Lectures On

Some Fixed Point Theorems

Of Functional Analysis

By
F.F. Bonsall

Tata Institute Of Fundamental Research, Bombay

1962
 Lectures On
Some Fixed Point Theorems
Of Functional Analysis

By
F.F. Bonsall

Notes by
K.B. Vedak

No part of this book may be reproduced in any

form by print, microfilm or any other means with-
out written permission from the Tata Institute of
Fundamental Research, Colaba, Bombay 5

Tata Institute of Fundamental Research

Bombay
1962
Introduction

These lectures do not constitute a systematic account of fixed point the-

orems. I have said nothing about these theorems where the interest is
essentially topological, and in particular have nowhere introduced the
important concept of mapping degree. The lectures have been con-
cerned with the application of a variety of methods to both non-linear
(fixed point) problems and linear (eigenvalue) problems in infinite di-
mensional spaces. A wide choice of techniques is available for linear
problems, and I have usually chosen to use those that give something
more than existence theorems, or at least a promise of something more.
That is, I have been interested not merely in existence theorems, but also
in the construction of eigenvectors and eigenvalues. For this reason, I
have chosen elementary rather than elegant methods.
I would like to draw special attention to the Appendix in which I
give the solution due to B. V. Singbal of a problem that I raised in the
course of the lectures.
I am grateful to Miss K. B. Vedak for preparing these notes and
seeing to their publication.

Frank F. Bonsall

iii
Contents

1 The contraction mapping theorem 1

2 Fixed point theorems in normed linear spaces 13

3 The Schauder - Tychonoff theorem 31

4 Nonlinear mappings in cones 43

5 Linear mapping in cones 51

6 Self-adjoint linear operator in a Hilbert space 79

7 Simultaneous fixed points 95

8 A class of abstract semi-algebras 103

v
Chapter 1

The contraction mapping

theorem

Given a mapping T of a set E into itself, an element u of E is called a 1

fixed point of the mapping T if T u = u. Our problem is to find condi-
tions on T and E sufficient to ensure the existence of a fixed point of T
in E. We shall also be interested in uniqueness and in procedures for the
calculation of fixed points.

Definition 1.1. Let E be a nonempty set. A real valued function d de-

fined on E × E is called a distance function or metric in E if it satisfies
the following conditions

i) d(x, y) ≥ 0, x, yεE

ii) d(x, y) = 0 ⇐⇒ x = y

iii) d(x, y) = d(y, x)

iv) d(x, z) ≤ d(x, y) + d(y, z)

A nonempty set with a specified distance function is called a metric

space.

1
 2 The contraction mapping theorem

Example . Let X be a set and E denote a set of bounded real valued

functions defined on X. Let d be defined on E × E by
 
d( f, g) = sup | f (t) − g(t)| : tεX , f, gεE.

Then d is a metric on E called the uniform metric or uniform dis-

tance function.

2 Definition 1.2. A sequence {xn } in a metric space (E, d) is said to con-

verge to an element x of E if

lim d(xn , x) = 0
n→∞

A sequence xn of elements of a metric space (E, d) is called a Cauchy

sequence if given ǫ > 0, there exists N such that for p, q ≥ N, d(x p , xq ) <
ǫ.

A metric space (E, d) is said to be complete if every Cauchy se-

quence of its elements converges to an element of E. It is easily verified
that each sequence in a metric space converges to at most one on point,
and that every convergent sequence is a Cauchy sequence.
Example. The space CR [0, 1] of all continuous real valued functions on
the closed interval [0, 1] with the uniform distance is a complete metric
space. It is not complete in the metric d′ defined by
Z 1
′
d ( f, g) = | f (x) − g(x)|dx f, gεCR [0, 1].
0

Definition 1.3. A mapping T of a metric space E into itself is said to

satisfy a Lipschitz condition with Lipschitz constant K if

d(T x, T y) ≤ Kd(x, y) (x, yεE)

If this conditions is satisfied with a Lipschitz constant K such that 0 ≤

K < 1 then T is called a contraction mapping.

Theorem 1.1 (The contraction mapping theorem). Let T be a contrac-

tion mapping of a complete metric space E into itself. Then
 The contraction mapping theorem 3

3 i) T has a unique fixed point u in E

ii) If xo is an arbitrary point of E, and (xn ) is defined inductively by

xn+1 = T xn (n = 0, 1, 2, . . .),
then lim xn = u and
n→∞

Kn
d(xn , u) ≤ d(x1 , xo )
1−K
where K is a Lipschitz constant for T .

Proof. Let K be a Lipschitz constant for T with 0 ≤ K < 1. Let xo εE

and let xn be the sequence defined by

xn+1 = T xn (n = 0, 1, 2, . . .)

We have
d(xr+1 , xs+1 ) = d(T xr , T xs ) ≤ Kd(xr , xs ) (1)
and so
d(xr+1 , xr ) ≤ K r (x1 , xo ) (2)
Given p q, we have by (1) and (2),

d(x p , xq ) ≤ K q d(x p−q , xo )

 
q
≤ K d(x p−q , x p−q−1 ) + d(x p−q−1 , x p−q−2 ) + · · · + d(x1 , xo )
 
≤ K q K p−q−1 + K p−q−2 + · · · + K + 1 d(x1 , xo )
Kq
≤ d(x1 , xo ) (3)
1−K
since the right hand side tends to zero as q → ∞, it follows that (xn ) is a 4
Cauchy sequence, and since E is complete, (xn ) converges to an element
u of E. Since d(xn+1 , T u) ≤ Kd(xn , u) → 0 as n → ∞,

T u = lim xn+1 = u.
n→∞
 4 The contraction mapping theorem

Kn
d(u, xn ) ≤ d(u, x p ) + d(x p , xn ) ≤ d(u, x p ) + d(x1 , xo ) for n < p, by
1−K
(3). Letting p → ∞, we obtain
Kn
d(u, xn ) ≤ d(x1 , xo )
1−K


Example . As an example of the applications of the contraction map-

ping theorem, we prove Picard’s theorem on the existence of solution of
ordinary differential equation.
Let D denote an open set in R2 , (xo , yo )εD. Let f (x, y) be a real val-
ued function defined and continuous in D, and let it satisfy the Lipschitz
condition:
| f (x, y1 ) − f (x, y2 )| ≤ M|y1 − y2 | ((x, y1 ), (x, y2 )εD)
Then there exists a t > 0, and a function φ(x) continuous and differen-
tiable in [xo − t, xo + t] such that i)φ(xo ) = yo , (ii)y = φ(x) satisfies the
differential equation
dy
= f (x, y) for xε[xo − t, xo + t]
dx
We show first that there exists an ǫ > 0 andRa function φ(x) contin-
x
5 uous in [xo − t, xo + t] such that iii)φ(x) = yo + x f (t, φ(t))dt (xo − t ≤
o
x ≤ xo + t), and ǫ(x, φ(x))εD(xo − ≤ x ≤ xo + t). Then it follows from the
continuity of f (t, φ(t)) that φ(x) is in fact differentiable in [xo − t, xo + t]
and satisfies (i) and (ii).
Let U denote a closed disc of centre (xo , yo ) with positive radius and
contained in the open set D, and let m denote the least upper bound of the
continuous function | f | on the compact set U. We now choose t, δ such
that 0 < t < M −1 , the rectangle |x−xo | ≤ δ is contained in U, and mt < δ.
Let E denote the set of all continuous functions mapping [xo − t, xo + t]
into [yo − δ, yo + δ]. With respect to the uniform distance function E is
a closed subset of the complete metric space CR [xo − t, xo + t] and is
therefore complete. We define a mapping T φ = ψ for φ ∈ E by
Z x
ψ(x) = yo + f (t, φ(t))dt
xo
The contraction mapping theorem 5

Clearly ψ(x) is continuous in [xo − t, xo + t]. Also |ψ(x) − yo | ≤

m|x − xo | ≤ mt < δ whenever |x − xo | ≤ t. Thus T maps E into it self.
Finally, T is a contraction mapping, for if φi ∈ E, ψi = T φ)i(i = 1, 2),
then
Z x 
|ψ1 (x) − ψ2 (x)| = | f (t′ , φ1 (t′ )) − f (t′ , φ2 (t′ )) dt′ |
xo
≤ |x − xo |M sup |φ1 (t′ )| − φ2 (t′ )| (xo − t ≤ t ≤ xo + t)|
≤ tMd(φ1 , φ2 )

Hence 6
d(ψ1 , ψ2 ) ≤ tMd(φ1 , φ2 )
As tM < 1, this proves that T is a contraction mapping. By the contrac-
tion mapping theorem, there exists φεE with T φ = φ i.e., with
Z x
φ(x) = yo + f (t, φ(t))dt
xo

This complete the proof of Picard’s theorem.

A similar method may be applied to prove the existence of solutions
of systems of ordinary differential equations of the form
dyi
= fi (x, y1 , . . . , yn )(i = 1, 2, . . . , n)
dx
with given initial conditions. Instead of considering real valued func-
tions defined on [xo − t, xo + t], one considers vector valued functions
mapping [xo − t, xo + t] into Rn .
In the following theorem we are concerned with the continuity of
the fixed point.

Theorem 1.2. Let E be a complete metric space, and let T and T n (n =

1, 2, . . .) be contraction mappings of E into itself with the same Lipschitz
constant K < 1, and with fixed points u and un respectively. Suppose
that lim T n x = T x for every xεE. Then lim un = u. By the inequality
n→∞ n→∞
in Theorem 1.1, we have for each r = 1, 2, . . .,
Kn
d(ur , T rn xo ) ≤ d(T r xo , xo ), xo εE
1−K
 6 The contraction mapping theorem

setting n = 0 and xo = u, we have 7

1 1
d(ur , u) ≤ d(T r u, u) = d(T r u, T u)
1−K −1
But d(T r u, T u) → 0 as r → ∞. Hence
lim d(ur , u) = 0
r→∞

Example . In the notation of the last example, suppose that yn is a real

sequence converging to yo and let T n be the mapping defined on E by
Z x
(T n φ)(x) = yn + f (t, φ(t))dt
xo

Then |(T n φ)(x) − yo | ≤ |yn − yo | + mε < δ for n sufficiently large

i.e. T n map E into itself for n sufficiently large. Also the mapping
T n , T have the same Lipschitz constant εM < 1. Obviously for each
φεE, lim T n φ = T φ. Hence if φn is the unique fixed point of T n (n =
n→∞
1, 2, . . .) then lim φn = φ. In other words, if φn is the solution of the
n→∞
differential equation
dy
= f (x, y)
dx
in [xo − t, xo + t] with the initial condition φn (xo ) = yn , then φn converges
uniformly to the solution φ with φ(xo ) = yo .
Remark. The contraction mapping theorem is the simplest of the fixed
point theorems that we shall consider. It is concerned with mappings of a
8 complete metric space into itself and in this respect is very general. The
theorem is also satisfactory in that the fixed point is always unique and is
obtained by an explicit calculation. Its disadvantage is that the condition
that the mapping be a contraction is a somewhat severe restriction. In the
rest of this chapter we shall obtain certain extension of the contraction
mapping theorem in which the conclusion is obtained under modified
conditions.

Definition 1.4. A mapping T of a metric space E into a metric space E ′

is said to be continuous if for every convergent sequence (xn ) of E,
lim T xn = T ( lim xn ).
n→∞ n→∞
The contraction mapping theorem 7

Theorem 1.3. Let T be a continuous mapping of a complete metric

space E into itself such that T k is a contraction mapping of E for some
positive integer k. Then T has a unique fixed point in E.

Proof. T k has a unique fixed point u in E and u = lim (T k )n xo , xo εE

n→∞
arbitrary. 

Also lim (T k )n T xo = u. Hence

n→∞

u = lim (T k )n T xo = lim T (T k )n xo
n→∞ n→∞
k n
= T lim (T ) xo (by the continuity of ) T
n→∞
= T u.

The uniqueness of the fixed point of T is obvious, since each fixed

point of T is also a fixed point of T k .

Example. We consider the non-linear integral equation 9

Z x
f (x) = λ K(x, y, f (y))dy + g(x) (1)
a

where g is continuous in [a, b] and K(x, y, z) is continuous in the region

[a, b] × [a, b] × R and satisfies the Lipschitz condition.

|K(x, y, z1 ) − K(x, y, z2 )| ≤ M|z1 − z2 |.

(The classical Volterra equation is obtained by taking K(x, y, z) =

H(x, y).z, with H continuous in [a, b] × [a, b]). Let E = C K [a, b] and T
be the mapping of E into itself given by
Z x
(T f )(x) = λ K(x, y, f (u))dy + g(x) ( f εE, a ≤ x ≤ b).
a

Given f1 , f2 in E it is easy to prove by induction on n that

1 n n
|(T n f1 )(x) − (T n f2 )(x)| ≤ |λ| M d( f1 , f2 )(x − a)n , (a ≤ x ≤ b)
n!
 8 The contraction mapping theorem

Then
1 n n
d(T n f1 , T n f2 ) ≤ |λ| M (b − a)n d( f1 , f2 )
n!
This proves that all T n and in particular T , are continuous and, for
1
n sufficiently large |λ|n M n (b − a)n < 1, so that T n is a contraction
n!
mapping for n large. Applying the theorem, we have a unique f εE with
T f = f which is the required unique solution of the equation (1).

10 Definition 1.5. Let (E, d) be a metric space and ε > 0. A finite sequence
xo , x1 , . . . , xn of points of E is called an ε - chain joining x0 and xn if

d(xi−1 , xi ) < ε (i = 1, 2, . . . , n)

The metric space (E, d) is said to be ε -chainable if for each pair

(x, y) of its points there exists an ε- chain joining x and y.

Theorem 1.4 (Edelstein). Let T be a mapping of a complete ε- chainable

metric space (E, d) into itself, and suppose that there is a real number K
with 0 ≤ K < 1 such that

d(x, y) < ε ⇒ d(T x, T y) ≤ Kd(x, y)

Then T has a unique fixed point u in E, and u = lim T n xo where xo

n→∞
is an arbitrary element of E.

Proof. (E, d) being ε chainable we define for x, yεE,

X
n
dε (x, y) = inf d(xi−1 , xi )
i=1

where the infimum is taken over all ε - chains x0 , . . . , xn joining x0 = x

and xn = y. Then d is a distance function on E satisfying

i) d(x, y) ≤ dε (x, y)

ii) d(x, y) = dε (x, y) for d(x, y) < ε

 The contraction mapping theorem 9

From (ii) it follows that a sequence (xn ), xn εE is a Cauchy sequence

11 with respect to dε if and only if it is a Cauchy sequence with respect
to d and is convergent with respect to dε if and only if it converges
with respect with respect to d. Hence (E, d) being complete, (E, dε ) is
also a complete metric space. Moreover T is a contraction mapping with
respect to dε . Given x, yεE, and any ε-chain xo , . . . , xn with xo = x, xn =
y, we have
d(xi−1 , xi ) < ε (i = 1, 2, . . . , n),
so that
d(T xi−1 , T xi ) ≤ Kd(xi−1 , xi ) < ε (i = 1, 2, . . . , n)
Hence T xo , . . . , T xn is an ε- chain joining T x and T y and

X
n X
n
dε (T x, T y) ≤ d(T ci−1 , T xi ) ≤ K d(xi−1 , xi )
i=1 i=1

xo , . . . , xn being an arbitrary ε- chain, we have

dε (T x, T y) ≤ Kdε (x, y)

and T has a unique fixed point uεE given by

lim dε (T n x0 , u) = 0 for xo εE arbitrary (1)

n→∞

But in view of the observations made in the beginning of this proof, (1)
implies that
lim d(T n xo , u) = 0
n→∞


Example . Let E be a connected compact subset of a domain D in the

complex plane. Let f be a complex holomorphic function in D which 12
maps E into itself and satisfies | f ′ (z)| < 1 (z, εE). Then there is a unique
point z in E with f (z) = z. Since f ′ is continuous in the compact set E,
there is a construct K with 0 < K < 1 such that | f ′ (z)| < K εE). For
each point ωεE there exists ρω > 0 such that f (x) is holomorphic in the
disc S (ω, 2ρω ) of center ω and radius 2ρω and satisfies | f ′ (z)| < K there.
 10 The contraction mapping theorem

E being compact, we can choose ω1 , . . . , ωn εE such that E is covered

by
S (ω1 , 2ρω1 ), . . . , S (ωn , 2ρωn )
Let ε = min{ρωi , i = 1, 2, . . . , n}. If z, z′ εE and |z − z′ | < ε then
z, z′ εS (ωi , 2ρω1 ) for some i and so
Z z
′
| f (z) − f (z )| = | f ′ (ω)dω| ≤ K|z − z′ |.
z′
This proves that Theorem 1.4 is applicable to the mapping z → f (z)
and we have a unique fixed point.

Definition 1.6. A mapping T of a metric space E into itself is said to be

contractive if
d(T x, T y) < d(x, y) (x , y, x, yεE)
and is said to be ε-contractive if
0 < d(x, y) < ε ⇒ d(T x, T y) < d(x, y)
Remark . A contractive mapping of a complete metric space into itself
13 need not have a fixed point. e.g. let E = {x/x ≥ 1} with the usual
1
distance d(x, y) = |x − y|,let T : E → E be given by T x = x + .
x
Theorem 1.5 (Edelsten). Let T be an ε-contractive mapping of a metric
space E into itself, and let xo be a point of E such that the sequence
(T n xo ) has a subsequence convergent to a point u of E. Then u is a
periodic point of T , i.e. there is a positive integer k such that
T ku = u
Proof. Let (ni ) be a strictly increasing sequence of positive integers such
that lim T ni xo = u and let xi = T ni xo . There exists N such that d(xi , u) <
i→∞
ε/4 for i ≥ N. Choose any i ≥ N and let k = ni+1 − ni . Then
d(xi+1 , T k u) = d(T k xi , T k u) ≤ d(xi , u) < ε/4
and
d(T k u, u) ≤ d(T k u, xi+1 ) + d(xi+1 , u) < ε/2

The contraction mapping theorem 11

Suppose that v = T k u , u. Then T being ε - contractive,

d(T u, T v)
d(T u, T v) < d(u, v) or < 1.
d(u, v)
T x, T y
The function (x, y) → is continuous at (u, v). So there exist
d(u, v)
δ, K > 0 with 0 < K < 1 such that d(x, u) < δ, d(y, v) < δ implies that
d(T x, T y) < Kd(x, y). As lim T K xr = T K u = v, there exists N ′ ≥ N
r→∞
such that d(xr , u) < δ, d(T xr , v) < δ for r ≥ N ′ and so 14

d(T xr , T T K xr ) < Kd(xr , T K xr ) (1)

d(xr , T K xr ) ≤ d(xr , u) + d(u, T K u) + d(T K u, T K xr )
ε ε ε
< + + = ε for r ≥ N ′ > N (2)
4 2 4
From (1) and (2)
d(T xr , T T K xr ) < Kd(xr , T K xr ) < ε for r ≥ N ′
and so T being ε-contractive,
d(T p xr , T p T k xr ) < Kd(xr , T K xr ) for n ≥ N ′ , p > 0 (3)
Setting p = nr+1 − nr in (3)
d(xr+1 , T k xr+1 ) < Kd(xr , T k xr ) for any r ≥ N ′
Hence d(xs , T k xs ) < K s−r d(xr , T k xr ) < K s−r ε and d(u, v) < d(u, xs ) +
d(xs , T k xs ) + d(T k xs , v) → 0 as s → ∞ This contradicts the assumption
that d(u, v) > 0. Thus u = v = T k u.

Theorem 1.6 (Edelstein ). Let T be a contractive mapping of a metric

space E into itself, and let xo be a point of E such that the sequence T n xo
has a subsequence convergent to a point u of E. Then u is a fixed point
of T and is unique.

Proof. By Theorem 1.5, there exists an integer k > 0 such that T k u = 15

u. Suppose that v = T u ,, u. Then T k u = u, T k v = v and d(u, v) =
d(T k u, T k v) < d(u, v), since T is contractive. As this is impossible,
u = v is a fixed point. The uniqueness is also immediate. 
12 The contraction mapping theorem

Corollary . If T is a contractive mapping of a metric space E into a

compact subset of E, then T has a unique fixed point u in E and u =
lim T n xo where xo is an arbitrary point of E.
n→∞
Chapter 2

Fixed point theorems in

normed linear spaces

In Chapter 1, we proved fixed point theorems in metric spaces without 16

any algebraic structure. We now consider spaces with a linear structure
but non-linear mappings in them. In this chapter we restrict our attention
to normed spaces, but our main result will be extended to general locally
convex spaces in Chapter 3

Definition 2.1. Let E be a vector space over. A mapping of E into R is

called a norm on E if it satisfies the following axioms.

i) p(x) ≥ 0 (x ∈ E)

ii) p(x) = 0 if and only if x = 0

iii) p(x + y) ≤ p(x) + p(y) (x, y ∈ E).

A vector space E with a specified norm on it called a normed space.

The norm of an element x ∈ E will usually be denoted by ||x||. A normed
space is a metric space with the metric d(x, y) = x − y (x, yεE) and the
corresponding metric topology is called the normed topology. A normed
linear space complete in the metric defined by the norm is called a Ba-
nach space. We now recall some definitions and well known properties

13
 14 Fixed point theorems in normed linear spaces

of linear spaces. Two norms p1 and p2 on a vector space E are said to

be equivalent if there exist positive constants k, k′ such that

p1 (x) ≤ kp2 (x), p2 (x) ≤ k′ p1 (x) (x ∈ E)

17 Two norms are equivalent if and only if they define the same topology.

Definition 2.2. A mapping f of a vector space E into R is called a linear

functional on E if it satisfies

i) f (x + y) = f (x) + f (y) (x, y ∈ E)

ii) f (αx) = α f (x) (x ∈ E, α ∈ R).

A mapping p : E → R is called a sublinear functional if

i)′ p(x + y) ≤ p(x) + p(y) (x, y ∈ E)

ii)′ p(αx) = αp(x) (x ∈ E, α ≥ 0).

Hahn- Banach Theorem. Let E0 be a subspace of a vector space E over

R; let p be a sublinear functional on E and let fo be a linear functional
on Eo that satisfies
fo (x) ≤ p(x) (x ∈ E0 ).

Then there exists a linear functional f on E that satisfies

i) f (x) ≤ p(x) (x ∈ E),

ii) f (x) = fo (x) (x ∈ E0 ).

[For the proof refer to Dunford and Schwartz ([14], p. 62) or Day [13,
p.9]].

Corollary. Given a sublinear functional on E and xo ∈ E, there exists a

linear functional f such that

f (x0 ) = p(x0 ), f (x) ≤ p(x) (x ∈ E).

 Fixed point theorems in normed linear spaces 15

In particular, a norm being a sublinear functional, given a point x0

18 of a normed space E, there exists a linear functional f on E such that

| f (x)| ≤ ||x|| (x ∈ E) and f (xo ) = ||xo ||

Definition . A norm p on a vector space E said to be strictly convex if

p(x + y) = p(x) + p(y) only when x and y are linearly dependent.

Theorem 2.1 (Clarkson ). If a normed space E has a countable every-

where dense subset, then there exists a strictly convex norm on E equiv-
alent to the given norm.

Proof. Let S denote the surface of the unit ball in E,

S = {x : ||x|| = 1}

Then there exists a countable set (xn ) of points of S that is dense in S .

For each n, there exists a linear functional fn on E such that

fn (xn ) = ||xn || = 1 and | fn (x)| ≤ ||x|| (x ∈ E). 

If x , 0, then fn (x) , 0 for some n. For, by homogeneity, it is

enough to consider x with ||x|| = 1, and for such x there exists n with
1
||x − xn || < . But then
2
fn (x) = fn (xn ) + fn (x − xn ) ≥ 1 − fn (x − xn )
1
≥ 1 − ||x − xn || >
2
( ) 12
P
∞
−n 2
We now take p(x) = ||x|| + 2 ( fn (x)) . It is easily verified 19
n=1
that p is a norm on E and that

||x|| ≤ p(x) ≤ 2||x||.

Finally p is strictly convex. To see this, suppose that

p(x + y) = p(x) + p(y),

 16 Fixed point theorems in normed linear spaces

and write ξn = fn (x), yn = fn (y). Then

∞  12  ∞  12  ∞  21

X −n
 

 

X 

 

X 

2 −n 2 −n 2 

 2 (ξ n + ηn ) 
 = 
 s ξ n 
 + 
 2 ηn 

     
n=1 n=1 n=1

and we have the case of equality in Minkowsiki’s inequality. It follows

that the sequence (ξn ) and (ηn ) are linearly dependent. Thus there exist
λ, ν, not both zero, such that

λξn + µηn = 0 (n = 1, 2, . . .)

