Foundations of Convex Geometry PDF
Foundations of Convex Geometry PDF
Foundations of Convex Geometry PDF
CAMBRIDGE
UNIVERSITY PRESS
PUBLISHED BY THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE
The Pitt Building, Trumpington Street, Cambridge CB2 lRP, United Kingdom
CAMBRIDGE UNIVERSITY PRESS
The Edinburgh Building, Cambridge CB2 2RU, UK http://www.cup.cam.ac.uk
40 West 20th Street, New York, NY 10011-4211, USA http://www.cup.org
10 Stamford Road, Oakleigh, Melbourne 3166, Australia
A catalogue record for this book is available from the British Library
Coppel, W. A.
Foundations of convex geometry I W. A. Coppel.
p. cm. - (Australian Mathematical Society lecture series: 12)
Includes bibliographical references and index.
ISBN 0 521 63970 0
1. Convex geometry. I. Title. II. Series.
QA639.5C66 1998
516'.08-dc21 97-32149 CIP
Preface XI
Introduction 1
I Alignments 3
II Convexity 27
1 Examples 27
2 Convex geometries 31
3 Additional axioms 36
4 Examples (continued) 43
5 Notes 47
III Linearity 49
1 Lines 49
2 Linear geometries 54
3 Helly's theorem and its relatives 66
vm Contents
IV Linearity (continued) 71
1 Faces 71
2 Polytopes 74
3 Factor geometries 81
4 Notes 84
VI Desargues 115
1 Introduction 115
2 Projective geometry 119
3 The Desargues property 124
4 Factor geometries 131
5 Notes 135
1 Coordinates 137
2 Isomorphisms 143
3 Vector sums 150
4 Notes 154
1 Introduction 157
2 Separation properties 164
3 Fundamental theorem of ordered geometry 172
4 Metric and norm 178
5 Notes 182
Contents IX
References 211
Notations 216
Axioms 217
Index 218
Preface
Euclid's Elements held sway in the mathematical world for more than two
thousand years. (For an English translation, see Euclid (1956).) In the nineteenth
century, however, the growing demand for rigour led to a re-examination of the
Euclidean edifice and the realization that there were cracks in it. The critical
reappraisal which followed was synthesized by Hilbert (1899), in his Grundlagen
der Geometrie, which has in tum held sway for almost a century.
In what ways may Hilbert's treatment be improved upon today? Apart from
technical improvements, a number of which were included in later editions of his
book, it may be argued that Hilbert followed Euclid too closely. The number of
undefined concepts is unnecessarily large and the axioms of congruence sit uneasily
with the other axioms. In addition, the restriction to three-dimensional space
conceals the generality of the results and seems artificial now.
For these reasons Hilbert's approach is often replaced today by a purely
algebraic one - the axioms for a vector space over an arbitrary field, followed by the
axioms for a vector space over the real field with a positive definite scalar product.
This permits a rapid development, but it simply begs the question of why it is
possible to introduce coordinates on a line. Only when one attempts to answer this
question does one realise that many important results in no way depend on the
introduction of coordinates. It is curious that some powerful advocates of
'coordinate-free' linear algebra have used real analysis to prove purely geometric
results, such as the Hahn-Banach theorem, without any apparent twinge of
conscience.
A system of axioms for Euclidean geometry in which the only undefined
concepts are point and segment was already given before Hilbert by Peano. We
follow his example in the present work, although some of our axioms differ. The
choice of a system of axioms is inherently arbitrary, since there will be many
equivalent systems. However, the purpose of an axiom system is not only to
provide a basis for rigorous proof, but also to reveal the structure of a subject.
XII Preface
From this point of view one axiom system may seem preferable to another which is
equivalent to it. The development should seem natural, almost inevitable.
Considerations of this nature have led us to isolate a basic structure, here called
a convex geometry, which is defined by two axioms only. Additional axioms are
chosen so that each individually, in conjunction with these, guarantees some
important property. Four such additional axioms define another basic structure,
here called a linear geometry, for which a rich theory may be developed.
Examples of linear geometries, in addition to Euclidean space, are hyperbolic space
and (hemi)spherical space, where a 'segment' is the geodesic arc joining two
points.
Apart from a dimensionality axiom, only seven axioms in all are needed to
characterize ordinary Euclidean or, more strictly, real affine space. However,
dimensionality plays a role in achieving this number, by eliminating some other
possibilities in one or two dimensions. Although the characterization of Euclidean
space may be regarded as our ultimate goal, it would be contrary to our purpose to
impose all the axioms from the outset. Instead we adjoin axioms successively, so
that results are proved under minimal hypotheses and may be applied in other
situations, which are of interest in their own right. For example, Proposition 111.17
establishes the theorems of Belly and Radon in any linear geometry, and Chapter IV
similarly extends the facial theory of polytopes. In this way each result appears, so
to speak, in its 'proper place'. When there is a branching of paths, we choose the
one which leads to our ultimate goal. This approach, I believe, has not previously
been pursued in such a systematic manner. I have found it illuminating myself and
hope that the illumination succeeds in shining through the present account.
We have spoken of the characterization of Euclidean space as our ultimate
goal, but we are actually concerned with characterizing only a convex subset of
Euclidean space. The greater freedom of the whole Euclidean space is
mathematically desirable, but it is reasonable to require only that our physical world
may be embedded in such a space. This type of non-Euclidean geometry was
already mentioned by Klein (1873) and was further considered by Schur (1909). In
view of the great historical importance of the parallel axiom, it is of interest that it
can play no role here, since it fails to hold in any proper convex subset of Euclidean
space (except a lower-dimensional Euclidean space).
The foundations of geometry have been studied for thousands of years, and
thousands of papers have been written on the subject. Since it is impossible to do
Preface xm
Andrew Coppel
Introduction
In the present work convex subsets of a real vector space are characterized by a
small number of axioms involving points and segments. At the same time a
substantial part of the basic theory of convex sets is developed in a purely geometric
manner. This contrasts with traditional treatments, in which the origin has a
distinguished role and in which use is made of such analytic devices as duality and
the difference of two sets, not to mention metric properties.
We owe to Pasch (1882) the recognition of the role of order in Euclidean
geometry, both the order of points on a line and a more subtle order associated with
points in a plane. A basic undefined concept in Pasch's work was the segment
determined by two distinct points. By assuming various axioms involving
segments he was able to define a line and the order of points on a line. An
additional undefined concept was that of the plane determined by three points not
belonging to the same line.
Peano (1889) avoided this additional undefined concept by replacing Pasch's
planar ordering axiom by two further axioms involving segments. Thus he was
able to give a system of axioms for Euclidean geometry in which the only undefined
concepts were point and segment. Some of Peano's axioms were consequences of
the remaining axioms. An independent system of axioms, involving only points
and segments, was given by Veblen (1904).
Two axioms which were used from the outset by all these authors will be
mentioned here. The first, called density in the present work, states that the
segment [a,b] determined by two distinct points a,b contains a point c ¢ a,b. The
second, called unendingness in the present work, states that for any two distinct
points a,b there is a point c ¢ a,b such that b is contained in the segment [a,c]
determined by a,c. In recent years discrete mathematics has attained equal status
with continuous mathematics, and it seems desirable to develop the foundations of
geometry as far as possible without using these two axioms. This is what is done
here.
2 Introduction
One consequence is that our segments are closed (endpoints included), rather
than open (endpoints excluded) as in the work of the authors cited. However, this
is not a novelty; see, for example, Szczerba and Tarski (1979). What is new here is
the gradual introduction of axioms and an obstinate insistence on establishing
results without the use of axioms which are not required.
I
Alignments
Although we will make no use of the theory of partially ordered sets and
lattices, we will at times use some of the concepts of these subjects and for
convenience of reference we state some definitions here.
4 I. Alignments
(01) x S: x (reflexivity),
(02) if x Sy and y S: x, then x = y (antisymmetry),
(03) if x S: y and y S: z, then x < z (transitivity).
Here a binary relation on X is just a subset R of the product set XxX and
x S: y denotes that (x,y) e R. It may be that neither x < y nor y S:x, in which case x
and y are said to be incomparable. A partially ordered set is totally ordered if no
two elements are incomparable.
If x S: y and x ~ y, we write x < y. Instead of x < y we may write y > x, and
instead of x < y we may write y > x.
We say that x e X is an upper bound for a subset Y of a partially ordered set X
if y < x for every y e Y, and a lower bound for Y if x < y for every y e Y. An
upper bound for Y is said to be a least upper bound, or supremum, for Y if it is a
lower bound for the set of all upper bounds. Similarly a lower bound for Y is said
to be a greatest lower bound, or bzfimum, for Y if it is an upper bound for the set of
all lower bounds. It follows from (02) that Y has at most one supremum and at
most one infimum.
A partially ordered set is a lattice if any two elements x,y have a supremum
x v y and an infimum x A y. It is a complete lattice if every subset Y has a
supremum and an infimum.
If X and Y are partially ordered sets, a map/: X -+ Y is said to be order-
preserving if x 1 < x 2 implies f(x 1) S:/(x2). It is said to be an order isomorphism
if, in addition, it is a bijection and /(x 1) < /(x2 ) implies x 1 < x 2 • An order
isomorphism between two lattices is a lattice isomorphism.
The preceding definitions are all rather basic. A few concepts of lesser
importance will be defined when they are encountered.
For us the most important example of a partially ordered set is the collection X
of all subsets of a given set C, where A S: B denotes that the subset A is contained in
the subset B. This partially ordered set is in fact a complete lattice, since the
supremum of any family {Aa} of subsets is their union UAa and the infimum is
their intersection n Aa.
One property of partially ordered sets will be proved here, since it will be used
repeatedly. This property may be formulated in several equivalent ways. We
1. Partially ordered sets 5
choose here a formulation due to Hausdorff (1927) and will always appeal to it.
Other formulations, including the well-ordering theorem and the popular Zorn's
lemma, may be found in Hewitt and Stromberg (1975).
Proof Let .srf. be the family of all totally ordered subsets of X. Evidently .srf. is not
empty, since it contains any singleton, i.e. any subset of X containing exactly one
element. If 5" is any subfamily of st which is totally ordered by inclusion, then the
union of all the (totally ordered) sets in 5" is again totally ordered and hence is in 91..
By the axiom of choice there exists a function f which associates to each
nonempty subset E of X an element/(£) of E. For any A e 91, let At be the set of
all x e X\A such that Aux e 91. We put
(i) A 0 e ~'
(ii) if 5" is any subfamily of ~ which is totally ordered by inclusion, then the
union of all the sets in 5" is again in ~'
(iii) if B e ~' then g(B) e ~.
satisfied by '<6 and by each ~(C). Evidently also (ii) is satisfied by '<6 and by each
~(C).
To show that ~(C) is a tower it remains to show that Be ~(C) implies
g(B) e ~(C), i.e. that B s C and g(C) s B each imply either g(B) s C or
g(C) s g(B). If g(C) s B then also g(C) s g(B), since B s g(B). If B = C, then
g( C) = g(B). If B c: C, then C cannot be a proper subset of g(B), since g(B) \ B
contains at most one element, and hence g(B) s C, since C e '<6 and g(B) e ~ 0 •
Thus ~(C) is a tower. Since ~o is minimal, it follows that ~(C) = ~o for
each C e '<6. Thus if B e ~o and C e '<6, then either B s C or g( C) s B. Thus
g(C) e '<6 and '<6 is a tower. Since ~o is minimal, it follows that '<6 = ~ 0. Hence,
by the definition of '<6, ~o is totally ordered. D
2 ALIGNED SPACES
Let X be a set and let '<6 be a collection of subsets of X. The collection '<6 will
be said to be an alignment on X if it has the following three properties:
(Hl) S s [SJ,
(H2) S s T implies [SJ s [T],
(H3) [[S]] = [S],
2. Aligned spaces 7
(H4) the convex hull of any set is the union of the convex hulls of all its
finite subsets.
Proof Properties (Hl) and (H2) follow immediately from the definition of
convex hull, and (H3) restates that [S] is convex. The proof of (H4) requires the
use of (A3) and is not so immediate.
A set S s X will be said to be 'good' if [S] = U [F], where F runs through all
finite subsets of S, and will be said to be 'excellent' if Su His good for every
finite set H s X. Thus every finite subset of X is excellent, and S u H is excellent
if S is excellent and H finite.
Let S be an arbitrary nonempty subset of X and let ~ be the family of all
excellent subsets of S. Then~ is not empty, since it contains every finite subset of
S. By Hausdorffs maximality theorem the family~, partially ordered by inclusion,
contains a maximal totally ordered subfamily ~. Put M = U Te'?J T. We are going
to show that Me~.
If His any finite subset of X, the family {[Tu H]: Te ~} is also totally
ordered by inclusion. Hence if we put CH= UTe'?J [Tu H], then CH is convex.
Since M u H s CH' it follows that [M u H] s CH· On the other hand, T s M
implies [Tu H] s [Mu H] and hence CH s [Mu HJ. Thus CH= [Mu H].
Consequently, if x e [M u H] then x e [Tu HJ for some Te ~. Since T is
excellent, it follows that x e [F u H] for some finite set F s T s M. Thus
M u H is good and M is excellent, as we wished to prove.
By the definition of~' no excellent subset of S properly contains M. Since
Mus is excellent ifs e S\M, it follows that M = S. Thus Sis indeed good. D
The properties (Hl)-(83) alone imply that, for arbitrary sets S,T s X,
(i) ss h(S),
(ii) S s T implies h(S) s h(T),
(iii) h(h(S)) = h(S).
We have shown that from any alignment on X we can derive an algebraic hull
operator. However, the process can also be reversed. If h: 2x-+ 2x is an algebraic
hull operator on X, then one sees immediately that the collection <(6 of all sets S s X
such that h(S) =Sis an alignment on X. In fact, there is a bijective correspondence
between alignments on X and algebraic hull operators on X. For our purposes it is
more convenient to take alignments as the starting-point.
An alignment <(6 on a set X may be said to be normed if it has the additional
property
(HO) [0] = 0.
The collection '<6 of all convex sets, partially ordered by inclusion, is a
complete lattice. For any family {Ca} of convex sets has an infimum, namely
nCa, and a supremum, namely the intersection of all convex sets which contain
UCa.
If S s X and e e S, then e is said to be an extreme point of the set S if
e ~ [S\e]. Clearly if e is an extreme point of S, then it is also an extreme point of
every subset of S which contains it. From the definition we obtain also
Since any intersection of convex sets is again a convex set, it follows from
Proposition 2 that if an arbitrary collection of extreme points is removed from a
convex set, then the remaining set is convex.
We will denote by E(S) the set of all extreme points of the set S. It can be
characterized in the following way:
Proof If e is an extreme point of Sand if Tis a subset of S with [T] = [S], then
e e T, since [S\e] is a proper subset of [S]. On the other hand, if e e Sis not an
extreme point of S then e e [S \ e]. Hence S s [S \ e] and [S] = [S \ e]. But
e ~ S\e. Cl
Obviously C is itself a face. The proper faces of a convex set C are the faces other
than c itself.
We will say that a set S s: Xis a Radon set if for every partition
we also have [S 1] n [S2] = 0. This implies that for any pair of disjoint subsets
S',S" of S we have [S1 n [S"] = 0, since we can extend the pair to a partition
S 1 =S', S2 = S \ S'. Consequently any subset of a Radon set is again a Radon set.
On the other hand, it follows from (H4) that an infinite set is a Radon set if every
finite subset is a Radon set. Hence the union of a totally ordered collection of
Radon sets is again a Radon set, and any Radon set is contained in a maximal
Radon set. Moreover any singleton is a Radon set, by (HO).
We will say also that a set S s: Xis a Helly set if
flses [S \ s] = 0.
Any subset of a Belly set is again a Belly set. For if Sis a Belly set and Ts: S,
then
= r\es [S\s] = 0.
We now show that an infinite set S is a Belly set if every finite subset is a
Belly set. Indeed if S is not a Belly set, there exists a point s0 e [S \ s] for every
s e S. Taking s =s0 , it follows that there exists a finite set F0 s: S \ s0 such that
s0 e [F0]. If F0 = {s 1, ••• ,snl then, in the same way, for each i = l, ... ,n there
exists a finite set F; s: S \ s; such that s0 e [F;]. It follows that the finite subset
12 I. Alignments
PROPOSffiON 9 Every Helly set is a Radon set, and every Radon set is an
independent set.
=
Proof Suppose that the set S X is not a Radon set. Then there exists a partition
*
S = S 1 u S2, S1 nS2=0 with [Si] n [S2] 0. Letx e [S 1] n [S2]. Ifs e S1,
then S 2 s S \s and hence x e [S \s]. Ifs e S 2, then S 1 =
S \s and again
x e [S\s]. Thus x e nses [S\s] and Sis not a Belly set.
On the other hand, by the definition of an extreme point, every point of a
Radon set R is an extreme point of R. Thus any Radon set is an independent set.
a
Let S be a Helly set with ISi = h + 1. (Here, and later, we denote the
cardinality of a set S by ISi.) If s 1, ••• ,sh+l are the elements of S, then the family of
convex sets C; = [S \s;] (i = l, ... ,h+l) has empty intersection, although every
subfamily of h of the sets has nonempty intersection. This statement admits a
converse:
Thus A is a Helly set containing at least h+ 1 points, and we can take S to be any
subset of A containing exactly h+ 1 points. D
(ii) Join: Let X 1, X 2 be disjoint sets and let '€k be a normed alignment on Xk
(k = 1,2). If X =X 1 u X2 and'€= {C: C = C 1 u C 2, where Ck e '€k (k = 1,2)},
then'€ is a normed alignment on X.
If S s X, then the extreme points of S in the aligned space (X,'€) are the
extreme points of S nX 1 in the aligned space (X 1,'€ 1), together with the extreme
points of Sn X2 in the aligned space (X2,'€ 2). Hence Sis independent in (X,'€) if
and only if Sn Xk is independent in (Xk,'€,J fork= 1,2.
[Du(X\Y)]nY =D
is an alignment on Y. The alignment ~ is normed if and only if X \ Y e "6.
14 I. Alignments
3 ANTI-EXCHANGE ALIGNMENTS
An aligned space X will be said to have the anti-exchange property, and its
alignment will be called an anti-exchange alignment, if the following additional
axiom is satisfied:
(.IE) for any distinct x,y e X and any S s X, if ye [S u x] but ye: [S], then
x e: [Su y].
This definition can be reformulated in several other ways:
subset of [S]. Let 9) be the family of all convex sets D such that C c: D c: [S].
Then 9), partially ordered by inclusion, contains a maximal totally ordered
subfamily ~. Evidently M = UDe'!f D is convex and M c: T, since e ~ M.
Consequently M is not a maximal proper convex subset of [S], which contradicts
the definition of ~. Thus (i) => (ii).
If C c: X is convex and x e X \ C, then x is an extreme point of the set C u x.
If (ii) holds, then xis also an extreme point of the set [Cu x] and hence [Cu x] \x
is convex, by Proposition 2. Thus (ii) => (iii).
Suppose S is a finite set such that [E(S)] c: [S]. Let T be a maximal subset of
S containing E(S) such that [T] c: [S]. Then there exists x e S\ [T]. If (iii) holds,
then C =[Tux] \xis convex. Since Tis maximal, it follows that [Tux]= [S]
and Sn C = T. Hence S =Tux and x ~ [S \x]. Since xis not an extreme point
of S, this is a contradiction. Thus (iii)=> (iv).
Suppose that for some x,y e X and some S s: X, ye [Su x] and x e
[Su y], but y ~ [S]. Then there exists a finite set F s: S such that ye [Fu x] and
x e [Fu y]. Put T =Fu {x,y}. If (iv) holds, then [T] = [E(nJ. Since x and y
are not extreme points of T, it follows that [T] = [F]. But this is a contradiction,
since [F] s: [S] and y ~ [S]. Thus (iv) => (lE). D
Throughout the remainder of this section we assume that the aligned space
X has the anti-exchange property and we derive some further simple properties.
Proof If T is a basis of S, then [T] = [S] and every point of T is an extreme point
of T. But, by Proposition 1 l(ii), T and S have the same extreme points. Hence
T = E(S) and [S] = [E(S)].
On the other hand if [S] = [E(S)] then E(S) is a basis of S, since it is
independent. a
It follows from Proposition 12 that a set has a basis if and only if its convex
hull has a basis, and the bases are then the same. Also, if S has a basis and
E(S) s: T, then [S] s: [T].
It will now be shown that if S is an independent set and T s: S, then
T = [T] n S.
16 I. Alignments
4 ANTIMATROIDS
(o) 0, Xe "6,
(i) C,D e 'l6 implies Cr.De '€.
Moreover (X,'€) has the anti-exchange property if and only if, in addition,
Proof Since Xis finite, (o) and (i) imply that the intersection of any family of sets
in ~ is again a set in ~' and it is trivially true that the union of any family of sets in
~ which is totally ordered by inclusion is again a set in~. Hence (X,~) is an
aligned space if and only if (o) and (i) hold
Suppose (X,~) is an aligned space with the anti-exchange property, Ce ~and
C -¢X. If C0: =X then, by Proposition ll(iv), C0 = [E(C0)]. Hence there exists a
point e 1 e E(C0) \C. Then, by Proposition 2, C 1 = C 0 \e 1 is also convex and
Cs C 1• If C -:1: C 1, let e2 e E(C 1)\C. Then in the same way C 2 = C 1 \e 2 is
convex and C s C2• If C '¢ C2 we can repeat the process. Since X is finite, there is
a least positive integer k such that C = Ck· Then C u ek is convex.
Suppose now that the aligned space (X,~) has the property (ii). If Ce ~and
x e X \ C, then there exists a chain of sets in~:
Hence the aligned space (X,~) has the anti-exchange property, by Proposition
11(iii). a
In the statement of Proposition 13 we can evidently omit the requirement
X e ~ in (o) when (ii) holds. It is worth noting also that, although the anti-
exchange property is obviously preserved under restriction of an aligned space to an
arbitrary subset, it is not obvious that the property (ii) is preserved.
Antimatroids can also be characterized in the following way:
Proof Suppose first that the aligned space (X,~) has the anti-exchange property,
and let C,D be convex sets with C ~ D. Since C is the convex hull of its extreme
18 I. Alignments
for some i, then 'Y =y 1 ... Y,,Xj is an admissible word for some xi e {x1, ••• ,xm}.
To see this, takej to be the least i for which X; E {y1,... ,ynl· Then there exist
distinct Y; 1, ••• ,y;i_1 e {y 1, ••• ,ynl such that x 1 = Y;1, ••• ,xi-I = Y;i-l . Since
we have xi e E(X \ {yi, ... ,ynD· Hence 'Y =y 1 ••• Y,,Xj is an admissible word.
A language with the properties (o)-(iii) is sometimes called a shelling
structure. It follows from Proposition 14 that, conversely, if X is the finite
alphabet of a shelling structure then an alignment with the anti-exchange property is
4. Antimatroids 19
obtained by taking " to be the collection of all sets C = X \ {x 1, ••. ,xm}, where
a =x 1 ••• Xm is an admissible word of the shelling structure.
Another result for finite aligned spaces with the anti-exchange property is the
following:
Proof Obviously we may assume that S itself is not convex. For each x e [S] \ S
we have x 4! [S \ s] for some s e S. Choose such an x so that the number of s e S
for which x 4! [S \ s] is a minimum and let e be one such s. Then e is an extreme
point of S, since x e [S].
Let T =x u (S \e), so that Ts X and ITI = ISi- Since [S] \ e is convex, we
have [T] s [S] \e. We claim that Tis a Helly set. To establish this we will show
that for each ye [T] we have y 4! [T\t] for some t e T. If y 4! [S\e] we can take
t =x, since T\x =S\e.
Thus we now suppose that y e [S \ e]. Assume first that for some t e S we
have y 4! [S\t]. If x e [S\t], then t e T andx is the only element of T\t which is
not in S\t. Sincex e [S\t], it follows that [T\t] s [S\t] and hence y 4! [T\t].
Hence we may now assume that x 4! [S \ t] for every t e S for which
y 4! [S \ t]. Since x 4! [S \ e] and ye [S \ e] this implies, by the choice of x, that
y ES.
Since y e S and S is a Helly set, y 4! [S \ t] for some t e S. Evidently this
implies that y = t and that y is an extreme point of S. Moreover y -:1: e, since
ye [S \e], and hence ye T. If we put D = [S \y], then x 4! D but x e [S] =
[y u D]. Hence, by the anti-exchange property, y 4! [x u D] and, a fortiori,
y 4! [T\y]. Thus our claim is now established.
In this way we have replaced the Helly set S by another Helly set T with the
same cardinality but with [T] c: [S]. If Tis not convex we can repeat the argument
Since Xis finite, we must eventually arrive at a convex Helly set which has the
same cardinality as S. a
It should be noted that, in any aligned space, a convex set is a Helly set if and
only if it is independent, by the definitions and Proposition 9.
