A crash course in p-adic analysis

W. H. Schikhof
It is a pleasure to make many acknowledgements. First to Marı́a
Soledad Alcaı́no A. and Marı́a Eugenia Heckmann G. for typing
the manuscript. My thanks are also due to Carla Barrios R. and
Tonino Costa A., graduate students of the Pontificia Universidad
Católica de Chile, who made a thorough revision of the typewrit-
ten text, and materially assisted in its preparation.
Contents

1 BASICS 4
1.1 Ultrametric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.1 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 Ultrametrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.3 Compact Ultrametric Spaces . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.4 Spherical Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Non-Archimedean Valued Fields . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Valued Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.2 Non-Archimedean Valued Fields . . . . . . . . . . . . . . . . . . . . . . 8
1.2.3 Sequences and series in K . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.4 Power Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.5 Continuous Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.6 Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Normed and Banach spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.1 Normed Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.2 Operators with an empty spectrum . . . . . . . . . . . . . . . . . . . . . 21
1.3.3 Commutation Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.4 Addendum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2 ORTHOGONALITY 23
2.1 Definition of orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Orthogonal bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 COUNTABILITY 32
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4 DUALITY 36

5 COMPACTOIDS 42
5.1 List of properties of compactoids . . . . . . . . . . . . . . . . . . . . . . . . . . 43

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 2

INTRODUCTION

In many branches of mathematics and its applications the fields R and C are playing a
fundamental role. For quite some time one has been discussing the consequences of replacing
in those theories R or C by a non-archimedean valued field. This story started in algebra
and number theory, where the p-adic fields, discovered by Hensel in 1909, could be used
successfully. Around 1930 analysis entered the picture through the work of Schnirlemann who
developed the basic theory of p-adic power series and analytic functions. From 1940 on the
Dutch mathematician A.F. Monna established the basics on p-adic Functional Analysis. At
this moment there are non-archimedean activities in all feasible disciplines such as Elementary
Calculus, Theory of C n - and C ∞ -functions, p-adic Lie Groups, Analytic Functions in one or
several variables, Algebraic and Analytic Number Theory, Algebraic Geometry, Functional
Analysis, ...
Of course, people are asking for applications.

1. There do exist several applications within mathematics.

2. In 1987 Igor Volovich raised the question as to whether at Planck distances (10−34 cm)
space should be disordered or disconnected. He suggested the use of p-adic numbers
to build more adequate quantum mechanical models. Right now serious mathematician
are working in this area.
Ultrametrics also appear speculatively in Psychology, Social Sciences, Economy (stock
market). Attempts are made to build a p-adic Probability Theory. A conference on
these applications was held, for the first time in history, in Moscow, October 2003.

Parts in the text, indicated ‘Background’are not needed for the course, but might be interesting
to hear about.

November, 2003.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 3

Chapter 1

BASICS

1.1 Ultrametric Spaces

In this section we recall well-known facts.

1.1.1 Metrics
A metric on a set X is a map d : X × X −→ [0, ∞) (d = distance) such that for all x, y, z ∈ X

(i) d(x, y) = 0 if and only if x = y

(ii) d(x, y) = d(y, x)

(iii) d(x, z) ≤ d(x, y) + d(y, z) (triangle inequality).

X = (X, d) is called a metric space. (If we relax (i) to just d(x, x) = 0 then we have a
semi-metric d)
For a ∈ X, r > 0 we set
B(a, r) := {x ∈ X : d(x, a) ≤ r},
called the ‘closed’ ball about a with radius r and

B(a, r− ) := {x ∈ X : d(x, a) < r}

called the ‘open’ ball about a with radius r.

A subset U ⊂ X is called open if for each a ∈ U there exists an r > 0 such that B(a, r− ) ⊂ U .
The collection of open sets form a topology on X which is called the topology induced by d.
Observe that an ‘open’ ball is open and that a ‘closed’ ball is closed.
The diameter of a non-empty set Y ⊂ X is

diam Y := sup{d(x, y) : x, y ∈ Y }

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 4

(possibly ∞). We set diam ∅ := 0. Y is called bounded if diam Y < ∞.
The distance between two non-empty sets Y, Z ⊂ X is

d(Y, Z) := inf{d(y, z) : y ∈ Y, z ∈ Z}.

For y ∈ X we write d(y, Z) rather than d({y}, Z).

A sequence x1 , x2 , . . . in X converges to x ∈ X, notation lim xn = x, if lim d(x, xn ) = 0.
n→∞ n→∞
A sequence x1 , x2 , . . . in X is called Cauchy sequence if lim d(xm , xn ) = 0.
m,n→∞
Exercise 1.A. A Cauchy sequence is bounded. Each convergent sequence is Cauchy, the
converse is not true.
Two metrics on a set X are called equivalent if they induce the same topology.
Exercise 1.B. Let d1 , d2 be equivalent metrics on a set X. Prove that a sequence is convergent
in (X, d1 ) if and only if it is convergent in (X, d2 ), but that a similar statement does not hold
for Cauchy sequences.
A metric space (X, d) is called complete if each Cauchy sequence converges.
Each metric space X can be embedded into its completion X # (a metric space X # that is
complete and contains X as a dense subset).
Let (X1 , d1 ), (X2 , d2 ) be two metric spaces. A map f : X1 −→ X2 is called an isometry if
d2 (f (x), f (y)) = d1 (x, y) for all x, y ∈ X1 .

1.1.2 Ultrametrics
A metric d on a set X is called ultrametric (and (X, d) is called an ultrametric space) if it
satisfies the so-called strong triangle inequality

d(x, z) ≤ max(d(x, y), d(y, z))

for all x, y, z ∈ X. (Clearly this implies the ‘ordinary’ triangle inequality). In the spirit of
above one defines semi-ultrametrics. We have the fundamental
ISOSCELES TRIANGLE PRINCIPLE:

If d(x, y) 6= d(y, z) then d(x, z) = max(d(x, y), d(y, z)).

Example: Subsets of a non-archimedean valued field (see next Section) with the ultrametric
(x, y) 7→ |x − y|.
[BACKGROUND: In fact we have this way all examples, since each ultrametric space can
isometrically be embedded into a non-archimedean valued field (Indag. Math. 46 (1984),
51-53)]
Exercise 1.C. Let (X, d) be a metric space. Then, among all semi-ultrametric that are ≤ d
there is a largest one. It is called the subdominant semi-ultrametric for d.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 5

The next Exercise is essential. For this course it is necessary that you have proved all facts
yourself once in your lifetime.
Exercise 1.D. Let (X, d) be an ultrametric space.

(i) Let a ∈ Y ⊂ X. Then diam Y = sup{d(x, a) : x ∈ Y }.

(ii) Each ball in X is both closed and open (‘clopen’).

(iii) Each point of a ball is a center. Each ball has an empty boundary.

(iv) The radii of a ball B form the set {r ∈ R, r1 ≤ r ≤ r2 }, where r1 = diam B, r2 =

dist (B, X \ B) (r2 = ∞ if B = X). It may happen that r1 < r2 , so that a ball may
have infinitely many radii.

(v) Two balls are either disjoint, or one is contained in the other.

(vi) If two balls B1 , B2 are disjoint, then dist (B1 , B2 ) = d(x, y) for each x ∈ B1 , y ∈ B2 .

(vii) Let ε > 0. The relation d(x, y) < ε (x, y ∈ X) is an equivalence relation and induces
a partition of X into ‘open’ balls of radius ε. A similar story holds for d(x, y) ≤ ε and
‘closed’ balls.

(viii) Let Y ⊂ X, B ball in X, B ∩ Y 6= ∅. Then, B ∩ Y is a ball in Y .

The topology induced by an ultrametric is zerodimensional, i.e. there is a base of the topology
consisting of clopen sets. Hence, an ultrametric space is totally disconnected.
If (X, d) is an ultrametric space and Y ⊂ X is dense then

{d(y1 , y2 ) : y1 , y2 ∈ Y } = {d(x1 , x2 ) : x1 , x2 ∈ X}.

Thus, completion of an ultrametric space does not create new values of the ultrametric.

1.1.3 Compact Ultrametric Spaces

Exercise 1.E. Let (X, d) be a compact ultrametric space. Show that {d(x, y) : x, y ∈ X} is
countable and has only 0 as possible accumulation point.
Let (X, d) be an infinite compact ultrametric space (if X is finite the process below breaks off).
By the above exercise the set of non-zero values of d is a sequence r1 > r2 > . . . tending to 0.
We shall make a picture of the collection of all balls in X as follows. We have X = B(a, r1 )
for any a ∈ X. The relation d(x, y) ≤ r2 (or d(x, y) < r1 if you want) decomposes X into
finitely many ‘closed’ balls of radius r2 :

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 6

 Level 1: rX
.

AD . . .

 DA ..

 D A
Level 2: r r DDr AAr


dots represent balls of radius r2 .
connecting lines indicate inclusion.

Each of the closed balls of level 2 decomposes into finitely many ‘closed’ balls of radius r2 :

r.

DA . . . .

 DA .

 DA
r.
r DDr AAr



 .. E AB@

 ..  E

 .  E
 BA@
 BA @
r r r r Er r r

  E  BBrAAr @

@r

etc. Continuing this way we arrive at a tree.

The dots represent the balls; the infinite paths starting from X, going below represent the
elements of X.
The tree determines X up to homeomorphisms.
Trees appear in social sciences (hierarchical structures), so no wonder why compact ultrametric
spaces enter the scene!

1.1.4 Spherical Completeness

An ultrametric space (X, d) is called spherically complete if each nested sequence of balls
B1 ⊃ B2 ⊃ . . . has a non-empty intersection. (It is not required that the diameters of the Bn
approach 0).
It is not difficult to see that a spherically complete ultrametric space is complete. Surprisingly,
the converse is not true, as we will see later.
Exercise 1.F. Prove that a compact ultrametric space is spherically complete.
The next exercise reveals the main reason why attention is devoted to spherical completeness.
Exercise 1.G. Let X be a spherically complete ultrametric space embedded in an ultrametric
space Y . Then each y ∈ Y has a best approximation in X, i.e.

min{d(y, x) : x ∈ X}

exists.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 7

1.2 Non-Archimedean Valued Fields
Again, most of this section is old stuff.

1.2.1 Valued Fields

Let K be a (commutative) field. A valuation is a map | | : K −→ [0, ∞) such that for all
λ, µ ∈ K

(i) |λ| = 0 if and only if λ = 0

(ii) |λµ| = |λ| |µ|

(iii) |λ + µ| ≤ |λ| + |µ| (triangle inequality).