But this implies that

fn (λx + µy) = 0 (n = 1, 2, . . .),

and so λx + µy = 0. This completes the proof.

Lemma 2.1. Let K be a compact convex subset of a normed space E

with a strictly convex norm. Then to each point x of E corresponds a
unique point Px of K at K at minimum distance from x, i.e., with

||x − Px|| = inf{||x − y|| : yεK}

20 and the mapping x → Px is continuous in E.

Proof. Let x ∈ E, and let the function f be defined on K by f (y) =

||x − y||. Then f is a continuous mapping of the compact set K into R
and therefore attains its minimum at a point z say of K

||x − z|| = inf{||x − y|| : y ∈ K}. 

Evidently for x ∈ K, z = x is uniquely determined. If x < K, suppose

that z′ is such that
0 , ||x − z|| = ||x − z′ || (1)
1
since K is convex, y = (z + z′ ) ∈ K and therefore
2
1 1
||x − y|| ≥ ||x − z|| = ||x − z′ || = ||x − z|| + ||x − z′ ||
2 2
Fixed point theorems in normed linear spaces 17

But
1 1
(x − y) = (x − z) + (x − z′ ),
2 2
1 1
so that ||x − y|| ≤ ||x − z || + ||x − z′ ||
′
2 2
1 1 1 1
Hence ||x − y|| = || (x − z) + (x − z′ )|| = || (x − z)|| + || (x − z′ )||
2 2 2 2
As the norm is strictly convex,

λ(x − z) + µ(x − z′ ) = 0

for λ, µ not both zero. By (1), |λ| = |µ| and so x − z = ±(x − z′ ). If

z + z′
x − z = −(x − z′ ), then x = ∈ K, which is not true. Hence
′ ′
2
x − z = x − z or z = z . This proves that the mapping x → Px = z is
uniquely on E. Given x, x′ ∈ E,

||x − Px|| ≤ ||x − Px′ || ≤ ||x − x′ || + ||x′ − Px′ ||,

and similarly ||x′ − Px′ || ≤ ||x − x′ || + ||x − Px||. So 21

| ||x − Px|| − ||x′ − Px′ || | ≤ ||x − x′ || (2)

Let xn ∈ E(n = 1, 2, . . .) converge to x ∈ E. Then the sequence Pxn

in the compact metric space K has a subsequence Pxnk converging to
y ∈ K. Then
lim ||xnk − Pxnk || = ||x − y|| (3)
k→∞

By (2)| ||xn − Pxn || − ||x − Px|| | ≤ ||xn − x|| → 0 as n → ∞, and

so ||x − y|| = ||x − Px||. Hence Px = yi.e. lim Pxnk = Px. Thus if (xn )
k→∞
converges to x, (Pxn ) has a subsequence converging to Px ans so every
subsequence of (Pxn ) has a subsequence converging to Px. Therefore
(Pxn ) converges to Px and P is continuous. 

Definition 2.3. The mapping P of Lemma 2.1 is called the metric pro-
jection onto K.
 18 Fixed point theorems in normed linear spaces

Definition 2.4. A subset A of a normed space is said to be bounded if

there exists a constant M such that ||x|| ≤ M (x ∈ A).

We now state without proof three properties of finite dimensional

normed spaces.
Lemma 2.2. Every finite dimensional normed space is complete.
Lemma 2.3. Every bounded closed subset of a finite dimensional norm-
ed space is compact.
Lemma 2.4 (Brouwer fixed point theorem). Let K be a non-empty com-
22 pact convex subset of a finite dimensional normed space, and let T be a
continuous mapping of K into itself. Then T has a fixed point in K.
The proofs of the first two of these Lemmas are elementary. (Refer
to Dunford and Schwartz [14, p. 244-245].) The Brouwer fixed point
theorem on the other hand is far from trivial. For a proof using some
elements of algebraic topology refer to P. Alexandroff and H. Hopf ([1],
p.376-378). A proof of a more analytical kind is given by Dunford ans
Schwartz ([14], p.467).

Theorem 2.2 (Schauder). Let K be a non-empty closed convex subset

of a normed space. Let T be a continuous mapping of K into a cumpact
subset of K. Then T has fixed point in K.

Proof. Let E denote the normed space and let T K ⊂ A, a compact subset
of K. A is contained in a closed convex bounded subset of E.
T (B ∩ K) ⊂ T (K) ⊂ A ⊂ B
so T (B ∩ K) is contained in a compact subset of B, K and there is no
loss of generality in supposing that K is bounded. If Ao is a countable
dense subset of the compact metric space A, then the set of all rational
linear combinations of elements of Ao is a countable dense subset of the
closed linear subspace Eo spanned by Ao and A ⊂ E0 . Then T (K ∩
E0 ) ⊂ T (K) ⊂ A, a compact subset of E0 , and K ∩ E0 is closed and
convex. Hence without loss of generality we may assume that K is a
bounded closed convex subset of a separable normed space E with a
strictly convex norm (Theorem 2.1). 
Fixed point theorems in normed linear spaces 19

1
Given a positive integer n, there exists a -net T x1 , . . . , T xm say in 23
n
T K, so that
1
min ||T x − T xk || < (x ∈ K) (1)
1≤k≤n n
Let En denote the linear hull of T x1 , . . . , T xm . Kn = K ∩ En is a
closed bounded subset of En and therefore compact (Lemma 2.3). Since
the norm is strictly convex, the metric projection Pn of E onto the con-
vex compact subset Kn exists. T n = Pn T is a continuous mapping of the
non-empty convex compact subset Kn into itself, and therefore by the
Brouwer fixed point theorem, it has a fixed point un εKn ,

T n un = un (2)

By (1), since T xk ∈ Kn (k = 1, 2, . . . , m), we have

1
||T x − T n x|| < (3)
n
The sequence {T un } of T K has a subsequence T unk converging to a
point v ∈ K. By (2) and (3), ||unk − v|| = ||T nk unk − v|| ≤ ||T nk unk − T unk || +
1
||T unk − v|| < + ||T unk − v||. Therefore, lim unk = u, and by continuity
n k→∞
of T, lim T unk = T v or T v = v.
k→∞

Example . Suppose that a function f (x, y) of two real variables is con-

tinuous on a neighbourhood of (xo , yo ). Then we can choose ε > 0 such
that f is continuous in the rectangle

|x − xo | ≤ ε, |y − yo | ≤ mε

and satisfies there the inequality 24

| f (x, y)| ≤ m.

Let E denote the Banach space CR [xo − ε, xo + ε], which is a Banach

space with the uniform norm
 
||ϕ|| = sup |Φ(t)| : |t − xo | ≤ ε
 20 Fixed point theorems in normed linear spaces

Let K be the subset of E consisting of all continuous mappings of

[xo − ε, xo + ε] into [yo − mε, yo + mε]. Then K is a bounded closed
convex subset of E. Let T be the mapping defined on K by
Z x
(T φ)(x) = yo + f (t, φ(t))dt (|x − xo | ≤ ε)
xo

Then T K ⊂ K. Also since

  Z 
(T φ)(x) − (T φ)(x′ ) ≤ 
x
   f (t, (t))dt ≤ m|x − x′ | (φ ∈ K),
x′

T K is an equicontinuous set. Since also T K is bounded, T K is contained

in a compact set by the Ascoli - Arzela theorem. Therefore, by Theorem
2.2, T has a fixed point φ in K i.e.,
Z x
φ(x) = yo + f (t, φ(t)) dt (|x − xo | ≤ ε).
xo

Then φ is differentiable in [xo − ε, xo + ε] and provides a solution

y = φ(x) there of the differential equation
dy
= f (x, y)
dx
25 with φ(xo ) = yo . This is Peano’s theorem. As a particular case of
Schauder’s theorem, we have

Theorem 2.3. Let K be a non-empty compact convex subset of a normed

space, and let T be a continuous mapping of K into itself. Then T has a
fixed point in K.

Remark . Theorem 2.2 and 2.3 are almost equivalent, in the sense that
Theorem 2.2, with the additional hypothesis that K be complete, follows
from Theorem 2.3. For, if K is a complete convex set and T K is con-
tained in a compact subset A of K, then the closed convex hull of A is a
compact convex subset Ko of K, and T K0 ⊂ K0 .

Definition 2.5. A mapping T which is continuous and maps each boun-

ded set into a compact set is said to be completely continuous.
Fixed point theorems in normed linear spaces 21

Theorem 2.4. Let T be a completely continuous mapping of a normed

space E into itself and let T E be bounded. Then T has a fixed point.

Proof. Let K be the closed convex hull of T E. Then K is bounded and

so T K is contained in a compact subset of K. By Theorem 2.2, T has a
fixed point in K. 

The Theorem 2.4 implies Theorem 2.3 is seen as follows. Let K

be a compact convex set and let T be continuous mapping of K into
itself. There is no loss of generality is supposing that the norm in E
is strictly convex. Let P be the metric projection of E onto K, and let 26
T̃ = T P. Then T̃ satisfies the conditions of theorem 24, and so there
exists u in E with T u = u. Since T maps E into K, we have u ∈ K and
so Pu = uT u = T Ru = u.

Lemma 2.5. Let K be a non-empty complete convex subset of a normed

space E, let A be a continuous mapping of K into a compact subset of
E, and let F be a mapping of K × K into K such that

(i) ||F(x, y) − F(x, y)|| ≤ k||y − y′ || (x, y, y′ ∈ K), where k is a constant

with 0 < k < 1,

(ii) ||F(x, y) − F(x′ , y)|| ≤ ||Ax − Ax′ || (x, x′ , y ∈ K). Then there exists
a point u in K with
F(u, u) = u.

Proof. For each fixed x, the mapping y → F(x, y) is a contraction map-

ping of the complete metric space K into itself, and it therefore has a
unique fixed point in K which we denote by T x,

T x = F(x, T x) (x ∈ K).
We have ||T x − T x || = ||F(x, T x) − F(x′ , T x′ )||
′

≤ ||F(x, T x) − F(x′ , T x)||

+ ||F(x′ , T x) − F(x′ , T x′ )||
≤ ||Ax − Ax′ || + k||T x − T x′ || 
 22 Fixed point theorems in normed linear spaces

1
Therefore ||T x− T x′ || ≤ ||Ax− Ax′ ||, (1) which shows that T k is
1−k
continuous and that T K ⊂ K is precompact since AK is compact, since
K is complete, T K ⊂ K is compact. By the Schander theorem, T has
fixed point u in K,
T u = u.
27 But then
F(u, u) = F(u, T u) = T u = u


Theorem 2.5 (Kranoselsku ). Let K be a non-empty complete convex

subset of a normed space E, let A be a continuous mapping of K into a
compact subset of E, let B map K and satisfy a Lipschitz condition

||Bx − Bx′ || ≤ k||x − x′ || (x, x′ ∈ k)

with 0 < k < 1 and let Ax + By ∈ K for all x, y in K. Then there is a

point u ∈ K with
Au + Bu = u

Proof. Take F(x, y) = Ax + By and apply Lemma 2.5. 

Corollary . Let K be a non-empty complete convex subset of a normed

space, let A be a continuous of K into a compact subset of K, let B map
K into itself ans satisfy the Lipschitz condition

||Bx − Bx′ || ≤ ||x − x′ || (x, x′ εK),

and let 0 < α < 1. Then there exists a point u ∈ K with

αAu + (1 − α)Bu = u

In general, under the condition of Schauder’s theorem, we have no

method for the calculation of a fixed point of a mapping. However there
is a special case in which this can be done using a method due to Kras-
noselsku.
Fixed point theorems in normed linear spaces 23

Definition 2.6. A norm p is uniformly convex if it satisfies

p(xn ) = p(yn ) = 1 (n = 1, 2, . . .), lim p(xn + yn )

n→∞
= 2 =⇒ lim p(xn − yn ) = 0.
n→∞

28

Lemma . Let p be a uniformly convex norm, and let εM be positive

constants. Then there exists a constant δ with 0 < δ < 1 such that

p(x) ≤ M, p(y) ≤ M, p(x − y) ≥ ε ⇒ p(x + y) ≤ 2δ max(p(x), p(y)).

Proof. For all x, y, we have

!
1 1 1
p (x + y) ≤ p(x) + p(y) ≤ max(p(x), p(y)). (1)
2 2 2

If there is no constant δ with the stated properties, there exist se-

quences (xn ), (yn ) with p(xn ) ≤ M, p(yn ) ≤ M,

p(xn − yn ) ≥ ε, (2)

and ! !
1 1
p (xn + yn ) > 1 − max(p(xn ), p(yn )). (3)
2 n
Let αn = p(xn ), βn = p(yn ), γn = max(αn , βn ). By (1) and (2),
1
γn ≥ , (4)
2
and so, by (1) and (3)
!
1 1
lim p (xn + yn ) = 1. (5)
n→∞ γn 2

It follows from (1) and (5), that

αn + βn
lim = 1. (6)
n→∞ 2γn
 24 Fixed point theorems in normed linear spaces

Since (γn ) is bounded, there exists a convergent sequence (γnk ), and

by (4)
lim γnk = γ ≥ εγ2 (7)
k→∞

lim (γnk − αnk ) + (γnk − βnk ) = 0,

k→∞

29 and, since each bracket is non-negative, each tends to zero. Therefore

lim αnk = lim βnα = γ (8)

k→∞ k→∞

By discarding some terms of the subsequence if necessary, we may

suppose that αnk ≥ 0 and βnk ≥ 0 for all k. Since
 ! !
 p 1 1 1 1 
 xnk + ynk − p xnk + ynk 
αnk βnk γnk γnk
! !
1 1 1 1
≤p − xnk + − ynk
αn γnk βnk γnk
( k )
1 1 1 1
≤M − + − ,
αnk γnk βnk γnk

it follows from (5), (7), (8), that

!
1 1
lim p xn + yn = 2.
k→∞ αnk k βnk k

Therefore, !
1 1
lim p xn − yn = 0.
k→∞ αnk k βnk k
and so
1
lim p(xnk − ynk ) = 0,
k→∞ γnk

which contradicts (2). 

30 Theorem 2.6 (Krashoselsku). Let K be a bounded closed convex set in

a Banach space E with a uniformly convex norm. Let T be a mapping
of K into a compact subset of K that satisfies a Lipschitz condition with
Fixed point theorems in normed linear spaces 25

Lipschitz constant 1, and let xo be an arbitrary point of K. Then the

sequence defined by
1
xn+1 = (xn + T xn ) (n = 0, 1, 2, . . .)
2
converges to a fixed point of T in K.

Proof. By Schauder’s theorem, there is a nonempty set F of fixed points

of T in K. We prove first that

||xn+1 − y|| ≤ ||xn − y|| (y ∈ F, n = 0, 1, 2, . . .)

In fact if y = T y , then

1 1
||xn+1 − y|| = || (xn + T xn ) − (y + T y)||
2 2
1 1
= || (xn − y) + (T xn − T y)||
2 2
1 1
≤ ||xn − y|| + ||T xn − T y||
2 2
≤ ||xn − y||

which is (1). 

Suppose that there exist an ε > 0 and N, such that

||xn − T xn || ≥ ε for all n ≥ N (2)

Then ||xn − y − (T xn − T y )|| ≥ ε for all n ≥ N, y ∈ F.

Also ||T xn − T y || ≤ ||xn − y|| ≤ ||xo − y||, by (1).
Since the norm is uniformly convex, this implies that there exists a
constant δ, 0 < δ < 1, such that
1 1
||xn+1 − y|| = || (xn + T xn ) − (y + T y)||
2 2
1 1
= || (xn − y) + (T xn − T y)||
2 2
≤ max {||xn − y||, ||T xn − T y||}
 26 Fixed point theorems in normed linear spaces

≤ ||xn − y|| for n ≥ N.

Therefore lim xn = y where T y = y. 31

n→∞
If there does not exist an ε > 0 for which (2) holds, there exists a
sequence nk of integers such that lim (xnk − T xnk ) = 0, and such that
k→∞
(T xnk ) converges. But this implies that lim xnk = u = lim T xnk and so
k→∞ k→∞
T u = u.
Hence ||xn+1 − u|| ≤ ||xn − u||, by (1). Since lim ||xnk − u|| = 0, we
k
have lim ||xn − u|| = 0 and the theorem is proved.
n→∞
The following theorem was proved by Altam by means of the con-
cept of ‘degree of a mapping’, but we can easily deduce it from schan-
der’s theorem.

Theorem 2.7 (Altman). Let E be a normed space, let Q be the closed

ball of radius r > 0,
Q = {x : ||x|| ≤ r}
and let T be a continuous mapping of Q into a compact subset of E such
that
||T x − x||2 ≥ ||T x||2 − ||x||2 (||x|| = r)

Then T has fixed point in Q.

Proof. Suppose T has no fixed point in Q then

||T x − x|| + ||x|| > ||T x|| (||x|| = r) (1)

For

(||T x − x|| + ||x||)2 − ||T x||2 = ||T x − x||2 + ||x||2 − ||T x||2
+2||x|| ||T x − x|| ≥ 2r||T x − x|| > 0

32 Let P be the mapping defined by




x (x ∈ Q)
Px = 

 r x (x < Q)
||x||

Plainly P is a continuous projection of E onto Q. 

Fixed point theorems in normed linear spaces 27

Let T̃ = PT
Then T maps Q continuously into a compact subset of Q. Hence, by
the Schauder theorem, T̃ has a fixed point u in Q,

PT u = u

If T u ∈ Q, then PT u = T u, and

Tu = u

If T u < Q, then ||T u|| > r and

r
u = PT u = Tu
||T u||
If follows that ||u|| = r, and we have
||T u||
||T u − u|| + ||u|| = || u − u|| + ||u||
!r
||T u||
= − 1 + 1 ||u|| = ||T u||
r
which contradicts (1)

supplementary results and exercises

(1) For further results connected with Theorem 2.6, see [19]

(2) Let A be a continuous mapping of a normed space E into itself

which maps bounded sets into compact sets and satisfies
||Ax||
lim =0
||x||→∞ ||x||

Then given arbitrary real λ > 0 and y in E, the equation

x = λAx + y

has a solution x in E 33
Consider the mapping T x = λAx + y
 28 Fixed point theorems in normed linear spaces

Clearly T has all the properties of A

Let S n = {x : x ∈ E, ||x|| ≤ n} (n = 1, 2, . . .)
Then
T S n ⊂ S n for some n (1)
Otherwise ||T xn || > n, for some xn ∈ S n n = 1, 2, . . . (2)
If {xn } were bounded, then {T xn } will be contained in a compact set
and therefore ||T xn || will be bounded which contradicts (2). Hence
||xn || → ∞ as n → ∞ But

||T xn || ||T xn ||
> 1 so lim ≥1
||xn || ||xn ||→∞ ||xn ||

As this is not true, T maps some S n into its compact subset; Schau-
der’s theorem then a gives a fixed point x which is the required
solution.
P
∞
(3) Let ak be a convergent series of non-negative real numbers and
k=1
let( fk ) be
a sequence of continuous mappings of the real line R into
itself such that

| fk (t)| ≤ ak (t ∈ R, k = 1, 2, . . .)

Given α ∈ R, there exists a convergent real sequence (ξk ) such that

(i) ξ = α
(ii) ξk+1 − ξk = fk (ξk ) (k = 1, 2, . . .)

consider the mapping T of (c) into itself given by

(T x)1 = α

X
n
(T x)n+1 = α + fk (ξk ) (n = 1, 2, . . .)
k=1

34 where x = (ξk ).
Fixed point theorems in normed linear spaces 29

(4) Let E be a Banach space with a uniformly convex norm, and let K
be a bounded closed convex subset of E. Then the metric projection
onto
E −−−−→ K exists and is uniformly continuous on each bounded
subset of E.

(5) Brodsku and Milman [10], give conditions under which a convex
set in a Banach space has a point invariant under all isometric self
mappings. In this connection see also Dunford and Schwartz ([14],
p.459).

(6) Browder (11) gives some generalization of the Schauder theorem

which appear to lie rather deep. Perhaps the most striking of these
results is the following generalization of theorem 2.4. Let T be
a continuous mapping of a Banach space E into itself that maps
bounded sets into compact sets. If, for some positive integer m, T m E
is bounded, then T has a fixed point. For a generalization of the
Schauder theorem of a different kind see Stepaneek (32).

(7) Aronszajn [2] gives general regularity condition on T sufficient to

establish that the set of its fixed points is an Rδ i.e. is a homeomor-
phic image of the intersection of decreasing sequence of absolute
retracts.
Chapter 3

The Schauder - Tychonoff

theorem

It this chapter we are concerned with non-linear operators in general 35

locally convex spaces.

Definition 3.1. A vector space E over R which is also a topological

space is called a linear topological space (l.t.s) if the mappings

(x, y) → x + y
(α, x) → αx

from E × E and R × E respectively into E are continuous. If also every

open set in E is a union of convex open sets, then F is said to be locally
convex.

We establish the elementary properties of a l.t.s E. Since the map-

ping (α, x) → αx is continuous, the mapping x → αx, with fixed α,
is continuous. Therefore, if α is a non-zero constant then the mapping
x → αx is a homeomorphism, and so
a) G open, α , 0 =⇒ αG open.
In particular
b) G open implies that - G is open.

31
 32 The Schauder - Tychonoff theorem

c) Similarly, G open ⇐⇒ y + G open, and so

d) V is neighbourhood of 0 if and only if y + V is neighbourhood of y.

36 Let V be a neighbourhood of 0, and let x ∈ E. Since the mapping
1
α → αx is continuous, and 0x = 0, we have x ∈ V for all suffi-
λ
ciently large λ, i.e.,

e) x ∈ λV for all sufficiently large λ.

We prove next that

f) The closure of a convex set is convex.

For 0 ≤ α ≤ 1, the mapping f : E × E → E given by

(x, y) → αx + (1 − α)y

is continuous and f (K × K) ⊂ K. Therefore f (K × K) ⊂ K̄, where K̄

denotes the closure of K. But K × K = K̄ × K̄ and so f (K̄ × K̄) ⊂ K̄
i.e., αa + (1 − α)b ∈ K̄ for a, b ∈ K̄.

g) The interior of a convex set is convex.

Let K0 be the interior of a convex set K, let a, b ∈ K0 and 0 < α < 1.
By (a), αK0 , (1 − α)K0 are open sets. By (c) αK0 + (1 − α)K0 is a
union of open sets and is therefore open. Since

αa + (1 − α)b ∈ K0 + (1 − α)K0 ⊂ K,

it follows that αa + (1 − α)b ∈ K0 . A subset A of a vector space E

over R is said to be symmetric if −A = A.

h) Let U be a neighbourhood of 0 in a locally convex l.t.s. Then there

exists a closed convex symmetric neighbourhood V of 0 with V ⊂ U.
Since 0 is an interior point of U and the space is locally convex,
1
37 there exists a convex open set G with 0 ∈ G ⊂ U. Let H = (G ∩
2
−G), and V = H̄. By (b) and ( f ) V is a closed convex symmetric
neighbourhood of 0. Finally V ⊂ U; for if v ∈ V, then v + H is an
The Schauder - Tychonoff theorem 33

open set containing and therefore has nonempty intersection with H,

ie there exists h, h′ in H with v + h = h′ . Since H is convex and
symmetric,

v = h′ − h ∈ 2H ⊂ G. Thus V ⊂ G ⊂ U.

Definition . Given an l.t.s. E over K, a subset A of E is said absorb

points if for every x in E,
x ∈ λA
for all sufficiently large λ.

Definition. Given a convex set K that absorbs points, the

Minkowski functional pK is defined by

pK (x) = inf{λ; λ > 0, and x ∈ λK}

Definition . A mapping p of E into R is called a seminorm on E if it

satisfies the axioms.

i) p(x) ≥ 0 (x ∈ E)

ii) p(αx) = |α|p(x) (x ∈ E, α ∈ R)

iii) p(x + y) ≤ p(x) + p(y)

Given a seminorm p, the seminorm topology determined by p is

the class of unions of open balls

S (x, ∈) = {y : p(y − x) < ε} (ε > 0)

With this topology E is a locally convex l.t.s. which is not in general a

Hausdorff space.
The Minkowski functional of a convex set K that absorbs points is 38
sublinear, and if K is also symmetric, then it is a seminorm. Also if
x ∈ λK and µ > λ, then x ∈ µK, for 0 ∈ K since K absorbs points and
! !
1 λ 1 λ
x= x + 1− ·0∈ K
µ µ λ µ
 34 The Schauder - Tychonoff theorem

i) If K is a closed convex symmetric neighbourhood of 0 in a l.t.s, the

p Minkowski functional pK is a continuous semi-norm in E, and

K = {x; pK (x) ≤ 1}

Conversely, if p is a continuous seminorm in E, then {x : p(x) ≤ 1}

is a closed convex symmetric neighbourhood K of 0, and pK = p.