20 1. Alignments
S EXCHANGE ALIGNMENTS
Throughout this section the convex sets of the aligned space X will be called
affine sets. Furthermore the convex hull of a set S s X will be called its affine
hull and will be denoted by <S>, rather than by [S]. The reason for these changes
in terminology and notation is that in Chapter Ill, where the results of this section
will first be used, we will be concerned with a set X which is equipped with two
different alignments and it will be necessary to distinguish between them.
An aligned space X will be said to have the exchange property, and its
alignment will be said to be an exchange alignment, if the following additional
axiom is satisfied:
(E) for any x,y e X and any S Sii X, if ye <Su x> but y i <S>, then
x e <S uy>.
Proof Suppose first that Sis an independent set and xi <S>. If Su xis not
independent then, for some ye Su x, ye <(Su x) \y>. In fact ye S, since
x i <S>, and y i <S \y>, since S is independent. If X has the exchange property it
follows that x e <(S \y) u y> = <S>, which is a contradiction. This proves that
(E) => (i).
Suppose next that T is a maximal independent subset of the set S. If (i) holds,
5. Exchange alignments 21
then S s <I> and hence <S> = <T>. Thus T generates S and actually, since it is
independent, Tis a basis of S. Thus (i) =>(ii).
Suppose now that Tis an independent subset of the set S. The family ?"; of all
independent subsets of S which contain T, partially ordered by inclusion, contains a
maximal totally ordered subfamily ~. The union B of all the sets in ~ is an
independent set containing T. Moreover B is a maximal independent subset of S.
Hence (ii) => (iii).
Suppose A is an affine set, x ~ A, and there exists an affine set A' such that
A c A' c <A u x>. If (iii) holds, then A has a basis B and A' has a basis
B' ::> B. Furthermore B' u x has a basis B" containing B'. Since <B'> =A' and
<B "> = <A u x>, we must have B" = B' u x. If y e B' \ B, then
y i <(B' u x) \y>, since B' u xis independent. But Bux s (B' u x) \y and
<B u x> = <A u x>. Since ye <A u x>, this is a contradiction. This proves that
(iii) => (iv).
Finally suppose y e <Su x> but y i <S>. Then x i <.S>. If (iv) holds, then
<.S u y> = <.Su .t> and hence x e <Su y>. Thus (iv) => (E). C
The property (iv) is known as the covering property, since in any partially
ordered set (in our case, the lattice of affine sets) an element B is said to cover an
element A if A < B, but there is no element C such that A < C < B.
Throughout the remainder of this section we assume that the aligned space
X has the exchange property. It follows from Proposition 16(ii) that any subset of
X has a basis. Furthermore,
Proof Evidently <Tu T'> = <S>. Let ?"; be the collection of all independent sets
U such that T s U s T u T'. Then ~ is not empty, since it contains T. If we
partially order ~ by inclusion then, by Hausdorffs maximality theorem, ?";
contains a maximal totally ordered subcollection V. The union V of all sets in Vis
an independent set such that T s V s Tu T'. Thus V = T u T", where T" s T'
and T n T" = 0. Since V is a maximal independent subset of T u T', we have
<V> =<Tu T'>, by Proposition 16(ii). C
22 I. Alignments
Proof Since Bi is independent, b 1 i <.B 1 \b 1> and hence, since <.B 2> =<S>,
there exists b 2 e B 2 such that b 2 i <.B 1 \b 1>. Since b2 e <(Bi \bi) u b 1>, it
follows from the exchange property that b1 e <.B 3>, where B3 = (B 1 \b 1) u b2.
Since also B 1\b 1 s <8 3>, it follows that <.B3> = <S>.
It remains to show that B3 is independent. Assume on the contrary that, for
some b 3 e B 1 \b 1, we have b 3 e <(B 1 \ {b 1 ,b 3 }) u b 2>. Since b 3 i
<.Bi\ {b 1,b3 }>, it follows from the exchange property again that b2 e <Bi \bi>,
which is a contradiction. a
Proof The result holds for n = l, by Proposition 16(i). We suppose n > 1 and use
induction on n. Then Bu S', where S' = {x1,... ,xn_iJ, is independent. Since
Xn i <.B u S'>, it follows that also B u s is independent. a
5. Exchange alignments 23
dim A2 - dim Alt is equal to the codimension of Ai in A2 • This holds also for
Ai = 0 if we assign to the empty set the dimension -1.
It follows directly from the definition that the restriction of an aligned space X
with the exchange property to an arbitrary subset Y again has the exchange
property. Furthermore the join of two disjoint aligned spaces with the exchange
property again has the exchange property.
6 NOTES
(A3)' the union of any nonempty directed family of sets in ~ is again a set in
~-
1 EXAMPLES
EXAMPLE 1 Let X be a vector space over the field R of real numbers. For any
x,y e X, define the segment [x,y] to be the set of all z e X which can be
represented in the form z =Ax+ (1 -A.)y, where A. e Rand 0 ~A.~ 1. Then a set
C =Xis said to be convex if x,y e C implies [x,y] s C.
For example, the subsets of R2 in Figure 1 are convex, but those in Figure 2
are not.
''''''''''''''''''''''
,,,,,,,,,,,,,,,,,,,,,,
''''''''''''''''''''''
,,,,,,,,,,,,,,,,,,,,,,
''''''''''''''''''''''
''''''''''''''''''''''
''''''''''''''''''''''
''''''''''''''''''''''
,,,,,,,,,,,,,,,,,,,,,,
''''''''''''''''''''''
''''''''''''''''''''''
,,,,, ,,,,,,
'''''
,,,,,
''''' ''''''
''''''
'''''
,,,,, ''''''
'''''
''''' ''''''
''''''
'''''
''''' ''''''
''''''
'''''
,,,,,
''''' ''''''
''''''
,,,,,,
''''' ''''''
,,,,,,,,,,,,,,,,,,,
,,,,,,,,,,,,,,,,,,,
'''''''''''''''''''
,,,,,,,,,,,,,,,,,,,
'''''''''''''''''''
'''''''''''''''''''
Many elementary properties of convex sets hold also in other structures, since
their proofs do not make full use of the preceding definition. Here are some
examples of the structures we have in mind. In each case we associate with any
two elements x,y of a set X a subset [x,y] of X containing them, and we then define
a set Cs X to be convex if x,y e C implies [x,y] s C.
EXAMPLE 2 In Example 1 the field R of real numbers can be replaced by the field
Q of rational numbers, leaving everything else unchanged. More generally, the
field R can be replaced by an arbitrary ordered division ring D.
A division ring differs from a field only in that multiplication need not be
commutative. A division ring D is ordered if it contains a subset P of positive
elements which is closed under addition and multiplication and is such that D is the
disjoint union of the sets {0}, P and - P = {- A.: A. e P}. If A.,µ e D we write
A. < µ when either µ - A. = 0 or µ - A. e P.
1. Examples 29
Any ordered field K can be embedded in a larger ordered field in the following
way. Let K(t) denote the field of all rational functions in one indeterminate t with
coefficients from K. An element of K(t) can be uniquely expressed in the form
f / g, where f and g are relatively prime polynomials in t with coefficients from K
and g has leading coefficient 1. If we define such an element to be 'positive' when
the leading coefficient off is positive, then K(t) also acquires the structure of an
ordered field. (This ordered field is 'non-archimedean ', since t > n for every
positive integer n.)
Any ordered field K can also be embedded in an ordered division ring which is
not a field. We sketch the construction, due to Hilbert, without giving detailed
proofs. The set L = L(K) of all formal Laurent series a = I, keZ aktk, where ak e K
and ak ~ 0 for at most finitely many k < 0, is a field if addition and multiplication
are defined by
where 'Yk = I,;+ j=k a;~j· Moreover, L is an ordered field if a = I, keZ aktk is
defined to be positive when, for some m e Z, am > 0 and ak = 0 if k < m. If
p e K, p > 0 and p ~ 1, then the map v: L ~ L defined by
'If (a) = I, keZ pkaktk if a = I, keZ aktk
is a nontrivial automorphism of the field L which preserves positivity.
Now let~= ~(K) be the set of all formal Laurent series A = I.veZ avsv,
where av e L and av ~ 0 for at most finitely many v < 0. If we define addition and
positivity as before, but define multiplication by
EXAMPLE 4 Let X be a partially ordered set. For any x,y e X, define the
segment [x,y] to be the set {x,y} if x and y are incomparable, to be the set of all
ze X such thatx < z Sy if x < y, and to be the set of all ze X such thaty < z <x if
y Sx.
EXAMPLE 6 Let X be a tree (i.e., a finite connected graph without circuits) and,
if x,y are vertices of X, define the segment [x,y] to be the set of all vertices z of X
which lie on the (unique) shortest path fromx toy.
EXAMPLE 7 Let X be a vector space over the field A = R(t) of all rational
functions in one variable with real coefficients. For any x,y e X define the segment
[x,y] to be the set of all z e X such that z = A.x + (1 - A.)y, where A. e A and
0 < A.(t) < 1 for all t e JR for which A.(t) is defined (i.e., for which the denominator
of A. does not vanish).
EXAMPLE 8 Let X be the set of all measurable functions x: T ~ JRn, where Tis a
given measure space. For any x,y e X, define the segment [x,y] to be the set of all
z e X such that z(t) e {x(t),y(t)} for all t e T, except a set of measure zero, i.e.
z ='XEX + (1 - 'XE)y, where 'XE is the characteristic function of some measurable set
Es T.
2 CONVEX GEOMETRIES
Let X be a set and suppose that with any unordered pair {a,b} of elements of X
there is associated a subset [a,b] of X containing a and b. The elements of X will be
called points and the subsets [a,b] segments. We will say that a convex geometry
is defined on X if the following two axioms are satisfied:
Throughout this section and the next we will assume that a set Xis given
on which a convex geometry is defined. We are going to study the consequences
of this assumption.
Proof Let c e [c 1,c 2]. From c 1 e [a,b] and c e [c1'c 2 ] we obtain, by (C),
c e [a,b1 for some b' e [b,c2]. From c2 e [a,b] and b' e [b,c2] we obtain, by (C)
and (Ll), b' e [a,b]. From b' e [a,b] and c e [a,b1 we obtain, by (C) and (Ll)
again, c e [a,b ]. a
We show next that the axioms (C) and (Ll) together imply a generalization of
axiom (C):
Proof Applying (C) twice, we obtain c e [a,b1 for some b' e [b 1,c 2 ] and
b' e [a,b] for some be [b 1,b2]. From c e [a,b1 and a,b' e [a,b] we now obtain
c e [a,b], by Proposition 1. a
[x,y]. However, convex geometries possess many properties not possessed by all
aligned spaces, as we will now see.
Proof Put
R = Uxe [SJ.ye [11 [x,y].
By Proposition 4,
- [x u (Ff"'\ A)]
Proof Let S be a Caratheodory set. Since Sis finite, it is enough to show that S\s
is also a Carath6odory set for every s e S. Moreover, since singletons are
Carath6odory sets, we may suppose that ISi > 2.
Assume, on the contrary, that for some s0 e S,
Then, by Proposition 4,
set, but it is not a Belly set, since (1,1,1) e [x,y] r. [y,z] r. [x,z]. It is easily
verified also that the four points (0,0,1), (1,2,3), (3,1,2), (2,3,0) form a Radon
set, but not a Carath6odory set, and that the six points (+ 1,0,0), (0, + 1,0),
(0,0, + 1) form an independent set, but not a Radon, Belly or Caratheodory set.
Obviously a set Sis convex if and only if K(S) =S. A set Sis said to be star-
shaped if K(S) ~ 0. An example in R2 of a set which is star-shaped, but not
convex, is the middle set in Figure 2 at the beginning of the chapter.
PROPOSITION 8 For any set S X and any x e S, the kernel K(Sx) of the set
5
of all points of S visible from x is the intersection of all maximal convex
subsets of S which contain x. Furthermore,
Proof Since singletons are convex, there exists a maximal convex subset of S
which contains x. Let y e K(Sx) and let M be any maximal convex subset of S
which contains x. Then M 5 Sx and hence, since y e K(Sx),
3 ADDITIONAL AXIOMS
Proof We may assume a :I: b, by (Ll). Then (a,b] and [a,b) are convex, by
Proposition 10 and Proposition 1.11. Hence (a,b) = (a,b] f"\ [a,b) is also convex.
a
Another property in which we are interested is the no branchpoint property:
PROPOSITION 12 If the convex geometry X satisfies (L3) then, for any finite
Radon set Sand any subsets S1,S2 of S,
Proof The result holds by the definition of a Radon set if S 1 and S 2 are disjoint.
Suppose now that S 1 f"\ S 2 = {s} is a singleton. We may clearly assume that
IS1I > 1, IS2I > 1, and then
Suppose z e [s,x] [s,y] for somex e [S 1\s] andy e [S2 \s]. Thenx E [S2] and
f"\
y E [Si], because Sis a Radon set. In particular, x E [s,y] and y E [s,x]. Hence,
by (L3), z = s. Thus [Si] f"\ [Si]= {s} and the result holds in this case.
Suppose next that S1 f"\ S2 contains m > 2 elements and assume that the result
holds whenever S 1 f"\ S2 contains less than m elements. Write S 1 n S2 = s u R,
wheres E R, and put Ti= Si\ (s u R) (j = 1,2). Then
[S 1l = Uxe [Ru Ti] [s,x] ' [S2] = Uye [Ru T2] [s,y].
Proof By Proposition 1.9 we need only show that every Radon set is a Belly set.
Moreover we may restrict attention to finite sets, since an infinite set is a Radon
(Belly) set if and only if every finite subset is a Radon (Belly) set. But if Sis a
finite Radon set then, by Proposition 12,
PROPOSITION 14 If the convex geometry X satisfies (L4) then, for any set
S s X and any c e [S],
[S] = Uses [c U (S\s)].
Proof It is sufficient to prove the result for finite sets S, by (84). Suppose
S = {s 1, ••• ,snl· The result is trivial for n = 1 and it holds for n = 2, by (L4). We
use induction and assume that the result holds for all finite sets containing at most n
elements, where n > 2. Let T =s0 u Sand supposed e [T]. Then, by Proposition
4, de [s 0 ,c] for some c e [S]. Hence, by Proposition 4 and the induction
hypothesis,
[T] = Uze[SJ [so,z] = Ui=t [cu (T\s;)].
But, by Proposition 4, [c u (T \ s;)] is the union of all segments [x,y], with
x e [s 0 ,c] and ye [S \s;]. Since [s 0 ,c] = [s 0 ,d] u [d,c], it follows from
Proposition 4 again that
Thus the result holds also for all finite sets containing n+ 1 elements. D
Proof If Sis finite and there exists a point x e nseS [S\s] then, by Proposition
14,
[S] = Uses [x U (S\s)] = Uses [S\s]. D
PROPOSITION 16 Suppose the convex geometry X satisfies both (L2) and
(L4). Let S = {s 1, .•• ,snl be a finite set and let d,e be distinct elements of [S].
Then [d,e] s [p,q], where p,q e U7= 1 [S\s;].
Proof By Proposition 14,
The axiom's label is chosen in honour of Peano (1889) who, developing the
work of Pasch, first used the axioms (C) and (P). Convex geometries satisfying
the axiom (P) possess a number of additional properties. Our first result is a
counterpart to Proposition 2:
Proof (i) There exists a point e e [b,d] n [c,c], by (P), and e = c, by (Ll).
(ii) Take d = b in (i).
(iii) If de [a,c] rt [b,c], thence [b,d], by (i), and hence d = c, by (ii). []
The preceding two results actually hold in any aligned space satisfying the
axiom (P). The following sand-glass property does require that the aligned space
be derived from a convex geometry - draw a picture in R.2 to see the reason for the
name!
Proof Since x e [a,a1 and ye [a,b] there exists a point z e [a',y] rt [b,x], by
(P). From x e [b,b1 and z e [b,x] we obtain, by Proposition 18(i), x e [b',z] s
[y,a',b1. Hence x e [y,y1 for some y' e [a',b1. D
Proof Obviously the right side is contained in the left. Suppose, on the other
hand, that xis any element of the left side. Then x e [d,y] for some de [a,c] and
y e [S], and x e [e,z] for some e e [b,c] and z e [S]. By Proposition 18(i),
c e [b,d] and also c e [d,e]. Consequently, by Proposition 19, x e [c,w] for some
we [y,z] Si [S]. []
As an application we prove
Proof Assume p e [b 1, ••• ,bk-1'ak,... ,an1 for some k. Then, by Proposition 20,
Since the assumption holds for k =1, the result follows by induction on k. a
A basic property of convex geometries with the property (P) is the following
separation theorem:
Proof Let ~ be the family of all convex sets C" which contain C but are disjoint
from D. Then ~ is nonempty, since it contains C. If we partially order ?:F by
inclusion then, by Hausdorffs maximality theorem,?; contains a maximal totally
ordered subfamily ?;0 • The union C' of all the sets in ?;0 is again a convex set
containing C but disjoint from D.
Since C' is maximal, for every x E C' we have
C' r'\ [x u D] = 0.
Assume on the contrary that, for some x E C ', there exists a point
c' e C' r'\ [x u D] and let d" e [x u C'] r'\ D . By Proposition 4 we have
c' e [x,d1 for some d' e D and d" e [x,c"] for some c" e C'. Hence, by (P),
and
[y u D1 r'\ C' *0
D' r'\ [y u C1 = 0.
42 II. Convexity
Since C' is maximal, it follows that y e C'. That is, C' is the complement of D '.
a
The axiom (P) is essential for the validity of Proposition 22. For suppose
z1 e [x,yi], z2 e [x,y 2] and [y 1,z2] n [y 2,zi] = 0. If C' and D' are convex sets
such that [y 1,z2 ] s C' and [y2 ,zi] s D', then x e C' implies z 1 e C' n D' and
x e D' implies z2 e C' n D'.
A set H s X is said to be a hemispace if both H and X \ H are convex.
Clearly X itself is a hemispace, and the complement of a hemispace is again a
hemispace. Sometimes another terminology is more convenient: two sets C,D are
said to be a convex partition of X if they are convex, nonempty and
CuD=X, CnD=0.
From Proposition 22 we can obtain a separation theorem for any finite number
of sets. The proof will be based on the following preliminary result:
LEMMA 23 Suppose the convex geometry X satisfies (P). Let A = {a 1, ••• ,an}
and B = {b 1, ••• ,bnl be subsets of X such that b; e [p,a;] (i = l, ... ,n)for some
p e X. If A;= A\ a; and B; = B \ b;, then ni:: 1[A; u q] ¢ 0, ni::1[B; u a;] ;t 0.
Proof For n = 1 the result is trivial. For n = 2 both formulae say that
[a 1,b2] n [a 2,bi] ¢ 0, and thus the result holds by (P). Suppose now that the
result holds for sets of n > 2 points, and that the hypotheses are satisfied by
A,= Au an+l and B' =Bu bn+l· If x E ni:1CA; u q] then
Since A 'n+l =A, this establishes the first formula unless [x,an+il n [A u bn+tl =
0. In the latter case there exists, by Proposition 22, a convex partition C,D of X
such that [Au bn+il s C and [x,an+d s D. Since p e D implies bn+l e D, we
must have p e C. Then bn e C and x e [An u bnl s C, which is a contradiction.
The second formula is proved similarly by taking a pointy E nl!:1CB; u a;]
and showing that [y,bn+il n [Bu an+il = 0 leads to a contradiction. a
PROPOSITION 24 Suppose the convex geometry X satisfies (P). If C1, ••. ,Cn
are convex subsets of X such that ni:: 1C; = 0, then there exist hemispaces
H 1, ••• ,Hn such that C; s H; (i = 1,... ,n), ni::1H; = 0 and LJi=1H; = X.
3. Additional axioms 43
Proof Let H 1, ••• ,Hn be maximal convex sets such that C; s H; (i = 1,... ,n), and
n7=1H; = 0. Thus any convex set properly containing Hi intersects nii!jH;
(j = l, ... ,n). Since there exists a hemispace containing Hi and disjoint from
n;i! j H; , by Proposition 22, Hi must itself be a hemispace. Assume there exists a
point p e X\ Uj= 1Hi. Then [Hi up] intersects nii!jH; in some point bi, and
bi e fp,aj] for some aj e Hi. Applying Lemma 23 to the sets A= {a 1, ••• ,an} and
B = {b 1, ••. ,bn}, we see that the sets [(B \bi) u aj] (j = 1,... ,n) have a common
point. Since (B \bi) u ai s Hi, this contradicts n}=t Hi = 0. D
(ii) Join: LetX =X 1 u X 2, where X 1,X 2 are disjoint sets on each of which a
convex geometry is defined. For a,b e X we define [a,b] = {a,b} if a e X 1,
be X2 or if a e X2, be X1, and [a,b] = [a,bh if a,b e Xk (k = 1,2). It is easily
verified that this defines a convex geometry on X. Indeed the axiom (C) holds
trivially if c e {a1,bi} or if de { c,b2 } and otherwise holds in X because it holds in
X 1 and X2 • Moreover this convex geometry on X satisfies those of the axioms
(L2),(L3),(L4),(P) which are satisfied by both the given convex geometries on
x l,X2.
4 EXAMPLES (continued)
It will now be shown that a real vector space, or indeed a vector space over
any ordered division ring, is a convex geometry with the usual definition of
segments. Since (Ll) is obviously satisfied, we need only prove (C). The
geometric significance of the axiom (C) in this case is illustrated in Figure 3. We
have d = 0c + (1 - 0)b 2 and c = A.a + (1 - A.)b 1, where 0 < 0,A. <l. Evidently
44 II. Convexity
Figure 3
we may assume that 0 < 0,A. < 1. Hence if we put v = 1 - 0A., then 0 < v < 1.
Moreover d = (1 - v)a + vb, where
We are going to show that segments in a vector space over an ordered division
ring also satisfy the axioms (L2 )-(L4) and (P). It is easy to verify (L2); for
suppose
b = A.c + (1 - A.)a, c = µb + (1 - µ)d,
where 0 S A.,µ < 1. Substituting the second equation in the first, we obtain
b= ea + (1 - 0)d, where e = (1 - A.µ)-1(1 - A.) and thus O < e < 1. Thus a vector
space over an ordered division ring is actually an aligned space with the anti-
exchange property.
We can now give a concrete illustration of the results of Chapter I, which was
first proved by Doignon (1973):
Proof Put X = [S] (') zd. Then [X] = [S], since S s X s [S], and hence
X = [X] (') Zd. As we saw in Chapter I,~ becomes an aligned space with the anti-
cxchange property if we define a 'convex' set in X to be the intersection with X of a
convex set in Rd. We wish to show that S is not a Belly set in this aligned space.
Assume that Sis a Belly set Then, by Proposition 1.15, there exists a Helly
set T s X with ITI = ISi and T = [T] (') X. Since ITI > 2d, there exist points
4. Examples (continued) 45
y,z e T whose corresponding coordinates are all congruent (mod 2). Then
x =(y + z)/2 E zd. But x E [T] Si [X] = [S]. Hence x E x, and so x E T. Thus T
is not independent, which contradicts Proposition 1.9. a
Since the set S of all points in zd whose coordinates are either 0 or 1 is an
independent set such that S =[SJ fl zd, it follows from the remark after the proof of
Proposition 1.15 that the bound 2d in Proposition 25 is actually sharp.
It is obvious that the axioms (L3) and (L4) are satisfied in a vector space over
an ordered division ring with the usual definition of segments. Finally it will be
shown that the axiom (P) is also satisfied. The geometric significance of the axiom
(P) in this case is illustrated in Figure 4.
Suppose
for some A.,µ with 0 S A.,µ ~ 1. If A. = µ = 1 then (P) obviously holds, since
c 1 =b 1 and c2 =b2. Otherwise we have A.µ< 1 and µA.< 1. Thus if we put
Figure4
We leave it as an exercise for the interested reader to verify that Examples 4-8
at the beginning of this chapter are all convex geometries. Furthermore, Examples
4-7 satisfy the axiom (L2). The axioms (L3),(L4) and (P) are in general not
satisfied in Example 4, but they are all satisfied if the set X is totally ordered. The
46 II. Convexity
axioms (L4) and (P) are satisfied in Examples 5--6, but the axiom (L3) is in
general not satisfied. The axiom (P) is satisfied in Examples 7-8. However, the
axioms (L3)-(L4) are not satisfied in Example 7, and the axioms (L2)-(L4) are
not satisfied in Example 8.
Finally we show that the axioms studied in this chapter are mutually
independent.