Obvious examples are Q, R, C with the ordinary absolute value function. The mapping
(λ, µ) 7→ |λ − µ| is a metric on K making it into a topological field (i.e. the basic arithmetic
operations are continuous).
The metric completion of K is in a natural way again a valued field.
Two valuations on a field K are called equivalent if they induce the same topology.
[BACKGROUND: e.g. G. Bachman, Introduction to p-adic numbers and valuation theory,
Academic Press, New York, 1964: Two valuations | . |1 , | . |2 on K are equivalent if and only
if there is a positive constant c such that | . |1 = | . |2 c .]
The following theorem essentially separates R, C from all other valued fields. It shows the
‘alternative character’ of our non-archimedean analysis.
[BACKGROUND: Bachman, see before, page 127:
Theorem Let (K, | . |) be a valued field. Then, there are only two possibilities. Either

(i) K is a subfield of C and | . | is equivalent to the absolute value function, restricted to

K, or

(ii) the valuation | . | is non-archimedean (n.a) i.e. it satisfies the strong triangle inequality

|λ + µ| ≤ max(|λ|, |µ|) (λ, µ ∈ K). ]

So, by excluding C and its (valued) subfields we obtain in return the strong triangle inequality.

1.2.2 Non-Archimedean Valued Fields

In this subsection (K, |.|) is a non-archimedean valued field. The expression ‘non-archimedean’
is explained by the fact that for each n ∈ N

|n · 1| = |1 + 1 + . . . + 1| ≤ max(|1|, |1|, . . . , |1|) = 1

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 8

contrasting with the Archimedean Axiom for R stating that N is unbounded.
We recall some facts that you know from a previous life.

– The completion of a non-archimedean valued field is non-archimedean valued.

– Let (K, |.|) be a non-archimedean valued field. Then K ∗ := {λ ∈ K, λ 6= 0} is a

multiplicative group and so is |K ∗ | = {|λ| : λ ∈ K ∗ }; it is called the value group of K.
This value group can either be dense in (0, ∞) or discrete.
In the last case, if the valuation is not trivial, ρ := max{r : r ∈ |K ∗ |, r < 1} exists. Any
element π ∈ K for which |π| = ρ is called a uniformizing element.
Furthermore, B(0, 1), the ‘closed’ unit disk, is a ring and B(0, 1− ), the ‘open’ unit disk,
is a maximal ideal in B(0, 1).
The quotient is a field:
k := B(0, 1)/B(0, 1− )
and is called the residue class field of K.

– The p-adic valuation on Q is determined by

|n|p := p−( number of factors p in n) (n ∈ N).

The completion of (Q, |.|p ) is called (Qp , |.|p ), the field of the p-adic numbers.
Its value group is {pn : n ∈ Z}, its residue field is the field Fp of p elements.
(Of course, in the above, p is a prime number)

p is a uniformizing element. The fields Q2 , Q3 , Q5 , Q7 , . . . are mutually non-isomorphic.

The closed unit disk of Qp is denoted Zp and called the ring of the p-adic integers. (This
name can be explained by the fact that Zp is the closure of Z).
Zp is compact. More generally, each ball in Qp is compact.
Qp is therefore spherically complete, separable, locally compact.
Qp is not algebraically closed.
|.|p can be extended uniquely to the algebraic closure Qap and the completion of (Qap , |.|p ) is
called Cp , the field of the p-adic complex numbers.
Cp is no longer locally compact, but separable and algebraically closed. Its value group is

{pr : r ∈ Q},

so the valuation is dense, and the residue class field is the algebraic closure of Fp , hence
infinite.
Cp is spherically complete! This follows from:

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 9

Theorem 1 Let (K, |.|) be a n.a. valued field. Suppose K is separable and has a dense
valuation. Then K is not spherically complete.

Proof There exist r1 > r2 > . . . in |K ∗ | such that r := limn→∞ rn > 0. Let {s1 , s2 , . . .} be a
countable dense subset. There is a ‘closed’ ball B1 of radius r1 such that s1 ∈ / B1 . Since B1
decomposes into infinitely many ‘closed’ balls of radius r2 (why infinite?) there is a ‘closed’
ball B2 ⊂ B1 such that s2 ∈ / B2 . Going on this way we find a nested sequence B1 ⊃ B2 ⊃ . . .
of closed balls with radius Bn = rn , and sn ∈ / Bn for each n.
T
Let B := n Bn . If B 6= ∅, it would be a ‘closed’ ball of radius r, hence open so the set
{s1 , s2 , . . .} must meet B. But on the other hand, by construction {s1 , s2 , . . .} ∩ B = ∅, a
contradiction.
FROM NOW ON IN THIS COURSE K = (K, |.|) IS A N.A. VALUED COMPLETE FIELD.
WE ASSUME THAT |.| IS NON-TRIVIAL I.E. THERE IS A λ ∈ K WITH |λ| =
6 0, |λ| =
6 1.
Exercise 1.H. Show that {λ ∈ K : |1 − λ| < 1} is a multiplicative subgroup of {λ ∈ K :
|λ| = 1}.

1.2.3 Sequences and series in K

We recall a few fundamental facts.

– Let a1 , . . . , an ∈ K and |ai | =

6 |aj | whenever i 6= j.
Then THE STRONGEST WINS:

|a1 + . . . + an | = max{|ai | : 1 ≤ i ≤ n}.

n
X
– Let a1 , a2 , . . . ∈ K. This sequence is called summable if lim ai exists. Then
n→∞
i=1
A STUDENT’S DREAM COME TRUE:

lim an = 0 =⇒ a1 , a2 , . . . is summable.
n→∞

– If lim an = a 6= 0 then |an | = |a| for large n. Thus, we have in Qp

n→∞

1
= 1 + p + p2 + . . .
1−p

∞
P
Because lim n! = 0 in every Qp the sum n! exists in every Qp . The following problem has
n→∞ i=0
been open since 1971.
P∞
PROBLEM: Can n! be rational for some prime p?
i=0

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 10

 ∞
P
Everybody believes that n! is irrational for every p. Several ‘proofs’ have proposed but
i=0
they all failed.
∞
P
It is not even known if n! 6= 0 in every Qp .
i=0
Recently Vladimirov (2002) introduced
n−1
X
!n := k!
k=0

and proposed the following related CONJECTURE:

g.c.d.(!n, n!) = 2.

∞
P
Exercise 1.I. Compute n (n!) in Qp .
n=0

1.2.4 Power Series

Like in the archimedean case, given a0 , a1 , . . . ∈ K then

R := (lim sup n |an |)−1

p

(where ∞−1 := 0 , 0−1 := ∞) is called the radius of convergence of the power series an xn .
P

Just like in the complex case one proves that a0 , a1 x, a2 x2 , . . . is summable for |x| < R, not
summable for |x| > R. The behaviour on the ‘boundary’ |x| = R is much easier that in the
complex case: a0 , a1 x, . . . is either summable everywhere on {x : |x| = R} or nowhere. This is
because the region of convergence {x ∈ K : a0 , a1 x, . . . is summable} = {x ∈ K : lim an xn =
n→∞
0}.
Exercise 1.J. ∞
xn is {x ∈ K : |x| < 1}.
P
(i) Prove that the region of convergence of
n=0
∞
P xn
(ii) Prove that, in Cp , n exists if and only if |x| < 1.
n=1
So, we can define the p-adic logarithm via the formula
∞ n
X x
− logp (1 − x) = (|x| < 1)
n
n=1

Exercise 1.K. Let n ∈ N be written in base p as

n = a0 + a1 p + . . . + as ps

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 11

where ai ∈ {0, 1, . . . , p − 1} and as 6= 0.
Define sn (the sum of digits) as a0 + a1 + . . . + as . Show that the number of factors p in n!
equals
n − sn
p−1
and prove with the help of this fact that in Cp
∞ n
X x
exp x :=
n!
n=0
1 1
is defined for |x| < p 1−p and that the series is not summable if |x| ≥ p 1−p .
1
From now on we denote {x ∈ Cp : |x| < p 1−p } by Ep .
Exercise 1.L. By brute force one can prove that

exp(x + y) = exp x · exp y for x, y ∈ Ep and

logp (xy) = logp x + logp y for |x − 1| < 1, |y − 1| < 1, and
exp logp x = x (x ∈ 1 + Ep )
logp exp x = x (x ∈ Ep ).

More surrealistic is the following. Prove that

| exp x − exp y| = |x − y| (x, y ∈ Ep )

| logp x − logp y| = |x − y| (x, y ∈ 1 + Ep ).

[BACKGROUND: One can prove that the p-adic logarithm maps {x ∈ Cp : |1 − x| < 1} onto
Cp and that logp x = 0 if and only if x is a root of unity.]

1.2.5 Continuous Functions

Let X be a topological space. The set of all continuous functions X −→ K is a K-vector
space (even a K-algebra) under pointwise operations, and is denoted C(X −→ K), or C(X)
if no confusion about the scalar field is to be expected.
We define the K-valued characteristic function of a subset Y ⊂ X by
(
1 if x ∈ Y
ξY (x) :=
0 if x ∈ X \ Y

It is easily seen that ξY is continuous if and only if Y is clopen.

A function f : X −→ K is locally constant if each point of X has a neighbourhood on which
f is constant. Locally constant functions are of course continuous.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 12

Theorem 2 Let X be a topological space, f ∈ C(X −→ K), ε > 0. Then there is a locally
constant function g : X −→ K such that |f (x) − g(x)| < ε for all x ∈ X.

Proof (Compare Exercise 1.D (vii)) The relation |f (x)−f (y)| < ε (x, y ∈ X) is an equivalence
S
relation and yields a partition of X into clopen sets: X = Ui .
i∈I
Choose ai ∈ Ui for each i and define

g(x) := f (ai ) whenever i ∈ I, x ∈ Ui

This g does the job.

1.2.6 Differentiability
Now we take for X a subset of K. To avoid problems, assume X is without isolated points.
A function f : X −→ K is called differentiable at a ∈ X if

f (x) − f (a)
f 0 (a) := lim
x→a x−a
exists (you can imagine how to define lim ). In the same spirit we define (everywhere) dif-
x→a
ferentiable function, derivative. The well-known rules for differentiation of sums, products,
quotients, compositions (‘chain rule’) hold also in the n.a. case. The proofs are classical.
A differentiable function is continuous. Rational functions without poles in X are differen-
tiable.
To be able to define analytic functions properly we need the following exercise on ‘double
sequences’.
Exercise 1.M. For m, n ∈ N, let amn ∈ K. We say that limm+n→∞ amn = 0 if, for each
ε > 0, the set {(m, n) ∈ N × N : |amn | ≥ ε} is finite. Show that the following statements
(α), (β), (γ) below are equivalent.