Proof. Let K be a closed convex symmetric neighbourhood of 0. 

Then pK is a seminorm on E, and so

|pK (x′ ) − pK (x)| ≤ pK (x′ − x) (x′ , x ∈ E)

Given ε > 0,

x′ ε, x + εK ⇒ x′ − xεK
⇒ pK (x′ − x) ≤ ε
⇒ |pK (x′ ) − pK (x)| ≤ ε

since x + εK is a neighbourhood of x, this shows that pK is continuous.

If xεK, then pK (x) ≤ 1, by the definition of pk . On the other hand, if
pK (x) ≤ 1, then x ε λK(λ > 1),
1
x ∈ K(λ > 1),
λ
and, since K is closed, x ∈ K.
39 Thus
K = {x : pk (x) ≤ 1}.
Conversely, let p be a continuous semi-norm on E, and let K = {x :
p(x) ≤ 1}. That K is a closed convex symmetric neighbourhood of the
origin is evident. We have

p(x) ≤ 1 ⇔ x ∈ K ⇔ pk (x) ≤ 1,

and, since p and pk are both positive-homogeneous, it follows that p =

pk .
The Schauder - Tychonoff theorem 35

(j) Let x be a nonzero point of a Hausdorff locally convex l.t.s.E. Then

there exists a continuous semi-norm p on E with p(x) > 0.

Proof. Since x , 0 and E is a Hausdorff space, there exists a neigh-

bourhood U of 0 such that x < U. By (h) there exists a closed con-
vex symmetric neighbourhood U of 0 with V ⊂ U. By (i), there
exists a continuous semi-norm p on E such that

V = {y : p(y) ≤ 1}

Hence p(x) > 1. 

(k) Let E be a vector space over K. Let p be a semi-norm on E, and let

N = {x : p(x) = 0}. Then N is a subspace of E, and the functional
E
q defined on the quotient space by
N
E
q( x̃) = p(x) (x ∈ x̃, x̃ )
N

is a norm on E/N

Proof. If x, y ∈ N, then 40

0 ≤ p(x + y) ≤ p(x) + p(y) = 0,

and so x + y ∈ N. Also p(x) = 0 implies p(λx) = 0, and so N is

a linear subspace of E. The definition of q( x̃) is in fact free from
ambiguity, for if x, x′ ∈ x̃, then x − x′ ∈ N, and so

|px − p(x′ )| ≤ p(x − x′ ) = 0,

p(x) = p(x′ ).

Finally that q satisfies the axioms of a norm is entirely straight-

forward. 
 36 The Schauder - Tychonoff theorem

Lastly, among these preliminary results, we need a proposition

which is a special case of a general theorem on uniform spaces. How-
ever, it is more convenient for our purposes to prove the special case
than to invoke the general theory.
(b) Let E, F be linear topological spaces, let K be a compact subset
of E and let T be a continuous mapping of K into F. Given a neigh-
bourhood U of 0 in F, there exists a neighbourhood V of 0 in E such
that
x, x′ ∈ K, x − x′ ∈ V ⇒ T x − T x′ ∈ U.

Proof. Let H be an open set containing 0 such that

H−H ⊂U

41 Given xǫK, there exists a neighbourhood G(x) of 0 such that x′ ∈ K ∩

(X + G(x)) ⇒ T x′ ∈ T x + H.
Let V(x) be an open neighbourhood of 0 in E such that

V(x) + V(x) ⊂ G(x),

since K is compact and is covered by open sets x + V(x), it has a finite

covering
x1 + V(x1 ), . . . , xn + V(xn ).
T
n
Let V = V(xi ).
i=1
Then V is a neighbourhood of 0 in E. Suppose x, x′ ∈ K and x − x′ ∈
V. Then there exists j with

x′ ∈ x j + V(x j ) ⊂ x j + G(x j )
x − x j = x − x′ + x′ − x j ∈ V + V(X j ) ⊂ V(x j ) + V(x j ) ⊂ G(x j )
since x, x′ ∈ x j + G(x j ),
we have T x ∈ T x j + H, T x′ ∈ T x j + H,
and so T x − T x′ ∈ H − H ⊂ U.

We are now ready to prove the main theorem by which we are able to de-
duce properties of operators in a locally convex linear topological space
 The Schauder - Tychonoff theorem 37

from the corresponding properties of operators in normed spaces. The

42 main idea of this theorem was derived from the proof of the Schauder-
Tychnoff theorem in Dunford and Schwartz [14] p.454. 

Theorem 3.1. Let K be a compact subset of a locally convex l.t.s E, T a

continuous mapping of K into itself, p◦ a continuous semi-norm on E.

Then there exists a semi-norm q on the linear L(K) of K such that

i) q(x) ≥ p◦ (x) (x ∈ L(K));
ii) q is continuous on K − K;
iii) K is compact with respect to the semi-norm topology given by q;
iv) T is uniformly continuous in K with respect to q i.e., given ǫ > 0,
there exists δ > 0, such that

x, x′ ∈ K, q(x − x′ ) < δ ⇒ q(T x − T x′ ) < ǫ

Remark . It would be better if one could prove the existence of a con-

tinuous semi-norm q on E satisfying (i) and (iv).
Proof. Since p◦ is bounded on K there is no real loss of generality in
supposing that
p◦ (x) ≤ 1 (x ∈ K).
It is convenient to introduce the following definition. We say that a set Γ
of continuous semi-norm dominates a set Γ of continuous semi-norms
if the following two conditions are satisfied. 

a) p′ (x) ≤ 1 (x ∈ K, p′ ∈ Γ′ ) 43

b) given p ∈ Γ and ǫ > 0, there exists p′ ∈ Γ′ and δ > 0 such that

x, x′ ∈ K, p′ (x − x′ ) < δ ⇒ p(T x − T x′ ) < ǫ.

We construct a countable self-dominating set containing p◦ . Given

a continuous semi-norm p, and a positive integer n, the set
( )
1
x : p(x) <
n
 38 The Schauder - Tychonoff theorem

is a neighbourhood of 0. Therefore, by proposition (2), there exists a

neighbourhood V of 0 in E such that
1
x, x′ ∈ K, x − x′ ∈ V ⇒ p(T x − T x′ ) < .
n
By (h), we may suppose that V is a closed convex symmetric neigh-
bourhood of 0, and then by (i), pV is a continuous semi-norm and
V = {x : pV (x) ≤ 1}.

Multiplying p by an appropriate positive constant δn , we obtain a

continuous semi-norm qn such that
qn (x) ≤ 1 (x ∈ K),
1
and such that x, x′ ∈ K, qn (x − x′ ) < δn ⇒ p(T x − T x′ ) < .
n
44 Plainly the set of semi-norms qn is a countable set dominating the set
(p).
It follows that given a countable set Γ of continuous semi-norms,
there exists a countable set Γ′ that dominates Γ. Now the set (p◦ ) is
dominated by a countable set Γ1 , Γ1 is dominated by a countable set Γ2 ,
and so on. Finally, we take
[
∞
Γ = (p◦ ) ∪ Γn .
n=1

Then Γ is a countable self-dominating set. Let (pn )∞

◦ be an enumer-
ation of Γ and take
X
∞
q(x) = 2−n pn (x) (1)
n=0
since
pn (x) ≤ 2 (x ∈ K − K),
the series (1) converges uniformly on K − K, and so q is continuous on
K − K. Also the series converges on L(K) (linear hull of K) and q is a
semi-norm there satisfying (i). Given xǫK, let

S (x, ρ) = {x′ ; x′ ∈ K and q(x − x′ ) < ρ}

The Schauder - Tychonoff theorem 39

since q is continuous on K − K, S (x, ρ) is an open subset of K in the

topology τ on L(K) induced from the initial topology on E. Hence each
open subset of K in the topology induced by τq (topology on L(K) de- 45
fined by q) is also open in the topology induced by τ. (iii) is now an
immediate consequence of the τ-compact-ness of K.
ǫ
Given ǫ > 0, we choose N with 2−N < since we have
4
X∞
1 X∞
1 ǫ
′
n
pn (x − x ) ≤ n+1
< (x, x′ ∈ K),
n=N+1
2 n=N+1
2 2

and so
XN
1 ǫ
q(x − x′ ) ⊂ n
pn (x − x′ ) + (x, x′ ∈ K) (2)
n=0
2 2
since T maps K into itself, (2) gives
XN
1 ǫ
′
q(T x − T x ) < p (T x − T x′ ) +
n n
(x, x′ ∈ K) (3)
n=0
2 2

since Γ is self-dominated, for each n, there exists kn and δn > 0 such that
ǫ
pkn (x − x′ ) < δn ⇒ pn (T x − T x′ ) < (x, x′ ∈ K) (4)
4
Let N ′ = max(k◦ , . . . , kN ), and
′
= 2−N min(δ◦ , . . . , δN ).
′
Then since pn ≤ 2N q for n ≤ N ′ , we have

q(x − x′ ) < δ ⇒ pkn (x − x′ ) < δn (n ≤ N)

and so, by (4)

ǫ
x, x′ ǫk, q(x − x′ ), < δ ⇒ pn (T x − T x′ ) < (n = 0, 1, . . . , N).
4
Therefore, by (3), 46

x, x′ ∈ K, q(x − x′ ) < δ ⇒ q(T x − T x′ ) < ε.

 40 The Schauder - Tychonoff theorem

Theorem 3.2 (Schauder-Tychonoff). Let K be a non- empty compact

convex subset of a locally convex Hausdorff l.t.sE, and let T be a con-
tinuous mapping of K into itself. Then T has a fixed point in K.

Proof. There is no loss of generality in supposing that L(K) = E. Sup-

pose that T has no fixed point in K. Then T x − T x , 0 (x ∈ K). 

It follows by proposition ( j) that for each point x of K there exists a

continuous semi-norm px such that

px (T x − x) > 0

By continuity of T and px , there exists a neighbourhood U x of x such

that
px (T y − y) > 0 (y ∈ U x )
Since K is compact, there is a finite covering of K by such neigh-
bourhood say

U x1 , . . . , U xm .
Let p = px1 + px2 + · · · + pxm .

47 Then p is a continuous semi-norm and

p(T x − x) > 0 (x ∈ K) (1)

Let q be the semi-norm constructed as in Theorem 3.1 with p0 = p.

Then q is defined on L(K) = E, q ≥ p, K is compact in the semi-norm
topology τq , and given ε > 0, there exists δ > 0 such that

x, x′ ǫK, q(x − x′ ) < δ ⇒ q(T x − T x′ ) < ε (2)

Let N = {x : q(x) = 0}. By lemma 3.4, E/N is a normed space with

the norm given by
|| x̃|| = q(x)
where x̃ is the coset of x. Let K̃ = {x : x ∈ K}. Then since mapping
x → x̃ is a continuous homomorphism from E with the topology τq to
The Schauder - Tychonoff theorem 41

E/N with the norm topology, K is a compact convex set in E/N. Also
by (2),
x, x′ ∈ K, q(x − x′ ) = 0 ⇒ q(T x − T x′ ) = 0 (3)
For each x̃ in K̃ there exists a point x in x̃ ∩ K, and we define T̃ x̃ by
taking
T̃ x̃ = Tfx
By (3), this definition is unambiguous, and T̃ maps K̃ into itself.
Also, by (2), given ε > 0, there exists δ > 0 such that x̃, x̃′ ∈ K, || x̃− x̃′ || <
δ ⇒ ||T̃ x̃ − T̃ x̃′ || < ǫ. For given x̃, x̃′ ǫK, there exist x ∈ K ∩ x̃ and 48
x′ ǫK ∩ x′ and q(x − x′ ) = || x̃ − x̃′ ||. Hence T̃ is a continuous mapping of
the compact convex subset K̃ of the normed space E/N. Applying the
Schauder fixed point theorem, T̃ has a fixed point ũ say

T̃ ũ = ũ.

Since ũǫ K̃, there exists u ∈ K ∩ ũ, and we have T̃ ũ = Tfu. Thus

T u − u ∈ N,
i.e., q(T u − u) = 0

It follows that p(T u − u) = 0, which contradicts (1) since u ∈ K.

Problem . It will be noticed that Theorem 3.2 generalizes theorem 2.3

rather than the full force of the Schauder theorem (2.2). It is not known
whether the following proposition is true.

Q. Let K be a closed convex subset of a locally convex Hausdorff

l.t.s. E, and let T be a continuous mapping of K into a compact subset
of K. Then T has a fixed point in K.
It is obvious that if T maps K into a compact convex subset H of K,
then T has a fixed point. For

TH ⊂ TK ⊂ H

and we can apply theorem 3.2 to H instead of K. In particular, Q will 49

hold if every compact subset of K is contained in compact convex sub-
set of K. By an elementary theorem of Bourbaki (Espaces Vectoriels
42 The Schauder - Tychonoff theorem

Topologiques, Ch.II, p.80) the convex hull of a precompact subset of a

locally convex Hausdorff l.t.s is precompact. Thus we can obtain a true
theorem from Q by supposing that K be complete instead of closed or E
quasi-complete. However, this is certainly unnecessarily restrictive. By
the Krein-Smulian Theorem [21], if E is a branch space with the weak
topology as the specified topology, then the closed convex hull of each
compact subset of E is compact, and so the proposition Q holds, even
though K need not be completes (in the weak topology).

Example. Let E be a reflexive Banach space, K a closed convex subset

of E, T a weakly continuous mapping of K into a bounded subset of K.
Then T has a fixed point in K.

For since K is norm closed and convex it is also weakly closed.

Also since E is reflexive, each bounded weakly closed subset of E is
weakly compact. Hence the weakly closed convex hull of T K is weakly
compact.
Chapter 4

Nonlinear mappings in cones

The theorems in this chapter are mainly due to Krein and Rutman [20] 50
and to Schaefer [28]. They may be regarded as a further step in the
transition from nonlinear to linear problems. We will be content with
considering normed spaces only, through theorems of the kind studied
here have been proved for general locally convex spaces by H. Schaeffer.

Definition 4.1. A subset C of a vector space E over R is called a positive

cone if it satisfies

(i) x, y ∈ C ⇒ x + y ∈ C

(ii) x ∈ C, α ≥ 0 ⇒ α ∈ C

(iii) x, −x ∈ C ⇒ x = 0

(iv) C contains non-zero vectors.

A vector space E over R with a specified positive cone is called a

partially ordered vector space , and we write x ≤ y( or y ≥ x) to denote
that y − x ∈ C. It is easily verified that this relation ≤ is a relation of
partial order in the usual sense i.e.,

(v) x ≤ x (x ∈ E),

(vi) x ≤ y, y ≤ z ⇒ x ≤ z,

43
 44 Nonlinear mappings in cones

(vii) x ≤ y, y ≤ x ⇒ x = y.
Also the partial ordering and the linear structure are related by the prop-
51 erties:
(viii) xi ≤ yi (i = 1, 2) ⇒ x1 + x2 ≤ y1 + y2 ,
(ix) x ≤ y, 0 ≤ α ≤ β ⇒ αx ≤ βy.
Conversely, given a non-trivial relation ≤ in E satisfying (v), . . . ,
(xi), the set {x : 0 ≤ x} is a positive cone in E to which the given relation
corresponds in the above manner.

Definition 4.2. Let C be a positive cone in a normed space E. A map-

ping T of a subset D of C into C is said to be strictly positive in D
if
xn ∈ D, lim T xn = 0 ⇒ lim xn = 0
n→∞ n→∞
A mapping T defined on C is said to be completely continuous in C if it
is continuous in C and maps each bounded subset of C into a compact
set.

Theorem 4.1 (Morgenstern [23]). Let C be a closed positive cone in a

normed vector space E such that the norm is additive on Ci.e.
||x + y|| = ||x|| + ||y|| (x, y ∈ C)
Let c > 0, let K = {x : x ∈ C, ||x|| = c}, and let T be a continuous
and positive mapping on K and strictly positive on K and map K into
a compact subset of C. Then there exists u in K and λ > 0 such that
T u = λu .

Proof. Since the norm is additive on C, K is a convex set. Since T is

strictly positive on K,
inf{||T x||, x ∈ K} > 0
and therefore the mapping A defined on K by
Ax = c||T x||−1 T x
52 is continuous and maps K into a compact subset A itself. By the
Schauder theorem, there exists u ∈ K with Au = u. 
Nonlinear mappings in cones 45

Corollary . Let C be a closed positive cone in a normed vector space

E such that the norm is additive on C. Let T be a strictly positive and
completely continuous mapping of C into itself. Then for each c > 0,
there exists uc in C and λc > 0 such that T uc = λc uc and ||uc || = c.

Definition 4.3. A positive cone C in a normed vector space is said be

normal if there exists a positive constant γ such that

||x + y|| ≥ γ||x|| (x, y ∈ C).

Theorem 4.2 (Schaefer). Let C be a closed

n normal
o positive cone in a
x∈C
normed space. Let c > 0 and let K = x|||x||≤c . Let T be a continuous
and strictly positive on K and map K into a compact set. Then there
exists u ∈ C, and λ > 0, such that T u = λu and ||u|| = c.

Proof. Since T K is contained in a compact set we can choose µ > 0,

such that µT k ⊂ K. Let A = µT , let y be a point of K with y = c, and let
B be the mapping defined on K by Bx = c−1 xAx + c−1 (c − x)y (x ∈ k),
since K is convex, we have BK ⊂ K. Also B is continuous in K, and
maps K into a compact set. Since T is strictly positive on K, there exists
c > 0 such that
1
x ∈ K, ||x|| ≥ c ⇒ ||Ax|| ≥ ε
2
since C is a normal cone, it follows that 53

1 1 1
x ∈ K, ||x|| ≥ c ⇒ ||Bx|| ≥ γc−1 cε = γε
2 2 2
On the other hand,

1 1 1
x ∈ K, ||x|| ≤ c ⇒ c−1 (c − ||x||)||y|| ≥ c ⇒ ||Bx|| γc
2 2 2
1
Therefore ||Bx|| ≥ γ min(ε, c) > 0 (x ∈ K).
2
It follows that the mapping

x → c||Bx||−1 Bx
 46 Nonlinear mappings in cones

is a continuous mapping of K into a compact subset of itself. Therefore,

there exists u ∈ K with
u = c||Bu||−1 Bu
Plainly ||u|| = c, and so Bu = Au, and we have

Au = λ, with λ = c−1 ||Au|| > 0.

The following theorem due to Krein and Rutman [20, Theorem 9.1]
marks a further transition towards a linear problem. 

Definition 4.4. Let E be a partially ordered vector space with positive

cone C, let T be a mapping of C into itself and let c be be positive real
number. T is said to be

54 (i) positive-homogeneous of

T (αx) = αT x (α ≥ 0, x ∈ C)

(ii) monotonic increasing if

x, y ∈ C, x ≤ y ⇒ T x ≤ T y

(iii) c-dominant if there exists a nonzero vector u in C with T u ≥ cu.

Lemma 4.1. Let C be a closed positive cone in a normed space E,

and let u be a point that does not belong to −C. Then there exists a
continuous linear functional f on E such that

(i) f (u) = d(u, −C) > 0,

(ii) f (x) ≥ 0 (x ∈ C),

(iii) || f || ≤ 1

Proof. Let
p(x) = d(x, −C) = inf{||x + y|| : y ∈ C}
Then p is a sublinear functional on E, with the properties
Nonlinear mappings in cones 47

(a) p(u) = d(u, −C) > 0,

(b) p(x) = 0(x ∈ −c),
(c) p(x) ≤ ||x|| (x ∈ E).


By the Hahn-Banach theorem there exists a linear functional f on E

with f (u) = p(u) and with

f (x) ≤ p(x) (x ∈ E).

Plainly f has the required properties. 55

Theorem 4.3 (Krein and Rutman). Let E be a partially ordered normed

vector space with a closed positive cone C. Let T be a completely con-
tinuous mapping of G in to itself which is positive-homogeneous, mono-
tonic increasing, and c-dominant for some c > 0. Then there exists a
nonzero vector v in C and a real number λ ≥ c such that T v = λv.

Proof. Since T is positive-homogeneous and c-dominant, there exists a

vector u in C with ||u|| = 1 and

T u ≥ cu (1)

since u < −C, Lemma 4.1 establishes the existence of a continuous

linear functional f on E with

f (u) > 0, f (x) ≥ 0 (x ∈ C) (2)

and || f || = 1 (3)

We now prove that

x ∈ C, α > 0, β > 0, T x = αx − βu ⇒ α > c. (4)

Let Γ denote the set of positive real numbers t with x ≥ tu. Since
β 1 β β
x = u + T x ≥ u, we have ∈ Γ. Also Γ is bounded above,
α α α α
for otherwise
1
x ≥ u (n = 1, 2, . . .),
n
 48 Nonlinear mappings in cones

and since C is closed, this gives 0 ≥ u, u = 0 with is not true.

Let m denote the least upper bound of Γ. Using again the fact that C 56
is closed, we have
x ≥ mu
and therefore
T x ≥ T (mu) = mT u ≥ mcu
Since T x = αx − βu, this gives
β + mc
x≥ u,
α
and therefore
β + mc
≤ m,
α
m(α − c) ≥ β > 0,
α>c

In the rest of the proof ε will denote a real number with

1
0<ε< (5)
2
Let Kε = {x : x ∈ E, || x ||≤ 1, xgeqε || x || u, f (x) ≥ ε f (u) Clearly K
is a closed, convex, bounded subset of E. Next we note that, for some
δ>0,

|| T x ||≥ δ ∈ f (u)0(x ∈ K∈ ) (6)

For x ∈ Kε , x ≥ ε || x || u gives

T x ≥ ε || x || T u

57 since T is positive homogeneous and monotonic increasing .

By (3), || T x ||≥ f (T x)
and by (2) f (T x) ≥ f (εc || x || u) = εc || x || f (u) (xǫK∈ )
i.e., || T x ||≥ εc || x || f (u) ≥ εe f (x) f (u)
Nonlinear mappings in cones 49

≥ δε f (u) (x ∈ Kε )

with δ = εc f (u) > 0.

Let Vǫ be the mapping defined by taking

Vǫ (0) = 0,
Vε (x) =|| x || . || x + 2ε || x || u ||−1 (x + 2ε || x || u), x , 0

V is well defined since

|| x + 2ǫ || x || u ||≥|| x || −2ε || x || || u ||
=|| x || (1 − 2ε) > 0 if || x ||, 0.

Plainly Vǫ is continuous in E and

|| Vǫ x ||=|| x || (7)

Also
x ∈ C, || x ||= 1 Vε x ∈ Kε (8)
For f (Vε x) = ||x||||x + 2ε||x||u||−1 { f (x) + 2ε||x|| f (u)}
2ǫ
≤ f (u) ≥ ε f (u)
1 + 2ǫ
Let Aε be the mapping defined on Kǫ by 58

Aε = Vε LT
x
when Lx = x,0
||x||
Then by (6) and (8)
Aε Kε ⊂ Kε
By (6), Vε L is continuous in T Kε , and Aε continuously into a compact
subset of Kε . Applying the Schauder theorem, we see that there exists a
point xε in Kε such that
Aε xε = xε ,
( )
Tx
i.e., Vε = xε (9)
||T xε ||
 50 Nonlinear mappings in cones

i.e. ||T xε ||−1 T xε + 2εU = || ||T xε ||−1 T xε + 2εu||xε This can be written in
the form
T xε = αε xε − βε u, (10)
where c < αε < (1 + 2ε)||T xε ||.
We now choose a sequence (εn ) such that lim εn = 0, and such
n→∞
that the sequences (T xεn ) and (αεn ) converges. Let v = lim T xεn and
n→∞
59 λ = lim αǫn . Then λ ≥ c, and since lim n → ∞βεn = 0, (10) gives
n→∞

1
lim xεn = v
n→∞ λ
By continuity and positive homogeneity of T ,
!
1 1
T v = T v = lim T xǫn = v
λ λ n→∞

Finally by (7) and (9), ||xǫ || = 1 and so v , 0. 