(C) Take X = {a,b 1,b2 ,c,d} with [a,bi] = {a,b 1,c}, [c,b2] = {c,b2,d}, and
[x,y] = {x,y} otherwise. The collection of all sets C Si X such that c e C if
a,b 1 e C, and de C if c,b 2 e C, is an alignment on X and, furthermore, it has the
anti-exchange property. Proposition 7 no longer holds, since {a,b 1,b 2 } is a
Caratheodory set but {b 1,b2 } is not.
(Ll) Take any setX such that IXI > 1 with [x,y] =X for allx,y e X.
(L2) Take X = {a~b,c,d} with [a,c] = {a,b,c}, [b,d] = {b,c,d} and [x,y] = {x,y}
otherwise.
An example which is 'continuous' rather than 'discrete' may also be given.
Let X = {z = ei& e C: 0 ~ 0 < x/3 or 2x/3 < 0 < x or 4x/3 ~ 0 < Sx/3}. No two
points of X are antipodal and, for all x,y e X, we define [x,y] to be the set of all
points of X on the shorter arc of the unit circle with endpoints x,y.
(IA) Take X = {a,b,c,d} with [a,b] = X and [x,y] ={x,y} otherwise. Proposition
14 no longer holds, since [a,b] :1: [a,c] u [b,c].
4. Examples (continued) 47
(P) Take X = {a,a',b,b',c} with [a,c] = {a,a',c }, [b,c] = {b,b',c }, and [x,y] =
{x,y} otherwise.
S NOTES
In this chapter we show how lines may be defined, and how the points of a line
may be totally ordered, using four of the axioms considered in Chapter II. The
notion of a linear geometry, which is defined by all six of the axioms of Chapter
II, is then introduced. A linear geometry possesses both an anti-exchange
alignment of convex sets and an exchange alignment of affine sets. This rich
structure makes it possible to establish many familiar geometrical properties,
including the existence of half-spaces associated with a hyperplane and the
property which Pasch showed to have been omitted by Euclid. Nevertheless a
linear geometry may in fact possess only finitely many points. However, it is
shown that all irreducible linear geometries which are not dense, in the sense of
Chapter V, have a rather simple form.
In conclusion we establish the theorems of Belly, Radon and Caratheodory in
any linear geometry. The extension to an arbitrary linear geometry of the theorems
of Bminy and Tverberg, which we prove for R.d, is left as an open problem.
1 LINES
Let X be a set and suppose that with any unordered pair {a,b} of elements of X
there is associated a subset [a,b] of X containing a and b. The elements of X will be
called points and the subsets [a,b] segments. Throughout this section we will
assume that the following four axioms are satisfied:
Proof (ii) Assume b -:I: c. Then be [a,a], by (L2), and hence b =a, by (Ll).
Thus c e [a,a] and hence c =a. Then b =c, which is a contradiction.
(iii) Supposed e [a,c] n [b,c]. Then de [a,b], by (L4). By (L4) also,
either c e [a,d] or c e [b,d]. In either event d = c, by (ii).
(i) Obviously we may assumed -:I: a,c. Then ct! [a,d], by (ii). By (L4),
de [a,b]. Since ct! [a,d], it follows from (L4) again that c e [b,d]. D
Proof Clearly we must have a -:1: b. It is sufficient to show that if c -:I: a,b then
<a,c> = <a,b>. In fact, by symmetry we need only show that <a,c> s <a,b>.
The following proof is arranged so as to appeal to Proposition 1(i) rather than
(L4), as far as possible.
Let x e <a,c>, so that either x e [a,c] or a e [c,.x] or c e [a,.x]. We wish to
show that either x e [a,b] or a e [b,.x] orb e [a,.x]. Evidently we may assume that
x -:I: a,b,c.
1. Lines 51
COROLLARY 3 If there exist three disti11ct poi11ts a,b,c such that cf! <a,b>,
the11 for a11y disti11ct points x,y there exists a point z such that z f! <x,y>. D
Points which lie on the same line will be said to be co/li11ear. We show next
how the points of each line may be totally ordered:
PROPOSITION 4 Given two distinct points x,y e X, the points of the line
e = <x,y> may be totally ordered so that x < y and so that, for any points
a,b E .f with a < b, the segment (a,b) Consists Of al/ CE .f SUCh that a < C < b.
Moreover, this total ordering is unique.
Proof For any a,b e e, we write a < b if either of the following conditions is
satisfied:
It should be noted that (i) and (ii) say the same thing if a = x and that they cannot
both hold if a ~ x.
The definition obviously impliesx < y. Also, from the definition of a line it is
clear that for any a E .f we have a < a. Let a,b be points of .f such that both a < b
and b <a. We wish to show that this implies b =a.
Assume first that a e [x,b] and either ye [x,b] orb e [x,y]. Since x ~ y, this
implies x f! [b,y]. Since b <a, it follows that b e [x,a]. Hence b =a, by
Proposition l(ii). The case b e [x,a] and either ye [x,a] or a e [x,y] may be
discussed similarly. Suppose finally that x e [a,y] and x e [b,y]. Then either
52 III. Linearity
a e [b,y] orb e [a,y], by (L3). We cannot have also ye [a,b], since this would
imply y =a or y =b, respectively, and hence x = y. Consequently we must have
be [a,y] and a e [b,y], which implies b =a.
Suppose next that a,b,c are distinct points of esuch that a < b and b < c. We
wish to show that this implies a< c.
Assume first that be [x,c] and either ye [x,c] or c e [x,y]. If a e [x,b] then
a e [x,c] and hence a< c. If x e [a,y] then ye [x,c] implies ye [a,c], by (L2),
and c e [x,y] implies c e [a,y], by (L4). In both cases a< c. Assume next that
x e [b,y] and either c e [b,y] or ye [b,c]. If x e [a,y] and be [a,y] thence [a,y]
or ye [a,c], respectively. Thus again a< c. All remaining cases may be discussed
similarly.
e
It remains to show that if a,b are distinct points of then either a< b orb< a.
Assume that neither relation holds. Then if a e [x,b] we must have x e [b,y], and
if x e [b,y] we must have be [a,y]. Hence be [x,y], by (L2), and b = x, by
Proposition l(ii). But this implies b =a, which contradicts our assumption.
Therefore a E [x,b]. Similarly we can show that b E [a,x]. Consequently
x e [a,b].
If x e [a,y] we must have a e [b,y] and hence x e [b,y], and if x e [b,y] we
must have b e [a,y]. Therefore x E [a,y], and similarly x E [b,y]. Thus either
a e [x,y] or y e [a,x]. But a e [x,y] would imply x e [b,y], by (L2), and
ye [a,x] would imply x e [b,y], by Proposition l(i). Thus we cannot escape a
contradiction.
This proves that the relation< is a total ordering of the line t such that x s y.
Suppose now that a< band c e [a,b]. If (i) holds, thence [x,b] and hence c < b.
Moreover a e [x,c] and either ye [x,c] or c e [x,y], so that a< c. Similarly it may
be seen that a < c < b if (ii) holds.
Suppose on the other hand that c e t and a< c < b. We will show that
c e [a,b]. Indeed, if a e [c,b] then, by what we have just proved, c <a Sb and
hence c =a. Similarly if be [a,c], then c = b.
Finally, suppose that < is any total ordering of t with the properties in the
e
statement of the proposition. Let a,b be any distinct elements of and assume first
that (i) holds. If b < x, then b <a< x and x Sb< y, since y E [x,b]. Hence b =a,
which is a contradiction. Consequently x < b and x < a < b. Assume next that (ii)
holds. Then a <x < y and either a< b < y or a< y < b, since y :I: a. In any event
a< b, and so the total ordering is the one originally defined. D
1. Lines 53
PROPOSITION 5 The union of two segments with more than one common
point is again a segment.
Proof Let [a,b] and [c,d] be two segments whose intersection contains the distinct
points x,y. Assume the line e =<x,y> totally ordered, as in Proposition 4, and
choose the notation so that a< b, c < d and a< c. Then y < b, since ye [a,b], and
c <x, since x e [c,d]. If d < b, then [c,d] s: [a,b]. On the other hand if b < d,
then a < c < x < y < b < d and hence
(a) Let a,b 1,b2 be non-collinear poi11ts and suppose c1 e (a,b 1), c2 e (a,b 2). If
de [b 1,c2] f"\ [b2,ci], then actually de (b 1,c2) f"\ (b2,c 1).
(b) Let a,b 1,b2 be non-collinear points and suppose ci e (a,b 1}, c2 e (a,b 2],
c e (c 1,c 2). If c e [a,b]for some be [b 1,b 2], then actually c e (a,b) and
b e (b 1,b2). D
(c) Let a,b 1,b2 be non-collinear points and suppose c 1 e (a,b 1), c 2 e (a,b 2],
be (b 1 ,b 2 ). If there exists a point c e [a,b] r.. [c 1 ,c 2 ], then actually
c e (a,b) r.. (c 1,c2). D
(d) Let a,a',b,b' be non-collinear points and let x e (a,a') r.. (b,b'). If x e [y,y1
for some ye (a,b) and y' e [a',b1, the11 actually x e (y,y') and y' e (a',b'). D
(P)' if a,bi ,b2 are non-collinear points and if c1 e (a,bi), c 2 e (a,b 2), the11
there exists a point d e (bi ,c2) r.. (b2,c1). D
PROPOSITION 11.2' Let a,bi ,b2 be 11on-collinear points and let c 1 e (a,b 1),
c2 e (a,b 2]. If c e (c 1,c2), then c e (a,b) for some be (b1,b2). D
PROPOSITION 11.17' Let a.bi ,b2 be non-collinear points and let c1 e (a,b1),
c2 e (a,b 2]. If be (b 1,b2), then there exists a point c e (a,b) r.. (c 1,c2). D
2 LINEAR GEOMETRIES
affine if x,y e A implies <x,y> s A. For example, the line <x,y> is itself affine,
by Proposition 2.
It follows at once from the definition that the collection SL of all affine sets is a
normed alignment on X. For any set S s X, we define its affine hull <S> to be the
intersection of all affine sets which contain S. Our notations are consistent, since
the affine hull of the set {x,y} is the line <x,y>. Evidently, for any set S s X, we
have S s [S] s <S>. The next result, inspired by Bennett and Birkhoff (1985), is
of fundamental importance:
Proof It is sufficient to show that the set A = Ux,x'eC < x, x'> is affine, since it is
certainly contained in any affine set which contains C. Evidently we may assume
that C contains more than one point. Thus we wish to show that if b e <xi ,x2>,
c e <x3,x4 >, and a e <b,c>, where Xi , ••• ,x4 e C and Xi -:1: x 2, x 3 -:1: x4 , b -:1: c, then
there exist x 5 ,x6 e C with x 5 -:1: x 6 such that a e <x5 ,x6>. This is evident if
c e <xi,x2>, since then also a e <x(,x2>. Consequently we assume c e <xi,x2>,
and similarly b ~ <x3,X4>.
By symmetry, and the convexity of C, we need only consider the following
five cases:
Consider first case (A). By (P) there exists a pointy e [a,x2] n [c,xi]. Since
c e <xi ,x2 >, we have y ~ x 2 and hence a e <x2,y>. Consequently we may
assume ye C. By (P) also there exists a point x 5 e [y,x 3 ] n [xi,x4]. Then
x5 e C and x5 e [a,x2,x3]. Hence x5 e [a,x6] for some x6 e [x2,x3]. Thus x6 e C
and a e <x5 ,x6> if x 5 -:1: x 6• We can now assume x5 = x 6 e [x2 ,x3]. If x 3 -:1: x 5
then x 2 e <x3,x5>, ye <x3,x5>, and hence a e <x3,x5>. Consequently we can
now assume x 3 = x 5 e [xi,x4]. Then, by (L2), x 4 e [c,xi]. Since ye C it
follows from (L4) that y e [c,x 4 ] and then from Proposition l(i) that
56 Ill. Linearity
It follows that all the results of Chapter I, Section 5 are valid for the alignment
of affine sets in a linear geometry. With the definition of dimension given there,
the empty set 0 has dimension -1, a point has dimension 0 and a line has
dimension 1. An affine set of dimension 2 will be called a plane. For finite-
dimensional X, an affine set H s X is a hyperplane if and only if
dim H =dim X - 1.
Pasch (1882) pointed out the incompleteness of Euclid's axioms for geometry
and introduced the following additional axiom: 'If a line in the plane of a triangle
does not pass through any of its vertices but intersects one of its sides, then it also
intersects another of its sides.' It will now be shown that Pasch 's axiom holds in
any linear geometry. (It is not difficult to show that, conversely, any convex
geometry is a linear geometry if it satisfies the axioms (L2 )-(L4) and Pasch 's
axiom.)
[a,b] ("\ <c,d> =0, [a,c] ("\ <b,d> =0, [b,c] ("\ <a,d> = 0,
then d E <a,b,c>.
Proof The hypotheses evidently imply that a,b,c are not collinear and
d E [a,b] u [a,c] u [b,c]. We will assumed e <a,b,c> and derive a contradiction.
It follows from Proposition 11.16 that if de [a,b,c], then de [a,e] for some
e e [b,c], which contradicts [b,c] ("\ <a,d> = 0. Thus we now suppose
d E [a,b,c].
By Proposition 7, de <p,q>, where p,q e [a,b,c] and p ~ q. Moreover we
may choose the notation so that q e (p,d). Furthermore, by Proposition 11.16, we
may take p,q e [a,b] u [a,c] u [b,c].
If p,q e [a,b], then de <a,b> and either a e (b,d) orb e (a,d). In either case
<a,d> ("\ [b,c] ~ 0. It is obvious also that <a,d> ("\ [b,c] ~ 0 if q =a andp e [b,c]
orifp=aandqe [b,c].
58 Ill. Linearity
Proof We show first that .e cannot intersect (a,b), (a,c) and (b,c). By symmetry it
is sufficient to show that if de (a,b), e e (a,c) and/ e (b,c), thenf E (d,e). But if
fe (d,e) then, by Proposition 11.2',/ e (a,g) for some g e (b,c). Hence a e <b,c>,
which is a contradiction.
It remains to show that .e intersects two 'sides' of the 'triangle' [a,b,c]. Since
a,b,c E .e, this follows at once from Proposition 11.16 if .e contains two distinct
points of [a,b,c]. Thus we now assume that .e contains a point e E [a,b,c]. Since
e e <a,b,c>, it follows from Lemma 9 that the points a,b,c may be named so that
there exists a point p e <a,e> ("\ [b,c]. Thus the hypothesis is now that .e contains a
point de (a,b) u (a,c) u (b,c).
Suppose first that de (a,b), which implies p :I: b. If a e (e,p), then de (eJ)
for some f e (b,p). If p e (a,e), then there exists a point/ e (d,e) ("\ (b,p). In both
cases/ e <d,e> ("\ (b,c).
Suppose next that de (b,c). We may assume that p :I: d and, without loss of
generality, that de (b,p). If a e (e,p), then there exists a point/ e (d,e) ("\ (a,b).
If p e (a,e), then de (el) for some/ e (a,b). In both cases/ e <d,e> ("\ (a,b).
The cased e (a,c) is reduced to the cased e (a,b) by interchanging band c.
D
Proof Obviously we can assume X3 ~xi and x 4 ~ x 2 • The lines <xi,x 2>,
<x2 ,x3> and <x3,x4> intersect H in unique points h1th 2 and h 3. Moreover
hi E (Xi,X2), h2 E (X2,X3) and h3 E (X3,X4).
Suppose first that xi,x2,x3 are collinear. Then hi= h2 and either Xie (h 2 ,x3)
or x 3 e (h 2,xi). If Xi e (h 2,x 3) then, by (P), there exists a point h e
[h 2,h3] n (xi,x4). If x 3 e (h2,xi) then, by (C), there exists a point he [xi,x4]
such that h 3 e [h 2 ,h]. If h 3 = h2, then h2 ·e (x 1,x4). If h3 ~ h 2 , then h e
H n (X1,X4).
Thus we may assume that x 1,x2 ,x3 are not collinear, and hence h 1 ~ h 2 •
Suppose now that x 1 ,x 3,x4 are collinear, so that h 3 e <x 1 ,x 3>. Since
<h 1,h 2 ,h3> ~<x 1,x 2,x3>, we must have h3 e <h1th 2>. By Proposition 10,
h 3 ~ (x 1,x3). Since x 1,x3,x4 are collinear, it follows that h 3 e (x1,x4).
Thus we may assume that x 1,x3 ,x4 are not collinear and indeed, by similar
arguments, we may assume that no three of the points x 1,x2,x3,x4 are collinear.
Suppose next that <x2,x4> n H ~ 0. Then the line <x2 ,x4> intersects Hin a
unique point h. We show first that hi!: (x 2,x4 ). Assume on the contrary that
h e (x2,x4). Since x 2,x3,x4 are not collinear, it follows from Proposition 10 that
h ~ <h2,h3>. Hence <h2,h3,h> = <x2,x3,x4>, which is a contradiction.
Thus either x 4 e (x 2 ,h) or x 2 e (x4,h). If x 4 e (x 2 ,h) then, by (P)', there
exists a point h' e (h,h 1) n (x1,x4). If x2 e (x4,h) then, by Proposition 11.2', there
exists a point h' e (x1,x4) such that h 1 e (h,h').
Thus we may assume that <x 2 ,x4 > n H = 0, and similarly also that
<x 1,x3> n H = 0. By (P)', there exist a point x e (x2,h 3) n (x4,h 2) and a point
h e (x,x3 ) n (h 2,h3). Since <x2 ,x> n H ~ 0, it follows from the previous part of
the proof (with x4 replaced by x) that (x 1,x) n H ~ 0. Since <x,x4> n H ~ 0, it
further follows from the previous part of the proof (with x 2 replaced by x) that
(X1,X4) n H ~ 0. D
PROPOSITION 12 Let H be a hyperplane such that X \ H is not convex. Then
there exist unique nonempty convex sets H + and H _ such that
X\H = H+ u H_. Furthermore,
(i) H+ n H_ = 0,
(ii) ify e H+ and z e H_, then (y,z) n H ~ 0,
(iii) Hu H+ and Hu H_ are also convex.
60 Ill. Linearity
Proof Since X \ H is not convex, there exist points a,b e X \ H and c e H such
that c e (a,b). Let H+ denote the set of all points x e X such that (b,x) r'\ H 0, *
and let H _ denote the set of all points x e X such that (a,x) r'\ H 0. Then *
a e H +' b e H _ and H + r'\ H = 0, H_ r'\ H = 0.
We are going to show that also H + r'\ H _ = 0. Assume on the contrary that
there exists a point x e H + r'\ H _. If x e <a,b> then c e (a,x) r'\ (b,x), since
<a,b> r'\ H = {c}.
But this is impossible, since also c e (a,b). Thus a,b,x are not
collinear. If (a,x), (b,x) intersect H in h,h' respectively, then c E <h,h'> by
Proposition 10. Hence <c,h,h'> = <a,b,x>, which is a contradiction.
We show next that H_ is convex. Assume on the contrary that there exist
points x',x" e H _ and a point x e (x',x") such that x E H _. Then there exist points
h',h" e H such that h' e (a,x'), h" e (a,x"). Hence a,x',x" are not collinear and
h' * h". By Proposition 11.17' the segment (h',h") contains a point he (a,x). Then
he Hand x e H_, which is a contradiction. Thus H_ is convex, and similarly also
H+.
It follows at once from Lemma 11 that if y e H + and z e H _, then
*
(y,z) () H 0. It will now be shown that X =Hu H+ u H_. Since His a
hyperplane, it is sufficient to show that X': =Hu H + u H_ is affine. In fact it is
enough to show that if x' e H +' x" e X' and x e <X',x">, then x e X'. Since
x' e H+, the segment (b,x') contains a point h' e H.
Suppose first that x" = h" e H. If h" e (x,x') then x e H_, since x' e H +· If
x e (x',h") then, by (P), there exists a point h e [h',h"] () (b,x) and hence x e H +·
If x' e (x,h") and b,x,x' are collinear, then h' = h" e (b,x) and hence x e H +· If
x' e (x,h") and b,x,x' are not collinear then, by Proposition 11.2', there exists a
point he (b,x) such that h' e (h,h") and again x e H+. This proves that a line
containing a point of H and a point of H + is entirely contained in X', and similarly a
line containing a point of Hand a point of H_ is entirely contained in X'.
Suppose next that x" e H+ u H_ and the line <x',x"> contains no point of H.
Then x" e H+, since x" e H_ would imply (x',x") () H * 0. Hence the segment
(b,x") contains a point h" e H. If x e (x',x") then x e H+, since H+ is convex.
Now consider the case x' e (x,x"). By (P) ', there exists a point
ye (b,x') r'\ (h",x). Moreover y E H. If h' e (b,y), then the segment (b,x)
contains a point h such that h' e (h,h'') and hence x e H +· H h' e (x',y), then the
segment (x,x') contains a point h such that h' e (h,h'') and hence x e X' by what we
2. Linear geometries 61
have already proved. Since the same argument applies in the case x" e (x,x1, this
completes the proof that X = H u H+ u H_.
It now follows that Hu H +and Hu H_ are convex. For if (say) x 1 e H and
x 2 e H+, then (x 1,x2) rt H =0 and hence (x 1,x2) rt H_ =0.
Finally, suppose X\H is the union of two nonempty convex sets G+ and G_.
We may assume the notation chosen so that G+ rt H+ :1: 0. But then G+ s H+, by
(ii). Hence G _ n H _ :1: 0 and so, in the same way, G _ s H _. Since
G+ u G_ =H+ u H_, we must actually have G+ =H+ and G_ =H_. D
Proposition 12 says that a hyperplane has two 'sides', if its complement is not
convex. The convex sets H + and H _ in the statement of Proposition 12 will be
called the open half-spaces associated with the hyperplane H, and the convex sets
H+ u Hand H_ u H will be called the closed half-spaces associated with H.
(When X \ H is convex, we may call X \ H the open half-space and X the closed
half-space associated with the hyperplane H. This situation arises if, for example,
X is itself the closed half-space associated with a hyperplane H of a real vector
space.) Our topological terminology will be given more justification in Chapter V.
Lemma 9 can be generalized to 'higher dimensions in the following way:
Proof Assume on the contrary that [x 1,x 2] rt H 3 = 0. Then x 1,x 2 ,x3 are not
collinear and hence Z :I: 0. Moreover, by Lemma 9,
(*)
where Yj e [x1,x2 ,zj] for some zi e Z U= 1,2). Moreover(*) implies that z1 :I: z2,
Yj ~ [x 1,x 2] for j = 1,2, and x 3 ~ fy 1,y2]. Without loss of generality we may
suppose y 1 e (y2,x3).
62 111. Linearity
Put
Then dim X' = 3 and x 3 e X'. Furthermore Hk' = <xk,z 1,z 2 > (k = 1,2,3) are
distinct planes inX' with the common line Z' = <z 1,z2>. Moreover
it follows that [y2,z2] (') <x1,x2,z 1> = 0. Thus y 2 and z2 lie in the same open half-
space of X' associated with the plane <x1,x2,z1>. Since
it follows that x 3 and z2 lie in different open half-spaces of X' associated with the
plane <x1,x2,z 1>. Thus there exists a point
The join of two convex geometries was defined in Chapter II and, in the
terminology of this chapter, it was shown there that the join of two linear
geometries is again a linear geometry. A linear geometry will be said to be
irreducible if it is not itself the join of two linear geometries. Since irreducible
2. Linear geometries 63
linear geometries are the building blocks of which all linear geometries are
composed, it is of interest to detennine their nature. To this task we now turn.
LEMMA 14 Let x,y 1,y2 be non-collinear points such that <x,y 1> :I:- {x,y 1} and
<x,y2> :I:- {x,y2}. If [y 1,y2] = {y 1,y 2 }, then there exist points z 1,z 2 such that
x e (y 1,z 1) () (y 2,z2) and the plane <x,y 1,y2> contains only the five points
x,yl ,y2,z1 ,z2.
Proof Let z" e <x,yk> with z" :I: x,yk (k = 1,2). Assume first that z2 e (x,y 2). If
y 1 e (x,z 1) then, by (P)', there exists a point we (y 1,y2) () (z 1,z2). If x e (y 1,z 1)
then, by Proposition 11.2', z2 e (y,z 1) for some ye (y 1,y2). If z 1 e (x,y 1) then, by
(P)', there exists a point u e (y 1,z2) n (y2 ,z 1) and, by Proposition 11.2', u e (x,v)
for some v e (y1,y2). Since (y 1,y2) =0 we conclude that (x,y2) =0, and likewise
(x,y 1) = 0.