(α) limm+n→∞ amn = 0.

(β) For each n, limm→∞ amn = 0; limn→∞ amn = 0 uniformly in m.

(γ) For each m, limn→∞ amn = 0; limm→∞ amn = 0 uniformly in n.

∞ P
P ∞ ∞ P
P ∞
If (α) − (γ) hold then amn and amn both exist and are equal.
m=1n=1 n=1m=1

Lemma 3 Suppose a power series in x − a converges in a neighbourhood of a ∈ K:

∞
X
f (x) := an (x − a)n (x ∈ B(a, r)) .
n=0

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 13

Then, for each b ∈ B(a, r), f can be expanded into a power series in x − b on the same ball:
there are b0 , b1 , . . . ∈ K such that
∞
X
f (x) = bn (x − b)n (x ∈ B(b, r) = B(a, r)).
n=0

Proof We may assume r ∈ |K ∗ |. Then lim |an |rn = 0.

n→∞
Put
n  
n n
X n
(x − a) = (x − b + b − a) = (x − b)k (b − a)n−k
k
k=0

and  
n
tkn := an (x − b)k (b − a)n−k .
k
Now clearly limk→∞ tkn = 0 for each n and since
 
 n  k n−k
|tkn | ≤ |an |   r r ≤ |an | rn
k 

( nk is an integer, hence  nk  ≤ 1)




we have that lim tkn = 0 uniformly in k. So, by the exercise we have

n→∞

∞ X
X ∞ ∞ X
X ∞ ∞
X
f (x) = tkn = tkn = bk (x − b)k
n=0k=0 k=0n=0 k=0

∞
n
(b − a)n−k .
P

where bk := an k
n=0
Definition Let B be a ‘closed’ ball in K with radius r ∈ |K ∗ |. A function f : B −→ K is
called analytic if there exists an a ∈ B and a0 , a1 , . . . ∈ K such that
∞
an (x − a)n
P
(∗) f (x) = (x ∈ B).
n=0

With the help of Lemma 3 one easily derives that sums and products of analytic functions on
B are analytic. Analytic functions are differentiable. If f is as in (∗) then
∞
X
0
f (x) = nan (x − a)n−1 .
n=1

If f is a complex analytic function defined on some bounded domain the maximum principle
states that |f | takes its maximum on the boundary. A similar result holds also for our n.a.
case; we include a proof since it is completely different from the ‘complex’ proof.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 14

Theorem 4 (Maximum Principle) Let K = Cp and B, r, f, a, an be as in the above defi-
nition. Then |f | attains its maximum on B and, moreover,

max{|f (x)| : x ∈ B} = max{|f (x)| : |x − a| = r}

= max{|an |rn : n ∈ {0, 1, 2, . . .}}.

Proof We only consider the case where B = B(0, 1). (The reader can furnish the missing
details). Thus, we have
∞
an xn
P
(∗∗) f (x) = (|x| ≤ 1).
n=0

We clearly have |f (x)| ≤ maxn |an xn | ≤ maxn |an |.

To prove that max{|f (x)| : |x| ≤ 1} ≥ max |an | we may assume that maxn |an | = 1 and
derive a contradiction from the assumption |f (x)| < 1 for all x ∈ B(0, 1). Let λ 7→ λ be the
homomorphism B(0, 1) −→ k (where k is the residue class field). Then from (**) we obtain
N
X
f (x) = 0 = an xn (x ∈ k)
n=0

for some N (since |an | < 1 for large n), for which aN 6= 0. Thus, a nonzero polynomial in k[x]
has each element of k as a root. But k is infinite, a contradiction.
Thus we have max{|f (x)| : |x| ≤ 1} = maxn |an |. To prove that also max{|f (x)| : |x| = 1} =
maxn |an | observe that |f (x)| < 1 for all |x| = 1 yields
N
X
0= an xn
n=0

for all x ∈ k \ {0}. Now reason as above, using that k \ {0} is infinite as well.
[BACKGROUND: Krull Valued Fields.
When looking at the requirements for a non-archimedean valuation:

(i) |x| = 0 if and only if x = 0

(ii) |xy| = |x| |y|

(iii) |x + y| ≤ max(|x|, |y|)

one notices that, unlike for the archimedean triangle inequality, addition of real numbers
does not play a role; one only needs multiplication and ordering. This leads to the following
generalized concept of valued fields.
Let G be a commutative, multiplicatively written, totally ordered group such that

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 15

 (∗) g1 ≤ g2 =⇒ hg1 ≤ hg2 for all h ∈ G.

(Example: (0, ∞) with the usual multiplication and ordering. Observe that R\{0} is a totally
ordered commutative group but that it does not satisfy (*)!)
Augment G with an element 0 and put for all g ∈ G

0·g =g·0=0·0=0
0<g

Definition A Krull valuation on a field K is a surjective map | · | : K −→ G ∪ {0} satisfying

(i), (ii), (iii) of above for all x, y ∈ K. G is called the value group of K.
The map (x, y) 7→ |x − y| behaves like an ultrametric, and induces a zerodimensional topology
on K in the usual way. However, we have to keep in mind that its values are in G ∪ {0}, not
in R, and that G can be ‘very big’ so that K may not be metrizable!
Now let us first construct a Krull valued field, different from the previous ones. For each
n ∈ N, let Gn be the infinite cyclic group with generator gn > 1. So Gn consists of the
ordered sequence . . . , gn−2 , gn−1 , 1, gn , gn2 , . . . (In fact, each Gn is isomorphic to the additive
group Z).
Now let G be the algebraic direct sum
M
G= Gn .
n

It becomes a totally ordered group if we order G antilexicographically: i.e. if

g = (g1n1 , g2n2 , . . . , gknk , 1, 1, . . .)

g 0 = (g1m1 , g2m2 , . . . , gkmk , 1, 1, . . .)

are two distinct elements of G we say that g < g 0 if nk < mk . Or: if nk = mk and nk−1 < mk−1 .
Or if nk = mk , nk−1 = mk−1 and nk−2 < mk−2 , etc.
Observe that this G is not isomorphic -as an ordered group- to a multiplicative subgroup of
(0, ∞) since it contains bounded non-trivial subgroups like

G1
G1 ⊕ G2
..
.
G1 ⊕ G2 ⊕ . . . ⊕ Gn
..
.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 16

Now let F := R(X1 , X2 , . . .) be the field of rational functions with real coefficients in countably
many variables X1 , X2 , . . .. It is easily seen that the requirements

|0| : = 0
|λ| : = 1 if λ ∈ R \ {0}
|Xn | = (1, 1, . . . , gn , 1, 1, . . .)
↑
n-th place

determine a Krull valuation | · | : F −→ G ∪ {0}.

The completion of (F, | · |) can be constructed in the natural way and leads to a complete
Krull valued field L. To see a striking difference with the ‘ordinary’ valuation, consider the
element X1−1 ∈ F . Then |X1−1 | = (g1−1 , 1, 1, . . .), so 0 < |X1−1 | < 1. But X1−n does not tend
to 0, as |X1−n | = (g1−n , 1, 1, . . .) > (1, g2−1 , 1, . . .) > 0 for all n!
We will return to this subject later on.]

1.3 Normed and Banach spaces

1.3.1 Normed Spaces
Let E be a vector space over K. A norm on E is a map k · k : E −→ [0, ∞) satisfying

(i) kxk = 0 if and only if x = 0,

(ii) kλxk = |λ|kxk,

(iii) kx + yk ≤ max(kxk, kyk)

for all x, y ∈ E, λ ∈ K.
(Requirement (iii) is not a logical necessity but there are good reasons for assuming it, see
also Exercise 1.N).
A complete normed space is, as customary, called a Banach space.
Two norms k.k1 and k.k2 on a K-vector space E are called equivalent if the metrics (x, y) 7−→
kx − yk1 and (x, y) 7−→ kx − yk2 induce the same topology. The following Proposition can be
proved as in the archimedean case.

Proposition 5 Two norms on a K-vector space E, say k · k1 and k · k2 , are equivalent if and
only if there exist constants 0 < c < C such that

ck · k1 ≤ k · k2 ≤ Ck · k1 .

On a finite-dimensional space all norms are equivalent, and the space is a Banach space with
respect to each norm.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 17

Examples (See also page 22)
n
P
1. Let e1 , . . . , en be a base of an n-dimensional space over K. Then λi ei 7−→ maxi |λi |
i=1
is a norm.

2. Let l∞ be the space (pointwise operations) of all bounded sequences in K. For x =

(λ1 , λ2 , . . .) ∈ l∞ , define
kxk∞ := sup{|λn | : n ∈ N}.
With this norm l∞ is a Banach space.

3. c0 := {(λ1 , λ2 , . . .) ∈ l∞ , limn→∞ λn = 0} is a closed subspace of l∞ , hence a Banach

space.

4. Let X be a topological space, let BC(X) be the space of all f ∈ C(X) that are bounded.
It is a Banach space with respect to the supremum norm

f 7→ kf k∞ := sup{|f (x)| : x ∈ X}.

5. Let M be a complete valued field containing K as a subfield. Then, M is a Banach

space over K. Thus, in particular, Cp is a Banach space over Qp . Observe that |Cp |
strictly contains |Qp | showing that nonzero vectors cannot always be normalized. This
leads to the following open problem.

PROBLEM (1960) Let (E, k·k) be a Banach space over K. Does there exist an equivalent
norm k · k0 such that kEk0 ⊂ |K|, i.e. for each nonzero x ∈ E there exists a λ ∈ K such
that kλxk = 1?

There are some partial answers and reductions, but the answer is so far unknown! A
good test case would be l∞ over Cp .

Exercise 1.N. A-NON-EXAMPLE.

∞
(i) Prove that l2 := {(λ1 , λ2 , . . .) ∈ K N : |λn |2 < ∞} is a K-vector space.
P
n=1
s
∞
|ξn |2 satisfies (i), (ii) of page 17 and
P
(ii) Show that x 7−→ kxk :=
n=1

kx + yk ≤ kxk + kyk (x, y ∈ l2 )

but not the strong triangle inequality (iii) of page 17.