Remark. The theorems in this chapter are unsatisfactory in that each of

them involves an adhoc condition (strict positivity and c- domi-nance).
It turns out that for linear mappings such an ad hoc condition can be
avoided, and I think that there is still scope for proving a better theorem
on non-linear mappings also.
Chapter 5

Linear mapping in cones

If A is a linear operator in R(n) with a matrix (ai j ) with non-negative 60

elements, ai j ≥ 0, then, by a famous theorem of Perron and Frobenius
(see for example Gantmacher, The theory of matrices), there exists an
eigen vector of A with non-negative coordinates and with eigenvalue ρ ,
such that all other eigenvalues satisfy |λ| ≤ ρ.
If we take E = R(n) and C to be the set of all vectors in E with non-
negative coordinates, then C is a positive cone in E , and n × n matrices
with non-negative elements correspond to linear operators in E that map
C into itself. Then theorems of the present chapter may be regarded as
generalizations of the Perron-Frobenius Theorem. A great many such
generalizations with various methods of proof have been published dur-
ing recent years, and our list of references is far from complete.
The idea of the method of proof adopted here is the use of the simple
concept of ’topological divisor of zero’.
Let U be a Banach algebra with a unit element e, and let a be a
frontier point of the set of invertible elements. Then a is a topological
divisor of zero, i.e., there exists a sequence (xn ) with ||xn || = 1(n =
1, 2, . . .) such that
lim axn = 0
n→∞

61

51
 52 Linear mapping in cones

Proof. There exists a sequence (an ) of invertible elements such that

lim an = a
n→∞

Then the sequence (||a−1

n ||) is unbounded. For otherwise

lim (an − a)a−1

n =0
n→∞
i.e., lim (e − aa−1
n )=0
n→∞

But this implies that aa−1

n is invertible for some n, and therefore a has a
right inverse. Similarly a has a left inverse, and a is invertible, which is
absurd since the set of invertible elements is open. 

We may therefore suppose that ||a−1 −1 −1

n || → ∞, and take xn = ||an ||
a−1
n . Then ||xn || = 1, and

lim axn = lim (a − an )xn + an xn = 0.

n→∞ n→∞

In particular, if λ is a frontier point of the spectrum of an element b,

then λe − b is a frontier point of the set of invertible elements of U and
so there exists (xn ), with ||xn || = 1 and

lim (λe − b)xn = 0

n→∞

62 If further we know that λ , 0, and bxnk → u for some subsequence

(xnk ). Then

λxnk → u,
λbxnk → bu,

so that bu = λu and u , 0, since ||u|| = λ.

Actually, our method is not quite so simple as this, for our Banach
algebra is a Banach algebra of operation on a Banach space X, and we
have to replace the sequence (xn ) of operators by a sequence of elements
of X.
Until we reach the statement of our main theorem (Theorem 5.1) we
shall use the following notation.
Linear mapping in cones 53

E will denote a normed and partially ordered vector space with norm
||x|| and positive cone C. We suppose that C is complete with respect to
||x||, and that
E =C −C
We do not suppose that E is complete with respect to ||x||. We denote by
B the intersection of C and the closed unit of E, i.e.,

B = {x : x ∈ C and ||x|| ≤ 1}

We denote by B0 the convex symmetric hull of B, i..e.

B0 = {αx − βy : x, y ∈ B, α ≥ 0, β ≥ 0, α + β = 1},

and by ||x||c the Mindowski functional of B0 , i.e., 63

||x||c = inf{λ : λ > 0, x ∈ λB0 }

Lemma 5.1. (α) ||x||c is a norm on E and satisfies

||x||c = ||x|| (x ∈ C), ||x||c ≥ ||x|| (x ∈ E)

(β) E is complete and C is a closed subset of E with respect to ||x||c .

Proof. (α) since E = C − C, B0 is an absorbing set for E, and so the
Mindowski functional ||x||c is defined on E. Since B0 is convex and
symmetric, ||x||c is a seminorm on E. If z ∈ E and ||z||c < 1, then z ∈ B0
i.e. z = αx − βy with x, y ∈ B and α ≥ 0, β ≥ 0, α + β = 1. Therefore

||z|| = ||αx − βy|| ≤ α||x|| + β||y|| ≤ α + β = 1.

This proves that

||x|| ≤ ||x||c (x ∈ E)
and completes the proof that ||x||c is a norm. Since B ⊂ B0 , we have

||x||c ≤ 1 (x ∈ B),

and therefore
||x||c ≤ ||x|| (x ∈ C)
 54 Linear mapping in cones

This completes the proof of (α). 64

(β) Let (zn ) be a Cauchy sequence in E with respect to xc . Then
there exists a strictly increasing sequence (nk ) of positive integers such
that
p, q ≥ nk ⇒ ||z p − zq ||c < 2−k
Let wk = znk (k = 1, 2, . . .). Then, in particular,

||wk+1 − wk ||c < 2−k (k = 1, 2, . . .)

Therefore
wk+1 − wk ∈ 2−k B0 ,
and so wk+1 − wk = αk xk − βk yk , with

αk ≥ 0, β ≥ 0, αk + βk = 1, xk , yk ∈ 2−k B

Let
X
n X
n
sn = αk xk , tn = βk yk .
k=1 k=1


Then p > q gives

p
X p
X
||s p − sq || ≤ αk ||xk || ≤ 2−k < 2−q ,
q+1 k=q+1

and similarly
||t p − tq || < 2−q .
65 Since C is complete, there exist s, t in C such that

lim ||s − sn || = 0, lim ||t − tn || = 0

n→∞ n→∞

Also, lim s p − sq = s − sq , and s p − sq ∈ C whenever p > q. Hence

p→∞
s − sn ∈ C (n = 1, 2, . . .). Therefore

||s − sn ||c = ||s − sn || (n = 1, 2, . . .),

Linear mapping in cones 55

and so lim ||s − sn ||c = 0.

n→∞

Similarly, lim ||t − tn ||c = 0, and so (wn ) converges with respect to

n→∞
||x||c to w1 + s − t. It is now easily seen that (zn ) converges with respect to
||x||c , and so E is complete with respect to ||x||c . That C is a closed subset
of E with respect to ||x||c is a simple consequence of the inequality

||x|| ≤ ||x||c (x ∈ E)

and the closeness of C with respect to ||x||, (in fact a larger norm gives a
stronger topology).

Definition 5.1. A linear operator in E is said to be positive if it maps C

into C and to be partially bounded if it maps B into a bounded set. The
partial bound p(T ) of a partially bounded linear operator T is defined
by
p(T ) = sup{||T || : x ∈ B}
Given partially bounded positive linear operators S and T , we have 66

p(S T ) ≤ p(S )p(T ),

and therefore the limit

 1
lim p(T )n n
n→∞

exists. It is called the partial spectral radius of T .

Lemma 5.2. A positive linear operator T is partially bounded if and

only if it is a bounded linear operator in the Banach space (E, ||x ||c ).
For such an operator T ,

p(T ) = ||T ||c = sup {||T x||c : x ∈ E and ||x||c ≤ 1} ,

and the partial spectral radius µ of T is equal to its spectral radius as

an operator in (E, ||x||c ), i.e.
1
µ = lim {||T n ||c } n
n→∞
 56 Linear mapping in cones

Finally, if λ > µ, there exists a partially bounded positive linear opera-

tor Rλ , such that

(λI − T )Rλ = Rλ (λI − T ) = I,

where I is the identity operator in E.

Proof. Let T be a partially bounded positive linear operator in E and let
x ∈ E with ||x||c < 1. Then x ∈ B0 , and so

x = αy − βz

67 with α ≥ 0, β ≥ 0, α + β = 1, y, z ∈ B. Then

T x = αT y − βT z,

and so

||T x||c ≤ α||T y||c + β||T z||c

= α||T y|| + β||T z||
≤ (α + β)p(T ) = p(T )

Thus T is a bounded linear operator in (E, ||x||c ) and

||T ||c ≤ p(T )

For the converse and the reversed inequality it is enough to note that

||T ||c ≥ sup {||T x||c : x ∈ C and ||x||c ≤ 1}

sup {||T x|| : x ∈ B} = p(T ).


1
That µ = lim {||T n ||c } n
is an obvious consequence of the fact that
n→∞
||T ||c = p(T ) for each partially bounded linear operator.
If λ > µ, the series
1 1 1
I + T + 3 T2 + · · ·
λ λ2 λ
 Linear mapping in cones 57

68 converges with respect to the operator norm for bounded linear operators
in the Banach space (E, ||x||c ) to a bounded linear operator Rλ , and
(λI − T )Rλ = Rλ (λI − T ) = I.

Since C is closed with respect to ||x||c , and the partial sums of the
series are obviously positive operators, it follows that Rλ is a positive
operator.

Definition 5.2. A positive linear operator T is said to be a normalising

operator if it satisfies the following condition:
αn ≥ 0, yn ∈ B, αn x ≥ T yn , lim αn = 0 lim ||T yn || = 0
n→∞ n→∞

If C is a normal cone, then every positive linear operator is obvi-

ously a normalising operator. We shall see later that a certain compact-
ness condition on T suffices to make T a normalising operator (without
restriction on C).
Lemma 5.3. Let T be a normalizing partially bounded positive linear
operator with partial spectral radius µ. Then

lim p(Rλ ) = ∞
λ→µ+0

Corollary. For each such operator T , the partial spectral radius µ is in

the spectrum of T regarded as an operator in the Banach space (E, ||x||c ).
Suppose that the conditions of the lemma are satisfied but that p(Rλ )
does not tend to infinity as λ decreases to µ. Then there exists a positive 69
constant M such that p(Rν ) ≤ M for some ν gather than and arbitrarily
close to µ.
The case µ = 0 is easily settled. For if µ = 0, then, from the formula
(λI − T )Rλ = I, it follows that

λR − λx ≥ x (λ > 0, x ∈ C) (1)
If we let λ tend to zero through values for which p(Rλ ) ≤ M, the left
hand side of (1) tends to zero, and, since C is closed, we obtain

−x ∈ C (x ∈ C).
 58 Linear mapping in cones

But this implies that C = (0) which was excluded by our axioms on C.
Suppose now that µ > 0. Then we may choose λ, ν with

0 < λ < µ < ν < λ + M −1 ,

and with p(Rν ) ≤ M. Since ||R||c = p(Rν ), it follows that the series

Rν + (ν − λ)R2ν (ν − λ)2 R3ν + · · ·

converges with respect to the operator norm || ||c to a partially bounded

positive linear operator S which is easily seen to satisfy

S (λI − T ) = (λI − T )S = I

70 This gives
S x = λ−1 x + λ−1 T S x (x ∈ C),
from which it follows by induction that

S x ≥ λ−(n+1) T n x (x ∈ C, n = 0, 1, 2, . . .) (2)
1
since λ < µ and lim ||T n ||cn = µ, we have
n→∞

lim ||λ−(n+1) T n ||c = ∞

n→∞

By the principle of uniform boundedness, there exists a point x in E

for which the sequence ||λ−(n+1) T n x||c is unbounded. Since E = C −
C, it follows that there exists a point w in C for which the sequence
||λ−(n+1) T n w|| is unbounded. But given any unbounded sequence {an } of
non-negative real numbers, there exists a subsequence (ank ) such that

ank > k (k = 1, 2, . . .) (3)

ank > a j ( j < nk , k = 1, 2, . . .) (4)

71 This is easily proved by induction. For if n1 , . . . , nr−1 have been

chosen so that (3) and (4) are satisfied for k = 1, 2, . . . , r − 1, we take nr
to be the smallest positive integer s for which

as > anr−1 + r
Linear mapping in cones 59

Hence we see that there exists a strictly increasing sequence (nk ) of pos-
itive integers for which

lim λ−(nk +1) T nk = ∞ (5)

k→∞

and
||λ−(nk +1) T nk || ≥ ||λ−nk T nk −1 w|| (6)
since
||T nk w|| ≤ p(T )||T nk −1 w||,
we also have
lim ||λ−nk T nk −1 w|| = ∞ (7)
k→∞

By (7), there is no loss of generality in supposing that T nk −1 , 0 for

all k, and we may take

yk = ||T nk −1 w||−1 T nk −1 w (8)

Then, by (2),

λnk ||T nk −1 w||−1 S w ≥ λ−1 T yk (9)

Since yk ∈ B, and T is a normalizing operator, it follows from (7) 72

and (9) that
lim λ−1 ||T yk || = 0 (10)
k→∞

But, by (6),
λ−1 ||T nk w|| ≥ ||T nk −1 w||,
which obviously contradicts (10). This contradiction proves the lemma.
To deduce the corollary, it is enough to appeal to the continuity of
the resolvent operator on the resolvent set.

Definition 5.3. Let τN denote the given norm topology in E, τ a second

linear topology in E, and A a subset of the positive cone C. We say that
τ is sequentially stronger than τN at 0 relative to A if 0 is a (τN )-cluster
point of each sequence of point of A of which 0 is a τ-cluster point.
 60 Linear mapping in cones

We recall that to say that 0 is a τ-cluster point of a sequence (an )

means that each τ-neighbourhood of 0 contains points an which arbi-
trarily large n.

Theorem 5.1. Let E be a normed and partially ordered vector space with
norm topology τN , positive cone C complete with respect to the norm,
and let B = {x : x ∈ C, ||x|| ≤ 1}. Let T be a partially bounded positive
linear operator in E with partial spectral radius µ, and let τ be a linear
73 topology in E with respect to which C is closed and T is continuous.

Let A be a subset of C that is contained in a countably τ-compact

subset of C, and let τ be sequentially stronger than τN at 0 relative to A.
If either

(i) A = T B and µ > 0,

or

(ii) A = B,
then there exists a non-zero vector u in C with T u = µu.

Proof. Since we can restrict our consideration to C − C, we shall

suppose that in fact E = C − C. 

Since T B ⊂ p(T )B, both (i) and (ii) imply

(iii) T B is contained in a subset of C that is countably compact with

respect to τ, and τ is sequentially stronger than τN at 0 relative to
T B.

We prove that under condition (iii), T is a normalizing operator. Let

αn x = T yn + zn

with αn ≥ 0, lim αn = 0, yn ∈ B and zn ∈ C. If ||T yn || does not converge

n→∞
to zero, we may select a subsequence (T ynk ) such that

||T ynk || ≥ ε > 0 (k = 1, 2, . . .) (1)

Linear mapping in cones 61

We have
T ynk + znk → 0 (τ),
and (T ynk ) has a τ-cluster point, v say, in C. It follows that −v is a τ- 74
cluster point of (znk ), and, since C is τ-closed, −v ∈ C, v = 0. This now
implies that 0 is a τ-cluster point of (T ynk ) and therefore is a τN -cluster
point of (T ynk ), which contradicts (1).
Thus, by Lemma 5.3, we have

lim p(Rλ ) = ∞,
λ→µ+0
i.e., lim ||Rλ ||c = ∞.
λ→µ+0

Applying the principle of uniform boundedness, we see that there

exists a sequence (λn ) converging decreasingly to µ and a vector w in C
with ||w|| = 1 such that
lim Rλn w = ∞,
n→∞

and we may suppose that ||Rλn w|| , 0 (n = 1, 2, . . .). Let αn = ||Rλn w||−1
and un = αn Rλn w. Then un ∈ B, ||un || = 1, lim αn = 0, and µun − T un =
n→∞
(µ − λn )un + (λn I − T )un = (µ − λn )un + αn w.
Suppose that condition (ii) in the statement of the theorem is satis-
fied. Since B is τ-countably compact and un ∈ B, it follows from (2)
that
lim µun − T un = 0 (τ) (3)
n→∞
75
Also (un ) has a τ-cluster point u in C, and (3) shows that µu−T u = 0.
We have u , 0, for otherwise a is 0 τN -cluster point of (un ), which
contradicts ||un || = 1.
Finally suppose that the condition (i) is satisfied. Then by (2).

(µ − I − T )T un = T (µI − T )un = (µ − λn )T un + αn T w

Since T B is contained in a τ-countably-compact subset of C, this shows

that
lim (µI − T )T un = 0 (τ)
n→∞
 62 Linear mapping in cones

and that (T un ) has a τ-cluster point v in K. By the τ-continuity of (µI −

T ), we have therefore
(µI − T )v = 0
If v = 0, then 0 is a τN -cluster point of (T un ). But, by (2),

lim µun − T un = 0 (τN )

n→∞

and therefore 0 is a τN -cluster point of (µun ). But since µ > 0, and

||un || = 1, this is absurd.
The statement of Theorem 5.1 is somewhat complicated in that it
76 seeks to combine generally and precision. A number of less complicated
but also less general theorems are easily deduced from it.

Theorem 5.2. Let C be a complete positive cone in a normed space E,

let B = {x : x ∈ C and ||x|| ≤ 1}, and let T be a positive linear operator
which is continuous in C, and maps B into a compact set and has a
positive partial spectral radius µ. Then there exists a non-zero vector u
in C with
T u = µu.

Proof. In Theorem 5.1, take τ = τN . 

Example 1. Let E = CR [0, 1] with the uniform norm, and let C be the
positive cone in E consisting of those functions f belonging to E that
are increasing, convex in [0, 1] and satisfy f (0) = 0.

Let 0 < k < 1, and let T denote the operator in E defined by

(T f )(x) = f (kx) ( f ∈ E, 0 ≤ x ≤ 1)

Plainly T is a bounded linear operator in E and maps C into itself.

For f in C we have || f || = f (1), and also since x = (1 − x)0 + x1,

f (x) ≤ (1 − x) f (0) + x f (1) = x f (1) (0 ≤ x ≤ 1)

since T n f (x) = f (K n x), it follows that

||T n f || = f (kn ) ≤ kn f (1) = kn || f || ( f ∈ C)

 Linear mapping in cones 63

77 and equality is attained with the function f (x) = x. Thus

p(T n ) = kn ,

and the partial spectral radius of T is k > 0. Also T maps B into a

compact set, for given a convex function f and 0 < x1 < x2 ≤ y1 < y2 ≤
1, we have
f (x2 ) − f (x1 ) f (y2 ) − f (y1 )
≤
x2 − x1 y2 − y1
Given f ∈ B, and 0 ≤ s < t ≤ 1, we therefore have
f (kt) − f (ks) f (1) − f (k) 1
0≤ ≤ ≤ ,
kt − ks 1−k 1−k
k
i.e., 0 ≤ (T f )(t) − (T f )(s) ≤ (1 − s)
1−k
This proves that T B is an equicontinuous set of functions, and, since
T B is also bounded, it is contained in a compact set.
For this particular operator T the conclusion of Theorem 5.2 is of
course trivial since the function u given by u(x) = x(0 ≤ x ≤ 1) plainly
satisfies T u = ku.
The example is however of interest in that it provides a simple ex-
ample of a bounded linear operator completely continuous in a cone that
is not a compact linear operator in any subspace of E that contains C. In
fact let gn , hn , fn be defined by

gn (x) = k−n x (0 ≤ x ≤ 1),




0 (0 < x < kn )
hn (x) = 

k−n (x − kn ) (kn ≤ x ≤ 1)
and fn = gn − hn .

Then gn , hn ∈ C, fn ∈ C − C and since 78



k−n x (0 < x < kn )

fn (x) =  
1 (kn ≤ x ≤ 1)
we have || fn || = 1 (n = 1, 2, . . .).
 64 Linear mapping in cones

Also,
T fn = fn−1 (n = 1, 2, . . .),
and for r > s,

|| fr − f s || ≥ fr (kr ) − f s (kr ) = 1 − kr−s > 1 − k.

It follows that no subsequence of (T fn ) converges.

Example 2. A slight modification of the last example yields a less triv-

ial application of Theorem 5.2. Let C denote the class of continuous,
increasing, convex functions f on [0, 1] with f (0) = 0, and let φ be an
element of C that satisfies

φ(1) < 1, φ1+ (0) > 0.

Then there exists an element g of C such that g ◦ φ = φ1+ (0)g. ( f ◦ g

79 denotes the composition ( f ◦ g)(x) = f (g(x))). As before we take E =
CR [0, 1], and consider the linear operator T given by

Tf = f ◦φ ( f ∈ E)

That Theorem 5.2 is applicable is proved as in the last example, except

for showing that the partial spectral radius µ is given by

µ = φ′+ (0)

To prove this we consider the sequence (φn ) of functions defined by

φn = T n φ (n = 1, 2, . . .).

Let k = φ (1). Then, for f in C,

||T f (x)|| = (T f )(1) = f (φ(1)) ≤ φ(1) f (1),

and so p(T ) ≤ k < 1.

lim φn (1) = 0.
n→∞
Hence
||tn φ|| φn (1) φ(φn−1 (1))
lim n−1
= lim = lim = φ′+ (0),
n→∞ ||T φ|| n→∞ φn−1 (1) n→∞ φn−1 (1)
Linear mapping in cones 65

1
and therefore lim ||T n φ|| n = φ′+ (0),
n→∞

µ ≥ φ′+ (0).

Finally (T n f )(x) = f (φn (x)), so that

p(T n ) ≤ φn (1),

and so µ ≤ φ′+ (0). This completes the proof that Theorem 5.2 is appli- 80
cable.
In this particular example, we can calculate an eigenvector g by an
iterative process. In fact, if we take gn defined by

φn (x)
gn (x) = (0 ≤ x ≤ 1, n = 1, 2, . . .)
φn (1)

Then the sequence (gn ) converges decreasingly to a function g with the

required properties.

Example 3. Let E, C and φ be defined as in the last example, and let

K(x, y) be a function continuous on the square

[0, 1] × [0, 1]

which belongs to C as a function of x for each fixed y in [0, 1]. Let T be

the operator defined on E by
Z 1
(T f )(x) = f (φ(x)) + K(x, y) f (y)dy (0 ≤ x ≤ 1).
0

Then T is a bounded linear operator mapping C into itself. T maps

B into a compact set, and its spectral radius µ satisfies

µ ≥ φ′+ (0) > 0.

Thus Theorem 5.2 is again applicable.

 66 Linear mapping in cones

Example 4. A variant on Example 3 is given by

Z x
(T f )(x) = f (φ(x)) + k(y) f (y)dy
0

where k is increasing non-negative and continuous in [0, 1]. Again The- 81

orem 5.2 is applicable and so there exists a non-zero function f in C
with Z x
µ f (x) − f (φ(x)) = k(y) f (y)dy (0 ≤ x ≤ 1),
0
where µ is the partial spectral radius of T . From this we see that µ f (x) −
f (φ(x)) is differentiable and we have a solution of the functional equa-
tion
d
{µ f (x) − f (φ(x))} = k(x) f (x)
dx
Example 5. That the conclusion of Theorem 5.2 need not hold if µ = 0
is seen by taking the following example. Let K(x, y) be continuous and
non-negative in the square [0, 1] × [0, 1], and suppose that

K(1, y) > 0 (0 ≤ y ≤ 1).

Let E = CR [0, 1], let C consist of all f ∈ E with

f (x) ≥ 0 (0 ≤ x ≤ 1),

and let T be the Volterra operator defined by

Z x
(T f )(x) = K(x, y) f (y)dy.
0

Then T satisfies the condition of Theorem 5.2 except that its spectral
radius is zero (and hence its partial spectral radius is zero).
82 Also, if f ∈ C and T f = 0, we have
Z 1
K(1, y) f (y)dy = 0,
0

and so f = 0.
Linear mapping in cones 67

Definition 5.4. Given a normed and partially ordered vector space X

with positive cone K, a non zero linear functional f such that f (x) ≥
0 (x ∈ K) is called a positive continuous linear functional.

The following theorem is quite easily deduced from Theorem 5.1.

Theorem 5.3. Let X be a normed and partially ordered vector space with
a closed positive cone K, and suppose that there exists a subset H of K
with the properties:

(i) Given xεX with ||x|| ≤ 1, there exists hH with −hxh,

(ii) H is contained in a compact set.

Let T be a partially bounded positive linear operator in X with partial

spectral radius µ.

Then there exists a positive continuous linear functional f and a real

number µ∗ with 0 ≤ µ∗ ≤ µ such that

f (T x) = µ∗ f (x) (x ∈ X).

If also K is a normal cone, then µ = µ∗

Proof. Let X ∗ be the dual space of X, and let C be the dual cone K ∗
consisting of all f in X ∗ that satisfy

f (x) ≥ 0 (x ∈ K)

We have seen in Lemma 4.1 (chapter 4) that since K is closed, K ∗ , 0. 83

By condition (i) in the theorem, X = K − K, and therefore K ∗ ∩ (−K ∗ ) =
(0). This proves that C = K ∗ is indeed a positive cone. It is clearly a
closed, and therefore complete, subset of the Banach space X ∗ . We take
E = C − C. 

Let M = sup {||h|| : h ∈ H}, and as usual, let B = { f : f ∈ C and

|| f || ≤ 1}. Given f ∈ B and x ∈ X with ||x|| ≤ 1, there exists h ∈ H with

−h ≤ x ≤ h,
 68 Linear mapping in cones

and therefore −T h ≤ T x ≤ T h,

− f (T h) ≤ f (T x) ≤ f (T h),
| f (T x)| ≤ f (T h) ≤ ||T h|| ≤ p(T )||h|| ≤ p(T ).M,

where p(T ) denotes the partial bound of T . Thus for each f ∈ E we

have an element T ∗ f of E given by

(T ∗ f )(x) = f (T x) (x ∈ X),

and we obtain a partially bounded positive linear operator T ∗ in E with

partial bound p(T ∗ ) satisfying

p(T ∗ ) ≤ M p(T ).