Assume next that y 2 e (x,z 2). If x e (y 1,z 1), then y 2 e (zltu) for some
u e (y 1,z2) and there exists a point v e (x,u) n(y 1,y2). If y 1 e (x,z 1), then there
exist a point u' e (y1'z 2 ) () (y 2 ,z 1) and a point v' e (x,u) n (y 1,y 2 ). Since
(y 1,y2) = 0 we conclude that x e (y2,z2), and likewise x e (y 1,z 1).
Assume now that the line <x,y1> contains a point z '#: x ,y 1,z 1. Then in the
same way x e (y 1,z) and by interchanging z and z1, if necessary, we may suppose
that z e (x,z 1). But then z e (y 2,z') for some z' e (z 1 ,z 2) and hence, by
Proposition 11.19', x e (y',z') for some y' e (y 1 ,y 2 ). We conclude that
<x,y 1> = {x,y1,z 1}, and similarly <x,y2> = {x,y2,z2 }.
Assume next that the line <y 1,y2> contains a point z '#: y 1,y2• Without loss of
generality we may suppose y 2 e (y 1,z). Then x e (z,u) for some u e (y1,z2) and
there exists a point v e (u,y 2 ) n (x,y 1). Since (x,y 1) = 0, we conclude that
<y1,Y2> = {y1,Y2}.
Assume finally that the plane <x,y 1,y2> contains a point z '#: x,y1,Y2,z 1,z2.
Then
[x,yi] n <y2,z> = 0,
COROLLARY 15 If not all points of the linear geometry X are collinear and if
every line contains at least three points then, for any distinct points a,b e X,
(a,b) -:1: 0. D
Proof It is clear thatX is irreducible in each of the cases (i)-(iv), and that no two
of these cases can coexist. To prove the proposition we will assume that none of
(i)-(iv) holds and derive a contradiction.
Since X is irreducible, our assumptions imply that X contains at least three
points and not all points are collinear. If A is a nonempty proper affine subset of X
then, for some a e A and some x e X \ A, the line <a,x> contains more than two
points (since otherwise X \ A would be affine and X would be the join of A and
X\A).
In particular some line in X contains at least three points. It follows from
Hausdorffs maxima!ity theorem that there is an affine subset A of X, every line of
which contains at least three points, but which is not properly contained in another
affine subset with the same property. Since (iii) does not hold, it follows from
Corollary 15 that A-:!: X. Hence there exist points x e X \A and z e A such that
<z,x> contains at least three points.
Assume there exist distinct points x 1 ,x 2 e <A u x> \ A such that
2. Linear geometries 65
<xi,..t2> = {xi,x2 }. Then <A u x> =<A u Xi> and it follows from Proposition 7
that x2 e <x',x">, where x' e [a',xi] and x" e [a" ,Xi] for some a',a" e A. Since
Xi e= <a',a"> and x 2 e= <a',a"> u <a',xi> u <a",xi>, we may suppose that
x' e (a',xi) andx" e [a",xi). It follows from Lemma 14 thatx' e (x2,x") and that
the plane <a',a",xi> contains only five points. Hence x" =a" and the line <a',a">
contains only two points, which is a contradiction.
We conclude that if xi,x 2 are distinct points of <Aux> \A, then
<xi,x2> :I:- {xi,x2 }. But, since A is maximal, there exist distinct Yi,y2 e <Aux>
such that <yi,y2 > = {yi,y2 }. We may choose the notation so that Yi= a e A and
y 2 e= A. If <a,x> ¢ {a,x}, then y 2 :1:-x and <y2,x> :I:- {y2,x}. Hence, by Lemma
14, there exists a pointy such that x e (a,y) and <y,y2 > = {y,y2 }. Since ye= A,
this is a contradiction. Thus we may take y2 = x.
Since the lines <z,x> and <z,a> contain at least three points, it now follows
from Lemma 14 that there exist points a',x' such that z e (a,a') n (x,x') and the
plane <z,a,x> contains only the points z,a,x,a',x'. Hence, by Corollary 15,
A = { z,a,a'}.
Choose any point be= <z,a>. If <z,b> :I:- {z,b}, then any affine subset
properly containing <z,b> has a line containing only two points. In particular the
plane <z,a,b> has a line containing only two points. The preceding argument, with
A replaced by <z,a> and x by b, shows that there exists a point b' such that
z e (b,b1 and <z,a,b> = {z,a,b,a',b'}.
Let C be the union of z with the set of all points c e X such that <z,c> ¢ { z,c}.
It follows from what has been said that C is affine. Moreover C :1:-X, since (iv)
does not hold. Hence there exist points p e X \ C and q e C such that
<p,q> :I:- {p,q}. Since <z,p> = {z,p}, we must have q :I:- z and <z,q> ¢ {z,q}.
Hence <z,q> = {z,q,q'}, where z e (q,q'). But since <p,q> :I:- {p,q}, it follows
from Lemma 14 also that q e (z,q1. Thus we have a contradiction. D
An irreducible linear geometry for which case (iv) of Proposition 16 holds may
be called a multi-cross with (uniquely determined) centre z, since each plane
containing z has exactly four other points, ui,u 2 ,vi,v 2 say, with
z e (ui,Vi) n (u2,v2). If Vis a vector space over an ordered division ring D with
basis {e;} ie/t then the restriction of V to the set X = {O,e;,--e;} ie/ is a multi-cross
with centre 0.
66 111. Linearity
Proof By Proposition 11.13 we need only show that every affine independent set is
a Radon set, and that every Belly set is affine independent. Moreover we may
restrict attention to finite sets.
Assume first that S = {s 1, ••• ,snl is an affine independent set, but not a Radon
set. Then we may choose the notation so that [Si] n [S 2 ] 0, where
':/:
S 1 = {s1 , •.. ,Sm} and S2 = {sm+l , ... ,sn} for some m with 1 < m < ti. Let
x e [Si] n [S 2]. Then x ':/: s 1, since [S2] s: <S\s 1> and Sis affine independent.
On the other hand x e <s 1, ••• ,sm>. Hence we can choose k, with 1 < k < m, so
that x ~ <s 1 , ••• ,sk> but x e <s 1 , ••• ,sk+t>. Then sk+l e <x,slt···,sk>, by
Proposition 8. Since x e <sm+lt···,sn>, it follows that sk+l e <S\sk+ 1>, which is
a contradiction.
To show that a finite Belly set is affine independent we use induction on the
cardinality of the set. Since the result is true for singletons, it is sufficient to
establish the following assertion:
(#) If T = {x 1, ••• ,xn} is an affine independent set and if Xn+l e <T> \ T, then
S =Tu Xn+l is not a Belly set
where u e [T \(xi u xk)]. Hence z' e [xk,xn+l•u], and so z' e [u, v], where
v e [xk,xn+il· Since u e [T\ (xj u xk)] and z' e [T\xk], it follows that
If v e T\xk, say v = X;, then Xn+l e <x;,X/r? and we are reduced to the case n =2.
If v t! T\xk then, by the induction hypothesis, the set
is not a Belly set. Thus there exists a point x e [T\xtJ such that x e [R \xi] for all
j e { 1, ... ,n} with j :I: k. Since v e [xk,xn+il' it follows that x e [S \x;] for
i = l, ... ,n+ 1. D
COROLLARY 18 For any set S s: X alld ally x e [S], there exists a finite
affine independent set F s: S such that x e [F] .
•
Proof Let F be a finite subset of S such that x e [F], but no proper subset of F
has the same property. Then F is a Caratheodory set. Hence F is a Helly set, by
Proposition 11.15, and thus also an affine independent set, by Proposition 17. D
Proposition 17 no longer holds if all the axioms for a linear geometry are
satisfied with the exception of (P). This is shown by the example illustrating the
independence of the axiom (P) in Example 11.10. In this example S = {a,a',b,b'} is
a Helly set, but it is not affine independent since a' e <a,c> s: <a,b,b'>.
If Sis an affine independent subset of Rd, then ISi S d + 1. Consequently, by
Proposition 17, any Belly set in Rd has at most d + 1 elements and hence, by
Proposition 1.10, a finite family of at least d + 1 convex sets in Rd has nonempty
intersection if every subfamily of d + 1 sets has nonempty intersection. This was
first proved by Helly (1923).
By Proposition 17 also, any Radon set in Rd has at most d + 1 elements.
Hence any subset S of R. d with ISi > d + 1 has disjoint subsets S 1,S 2 such that
[Si] n [S 2] :1: 0. This was first proved by Radon (1921), who used it to give the
first published proof of Helly's theorem.
Similarly it follows from Corollary 18 that if S s: Rd and x e [S], then there
exists a subset F of S with IFI < d + 1 such that x e [F]. This was first proved by
Caratheodory (1911).
68 Ill. Linearity
Proof We may assume, without loss of generality, that x = 0 and that JRd is
equipped with the Euclidean norm. Furthermore, by Caratheodory's theorem, we
may suppose that each set S; contains at most d+ 1 points. Then the collection 9' of
all sets S = {x 1, ••• ,xd+d withx; e S; (i = l, ... ,d+l) is finite. For any Se 9', put
Since [S] is compact, we have d(S) = llzll for some z e [S]. We wish to show that
d(S) =0 for some S e 9'. Since 9' is finite, it is sufficient to show that if d(S) > 0,
then there exists an S' e 9' such that d(S') < d(S).
If S is not affine independent then, by Caratheodory's theorem, z is a convex
combination of at most d points of S. If Sis affine independent, then the nearest
point of [S] to the origin lies on some proper face. In either case z e [S \xi] for
some j e { l, ... ,d+l }.
The origin cannot be separated from Si by any hyperplane, since 0 e [Sj].
Consequently the open half-space {x e Rd: (x - z,z) < 0}, which contains the
origin, must also contain a point x/ e Si. Let S' be the set which is obtained from S
by replacing xi by x/ and retaining the other x;. Then [S1 contains the segment
[x/.z], since it contains both x/ and z. But, for small A.> 0,
3. Helly's theorem and its relatives 69
Let S 1, ••• ,Sm be subsets of a linear geometry X such that dim <S 1, ••• ,Sm>
< m. If x e [S 1] r. ... r. [Sm], does there exist an affine independent set
{x 1, ••• ,xn}, with Xj e S ;ifor distinct illtegers i1 , ••• ,in e { l, ... ,m }, such that
XE [X1, ..• ,Xn]?
Proof We can identify JRd with the hyperplane ~ 1 + ... + ~d+t = 1 in JRd+l. Let
v0 ,v 1, ••• ,vn, where n = (r-l)(d+l), be the points of Sin this identification. With
each point v; e S we associate a set S;' = {M;i, ... ,M;,} of (d+l)x(r-1) matrices
defined in the following way: if 1 < k < r the matrix M;k has v; as its k-th column
and zeros elsewhere, whereas every column of M;, is-v;. If we also identify the
set of all (d+l)x(r-1) matrices with JRn, we haven+ 1 sets S;' s JRn with IS;'I = r.
Moreover, the origin is contained in the convex hull of every set S; ', since
M ;1 + ... + M ;, = 0. Consequently, by Barany's theorem, there exist a matrix
M;k; e S;' and A;> 0 with I,~ 0 A;= 1 such that
L~o A.;M;k; = 0.
Let Ii denote the set of all i e {0, 1,... ,n} for which k; = j. Then we can rewrite the
last equation in the form
Equivalently,
70 Ill. Linearity
Put µj = I, ieli A;. Since each v; lies on the hyperplane ~ 1 + ... + ~d+l = 1 and
L7=o A;= 1, it follows that µ1 =µ2 = ... =µ, = l/r. Hence, if we denote by sj
the set of all v; with i e Ji, the given set S has a partition into pairwise disjoint
subsets S1, ... ,S, whose convex hulls [S 1], ... ,[S,] have a common point. D
We show here that faces may be given yet another characterization in a linear
geometry, and that polytopes possess most of their familiar properties in 1Rd. A
notable omission is the Euler-Poincare relation. It is shown also that from a given
linear geometry further factor geometries may be derived.
1 FACES
Throughout this chapter we assume that a set Xis given on which a linear
geometry is defined. The notion of 'face' of a convex set was defined in any
aligned space in Chapter I, and in Chapter II a simpler characterization was given in
any convex geometry. Yet another characterization will now be given in any linear
geometry:
Proof Let A be a set such that Cr. <A>= A and C\A is convex. Then A s C
and A is convex, since it is the intersection of two convex sets. If a e (c,c1, where
a e A and c,c' e C, then at least one of c,c' is in A, since C \A is convex, and in
fact they both are, since C r. <A> =A. Thus A is a face of C, by Proposition 11.6.
On the other hand, if A is a face of C then C\A is convex, by Proposition 1.5.
Put B =Cr. <A>. Then A is a face of B, by Proposition 1.7, since B is convex
and A s B s C. By Proposition 111.7, if be B then be <a,a'> for some
a,a' e A. If be [a,a1 then be A, since A is convex. If a' e (a,b) or a e (a',b)
then be A, since A is a face of B. Thus B =A. D
72 IV. Linearity (continued)
Proof Put D = C n H and let H+ be the open half-space of <C> associated with H
such that Cs Hu H+. Then C n <D> =D, since D s C n <D> s C n H, and
C\D is convex, since C\D = C n H+. D
PROPOSITION 5 Let C be a convex set and S any subset of C. Then the subset
As of C, consisting of the union of C n <.S> with the set of all points x,x' e C
such that (x,x') n <S> '¢ 0, is a face of C. Moreover As is contained in every
face of C which contains S.
Proof We show first that As is convex. Suppose z e (x,y), where x,y are distinct
elements of As with x E <.S> and z E <S>. Then z e C and there exist points
x' e As, a e <.S> such that a e (x,x'). If a,x,y are collinear, then a e (z,z') for
some z' e {x,y,x'} s C and so z e As. Hence we now suppose that a,x,y are not
1. Faces 73
collinear. Then, by (P)', there exists a point c e (a,y) n (x',z). If ye <S>, then
c e <S> and hence z e As. If y E <S>, there exist y' e As and b e <S> such that
b E (y,y1.
If b =a then, by Proposition II.19', a e (z,z1 for some z' e (x',y1 and hence
z e As. Thus we now suppose b ':/!a and hence y E <a,b>. By (P)', there exists a
point de (a,b) n (c,y1. Moreover de C n <S>, since a,b e C n <S>. Thus we
may suppose d ':/! z. We may also suppose that y' E <x',z>, since otherwise
de (z,x1 or de (z,y1. Then, by Proposition 11.2', there exists a point e e (x',y1
such that de (e,z). Thus z e As in every case.
We now show that As is a face of C. Suppose x e As and x e (c,c1, where
c,c' e C. If x e <S> then c,c' e As, by the definition of As. If x E <S>, then there
exist x' e As and a e <S> such that a e (x,x1. If a,c,c' are collinear, then c =a or
a e (c,x) or a e (c,x1. In any event c e As, and likewise c' e As. Thus we now
suppose that a,c,c' are not collinear. Then, by Proposition 11.2', a e (c,y) for some
y e (c',xl Hence c e As, and likewise c' e As.
It follows at once from Proposition 1 that any face of C which contains S must
also contain As. D
Proof Assume on the contrary that three distinct facets F 1,F2,F3 of a convex set C
have a common nonempty facet F. Then <F 1>,<F 2 >,<F 3 > are distinct
hyperplanes of <C> with the common hyperplane <F>. Choose xk e Fk \F
(k = 1,2,3). Since C n <Ft> =Fk and C \ Fk is convex, we have
2 POLYTOPES
Proof By Proposition 1.11, P = [S]. We show first that if Fis a face of P, then
F = [71 for some T s S such that [S \ T] r. <T> =0. Evidently we may assume
that F ¢ 0, P. Let T denote the set of all extreme points of P which are contained
in F. Then [S\71 s P\F, since P\F is convex, and thus Tis a nonempty proper
subset of S. If x e F, then x e [y,z] for some y e [S \ 71 and z e [T]. Thus
ye P \F and z e F. Since x ¢ z would imply ye <.F>, which contradicts
Pr. <F> = F, we must have x = z e [T]. Hence F = [T] and [S\ T] r. <T> = 0.
We show next that [T] is a face of P if T is a subset of S such that
[S \ n r. <T> = 0. Let x E pr. <T>. Then x E [y,z], where y E [S \ n and
z e [71. In fact x = z, since x ¢ z would imply y e <T>. Thus P r. <T> = [T]. It
remains to show that P \ [T] is convex.
Suppose z e (x,y), where x e [S], ye [S \ T] and z e [T]. Then x e [y',z1,
where y' e [S \ 71 and z' e [T]. Since [S \ T] r. <T> = 0, we must actually have
x e (y',z') and x,y,z' are not collinear. Hence there exists a point z" e (y,y') such
that z e (z',z"). Since z" e [S \ T] r. <T>, this is a contradiction.
If P \ [T] is not convex then, for some points x' ,x" e P \ [T], there exists a
point z e (x',x") r. [T]. Then x' e [y',z'], where y' e [S \ T] and z' e [T].
Moreover, by what we have already proved, x' e (y',z') and x",y',z' are not
collinear. Hence there exists a point z" e (x",y') such that z e (z',z"). Since
z" e P r. <1'> = [T] this yields a contradiction, as we have already seen. a
COROLLARY 9 If Pis a po/ytope and Fa face of P, then Fis a polytope and
E(F) = E(P) r. F. D
COROLLARY 10 A polytope has only finitely many faces. More precisely, a
polytope with n extreme points has at most 2n faces. D
2. Polytopes 75
Proof Suppose P = [S], where S = {s 1, ••• ,sn}. Since the result is obvious if
n < 2, we assume that n > 2 and the result holds for all smaller values of n.
Obviously we may assume also that t contains two distinct points x,y of P. The
points of the line t may be totally ordered, as in Proposition IIl.4. By the induction
hypothesis, [S \s;] (') t either is empty or has the form [a;,b;], with a;< b;, for
each i e {l,... ,n}. If we put a= min; a; and b =max; b;, then [a,b] s P(') t.
Moreover, by Proposition 11.16, P cannot contain a point oft\ [a,b]. D
It may be noted that the axiom (P) is not required in the proof of Proposition
12.
Proof Suppose P = [S], where S = {x 1, ••• ,xn}. Since the result is obvious if
n = 1, we assume that n > 1 and the result holds for all smaller values of n. Let
p e P (')L. By Proposition Il.14, we have P = Ui:1P;, where P; = [p u
(S\x;)].
It is enough to show that the sets P; (') L (i = 1,... ,n) are all polytopes. For if
P; (')Lis the convex hull of a finite set S; (i = l, ... ,n), then
The right side is certainly contained in the left, since p u P' = (p u P 11 n Land
P 1 = [jJ u P 11. On the other hand, the left side is contained in the right. For if
x e P 1 \ p then x e (p ,y] for some y e P 1 ', and if x e (P 1 n L) \ p then
ye P 1' n L. a
(#) Let S be the set of extreme points of P, let S 1 be a subset of S such that
L 1: = <S 1> c <S> and let F 1 be a facet of the polytope P 1 =P n L 1• If
L 2: =<b,S1>, where be S\L 1, then the polytope P2 = P n L2 has a facet F2 such
that F 1 c F 2•
If F 2 ' = P (')Li is a facet of P 2, then F 2 (') F 2' =Fi since H 2 (')Li= Hi. If
P (') Li is not a facet of P 2, then T (') N ~ 0 and we can in the same way choose
points d1t···,d,, e T (') N and show that F 2' = P (') H2', where H 2' = <d,,,Hi>, is
also a facet of P 2 containing Fi· Since the segment (bm,d,,) contains a point x e Li,
and H 2 ' = H 2 implies x e Hi' the facets F 2 and F 2' will be distinct if
(bm,d,,) (')Hi= 0. Thus they are certainly distinct if Fi is a face of P 2. Again,
F 2 (') F 2' =Fi since H2 (') H 2' =Hi. If we take Fi= F to be a face of P such that
dim F =dim P - 2, then P 2 = P and F is the intersection of the facets F 2,F2' of P.
Since Fis contained in at most two facets of P, by Proposition 7, this completes the
proof of the second statement of the proposition.
To prove the first statement of the proposition, take F 1 = F to be a face of P
such that dim F <dim P -2 and put Hi= <Fi>. Then Hi= <R>, where R is the
set of extreme points of P which are contained in F. If Li = <a,R>, where
a e S\Hi, then Fi is a facet of Pi= P (')Li. By repeatedly applying(#), we see
that Fis contained in a facet of P. D
n. If m =dim G - dim F then, since n > 2, either m > 1 or m < n - 1 (or both).
Without loss of generality suppose m > 1 and let Fi be a face of P such that
F c Fi c G and F is a facet of Fi· Then by the induction hypothesis there is a
face G" of P such that Fi s G" s H, Fi= G (') G" and every face which contains
both G and G" also contains H. Since G" ~ H, there also exists a face G' of P such
that F s G's G", F =Fir'\ G' and every face which contains both F 1 and G' also
contains G". Since G' s G",
and, since Fi s G, any face which contains both G and G' also contains G" and
hence also contains H. D
PROPOSITION 19 Let P be a polytope and Fi, ... ,Fm its facets. If Q; is the
closed half-space of <P> associated with the hyperplane H; = <F;> which
contains P (i = l, ... ,m}, then
Proof Put Q = nf!1 Q;. Then obviously P s Q. We will assume that there exists
a point a e Q \P and derive a contradiction. Since a e <P>, we have a e <x,y>
for some distinct points x,y e P, by Proposition III.7. Moreover, by Proposition
12, we may choose x,y so that P (') <x,y> = [x,y] and we may choose the notation
so that x e (a,y). Then x and y belong to proper faces of P, by Corollary 6.
Hence, by Proposition 15, x and y belong to facets of P. Thus x e H k for some k.
Since a and y cannot lie in different open half-spaces of <P> associated with the
hyperplane Hk' it follows that a,y e Hk and hence y belongs to the same facets of P
as x. Hence, by Proposition 17, y belongs to every face of P which contains x. In
particular, by Corollary 6 again, x e (y,y') for some y' e P. But this is a
contradiction. a
From Propositions 14 and 19 we immediately obtain
Proof Since the result is obvious if 11 = 3, we assume that n > 3 and the result
holds for two-dimensional polytopes with fewer than n extreme points. Let S
denote the set of extreme points of P. If e e S then, by Proposition 15, there exist
distinct e',e" e S such that [e,e1 and [e,e"] are facets of P. Suppose/e Sand
f-::1: e,e',e". Then <e,e'> n [e"J] = 0 = <e,e"> n [e'J]. Hence, by Lemma IlI.9,
<el> n [e',e"] -:I: 0. Since all points involved are extreme points of P, we must
actually have {e/) n (e',e") -:I: 0. It follows that [e',e"] is a facet of the polytope
[S\e]. The result now follows from the induction hypothesis. a
With any polytope P there is associated a graph, whose vertices are the
extreme points of P and whose edges are the edges of P. It will now be shown that
the graph of a d-dimensional polytope is d-co11nected, i.e. it is connected and
remains so whenever less than d vertices are removed. This was first proved, for
polytopes in Euclidean space, by Balinski (1961).
Proof Fix e e E. It follows from Lemma 22, with F = {e}, that there exists an
affine independent set Vof d+l extreme points of P, including e, such that [eJ] is
an edge of P for every f e V \ e. Hence we may assume that e' ~ V and that the
result holds for all d-dimensional polytopes with fewer extreme points than P.
Since <V> =<P>, it follows that e' e <S \ e'>.
By applying Proposition 19 to the polytope Q =[S \ e1, we see that there
exists a facet F of Q such that e' does not lie in the closed half-space of <P> = <Q>
associated with the hyperplane <.F> which contains Q. We have F = [T], where
Tc: S\e'. If/ e T then [e'J] is an edge of the polytope [e' u T], by Proposition
8, and hence also of P, by Proposition 4. Thus we may assume that e e U, where
U = S\ (e' u n. By the induction hypothesis any two distinct extreme points of Q
which are in E can be connected in E by edges of Q. Hence, by Proposition 4, e
can be connected in E to some extreme point f e T (') E by edges of P.
Consequently e can also be connected to e'. a
3 FACTOR GEOMETRIES
In this section we show how new linear geometries may be created from a
given one. This property, and the theory of polytopes constructed in the previous
section, are persuasive evidence that the concept of linear geometry has intrinsic
significance.
Let X be an arbitrary linear geometry, let H+ be an open half-space associated
with a hyperplane Hof X, and letL be any nonempty affine subset of H. For any
a e H +' we denote by Pa the closed half-space of <a u L> associated with the
hyperplane L which contains a. In symbols, Pa =<au L> (') (H + u H).
Consequently Pa \L =<au L> (') H+ and <Pa>= <au L>. Evidently Pa·= Pa
for any a' e Pa \L.