(iii) Compute the largest norm on l2 that is ≤ k · k (compare Exercise 1.C)

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 18

Exercise 1.O. Let (E, k · k) be a normed space. Show that kEk \ {0} := {kxk : x ∈ E, x 6= 0}
is a union of multiplicative cosets of |K ∗ | in (0, ∞), and that, conversely, each such set equals
kEk \ {0} for some normed space (E, k · k).
Consider in a normed space (E, k · k) the ball with radius r > 0 about 0:

B(0, r) := {x ∈ E : kxk ≤ r}.

Clearly x ∈ B(0, r), λ ∈ K, |λ| ≤ 1 =⇒ λx ∈ B(0, r), but also because of the strong triangle
inequality, x, y ∈ B(0, r) =⇒ x + y ∈ B(0, r). So, in algebraic language, B(0, r) is a module
over the ring {λ ∈ K : |λ| ≤ 1}! Such submodules are called absolutely convex.
[BACKGROUND: a seminorm on a K-vector space E is a map q : E −→ [0, ∞) satisfying
q(λx) = |λ|q(x), q(x + y) ≤ max(q(x), q(y)) but q(x) is allowed to be 0 for nonzero x. Like
in the classical case one can define a locally convex topology on E as the weakest topology
for which each member of a collection P of seminorms is continuous. A subbase of zero
neighbourhoods is formed by the sets {x ∈ E : q(x) < ε} where q ∈ P, ε > 0. By the above
observation these sets are absolutely convex. There is an extensive theory on locally convex
spaces on K.]
[BACKGROUND: Normed spaces over Krull valued fields?
Let (K, | · |) be a Krull valued field with value group G. A norm k · k on a K-vector space E
should satisfy (i), (ii), (iii) of page 17 but it is not at once clear what a ‘natural home’ for norm
values should be. Only admitting G ∪ {0} turns out to be too restrictive. This leads to the
concept of a so-called G-module; this is an ordered set X on which G acts monotonically. For
full details see p-adic Functional Analysis, Lecture Notes in Pure and Applied Mathematics
207, 233-293. Marcel Dekker. New York (1999).
Remark So far, nobody has attempted constructing a theory of locally convex spaces over
Krull valued fields!]
Let us return to the basic theory.

Proposition 6 Let E, F be normed spaces over K, let T : E −→ F be linear. Then, if T is

continuous at 0 then T is Lipschitz i.e. there is an M > 0 such that kT x − T yk ≤ M kx − yk
for all x, y ∈ E.

Proof There is a δ > 0 such that x ∈ E, kxk ≤ δ implies kT xk ≤ 1. Choose π ∈ K, 0 <

|π| < 1, and let x ∈ E be arbitrary, nonzero. Then there is a λ ∈ K such that

(∗) |π|δ ≤ kλxk ≤ δ

(take the trouble to verify this!). Then kT (λx)k ≤ 1, so kT xk ≤ |λ|−1 which is, by (*),
≤ |π|−1 δ −1 kxk, and we have

kT xk ≤ M kxk (x ∈ E)

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 19

with M := |π|−1 δ −1 . Then by linearity, for x, y ∈ E, kT x − T yk = kT (x − y)k ≤ M kx − yk.
[BACKGROUND: In theory of normed spaces over Krull valued fields (*) is not always possible
and there are, indeed, continuous linear maps that are not Lipschitz!]
Let L(E, F ) be the space of all continuous linear operators E −→ F . For each T ∈ L(E, F ),
among all M ≥ 0 such that kT xk ≤ M kxk for all x ∈ E there is a smallest one which is called
kT k.
We have kT xk ≤ kT kkxk for all x ∈ E and the map T 7−→ kT k is a norm on L(E, F ).
We write L(E) for L(E, E) and E 0 (dual space) for L(E, K), where K is normed by the
valuation | · |.
Exercise 1.P. In complex functional analysis the dual space of c0 is isomorphic to l1 . In the
non-archimedean theory the situation is radically different:
For x ∈ c0 , y ∈ l∞ , say x = (λ1 , λ2 , . . .), y = (µ1 , µ2 , . . .), put
∞
X
< x, y >:= λn µn
n=1

(Show that the sum exists!). For each y ∈ l∞ define T y ∈ c00 by

(T y)(x) :=< x, y > (x ∈ c0 ).

Prove that, indeed, T y ∈ c00 and that T maps l∞ linearly and isometrically onto c00 . Hence,
in popular form: ‘the dual of c0 is l∞ ’.

Like in the classical theory one can prove that L(E, F ) is a Banach space as soon as F is a
Banach space. Hence, E 0 , E 00 , E 000 , . . . are all Banach spaces.
Here is a construction that works in both classical and n.a. context, but that you may not
have seen yet: forming of quotients.
Let E be a normed space over K, let D ⊂ E be a closed subspace. We define a natural norm
on E/D as follows. Let π : E −→ E/D be the canonical map (assigning to each x ∈ E the
coset x + D). Set
kπ(x)k := dist(x, D) = inf{kx − dk : d ∈ D}.
Straightforward verification shows that kλzk = |λ|kzk, kz + uk ≤ max(kzk, kuk) for all z, u ∈
E/D, λ ∈ K. Suppose kzk = 0; we prove that z = 0.
z has the form x + D, so kzk = inf{kx − dk : d ∈ D} = 0. Hence, there are d1 , d2 , . . . ∈ D
with limn→∞ kx − dn k = 0 i.e., x = limn→∞ dn . Since D is closed, x ∈ D which implies that
x + D is the zero element of E/D.

Theorem 7 If E is a Banach space, D is a closed subspace then E/D is a Banach space.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 20

Proof Let z1 , z2 , . . . be Cauchy in E/D. Then limn→∞ (zn+1 − zn ) = 0, so there are v1 , v2 , . . .
∞
P
in E such that π(v1 ) = z1 , π(vn+1 ) = zn+1 −zn (n ≥ 1), and limn→∞ vn = 0. Then, x := vn
n=1
∞
P
exists in E and π(x) = π(vn ) = limn→∞ zn .
n=1

The Uniform Boundedness Principle, the Banach Steinhaus Theorem, the Closed Graph The-
orem and the Open Mapping Theorem all rest on the Baire Category Theorem and some
linearity considerations and therefore remain valid in our n.a. theory.
We like to conclude this introductory chapter by signalling two striking differences with the
classical case.

1.3.2 Operators with an empty spectrum

Let E be a Banach space, T ∈ L(E). Like in the classical case one defines the spectrum of T ,
σ(T ) as
{λ ∈ K : T − λI has no inverse }
(If dim E < ∞, σ(T ) is the collection of eigenvalues). A famous classical result is that σ(T )
is non-empty if the scalar field is C.
Now take the Banach space M of Example 5 in page 18 and let a ∈ M \ K and put
T x := xa (x ∈ M )
Then T ∈ L(M ) but for λ ∈ K we have (T − λI)x = (a − λ)x, hence T − λI has an inverse
x 7−→ (a − λ)−1 x. Hence, σ(T ) = ∅.
PROBLEM. Develop a reasonable spectral theory for operators on a Banach space.
(So far, only partial results have been obtained for so-called compact operators.)

1.3.3 Commutation Relations

It is well-known that in a complex Banach space E there do not exist A, B ∈ L(E) satisfying
(∗) AB − BA = I
an important relation in Quantum Mechanics. Turning to the non-archimedean situation, let
E := c0 . In 1996, A. Khrennikov and A. Kochubei independently found the following. Set
A(λ1 , λ2 , . . .) = (λ2 , λ3 , . . .)
B(λ1 , λ2 , . . .) = (0, λ1 , 2λ2 , 3λ3 , . . .).
One verifies immediately that A, B ∈ L(E) (B is bounded since |nλn | ≤ |λn |!) and that (*)
holds.
PROBLEM. Try to find all solutions of (*). (Last October 2003 I asked various experts about
this but nothing seems to be known apart from the above example!)

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 21

1.3.4 Addendum
On page 18 I should have included the following example of a space of analytic functions. Let
K = Cp (more generally, let K have infinite residue class field), let r ∈ |K ∗ | and consider the
space Hr of all analytic functions B(0, r) −→ K (see page 14).
∞
an xn (|x| ≤ r):
P
From the maximum principle (Theorem 4) it follows that for f ∈ Hr , f (x) =
n=0

kf k∞ := max{|f (x)| : |x| ≤ r} = max |an |rn

n

(so the sequence a0 , a1 , . . . is uniquely determined by f ) and the space Hr is (linearly) iso-
metrically isomorphic to {(a0 , a1 , . . .) : |an |rn −→ 0}, with the max norm.
For r = 1 we see that H1 ∼ = c0 . (Show that also Hr ∼ = c0 if r ∈ |K ∗ |.) So, Hr is a Banach
space.
COROLLARY. The uniform limit of a sequence of analytic functions on B(0, r) (r ∈ |K ∗ |) is
again analytic.
Exercise 1.Q. Show that k · k∞ on Hr is multiplicative i.e. for f, g ∈ Hr

kf gk∞ = kf k∞ kgk∞

Hence, the ring Hr has no zero divisors!

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 22

Chapter 2

ORTHOGONALITY

2.1 Definition of orthogonality

The finest examples of classical Banach spaces are Hilbert spaces, so one may look for a non-
archimedean pendant.
An inner product ( · , · ) in a space over C satisfies

(i) x 7−→ (x, y) is linear in x, for each y,

(ii) (x, y) = (y, x),

(iii) (x, x) > 0 whenever x 6= 0.

In the non-archimedean case we replace complex conjugation by a field automorphism λ 7→ λ∗

for which λ∗∗ = λ and |λ∗ | = |λ| for all λ ∈ K. (We allow ∗ to be the identity). As K is not
ordered, we replace (iii) by (x, x) 6= 0 whenever x 6= 0. So, on a K- vector space E, let us say
that a map ( · , · ) : E × E −→ K is an inner product if

(i)’ x 7→ (x, y) is linear for each y ∈ E

(ii)’ (x, y) = (y, x)∗ for all x, y ∈ E

(iii)’ (x, x) 6= 0 whenever x 6= 0, x ∈ E.