84 similarly
p(T ∗n ) ≤ M p(T n ),
and therefore
µ∗ ≤ µ
where µ∗ denotes the partial spectral radius of T ∗ . We take τ to be the
weak topology in E. Plainly C is τ-closed, B is τ-compact and T ∗ is
τ-continuous. In order to apply Theorem 5.1 it only remains to prove
that τ is sequentially stronger than τN at 0 relative to B. To prove this,
suppose that fn ∈ B(n = 1, 2 . . .) and that 0 is a τ-cluster point of the
sequence ( fn ).
Since H is contained in a (norm) compact set, given ε > 0. there
exists h1 , . . . , hr in H such that for each h ∈ H there is some k(1 ≤ k ≤ r)
with
ε
||h − hk || < (1)
2
Since 0 is a weak ∗-cluster point of ( fn ), there exists an infinite set
Λ of positive integers such that
ε
| fn (hk )| < (k = 1, 2, . . . , r; n ∈ Λ) (2)
2
Therefore by (1) and (2) and the fact that || fn || ≤ 1,

| fn (h)| < ε (h ∈ H, n ∈ Λ) (3)

 Linear mapping in cones 69

Given x ∈ X with ||x|| < 1, there exists h ∈ H with −h ≤ x ≤ h, and

85 so, by (3),
| fn (x)| ≤ | fn (h)| < ε (n ∈ Λ)
Therefore
|| fn || ≤ ε (n ∈ λ),
and 0 is a τN -cluster point of the sequence ( fn ).
Suppose now that K is a normal cone, i.e. for some γ > 0,

||x + y|| ≥ γ||x|| (x, y ∈ K).

Then, for each point x in K,

d(x, −K) = inf ||x + y|| > γ||x||

y∈K

Therefore, for each x in K, there exists f ∈ K ∗ with || f || ≤ 1 and

f (x) ≥ γ||x||

In particular, given ε > 0, there exists x0 ∈ K with ||x0 || ≤ 1, and

||T x0 || > p(T )−

Then there exits f0 ∈ B with

f0 (T x0 ) ≥ γ||T x0 || > γ(p(T ) − ε).

Therefore

||T ∗ f0 || > γ(p(T ) − ε),

and p(T ∗ ) ≥ γp(T ).

86
Since, similarly,

p(T ∗n ) ≥ γp(T n ),
we have µ∗ ≥ µ,

and the proof is complete.

As a special case of Theorem 5.3, we have
 70 Linear mapping in cones

Theorem 5.4 (Krein and Rutman ). Let X be a normed and partially or-
dered over space with a closed normal positive cone K with non-empty
interior. Let T be a positive linear operator in X. Then
(i) T is a bounded linear operator in X.
(ii) There exists a positive continuous linear functional f such that
f (T x) = ρ f (x) (x ∈ X),
where ρ is the spectral radius of T .

Proof. Since K has non-empty interior, there exists a point e of K such

that
||x|| ≤ 1 ⇒ e + x ∈ K
87 since || − x|| ≤ ||x||, we have

||x|| ≤ 1 e ± x ∈ K ⇒ −e ≤ x ≤ e (1)

Thus conditions (i) and (ii) of Theorem 5.3 are satisfied with
H = (e). 

Give a positive linear operator T and ||x|| ≤ 1, we have −e ≤ x ≤ e,

and so −T e ≤ T x ≤ T e. Since K is a normal cone, there exists a positive
constant γ with
||y + z|| ≥ γ||y|| (y, z ∈ K) (2)
We have T e ± T xεK, and so

2||T e|| = ||(T e + T x) + (T e − T x)|| ≥ γ||T e − T x|| ≥ {||T x|| − ||T e||} ,
2+γ
and so ||T x|| ≤ ||T e||,
γ
which proves that T is bounded, and also gives
2+γ
||T || ≤ ||T e|| (3)
γ
It follows from (3) that
2+γ
||T || ≤ ||e|| p(T ),
γ
Linear mapping in cones 71

and similarly, for any positive integer n,

2+γ
||T n || ≤ ||e|| p(T n ),
γ
Therefore 88

ρ = lim ||T n ||1/n ≤ lim p(T n )1/n = µ

n→∞ n→∞

On the other hand it is obvious that µ ≤ ρ, and so the theorem now

follows from Theorem 5.3.
As a special case of Theorem 5.2 we have the following theorem.

Theorem 5.5 (Krein and Rutman theorem 6.1). Let X be a partially or-
dered Banach space with a positive cone K such that X is the closed
linear hull of K. Let T be a compact linear operator in X that maps K
into itself and has a positive spectral radius ρ. Then there exists a non-
zero vector u in K and a positive continuous linear functional f such
that
T u = ρu, T ∗ f = ρ f.

The proof depends on the following lemma concerning compact

liner operators.

Lemma 5.4. Let T be compact linear operator in a normed space X,

and let T have a positive spectral radius. Then there exists a vector x in
X such that
limsup ||T n ||−1 ||T n x|| > 0.
n→∞

Proof. Let ε > 0, and suppose that the lemma is false. Then,for each x
in X there exists a positive integer N x such that
ε
n ≥ N x ⇒ ||T n x|| < ||T n ||
2
89
Also
ε ε
||x′ − x|| < ⇒ ||T n x′ − T n x|| < ||T n ||,
2 2
 72 Linear mapping in cones

ε
and so ||T n x′ || < ε||T n || (n ≥ N x , ||x′ − x|| < )
2
Let S denote the closed unit ball in X.Then T S is compact and so has a
finite covering by open balls of radius 2ε and centers x1 , . . . xm say. Let

N max(N x1 , . . . , N xm ).

Then

||T n x|| < ε||T n || (n ≥ N, x ∈ T S ),

||T n+1 x|| ≤ ε||T n || (n ≥ N, x ∈ S ),

and so
||T n+1 || ≤ ε||T n || (n ≥ N),
from which it follows that lim ||T n ||1/n ≤ ε. 
n→∞

Proof of Theorem 5.5. By Lemma 5.4 and the fact that X = K − K,

there exists a vector x in C with

lim sup ||T n ||−1 ||T n x|| > 0.

n→∞

90 It easily follows from this that

µ lim p(T n )1/n = lim ||T n ||1/n = ρ

n→∞ n→∞

Applying Theorem 5.2 with E = X and C = X we see that there exists a

non-zero vector u in K with

T u = ρu

As in the beginning of the proof of Theorem 5.3, the set K ∗ of continu-

ous linear functions f with

f (x) ≥ 0 (x ∈ K)

is a positive cone in the dual space X ∗ . Also, T ∗ is a compact linear

operator in X ∗ and maps K ∗ into itself. Applying Theorem 5.2 with
Linear mapping in cones 73

E = X ∗ and C = K ∗ , we conclude that there exists a positive continuous

linear functional f with
T ∗ f = µ∗ f,
where µ∗ is the partial spectral radius of T ∗ . Since the spectral radius of
T ∗ is equal to that of T , we have µ∗ ≤ ρ. It only remains to prove that
µ∗ ≥ ρ. There exists u ∈ K with ||u|| = 1 and T u = ρu. We have
T un = ρn u (n = 1, 2, . . .).
Since δ = d(u, −K) > 0, there exists φ in K ∗ with ||φ|| ≤ | and 91
φ(u) = δ. Then
φ(T n u) = φ(ρn u) = ρn δ,
n
and so we have in turn (T ∗ φ)(u) = ρn δ,
n
||T ∗ φ|| ≥ ρn δ,
n
p(T ∗ ) ≥ ρn δ, (n = 1, 2, . . .)
∗
µ ≥ρ
Remark. It is in fact enough in Theorem 5.5 that K be a complete cone
in a normed space (rather than a closed cone in a complete space).
In our next theorem we give a formula for the calculation of positive
eigenvectors corresponding to ρ valid under the conditions of Theorem
5.5. For its proof we shall need some result from the classical Riesz
Schauder theory of compact operators in Banach spaces. We shall state
these results without proof (see Dunford and Schwartz [14]).
Let T be a compact linear operator in a complex Banach space X.
The spectrum σ(T ) is the set of all complex numbers λ for which λI − T
is not a one-to-one mapping of X onto itself. Then σ(T ) is a countable
set contained in the disc |ζ|| ≤ ρ where ρ = lim ||T n ||1/n . Also each
n→∞
element of σ(T ) other than 0 is an eigenvalue, and is an isolated point
of σ(T ), i.e. has a neighbourhood containing no other point of σ(T ). 92
Let λ be a non-zero point of σ(T ). Then there is a positive integer
γ = γ(λ) called the index of λ which is the smallest integer n with the
property that
(T − λI)n+1 x = 0 ⇒ (T − λI)n x = 0.
 74 Linear mapping in cones

[The null space of (T −λI)n increases with n, but eventually we come

to an integer after which it remains constant.] Let

N = x : (T − λI)γ x = 0 ,

and let
M = (T − λI)γ X.
Then N and M are closed subspaces of X and

X=N⊕M (1)

(i.e., each vector is x has a unique expression x = y + z with y ∈ N and

z ∈ M.)
Also M and N are invariant subspaces for T ,

T (M) ⊂ M, T N ⊂ N.

There exist bounded linear projections P and Q orthogonal to each

other and with ranges M and N respectively.

I = P + Q, PQ = QP = 0, P2 = P, Q2 = Q (2)
PX = M, QX = N.

93 Let T M denote the restriction of T to M. Then T M is a compact

linear operator in M and

α , λ, α ∈ σ(T ) ⇐⇒ α ∈ σ(T M ) (3)

Theorem 5.6. Let X, K, T, ρ satisfy the conditions of Theorem 5.5 and

let γ be the index of ρ. Let Q be the projection onto the null space of
(T − ρI)γ which is orthogonal to the range of (T − ρI). Then
!−1
−n n
(i) lim (1 + ρ) (I + T )n = (1 + ρ)1−ν (T − ρI)−1 Q,
n→∞ γ−1

(ii) lim ||(I + T )n ||−1 (I + T )n = ||(T − ρI)ν−1 Q||−1 (T − ρI)−1 Q,

n→∞

the convergence being with respect to the operator norm.

Linear mapping in cones 75

Proof. Since there is a natural isometry between bounded linear oper-

ator in a real Banach space and their complexifications, and since this
isometry preserves compactness of operators, there is no loss of gener-
ality in supposing that X is a complex Banach space. 

Taking λ = ρ in the above considerations, we have continuous linear

projections P, Q on to the range M and null space N of Aγ , γ being the
index of ρ, where A = T − ρI, and (1), (2), (3) hold.
Suppose α is a non-zero point in the spectrum of P + T P. We prove
that α − 1 is in the spectrum of T M (the restriction of T to the subspace
M.) In fact
S = αI − (P + T P)
is not a (1 − 1) mapping of X onto itself. Either 94

i) the mapping S is not (1 − 1),

or
ii) the range of the mapping S in not the whole of X.
In case (i) there exists a non-zero vector x in X with

(P + T P)x = αx.

It follows that x ∈ PX = M, x = Px, and so

(I + T )x = αx, T x = (α − 1)x, α − 1 ∈ σ(T M ).

In case (ii), since S M ⊂ M, S N ⊂ N and X = M+N, either S M , M

or S N , N. But as P is zero on N, and α , 0, we have S N = N. Hence
S M , M. But
S x = {(α − 1)I − T }x, (x ∈ M)
and so α − 1 ∈ σ(T M ). It follows from this and (3) that if α is a non-zero
point of the spectrum of P + T P, then

α − 1 ∈ σ(T ) (ρ).

Therefore

lim ||(P + T P)n ||1/n ≤ sup {|1 + ζ||ζ ∈ σ(T ), ζ , ρ}

n→∞
 76 Linear mapping in cones

Since all points ζ of σ(T ) satisfy |ζ| ≤ ρ, and since ρ is as isolated point 95
of σ(T ), it follows that

lim ||(P + T P)n ||1/n = k < 1 + ρ

n→∞

since T M ⊂ M, we have PT P = T P, and so

lim ||(I + T )n P||1/n = k.

n→∞

We choose η with k < η < 1 + ρ. Then there exists n0 with

||(I + T )n P|| < ηn (n ≥ n0 ) (1)

We have

I + T = (1 + ρ)I + A
and An Q = 0 (n ≥ ν).
Hence
( ! ! )
n n n n−1 n n−ν+1 ν−1
(I + T ) Q = (1 + ρ) I + (1 + ρ) A + · · · + (1 + ρ) A Q
1 −1

It follows that
!−1
−n n
lim (1 + ρ) (I + T )n Q = (1 + ρ)1−ν Aν−1 Q (2)
n→∞ ν−1
Also (1) gives,
!−1
−n n
lim (1 + ρ) (I + T )n P = 0 (3)
n→∞ ν−1
96 and, since (I + T )n = (I + T )n P + (I + T )n Q, (2) and (3) give
!−1
n
lim (1 + ρ)−n (I + T )n = (1 + ρ)1−ν Aν−1 Q (4)
n→∞ ν−1
Taking norms, we have
!−1
−n n
lim (1 + ρ) ||(I + T )n || = (1 + ρ)1−ν ||Aν−1 Q|| (5)
n→∞ ν−1
Linear mapping in cones 77

By definition of ν, the null space of Aν−1 is not the whole of N, and

so Aγ−1 Q , 0. Thus (4) and (5) combine to give (ii).
Further results connected with Theorem 5.5 are given by Krein and
Rutman [20]. In particular very precise results are proved (Theorem 6.3
[20]) for an operator T which satisfies the conditions of Theorem 5.5
and also maps each non-zero point of K into the interior of K. In this
case ν = 1, and the result of Theorem 5.6 takes the specially simple
from
lim (1 + ρ)−n (I + T )n = Q.
n→∞

Here Q is a dimensional operator

Qx = f (x)u,

where u is a positive eigenvector of T, f a positive eigenvector of T ∗

normalized by taking f (u) = 1.
Chapter 6

Self-adjoint linear operator

in a Hilbert space

It would be foolish of me to attempt to give in these lectures an account 97

of the many methods that have been developed for the study of the spec-
tral resolution of a self adjoint operator. I shall limit myself to giving an
account of a certain explicit formula for the projections belonging to the
spectral resolution. In general this is a theorem about projections; in the
case of a compact operator it becomes a theorem about eigenvectors.

Definition 6.1. A complex (real) Hilbert space is a vector space over

C(R) with a mapping
H × H → C(R)
called the scalar product and denoted by (x, y) which satisfies the fol-
lowing axioms

i) (x, y) = (y, x)

ii) (x1 + x2 , y) = (x1 , y) + (x2 , y) (x1 , x2 , y ∈ H)

iii) (αx, y) = α(x, y) (x, y ∈ H, α ∈ C(R))

iv) (x, y) > 0 for x , 0; (x, x) = 0 for x = 0 (x ∈ H)

1
v) H is a Banach space with the norm ||x|| = (x, x) 2 .

79
 80 Self-adjoint linear operator in a Hilbert space

Let H be a real or complex Hilbert space, and let S denote the class
of all bounded symmetric operators in H, i.e., bounded linear mappings
of T into itself such that

(T x, y) = (x, T y) (x, y ∈ H)

98 S is the class bounded self adjoint operators, T = T ∗ . A relation ≤ is

introduced into S by writing A ≤ B or B ≥ A to denote that

(Ax, x) ≤ (Bx, x) (x ∈ H).

In particular, the operators T belonging to S that satisfy

T ≥0

are called positive operators .

Note that this definition of ‘positive’ has nothing to do with the prop-
erty of mapping a cone into itself that we have considered in earlier
chapters.
We establish a few elementary properties of the relation. First we
recall a few obvious properties of commutants.
Let B denote the class of all bounded linear operators in H, and
given E ⊂ B, let

E ′ = {T : T ∈ B and AT = T A(A ∈ E)}

E ′ is called the commutant of E.

(i) E1 ⊂ E2 ⇒ E2′ ⊂ E1′ (obvious)

(ii) E ′ is strongly closed.

Let T n ∈ E ′ converge strongly towards T ∈ B. Then T n A =
AT n (n = 1, 2, . . .)

T A x = lim T n Ax = lim AT n x = AT x (x ∈ H)
n→∞ n→∞

99
Self-adjoint linear operator in a Hilbert space 81

(iii) E self-adjoint ⇒ E ′ self-adjoint.

T ⊂ E ′ ⇒AT = T A (A ∈ E)
∗ ∗
A T = TA (A ∈ E)
∗ ∗
T A = AT (A ∈ E)

(iv) E ′ is a complex linear algebra (obvious)

(v) E = (T ) ⇒ E ′′ is a strongly closed commutative algebra contain-

ing T .

(T ) ⊂ (T )′ ⇒ (T )′′ ⊂ (T )′
A1 , A2 ∈ (T )′ ⇒ A1 ∈ (T )′′ , A2 ∈ (T )′ A1 A2 = A2 A1 .

We next establish some well known elementary propositions con-

cerning positive operators.

(a) For any positive operator T , the generalized Schwartz inequality

holds i.e.,
|(T x, y)|2 ≤ (T x, x)(T y, y)

Proof. B(x, y) = (T x, y) is a positive semi-definite symmetric bi-

linear form and so the generalized Schwarz inequality for this form.


(b) If T is a positive operator, then

||T || = sup{(T x, x) : ||x|| ≤ 1}

Proof. Let T be a positive operator and let 100

M = sup{(T x, x) : ||x|| ≤ 1}

By the Schwarz inequality

|(T x, x)| ≤ ||T x|| ||x||,

 82 Self-adjoint linear operator in a Hilbert space

and so M ≤ ||T ||. On the other hand putting y = T x in the general-

ized Schwarz inequality, we have

||T x||4 = (T x, T x)2 = |(T x, y)|2 ≤ (T x, x)(T y, y) ≤ M 2 ||x||2 ||T x||2 ,

and so ||T || ≤ M. 

(c) The set of positive operators is a positive cone in S .

Proof. If T and −T are both positive, we have

(T x, n) = 0 (x ∈ H)

By (b), this gives T = 0. The other properties of the cone are obvi-
ous. 

(d) Let (T n ) be a bounded increasing sequence of elements of S , i.e.,

T n ≤ T n+1 ≤ M.I (n = 1, 2, . . .)

Then (T n ) converges strongly to an element T of s i.e.

lim T n x = T x (x ∈ H)
n→∞

101 Proof. For m < n, let Am,n = T n − T m . By the generalized Schwartz

inequality (a) with T = Am,n and y = Am,n x, we have ||Am,n x||4 =
(Am,n x, Am,n x)2 = |(Am,n x, y)|2 ≤ (Am,n x, x)(Am,n y, y). Since 0 ≤ Am,n ≤
MI, we have (Am,n y, y) ≤ M 3 ||x||2 . Hence ||T n x − T m x||4 ≤ M 3 ||x||2
 
T n x, x) − T m x, x) . Since the sequence (T n x, x) is a bounded increas-
ing sequence of real numbers, it follows that (T n x) is a Cauchy sequence
which converges to an element T x ∈ H in view of the axiom (v) in defi-
nition 6.1.

(e) T ≥ 0 ⇒ T n ≥ 0 (n = 1, 2, . . .).


Self-adjoint linear operator in a Hilbert space 83

Proof. (T 2K x, x) = (T K x, T K x) ≥ 0

(T 2k+1 x, x) = (T.T x , T k x) ≥ 0.
1
(f) Each positive operator T is the square of a positive operator T 2 , and
1
T 2 belongs to the second commutant (T )′′ of T .

Proof. Suppose that 0 ≤ A ≤ I, and let B = I − A, so that also

0 ≤ B ≤ I.

Let Yn be the sequence defined inductively by

1
Y0 = 0, Yn+1 = (B + Yn2 ) (n = 0, 1, 2, . . .).
2
By induction we have ||Yn || ≤ 1, and so

0 ≤ Yn ≤ I.

Also, since 102

1 2 2 1
Yn+1 − Yn = (Yn − Yn−1 ) = (Yn + Yn−1 )(Yn − Yn−1 ),
2 2
we see by induction that Yn+1 −Yn is a polynomial in B with non-negative
real coefficients. Since Bn ≥ 0, for every n, it follows that (Yn ) is an in-
creasing sequence. Hence (Yn ) converges strongly to a positive operator
Y, and we have
1
0 ≤ Y ≤ I, Y = (B + Y 2 )
2
Let X = I − Y. Then X is a positive operator and

X 2 = A.

If a bounded linear operator commutes with A, it commutes with

each polynomial in A, hence it commutes with Yn , and therefore with X.
 84 Self-adjoint linear operator in a Hilbert space

1
If 0 ≤ T ≤ MI, A = .T satisfies 0 ≤ A ≤ I and so the proposition
M
holds for T .
1
The positive square root T 2 is in fact unique but we do not need this
fact.

(g) A ≥ 0, B ≥ 0, AB = BA ⇒ AB ≥ 0.
1
Proof. Since A ∈ (B′ ), we have B 2 ∈ (A)′ and so
1 1 1 1
AB = AB 2 B 2 = B 2 AB 2 .


1 1 1 1
Finally, (B 2 AB 2 x, x) = (AB 2 x, B 2 x) ≥ 0.
(h) T ≥ 0 I + T is invertible, (I + T −1 ) ≥ 0, and

(I + T )−1 ∈ (T )′′ .

103

Proof. We have

I ≤ I + T ≤ (1 + M)I,
1
≤ A ≤ I,
1+M

1
where A = (I + T ). Therefore
1+M
!
1 M
||I − A|| ≤ || 1 − I|| = < 1.
1+M 1+M


Therefore the Neumann series

I + (I − A) + (I − A)2 + · · ·
Self-adjoint linear operator in a Hilbert space 85

converges in operator norm and since (I − A)k ≥ 0 its sum is a positive

operator B. We have

I = B(I − A) = I + (I − A)B = B,

and so AB = BA = I.
Finally (1 + M)−1 B is the required positive inverse of I + T . By a
projection we mean an operator P belonging to S with P2 = P.

(i) Each projection P satisfies 0 ≤ P ≤ I.

Proof. Since P ∈ S and P = P2 , we have P ≥ 0. 

Since I − P ∈ S and (I − P)2 = I − P we have I − P ≥ 0.

(j) For projections P1 , P2

P1 ≥ P2 ⇐⇒ P2 = P2 P1 ⇐⇒ P2 = P1 P2

104

Proof. P1 P2 = P2 ⇒ P2 = P∗2 = (P2 P1 )∗ = P∗1 P∗2 = P1 P2

P1 P2 = P2 P2 = P∗2 = (P2 P1 )∗ = P∗2 P∗1 = P2 P1 .

Thus, if P2 = P2 P1 , we also have P1 P2 = P2 , and so

(P1 − P2 )2 = P21 − P1 P2 − P2 P1 + P22 = P1 − P2 ,

and therefore P1 ≥ P2 .
Finally suppose that P1 ≥ P2 . If P1 x = 0, then P2 x = 0, for

(P2 x, P2 x) = (P22 x, x) = (P2 x, x) ≤ (P1 x, x) = 0.

Since P1 (I − P1 ) = 0,

P2 (I − P1 ) = 0,
i.e., P2 = P2 P1 .
 86 Self-adjoint linear operator in a Hilbert space

Lemma 6.1. Let A ≥ 0, and let B = 2A2 (I + A2 )−1 . Then

i) B ∈ A′′ ,
ii) 0 ≤ B ≤ A,
iii) I − B = (I − A)(I + A)(I + A2 )−1 ,
iv) if P is a projective permutable with A and P ≤ A, then P ≤ B.
Proof. Proposition (h) implies (i). 

105 That B ≥ 0 is clear since A2 and (I + A2 )−1 are permutable. Also

(I + A2 )(A − B) = A + A3 − 2A2 = A(I − A)2 ≥ 0,

and so, using the permutability of the operators,

A − B = (I + A2 )−1 (I + A2 )(A − B) ≥ 0

This proves (ii), and (iii) is straight forward.

Let P be a projection such that P ∈ A′ and P ≤ A. We have

P = P2 ≤ PA ≤ A2 ,

and therefore
P = P2 ≤ A2 P.
Therefore

(I + A2 )(B − P) = 2A2 − (I + A2 )P ≥ 2A2 − 2A2 P

= 2A2 (I − P) ≥ 0

Finally, since (I + A2 )−1 is permutable with all the operators concerned,

B − P ≥ (I + A2 )−1 2A2 (I − P) ≥ 0.

Theorem 6.1. Let A be a positive operator, and let the sequence (An ) be
defined inductively by

A1 = A, An+1 = 2A2n (I + A2n )−1 (n = 1, 2, . . .)