The collection H + : L of all such, closed half-spaces Pa with a e H + can be
given the structure of a linear geometry, the factor geometry of H + with respect to
L, in the following way. For any p 1,p 2 e H+: L, we define the sector [p 1,p2] to
be the set of all p e H+: L such that a e [a 1,a2] for some a e p \L, a1 e Pt \Land
a2 e p2 \L. As we now show, we may actually fix a1 and a2.
82 IV. Linearity (continued)
Proof Evidently we may assume p 1 if: p2 and p if: p 1,p2. Then <P1> if: <P2>. By
hypothesis there exist points a' e p \L, a 1' e p 1 \L and a 2' e p2 \L such that
a' e [a 1 ',a 2 1. In fact a' e (a 1',a 2 '), since p if: p 1,p 2. If we put L; = <p ;> =
<a;' u L> (i = 1,2), then L 1 and L 2 are hyperplanes of L* = <a 1',a 2 ',L>.
Moreover L 1 n L 2 = L, by the exchange property. Since a' e (a1',a2 1, it follows
that L' = <p> =<a' u L> is also a hyperplane of L* and L' il:L 1,L2. Moreover a 1'
and a 2' lie in different open half-spaces of L* associated with the hyperplane L'.
From L 1 n L' =Land [a 1,a 11nL=0 we obtain [a 1,a 11nL'=0. In the same
way [a 2,a21nL'=0. Hence also a 1 and a 2 lie in different open half-spaces of
L* associated with the hyperplane L'. Consequently there exists a point a e
(a 1,a 2) n L'. Evidently a and a 2 lie in the same open half-space of L* associated
with the hyperplane L 1• But a and a' also lie in the same open half-space of L *
associated with the hyperplane L 1, since L 1 n L' = L. Therefore [a',a2] n L 1 = 0.
Assume a E p. Then ~here exist a point x e (a,a1 n L and a pointy e (a',a 2) such
that x e (a 1,y). Since ye [a',a2] n L 1, this is a contradiction. Thus a e p. 0
Proof By Proposition 11.16 there exist points y,z e [a,b] u [a,c] u [b,c] such that
a,x e [y,z]. Moreover a e {y,z}, since a is an extreme point of [a,b,c]. Without
loss of generality assume a= z. Then ye (b,c), since x E [a,b],[a,c], and x if: y,
since x E [b,c]. 0
Proof We will prove that H+: L satisfies each of the axioms (C), (Ll)-(L4) and
(P). In only one case, namely (L3), does the proof present difficulty.
(C): Suppose p' e [p,pi) and p" e [p',p2]. Choose a e p \L, at e Pt \L and
o 2 e p2 \L. Then, by Lemma 24, there exist a point o' e p'\L such that
o' e [o,oi) and a point o" e p"\L such that o" e [o',o2]. But o" e [o,d] for some
a E [Ot,a2]. If a e: L, it follows that p" E [p,p] for some p E Cpt,P21· On the
other hand, if a EL then o" E p \Land p" = p.
(P): Suppose Pt' e [p,pi] and p 2' e [p,p2]. Choose o e p \L, o1 e Pt \Land
a 2 e p2 \L. Then there exist a point at' e Pt'\L such that Ot' e [o,oi] and a
point o 2 ' e p 2'\L such that o 2 ' e [o,o 2]. Hence there exists a point
o'e [01'021 (') Cot',02]. Moreover o'e H+, since Ot,02'e H+. Consequently
Pa·E [pt,P21 (') [Pt',p2].
a' -:1: a", since a 1,a2 ,a3 are not collinear. Since a' e <a",L>, it follows from
Proposition 111.7 that a' e <a",/',/"> for some/',/" e L. Moreover <a',a"> n L =
0, since a',a" e <a 1,a2 ,a3>, and hence/' -:I:/".
Put Y = <l',l",a 1,a3>. Since a" e Y, we must have a' e Y and hence also
a2 e Y. Since Y contains the affine independent points a1,a2,a3,I', it follows that
dim Y =3. But, since Hk =<l',l",ak> (k = 1,2,3) are distinct planes in Y with the
common line Z = <1',I">, this contradicts Proposition Ill.13. D
4 NOTES
The faces of arbitrary convex sets do not share many of the properties
possessed by the faces of polytopes (or of polyhedra, which will be considered in
Chapter V). For example in R 3, with the usual definition of convexity, let
C =[p,q,S] be the convex hull of the set consisting of the circle S = {x =(~ 1 ,~ 2 ,~3):
(~ 1 - 1)2 + ~ 2 2 = l, ~ 3 = 0} and the two points p = (0,0,l), q = (0,0,-1). The
extreme points of C are p,q and all points of S except the origin. The edges of Care
the segments whose endpoints are distinct extreme points, not both of which are in
S. However, Chas no facets. (It may be noted also that, although C is compact,
the set of all extreme points is not closed.)
Polytopes in Euclidean space are extensively treated in the books of Br~ndsted
(1983), Grilnbaum (1967) and Ziegler (1995). The present treatment follows
Coppel (1995), but the key Proposition 15 has a stronger formulation here. It is
natural to ask if the Euler-Poincare relation, which connects the number of faces of
different dimensions of a polytope in JRd, can be generalized to any linear geometry.
Factor geometries are considered by Prenowitz (1961), Rubinstein (1964) and
Prenowitz and Jantosciak (1979), but they assume the axioms (D) and (U) which
will be considered in the next chapter.
v
Density and Unendingness
In this chapter two new axioms are introduced. A linear geometry is dense if no
segment contains exactly two points. The notions of intrinsic interior and convex
closure of a convex set are defined, and it is shown that in a dense linear geometry
they have many basic properties. In particular, any finite-dimensional convex set
has a nonempty intrinsic interior. A linear geometry is unending if any segment is
contained in another segment with different endpoints. An unending linear
geometry of dimension greater than one is necessarily also dense. In an unending
linear geometry the affine hull of the union of two affine sets A,B is the union of
the affine hulls of all sets {a,b}, where a e A and b e B, provided A (') B -:1: 0.
Furthermore a dense unending linear geometry may be given the structure of a
topological space, and actually of a Hausdorff space.
Following Prenowitz, we define products of arbitrary nonempty sets in a
dense linear geometry, and quotients also in a dense unending linear geometry.
Products and quotients possess many simple algebraic properties, and their
usefulness in elementary geometry has perhaps not been fully appreciated. We
further show that a substantial part of the usual theory of cones and polyhedra in
R.d carries over to any dense unending linear geometry.
It is remarkable that this axiom actually implies the axiom (P), when all the other
axioms of a linear geometry are satisfied:
86 V. Density and unendingness
By Corollary III.15, a linear geometry is dense if not all points are collinear
and every line contains at least three points. Furthermore, a dense linear geometry
is irreducible. Throughout the remainder of this section we will assume that a
set X is given on which a dense linear geometry is defined.
Proof We need only show that every nonempty finite affine independent set is a
Caratheodory set, by Propositions 11.15 and llI.17. Let S = {s 1, .•• ,snl be an affine
independent set. By (D) we can choose x 1 e (s 1,s2) and then, inductively,
xk e (xk-t,Sk+t) (k = 2, ... ,n - 1). Evidently Xn-l e [S] and, for each j < n,
Xj-l e Cstt···,sj] s <S\si+t>· However, xi~ <S\sj+t>, since Xj e <S\sj+t>
would imply Sj+t e <xi_ 1,xi> s <S\sj+t>, which contradicts the affine
independence of S. In particular,xn-1 ~ [S\sJ.
Assume, contrary to the proposition, that [S] = U7=1[S\s;]. Then Xn-1 e
[S \sm1 for some m < n. But if for some k with m < k < n we have xk e <S\sm>
then, since sk+l e S\sm, we also have xk-t e <xk,sk+t> s <S\sm>. It follows
that Xm e <S\sm> and, if m > l, even Xm-l e <S\sm>. Since the latter is a
contradiction, we conclude that m = 1 and x 1 e <S \ s 1>. But this is also a
contradiction, since St E <x1,S2> S <S\s1>. C
1. Dense linear geometries 87
The axiom (D) is essential for the validity of Proposition 2. For if x,y are
distinct points of a linear geometry such that [x,y] = {x,y}, then {x,y} is an affine
independent set but not a Caratheodory set.
PROPOSITION 3 If C is a convex set and Ai ,... ,Am affine sets such that
Proof The result is obviously true if m = 1. We assume that m > 1 and the result
holds for all smaller values of m. If C is not contained in Ai for some i then, by the
induction hypothesis, there exist a point a e C \ (U ~ 2 A;) and a point
be C\ (LJ~iiAi). Then a e Ai and be Am. If x e (a,b), then x e: Aiu Am.
Hence m > 2 and x e A 2 u ... u Am-i· Thus if we fix c e (a,b), then [a,c] s
Aiu ... u Am-i· Hence, by the induction hypothesis, [a,c] s Ai. Thus c e Ai,
which is a contradiction. D
set of ally e X such that [x,y) s Ci for some x e Ci. In either event C s C, by
Proposition 4.
In the preceding definition of 'convex closure' we can replace 'some' by
'every'. For suppose [x',y) s Ci and x e Ci. We wish to show that also
[x,y) s Ci. We may assume that x,x',y are not collinear, by Proposition 4. Since
x e Ci, there exists x" e C such that x e (x',x"). If z e (x,y), then z e (x",z~ for
some z' e (x',y). Since z' e Ci and x" e C, it follows that z e Ci.
The definitions of 'intrinsic interior' and 'convex closure' make sense, but
have little significance, in an arbitrary convex geometry and the preceding properties
hold in an arbitrary linear geometry. However, the following results will make
essential use of the axiom (D).
Proof To prove the first assertion we may assume that Ci contains more than one
point. Suppose x,y e Ci, where x '¢ y. Then there exists z e C such that x e (y,z).
Moreover, by (D), there exists some w e (x,z). Then w e Ci and, by Proposition
III. l(i), x E (y,w]. Hence w '¢ y and x E (y,w). Thus x E Cii and Cii = c;.
Since C; '¢ 0, u e C if and only if (u,v] s: Ci for every v e Ci. Since
cu= Ci, u e C; if and only if the same condition is satisfied. D
PROPOSITION 9 If C is a convex set and c e C;, then C n (c,x) '¢ 0 for each
XE <C>\c.
Proof It is enough to show that Ci s: Di, since the relation C s: D then follows
from the definition of convex closure. Obviously we may assume that Ci '¢ 0.
Let x e Ci and suppose first that there exists a pointy e C n Di, where y '¢ x.
Since ye C, there exists a point z e C such that x e (y,z). Since z e D and y e Di,
it follows thatx e Di.
Suppose next that <C> = <D> and let ye D \C. Then there exists a point
x' e C n (x,y), by Proposition 9, and there exists a point y' e C such that
x e (x',y'), since x e Ci. Since y' e D and x e (y,y'), it follows that x e Di. D
PROPOSITION 11 For any convex sets C,D with Ci'¢ 0, Di'¢ 0, the following
statements are equivalent:
Proof It follows at once from Propositions 7 and 8 that (i) and (ii) are equivalent.
Also it is obvious that Ci= Di implies Ci s: D, and that C = D implies D s: C.
90 V. Density and unendingness
P n <x,y> = [a,bJ. Then ye= [a,bJ, x e (a,b) and we may choose the notation so
that be (x,y). Hence be pi, which is a contradiction. CJ
Proof The formula for pi follows from Proposition IV.11 and the definition of
intrinsic interior. Since S is a Caratheodory set, by Proposition 2, it now follows
that pi'¢ 0. CJ
Proof Let S be a maximal affine independent subset of D and put C = [SJ. Then
Ci'¢ 0, by Proposition 15. But <C> = <D> and so, by Proposition 10, Ci s Di.
CJ
Proof Choose c e (a,b) and consider the family ~ of all affine sets which contain
c but not a. If H is the union of a maximal totally ordered subfamily of ~, then
He ~. Moreover, if x e X\H then a e <Hux> and hence x e <Hu a>, by
Proposition III.8. Thus <H u a> = X and H is a hyperplane. D
According to our definition, ab is not a segment if a '¢ b, since a,b ~ ab. However,
we will show that the analogues of the axioms (C) and (P) are satisfied.
We show first that if c e abi and de cb 2, then de ab for some b e bib2. If
a ,bi ,b2 are not collinear this follows from Proposition 11.2 ', and if a,bi ,b2 are not
all distinct it follows readily from (D). Suppose then that a,b 1,b2 are distinct but
collinear. If a e (bi,b2) then either d =a and we can take b =a, or de (a,c) and
we can take any be (c,d), or de (a,b 2) and we can take any b e (d,b 2). If
b 2 e (a,bi) we can take any be (bi,b 2) (') (c,b 1). If bi e (a,b 2) we can take any
b E(bi ,b2) (') (d,b2).
We show next that 1f Ci e abi and c2 e ab 2, then bic 2 (') b 2ci '¢ 0. If
a,b 1,b2 are not collinear this follows from (P)', and if a,b 1,b2 are not all distinct it
follows readily from (D). Suppose then that a,b 1,b2 are distinct but collinear. If
a e (bi,b 2), then a e (b 1,c2) (') (b 2,ci). Suppose b 2 e (a,bi). If ci e (bi,c 2) then
(b 2,ci) 5 (bi ,c2), and if c2 e [ci,b 1) then (b 2,c2) 5 (b 1,c2) (') (b 2,ci). The same
argument applies if b 1 e (a,b 2).
We will refer to these analogues of the axioms (C) and (P) as (C)" and (P)".
It is easily verified that the analogues of Propositions 11.2, 11.17 and 11.19 also
remain valid, i.e.
When required, these analogues will be referred to as Propositions 11.2", II.17" and
11.19".
If A and B are nonempty subsets of X, we define their product AB by
AB = U aeA,beB ab.
In other words, AB is the union of A n B with the set of all points c e X such that
c e (a,b) for some distinct points a e A and be B. The reason for the terminology
and notation is that products possess many of the properties of products of positive
numbers, as we now show.
First of all AB'#: 0, on account of (D), and it is obvious that, if A s C and
B s D, then AB s CD. Evidently also AB= BA. Furthermore the associative
law holds: for any nonempty sets A,B,C,
(AB)C = A(BC).
Proof The statement (i) is just a refonnulation of Proposition 5, and (ii) follows at
once from (i) and Proposition 4.
If Ci:;: 0, then (ii) implies that Cic s Ci for any c e C. On the other hand, if
x e Ci and c e C, then x e cc' for some c' e C. If we choose x' e ex, then x' e Ci
and x e ex'. Hence Ci s: Cic. This proves (iii). CJ
[A u B]i = (AB)i.
Proof We may assume that [A u B] is not a singleton, and then also AB is not a
singleton by Proposition 21.
We show first that [A u B]i s: (AB)i. If x e [A u B]i, y e AB and y :;: x,
then x e (y,z) for some z e [A u B]. If z' e (x,z), then z' e (y,z) and hence
z' e AB by Proposition 21. Since x e (y,z'), it follows that x e AB and actually
x E (AB)i.
We show next that (AB)i s: [Au B]i. Suppose x e (AB);, ye [Au B] and
y:;: x. If y' e (x,y), then y' e AB by Proposition 21. Hence x e (y',z) for some
z e AB. Since AB s: [A u B] and x e (y,z), it follows that x e [A u B]i. D
PROPOSITION 23 If A and Bare convex sets such that Ai:;: 0 and Bi:;: 0,
then
[Au B]i = (AB); = AiBi :;: 0.
Proof By Proposition 22 we need only show that (AB); = AiBi. We divide the
proof into two parts.
1. Dense linear geometries 95
We show first that (AB)i s AiBi. Let x e (AB)i and let ye AiBi with y -:1= x.
We show next that AiBi s (AB)i. Letze AiBi and let c e AB. Then z e xy
for some x e Ai and ye Bi, and c e ab for some a e A and b e B. Hence x e aa'
for some a' e A, and y e bb' for some b' e B. Thus
z e (aa')(bb') = (ab)(a'b').
But abc =ab, since c e ab and ab= (ab)i. Hence z e c(AB)(AB) = c(AB). Thus
z e (AB)i, by Proposition 20(i). D
PROPOSITION 24 If A and Bare convex sets such that Ai ":I= 0 and Bi ":I= 0,
then
[AuB] = AB. CJ
This axiom also implies the axiom (P), when all the other axioms of a linear
geometry are satisfied. Moreover, as we now show, an unending linear geometry
is 'generally' dense:
96 V. Density and unendingness
Proof By Proposition III.6 and Proposition l, we need only show that the axiom
(D) is satisfied if not all points of X are collinear.
Let y 1,y2 be any two distinct points and choose z e: <y 1,y2>. By (U) we can
now choose z1 so that z e (y2,z 1) and then x so that z1 e (x,y 1). It follows from
Proposition 11.2' that z e (x,y) for some ye (y 1,y2). D
An unending linear geometry need not be dense if all points are collinear; for
example, take X = lR \ (0,1) and define the segment [x,y] to be the intersection with
X of the corresponding segment in JR.
It may be shown also that a convex geometry is a linear geometry if it
satisfies the axioms (L2),(L3),(P) and (U), i.e. these axioms imply the axiom
(L4). For suppose c,d e [a,b]. We wish to show that de [a,c] u [c,b].
Evidently we may assume a '¢ b, c '¢ a,b and d '¢ a,b,c. By (U) we may choose e
so that be (a,e). Since c,d e [a,b], it follows from Proposition II.18(i) that
be [c,e] fl [d,e]. Since b '¢ e, it now follows from (L3) that c e [d,e] or
de [c,e]. If de [c,e] then, since be [d,e], it follows from Proposition II.18(i)
that de [b,c]. Thus we now suppose c e [d,e]. From be [d,e] and de [a,b] we
obtain, by (L2), de [a,e]. From de [a,e] and c e [d,e] we obtain de [a,c], by
Proposition II. l 8(i) again.
Throughout the remainder of this section we assume that a set Xis given
on which a dense, unending linear geometry is defined.
An immediate consequence of this assumption is that, if H is a hyperplane of
X, then X \ H is not convex. Hence Proposition 111.12 and Proposition 17 now
apply to arbitrary hyperplanes. Another consequence of this assum~tion is that the
definition of intrinsic interior may be reformulated in the following way:
Proof The necessity of the condition has already been proved in Proposition 9.
Suppose now that C fl (c ,x) '¢ 0 for every x e <C> \ c. For any d e C \ c, choose
x so that c e (d,x). If e e C fl (c,x), thence (d,e) and hence c e Ci. CJ
2. Unending linear geometries 97
Thus c e alb if and only if a e be. More generally, for any nonempty subsets A,B
of X we define the quotient AIB by
98 V. Density and unendingness
(i) A s B(AIB),
(ii) A s (AB)IB,
(iii) A s Bl(BIA).
Proof Since these properties are really just reformulations of the definitions, we
only prove (i): if a e A, be B and c e ab, thence AB and a e (AB)/B. D
For any nonempty set A s: X we have A s: A/A. It will now be shown that
affine sets are characterized by equality in this relation:
Proof If x e A/A, then there exist a,a' e A such that a e xa' and hence x e <A>.
100 V. Density and unendingness
On the other hand if x e <A> then, by Proposition III.7, x e <a,a'> for some
a,a' e A. Hence either x e A or a e (x,a~ or a' e (x,a). In every case x e A/A.
a
PROPOSITION 35 If A is a convex set and x e Ai, then <A> = Alx = x/A.
Proof It is obvious that Alx s <A> and x/A s <A>. Moreover x e Alx and
x e x/A. Suppose ye <A> and y :I: x. Then, by Proposition 9, there exists a point
z e A n (x,y). Hence y e zlx s Alx. Furthermore, since x e Ai, we have
xe (z,z~ for some z' e A. Hence x e (z',y) and ye xlz' s x/A. D
(i) x E Ai,
(ii) A s x/A,
(iii) x/A is affine,
(iv) Alx is affine.
Proof It follows from Proposition 35 that (i) :::) (ii),(iii),(iv). To complete the
proof we show that (ii),(iii),(iv) :::) (i). We may assume that IAI > 1. Let y e A \x
and choose z,w e X so that x e (y,z), ye (x,w). Then z e x/A and we Alx.
If A s x/A, then ye x/A. Thus x e (a,y) for some a e A, and hence x e Ai.
If x/A is affine then ye x/A, since x e x/A. Hence A s x/A and x e Ai.
If Alx is affine then z e Alx, since x e Alx. Thus there exists a point
a' e A n (x,z). Then x e (a',y), and hence x e Ai. D
Hence AIB is affine, by Proposition 33, and the result now follows from
Proposition 37. D
PROPOSITION 39 If A and B are convex sets such that Ai '¢ 0 and Bi '¢ 0,
then
(AIB)i =Ai/Bi-¢ 0.
Proof We show first that (AIB)i s Ai/Bi. Let x e (AIB)i and let y e Ai/Bi with
y '¢ x. Since Ai/Bi s AIB, we have x e (y,z) for some z e A/B. Hence, by
Lemma 31 and Proposition 20,
We show next that Ai/Bi s (AIB)i. If z e Ai/Bi, then z e xly for some
x e Ai, ye Bi. Hence x e Ai n (zBi) =Ai n (zB)i, by Proposition 23. It now
follows from Proposition 38 that
The next result may be regarded as an affine analogue of Proposition 11.4, but
it requires a supplementary hypothesis.
102 V. Density and unendingness
Proof By Proposition 1.22, we need only show that if Ai has finite codimension
min <Aiu A 2>, then A 1 n A 2 has finite codimension n in A 2, where n < m. Let
B be a basis for Ai n A 2 • Then Ai has a basis of the form B u Bi and
<Aiu A 2> has a basis of the form Bu Bi u B 2', where B 2' s A 2 and IB2 1= m.
By Proposition 41, with A,B,C replaced by Ai,A 2,<B2'>, we have
2. Unending linear geometries 103
3 CONES
convenient to denote by (z,x> the 'open' ray [z,x> \ z. Evidently if x' e (z,x>, then
[z,x'> = [z,x> and (z,x'> = (z,x>.
A set K s X is said to be a cone with vertex z if x e K implies [z,x> s K.
The vertex need not be uniquely determined; the set of all vertices of a cone will be
called its vertex set.
For example, the empty set is a cone with X as its vertex set, and the singleton
{z} is a cone with vertex z. An arbitrary union of rays from z is a cone with vertex
z. A nonempty set Lis affine if and only if it is a cone with vertex set L.
The definition of a cone does not require it to be convex, but the convex case is
of greater interest. Thus if L is a nonempty affine set and x e X \ L then, by
Proposition 111.12, the closed half-space of <x u L>, which is associated with the
hyperplane Land which contains x, is a convex cone with every point of Las a
vertex.
It follows at once from the definition that if z is a vertex of a nonempty cone K,
then z e K. Moreover, if z and z' are vertices of K, then <z,z'> s K. In fact, as
we now show, every point of <z,z'> is a vertex.
It follows that if K is a nonempty convex cone and L the affine set of all
vertices of K, then K\L is convex.
Let 'XL denote the collection of all convex cones whose vertex set contains L,
where L is a nonempty affine set. It is easily seen that 'XL is a normed alignment on
X. We define the L-conic hull [L,S> of any set S s X to be the intersection of all
convex cones in 'XL which contain S.
Evidently [L,S> e 'Xv Furthermore our notations are consistent, since for
L = {z} and S = {x} the L-conic hull of Sis the ray [z,x>. From the definition of
L-conic hull we immediately obtain the following properties:
K = LK = KIL = LKIL.
Proof It follows directly from the definition of a cone and (U) that, if K is a cone
whose vertex set contains L, then K = LK and K =KIL. These two relations
obviously imply K =LKIL. Suppose on the other hand that K = LKIL and L s K.
Letze Land x e K\z, so thatx e y'lz'for some y' e LK and z' e L. We wish to
~how that [z,x> s K.
PROPOSITION 48 For any nonempty convex set C and any nonempty affine
set L,
[L,C> = [L u C]/L = LCIL u L.
On the other hand, since L s: LKIL and LC!L s: LKIL, we also have K s: LKIL.
Consequently K = LKIL and K is a cone whose vertex set contains L, by
Proposition 47. Furthermore, if K' is any convex cone containing C whose vertex
set contains L, then
K = LCIL u L s LK'IL = K'.
3. Cones 107
Consequently K = [L,C>. D
We consider next the structure of the intrinsic interior and convex closure of a
convex cone. If some vertex of a convex cone K is contained in its intrinsic interior
Ki then, by Proposition 45, every point of K is a vertex. Hence K is affine and
K =Ki = K. We now treat the general case:
Proof Obviously we may assume that Ki~ 0. Then H: =Ki u L is convex, since
Ki and L are convex and L s K. To show that H is a cone whose vertex set
contains L, we need only show that if z e L, x e Ki\L and x e (y,z), then ye Ki.
