to 0 as
n → ∞. It follows that a = i=1 αi ei . This proves that {en }n∈N is a Schauder
basis for `p (K). P∞
For `∞ (K), let a = (1, 1, . . . ) ∈ `∞ (K). If αi are such that a = i=1 αi ei
then necessarily αi = 1 for all i. But

ka − sn k`∞ = k(0, . . . , 0, 1, 1, . . . )k`∞ = 1,

a contradiction. Hence, {en }n∈N is not a Schauder basis for `∞ (K).

It can also be shown that `∞ does not possess any Schauder basis. In this
sense, `∞ is “much bigger” than the other `p spaces.