106 Then
Self-adjoint linear operator in a Hilbert space 87

i) 0 ≤ An+1 ≤ An (n = 1, 2, . . .),

ii) the sequence (An ) converges strongly to a projection Q belonging

to A′′ .

iii) Q ≤ A,

iv) (I − A)(I − Q) ≥ 0,

v) Q is maximal in the sense if P is a projection permutable with A

and satisfying P ≤ A, then P ≤ Q.

Proof. (i) This follows at once from Lemma 1. (ii) and (iii). It follows
from (i) and Proposition (d) that (An ) converges strongly to a positive
operator Q with Q ≤ A, and that Q ∈ (A)′′ . It remains to prove that Q is
a projection. 

Since 0 ≤ An ≤ A, we have

||An || ≤ ||A|| (n = 1, 2, . . .);

and therefore, since

lim An x = Qx (x ∈ H),
n→∞

we have in turn,

lim A2n x = Q2 x (x ∈ H),

n→∞

lim An+1 (I + Q2 )x − (I + A2n )x = 0 (x ∈ H),
n→∞
lim An+1 (I + Q2 )x = lim 2A2n x = 2Q2 x (x ∈ H).
n→∞ n→∞

But 107

lim An+1 (I + Q2 )x = Q(I + Q2 )x (x ∈ H),

n→∞
and so Q(I + Q2 ) = 2Q2 .

Therefore
(Q − Q2 )2 = 0.
 88 Self-adjoint linear operator in a Hilbert space

But since Q − Q2 is symmetric, this gives

(Q − Q)2 = 0,

i.e., Q is a projection.
(iv) By Lemma 1 (iii),

(I − An ) = (I − An−1 )(I + An−1 )(I + A2n−1 )−1 ,

Y
n−1
= (I − A) (I + Ak )(I + A2k )−1 ;
k=1

and so (I − A)(I − An ) ≥ 0 (n = 1, 2, . . .),

which gives (I − A)(I − Q) ≥ 0.

(v) Let P be a projection permutable with A such that P ≤ A.

108 By repeated application of Lemma 1 (iv), P is permutable with An
and P ≤ An . In the limit we have P ≤ Q.

Theorem 6.2. Let T be a bounded symmetric operator with

mI ≤ T ≤ MI.

Let E M = I, and for λ < M let Eλ be the projection Q of Theorem 1

corresponding to A given by
1
A= (MI − T )
M−λ
Then

i) the projections E belongs to (T )′′ ;

ii) Eλ = 0 (λ < m), E M = I;

iii) Eλ ≤ Eµ (λ ≤ µ);

iv) λ(Eµ − Eλ ) ≤ T (Eµ − Eλ ) ≤ µ(Eµ − Eλ ) (λ ≤ µ);

v) the maily (Eλ ) is strongly continuous on the right.

Self-adjoint linear operator in a Hilbert space 89

Proof. i) follows from Theorem 1 (ii).

ii) If Eλ , 0, there exists a non-zero x0 with Eλ x0 = x0 .



By Theorem 1 (iii), we have

1
Eλ ≤ (MI − T ),
M−λ
and so
1
(x0 , x0 ) = (Eλ x0 , x0 ) ≤ ((MI − T )x0 , x0 ).
M−λ
Since MI − T ≤ (M − m)I, this gives

M−m
(x0 , x0 ) ≤ (x0 , x0 ),
M−λ
and so λ ≥ m. This proves (ii). 109

(iii) This is obvious except when λ < µ < M. In this case, since
1
Eλ is a projection permutable with (MI − T ), and
M−µ

1 1
Eλ ≤ (MI − T ) ≤ (MI − T ),
M−λ M−µ
Then 1 (v) shows that
Eλ ≤ Eµ

(iv) Suppose that λ < µ ≤ M. By Theorem 1 (iv),

!
1
I− (MI − T ) (I − Eλ ) ≥ 0,
M−λ
i.e., (I − λI)(I − Eλ ) ≥ 0.

Since Eλ ≤ Eµ , we have Eλ Eµ = Eλ , and so this gives

(T − λI)(Eµ − Eλ ) ≥ 0,
 90 Self-adjoint linear operator in a Hilbert space

which is the left hand inequality in (iv). The right hand inequality is
obvious if µ = M ( since T ≤ MI); and if µ < M we have
1
Eµ ≤ (MI − T ),
M−µ
i.e., T ≤ MI − (M − µ)Eµ .
Since Eµ (Eµ − Eλ ) = Eµ − Eλ , this gives.
T (Eµ − Eλ ) ≤ µ(Eµ − Eλ ),
110 and (iv) is proved.
(v) Suppose µ < M. If (Eλ ) is not strongly continuous on the right at
µ, there exists a sequence (λn ) convergent decreasing to µ but such
that Eλn does not converge strongly to Eµ .
Since (Eλn ) is a decreasing sequence of operators, with Eλn ≥ Eµ ,
there exists a positive operator J such that J ≥ Eµ and (Eλn ) converges
strongly to J. Then J is a projection, permutable with T , and
1
Eµ ≤ J ≤ Eλn ≤ (MI − T ) (n = 1, 2, . . .)
M − λn
It follows that
1
Eµ ≤ J ≤ (MI − T ),
M−µ
and so by the maximal property of Eµ ,
J ≤ Eµ .
This completes the proof of the theorem.
Corollary. (The spectral theorem).
ZM
T = λdEλ (ε > 0)
M−ε

The integral being the uniform limit of its Riemann-Stieltjes sums. In

fact let
m − ε = λ0 < λ1 < · · · < λn = M.
111
Self-adjoint linear operator in a Hilbert space 91

Then

λk−1 (Eλk − Eλk−1 ) ≤ T (Eλk − Eλk−1 ) ≤ λk (Eλk − Eλk−1 )

and so since Eλn = I and Eλ0 = 0,

X
n X
n
λk−1 (Eλk − Eλk−1 ) ≤ T ≤ λk (Eλk − Eλk−1 ),
k=1 k=1
X
n X
n
0≤T− λk−1 (Eλk − Eλk−1 ) ≤ (λk − λk−1 )Eλk − Eλk−1 )
k=1 k=1
≤ max(λk − λk−1 )I

Hence
X
n
||T − λk−1 (Eλk − Eλk−1 )|| → 0 as max(λk − λk−1 ) → 0
k=1

Moreover
ZM
r
T = λr dEλ (r = 0, 1, 2, . . .)
m−ε
To see this we rewrite (iv) in the form

(M − µ)(Eµ − Eλ ) ≤ (MI − T )(Eµ − Eλ ) ≤ (M − λ)(Eµ − Eλ )

since MI − T ≥ 0, M − λ ≥ 0, and M − µ ≥ 0, it follows that

(M − µ)r (Eµ − Eλ ) ≤ (MI − T )r (Eµ − Eλ ) ≤ (M − λ)r (Eµ − Eλ )

Therefore, as in the preceeding argument, 112

ZM
r
(MI − T ) = (M − λ)r dEλ (r = 0, 1, 2, . . .)
m−ε

and the required result follows by induction.

In the next theorem we ∞nsider the special simplification which
occurs when the operator is also compact. We need a simple lemma.
 92 Self-adjoint linear operator in a Hilbert space

Lemma 6.2. Let A be a positive operator, and let (An ) be the sequence
constructed as in Theorem 7. Then An = A2 Bn (n = 2, 3, . . .), where
each Bn belongs to (A)′′ and

0 ≤ Bn+1 ≤ Bn (n = 2, 3, . . .)

Proof. We have
A2 = 2A2 (I + A2 )−1 = A2 B2 ,
with B2 = 2(I + A2 )−1 . If An = A2 Bn with Bn ≥ 0 and Bn ∈ (A)′′ , then

An+1 = 2A4 B2n (I + A4 B2n )−1

= A2 Bn+1 ,

with Bn+1 = 2A2 B2n (I + A4 B2n )−1 . Then


Bn − Bn+1 = (I + A4 B2n )−1 Bn (I + A4 B2n ) − 2A2 B2n
= (I + A4 B2n )−1 Bn (I − A2 Bn )2 ,

113 so that 0 ≤ Bn+1 ≤ Bn . 

Theorem 6.3. Let A be a compact positive operator, and let (An ) and Q
be the corresponding and projection defined as in Theorem 1. Then

(i) (An ) converges uniformly to Q;

(ii) Q has finite rank;

(iii) the range of Q is spanned by eigenvectors of A corresponding to

eigenvalues λ with λ ≥ 1, and all such eigenvectors lie in the range
of Q.

Proof. Let Bn be the sequence defined in Lemma 2. Then (Bn ) con-

verges strongly to an operator B in (A)′′ . Since An = A2 Bn and An
converges strongly to Q, we have

Q = A2 B.


Self-adjoint linear operator in a Hilbert space 93

Let K be the unit ball in the Hilbert space H, and let E = (AK).
Then E is a norm compact set. Since lim Bn x = Bx (x ∈ H), we have
n→∞

lim (Bn x, x) = (Bx, x) (x ∈ E).

n→∞

With respect to the norm topology in E, the functions (Bn x, x) are

continuous real functions converging decreasingly to the continuous real 114
function (Bx, x). Therefore, by Dini’s theorem, the convergence is uni-
form on E. Since
(An x, x) = (A2 Bn x, x) = (Bn Ax, Ax)
It follows that (An x, x) converges uniformly on K to (BAx, Ax) =
(A2 Bx, x) = (Qx, x). Therefore An → Q uniformly (i.e., with respect to
the operator norm).
Since Q = A2 B and A is compact, we know that Q is compact, and
therefore has finite rank, i.e., its range HQ is finite dimensional; for Q
is the identity operator in the Banach space HQ and a Banach space in
which a ball is compact is finite -dimensional.
The rang HQ of Q is a finite dimensional Hilbert space and A maps
Q into itself (since QA = AQ). By the elementary theory of symmetric
matrices, HQ is spanned by eigenvectors u1 , . . . , ur with real eigenvalues
λ1 , . . . , λr ,
Ai ui = λi ui
Since Qui = ui , and A ≥ Q, we have
λi (ui , ui ) = (Aui , ui ) ≥ (Qui , ui ) = (ui , ui ),
and so λi ≥ 1.
Conversely, let u be an eigenvector of A with Au = λu, λ ≥ 1. We
may suppose that ||u|| = 1 and then define a projection P by taking 115

Px = (x, u)u.
APx = (x, u)Au = λ(x, u)u = (x, λu)u = (x, Au)u = (Ax, u) = PAx. Also
P ≤ A, for given x ∈ H, we have x = H, we have x = ξu + v with
(u, v) = 0. Then Ax = λξu + Av, and
(u, Av) = (Au, v) = λ(u, v) = 0. So
94 Self-adjoint linear operator in a Hilbert space

(Ax, x) = λ|ξ|2 (u, u) + (Av, v)2

≥ λ|ξ|2 (u, u) = λ(Px, x) ≥ (Px, x).

By the maximal property of Q, P ≤ Q. Hence P = QP, u ∈ HQ .

The iterative method, given here can also be applied to construct
the projections belonging to the spectral family of an unbounded self-
adjoint operator, and details of this may be found in my paper.
Chapter 7

Simultaneous fixed points

In this brief chapter we are concerned with the existence of a simultane- 116
ous fixed point of a family F of mappings

T u = u (T ∈ F ).

We first state without proof two well known theorems on this question
proofs of which will be found in Dunford and Schwartz [14] pp.456-
457. We then prove a theorem on families of mappings of a ∞ne into
itself. Some further results in the present context will appear in the next
chapter in the theory of a special class of semialgebras.

Theorem 7.1 (Markov- Kankutani). Let K be a compact convex subset

of a Hausdorff linear topological space. Let F be a commuting family
of continuous affine mappings of K into itself. Then there exists a point 117
u in K with
T u = u (T ∈ F )

A mapping T is said to be affine if

T (αx + (1 − α)y) = αT x + (1 − α)T y (x, y ∈ K, 0 ≤ α ≤ 1)

Theorem 7.2 (Kakutani). Let K be a compact convex subset of a Haus-

dorff locally convex space and let G be a group of linear mappings which

95
 96 Simultaneous fixed points

is equi-continuous on K and satisfies GK ⊂ K. Then there exists u in K

with
T u = u (T ∈ G)
G is said to be equi-continuous on K if given a neighbourhood V of 0,
there exists a neighbourhood U of 0 such that k1 , k2 ∈ K, k1 − k2 ∈ U
implies that T k1 − T k2 ∈ V(T ∈ G).

We have chosen to include a theorem on mappings of cones into

themselves in this chapter because of the opportunity it gives to employ
a method of proof quite different from the methods used in the rest of
these lectures. The method depends on the concept of ideal in a partially
ordered vector space.

Definition 7.1. Let E be a partially ordered vector space with positive

cone C. A subset J of E is called an ideal if
(i) J is a linear subspace
(ii) j ∈ J ⇒ [0, j] ⊂ J,

Here [a, b] = x : a ≤ x ≤ b ; and conditions (i) and (ii) are equivalent
to (i) and
(ii)′ j ∈ J ⇒ [− j, j] ⊂ J
For let j ∈ J and − j ≤ x ≤ j. Then 0 ≤ x + j ≤ 2 j. By (i), 2 j ∈ J
and by (ii), x + j ∈ J. Again by (i), x ∈ J.

Example . If E◦ is a linear subspace of E with E◦ C = (0). Then E◦ is

an ideal. ( For then [ 0, j ] for j ∈ J is empty unless j = 0).

118 Definition 7.2. An element e of C is called an order unit if [−e, e] ab-

sorbs all points of E, i.e., if for every x in E
−λe ≤ x ≤ λe
for all sufficiently large λ.

By definition of a positive cone, C contains non -zero elements, and

so certainly
e > 0.
Simultaneous fixed points 97

We establish a few elementary properties of ideals. An ideal J is

proper if 0 , J , E.

a) An element of C is an order unit of E if and only if it is not contained

in any proper ideal of E.

Proof. Let J be an ideal containing an order unit e. Then λe ∈ J and

[
E⊂ [−λe, λe] ⊂ J
λ>0

Hence E = J is not a proper ideal. 

Conversely, if e ∈ C is not contained in any proper ideal, then the

S
ideal [−λe, λe] which contains e is the whole of E i.e., e is an order
λ>0
unit.

(b) If E has an order unit, then each proper ideal of E is contained in a

maximal proper ideal of E.

Proof is immediate using (a) and Zorn’s lemma.

(c) If E has an order unit e, and J is a proper ideal of E, then E/J is a 119
partially ordered vector space with order unit.

Proof. Since J is a linear subspace, E/J is a vector space whose ele-


ments are the cosets x̃ = x + J, x ∈ E. The set C̃ = x̃ : x ∈ C is a
positive cone in E/J. It contains the non-zero element ẽ, and x̃, − x̃ ∈ C̃
implies that there exist j, j′ ∈ J such that

x + j, − x + j′ ∈ C.

This gives

0 ≤ x + j ≤ j′ + j
and so x + j ∈ [0, j + j] ⊂ J,
x ∈ J, x̃ = 0.
 98 Simultaneous fixed points

Finally ẽ is an order unit. For

−λe ≤ x ≤ λe
implies that − λẽ ≤ x̃ ≤ λẽ

which proves that ẽ is an order unit since x → x̃ is an ‘onto’ mapping.



120 (d) If E has an order unit, and M is a maximal proper ideal of E, then
E/M has no proper ideals.

Proof. If J˜ is an ideal of E/M, then


J = x : x̃ ∈ J˜

is an ideal of E containing M. Hence J = M or J = E i.e. J˜ = (0) or

J˜ = E/M. 

(e) If E has no proper ideals then E ≃ R.

Proof. C contains a non-zero element e. Since there are no proper ide-

als, e is an order unit. 

Let
p(x) = inf[ξ : x ≤ ξe] (x ∈ E) (1)
Let y = p(x)e − x.
If x ∈ E, then either x ∈ C or −x ∈ C, for otherwise (x) is an ideal.
Hence y ∈ C or −y ∈ C i.e. y ≥ 0 or y < 0. If y < 0 for some x, then
−y ≥ εe for some ε > 0, and x ≥ (p(x) + ε)e which contradicts (1). If
y > 0, for some x ∈ E, then y is an order unit and so

y ≥ εe for some ε > 0

But then (p(x) − ε)e ≥ x which also contradicts (1). Hence

y = p(x)e − x = 0 x ∈ E)

121 i.e. e , 0 spans the whole space E and so E is isomorphic to R.

Simultaneous fixed points 99

(f) Let E have an order unit e, and let M be a maximal proper ideal.
Then there exists a linear functional f on E with

(i) f (x) ≥ 0 (x ∈ C),

(ii) f (e) = 1,
(iii) f (x) = 0, (x ∈ M).

(i.e., M is the null space of a normalised positive linear func-

tional).
Proof. E/M has no proper ideals, and ẽ > 0 where x → x̃ is the canoni-
cal mapping E → E/M. By (e), for each x̃, there exists a real number ξ
with
x̃ = ξ ẽ
Let f (x) = ξ. Then it is easily verified that f is a linear functional
with the required properties. 

(g) Let E have an order unit and have dimension greater than one, and
let T be a positive linear mapping of E into itself.
Then there exists a proper T -invariant ideal, i.e. a proper ideal J
with T J ⊂ J.

Proof. Let e be the order unit, let p be the Minkowski functional of

[−e, e], and let
N = {x : p(x) = 0}.
The set N is an ideal in E. For if j ∈ N then 122

−ε ≤ j ≤ εe for all ε > 0

Hence, if 0 ≤ x ≤ j, then

−εe ≤ x ≤ εe (ε > 0)

and so p(x) = 0 and x ∈ N. N , E since p(e) = 1. Also T N ⊂ N. For

x ∈ N, we have

−εe ≤ x ≤ εe (ε > 0)
 100 Simultaneous fixed points

and so − εT e ≤ T x ≤ εT e (ε > 0)

But

−αe ≤ T e ≤ αe for some α>0

so that − εαe ≤ T x ≤ εαe for some α > 0 and all ε > 0

Hence
αp(T x) = 0 i.e. T x ∈ N.


Thus if N , (0), it is a T -invariant proper ideal. Suppose N = (0), so

that p is a norm. Let H denote the set of all positive linear functionals
f with f (e) = 1. As E is of dimension greater than one, by (e), (b)
and ( f ), H is nonempty. If for some f ∈ H, we have f (T e) = 0, then
f (T x) = 0 for all x ∈ E and so the null-space of f is a proper T-invariant
ideal. Suppose then that f (T e) , 0 for all f in H. Clearly H is a convex
weak * closed subset of the dual space E ∗ of the normed space (E, p).
Also H is contained in the unit ball of E ∗ .
123 For we have

−p(x)e ≤ x ≤ p(x)e (x ∈ E)
so that − p(x) f (e) ≤ f (x) ≤ p(x) f (e) (x ∈ E, f ∈ H)
i.e., − p(x) ≤ f (x) ≤ p(x) (x ∈ E, f ∈ H)
or | f (x)| ≤ p(x)

Hence H is weakly compact.

As
−p(x)T e ≤ T (x) ≤ p(x)T (e) (x ∈ E),
T is a bounded linear transformation of (E, p) into itself with ||T || ≤
T (e). Therefore its transpose T ∗ is a weak* continuous mapping of E ∗
into itself. Thus the mapping S defined by

1
Sf = T∗ f
f (T e)
Simultaneous fixed points 101

is a weak* continuous mapping of the convex, weak* compact subset H

of E ∗ into itself. By the Schander-Tychnoff fixed point theorem S has a
fixed point f0 in H, and the null space of f0 is a proper T -invariant ideal.

(h) Under the condition of (g) there exists a maximal proper ideal M
and a non-negative real number such that

T x − µx ∈ M (x ∈ E).

Proof. By (g) and Zorn’s lemma, there exists a maximal proper T -

invariant ideal M. In fact M is a maximal proper ideal, for otherwise 124
E/M has dimension greater than 1 and so there is a proper T̃ -invariant
ideal M̃1 in E/M, where T̃ is the mapping on E/M given by

T̃ x̃ = T̃ x,

x → x̃ being the canonical mapping E → E/M. Then the inverse image

M1 of M̃1 by this mapping is a proper T -invariant ideal containing M
strictly which contradicts the definition of M.
The maximal proper ideal M is the null space of a normalised posi-
tive linear functional f .
Since f (e) = 1, we have

x − f (x)e ∈ M (x ∈ E),
and so T x− = f (x)T e ∈ M (x ∈ E)
f {T x − f (x)T e} = 0 (x ∈ E)

Writing µ = f (T e), we have

f (T x − µx) = 0 (x ∈ E),

and so T x − µx ∈ M (x ∈ E). 

Theorem 7.3. Let E be a partially ordered vector space with an order

unit e and with dimension greater than one. F , be a commuting family
of positive linear mappings of E. Then there exists a maximal proper 125
ideal which is T -invariant for all T in.
102 Simultaneous fixed points

Proof. We prove that there is a proper ideal that is F invariant (i.e. T -

invariant for every T ∈ F ), and then the proof is completed by applying
Zorn’s lemma as in (h). 

If every T ∈ F is a constant multiple of the identity mapping, then

this assertion is obvious, for every proper ideal is then F -invariant. Sup-
pose then that T 0 ∈ F is not a constant multiple of the identity. Then by
(h), there exists a maximal proper ideal M and a constant real number
such that
T 0 x − µx ∈ M (x ∈ E).
Let E0 = {T 0 x − µx : x ∈ E}. Then E0 is a proper subspace of E and
since
T (T 0 x − µx) = T 0 (T x) − µ(T x) (T ∈ F ),
E0 is F -invariant. If E0 ∩ C = (0) then E0 is the required proper F -
invariant ideal.
Otherwise, let [
J= [−y, y]
y∈E0

Then (0) , J ⊂ M, and J is an F invariant ideal.

Corollary . There exists a normalized positive linear functional f such

that
f (T x) = f (T e) f (x) (T ∈ F , x ∈ E).
Chapter 8

A class of abstract
semi-algebras

The present chapter is somewhat of an intruder in this course of lectures. 126

It has some incidental bearing on the Perron-Frobenius theorem, but our
main purpose is to establish some algebraic properties of a certain class
of semi-algebras.

Definition 8.1. A real Banach algebra is a linear associative algebra over

R together with a norm under which it is a Banach space and which
satisfies
||xy|| ≤ ||x|| ||y|| (x, y ∈ B)

Definition 8.2. A non-empty subset A of a real Banach algebra B is

called a semi-algebra if
(i) x, y ∈ A, α ≥ 0 ⇒ x + y, αx ∈ A and xy ∈ A.
A semi-algebra A is called a locally compact semi-algebra if it
satisfies the additional axioms
(ii) A contains non zero vectors;
(iii) the set of elements x of A with ||x|| ≤ 1 is a compact subset of B.
It is easily seen that if A is a locally compact semi-algebra, then
the intersection of A and each closed ball in B with its center at

103
 104 A class of abstract semi-algebras

the zero vector is compact, and hence that A is a closed subset

127 of B, and that each closed bounded subset of A is compact. It is
easily seen that axioms (i), (ii) and (iii) are equivalent to (i), (ii)
and

(iii)′ A with the relative topology induced from the norm topology in
B is a locally compact space.

This is our justification for the use of the term locally compact in the
present sense. Axiom (ii) of course merely excludes trivial exceptional
cases.
If the Banach algebra B has finite dimensions, then its closed unit
ball is compact, and therefore every nontrivial closed semi-algebra in
B is locally compact. In particular, each closed semi-algebra of n × n
real matrices is of this kind. However, the axioms do not imply that
every locally compact semi-algebra is contained in a finite dimensional
algebra, as the following example shows.

Example. Let E be the subset of the closed unit interval [0, 1] consisting
1
of the closed interval [0, ] together with the point 1, and let E be given
2
the topology induced form the usual topology in [0, 1], so that E is a
compact Hausdorff space.