We certainly have ye K. Lety' e K\y and choose y" e X so thaty e y'y". Then
x e y'y"z and hence x e y'z' for some z' e y"z. Since x e Ki, it follows that
x e y'x' for some x' e xz' n K. Since x e yz and z' e y"z, it follows that x' e x"z
for some x" e yy". Moreover x" e K, since x' e Kand z is a vertex of K. Since
y e y'x", it follows that y e Ki.
To prove that K is also a convex cone whose vertex set contains L, we need
only show that K contains the ray [z,y> for any ye K \Kand z e L. Choose any
y' e [z,y>, x e Ki and x' e xy'. Suppose first that y' e (z,y). Then, by (C) ",
x' e zz' for some z' e xy. Consequently z' e Ki and hence x' e Ki. This proves
that y' e K. Suppose next that ye (z,y'). Then, by (P)", there exists a point
z" e zx' n xy. Consequently z" e Ki and hence x' e Ki. by the first part of the
proof. Thus again y' e K. D
Proof We may assume K ~L. Letze Land put Kz = z/K. Thus Kz is the union
of {z} with the set of all points x e X such that z e (x ,y) for some y e K.
Evidently Kz is a cone with vertex z, and Kz ~ L since K ~ L. Moreover Kz is
convex, by Proposition 11.19", and Ls: Kz since Ls: K.
108 V. Density and unendingness
In fact L =Kn Kz. For if z' e Kn Kz and z' -:I: z, then z e (z',z") for some
z" e Kand hence z' e L, since Lis a face of K by Proposition 45. Since it is clear
from the definition that (~)z = K, it follows that the vertex set of~ is ~ n K = L.
It only remains to show that Kz = LI K. In fact, since it is obvious that
Kz s L/K, we need only show that if x' e L/K, then x' e Kz. There exist y' e K
and z' e L such that z' e xy'. If we choose x" so that z e xx", then there exists a
pointy e zy' n z'x" by (P)". It follows that ye K, since K is convex, and hence
thatx" e K, since z' e L. Consequently x' e Kz. D
Proof We show first that the vertex set of Ki u Lis exactly L. By Proposition 49
the vertex set of Ki u L contains L. Assume that Ki u L has a vertex w e Ki \L.
Then for some x e K the ray [w,x> is not wholly contained in K. However, the
line <x,w> contains points w',w" e Ki such that we (w',w"). Moreover w',w" are
vertices of Ki u L, since the set of all vertices is a face of Ki u L. Hence
<x, w> = <w', w"> s Ki u L, which is a contradiction.
If Ki -:1: 0 then, by Proposition 39, (Kt); = LI Ki -:1: 0. Since the relation
between Kand Kt is symmetric, this proves (i). Furthermore, since LIL= L,
4 POLYHEDRA
(*)
The representation (*) will be said to be redundant if some term Hi+ can be
omitted and irredundant otherwise. Evidently if the representation (*) is
redundant, then an irredundant representation can be obtained from it by omitting
certain terms.
110 V. Density and unendingness
Proof Put
Mk = nb~k Hi+.
Fk = nj~k Hi+ n Hk
sincey e nj:1 Hi+, we can choose z e nj:1 Hi+, by (D). Then z e P. Since F
is a face of P, it follows that y e F. Thus F =P, which is a contradiction.
This shows, in particular, that F 1, ••• ,Fm are the only facets of P. Since the
representation(*) is irredundant, they are certainly distinct.
It is obvious that nJ'= 1 Hi+ s Pi. Since no element of pi is contained in a
facet Fi, we must actually have pi = n j:1Hi+. Moreover Pi -¢ 0 since, by
Proposition 3, P is not contained in UJ'=1H j· If y e P and x e pi, then
(x,y) s Hi+ (j = 1,... ,m) and hence ye P. Thus P = P.
Let F be any nonempty proper face of P. As we have seen, F s F k for some
k. If F = Fk' then F is a polyhedron. If F '¢ Fk' then Fk is not affine. By what we
have already proved, the facets of Fk have the form
for some i '¢ k. Since F is a proper face of F k' it follows that F s F k n Fi for
some i '¢ k such that Fk n Fi is a facet of Fk. Moreover F = Fk n F; if and only if
F is a facet of F k and in this case F is a polyhedron. If F '¢ F k n F;, then Fk n F;
is not affine and the argument can be repeated. Since the process must eventually
terminate, this yields (ii) and (iii). Cl
It may be noted that, in Corollary 54, the length of the sequence is the
codimension of <F'> in <G> and consequently does not depend on the choice of
sequence.
Proof The proof of the corresponding Proposition IV.18 for polytopes remains
valid if we replace differences of dimension by codimension. a
PROPOSITION 57 If P is a polyhedron and A an affine set, then P n A is a
polyhedron.
Proof Since P r'\ A =P n A r'\ <.P r'\ A>, we may assume that A = <P n A>.
Evidently we may assume also that A'¢ <P>, that P n A'¢ 0,A and that Pis not
affine. Let P have the irredundant representation (*). Then A is not contained in
Hi+ for at least one j, and we may choose the notation so that A s Hi+ if and
only if j > n. Then
Pr'\ A = nj= 1 Hi+ r'\ A
and, by Proposition 43, Hi+ r'\ A is a closed half-space of A associated with the
hyperplane Hin A for each j e { l, ... ,n}. Since A= <P n A>, it follows that
P r'\ A is a polyhedron. D
S NOTES
The concepts of intrinsic interior and convex closure, for convex sets in a real
vector space, are discussed (among other things) by Bair and Fourneau (1975/80).
However, the definition of convex closure used here is slightly different from theirs
and has the advantage that in Proposition 7, for example, it need not be assumed
that Ci -:1: 0.
The properties of products and quotients in 'join spaces' are extensively treated
by Prenowitz (1961) and Prenowitz and Jantosciak (1979). They not only replace
geometry by algebra, but also obviate special consideration of degenerate cases.
Proposition 26 is already contained in Veblen (1904). Proposition 41 plays a
role in the lattice-theoretic characterization of affine geometry, which was initiated
by Menger (1936); see, for example, J6nsson (1959), where 'weakly modular' is
called 'special', and Maeda and Maeda (1970). Proposition 42, for the finite-
dimensional case, is already contained in Grassmann (1844), §126. Cones are
discussed for more general situations than Euclidean space by Prenowitz and
Jantosciak (1979) and Rubinstein (1964).
The idea of basing the theory of polyhedra on Proposition 3 is taken from
Griinbaum (1967). The usual analytic treatment replaces the theory of polyhedra by
the theory of linear inequalities; see, for example, Stoer and Witzgall (1970),
Tschemikow (1971) and Schrijver (1989). It should be mentioned, however, that
two significant results of this theory do not carry over to the more general situation
considered here. This is illustrated by the following simple example.
Let X be the open unit disc of R2, with the usual definition of segments, and
let P be the set of all points of X in the (closed) first quadrant. Then P is a
polyhedron with two facets. However, P is not the convex hull of its facets,
although it is neither affine nor a closed half-space. Furthermore, P is not the
convex hull of finitely many points and rays, although it is a polyhedral cone with
the origin as vertex.
VI
Desargues
There are many examples of the dense unending linear geometries studied in the
previous chapter, besides vector spaces over an ordered division ring. We give
several such examples here, including the projective plane over the field of rational
numbers. This example is not atypical, since in the next chapter we will show that
any dense linear geometry of dimension greater than 2 is isomorphic to a subset of a
projective space over an ordered division ring. In the present chapter we lay the
groundwork for this result by giving a brief introduction to projective geometry and
by showing that any dense linear geometry of dimension greater than 2 has the
Desargues property. Many dense two-dimensional linear geometries do not have
the Desargues property, but we establish several properties of those which do. We
show also that a factor geometry of any dense linear geometry is a dense linear
geometry with the Desargues property.
1 INTRODUCTION
The results which have already been established show that dense, unending
linear geometries closely resemble vector spaces over an ordered division ring,
which are indeed a special case. Nevertheless, there are several other examples:
EXAMPLE 1 Let X be any dense totally ordered set which has no least or greatest
element. For any x,y e X define the segment [x,y] to be the set of all z e X such
that x < z Sy or y < z S x, according as x Sy or y < x.
For example, we may take X to be the set of all ordered pairs (~ 1 ,~ 2) of real
numbers with the total ordering defined by (~ 1 ,~ 2) -< (11 1,11 2) if ~ 1 < 11 1, or if
~1 = 111 and ~2<112·
116 VI. Desargues
, ,•
,,
,,
, ,,' y
,,
0
(i) for each subset !J e <§, there is a continuous injective map t ~ g(t) of R onto
e R}, and d(g(t),g(O)) ~ oo as t ~ +oo;
9, so that !J = {g(t): t
(ii) for any two distinct points a,b e R2, there is exactly one subset !Jab e <§which
contains both a and b.
The hypothesis (i) says that each !J e <§ is homeomorphic to R and is a closed
subset of R2. By adjoining to R2 a point at infinity and applying the Jordan curve
theorem (see, for example, Wall (1972)) to the resulting sphere s2, we see that
R2 \ !J has exactly two connected components and !J is their common boundary. We
will call these two connected components the 'sides' of 9.
If a= gab(a) and b = gab(p), we define the segment [a,b] to be the set of all
points c e !Jab with c =gab(t) for some t such that a St SP if a< p, or such that
p St Sa if p <a. We also put [a,a] ={a} for any a e R2. It may be
immediately verified that, with segments so defined, the axioms (Ll)-(L4) are all
satisfied. Thus we are free to use the results of Chapter 111, Section l, and !Jab is
what is there denoted by <a,b>. This notation will now also be used here.
We now show that the axiom (C) is also satisfied. Without loss of generality,
suppose c e (a,b 1) and de (c,b 2). By Proposition 111.6 we may assume that
b2 E <a,b 1>. Then b1 and b 2 are on opposite sides of <a,d>, since c and b2 are on
opposite sides of <a,d> but c and b 1 are on the same side. Thus there exists a point
b E (b1,b~ ~ <a,d>.
118 VI. Desargues
Assume first that be (a,d). Then c and dare on opposite sides of <b1,b2>,
since a and d are on opposite sides of <b 1,b2> but a and c are on the same side.
Since de (c,bi), this is a contradiction.
Assume next that a e (d,b). Then band b 1 are on opposite sides of <c,d>,
since a and b 1 are on opposite sides of <c,d> but a and b are on the same side.
Thus there exists a point b' e (b 1,b) n <c,d>. Since b' e (b 1,b2) and b2 e <c,d>,
this is a contradiction.
We conclude that de (a,b). Thus R2 is a convex geometry with the present
definition of segments. Since the axiom (U) is evidently satisfied, it follows from
Proposition V.26 that the axioms (P) and (D) are also satisfied.
As a specific example, consider the family CS consisting of all subsets
which implies
Since ~ 1 '¢11 1 , this is a cubic equation for a with at least one real root. In fact it has
a unique real root, since the polynomial in a on the right side has a nonvanishing
derivative. Having determined a, we immediately obtain (3 = ~ 2 - (~ 1 -a)4 =
112 - (111 - a) 4 •
As another example, let 9 = {x = (~ 1 ,~ 2 ) e R2: ~ 2 > 0, ~ 1 = - ~ 2- 1 } be the
branch of the hyperbola ~ 1 ~ 2 = - 1 in the upper half-plane. Given any positive
numbers µ,p, there exist uniquely determined points x = (~ 1 ,~ 2), y = (11 1,112) on 9,
1. Introduction 119
with ~ 1 <11 1 , such that the straight line segment joining x and y has slopeµ and
length p. In fact, if we put a= p/2(1 + µ2)112, then
where a,p e R, be an arbitrary translate of II· Then it follows that, given any two
distinct pointsx =(~ 1 ,~ 2 ) andy =(11 1,11 2) ofR2 such that the straight line segment
joining x and y has slope µ > 0 and length p > 0, there is exactly one translate !lap
which contains both x and y; in fact we must take
where y,B,µ e R and µ < 0, satisfies the conditions (i) and (ii).
2 PROJECTIVE GEOMETRY
(i) any two distinct 'points' a,b are contained in exactly one 'line', denoted here
by <a,b>,
(ii) every 'line' contains at least three distinct 'points',
(iii) if <a,b> and <c,d> are distinct 'lines' with a common 'point' e '* a,b,c,d, then
<a,d> and <b,c> are also distinct 'lines' with a common 'point' f '* a,b,c,d (see
Figure 2).
120 VI. Desargues
Two 'lines' will be said to intersect if they have a common 'point' and 'points'
will be said to be collinear if they are all contained in a common 'line'. (Inverted
commas are used for several concepts in projective space to distinguish them from
the corresponding concepts in a vector space.) A projective plane is a projective
space in which any two 'lines' intersect.
For example, with any vector space V over a division ring D there is associated
a projective space P(V) whose 'points' are the one-dimensional vector subspaces of
V and whose 'lines' are the two-dimensional vector subspaces of V. If Wis a
vector subspace of V, then P(W) is said to be a 'subspace' of P(V), and if Wis a
maximal proper vector subspace of V, then P(W) is said to be a 'hyperplane' of
P(V).
Figure 2
We can regard Vas a subset of the projective subspace P(V*) by identifying the
vector v e V with the one-dimensional subspace of V* containing the couple (l,v).
The remaining 'points' of P(V*), i.e. the one-dimensional subspaces of V*
containing a couple (O,v) for some v e V\ {0}, form a 'hyperplane' of P(V*) - the
2. Projective geometry 121
If <a 1,bi>,<a 2 ,b2>,<a 3 ,b3> are three distinct 'lines' of P(V) with a
common 'point' p :I! a;,b; (i = 1,2,3), then the 'points' of intersection
are collinear.
We sketch the algebraic proof. Suppose the 'points' p,a;,b; of P(V) are
represented by the vectors x,cx;,P; (i = 1,2,3) of V. Since each is determined only
up to a non-zero scalar multiple, we may assume that
Then
cx.2 - cx.3 = P2 - p3, cx.3 - cx.1 = P3 - P1,
. cx.1 - a2 = P1 - P2·
Hence the 'points' c 1,c2,c 3 are represented by the vectors cx. 2 - cx 3,cx 3 - cx. 1,
cx. 1 - a 2. Since the three vectors have zero sum, they are contained in a two-
dimensional subspace of V and the three 'points' are collinear.
122 VI. Desargues
Desargues' theorem implies its own converse (cf. the proof of Proposition 3
below):
If a 1,a2 ,a3 and b 1,b2,b3 are non-collinear triples of 'points' in P(V) such
that
are distinct collinear 'points', then the 'lines' <a 1,b1>,<a2,b2>,<a 3,b3> have a
common 'point' p.
Suppose now that P(V) is the projective space associated with a vector space V
over an ordered division ring D. If a,b are distinct 'points' of P(V), represented by
the vectors a,p of V, then the 'line' <a,b> is the disjoint union of the 'points' a,b
and two open segments, represented by the vectors A.a+ µp, where A.,µ e D and
A.µ> 0 for one segment, A.µ< 0 for the other. The corresponding closed segments
are obtained by adjoining the two endpoints a,b.
It is readily seen that if <a;,b;> (i = 1,2,3,4) are four distinct 'lines' with a
common 'point' p '¢ a;,b; (i = 1,2,3,4), and if a3,a4 e <a 1,a2>, b3,b4 e <b 1,b2>,
then b3 ,b4 belong to the same open segment with endpoints b 1,b2 of the 'line'
<b1,b2> if a3,a4 belong to the same open segment with endpoints a 1,a2 of the 'line'
<a1,a2>.
EXAMPLE 6 The projective plane X over the field Q of rational numbers may be
given the structure of a dense, unending linear geometry in the following way.
Regard X as a subset of the real projective plane X' and choose a 'line' A.' in X'
which contains no point of X. (For example, if a 1,a2 e Rare such that l,a 1,a2
are linearly independent over Q, then a 1a 2 '¢: 0 and we can take A.' to be the 'line'
determined by the 'points' (l,a 1-1,0) and (l,O,a2-1).) Any two points x,y e X
with x '¢ y determine a unique 'line' l's X', and l' intersects A.' in a unique point
z' e: X. The points of I.' other than x and y form two disjoint open segments. We
define the segment [x,y] of X to be the intersection with X of the open segment of l'
which does not contain z', together with x and y.
Once again, the only axiom which needs to be verified is (C). One of the
configurations which can arise in the course of this verification is illustrated in
Figure 3.
2. Projective geometry 123
''
''
''
''
''
', A.'
''
''
''
''
b....... ....... .......
' ' ',
....... ....... '
....... ....... ....... ' ',
/
- - - - -- - - .......- ~
~-a--==:
''
/
Figure 3: Veblen's projective plane
EXAMPLE 7 For any ordered division ring D and any positive integer n, the
n-dimensional projective space pn = pn(D) can be given the structure of a dense,
but not unending, linear geometry in the following way. We regard pn as the set of
(n + 1)-tuples of elements of D, not all zero, with proportional (n + 1)-tuples
identified, and Pi (0 < j < n) as the set of all (n + 1)-tuples (~ 0 ,~ 1 , .•• ,~n) with
~o = ~ 1 = ... = ~n-j-l = 0. In particular, po contains the unique 'point'
Po = (O, ... ,O,~n>·
For any x,y e pn with x :F- y put [x,x] = {x) and define [x,y] inductively in the
following way. If x,y e pl and x,y ':I: p 0 , take [x,y] to be the segment with
124 VI. Desargues
endpoints x,y of the 'line' <t,y> which does not contain p0; if x,y e pl and x =Po
or y =Po take [x,y] to be the segment with endpoints x,y of the 'line' <t,y> which
contains all (n + 1)-tuples A.x + µy with A.µ > 0. Now assume that [x,y] has been
defined for x,y e pk-1, where k > 1. If x,y e pk and x,y E pk-1, take [x,y] to be
the segment with endpoints x,y of the 'line' <t,y> which does not contain a point of
pk-1; if x,y e pk and exactly one of x,y is in pk-1, take [x,y] to be the segment with
endpoints x,y of the 'line' <t,y> which contains all (n + 1)-tuples A.x + µy with
A.µ> 0.
With these definitions X = pn is a linear geometry. The verification of the
axioms (L2)-(L4) reduces to their verification for n = l, and presents no
difficulty. To verify now the axiom (C) we may assume that a,b 1,b2 are not
collinear. If for any de [c,b2] we put d' = <a,d> n <b 1,b2>, then the map d ~ d'
is a bijection of [c,b2] onto a segment with endpoints b 1,b2 of the 'line' <b 1,b2>.
By considering separately the various possibilities, it may be seen that in all cases
this segment is [b 1,b2]. Since Xis obviously dense, the axiom (P) follows from
Proposition V.1. However, the dense linear geometry Xis not unending, since if
b :1: Po there is no c :1: Po such that Poe [b,c].
A linear geometry will be said to have the Desargues property if the following
condition (illustrated in Figure 4) is satisfied:
(A) Let <a 1,b 1>, <a 2 ,b2>, <a 3,b3 > be three distinct lines with a common point
p :1: a;.b; (i = 1,2,3). If the lines <aj,ak> and <bj,bk> intersect for (j,k) = (1,2),
(1,3),(2,3), then the points of intersection c 3 ,c 2 ,c 1 (which are necessarily
uniquely determined and distinct) are collinear.
the lines <a 1,a 3 > and <b 1,b 3 > intersect in the point c 2 = (160/29,36/29).
However, c2 does not lie on the line <c 1,c3>. (Only one of the lines mentioned,
namely <b 1,b3>, is a 'refracted' line. Since the ordinary affine plane has the
Desargues property, the refracted plane cannot.)
Similarly we can choose y e pi with x e (p,y) so that <a3,.x> and <b3 ,y'> intersect
within pi for any y' e (x,y].
Choose such a y e (p,x) if b 3 e (p,a 3}, and such a y with x e (p,y) if
a 3 e (p,b 3 ) or p e (a 3 ,b3). If z e <a3,x> n <b3 ,y>, then z e pi and in every case
x e (a 3 ,z). Since z e <a 3 ,a 1> = <c 2 ,a 1>, we can in the same way choose
a 3 ' e (x,z) so that <a 1 ,a 3 '> and <c 2 ,z> intersect in Pi. Similarly, since
z e <a 3,a 2> = <c 1,a2>, we can choose a 3 " e (x,z) so that <a2,a3 "> and <c 1,z>
intersect in Pi. Moreover, by choosing the one which is nearer to z, we may
suppose a 3 " = a 3 '. In the same way we can choose b 3 ' e (y,z} so that <b 1,b3 '>
intersects <c2,z> in pi and <b2,b3 '> intersects <c 1,z> in Pi. Furthermore, we may
now modify our choice of a 3 ' and b3 ' to ensure that p e <a3',b3 '>.
The line <c2,z> is not contained in the plane A'= <JJ,a 1,a3 '>. For otherwise
we would have x,a 3,b3 ,a 1,b 1 e A' and hence, since A is a plane, A =A', which
contradicts x e A. Since the line <c2,z> intersects both <a 1,a3 '> and <b 1,b3 '>, it
follows that the two points of intersection must be the same. Thus there exists a
point c 2 ' e <a 1 ,a 3 '> n <b 1 ,b 3 '>, and similarly there exists a point c 1 ' e
<a 2 ,a 3 '> n <b 2 ,b 3 '>. The points p,a;,bi,c;' (i = 1,2) and a 3 ',b3 ',c3 satisfy the
hypotheses of (A) and the pointsp,a 1,a 2 ,a 3 ' are not coplanar, since a2 e A'.
Hence, by the first part of the proof, c 1',c2 ',c3 are collinear. Thus c 3 e <c 1,c2,z>.
Since ze A and <c1tc 2 ,c3 > s A, it now follows from Proposition 111.8 that
C3 E <C1 ,C2>. 0
Although the Desargues property does not hold in all two-dimensional dense
linear geometries, Proposition 1 shows that it must hold in such a geometry if it can
be embedded in a dense linear geometry of higher dimension. Consequently we
will assume throughout the remainder of this section that X is a dense linear
geometry of dimension> 2 with the Desargues property.
We show first that a modified form of the Desargues property also holds:
PROPOSITION 2 Let <a 1,b 1>, <a 2,b2>, <a 3 ,b3 > be three distinct lines with a
*
common point p ai,b; (i = 1,2,3). If
3. The Desargues property 127
Proof The hypotheses imply that c1 -:1: a2,a3,b2,b3 .and c2 -:1: a 1,a3 ,b 1,b3• Moreover
c 1 -:1: c2 • Hence the lines <a 1,c2 >, <a 2 ,c 1>, <p,b 3> are distinct and have the
common point a3• It now follows from (A) that c3 e <b1,b2>. a
exist and are distinct collinear points. If p e <a 1,b 1> r. <a 2 ,b2>:, then also
p E <a3,b3>.
Proof Since the points c 1,c2 ,c3 are distinct, we have a; -:1: b; (i = 1,2,3). Since
a 1,a2,a3 and b 1,b2 ,b3 are not collinear, we also have c; -:I: a;,b; (i = 1,2,3). We
may further assume that c 1,c 2 -:I: a 1,a 2 ,b 1 ,b 2 • For suppose c 2 = a 1 • Then
a 2 e <c2,c3> = <c 1,c2>. Since a1,a2,a3 do not all lie on the line <c 1,c2>, it follows
that c 1 = a2. Thus a 1 e <b 1,b3>, a2 e <b2,b3>, and hence p =b3.
With this assumption the lines <a 1,a2>, <b 1,b2>, <c2,c 1> are distinct, and by
hypothesis they have the common point c 3 • We may assume in addition that
c3 ':/! a 1,a 2,b 1,b2 . For suppose c3 =a 1• Then a3 e <c2 ,c 3> = <c 1,c3> and hence
c 1 = a3, since a 1,a2,a3 are not collinear. Moreover <a 1,b 1> = <b 1,b2> and hence
p = b2• Since c 1 e <b2,b3> and c 1 = a 3 -:I: b3 , we obtain p e <a3 ,b3>.
With this additional assumption we have
PROPOSITION 4 Let a 1,a 2 ,a 3,a4 and a 1',a 2 ',a3',a 4 ' be two quadruples of
points, such that no three points of the same quadruple are collinear and none
of the points lies on a given line l. If <aj,ak> and <a/,ak'> intersect l in the
same point ajk,for (j,k) = (l,2},(l,3},(l,4},(2,3},(3,4}, and if <a 2 ,a4 > intersects
l in a 24 , then <a 2 ',a4 '> also intersects l in a24.