Exercise 2.A. Requirement (iii)’ seems to be a kind of arbitrary relaxation of (iii), born out
of emergency. But it is not such a ‘wild’ generalization as one may think. In fact, prove the
following. In a vector space over C, any form ( · , · ) satisfying (i), (ii) and (iii)’ is either an
inner product or else −( · , · ) is an inner product!
For E as above and x ∈ E we put
p
kxk = |(x, x)|.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 23

(Notice the valuation signs, compare with Exercise 2.A.)
¿From this point we could proceed with the theory like in classical theory. However, we will
never reach the goal we had in mind, according to the following theorem which we state
without proof.

Theorem 8 Let ( · , · ) be an inner product on a Banach space E over K such that |(x, x)| =
kxk2 for all x ∈ E. Suppose that each closed subspace has an orthogonal complement. Then
dim E < ∞.

[BACKGROUND: Theorem 8 is not the last word that is to be said about the subject. The
picture changes completely if we allow Krull valuations on the scalar field. In fact, let L be
the completion of the Krull valued field R(X1 , X2 , . . .) as constructed on pages 16–17. Put
X0 = 1. Let E be the L- vector space of all sequences (ξ0 , ξ1 , . . .) ∈ LN for which ∞ 2
P
i=0 ξi Xi
converges in L. For x, y ∈ E, say x = (ξ1 , ξ2 , . . .) and y = (η1 , η2 , . . .) put
∞
X
(x, y) := ξi ηi Xi .
i=0

H. Keller showed in 1980 that ( · , · ) is an inner product in the sense of (i)0 , (ii)0 , (iii)0 of page 23
(where λ∗ = λ for all λ ∈ L), but, what is more, he proved has every closed subspace D of E has
an orthogonal complement, i.e. D + D⊥ = E when D⊥ = {x ∈ E : (x, y) = 0 for all y ∈ D}.
The study those ‘non-classical Hilbert spaces’ is in full progress.]
So, for our theory to develop, inner products do not seem appropriate. But we do have a
powerful non-archimedean concept, valid in any normed space:
Definition Let x, y be elements of a normed space E over K. We say that the vector x is
(norm-)orthogonal to y if the distance of x to the space Ky is precisely kxk:

kxk = min{kx − λyk : λ ∈ K}

(so the minimum is attained for λ = 0. Draw a picture). We denote this by x ⊥ y.

Before continuing we state

VAN ROOIJ’S PRINCIPLE:

If x, y ∈ E, kx − yk ≥ kxk, then kx − yk ≥ kyk.

The proof is obvious.

It can be used for

Proposition 9 If x ⊥ y then y ⊥ x.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 24

Proof We only have to prove that ky − µxk ≥ kyk for all µ ∈ K. This is clear for µ = 0. If
µ 6= 0 we have ky − µxk = |µ| kx − µ−1 yk, which by x ⊥ y is ≥ |µ|kxk = kµxk.
By van Rooij’s Principle, ky − µxk ≥ kyk.
It is the symmetry that makes ⊥ important. It is revealed in the following alternative defini-
tion.

Proposition 10 x ⊥ y if and only if for each λ, µ ∈ K

kλx + µyk = max(kλxk, kµyk).

Proof Left to the reader.

[BACKGROUND: The relation ⊥ can in the same way be introduced in archimedean normed
spaces, so one may wonder why you have never met it in the courses you took. The reason is
that in classical theory ⊥ is not symmetric. Essentially the only spaces where ⊥ is symmetric
are inner product spaces and there x ⊥ y is equivalent to (x, y) = 0.]
The following two Exercises show that one has to be careful by not automatically assuming
‘form orthogonality’ properties to hold for our norm orthogonality.
Exercise 2.B. Let K 2 be normed by

(λ1 , λ2 ) 7−→ max(|λ1 |, |λ2 |).

Prove that (1, 0) ⊥ (0, 1) (not surprising), but also that (1, 1) ⊥ (1, 0) and (1, 1) ⊥ (0, 1).
Exercise 2.C. Let E be a normed space. Prove that 6⊥ (‘not orthogonal’) is an equivalence
relation on E \ {0}. Thus, if x ⊥ y and x 6⊥ z then y ⊥ z!
Inspired by Proposition 10 we now define:
Definition A sequence e1 , e2 , . . . of non-zero vectors in a normed space E is called orthogonal
if for each n ∈ N, λ1 , . . . , λn ∈ K
n
X
λi ei = max{kλi ei k : 1 ≤ i ≤ n}



i=1

(i.e. each ei is orthogonal to every vector in the linear span of {ej : j 6= i}). The sequence is
orthonormal if, in addition, ken k = 1.
The difference with classical orthogonality is demonstrated again in the following

Theorem 11 (Perturbation theorem) Let e1 , e2 , . . . be an orthogonal sequence in a nor-

med space E and let f1 , f2 . . . ∈ E be such that kfn − en k < ken k for all n. Then f1 , f2 , . . . is
orthogonal.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 25

 n
X n
X
Proof Choose λ1 , . . . , λn ∈ K. Set x = λi ei , y = λi fi .
i=1 i=1
n
X
By assumption we have kfi k = kei k for all i. Hence, ky−xk = k λi (fi −ei )k ≤ max |λi |kfi −
i
i=1
ei k < (if not all λi = 0) < maxi |λi |kei k = kxk. It follows that kyk = kxk = maxi |λi |kei k =
maxi |λi |kfi k.

2.2 Orthogonal bases

In this section E is a Banach space over K. It is not hard to see that, for an orthogonal
sequence e1 , e2 , . . . and λ1 , λ2 , . . . ∈ K with kλn en k −→ 0 we have
∞

X
λi ei = max kλi ei k.


i
i=1

∞
X
Definition e1 , e2 , . . . is called an orthogonal base for E if x ∈ E has an expansion x = λi ei
i=1
for some suitable λi ∈ K.
Clearly such an expansion is unique.

Proposition 12 If e1 , e2 , . . . is orthogonal and [e1 , e2 , . . .] (K-linear span) is dense in E then

e1 , e2 , . . . is an orthogonal base.

Proof Let π ∈ K, 0 < |π| < 1. Then there exists µ1 , µ2 , . . . ∈ K such that |π| ≤ kµn en k ≤ 1
∞
X
for all n ∈ N. For (λ1 , λ2 , . . .) ∈ c0 , put T (λ1 , λ2 , . . .) = λn µn en .
n=1
One verifies immediately that T is a linear map c0 −→ E, that, for x = (λ1 , λ2 , . . .) ∈ c0 :

|π|kxk = |π| max |λn | ≤ kT xk ≤ kxk,

n

so that T is a homeomorphism, hence T c0 is complete. But also T c0 contains all finite linear
combinations of e1 , e2 , . . ., hence T c0 is dense. It follows that T c0 = E, which is what we
wanted to prove.
The most famous example of an orthogonal base is the so-called Mahler base of C(Zp −→ Cp ):
For x ∈ Zp and n ∈ {0, 1, 2, . . .} we set
 
x x(x − 1) . . . (x − n + 1)
en (x) := :=
n n!

Then we have

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 26

 1. en is a polynomial of degree n,

2. en (m) = 0 if n > m, en (n) = 1,

3. en (−1) = (−1)n ,

4. ken k∞ = 1 for all n,

5. e0 , e1 , . . . is an orthonormal sequence.

(Proof. Let λ0 , λ1 , . . . , λn ∈ K. Then by taking the value at 0, we get

kλ0 e0 + . . . + λn en k∞ ≥ |λ0 e0 (0)| = |λ0 |ke0 k∞ .

By van Rooij’s Principle,

kλ0 e0 + . . . + λn en k∞ ≥ kλ1 e1 + . . . + λn en k∞ .

By substituting the value x = 1 we get:

kλ1 e1 + . . . + λn en k∞ ≥ |λ1 e1 (1)| = |λ1 |ke1 k.

Continuing this way we obtain

kλ0 e0 + . . . + λn en k∞ ≥ max{kλi ei k : 0 ≤ i ≤ n}
= max{|λi | : 0 ≤ i ≤ n}).

To conclude the proof that e0 , e1 , . . . is an orthonormal base some work has to be done. First,
we need
Exercise 2.D. (Continuation of Exercise 1.K)

(i) Show that for j ∈ {1, 2, . . . , pn }  n 

 p  −n
 j =p .
 

(ii) Show that spj = sj for each j ∈ N (here again sm is the sum of digits of n in base p)

(iii) Use (i) and (ii) for  n 

 p  |pn |
 j  = |j|
  (j ∈ {1, 2, . . . , pn }).

To get an idea for the proof we imagine that f ∈ C(Zp −→ Cp ) has an expansion. Then what
would be the candidates for the coefficients? So suppose f (x) = a0 x0 + a1 x1 + . . .

By taking x = 0 we find f (0) = a0 .

By taking x = 1 we find f (1) = a0 + a1 .

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 27

By taking x = 2 we find f (2) = a0 + 2a1 + a2
..
.
So, a0 = f (0), a1 = f (1) − f (0), a2 = f (2) − 2f (1) + f (0).
In general we proceed as follows: Let E : C(Zp −→ Cp ) −→ C(Zp −→ Cp ) be the shift
operator
(Ef )(x) = f (x + 1) (x ∈ Zp ).
Then
∞  
X x+1
(Ef )(x) = an .
n
n=0

By using      
x+1 x x
= + (n ≥ 1)
n n n−1
we find
∞  X ∞  
X x x
(Ef )(x) = an + an
n n−1
n=0 n=1
∞  
X x
= f (x) + an+1
n
n=0

so that
∞  
X x
(E − I)f (x) = an+1
n
n=0

(I =identity operator). By applying E − I k times

∞  
k
X x
(E − I) f (x) = an+k
n
n=0

and by putting x = 0:
(E − I)k f (0) = ak .
n  
X n
Because (E − I)n = (−1)n−k E k we have
k
k=0
n  
X n
(*) an = (−1)n−k f (k)
k
k=0

and
n  
k
X n
(E − I) f (x) = (−1)n−k f (x + k).
k
k=0

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 28

Since
n   n  
X n X n
0 = (1 − 1)n = (−1)n−k = (−1)n−k f (x)
k k
k=0 k=0
we get
n  
n
X n
(**) (E − I) f (x) = (−1)n−k (f (x + k) − f (x)).
k
k=1

We will prove that for f ∈ C(Zp −→ Cp ), k(E − I)n f k∞ −→ 0.

n
Since kf k∞ ≥ k(E−I)f k∞ ≥ k(E−I)2 f k∞ ≥ . . . it suffices to show that k(E−I)p f k∞ −→ 0.
Let ε > 0. By uniform continuity of f there is an l such that |k| < p−l implies |f (x+k)−f (x)| <
ε for all x.
Choose n so largethatp−n < p−l ε. Let k ∈ {1, . . ., pn }.
 pn n
If |k| < p−l then  (−1)p −k (f (x + k) − f (x)) < ε.

 kn    n  n
 kf k∞ ≤ |p | kf k∞ (by Exercise
−l
 p pn −k
  p 
If |k| > p then   (−1) (f (x + k) − f (x)) ≤ 
k k  |k|
−n+l
2.D.) < p kf k∞ < εkf k∞ . Thus, we have proved that k(E − I)n f k∞ −→ 0, using (∗∗).
From this fact and (∗) it follows that an −→ 0 for every f ∈ C(Zp −→ Cp ). This means that
∞  
X x
g : x 7−→ an
n
n=0

is a well-defined continuous function on Zp .