Let A′ denote the class of all functions belonging to CR [0, 1] that are
non-negative, increasing, and convex in [0, 1]; and let A denote the class
of all functions on E that are restrictions to E of functions belonging to
A′ .
It is obvious that A′ is a semi-algebra in CR [0, 1]. We prove that A
128 is a closed subset of CR (E). Each element f of A has a unique extension
1
f ′ ∈ A′ which is linear in [ , 1] defined by
2
1
f ′ (x) = f (x), 0 ≤ x ≤
! 2
1 1
f ′ (x) = α f + (1 − α) f (1) for x = α. + (1 − α)1 , 0 ≤ α ≤ 1
2 2
A class of abstract semi-algebras 105

Let fn be a sequence of elements of A that converges in norm to an

element f of CR (E). Then the sequence { fn′ }, where fn′ ∈ A′ is the exten-
1
sion of fn to [0, 1] which is linear in [ , 1], converges uniformly in [0, 1]
2
and since A′ is a closed subset of CR [0, 1], the limit function f ′ belongs
A′ . But f is the restriction of f ′ to E, and so f ∈ A. Hence A is a closed
semi-algebra in CR (E). To prove that it is locally compact it is enough
to prove that A intersects the unit ball of CR (E) in an equicontinuous set.
If f ∈ A, and || f || ≤ 1, then
!
1
0≤ f ≤ f (1) ≤ 1;
2

and so, for any pair of points x1 , x2 with

1
0 ≤ x1 < x2 ≤ ,
2
we have, by the ∞nvexity of f ,
 
1
f (x2 ) − f (x1 ) f (1) − f 2
0≤ ≤ ≤2
x2 − x1 1 − 21

Thus the set of all such f is equi-continuous, and A is a locally compact 129
semi-algebra. Finally it is obvious that A is not contained in any finite
dimensional subspace of CR (E).
Our principal results are concerned with the existence and properties
of idempotents in a locally compact semi-algebra, and may be regarded
as analogues of classical theorems of Wedderburn. As biproducts we
obtain an abstract characterization of the semi-algebra of all n × n ma-
trices with non-negative entries, and some results related to the Perron-
Frobenius theorems.
Throughout this chapter we will denote by S A the intersection of A
with the surface of the unit ball in B, i.e.,

S A = {x : x ∈ A and ||x|| ≤ 1}.

Obviously the set S A is compact.

 106 A class of abstract semi-algebras

Definition 8.3. Given a subset E of a semi-algebra A, Er will denote the

right annihilator of E in A, i.e.

Er = {x : x ∈ A and ux = 0 (u ∈ E)}.

In particular (a)r denotes the right annihilator of the set consisting

of a only; and the left annihilators El and (a)l are similarly defined.

Definition 8.4. Given subsets P and Q of A, PQ will denote the set of

all finite sums
p1 q1 + · · · + pn qn
130 with pi ∈ P and qi ∈ Q; and P2 will denote PP.

Definition 8.5. A semi-algebra J contained in A is called a right ideal

of A if
a ∈ A, j ∈ J ⇒ ja ∈ J
Left ideals are similarly defined with ja replaced by a j, and semi-algebra
contained in A is called a two sided ideal if it is both a left ideal and right
ideal.

A closed right ideal J is called a minimal closed right ideal if

J , (0) and if the only closed right ideals contained in J are (0) and
J. Similar definitions apply for minimal closed left and two-sided ide-
als. (Closed ideal means an ideal which is a closed subset of A in relative
topology.)

Theorem 8.1. Each non-zero closed right ideal of a locally compact

semi-algebra contains a minimal closed right ideal.

A similar statement holds for left and two-sided ideals.

Proof. Given a non-zero closed right ideal J, there exists, by Zorn’s

lemma, a maximal family ∆ of non zero closed right ideals contained in
J and totally ordered by the relation of set theoretic inclusion. The sets
I ∩ S A with I ∈ ∆ are compact and have the finite intersection property.
Hence their intersection is non-empty and therefore the intersection I0
A class of abstract semi-algebras 107

of the ideals I in ∆ is non-zero. Clearly I0 is a minimal closed right

ideal. 

It is clear that similar results hold for left and two sided closed ide- 131
als.

Lemma 8.1. Let E be a closed subset of a locally compact semi-algebra

A such that αx ∈ E whenever α ≥ 0 and x ∈ E. Let a be an element of
A such that (a)r ∩ E = (0). Then aE is closed.

Proof. Let y = lim axn , (xn ∈ E). 

n→∞

If y = 0 there is nothing to prove since 0 ∈ E and so 0 ∈ aE.

Let y , 0. Then we can assume that xn , 0 (n = 1, 2, . . .). The
xn
sequence sn = in the compact set E ∩ S A has a subsequence (sni )
||xn ||
that converges to an element s ∈ E ∩ S A . We have

asni , 0 and lim asni = as , 0,

i→∞
since sni , 0, s , 0 and (a)r ∩ E = (0).

Hence
|| asni || > m > 0 (i = 1, 2, . . .).

Also since
lim ||xni ||asni = lim axni = y,
i→∞ i→∞

||xn ||asni is a bounded sequence. It follows that the sequence (||xni ||) is
bounded, and therefore has a subsequence convergent to λ > 0 say. Then

y = λ as = a(λs) ∈ aE

and the lemma is proved.

Theorem 8.2. Let M be a minimal closed right ideal of a locally com- 132
pact semi-algebra A with M 2 , (0). Then M contains an idempotent e
and M = eA.
 108 A class of abstract semi-algebras

Proof. The proof begins on familiar algebraic lines. As M 2 , (0), there

exists a ∈ M with aM , (0). Hence

M ∩ (a)r , M

Since M ∩ (a)r is a closed right ideal contained in M, the minimal prop-

erty of M implies that
M ∩ (a)r = (0) (1)


Hence, by Lemma 1, aM is a closed right ideal. We have

0 , aM ⊂ M

and therefore
aM = M.
In particular, there exists an element e ∈ M with

ae = a (2)

The complication of the rest of the argument is forced on us by the

fact that we cannot assert at this point that e2 − e belongs to A. Our next
is to prove that
lim ||en ||1/n = 1 (3)
n→∞
133 By (2), we have

aen = a (n = 1, 2, . . .)
||a||||e || ≥ ||aen || = ||a||,
n

||en || ≤ 1 (n = 1, 2, . . .)

so that
lim ||en ||1/n ≥ 1
n→∞

In order to prove that lim ||en ||1/n ≤ 1, it suffices to show that (||en ||) is
n→∞
bounded.
Let K = inf{||am|| : m ∈ M ∩ S A }.
A class of abstract semi-algebras 109

Since M ∩ S A is compact, this infimum is attained; there exists m0 ∈

M ∩ S A with ||am0 || = K. Therefore, by (1), K , 0. Thus K > 0, and we
have
||ax|| ≥ K||x|| (x ∈ M)
1 1
In particular ||en || ≤ ||aen || = ||a|| (n = 1, 2, . . .). This completes
2 2
the proof of (3).
Suppose that λ > 1, and let

1 1
bλ = e + 2 e2 + . . .
λ λ
The convergence of the series is established by (3), and we have
bλ ∈ M. Also,
λbλ − ebλ = e ∈ M,
and therefore bλ , 0. Let λn > 1(n = 1, 2, . . .) and lim λn = 1. By what 134
n→∞
we have just proved, there exists for each n, an element mn of M ∩ S A
such that
λn mn − emn ∈ M
Therefore, by the compactness of M ∩ S A , there exists an element m of
M ∩ S A such that
m − em ∈ M
Let
J = {x : x ∈ M, x − ex ∈ M}.
We have J , (0) since m ∈ J. Also, J is a closed right ideal con-
tained in M, and therefore J = M i.e.,

x − ex ∈ M (x ∈ M)

But, by (2)
a(x − ex) = 0 (x ∈ M),
and so
x − ex ∈ M ∩ (a)r (x ∈ M)
 110 A class of abstract semi-algebras

Therefore, by (1),

x − ex = 0 (x ∈ M)

In particular e = e2 , and also 135

M = eM = eA.

Definition 8.6. An idempotent e in a semi-algebra A for which eA is a

minimal closed right ideal is called minimal idempotent.

A semi-algebra A is called a division semi-algebra if it contains a

unit element different from zero and if every non-zero element of A has
an inverse in A.

Theorem 8.3. Let e be a minimal idempotent in a locally compact semi-

algebra A. Then eAe is a closed division semi-algebra.

Proof. Let A0 = eAe. Then A0 is a semi-algebra with unit element e,

and is closed since

A0 = {x : x ∈ A and x = ex = ex}.

Let eae e a non-zero element of A0 . Then

e < (eae)r ∩ eA, e ∈ eA.

Since eA is a minimal closed right ideal and (eae)r ∩ eA is a closed right

ideal properly contained in it, we have

(eae)r ∩ eA = (0).

It follows, by Lemma 1, that (eae)r ∩ eA is a closed right ideal.

Since it contained eae and is contained in eA it coincides with eA, and
therefore
(aea)A0 = A0 .
136 This proves that every non-zero element of A0 has a right inverse,
A class of abstract semi-algebras 111

and a routine argument now completes the proof.

Given x , A0 , with x , 0, there exists y ∈ A0 with xy = e. It follows
that y , 0, and so there exists z ∈ A0 with yz = e. But then

x = x(yz) = (xy)z,

and so yx = e and x has an inverse y.

Definition 8.7. A semi-algebra A is said to be strict if x, y ∈ A, x + y =

0 ⇒ x = 0.

Theorem 8.4. Let A be a closed strict division semi-algebra. Then

A = R+ e,

where e is the unit element of A and R+ is the set of all non-negative real
numbers.

Proof. We prove first that if x, y ∈ A, y , 0, and k x k is sufficiently

small then y − x ∈ A. 

Since y , 0, it has an inverse y−1 in A and for sufficiently small k x k,

we have k z k< 1, where z = y−1 x. Since A is a closed semi-algebra in a
Banach algebra, the series

e + z + z2 + · · ·

converges to an element a of A, and (e − z) = e. This shows that a , 0,

and it therefore has an inverse b in A therefore

e − z = (e − z)ab = b ∈ A.

137
Finally
y − x = y(e − y−1 x) = y(e − z) ∈ A.
Suppose now that u ∈ A, u , 0, and let

µ = sup{λ : e − λu ∈ A}.
 112 A class of abstract semi-algebras

By what we have just proved, we have µ > 0. Also, the strictness of

A implies that µ is finite, for otherwise we have
1
e−u ∈ A (n = 1, 2, . . .),
n
and so −u ∈ A, since A is closed; and then u = 0 (as A is strict) which is
not true.
Let y = e − µu. Since A is closed we have y ∈ A. If y , 0, then, for
sufficiently small λ > 0, we have

(e − µu) − λu ∈ A,
i.e., e − (µ + λ)u ∈ A,

which is absurd. Therefore e − µu = 0,

A = R+ e.

Remark . It is of interest to consider what other division semi-algebras

there are besides R+ . In any semi-algebra A, A ∩ (−A) is an ideal. Hence
if A is a division semi-algebra either A ∩ (−A) = (0) and A is strict, or
138 A ∩ (−A) = A and A is a division algebra. Thus the only non-strict divi-
sion semi-algebras are the familiar division algebras. On the other hand
there are many strict (nonclosed) semi-algebras. For example, let E be
a compact Hausdorff space and let A be the subset of CR (E) consisting
of those functions f ∈ CR (E) such that either

f (t) = 0 (t ∈ E),
or f (t) > 0 (t ∈ E).

It is easily seen that each such A is a strict division semi-algebra.

Definition 8.8. A semi-algebra A is said to be semi-simple if the zero

ideal is the only closed two-sides ideal J with J 2 = (0).

Lemma 8.2. Let A be a semi-simple semi-algebra, and let I be an ideal

(left, right, or two-sided) of A such that I n = (0) for some positive integer
n. Then I = (0).
A class of abstract semi-algebras 113

Proof. We first show that if J is any left ideal with J 2 = (0) then

J = (0) (1)

Let H = (JA). Then H is a closed two-sides ideal, and since

(JA)(JA) = J(AJ)A ⊂ J 2 A = (0),

we have H 2 = (0) and so H = (0), JA = (0). This gives J ⊂ A1 , and 139

since A1 is a closed two-sides ideal with

A21 ⊂ A1 A = (0),
we haveA1 = (0), J = (0)

If now I is a left ideal and n is the least positive integer with I n = (0),
then I n−1 I = 0, and so n > 1 would give (I n−1 )2 = (0), and so by (1)

I n−1 = 0

Hence n = 1, I = (0).
A similar argument applies to right ideals. 

Theorem 8.5. Let A be a semi-simple locally compact semi-algebra, and

let e be an idempotent in A. Then e is a minimal idempotent if and only
if eAe is a division semi-algebra.

Corollary . eA is a minimal closed right ideal if and only if Ae is a

minimal closed left ideal.

Proof. That eAE is a division-algebra if e is a minimal idempotent was

proved in Theorem 8.3. To prove the converse suppose that eAe is a
division semi-algebra. Since

eA = {x : x ∈ A and x = ex},

eA is a closed right ideal. Since it contains e it is non-zero, and it there-

fore contains a minimal closed right ideal M. Since A is semi-simple, 140
 114 A class of abstract semi-algebras

Lemma 2 shows that M 2 , (0); and therefore, by Theorem 8.2, M con-

tains an idempotent f with M = f A. Since

( f A)2 ⊂ ( f A)(eA),

and A is semi-simple, we have

f Ae , 0

Let a be an element of A with f ae , 0. Then f ae is non-zero element

of eAe and therefore has an inverse b in eAe,

f aeb = e.

It now follows that eA ⊂ f A; and so by the minimal properly of f A, eA

is a minimal closed right ideal. 

The Corollary is evident from the symmetry of the conditions on A

and eAe.

Theorem 8.6. Let A be a semi-simple locally compact semi-algebra, and

let I be the set of all minimal idempotents in A. If e ∈ I and a ∈ A,
then there exists f ∈ I and b ∈ A with ae = f b.

Corollary. I A is a two-sides ideal.

Proof. Let e ∈ I and a ∈ A. If ae = 0, we take f = e and b = 0. 

Suppose ae , 0. Then e < (ea)r , and therefore the closed right ideal

(ea)r ∩ eA

141 is a proper subset of eA. Therefore, by the minimal property of eA,

(ae)r ∩ eA = (0)

By Lemma 8.1, it follows that

aeA = (ae)(eA)
A class of abstract semi-algebras 115

is closed right ideal. It is a minimal closed right ideal, for if J is non-zero

closed ideal properly contained in aeA, then {ex : aex ∈ J} is a non-zero
closed right ideal properly contained in eA. Since A is semi-simple and
locally compact, there exists f ∈ I with aeA = f A. In particular

ae = f b for some b ∈ A.

The corollary is obvious.

Theorem 8.7. Let A be a semi-simple locally compact semi-algebra.

Then the set of minimal closed two-sides ideals of A is finite and non-
empty.

Proof. By Theorem 8.1 and the fact that A is a non-zero closed two-
sided ideal of itself, A has at least one minimal closed two-sided ideal.


Suppose that A has an infinite set {Mα : α ∈ ∆} of minimal closed

two-sides ideals. Then

Mα ∩ Mβ = (0) (α , β),
and so Mα Mβ (0) (α , β).

142
For each α ∈ ∆, choose mα ∈ Mα ∩ S A . By the compactness of S A ,
there exists a sequence (αn ), of distinct elements of ∆ such that (mαn )
converges to an element m say of S A . Given α ∈ ∆, we have

Mα mαn = (0) for all n such that α , αn

and therefore Mα m = (0) (α ∈ ∆)
T
Let J= (Mα )r
α∈∆
Since Mα is a two-sided ideal, (Mα )r is a closed two-sided ideal
and so J is a closed two-sided ideal and is non-zero since m ∈ J. By
Theorem 8.1, J contains a minimal closed two-sided ideal Mβ , say. But
Mβ2 = (0), contradicting the semi-simplicity of A.
 116 A class of abstract semi-algebras

Theorem 8.8. Let A be a semi-simple locally compact semi-algebra and

let its minimal closed two-sided ideals be denoted by M1 , M2 , . . . , Mn .
Let I be the set of all minimal idempotent in A and let Ik = I ∩
Mk (k = 1, 2, . . . , n). Then

i) the sets Ik are disjoint and their union is I ,

ii) For each k, Ik A is a two-sided ideal,

Ik A = AIk = Ik AIk ,

143 iii) Mk = d(Ik A).

Proof. Given e ∈ I , either eA ∩ Mk = eA or eA ∩ Mk = (0). In the first

case e ∈ Mk , e ∈ Ik . In the second case, since

eAMk ⊂ eA ∩ Mk ,
S
n
we have eAMk = (0), e ∈ (Mk )l . Thus if e ∈ I but e < Ik , then
k=1

\
n
e∈ (Mk )l = J.
k=1

Since J is a non-zero closed two-sided ideal, it contains one of the min-

imal closed two-sided, M j say. But this lead to M 2j = (0), which is
impossible as A is semi-simple. Thus
[
n
I = Ik .
k=1

The disjointness of the Ik follows from the fact that M j ∩ Mk = (0) ( j ,

k). 

Given e ∈ Ik , and a ∈ A, we have ae = f b with f ∈ I and b ∈ A.

If ae = 0, we can take f = e ∈ Ik . Since ae ∈ Mk , we have f b ∈ Mk ,
and so if ae , 0,
f A ∩ Mk , (0).
A class of abstract semi-algebras 117

But this gives f A ∩ Mk = f A, and so f ∈ Mk . This proves that

Ik A is a two-sided ideal, and it is plainly the smallest two-sided ideal
containing Ik . Similarly AIk and Ik AIk are both this smallest two-
sides ideal, and so (ii) holds.
Finally, each Ik is non-empty, for Mk being a non-zero closed right 144
ideal contains a minimal idempotent. Therefore d(Ik A) is non-zero
closed two-sided ideal containing Mk ; and the minimal property of Mk
gives (iii).

Theorem 8.9. Let e be a minimal idempotent in a semi-simple locally

compact semi-algebra A. Then eA is a minimal right ideal and Ae is a
minimal left ideal.

Remark . Of course, we know that eA and Ae are minimal closed right

and left ideals. But, a priori, they might contain smaller non-closed
ideals.

Proof. With the notation of Theorem 8.8, e ∈ Ik for some k. Let J be a

non-zero right ideal contained in eA, and choose u ∈ J with u , 0. 

If Ik ⊂ (u)r , then Mk ⊂ (u)r by Theorem 8.8, and so u ∈ (Mk )l . But

since u ∈ eA ⊂ Mk , this implies that

Mk ∩ (Mk )l , (0),

and therefore Mk2 = (0), which impossible. Therefore there exists f ∈

Ik with u f , 0. Since f A is a minimal closed right ideal, it follows that

(u)r ∩ f A = (0)

Therefore, by Lemma 8.1, u f A is a closed right ideal. It is non-zero

since it contains u f 2 = u f , and is containing in eA since u ∈ eA. There- 145
fore

eA = u f A ⊂ J,
J = eA
 118 A class of abstract semi-algebras

Lemma 8.3. Let e be a minimal idempotent in a strict locally compact

semi-algebra A. Then
eAe = R+ e.

Proof. This is an immediate consequence of Theorem 8.3 and 8.4 

Lemma 8.4. Let A be a semi-simple strict, locally compact semi-

algebra, and let e, f be minimal idempotent for which f Ae , (0). Then
there exists an element ω of eA f such that

eA f = R+ ω,

and either ω2 = ω or ω2 = 0.

Proof. Choose a non-zero element v of f Ae, say v = f ae. since ve =

v , 0, we have e < (v)r . Using Lemma 1 and the fact that eA is a
minimal closed ideal, we deduce that veA is closed right ideal. Since it
contains v and is contained in f A, we have

veA = f A,
veA f = f A f.

Thus there exists u ∈ eA f for which vu = f . Given x ∈ A, Lemma 8.3,

gives
ex f v = ex f ae = λe

146 for some λ ∈ R+ . Therefore

ex f = ex f vu = λeu = λu

If u2 = 0, we ω = u. If u2 , 0, then u2 = αu with α > 0, and we take

1
ω = u. 
α

Lemma 8.5. Let e, f be minimal idempotents in a semi-simple locally

compact semi-algebra A. Then f Ae , (0) if and only if e and f belong
to the same closed two-sided ideal.
A class of abstract semi-algebras 119

Proof. Suppose that e ∈ Mi and f ∈ M j with Mi and M j minimal closed

two-sides ideals. Then

f A ⊂ Mj
and so, if Mi , M j ,
f Ae ⊂ M j Mi = (0)

On the other hand, suppose f Ae = (0). Then f ∈ (Ae)l since (Ae)l is

closed two-sided ideal and its intersection with M j contains f , it follows
that
M j ⊂ (Ae)l
Therefore M j e = (0), and so e < M j . 

Theorem 8.10. Let A be a semi-simple, strict, locally compact semi-

simple, and let Mk and Ik be defined as in Theorem 8.8. For each pair
e, f of minimal idempotent belonging to Ik , there exists an element ωe, f
of eA f such that eA f = R+ ωe, f and either ω2e, f = ωe, f or ω2e, f = 0. Also 147
X
J= R+ ωe, f
e, f ∈Ik

is a two-sided ideal contained in Mk , and Mk = clJ. Finally, for all

idempotents e, f, g, h in Ik ,

ωe, f ωg,h = λωe,h , for some λ ∈ R+ .

Proof. Let e, f be idempotent belonging to Ik . By Lemma 8.5, f Ae ,

0, and so Lemma 8.4, there exists ωe, f in eA f such that

eA f = R+ ωe, f ,

and ω2e, f = ωe, f , or ωe, f = 0. That J is a two-sided ideal the closures of

which is Mk , now follows from Theorem 8.8. Finally

ωe, f ωg,h = ea f ghk ∈ eAh = R+ ωe,h


 120 A class of abstract semi-algebras

We now consider the question: when does a semi-algebra contain

exactly one minimal closed two-sided ideal?

Definition 8.9. We say that a semi-algebra A is prime if I J , (0) when-

ever I and J are closed non-zero two-sided ideals.

A prime semi-algebra is obviously semi-simple. It is also clear that

in a prime semi-algebra, if J is a non-zero left ideal, then Jl = (0), and
148 if J is a non-zero right ideal, then Jr = (0).

Theorem 8.11. Let A be a locally compact semi-algebra. If A is prime,

then A has exactly one minimal closed two-sided ideal. Conversely, if A
is semi-simple and has exactly one minimal closed two-sided ideal, then
A is prime.

Proof. Minimal closed two-sided ideals annihilate each other, and there-
fore if A is prime there is exactly one such ideal. 

Suppose on the other hand that A is semi-simple and has exactly one
minimal closed two-sided ideal, and let H, J be closed two-sided ideals
with H J = (0). Then (H ∩ J)2 = (0) and so, by semi-simplicity of A,

H ∩ J = (0)

Then either H = (0) or J = (0), for otherwise by Theorem 8.1, they

contain minimal closed two-sided ideals which are distinct since they
have zero intersection.
Our next theorem is concerned with an abstract characterization of
the semi-algebra of all n × n matrices with non-negative real entries.

Definition 8.10. We say that a semi-algebra A is simple if it has no two-

sided ideals other than (0) and A.

Theorem 8.12. A simple, strict, locally compact semi-algebra with a

unit element is isomorphic to the semi-algebra Mn (R+ ) of all n × n ma-
149 trices with non-negative real entries where n is some positive integer.
Conversely, for each positive integer n, the semi-algebra Mn (R+ ) is sim-
ple, strict, locally compact and has a unit element.
A class of abstract semi-algebras 121

Proof. Since A has a unit element, A2 , (0). But A is the only non-zero
two-sided ideal, and therefore A is semi-simple and indeed prime. Let
I denote the class of all minimal idempotents, and let 1 be the unit
element. Then I A is a non-zero, two-sided ideal in A, and so

A = I A.

In particular there exist e1 , e2 , . . . , en ∈ I and a1 , a2 , . . . , an ∈ A

such that
1 = e1 a1 + e2 a2 + · · · + en an , (1)
and we may suppose that the expression (1) has been chosen so that n is
as small as possible. From (1), we obtain

e1 = e1 a1 e1 + e2 a2 e1 + · · · + en an e1 (2)

By Lemma 8.3, e1 a1 e1 = λe1 with λ ∈ R+ . We have λ ≥ 1; for if λ < 1,

then (2) gives
(1 − λ)e1 = e2 a2 e1 + · · · + en an e1 ,
and we could rewrite (1) in the form,

1 = e2 b2 + · · · + en bn ,

contradicting our hypothesis that n was as small as possible. Therefore 150

λ ≥ 1; and rewriting (2) in the form

(λ − 1)e1 + e2 a2 e1 + · · · + en an e1 = 0

and using the strictness of A, we obtain

λ = 1, e j a j e1 = 0( j , 1).