Figure 5
We consider first the case in which the two quadruples have a corresponding
point in common. Since ai =a/, ak = ak' for some j,k e {1,2,3,4} with j -:1: k
implies a;= a;' for all i e {1,2,3,4 }, we need only consider the case in which
a;=a;' for exactly one ie {l,2,3,4}. If a 1 =a 1 ', then the lines <a 2 ,a 2 '>,
<a4 ,a4 '>, <a 3 ,a 3 '> have the common point a 1 and hence, by Proposition 2,
a 24 e <a 2 ',a4 '>. If a 3 = a 3 ', then the lines <a4 ,a4 '>,<a 2,a2 '>,<a 1,a 1'> have the
common point a 3 and hence, in the same way, a 24 e <a2 ',a 4 '>. If a 2 =a 2 ' we
obtain a 24 e <a 2 ',a 4 '> by applying Proposition 3 to the triples a 3 ,a 1,a4 and
3. The Desargues property 129
<c3,a3> intersects (ai,a 12 ) \bi in a point p such that b 2 E [bi,p], and the line
<c3,a23 > intersects (ai,a 12 ) \bi in a point c 2 such that b 2 E [bi,c 2]. Evidently
bi,c2,c3 are not collinear. If ai 4 e (ai 3 ,a34) take c4 = (bi,ai 4) n (c3 ,a34), and if
ai 3 e (ai 4 ,a34) take c4 e (bi,ai 4) so that c3 e (c4 ,a34). In either case bi,c3,c4 are
not collinear and c4 = (bi,ai 4) n <c3,a34>. Indeed if we put ci =bi, then no three
points of the quadruple Ci ,c2,c3,c4 are collinear and ajk e <cj,ck> for (j,k) =
(1,2),(l,3),(1,4),(2,3),(3,4). Furthermore c; ¢ b;, not only for i = 2,3 but also for
i = 4, since
a24 e <b2',b/>. On account of the arbitrary nature of b 1',b3 ' we may suppose that
<b/,bk'> :I: <a/,ak'> for (j,k) = (l,2},(1,4),(2,3},(3,4), but in addition
<b 1',b 3 '> = <b 1,b 3> :I: <a 1 ',a 3 '>. The quadruple b 1 ',b 2 ',b 3 ',b 4 ' has b 3 ' e
(b 1 ',a 13 ). If a 3' e (a 1 ',a 13 ), we are finished. To complete the proof we now
construct another quadruple b 1",b 2",b 3",b4 " with the same properties but with
b1" E (b3 ",a13).
We have b 3' e (b 1,a 13 ) and b 1 e (a 1,a 12). Choose any b 1" e (b 3',a 13). We
are going to choose ye (b 1",b3 ') so that, for each x e (b 1",y), the line <x,a 34>
intersects the line <b 1",a 14>. If a 13 e (a 14 ,a 34 ), choose any ye (b 1",b 3'). If
a 14 e (a 13 ,a 34 ) we choose y in the following way. If a 13 e (a 12 ,a 34 ), then
b 1" e (u,a 14 ) for some u e (b 1,a 12) and we take y' = (b 1,a 13 ) r'\ (u,a 34 ). If
a13 e (a 12 ,a34), then b 1" e (v,a 14) for some v e (a 1,a 13) and we take y' e (v,a34)
so that b 1" e (y',a 13 ). Then in either of the last two cases we choose any
Y E (b1 ",b3 ') ('\ (b{,y').
If a 12 e (a 13 ,a 23 ), take z e (a 1,a 13) so that b 1" e (z,a 12 ) and then choose
z" e (b 1",z) so that y" = (z",a 23 ) r'\ (b 1",b 1) satisfies y" e (b 1",y). Choose any
b 2" e (b 1",z") and then take b 3" = (b 1",y'') r'\ (b2",a23 ).
If a 13 e (a 12 ,a 23 }, take z e (y,a 23 ) so that b 1" e (z,a 12 ), and if
a23 e (a12,a13}, take z = (b 1 ",a 12 ) r'\ (y,a 23 ). In either case choose any
b 2" e (b 1",z) and then take b3" e (b{,y) so that b2" e (b3",a23).
In every case we now take b4 " = <b3",a34> r'\ <b1",a 14>. By construction no
three points of the quadruple b 1",b 2",b3",b 4 " are collinear, b 1" e (b3 ",a 13) and
ajk e <b/',bk"> for (j,k) = (l,2),(1,3),(1,4),(2,3),(3,4). In the same way as before
it follows that a24 e <b2",b4 "> and we may choose b 1",b2 " so that <b/',bk"> :I:
<a/,ak'> not only for (j,k) = (1,3), but also for (j,k) = (l,2},(l,4),(2,3),(3,4). D
4 FACTOR GEOMETRIES
In this section we study the way in which the Desargues property, and also the
properties of density and unendingness considered in Chapter V, are inherited by
the factor geometries of a linear geometry which were defined in Chapter IV.
132 VI. Desargues
spaces. Hence band b' lie in different open half-spaces and there exists a point
x e (b,b1 n <au M>. Since x e a 0 nab, but x E M, this is a contradiction.
Suppose now that a 0 <ab and ab< a 0 • Then there exist points a' e a 0 \M
and b' e ab\ M such that a' e [d,b] and b' e [d,a]. Hence there exists a point
ye [a,a1 n [b,b1. Since ye a 0 nab and y E M, it follows that a 0 =ab.
Suppose next that a 0 S ab and ab < ac. Then there exists a point b' e ab\ M
such that b' e [d,c]. Also, since ab.= ab, there exists a point a' e a 0 \M such that
a' e [d,b1. Then a' e [d,c] and hence aa < ac.
It remains to show that, for any a,b e H +' either a 0 <ab or ab < a 0 •
Obviously we may assume b E a 0 and a E ab. Let C',C" be the closed half-spaces
of X associated with the hyperplanes H' = <a u M>, H" = <b u M> which
contain d. We intend to show that either a e C" orb e C'.
Assume on the contrary that a E C" and b E C'. Then there exist points
u e (d,a) n H" and v e (d,b) n H'. Hence d,a,b are not collinear and there exists
a point w e (a, v) n (b,u). Then w e H' n H" = M and there exists a point
t e (a,b) such that we (d,t). Since t e <d,w> s H and (a,b) s H +' this is a
contradiction.
From c" e H\ M it follows that c" E H',H". Let B ',B" be the closed half-
spaces of X associated with the hyperplanes H',H" which contain c". Then in the
same way we can show that either a e B" orb e B'. Since b E a 0 we cannot have
b e H' = C' n B', and since a E ab we cannot have a e H" = C" n B". Hence
a e C" implies b E C', and be C' implies a E C". But if b E C', then (d,b) n H' ~
0 and hence a 0 <ab. Similarly if a E C", then ab< a 0 •
We now proceed with the proof that H+: L has the Desargues property. Let
<p1,P1'>, <p2,P2'>, <p3,p3'> be three distinct 'lines' in H+: L with a common
'point' p ¢. p;,p;' (i = 1,2,3). Suppose also that the intersections
exist. Choose a; e p; \L, a;' e p;'\L (i = 1,2,3). We begin by showing that there
exists h' e H + such that the hyperplane H' = <h' u M> of X intersects (d,a;],
(d,a;1 (i = 1,2,3).
Let a;,a;' be the closed half-space of <a; u M>, resp. <a;' u M>, associated
with the hyperplane M which contains a;, resp. a;'. By what we have just proved
we may choose the notation so that a 1 ~ a;,a;' (i = 1,2,3). Thus if we take h' = a 1
and H' = <h' u M>, then [d,a;] n a 1 ¢. 0, [d,a;1 n a 1 ¢. 0 (i = 1,2,3). Since
4. Factor geometries 135
d E H', it follows that H' contains points of (d,a;], (d,a;1 (i = 1,2,3). By changing
notation we may now assume that a;,a;' e H' (i = 1,2,3).
Suppose first that L = {d}. Then any element of H + : L intersects H' in at
most one point. Since p e <p;,p;'> and p :I: p;,p;' (i = 1,2,3), it follows from
Lemma N.24 and the definition of a line that there exists a point P; e p \L such that
P; e <a;,a;'> and p; :I: a;,a;'. Moreover, by the preceding remark, P; = p is
independent of i. Furthermore the lines <a;,a;'> are distinct, since the 'lines'
<p;,p;'> are distinct. In the same way 'Yi contains both a point hi' e <a2 ,a3> and
a point hi" e <a2',a3 '>,and we must have hi'= hi". Thus the intersections
exist. Hence, by the Desargues property in X, hi ',h 2 ',h 3 ' are collinear.
Consequently y1,y2,y3 are collinear in H + : L.
Suppose next that L :I: { d} and put i = <D \ d>. Any a e H + is contained in a
unique p e H+: Land a unique p e H +: i. It is easily seen that if also a e H',
then p = p n H'. Hence if we put P; = p; n H', p;' = p;' n H' (i = 1,2,3), then
the lines <pi, Pi'>,<p2 , p 2 '>,<p3 , p3 '> in H + : i have a common point :I: f);, P;'
(i = 1,2,3). Furthermore the intersections
- -- -,-,,
"f3 =<pi, P2,> n <P1, P2 >
5 NOTES
Among the many texts on projective geometry, two which are particularly
concerned with its foundations are Heyting (1980) and the classic Veblen and
Young (1910/18). A basic theorem says that if Pis a projective space containing a
pair of disjoint 'lines', then there is a vector space Vover a division ring D such that
P = P(V). Moreover the same conclusion holds if no two 'lines' of P are disjoint
(i.e. if P is a projective plane), provided Desargues' theorem holds in P. In the
136 VI. Desargues
next chapter we will prove the analogues of these theorems for linear geometries, in
which case the division ring D is ordered.
Desargues' theorem dates from 1636. Smith (1959) gives a short account of
the life of Desargues, together with an English translation of the statement of his
theorem in the Oeuvres de Desargues. Desargues' theorem provides an interesting
combinatorial configuration of ten points and ten lines, each line containing three of
the points and each point contained in three of the lines. The ubiquitous Petersen
graph is obtained by taking the ten points as vertices, with two points joined by an
edge if they are not both contained in one of the ten lines.
Segre (1956) has given an 'algebraic' example of a projective plane which
does not have the Desargues property. Its 'points' are the 'points' (~0 ,~ 1 ,~ 2) of the
real projective plane, but its 'lines' are the two-parameter family of cubic curves
where e is a sufficiently small positive constant and p 0,p 1,p2 are homogeneous
parameters specifying the curve.
Many dense unending two-dimensional linear geometries which do not have
the Desargues property are provided by Example 5, in particular the two specific
cases discussed there. (Even Moulton planes with different refractive indices µ are
non-isomorphic.) Rather than attempt to enumerate all such geometries, we will
show in Chapter VID that they can still possess nice 'Euclidean' properties.
The proof of Proposition 4 is based on Sperner (1938), but some extra
argument is needed because we do not require that the linear geometry be unending.
VII
Vector Spaces
Here the results of the previous chapter are used to introduce coordinates, using a
splendid idea of von Staudt, as modified by Hilbert and then generalized by
Sperner. We define isomorphism of linear geometries and establish a result of
Doignon, that any dense linear geometry of dimension > 2 with the Desargues
property can be isomorphically embedded in a projective space over an ordered
division ring. In general such a geometry cannot be isomorphically embedded in a
vector space over an ordered division ring, but an open half-space of such a
geometry can be embedded in this way.
We also study properties involving vector sums of convex sets in a vector
space over an ordered division ring. In particular, we establish a useful algebraic
version of RAdstrom's cancellation law, which will be encountered in Chapter IX.
1 COORDINATES
w"
Figure 1: Addition
z w
z a C Q r=S w
To define the product of two points a,b e (z,w), we fix a point e e (z,w).
There exists a point z' e (z,w1 such that the lines <e,z'> and <a,z'> both intersect
the segment (w,w"). Indeed we can take any z' e (z",w'), where
z" = (e,w") rt (z,w1 if a e (z,e] and z" = (a,w") rt (z,w1 if a e [e,w). Put
140 VII. Vector spaces
z e P a w
Figure 4: Multiplication
It will now be shown that the points of the segment (z,w) form a group under
multiplication. Since multiplication is associative and e acts as an identity element,
we need only show that for each point a e (z,w) there exists a point be (z,w) such
that a·b = e. But if e' = (e,w0 ) n (z,w') and b = <e',we> n (z,w), then a·b = e.
We prove next the distributive laws: (a+ b)·c =a·c + b·c and c·(a + b) =
c·a + c·b. Again put p =a + _b and choose z' e (z,w') so that the lines <e,z'>,
<a,z'>,<b,z'>,<p,z'> all intersect the segment (w,w"), in points we,wa,wb,wp say.
If q' = (c,we) n (z,w') then, by the definition of multiplication,
a·c = <q',w0 > n (z,w), b·c = <q',wb> n (z,w), p·c = <q',wp> n (z,w).
It is easily verified that these definitions do not depend on the choice of elements
within an equivalence class, i.e., if (a 1',a21 = (a1'a2) and (b 1',b2 1 = (b 1,b2), then
Furthermore, the commutative and associative laws for addition, the associative law
for multiplication and the two distributive laws hold in D, since they hold in P.
The elements 0 =(a,a) and 1 =(e+a,a) act as identity elements for addition
and multiplication respectively. Moreover any element (a 1,a2) of D has an additive
inverse (a 2,a 1). We will show that any element (a 1,a2) :I= 0 has a multiplicative
inverse. Since a 1 :I= a2, either a 1 e (z,a 2) or a2 e (z,a1). If a 1 e (z,a2), there exists
x e P such that a 1 + x =a2 and hence (a 1,a2)·(a,x-1+a) = 1. If a 2 e (z,a 1), there
exists ye P such that a2 + y = a 1 and hence (a 1,a2)-(y-1+a,a) = 1.
Thus we have now shown that D is a division ring. Since the map
x--+ (x+a,a) of PintoD preserves sums and products, we may identify the point
x e P with the element (x+a,a) of D. In this way Pis embedded in D. Since Pis
closed under addition and multiplication, and Dis the disjoint union of the sets P,
{0} and -P, it follows that D is an ordered division ring with P as the subset of
positive elements.
We now show that if a and bare distinct points of P, then the 'open segment'
(a,b) is the set of all points ce P which can be represented in the form
c = A.a + (1 - A.)b, where 0< A. < 1. Without loss of generality assume a < b, i.e.,
a e (z,b), and suppose first that c admits such a representation. Since A.a< Ab we
have c < A.b + (1 - A.)b = b, and since (1 - A.)a < (1 - A.)b we have
a= A.a+ (1 -A.)a < c. Thus c e (z,b) and a e (z,c), which implies c e (a,b).
Suppose on the other hand that c e (a,b). Since a e (z,c) we have c =a+ a,
where a e P, and since c e (z,b) we have b = c + "(, where "( e P. Thus
b =a + p, where p =a + "( and hence "( < p. If we put A. =yp-1, then 0 < A. < 1
and
c = a + a = a + (1 - A.)p = A.a + (1 - A.)b.
2. Isomorphisms 143
2 ISOMORPHISMS
Proof Suppose first that b 1 = a 1 and b 2 ~ <a 1,a 2>. Choose some c e (a 2 ,b2).
For any a e (a 1,a 2), there exist a unique a' e (b 2 ,a) ri (b 1,c) and a unique
be (b1,b2) such that a' e (a 2,b). Moreover every be (b 1,b2) may be obtained in
this way. If we put /(a) = b,f(a 1 ) = b 1 ,f(a 2 ) = b 2 , then it follows from
Propositions 11.17' and 11.2' that/ is an isomorphism of [a 1,a2] onto [b1,b2].
Suppose next that <a 1,a2> = <h 1,b2> and choose c2 ~ <a 1,a2>. By what we
have already proved, [a 1,a2] is isomorphic to [a 1,c2], [a 1,c2] is isomorphic to
[b 1,c2 ], and [b 1,c2 ] is isomorphic to [b 1,b 2 ]. Hence [a 1,a 2] is isomorphic to
[b1 ,b2].
Suppose finally that <a 1,a2> :I: <h 1,b2>. We may choose the notation so that
b 1 ~ <a 1,a2> and a 1 ~ <b1,b2>. Then, by what we have already proved, [a1'a 2]
is isomorphic to [a 1,bi] and [a1tbtl is isomorphic to [b1tb2]. Hence again [a1'a2]
is isomorphic to [b 1,b2]. Cl
144 VII. Vector spaces
coplanar and hence i' = i. We now complete the definition off by putting
/(z) = i. It is not difficult to verify that/ is an isomorphism of C onto C with the
required properties.
It remains to show that our initial assumption may be removed. Choose
a 0 ' e Ao' and then c 0 e (a 0 ,a 0 ') and a 1 ' e (c 0 ,a 1). Also, let a0 ' = / 0 (a 0 '),
c0 = / 0(c0), and choose d1' e (c0 , d1). Then a 1' e (a 0,b) for some be (a0 ',a 1)
and d1 ' e (do, b) for some be (d0 ', d1). By what we have already proved, there
exists an isomorphism f' of C' =[Ao u a 11 onto C' = [Ao u a11 such that
f'(a 1 ') = d1 ' andf'(x) = f 0 (x) for x e C0 • Clearly it is enough now to show that
there exists an isomorphism cl> of C onto C' such that cj>(a 1) = a1' and cj>(x) = x for
x e C0. We will merely give the definition of cl> at other points.
For any x e C \ C0 with x ::1: a 1, there is a unique point x 0 e C 0 such that
x e (a 1,xo). If Xo ::/:co we put cj>(x) = x', where x' = (a 1 ',xo) ri (x,c 0 ). If
x e (a 1,c0 ), there is a unique x" e (a 1,a0) such that x e (x",a 0 ') and we define
cj>(x) = (cj>(x"),a0 ') ri (a 1',co). D
Proof We have already shown that there exists an ordered division ring D such
that, for any affine independent set A s X with IAI = 2, there is an affine
independent set A s D with IAI = 2 such that [A] is isomorphic to [A].
Suppose now that A is an affine independent set with IAI > 2. Let ?J be the
family of all triples (B, i. J), where B s A, i. is a vector space over D and/ is an
isomorphism of [B] onto a simplex in X. We partially order ?J by writing
(B, i. J) ~ (B', i. ',g) if B s B', i. is a subspace of X 'and g(x) = f(x) for x e [B].
Let ~ = { (B ;. X; /;)} be a maximal totally ordered subfamily of ?J and put
B =U B ;. X = U X;. If x e [B], then x e [B ;] for some i and we can define
/: [B] ~ i. by f(x) =f;(x). Moreover f(B) is an affine independent set, since any
finite subset is affine independent, and hence f is an isomorphism of [B] onto the
146 VII. Vector spaces
Let 8' be the image under/ of 8 n C, where 8 =<d1,d2>. Since the intersections
exist and are points of C, and since the lines <a 1,a2>,<c 1,c2>,<d1,d2> have the
common point x, it follows from the Desargues property that u, v, w are collinear
(see Figure 5). Under f the points a;, ... ,d; are mapped into points a/, ... ,d;' of C'
Figure 5
with preservation of segments and collinearity. Thus u',v',w' are collinear points of
V. Since V has the converse Desargues property (Proposition Vl.3), it follows
that the lines <a'>,<y'>,<8'> have a common point of intersection. Since the
intersection u = (b 1,c 1) n (b 2,c2) also exists and is a point of C, it follows in the
same way that the lines <P'>,<y'>,<8'> have a common point of intersection.
Hence the lines <a'>,<P'>,<y'> have a common point of intersection.
We now extend the definition off to X by putting/(x) = x'. We are going to
show next that distinct collinear points of X are mapped by f into distinct collinear
points of V. This implies, in particular, that/ is injective.
Let x,y,z be points of X such that x e (y,z) and let x',y',z' be their images
under/. There exists a finite subset {elt···,en} of E, with n ~ 3, such that
y,z e <e 1, ••• ,en>. Put P = [e 1, ••• ,en1 and choose p e pi with pi! <y,z>. There
exist points i, y, z e C such that p e (x, i), ye (p,y), ze (p,z). There exist also
148 Vil. Vector spaces
points x = (p,x) (y, z), y 1 e (i, z) such thatp e (y,y 1), and z1 e (i, y) such
f"\
thatp e (z ,z 1). Choose ye (y,z 1) and let x 1 e (i,y 1) be such that ye (y,x 1).
If we put a 1 = p, then there exist points
Let a2 e (z,c 1) be such that b 1 e (y,a2), let y2 e ( z ,y 1) be such that a 2 e (b 1,y 2),
and let z2 e (i,z 1) be such that c 1 e (a 2,z2). Finally, if we choose x 2 e (i,z 2),
there exist points
y x z
Figure 6
a; = (y,y;) f"\ (z,z;), b; = (z,z;) f"\ (x,x;), C; = (x,x;) f"\ (y,y;) (i = 1,2)
and
2. Isomorphisms 149
are distinct collinear points, it follows from Proposition VI.3 that the segments
(a 1,a2), (b1,b2), (c 1,c2) have a common point q. Since i, y, ze C, it follows that
x = <b t ,c1 > n <b2 ,c2 >, Y = <a1 ,c1 > n <a2 ,c2 i>, z = <a1 'bi
I I i I i I I i I I
' t > n <a2''b 2 > I
and <a 1',a2'>, <b 1',b2'>, <c 1',c2'> have the common point q' ~ a;',b;',c;' (i = 1,2).
Hence x',y',z' are distinct and collinear, by the Desargues property in V.
It will now be shown that/ maps the segment [y,z] of X onto the segment of
the line <y',z'> in V which has endpoints y',z' and contains x'. Suppose that
w e (y,z), where w ~ x, and put w = (p, w) n (y, z). Then the lines <y', y '>,
<z',z'>,<x',x'>,<w',w'> have the common point p' and x',w' e <y',z'>,
x',w' e <y',z'>. Since x ', w 'lie in the same open segment with endpoints y', z ',
it follows that x', w' lie in the same open segment with endpoints y',z'.
On the other hand, if w' is any point of the open segment with endpoints y',z'
which contains x', the point w' = <p',w'> n <y', z'> lies in the open segment
with endpoints y ', z' which contains x '. Hence w ' = /( w ) for some w e ( y, z)
and there exists a point we (y,z) such that w e (p,w). It follows that w' = /(w).
Thus f maps X isomorphically onto X' = f(X). D
Proof Let E = {e;: i e /} be an affine basis for X such that H = <E\e 1> and
e 1 e H+· By Theorem 3 there exists an isomorphism/ of the simplex C = [E] onto
a simplex C' = [E1 in a vector space Vover an ordered division ring D. Moreover
we may assume that V = <E'>. By Theorem 4 and its proof we may extend f to an
isomorphism of X into the projective completion V of V. The hyperplane
<E' \ e 1 '> of V extends to a hyperplane H' of V. We will show that
f<H+) n H' = 0.
Assume on the contrary thatx': =f(x) e H' for somex e H+. Let {e 1, ••. ,en}
be a finite subset of E, with n > 3, such that x e <e 1, ••• ,en>. Put P = [e 1, ••• ,enl
150 Vil. Vector spaces
3 VECTOR SUMS
The set A+ B will be called the vector (or Minkowski) sum of A and B, and aA
will be called a scalar multiple of A.
It is easily verified that the following properties hold for any nonempty sets
A,B ,C ~ X and any a,~ e D:
(iii) A + 0 = A, OA = 0.
3. Vector sums 151
Furthermore,
(v) (A. + µ)A = A.A + µA for any A.,µ e D with A.,µ > 0.
Proof Since A,B are convex, for any A.,µ > 0 with A. + µ = 1 we have
A.A+ µA~ A and A.B + µB ~ B. Hence
AA + µA. Hence to prove the last statement of the lemma we need only show that
if A is convex and if A.,µ > 0, A. + µ > 0, then A.A +µA !;:; (A. + µ)A. But if
x ="A.a.'+ µa", where a',a" e A, then
It may be noted that A is necessarily convex if (A.+ µ)A= A.A+ µA for all
A.,µ ~ 0, and that if IAI > 1, A. > 0 then
It follows by induction from Lemma 6 that, for any nonempty convex set A
and any positive integer n,
nA = A+ ... + A,
Proof If x e [S + 71, then there exist Sj e Sand tie T (j = l, ... ,n) such that
152 VII. Vector spaces
(A + B)i = Ai + Bi.
Proof We show first that Ai+ Bis;: (A + B)i. Let z = x + y, where x e Ai and
ye Bi, and let c =a+ b, where a e A, be Band c '¢ z. There existsµ> 1 such
that, for 1 < 'A < µ,
Proof Put
3. Vector sums 153
and hence
a+ q = c + aq" + a(l - p)-l(c' - a) + (1 - a)q'.
Consequently
PROPOSITION 13 If Pis a polytope and A,B nonempty convex sets such that
A+ B = P, then A and Bare also polytopes.