Inductively, from (E −I)n f (0) = (E −I)n g(0) for each n, we find f (0) = g(0), f (1) = g(1), . . ..
Since N is dense in Zp we have f = g. So we have proved:

Theorem 13 (Mahler) The functions e0 , e1 , . . . given by

 
x
en (x) =
n

form an orthonormal base of C(Zp −→ Cp ). If f ∈ C(Zp −→ Cp ) has the expansion

∞
X
f= an en
n=0

then
n  
X n
an = (−1)n−k f (k).
k
k=0

Corollary 14 (Weierstrass, p-adic version) Continuous functions on Zp can uniformly

be approximated by polynomial functions.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 29

[BACKGROUND: One can prove the following generalization.
STONE–WEIERSTRASS: Let X be a compact topological space, let H be a subalgebra of
C(X −→ K) that contains the constants and separates the points of X. Then H is uniformly
dense in C(X −→ K).]
Exercise 2.E. (Another application: the ‘indefinite sum’).
Rx
In classical analysis a function of the form x 7−→ a f (t)dt is often called ‘indefinite integral’
of f . Show that in the p-adic case we have an ‘indefinite sum’ for an f ∈ C(Zp −→ Cp ):
There is an F ∈ C(Zp −→ Cp ) such that

F (n) = f (0) + f (1) + . . . + f (n − 1) (n ≥ 1)

F (0) = 0.

Also prove the next

Corollary 15 (No p-adic Haar integral) Let ϕ ∈ C(Zp −→ Cp )0 have the property that
ϕ(f1 ) = ϕ(f ) (f ∈ C(Zp −→ Cp ), where f1 is the ‘shift’ f1 (x) = f (x + 1). Prove that ϕ = 0.
1
Exercise 2.F. Let α ∈ Cp , |α| < p 1−p . Then exp αx is defined for all x ∈ Zp . Show that its
Mahler expansion is
∞  
X x
(exp α − 1)n .
n
n=0

[BACKGROUND: One can prove that if a Banach space E has an orthogonal base then so has
every closed subspace D. However it is not always true that D has an orthogonal complement
i.e. a closed subspace S with D + S = E and D ⊥ S. It depends on K and kEk.]
In the next Chapter we will use the following version of ‘approximate orthogonality’.
Definition Let t ∈ (0, 1]. A sequence e1 , e2 , . . . in a normed space is called t-orthogonal if for
all n ∈ N, λ1 , . . . , λn ∈ K

Xn
λi ei ≥ t max{kλi ei k : i ∈ {1, . . . , n}}.



i=1

In the same spirit we have the notion of t-orthogonal base of a Banach space.
1- orthogonal = orthogonal.
If e1 , e2 , . . . is a t-orthogonal base of E and we introduce a new norm k k∼ by
∼
∞

X
λ n n = max kλn en k
e


n
n=0

then e1 , e2 , . . . is an orthogonal base of E with respect to k k∼ , and the norms k k and k k∼

are equivalent.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 30

Thus, the fact that E has a t-orthogonal base for some t ∈ (0, 1] means nothing else but: there
is an equivalent norm for which E has an orthogonal base.
Exercise 2.G. Verify the t-version of Van Rooij’s Principle: If t ∈ (0, 1] and kx − yk ≥ tkxk
then kx − yk ≥ tkyk.
Prove also a t-version of Proposition 12.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 31

Chapter 3

COUNTABILITY

3.1 Introduction
A topological space is called separable if it has a countable dense subset. If we have a metric
space that is separable then any subset is also separable.
In classical Functional Analysis some role is played by the separable Banach spaces.
In our n.a. case separability is not the right notion to work with. In fact, if K happens to be
non-separable then each onedimensional space has the same property.
To overcome this problem one has “linearized”the definition of separability as follows.
Definition A normed space E is called of countable type if it has a countable subset whose
linear hull is dense.
It is an easy exercise to show that if K itself is separable then a normed space over K is of
countable type if and only if it is separable.
It has been an open question for quite some time as to whether each separable Banach space
∞
X
over C has a Schauder base e1 , e2 , . . . i.e. each vector has a unique expansion λn en , where
n=1
λn ∈ C. Only in 1974 Enflo proved that the answer was negative.
Further, in the complex case there are many non-isomorphic separable Banach spaces e.g. lp
for 1 ≤ p < ∞. In the non-archimedean case the situation is much simpler as we will show
now.

Lemma 16 Every finite-dimensional space has, for each t ∈ (0, 1), a t-orthogonal base.

Proof Let E have dimension n, let x1 , . . . , xn be an algebraic base for E. Choose t1 , . . . , tn−1 ∈
√
(0, 1) such that their product t1 t2 . . . tn−1 ≥ t (e.g. ti = n−1 t).
Set e1 = x1 , D1 := [e1 ] = [x1 ] ([·] indicates linear span). Now D1 is closed, x2 ∈ / D so
dist(x2 , D2 ) > 0.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 32

Maybe this distance is not attained but certainly there is a d1 ∈ D1 such that

kx2 − d1 k ≤ t−1
1 dist(x2 , D1 ).

Set e2 = x2 − d1 . Then D2 := [e1 , e2 ] = [x1 , x2 ] and dist(e2 , D1 ) = dist(x2 , D1 ), so

ke2 k ≤ t−1
1 dist(e2 , D1 ).

Continuing this way we can find an d2 ∈ D2 with

kx3 − d2 k ≤ t−1
2 dist(x3 , D2 ).

By putting e3 = x3 − d2 we have D3 := [x1 , x2 , x3 ] = [e1 , e2 , e3 ] and

ke3 k ≤ t−1
2 dist(e3 , D2 ).

Inductively we arrive at a base e1 , . . . , en of E such that



 ke2 k ≤ t−1
1 dist(e2 , [e1 ])
 ke3 k ≤ t−1 dist(e3 , [e1 , e2 ])


2
(∗) .

 .
.

 ke k ≤ t−1 dist(e , [e , . . . , e

n n n 1 n−1 ]).

We claim that e1 , . . . , en is t-orthogonal. In fact, let λ1 , . . . , λn ∈ K. We have to prove

kλ1 e1 + . . . + λn en k ≥ t max{kλi ei k : 1 ≤ i ≤ n}.

By the last line of (∗) we have if λn 6= 0

kλ1 e1 + . . . + λn en k = |λn ||λ−1 −1

n λ1 e1 + . . . λn λn−1 en−1 + en k
≥ |λn | dist(en , [e1 , . . . en−1 ])
≥ tn |λn |ken k,

so we have
kλ1 e1 + . . . + λn en k ≥ tn kλn en k
(this formula is also true for λn = 0). By the extension of van Rooij’s Principle, Exercise 2.G,
we have
kλ1 e1 + . . . + λn en k ≥ tn kλ1 e1 + . . . + λn−1 en−1 k.
Now with the same trick as above we can prove

kλ1 e1 + . . . + λn−1 en−1 k ≥ tn−1 kλn−1 en−1 k

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 33

leading to
kλ1 e1 + . . . + λn en k ≥ tn tn−1 kλn−1 en−1 k.
Inductively we arrive at

kλ1 e1 + . . . + λn en k ≥ tn kλn en k
kλ1 e1 + . . . + λn en k ≥ tn tn−1 kλn−1 en−1 k
..
.
kλ1 e1 + . . . + λn en k ≥ tn tn−1 . . . t1 kλ1 e1 k.

Now tn , tn tn−1 , . . . are all ≥ t and we obtain

kλ1 e1 + . . . + λn en k ≥ t max{kλi ei k : 1 ≤ i ≤ n},

which was to be shown.

Remark. One can prove that, if K is spherically complete, all distances in the above proof
are attained and hence that each finite-dimensional space has an orthogonal base! But, if K is
not spherically complete one can construct a two-dimensional space without orthogonal base.
[It can be done with the knowledge built up so far. In fact, let B(λ1 , r1 ) ⊃ B(λ2 , r2 ) ⊃ . . . be
balls in K with empty intersection. Define for (α1 , α2 ) ∈ K 2

k(α1 , α2 )k = lim |α1 − α2 λj |

j→∞

(show first that this limit exists!). Verify that k · k is a norm in K 2 and that k(α1 , α2 )k ∈ |K|
for all (α1 , α2 ) ∈ K 2 . Show that (α1 , α2 ) ⊥ (1, 0) =⇒ α1 = α2 = 0. Finally, show that two
arbitrary non-zero vectors in K 2 are not orthogonal.]

Theorem 17 Let t ∈ (0, 1). Then each Banach space of countable type has a t-orthogonal
base.

Proof Let E be an infinite-dimensional Banach space of countable type.

∞
Y
Choose t1 , t2 , . . . ∈ (0, 1) such that ti ≥ t (show that this can be done). Let x1 , x2 , . . . be
i=1
such that [x1 , x2 , . . .] is dense in E. We may assume that x1 , x2 , . . . are linearly independent.
By continuing the induction of the proof of Lemma 16 to infinity we obtain e1 , e2 , . . . ∈ E such
that [e1 , e2 , . . .] = [x1 , x2 , . . .] is dense in E and such that e1 , . . . , en is t1 . . . tn - orthogonal for
each n. Then it follows that e1 , . . . , en is t-orthogonal for each n i.e. e1 , e2 , . . . is t-orthogonal.
Now apply Exercise 2.G.

Corollary 18 Each infinite-dimensional Banach space of countable type is linearly homeo-

morphic to c0 .