By applying a similar argument with ei in place of e1 , we obtain the

formula
ei ai ei = ei , e j a j ei = 0 (i , j) (3)
 122 A class of abstract semi-algebras

We take ui = ei ai (i = 1, 2, . . . , n). Then (3) gives

u2i = ui (i = 1, . . . , n), ui u j = 0 (i , j), (4)

and we also have

1 = u1 + · · · + un (5)
Since ui A = ei A each ui is a minimal idempotent. And for each i, j, ui Au j
is non-zero and is of the form

ui Au j = R+ ei j

for some element ei j of ui Au j . We choose the elements ei j in such a way

that

eii = ui (i = 1, 2, . . . , n) (6)
ei j e jk = eik (i, j, k = 1, . . . , n), (7)
ei j ekl = 0 ( j , k) (8)

151
In the first place we have ui Aui = R+ ui , and so we can take eii =
ui (i = 1, 2, . . . , n). Next, for j = 2, . . . , n we take ei j to be an arbitrary
non-zero element of u1 Au j . Then we have

u1 Au j = R+ ei j ( j = 1, . . . , n).

Since ei j , (0), (Auk )l = (0), we have

ei j u j Auk = e1 j Auk , (0),

and so e1 j u j Auk = u1 Auk = R+ eik ( j, k = 1, . . . , n).

Therefore, for j = 2, . . . , n and k = 1, . . . , n, we can select e jk such

that
e1 j e jk = e1k
Since e11 = u1 , this holds also for j = 1 i.e.

e1 j e jk = e1k ( j, k = 1, . . . , n) (9)
A class of abstract semi-algebras 123

We have now chosen e jk for all j, k with ui Au j = R+ e jk , and with

(6) and (9) holding. To prove (7), we note that

ei j e jk ∈ ui Auk
and so ei j e jk = λeik with λ ≥ 0.

multiplying by e1i , we obtain by (9), 152

e1k = e1 j e jk = e1i ei j e jk = e1i eik = λe1k

Therefore λ = 1 and (7) is proved. Finally (8) is obvious since u j uk =

0 ( j , k).
Given x ∈ A, we have
X
n
x = 1.x.1 = ui xu j
i, j=1
X n
= ξi j ei j ,
i, j=1

with ξi j ≥ 0. Similarly, for y ∈ A,

X
n
y= ηi j ei j
i, j
 
n 
X Xn 


 

and xy = 
 ξi j η jl 
 eil .

 

i,l=1 j=1

It is now easily seen that the mapping x → (ξi j ) gives an isomor-

phism between the semi-algebra A and Mn (R+ ).
Conversely, with a given positive integer n, let A = Mn (R+ ). A is a
closed semi-algebra in the Banach algebra M = Mn (R) of all n × n real
matrices (with an arbitrary Banach algebra norm), and is locally com-
pact since B has finite dimension. That A is strict and has unit element
is obvious.
Let urs be the matrix with (i, j) the element urs r s
i j = δi δ j , where 153
 124 A class of abstract semi-algebras


1 if
 α=β
δαβ = 

0 if α,β

The matrix urs belongs to A, and if a = (αi j ) is an arbitrary element

of A, we have
X
(urs aulm )i j = δri δµs αµγ δlγ δmj
= δri α sl δmj
= α sl urm
ij ,

so that
m
urs aulm = α sl ur .
It follows that every non-zero two-sided ideal of A contains all the
matrices urs , and so is the whole of A. Thus A is simple and the proof is
complete.
In the case when A is commutative our theorems on idempotents
take a particularly simple form. In the first place we can determine semi-
simple and prime commutative semi-algebras by annihilation properties
of individual elements.

Lemma 8.6. Let A be a closed commutative semi-algebra. If A is semi-

simple, then

a ∈ A, a , 0 ⇒ an , 0 (n = 2, 3, . . .)

154 Conversely if
a ∈ A, a2 = 0 ⇒ a = 0,
Then A is semi-simple.

Also, A is prime if and only if it has no divisors of zero ie. a, b ∈

A, ab = 0 ⇒ a = 0, b = 0.

Proof. Entirely straight-forward using Lemma 8.2. In fact we can prove

a stronger result than Lemma 8.6, that is analogous to a well-known fact
about semi-simple commutative Banach algebras. 
A class of abstract semi-algebras 125

Theorem 8.13. Let a be a non-zero element of a semi-simple commu-

tative locally compact semi-algebra. Then

lim ||an ||1/n > 0

n→∞

Proof. Given a non-zero closed ideal J, let

 
K J = inf ||ax|| : x ∈ J ∩ S A .

By the compactnes of J ∩ S A , this infimum is attained, and so

K J = 0 ⇒ J ∩ (a)r , (0)

For some J, we have K J > 0. For otherwise every non-zero closed ideal
J satisfies J ∩ (a)r , (0), and therefore

M ∩ (a)r = M.

for every minimal closed ideal M i.e., aM = (0) for every such M. Let 155
I = cl(aA). Then I is a non-zero closed ideal and I M = (Q). for every
minimal closed ideal M. Since I contains a minimal closed ideal, this
would contradict the semi-simplicity of A.
Let J be a closed ideal with K J > 0. Then

||ax|| ≥ K J ||x|| (x ∈ J),

and, since axn−1 ∈ J,

||an x|| ≥ K J ||an−1 x|| (x ∈ J, n = 1, 2, . . .)

Therefore

||an || ||x|| ≥ K Jn ||x|| (x ∈ J),

n 1/n
and so lim ||a || ≥ K J > 0.
n→∞
 126 A class of abstract semi-algebras

Theorem 8.14. Let A be a semi-simple, commutative locally compact

semi-algebra. Then the set I of minimal idempotents of A is a fi-
nite non-empty set e1 , . . . , en . Each ideal ek A is a closed division semi-
algebra with unit element ek , and ek e j = 0 (k , j). If also A is strict,
then
ek A = R+ ek (k = 1, 2, . . . , n)

Remark. The elements ek are simultaneous eignvectors for A,

aek = λa ek (a ∈ A)

156 Proof. All ideals of A are two-sided, and so, by Theorem 8.7 A has only
finitely many minimal closed ideals M1 , . . . , Mn and we have

Mi M j = (0) i , j.

Since Mi is a minimal closed right ideal, it is of the form ei A with

ei ∈ I . Since ei A = ei ei A = ei Aei , (A being commutative), ei A is a
closed division semi-algebra with unit element ei .
Suppose now that e is a minimal idempotent. Then eA is a minimal
closed ideal, and so eA = Mi = eAi for some i. Then e and ei are both
unit elements for Mi , and so e = ei .

Theorem 8.15. Let A be a semi-simple, strict, commutative, locally

compact semi-algebra, and let e1 , . . . , en be the minimal idempotents of
A. Then, for each element a of A there exists non-negative real numbers
λ1 , . . . , λn such that

aei = λi ei (i = 1, 2, . . . , n),
and max {λi : 1 ≤ i ≤ n} = ρa
where ρa = lim ||ak ||1/k , and ρa > 0 (a , 0)
k→∞

Corollary. If also A is prime then there exists exactly one minimal idem-
potent e and
ae = ρa e (a ∈ A).
 A class of abstract semi-algebras 127

Proof. We have already proved in Theorem 8.13 and 8.14 everything

157 except that
max λi = ρa .


Let e = e1 + · · · en . Since A is strict, we have

\
n
(e)l = (ei )l
i=1

T
n
Also (ei )l is a closed ideal, and so if it were non-zero it would
i=1
contain one of the minimal idempotents ei , which is absurd. Therefore
(e)l = (0).
Let a ∈ A,

K = inf{||xe|| : x ∈ S A }, λ = max(λ1 , . . . , λn ),
µ = ||e1 || + . . . + ||en ||

Since K is attained and (e)l = (0), we have K > 0, and so

||x|| ≤ K −1 ||xe|| (x ∈ A)

For every positive integer k, we have

ak e = λk1 e1 + · · · + λkn en ,
and so ||ak e|| ≤ λk µ.

Therefore
||ak || ≤ K −1 λk (k = 1, 2, . . .)
On the other hand, for some i we have λ = λi , 158

ak ei = λk ei ,
and therefore ||ak || ≥ λk

Therefore λ = ρa , and the proof is complete.

 128 A class of abstract semi-algebras

We now consider some concrete semi-algebras. First some semi-

algebras of matrices.
Let n be a positive integer, Mn (R) the Banach algebra of all n× n real
matrices, Mn (R+ ) the semi-algebra of all matrices belonging to Mn (R)
with all their entries non-negative. Let X = R(n) and let C be the positive
cone in R(n) consisting of all x = (ξ1 , . . . , ξn ) such that ξi ≥ 0 (i =
1, 2, . . . , n). Let ui be the vector (0, 0, . . . , 1, 0, . . . , 0) with 1 in this ith
place and 0 elsewhere.
Given a subset ∆ of (1, 2, . . . , n), let C∆ be the set of all vectors

x = (ξ1 , . . . , ξn ) = ξ1 u1 + · · · + ξn un ,

with ξi ≥ 0 (i = 1, 2, . . . , n), ξi = 0(i < ∆). We call each such cone C∆

a basic cone, and call C a proper basis cone if ∆ is a non-empty proper
subset of (1, 2, . . . , n).
Each matrix a ∈ Mn (R+ ) may be regarded as a linear operator in X
159 that maps C into itself. Such a matrix is said to be reducible if there
exists a proper basic cone C with

aC∆ ⊂ C∆ .

In term of the entries αi j in the matrix a, this is equivalent to

i < ∆, j ∈ ∆ αi j = 0,

for it is equivalent to
au j ∈ C ( j ∈ ∆),
and au j is the vector (αi j , α2 j , . . . , αn j ). For example, if ∆ = (1, 2, . . . , r),
then (αi j ) is of the form !
b c
o d
where b is an r × r matrix.
A matrix a ∈ Mn (R+ ) is said to be irreducible if it is not reducible.
Given a subset E of Mn (R+ ), let

N(E) = {x : x ∈ C and E x = (0)}.

A class of abstract semi-algebras 129

It follows that if a is an irreducible matrix and A is a semi-algebra

with
a ∈ A ⊂ Mn (R+ ),
then A is prime. 160
For, let I, J be non-zero two-sided ideals of A with

I J = (0)

Since J ⊂ N(I), we have N(I) , (0). Since I , (0), we have N(I) ,

C. Therefore
 N(I) is a proper basic cone. But since Ia ⊂ I, we have
aN(I) N(I), so that this would imply that a is reducible.

Theorem 8.16. Let a0 be an irreducible matrix belonging to Mn (R+ ),

and let A(a0 ) denote the smallest closed semi-algebra in Mn (R+ ) that
contains a0 . Then there exists an idempotent e of rank 1 in A(a0 ), such
that
ae = ρa e,
with ρa = lim ||an ||1/n , for every element a of Mn (R+ ) that is permutable
n→∞
with a0 .

Proof. Let A be a closed commutative semi-algebra with

a0 ∈ A ⊂ Mn (R+ ).

Then A is a strict, prime, commutative, locally compact semi-

algebra, and therefore there exists a unique minimal idempotent eA in
A and
aeA = ρa eA (a ∈ A)
In particular
a0 eA = ρa0 eA .
By a theorem of Function (Gantmacher, Theory of matrices, Vol. 2), 161
ρa0 is a simple eigenvalue of the irreducible matrix a0 . Therefore

X = (u) + Y
 130 A class of abstract semi-algebras

where (u) is the one-dimensional null-space of a0 − ρa0 .1, and Y is the

range of this matrix. By a straight-forward argument, using the fact that
the restriction of a0 − ρa0 .1 to Y is nonsingular, we have
eA u = u, eA Y = (0)
In fact
(ao − ρao .1)eA = 0
and so (a0 − ρa0 .1)eA X = (0)
eA X ⊂ (u)
Since eA , 0, this gives eA X = (u), (u) being one dimensional. Also,
since
eA (a0 − ρao .1) = 0, and (a0 − ρa0 .1)Y = Y,
we have eA Y = eA (a0 − ρa0 .1)Y = (0).
162 This proves that eA is the unique projection with eA u = u and eA Y =
(0).
Therefore eA is independent of the choice of the commutative semi-
algebra A containing ao , and so
eA ∈ A(a0 ).
Finally given any matrix a in Mn (R+ ) with aa0 = a0 a, there exists a
closed commutative semi-algebra A in Mn (R+ ) that contains a and a0 .
Example. For n = 2, if !
1 1
a0 =
1 1
1
then e = a0 is a minimal idempotent in A(a0 ), A(a0 ) = R+ a0 .
2
!
α α12
a = 11 commutes with a0 if and only if α11 = α22 and
α21 α22
α12 = α12 , i.e., if and only if
!
α β
a=
β α
A class of abstract semi-algebras 131

Then ae = (α + β)e. (Note all such a are in A(a0 )).

Examples of semi-algebras of functions.
Let E denote a topological space, B(E) the Banach algebra of all
bounded continuous real functions on E, and let A be a locally compact
semi-algebra in B(E). At the beginning of this chapter we saw an exam-
ple of such a semi-algebra A that had infinite dimension. In this particu-
lar example E was not connected and A was a type 2 semi-algebra. (We 163
fn
say that a semi-algebra A in B(E) is of type n if f ∈ A ⇒ ∈ A.)
1+ f
In both these respects our example was as simple as possible. Such
an example cannot have E connected and cannot be of type 1. In fact we
can prove that the following propositions hold for any locally compact
semi-algebra A in B(E).

a) If A contains a non-constant function, then E is not connected.

b) If A is of type 1, then each element of A is constant on each connected

subset of E, A contains a finite set χ1 , . . . , χn of characteristic func-
tions of subsets of E, and each element f of A is a linear combination
of these characteristic functions with non-negative coefficients.
X
n
f = λi χi , (i = 1, 2, . . . , n).
i=1

It is obvious that any semi-algebra in B(E) is semi-simple and com-

mutative. Let A contain a non-constant function. Since A is semi-simple
it contains a minimal idempotent χ, and is the characteristic function of
a set that is both open and closed. We have χ , 0, and so if χ , 1, then
E is not connected.
Suppose that χ = 1 and that E is connected. Then

A = Aχ

is a division semi-algebra. If f ∈ A and f is not constant, then f has 164

an inverse in A, and so, for every t, f (t) , 0. Since E is connected, we
have either f (t) > 0, for all t or f (t) < 0 for all t. The second possibility
cannot occur since f + λχ ∈ A (λ ≥ 0). Therefore all non-constant
 132 A class of abstract semi-algebras

functions in A are contained in B+ (E). Since A contains a non-constant

function, it follows that in fact A ⊂ B+ (E), for if A contains a negative
constant function it also contains a non-constant function that is not in
B+ (E).
Since A ⊂ B+ (E), A is strict, and so

A = Aχ = R+ χ,

i.e., A contains only constant functions and (a) is proved.

b) Suppose A is of type 1. Given f ∈ A, and α > 0, let

E( f, α) = {x : f (x) ≥ α},

and let χ f,α denote the characteristic function f of E( f, α).

1
We prove that χ f,α ∈ A. Let g1 = f , and
α
2g2n
gn+1 = (n = 1, 2, . . .)
1 + g2n
Then lim gn (t) = χ f,α (t) (t ∈ E)
n→∞

Since (gn ) is a bounded sequence of elements of A, it has a uniformly

convergent subsequence, and so χ f,α ∈ A.
165 Let E0 be a subset of E, and f an element of A that is not constant
on E0 , then we can choose points s, t in E0 and a real number α with

f (s) < α < f (t).

We have α > 0, and since χ f,α ∈ A, E( f, α) is both open and closed. But
t ∈ E( f, α) and s < E( f, α), and so E0 is not connected.
Let △ denote the set of all characteristic functions that belongs to
A. Then △ is a finite set, for if (χn ) were an infinite sequence of distinct
elements of △, it would have a uniformly convergent subsequence which
is absurd, since
||χ p − χq || = 1 (p , q)
We show that each element of A is a non- negative linear combination
of the elements of
A class of abstract semi-algebras 133

Give f ∈ A, we have χ f,α ∈ △ for every α > 0, and so f takes only a

finite set of different values α1 , . . . , αn say with 0 < α1 < · · · < αn , and
perhaps also the value 0. Let

hi = χ f,αi (i = 1, . . . , n),

and consider the function

X
n
h = α1 h1 (αi − αi−1 )hi
i=2

If f (t) = αk , then 166




0 (i > k)
hi (t) = 

1 (i ≤ k)
X
k
and so h(t) = α1 + (αi − αi−1 ) = αk .
i=2

Also, if f (t) = 0, then hi (t) = 0 (i = 1, 2, . . . , ), and so h(t) = 0.

Thus f = h, and we have proved (b)
Appendix The Schauder
theorem for locally convex
spaces

In Chapter 3, I asked whether the Schauder fixed point in its full gen- 167
erality (Theorem 2.2) is true for locally convex spaces, and pointed out
that this question did not seem to be answered in the literature. I am very
much indebted to B.V. Singbal who showed that this question could be
settled affirmatively by using a technique due to Nagumo. The result-
ing proof of the general Schauder fixed point theorem is in my view the
simplest proof even for the special case of normed spaces.

Lemma. (Nagumo [24]). Let A be a compact subset of a locally convex

l.t.s E, and V be a neighbourhood of 0 in E. Then there exists a finite set
a1 , . . . , am of points of A and continuous mapping S of A into the convex
hull of a1 , . . . , am such that

S x − x ∈ V (x ∈ A).

Proof. Let W be an open convex symmetric neighbourhood of 0 with

W ⊂ V. Since A is compact, there exists a finite set a1 , . . . , am in A such
that
[n
A⊂ (ai + W). (1)
i=1

135
 136 A class of abstract semi-algebras

Let PW denote the Minkowski functional of W. Since W is open, 168

x ∈ W ⇔ pW (x) < 1 (2)

By (1) and (2), for each x ∈ A, there exists some i with

PW (x − ai ) < 1.

Let S be the mapping of A into E defined by

 m −1 m
X  X
S x =  qi (x)
 qi (x)ai (x ∈ A),
i=1 i=1
where qi (x) = max{1 − pw (x − ai ), 0}.

For each x in A, there is at least one i with qi (x) > 0. Since also
qi (x) ≥ 0 for all i, it follows that S is defined and continuous on A
and that it maps A into the convex hull of a1 , . . . , am . For any i with
pw (ai − x) ≥ 1, we have qi (x) = 0, and therefore

pW (S x − x) < 1 (x ∈ A),
i.e., S x − x ∈ W ⊂ V (x ∈ A).

The mapping S constructed in the above Lemma serves essentially

the purpose for which we used the metric projection in Chapter 2.

169 Theorem. (Singbal). Let E be a locally convex Hausdorff l.t.s., K a non

empty closed convex subset of E, T a continuous mapping of K into a
compact subset of K. Then T has a fixed point in K.

Proof. Let T map K into A ⊂ K, with A compact. For each neighbour-

hood V of 0, there exists a convex hull KV of a finite subset of A and a
continuous mapping S V of A into KV such that

S V x − x ∈ V (x ∈ A). (1)


A class of abstract semi-algebras 137

Since K is convex and A ⊂ K, we have KV ⊂ K. Let T V be the

mapping
T V = S V oT
of KV into itself. By the Brower fixed point theorem T V has a fixed point
xV in KV ,
T V xV = xV (2)
Since xV ∈ KV ⊂ K, we have T xV ∈ A. Since A is compact, there exists
a point u of A such that every neighbourhood of u contains points xV
corresponding to arbitrarily small V, i.e. given neighbourhoods G, H
of 0, there exists a neighbourhood V of 0 such that
(i)V ⊂ G, (ii)xV ∈ u + H (3)
Since u ∈ K and T is continuous in K, given an arbitrary neighbour- 170
hood G of 0, there exists a neighbourhood H of 0 such that
x ∈ (u + H) ∩ K ⇒ T x ∈ T u + G, (4)
and by (3) there exists a neighbourhood V of 0 such that
V ⊂ G, xV ∈ u + (H ∩ G). (5)
Since xV ∈ KV ⊂ K, it follows from (4) that we then have, for such V,
T xV ∈ T u + G. (6)
Since T K ⊂ A, (1) gives
S V T x − T x ∈ V (x ∈ K),
and, in particular,
xV − T xV = T V xV − T xV = S V T xV − T xV ∈ V ⊂ G (7)
Since
u − T u = (u − xV ) + (xV − T xV ) + (T xV − T u),
(5), (6) and (7) give
u − T u ∈ −G + G + G.
Since G is an arbitrary neighbourhood of 0, it follows that u−T u = 0.
Bibliography

[1] P. Alexandroff and H. Hopf, Topologie. 171

[2] N. Aronszajn, Le correspondant topologique de l’ unicite dans las

theorie des equations differentialles, Ann. of Math. (2), 43 (1942),
730-738.

[3] F. F. Bonsall, Sublinear functionals and ideals in partially ordered

vector spaces, Proc. London Math. Soc. 4 (1954), 402 - 418.

[4] ———–, Endomorphisms of partially ordered vector spaces, J.

London Math. Soc 30 (1955), 133-144.

[5] ———–, Endomorphisms of partially ordered vector spaces with-

out order unit, J. London Math. Soc. 30 (1955), 144-153.

[6] ———–, Linear operators in complete positive cones, Proc. Lon-

don Math. Soc. (3) 8 (1958), 53 - 75.

[7] ———–, The iteration of operator mapping a positive cone into

itself, J. London Math. Soc. 34 (1959), 364-366.

[8] ———–, A formula for the spectral family of an operator, J. Lon-

don Math. Soc. 35 (1960), 321-333.

[9] ———–, Positive operators compact in an auxiliary topology, Pa-

cific J. Math. 1960, 1131 - 1138.

[10] M. S. Brodsku and D. P. Milman, On the centre of a convex set, 172

Doklady Akad. Nauk S. S. S. R (N.S), 59 (1948), 837-840.

139
 140 BIBLIOGRAPHY

[11] F. E. Browder, On a generalization of the Schauder fixed point

theorem, Duke Math. J., 26 (1959), 291-303.
[12] J.A. Clarkson, Uniformly convex spaces, Trans. Amer. Math. Soc.,
40 (1936), 396-414
[13] M. M. Day, Normed lineat spaces, Berlin (1958).
[14] N. Dunford and J.T. Schwartz, Linear operators, Part I, New York
(1958).
[15] M. Edelstein, An extension of Banach contraction principle, Proc.
Amer. Math. Soc (1961).
[16] M. Edelstein, On fixed points and periodic points under contrac-
tion mappings, J. London Math. Soc., 37, 74-79.
[17] R. V. Kadison, A representation theory for a commutative topolog-
ical algebra, Memoirs Amer. Math. Soc 7 (1951).
[18] S. Karlin, Positive operators, J. Math. Mech 8 (1959), 907-937.
[19] M.A. Krasnoselsku, Two remarks on the method of successive ap-
proximations, Uspehi, Math. Nauk (N.S), 10, No 1 (63), (1955),
123 - 127 (Russian).
[20] M.G. Krein and M.A. Rutman, Linear operators leaving invari-
ant a cone in a Banach space, Uspehi, Math - Nauk (N.S) 3, No.
1 (23), (1943), 3-95, Russian (English Translation: Amer. Math.
Soc. Tran. no 26).
173 [21] M. Krein and V. Smulian, On regularly convex sets in the space
conjugate to a Banach space, Ann. of Math. 41 (1940), 556 - 583.
[22] A.C. Mewborn, Generalization of sometheorems on positive ma-
trices to completely continuous linear transformations in a normed
linear space, Duke Math. Journal, 27 (1960), 273 - 281.
[23] D. Morgenstern, Thesis (T. U Berlin 1952, quoted in Algebraische
Integral gleichugen II, by W. Schmeidler, Math. Nachr. 10 (1953),
247-(255)
BIBLIOGRAPHY 141

[24] M. Nagumo, Degree of mapping in convex linear topological

spaces, Amer. J. Math. 73, 497-511 (1951).

[25] H.H. Schaefer, Positive transformationen in halbgeordneten

lokalkonvexen vekorraumen, Math. Ann. 129, (1955) 323-329.

[26] ———-, Uber die Methode Sukzessver approximation, Jahres

Deutsch Math. Verein 59 (1957), 131-140.

[27] ———-, Halbeordnete lokalkonvexe Vektorraume, Math. Ann.

135 (1958), 115-141.

[28] ———-, Halbgeordnete lokalkonvexe Vktorraume II, Math. Ann.

138 (1959), 259-286.

[29] ———-, On nonlinear positive operators, Pacific J. Math. 9

(1959), 847-860.

[30] ———-, Some spectral properties of positive linear operators, Pa-

cific J. Math. (1960), 1009-1019.

[31] J. Schauder, Zur Theorie stetiger Abbildun gen in Function alrau- 174
men, Math. Z. 26 (1927), 47-65, 417-431.

[32] ———, Der fixpunktsatz in Funktionalramen, Studia Math. 2

(1930), 171-180.

[33] K. L. Stepaniuk, uelques generalisations du principe du point sta-

tionnarire, Ukrain Mat. Z. 9 (1957), 105-110 (Russian).