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 34

A projection is an operator P ∈ L(E) with P 2 = P . If P 6= 0 then kP k ≥ 1. P is an
orthoprojection if kP k ≤ 1. Then Im P ⊥ Ker P and Im P ⊕ Ker P = E.

Corollary 19 Let E be a Banach space of countable type, let D ⊂ E be a closed subspace.

Then D and E/D are of countable type. For each ε > 0 there is a projection P with P E = D
and kP k ≤ 1 + ε. Let 0 < t0 < t < 1. Then each t-orthogonal base of D can be extended to a
t0 -orthogonal base of E.

Proof By Theorem 7 the space E/D is Banach. That it is of countable type is obvious. By
p
Theorem 17 it has a t0 /t-orthogonal base g1 , g2 , . . .. Let π : E −→ E/D be the canonical
map. By definition of the quotient norm there are f1 , f2 , . . . ∈ E with π(fi ) = gi and kfi k ≤
p
(t/t0 )kgi k for each i. It is easily seen that the map
∞
X ∞
X
Q: λi gi 7−→ λi fi (λi gi → 0)
i=1 i=1

is in L(E/D, E) and that π ◦ Q is the identity on E/D, and that kQk ≤ (t0 )−1 t. Then put
P := I − Q ◦ π; one verifies that kP k ≤ (t0 )−1 , that P is a projection onto D. Finally, if
e1 , e2 , . . . is a t-orthogonal base of D then e1 , f1 , e2 , f2 , . . . is easily seen to be a t0 -orthogonal
base of E.

Corollary 20 ((1 + ε)-HAHN-BANACH) Let E be a Banach space of countable type, let

D be a subspace, let f ∈ D, let ε > 0. Then f can be extended to an f ∈ E 0 such that
kf k ≤ (1 + ε)kf k.

Proof We may assume that D is closed. Then choose f := f ◦ P , where P is a projection

onto D with norm ≤ 1 + ε.
[BACKGROUND: This is probably the strangest proof of the Hahn-Banach you ever encoun-
tered! The classical proof for real scalars uses the fact that a collection of closed bounded
intervals with the finite intersection property has a non-empty intersection. If K is spheri-
cally complete one can use this proof, with obvious modifications, to arrive at a Hahn-Banach
Theorem, the one you are used to (E arbitrary, ε = 0. See next Chapter). But if K is not
spherically complete then taking ε = 0 in Corollary 20 leads to a falsity. In fact one can prove
that in K 2 , normed like in Remark, page 34, the linear function f : (λ, 0) 7−→ λ cannot be
extended to an f ∈ (K 2 )0 with kf k = 1].

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 35

Chapter 4

DUALITY

Theorem 21 (HAHN-BANACH THEOREM) Let K be spherically complete, let E be

a normed space over K, let D be a subspace and f ∈ D0 . Then f can be extended to an f ∈ E 0
such that kf k = kf k.

Proof Like in the classical case, through Zorn’s Lemma it suffices to prove it for the case
E = D + Ka where a ∈ E \ D. By linearity, if we fix η := f (a) then f is determined:

λa + d 7−→ λη + f (d) (λ ∈ K, d ∈ D).

The requirement kf k ≤ kf k means

(*) |λη + f (d)| ≤ kf kkλa + dk (λ ∈ K, d ∈ D).

We notice that (*) is true for λ = 0. Then, to have (*) for λ 6= 0 it suffices to have (*) for
λ = 1:
|η + f (d)| ≤ kf kka + dk (d ∈ D).
Thus
η ∈ B(−f (d), kf kka + dk) (d ∈ D).
To be able to choose η this way we need that the balls have a non-empty intersection. By
spherical completeness it suffices to prove that each two of them have a non-empty intersection,
i.e. we have to see if two ‘centers’−f (d1 ), −f (d2 ) have distance ≤ the maximum of kf kka +
d1 k, kf kka + d2 k. But that is true:

|f (d1 ) − f (d2 )| ≤ kf kkd1 − d2 k ≤ kf kkd1 + a − a − d2 k ≤ max(kf kka + d1 k, kf kka + d2 k).

For each normed space E over K we have, as in the complex case, the canonical map jE :
E −→ E 00 given by
jE (x)(f ) = f (x) (f ∈ E 0 , x ∈ E).

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 36

Exercise 4.A. (Has nothing to do with non-archimedean aspects). Prove that jE is continu-
ous, linear and that kjE k ≤ 1. Also, let T ∈ L(E, F ), where F is a second normed space, let
T 0 : F 0 −→ E 0 be defined as T 0 (f ) := f ◦ T (f ∈ F 0 ). T 0 is called the adjoint of F . Show that
the diagram:

E T -F
jE jF
? ?
E 00 - F 00
T 00

is commutative.
Exercise 4.B. Apply Corollary 20 and Theorem 21 for one-dimensional subspaces D to prove
the following: Let E be a Banach space of countable type or let K be spherically complete.
Then the canonical map jE : E −→ E 00 is isometrical.
A normed space E is called reflexive if jE is an isometrical bijection E −→ E 00 . Reflexive
spaces are complete.
One proves easily that finite - dimensional spaces are reflexive. We first study reflexivity for
spherically complete K.

Lemma 22 Let K be spherically complete, let E be a reflexive Banach space over K. Then
every closed subspace is also reflexive.

Proof Let D be a closed subspace, let i : D −→ E be the inclusion and π : E −→ E/D the
quotient map.
By Exercise 4.A. we have the commutative diagram:

D i -E π - E/D
jD jE jE/D
? ? ?
D00 - E 00 - (E/D)00
i00 π 00

Observe that π ◦ i = 0, hence π 00 ◦ i00 = 0.

By Exercise 4.B., we only have to prove that jD is surjective. Let θ ∈ D00 . Then i00 (θ) ∈ E 00 .
By reflexivity of E there is an x ∈ E such that jE (x) = i00 (θ). Applying π 00 at both sides we
find 0 = π 00 i00 (θ) = π 00 jE (x) = jE/D π(x). By injectivity of jE/D we have π(x) = 0 i.e. x ∈ D
i.e. x = i(d) for some d ∈ D. So i00 ◦ jD (d) = i00 (θ) i.e. jD (d) − θ ∈ Ker i00 . Now observe
that by the Hahn- Banach Theorem i0 is surjective so that i00 is injective and it follows that
θ = jD (d): SURJECTIVITY !

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 37

Exercise 4.C. Show that if E and F are linearly homeomorphic and E is reflexive then so is
F (K is spherically complete).

Lemma 23 Let K be spherically complete. Then c0 is not reflexive.

Proof We will construct a θ ∈ c000 that is not in jc0 (c0 ). Let, for n ∈ N, δn be the n-th
coordinate function c0 −→ K :
(ξ1 , ξ2 , . . .) 7−→ ξn .
∞
X
Then δn ∈ c00 . Let D = [δ1 , δ2 , . . .] ⊂ c00 . The function (ξ1 , ξ2 , . . .) 7−→ ξn is in c00 but not
n=1
in D, so we have D 6= c00 . Hence c00 /D 6= {0} and there is a nonzero ω ∈ (c00 /D)0 (Here we
used that K is spherically complete). Combining ω with the quotient map c00 −→ c00 /D:
ω
θ : c00 −→ c00 /D −→ K

we find a θ ∈ c000 , θ 6= 0 on D. We claim that this θ is not in Im jc0 . In fact, suppose we would
have an x ∈ c0 with jc0 (x) = θ i.e. θ(f ) = f (x) for all f ∈ c00 , say x = (ξ1 , ξ2 , . . .). By taking
for f = δn we find
ξn = δn (x) = θ(δn ) = 0,
implying x = 0 hence θ = 0, a contradiction.
This leads to the disappointing result:

Theorem 24 Let K be spherically complete. Then, the only reflexive spaces are the finite-
dimensional ones.

Proof Let E be infinite-dimensional. Then we can take a linearly independent sequence

x1 , x2 , . . .. If E were reflexive then so would [x1 , x2 , . . .] by Lemma 22. But the latter space
is of countable type so it is isomorphic to c0 by Corollary 18. By Exercise 4.C. then c0 would
be reflexive conflicting Lemma 23.
[BACKGROUND: The situation becomes more varied if we move to the locally convex theory:
here there are big classes of reflexive spaces over spherically complete K.]
Next, we investigate reflexivity for non-spherically complete scalar field K. The results will
be surprising.

Lemma 25 Let E be a spherically complete Banach space (no condition on K). Suppose D
is a closed subspace of E. Then E/D is also spherically complete.

Proof Let B(y1 , r1− ) ⊃ B(y2 , r2− ) ⊃ . . . be a nested sequence of ‘open’ balls in E/D. It suffices
to show that their intersection is non-empty. Let π : E −→ E/D be the quotient map. Choose
an x1 ∈ E with π(x1 ) = y1 . Then, from the way the norm on E/D was defined we infer that

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 38

π(B(x1 , r1− )) = B(y1 , r1− ). Then we can choose an x2 ∈ B(x1 , r1− ) with π(x2 ) = y2 and form
B(x2 , r2− ), whose image under π is exactly B(y2 , r2− ), etc. By spherical completeness of E
there is an x ∈ B(xn , rn− ). Then π(x) ∈ B(yn , rn− ).
T T
n n

Lemma 26 (ANTI-HAHN BANACH THEOREM) Let E be a Banach space over K.

If E is spherically complete and K is not then E 0 = {0}.

Proof Let f ∈ E 0 , f 6= 0. Decompose f as follows.

f -
E K (On E/Ker f the canonical norm.
A ρ is the unique map making the
A f1
ρ



diagram conmute.)
U
A

E/Ker f

By Lemma 25, the space E/Ker f is spherically complete. But ρ is a linear homeomorphism
between two one-dimensional spaces viz. E/Ker f and K. Then there is a constant c > 0
such that |ρ(x)| = ckxk for all x ∈ E/Ker f . Hence, ρ maps balls onto balls and so K must
be spherically complete as well. Contradiction.

Proposition 27 For any K, the space l∞ /c0 is spherically complete.

Proof Let π : l∞ −→ l∞ /c0 be the quotient map. For x = (ξ1 , ξ2 , . . .) ∈ l∞ we have

kπ(x)k = lim sup(|ξn |, |ξn+1 |, . . .)

n→∞

(verify this!) Like in the proof of Lemma 25 we can find, for given balls B(y1 , r1− ) ⊃
B(y2 , r2− ) ⊃ . . . in l∞ /c0 , balls B(x1 , r1− ) ⊃ B(x2 , r2− ) ⊃ . . . in l∞ such that π(xn ) = yn ,
π(B(xn , rn− )) = B(yn , rn− ) for each n. Now write

x1 = (x11 , x12 , . . .)
x2 = (x21 , x22 , . . .)
..
.

and take the diagonal sequence

a := (x11 , x22 , . . .).
Clearly |xnn | ≤ kxn k∞ for all n, so kak∞ ≤ supn kxn k∞ < ∞ and we see that a ∈ l∞ . We
claim that π(a) ∈ B(yn , rn− ) for each n (which will finish the proof), i.e., we have to show

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 39

kπ(a) − π(xn )k < rn . Now

kπ(a) − π(xn )k = lim sup(|xkk − xnk |, |xk+1,k+1 − xn,k+1 |, . . .)

k→∞
≤ sup(|xn+1,n+1 − xn,n+1 |, |xn+2,n+2 − xn,n+2 |, . . .)
≤ sup(kxn+1 − xn k∞ , kxn+2 − xn k∞ )
≤ sup(kxn+1 − xn k∞ , rn+1 , rn+2 , . . .)
< rn

and we are done.

Corollary 28 Let K be not spherically complete. Then (l∞ /c0 )0 = {0}. If f ∈ (l∞ )0 and
f = 0 on c0 then f is identically zero.

Proof Combine Lemma 26 and Proposition 27. For the second statement, observe that f = 0
on c0 implies that (with π : l∞ −→ l∞ /c0 the canonical map)

π(x) 7−→ f (x) (x ∈ l∞ )

defines an element of (l∞ /c0 )0 which has to be 0, so f = 0.

Proposition 29 Let K be not spherically complete. Then

∞
X
f : (ξ1 , ξ2 , . . .) 7−→ ξn ((ξ1 , ξ2 , . . .) ∈ c0 )
n=1

cannot be extended to an element of (l∞ )0 .

Proof With an ingenious trick. Suppose we could extend the given functional to a h ∈ (l∞ )0 .
Define the shift operator Ω : l∞ −→ l∞ by the formula

Ω(ξ1 , ξ2 , . . .) = (0, ξ1 , ξ2 , . . .).

Then Ω maps c0 into c0 and f ◦ Ω = f , so h ◦ Ω is also an extension of f . By Corollary 28 we

have h ◦ Ω = h. Now take any x ∈ l∞ , x = (ξ1 , ξ2 , . . .). Set

s = (ξ1 , ξ1 + ξ2 , . . .)

Then s ∈ l∞ and Ωs − s = x. Then h(x) = (h ◦ Ω − h)(s) = 0, a contradiction.

The idea of the proof of Proposition 29 can be generalized:

Proposition 30 Let K be not spherically complete. Let (a1 , a2 , . . .) ∈ l∞ . Then

∞
X
f : (ξ1 , ξ2 , . . .) 7−→ ξn an ((ξ1 , ξ2 , . . .) ∈ c0 )
n=1

can be extended to an element of (l∞ )0 if and only if (a1 , a2 , . . .) ∈ c0 .

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 40

Proof We prove that if (a1 , a2 , . . .) ∈
/ c0 then f cannot be extended.
There is an ε > 0 such that X := {n ∈ N : |an | ≥ ε} is infinite. Let n1 , n2 , . . . be an
enumeration of X. Define a continuous linear map T : l∞ −→ l∞ by

T (ξ1 , ξ2 , . . .) = (b1 , b2 , . . .),

where (
0 if n ∈
/X
bn =
a−1
ni ξi if n = ni .
We have for (ξ1 , ξ2 , . . .) ∈ c0
P P∞ P∞
(∗) (f ◦ T )(ξ1 , ξ2 , . . .) = f (b1 , b2 , . . .) = n∈X an bn = i=1 ani bni = i=1 ξi .

If f had an extension f ∈ (l∞ )0 then f ◦ T would be an extension of f ◦ T . But this is

impossible by the previous Proposition and (∗).

Theorem 31 (c0 and l∞ are reflexive) Let K be not spherically complete. For each x ∈ c0
define fx ∈ (l∞ )0 by
∞
X
fx (y) =< x, y >= ξn ηn (y ∈ l∞ )
n=1

(x = (ξ1 , ξ2 , . . .), y = (η1 , η2 , . . .)). Then x 7→ fx is an isometrical isomorphism. c0 ' (l∞ )0 .

Proof By Proposition 30, x 7→ fx is surjective. Obviously it is linear. We have for x ∈ c0 ,

y ∈ l∞ :
X 
|fx (y)| = | < x, y > | =  ξn ηn  ≤ max |ξn ||ηn |
 

≤ kxk∞ kyk∞

so kfx k ≤ kxk. But also, if we put y = en (n-th unit vector) we get

kfx k = ken kkfx k ≥ |fx (en )| = |ξn |

so, kfx k ≥ maxn |ξn | = kxk.

[BACKGROUND: The conclusion of Lemma 22 is not true if K is not spherically complete,
as one can construct non-reflexive subspaces of l∞ . The class of reflexive spaces over non-
spherically complete K is, so far, not yet well-described.]

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 41

Chapter 5

COMPACTOIDS

This little chapter treats a third important notion (next to ‘orthogonality’and ‘of countable
type’), namely that of a compactoid set. We do not have much time to explain it in the
lectures, so the only purpose of this chapter is to give you an idea about these notions, its
impact on the theory but without details and without proofs.
A subset X of a Banach space E (over K or over R or C) is called precompact if, for any ε > 0,
X can be covered by finitely many ε-balls. In other words, X is precompact if for every ε > 0
there is a finite set F ⊂ E such that

X ⊂ B(0, ε) + F

One proves by classical means that X is precompact if and only its closure X in E is compact.
(For this, it is necessary that E is complete).
In classical theory convex compact sets play an important role in Functional Analysis
(e.g. Alaoglu-Bourbaki Theorem, Krein-Milman Theorem, Choquet Theory). How are the
prospects in the non-archimedean case? Recall that a subset C of a K-vector space is absolutely
convex if 0 ∈ C and for x, y ∈ C, λ, µ ∈ K, |λ| ≤ 1 |µ| ≤ 1 we have λx + µy ∈ C. So, if K is
not locally compact the only absolutely convex compact subset of a Banach space is {0}. This
difficulty is similar to the separability problem on page 32, and we will solve it in a similar
way; this time by ‘convexifying’ the definition of precompactness.
Definition A subset X of a Banach space E over K is called compactoid if
for each ε > 0 there is a finite set F ⊂ E such that

X ⊂ B(0, ε) + aco F,

where ‘aco’ stands for ‘absolutely convex hull’, the smallest absolutely convex set containnig
the given set. In our case if F = {a1 , . . . , an } then aco F = {λ1 a1 +. . .+λn an : λi ∈ K, |λi | ≤ 1
for each i}.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 42

One may hope now that absolutely convex complete compactoids will take over the role played
by absolutely convex compact sets in classical analysis. The following results show that this
is the case to a certain degree.

5.1 List of properties of compactoids

Compactoids are bounded.
The absolutely convex hull of a compactoid is a compactoid.
The closure of a compactoid is a compactoid.
The continuous linear image of a compactoid is a compactoid.
If K is locally compact then a subset of a K-Banach space is a compactoid if and only if it is
precompact.
Those were easy properties. Now a few less trivial ones (we do not give proofs).
Let E be a K-Banach space with norm k · k.

– X ⊂ E is a compactoid and only if there exists a sequence x1 , x2 , . . . tending to 0 such

that X ⊂ aco {x1 , x2 , . . .}.

– If k k0 is a norm weaker than k k and A is a closed absolutely convex compactoid in E,

then the topologies induced by k k and k k0 coincide on A.

– Let A ⊂ E be a complete compactoid, absolutely convex. Let λ ∈ K, |λ| > 1 if the

valuation is dense. λ = 1 otherwise. Let t ∈ (0, 1) if K is not spherically complete. Let
t ∈ (0, 1] otherwise. Then there exists a t-orthogonal sequence e1 , e2 , . . . in λA such that

A ⊂ aco {e1 , e2 , . . .} ⊂ λ A

– Let X ⊂ E be bounded. Then X is a compactoid if and only if for every t ∈ (0, 1] each
t-orthogonal sequence in X tends to 0.

– Let X ⊂ E. Then X is a compactoid if and only if for each sequence x1 , x2 , . . . in X

p
lim n Vol(x1 , x2 , . . . , xn ) = 0
n→∞

(Here, Vol(x1 , . . . , xn ) := kx1 k dist(x2 , [x1 ]) dist(x3 , [x1 , x2 ]) . . . dist(xn , [x1 , . . . xn−1 ])).

– (Arzela-Ascoli) Let F ⊂ C(X) where X compact.

Then F is bounded and equicontinuous if and only if F is a compactoid.

– If A ⊂ E is an absolutely convex closed compactoid then (T A)e :=

T
λ(T A) is closed
|λ|>1
(where T ∈ L(E)).
(Not always T A is closed)

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 43

 – If T ∈ L(E) and A is a complete absolutely convex compactoid and z1 , z2 , . . . ∈
T A, zn → 0 then for each λ ∈ K, |λ| > 1 there exist x1 , x2 , . . . ∈ λA with T xn = zn for
each n and xn → 0 (The map T |A is ‘almost’ open).

An A ∈ L(E) is called compact operator if it maps the unit ball into a compactoid.
One can derive a theory a la Riesz for compact operators:

– non-zero elements of the spectrum are eigenvalues.

– 0 is the only possible accumulation point of the spectrum.

– Each eigenspace of a λ 6= 0 is finite-dimensional.

– Alternative of Fredholm.
..
.

One may think of alternative ways of ‘convexifying’ notions of compactness. The following
was introduced by Springer.
Let A ⊂ E be convex. We call A c-compact if the following holds. A is closed and for each
collection of closed convex sets {Ci : i ∈ I} in A such that Ci,1 ∩ . . . ∩ Ci,n 6= ∅ for any
T
i1 , . . . , in ∈ I we have Ci 6= ∅ (This translates a well-known property of closed subset of a
i
compact set).
However:

– If K is not spherically complete each c-compact set is either empty or a singleton set.

But:

– If K is spherically complete and A ⊂ E is convex and bounded then

A is a complete compactoid ⇔ A is c-compact.

P. Universidad Católica de Chile – Fac. Matemáticas – MECESUP PUC-0103 – FONDECYT 7020710 44