Proof Let P = [p 1, ... ,pn1· Then Pk= ak +bk for some ak e A and bk e B
(k = l, ... ,n). If A'= [a 1, ••• ,an] and B' = [b 1, ••• ,bn], then A'!::: A, B' ~Band
hence A'+ B's:: P. On the other hand, Pk e A'+ B' (k = l,... ,n) and A'+ B' is
convex, by Lemma 6. Hence P s:;: A ' + B '. Thus P = A ' + B ', and so
A + B' ~A'+ B'. Hence A ~A', by Proposition 11. Consequently A =A', and
similarly B = B'. D
4 NOTES
Suppose a 1,b 1,c 1 are distinct collinear points, and a 2 ,b 2 ,c 2 are distinct
collinear points not on the same line; if <a 1,b 2> intersects <a 2 ,b 1>, <b 1,c 2 >
intersects <b 2 ,c 1>, and <c 1,a 2 > intersects <c 2 ,a 1>, then the three points of
intersection are collinear.
Furthermore the Pappus property actually implies the Desargues property. (It is
pointed out in Seidenberg (1976) that the original proof of this by Hessenberg
(1905) was incomplete.) Proofs of these results for ordinary projective space are
given in Heyting (1980), for example. We do not establish them here, since
commutativity of multiplication will be a consequence of the additional axiom (S) in
Chapter VIIl.
Proposition 10 appeared in Klee (1959). Proposition 11 is a topology-free
version of RAdstrom's cancellation law (Proposition IX.4). It may be noted that
Proposition 11 no longer holds if 'polytope' is replaced by 'polyhedron'. For
156 VII. Vector spaces
A+PcP=C+P.
Bair and Fourneau (1975/80) contains additional results on vector sums, and
additional references.
VIII
Completeness
1 INTRODUCTION
(S) if a,b e X and if C is a convex subset of [a,b] such that a e C, then there
exists a point c e [a,b] such that either C = [a,c] or C = [a,c).
The axiom's label is chosen to indicate that c 'separates' the points of [a,b]
which are in C from those which are not. It is easily seen that the point c is
uniquely determined if the linear geometry X is dense. On the other hand, if
[a,b] = {a,b} for some distinct a,b e X and if C ={a}, then we can take either c =a
or c = b.
The axiom (S) may be regarded as an extension to linear geometries of the
Dedekind cut property for the real line JR. Any interval of JR is a complete dense
linear geometry. However, a segment in a complete dense linear geometry need not
be isomorphic to an interval of lR:
EXAMPLE 1 Let W be an uncountable set, e.g. the set of all sequences (an),
where an e {0, 1}. By the well-ordering principle, there is a total ordering of W
such that every nonempty subset of W has a least element. (The well-ordering
principle may be deduced from Hausdorffs maximality theorem; see, e.g., Hewitt
and Stromberg (1975).) Let Z be the set of all z e W for which the set
{w e W: w < z} is uncountable. If Z is nonempty, it has a least element ~- Put
Y = {y e W: y < ~} if Z ~ 0 and Y = W otherwise. Then Y is an uncountable
totally ordered set with the following properties:
The long line A is defined to be the set of all pairs (y ,p ), where y e Y and
p e (0,1), except for the pair (y0 ,0) where Yo is the least element of Y, with the
lexicographic ordering defined by (y 1,p 1) -< ()'2,P2) if Y1 < Y2, or if Y1 = Y2 and
p 1 < p 2. Since A is a dense totally ordered set with no least or greatest element, it
is a dense unending linear geometry with the definition of segments given in
Example VI.1. Furthermore the separation axiom (S) is satisfied. Since A has no
countable dense subset, it is not isomorphic to JR. However, any segment of A
containing more than one point is isomorphic to the interval (0,1] of JR.
Finally define a totally ordered set X = U ~=l X;, where X 1 and X 3 are copies
of the long line A, X2 = {p} and Xj < xk if xi e Xi, xk e Xk U < k). Then Xis a
1. Introduction 159
complete dense unending linear geometry, but the segment [a,b] is not isomorphic
to an interval of R. if a e X 1 and b e X \ X 1•
Proof Let {s 1,s2, ••• } and {r1,r2 , ••• } be enumerations of Sand Q respectively.
We wish to construct a bijective map <p: S ~ Q such that <p(sj) ~ <p(sk) according
as sj ~ sk for all distinctj,k e N. The construction is inductive.
Put <p(s 1) = r 1 and write sl = s 1, rl = r 1• Suppose we have defined subsets
sn = {s•, ... ,sn}, Qn = {r•, ... ,rn} of S, Q respectively and a bijective map
cp: Sn ~ Q n such that <p (si) ~ cp (sk) according as si ~ sk for all disinct
j,k e { l ,... ,n}. If n is odd define rn+ 1 to be the rational number rk with least k
which is not in Qn. Then either rn+ 1 is less than all elements of Qn, or rn+ 1 is
greater than all elements of Qn, or there exist two elements ri,rk of Qn such that
ri < ,.n+I < rk, but ri <,.; < rk for no ri e Qn. We define sn+l to be the element sk of
S with least k which is correspondingly less than all elements of sn, or greater than
all elements of Sn, or satisfies si < sn+l < sk, and we put cp(sn+l) = rn+l. If n is
even we interchange the roles of Sand Q. Thus we define sn+l to be the element sk
of S with least k which is not in Sn and we define ,.n+ I = <p(sn+ I) to be the rational
number rk with least k which has the same order relation with respect to Qn as sn+l
has with respect to Sn. Since Sand Qare both dense, and have no least or greatest
element, this construction is always possible. Moreover each element of S and Q is
ultimately included. a
PROPOSITION 2 If X is a complete dense linear geometry with dim X > 2,
then every nondegenerate segment of X is isomorphic to the interval [0, 1] of R..
Proof Let a and b be distinct points of X and choose Yoe (a,b). Define a
sequence {Yml recurrently by Ym e (a,ym-l) (m > 1). If we suppose the segment
[a,b] totally ordered in accordance with Proposition III.4 so that a< b, then {Yml is
a decreasing sequence. Denote by C the set of all points x e [a,b] such that x < Ym
for all m. Then C is convex, a e C and be: C. Hence, by (S), there exists a point
160 VIII. Completeness
x y z
Figure 1: z = /y(x)
For any ye (a,b), the iterates y(O) = y, y(l) = !y(y), y(2) = f//y(y)), ... form an
increasing sequence. We now show that, for each x e (a,b), there exists a non-
negative integer n such that y(n) > x. Assume on the contrary that y(n) < x for all
n > 0. It follows from (S) that there exists z e (a,b) such that y<n> S z for all n > 0,
1. Introduction 161
but y<n> > z' for some n > 0 if z' < z. Since z e (y,b), there exists we (a,b) such
that fy(w) = z. Then w < z and hence y(n) > w for some n > 0. Consequently
y(n+l) > /y(w) = z, which is a contradiction.
We show next that the set E of all points Ym(n), where {Yml is the sequence
constructed at the outset and n is any non-negative integer, is dense in [a,b].
Suppose u,v e (a,b) and u < v. Then there exists t e (a, v) such thatf,(u) = v.
Choose m so that y m < t and y m < u. There is an integer p > 0 such that
Ym(p) < u < Ym(p+l). Then Ym(p+l) <fy(u) <f,(u) = v. Thus Ym(p+l) e (u,v), as we
wished to prove.
Thus (a,b) is a totally ordered set with no least or greatest element and with a
countable dense subset E. Consequently, by Lemma 1, there is a bijective order-
preseiving map cp of E onto the set Q of rational numbers. For any c e (a,b) \ E,
put
A = {x e E: x < c}, B = {x e E: x > c}.
Suppose X is the open unit disc in the Euclidean plane, with the usual
definition of segments. Through any point not on a given line there pass infinitely
many lines which do not intersect the given line. However, there are two 'special'
ones, namely the intersections with X of the lines in R 2 which pass through the two
endpoints on the unit circle of the given line. This example may be borne in mind in
the statement and proof of the following proposition.
Furthermore <a 1,a2> t\ <.b 1,b2> = 0 and, if the lines <a 1,a2>,<b 1,b2> are
ordered so that a 1 < a 2, b 1 < b2, then for any a',a" e <a 1,a 2> with a'< a" and
any b',b" e <b 1,b2> with b' < b", b" has the same properties with respect to
a',a",b' as b 2 has with respect to a 1,a2,b 1.
Proof Choose w so that b 1 e (a 1,w). Then, for any x e H +' the ray [b 1 ,x>
contains a unique point c e (a 1,a2) t\ [a 2,w). We wish to show that there exists
b 2 e (a 2,w) such that [b 1,c> intersects [a 1,a2> if and only if c e [a 2,b2).
Let C be the set of all c e [a 2,w) such that [b 1,c> intersects [a 1,a 2>.
Evidently a2 e C and C is convex. If we choose a 3 so that a2 e (a 1,a3), then there
exists a point u e (b 1 ,a 3 ) t\ (a 2 ,w) and u e C. If we choose a 0 so that
a 1 e (a 0 ,a2), then b 1 e (a 0 ,v) for some v e (a 2,w) and v e: C. It now follows
from (S) that there exists b 2 e (a 2,w) such that C = [a 2,b2] or C = [a2 ,b 2). It
remains to show that b2 e: C.
Assume on the contrary that there exists a pointp e [a 1,a2> t\ [b 1,b2>. Then
b 2 e (b 1,p) and a2 e (a 1,p). If we choose a 3 so thatp e (a2,a3), then there exist a
point q e (a3,b 1) t\ (w,p) and a point re (b 1,q) t\ (w,b 2). Hence re Ct\ (TI\ C),
which is a contradiction.
Suppose the lines t = <a 1,a2>, m = <.b 1,b2> ordered so that a 1 < a2, b 1 < b2•
Since b2 E H+, we have both [a1,a2> (\ [b1,b2> = 0 and e(') [b1,b2> = 0. We
will show that actually e(\
m = 0.
Assume on the contrary that there exists a point p e t (') m. Then p < b 1 on
m, and hence p <al on .f. If we choose ao E .f SOthat ao < p, then b1 E (ao,Q) for
some q E (a2,bi). Hence there exists a point r E e(') [b1 ,q>. Since b1 E (ao,r)
and b 1 e: t, this is a contradiction.
Let a',a" be points oft with a'< a" and let b',b" be points of m with b' < b".
It is evident that if a'= a 1 and b' = b 1, then b" has the same properties with respect
to the ordered triple a',a",b' as b 2 has with respect to the ordered triple a 1,a2,b 1.
The same conclusion holds if b' = b 1 and a 1 <a', since [a',a"> (') [b 1,x> = 0 for
any x e H_ in the open half-plane of TI associated with the line <a',b 1> which
contains a". The same conclusion holds also if a'= a 1 and b 1 < b', since
[a 1,a"> (') [?J',x> "# 0 for any x e H_ in the open half-plane of TI associated with
the line <a 1,b'> which contains a". Since t "m = 0, it follows from what has
already been proved and the total ordering of the lines t,m that the same conclusion
holds also in the general case. []
1. Introduction 163
The uniquely determined ray [b1tb2> will be called the asymptote through b 1
to the ray [a 1,a2>. The second part of Proposition 3 shows that it depends only on
the oriented line <a 1,a2>, and not on the choice of a1,a2• It further shows that, in a
sense, it depends only on the oriented line <b 1,b2>.
In the same way, reversing the orientation on <a1ta 2>, there exists an
asymptote [b1tb 2 t> through b 1 to the ray [a 2 ,a 1>. In Euclidean space
<b 1,b2> =<b 1,b2t>, but the restriction of the Euclidean plane to the open unit disc
already shows that this need not be true in general.
Proposition 3 shows that, for any line eand any point p ~ e' there is at least
one line m in the plane <p u .f> such that p e m and t r. m = 0. Thus we may
regard Proposition 3 as a replacement for Euclid's parallel axiom. However, unlike
parallelism in Euclidean space, the asymptote relationship may be neither symmetric
nor transitive. To see this, in the second specific case of Example VI.5 take
a 1 = (-1,1), a2 = (-1,2) and b 1 = (0,1), b2 = (0,2). Then <a 1,a2> r. <b 1,b2> = 0
since, in the notation of Chapter VI, <a 1,a 2 > = 11,- 1 and <b 1,b 2 > = IL 0 •
Furthermore, if a e (a 1,a2> then [a,bi] is the usual straight line segment, since
a = (-1,a) for some a > 1 and hence <a,b 1> = .fl-a,l. It follows that [b1tb 2> is
the asymptote through b 1 to the ray [a 1,a2>. However, if a 3 = (-1/2,2), it is easily
seen that the asymptote through a 1 to [b 1,b2> is the ray [a 1,a3> of 9.
Proof It follows from Proposition 2 that, for any distinct points a,b e X, (a,b) is
order isomorphic to (0,1) and hence to R. We need only show that an arbitrary line
<a 1,a2> of X is order isomorphic to R.
Let b 1 be a point not on the line <a 1,a2>, choose w so that b 1 e (a 1,w), and
let b2 e (a 2,w) be defined as in the proof of Proposition 3. Also, choose a',a" so
that a 2 e (a 1 ,a ') and a 1 e (a 2 ,a"). By the proof of Proposition 3,
(a 2,a'> = [a 2,a'> \ a2 is order isomorphic to (a 2 ,b2), and hence to Rand to (1, 00) .
Similarly (a 1,a"> is order isomorphic to (a 1,b2t), and hence to Rand to (- oo,0).
Since [a 1,a2] is order isomorphic to [0,1], the result follows. a
164 VIII. Completeness
2 SEPARATION PROPERTIES
such that (x,y] s C and (x,z] s C'. In fact, if x = (xn), where Xn =0 for n > m,
we can take y = (yn) and z = (zn), where Yn =Zn =Xn for n :I: m+ 1 and Ym+l = 1,
Zm+l = -1.
Uoder our present hypotheses there is a simple characteriution of affine sets:
Proof Suppose first that C is affine. Then C is certainly convex and, since Xis
unending, Ci= C. It follows at once that also C = C.
Suppose next that C is convex and Ci = C. If C is not affine, there exist
distinct points c,c' e C and x e X \ C such that c' e (c,x). By (S), there exists
ye (c,x] such that (c,y) s C, (y,x) s X \ C. Since Ci= C = C, it follows that
ye C, y :1:-x and (y,x) contains a point c" e C, which is a contradiction. D
Proof By Proposition 6, H :I: 0. Since Ci= Ci, it follows that C is not affine
and thus C c <C>. Let c e Ci and a e <C> \ C. Then, by Proposition V .9, there
exists a point c' e (c,a) (')Ci. Hence, by (S), there exists a point h e (c,a] such
that (c,h) s Ci and (h,a] s X \Ci. Then he H, since he C, h :I: a and hence
h E Ci.
We now show that D = <C> \ Ci is convex. Assume on the contrary that there
exists c e Ci such that c e (d,d') for some distinct d,d' e D. If de C choose a so
that de (c,a), and if d E C put a = d. Similarly if d' e C choose a' so that
d' e (c,a'), and if d' e= C put a'= d'. Then a,a' E C. By the argument of the
preceding paragraph, the segments (c,a) and (c,a') each contain a point of H. Since
H is convex, it follows that c e H, which is a contradiction.
Thus Ci ,D is a convex partition of <C>. By Proposition ~it now need only be
shown that H =D \Di. As in the proof of Proposition 5, Ci = <C> \Di. But
C =Ci, since Ci :1:-0, and H = C (')D. Hence H =D \Di. D
166 VIII. Completeness
PROPOSmON 8 Let A and B be convex sets such that Ai*' 0, <A>= X and
B ¢. 0. Then A and B can be separated by a hyperplane if and only if
Ai "' B = 0, and Ai is then contained in one of the open half-spaces associated
with this hyperplane.
In Proposition 8 the hypothesis <A> = X is rather strong and the convex sets
A and B do not appear on an equal footing. These drawbacks are removed in the
following deeper-lying second separation theorem:
PROPOSITION 9 Let A and B be convex sets such that Ai¢. 0 and Bi¢. 0.
Then A and B can be properly separated by a hyperplane of X if and only if
Ai"' Bi= 0.
Proof Suppose first that A and Bare properly separated by a hyperplane H, and
assume that there exists a point x e Ai"' Bi. Then x e Hand there exists also a
point y e (A u B) \ H. Without loss of generality assume y e A \ H. Since
x e Ai, there exists a point z e A such that x e (y,z). Then y and z lie in different
open half-spaces associated with H, which is a contradiction.
Suppose next that Ai"' Bi= 0. We show first that if a hyperplane H properly
separates Ai and Bi, then it also properly separates A and B. Indeed Ai and Bi are
not both contained in H. If Ai, say, is not contained in H then, since Aii =Ai, Ai is
2. Separation properties 167
PROPOSITION 12 Let C be a convex set with Ci¢ 0 and let x e C \Ci. Then
there exists at least one hyperplane Hx of X, with x e H X' such that Ci is
contained in one of the open half-spaces associated with H x and C is
contained in the corresponding closed half-space.
Moreover, C is the intersection of all such closed half-spaces and Ci the
intersection of all such open half-spaces, for varying x e C \Ci.
PROPOSITION 13 Let C1, ... ,Cn (n > 1) be convex proper subsets of X such
that Ck;~ 0 (1 S k Sn) and
Then there exist hyperplanes Hk with associated open half-spaces Hk+ ,Hk-
such that Ck s Hk + (1 < k < n), Ck~ Hkfor at least one k, and
and
We have just shown that this is true for k = 2 and we wish to show that if it holds
as written, then it also holds when k is replaced by k + 1. Let
and let
Since Hn s Ln+ and z e Ln-, by supposition, there exists a point we (y,z) such
that w e Ln. Then
Hence there exists z' e (x,w) such that z' e Hk. Then
Since (x,z) s H 1+ (') ... (') Hk+ (') Hk+l (') .•. (') Hn_ 1, we must actually have
PROPOSITION 14 Let D be an ordered division ring. If, for any convex set
C s [0,1] with 0 e C and 1 E C, there exists c e [0,1] such that [0,c) s C and
(c,l] s D \ C, then there is a bijective map of D onto the real field R which
preserves sums, products and order.
Proof The interval [0,1) may be linearly mapped onto the interval [a,b], for any
a,b e D with a ':I: b. Consequently the property in the statement of the proposition
continues to hold if 0 is replaced by a and 1 by b.
In any ordered division ring D, nl = 1 + ... + 1 (n times) is positive for any
positive integer n. It follows that the set of all elements ml·(nl)-1, where mis any
integer and n is any positive integer, is a field isomorphic to the field Q of rational
numbers. By identifying ml·(nl)-1 with mn-1, we may regard Q as embedded in
D.
We show first that for each a e D there exists a positive integer n such that
n > a. Assume on the contrary that n < a for every positive integer n. Then a > 0
and the set C of all be [O,a] such that n < b for every positive integer n is convex
and contains a, but not 0. Hence there exists c e [0,a] such that (c,a] s C and
[0,c) s D \ C. Moreover 1 < c, since 2 E C. Thus if we choose de D so that
1 + d = c, then 0 < d < c. Since d E C, we have m > d for some positive integer m.
Hence 1 + m > 1 + d =c, which is a contradiction.
3. Fundamental theorem of ordered geometry 173
Consequently, for any a,b e D with a> 0, there is a positive integer n such
that n > ba-l, and then na > b.
We show next that, for any a,b e D with b > a, there exists x e Q such that
b > x > a. Let n be a positive integer such that n(b - a) > 1 and let m be the least
integer such that m > na. Then nb > na + 1 > (m - 1) + 1 = m > na and hence
b > mn-l >a.
It will now be shown that D is a field. It is sufficient to show that ab = ba for
all a> 0, b > 0. Assume on the contrary that there exist a> 0, b > 0 such that
ab > ba. Then ab > x > ba for some x e Q. Thus b-lx > a and a > xb-l, hence
b-lx > xb-l. Then b-lx > y > xb-l for some ye Q. Thus b-1 > yx-l and
x-ly > frl. Sinceyx-1 =x-ly, this is a contradiction.
For any a e D, the set of all x e Q with x <a is a 'cut' in Q. Moreover, the
hypothesis of the proposition ensures that all cuts in Q may be obtained in this way.
It is readily verified that the mapping which makes correspond to a the real number
defined by this cut in Dedekind's construction of the reals from the rationals has all
the required properties. D
isomorphism/ are disjoint nonempty convex sets with union C. Moreover Y+; = Y+
and y_i = Y_, since H+i = H+ and H_i = H_.
The theorem will be proved if we show that there is a 'hyperplane' of V
disjoint from C. Since it is sufficient to show that there is a 'hyperplane' disjoint
from C in the projective subspace of V generated by C, we may assume that C itself
generates V. Then there is a unique 'hyperplane' K of V containing Y and Y+,Y_
are contained in the affine space A = V \ K. Since A is the affine hull of Y+ and of
Y_, it follows from Proposition 8 that there is a hyperplane L of A such that Y+ and
Y_ are contained in different open half-spaces of A associated with the hyperplane
L. The hyperplane L of the affine space A extends to a 'hyperplane' L of V.
Evidently Y+ and Y_ are disjoint from L, since Lis obtained from L by adjoining
points of K. It only remains to show that Y is also disjoint from L. Suppose
c' e Y, so that c' = f(c) for some c e H. Thence (a,b) for some a e H+, be H_.
If a'= f(a), b' =f(b), then the line <a',b'> of the affine space A contains a point of
L. Since c' e Land a',b' e L, it follows that c' e L. D
and since an irreducible polynomial in F[x] has degree dividing [C: F] = p, the
irreducible factors of xP - 1 in F[x] are all linear. Thus F contains a primitive p-th
root of unity ~·
We are going to show that C = F(a), where aP e F. Let 'Ye C \ F, so that
C = F(y). If
CX = "( + ~g("() + ... + ~p-1 gP-l(y),
then
g(a) = g(y) + ~g2(y) + ... + ~P-lgP(y) = ~-la.
Thus a E F, since g(a) ':I: a. But g(a.P) = ~-Pcx,P = a.P. Hence a.I' e F and
C = F(a).
Choose p E c so that pP =a.
If b = n
'l=l gi-1(p), then g(b) = b and hence
be F. Moreover bP = Il ~=l gi-l(a) and gi-l(a) = ~-j+lcx, (j = l, ... ,p). If p -¢: 2,
then 1 + 2 + ... + (p - 1) = p(p - 1)/2 is divisible by p and hence bP = aP. Since
a E F, be F and F contains all p-th roots of unity, this is a contradiction. Hence
p = 2, ~ = -1 and b2 = - cx,2. Since (ib)2 = cx,2 and a E F, it follows that i E F. A
fortiori, i E K. If K(i) ':I: C, the same argument can be applied with K(i) in place of
K to obtain the contradiction i E K(i). Hence K(i) = C and n = 2. D
Proof The proof will be based on simple facts about filters. Let X be a nonempty
set. A nonempty collection~ of subsets of Xis said to be afilter on X when the
following properties hold:
Properties (Fl) and (F3) imply that the intersection of finitely many sets in ~
is nonempty and, since~ is nonempty, property (Fl) implies thatX e ~.
176 Vil/. Completeness
an ultrafilter, it follows that the set {x e X: cp(x) '# 0} e oU. Define v: X--+ D by
'lf(X) = cp(x)-1 if cp(x) '# 0, = 1 otherwise. If 'If E \ji, then ~· \ji = \ji· ~ = i.
Suppose ~' \ji e t and ~ ~ \ji. We wish to show that '\fl < ~· If cp e ~'
'II e 'if, then the set {x e X: cp(x) < 'lf(X)} e oU. Since oU is an ultrafilter, it
follows that the set {x e X: 'lf(X) < cp(x)} e oU, and hence also the set
{x e X: 'lf(X) S cj>(x)} e oU. Thus \ji S ~.
For any a e D, define cl>a:X--+ D by ct>a(x) =a for every x e X. If
cl>a e {~a}, then the map a --+ ~a is an injection of D into t which preserves
addition, multiplication and order. Consequently we may regard D as embedded in
r.
Suppose now that the ordered division ring D has a gap; i.e., there exist
nonempty subsets A,B of D such that a< b for every a e A, be B, but there is no
c e D such that a< c < b for every a e A, be B. This implies that A(') B = 0.
We wish to construct an ordered division ring D' :::> D in which the gap has been
removed, i.e. there exists c' e D' such that a< c' < b for every a e A, be B.
For any a e A, let Fa = {x e A: a < x} and le~ '!Ji denote the collection of all
sets F s A which contain a set Fa for some a e A. Evidently '!Ji is a filter on A.
Let oU be an ultrafilter on A such that '!Ji s oU, and now take t = t (A, oU) to be the
ordered division ring constructed with X =A and with this ultrafilter. Define the
map ·Y: A --+ D by 'Y(x) = x for every x e A and let y be the corresponding element
oft. For each a e A, the set {x e A: a S y(x)} e oU, and for each be B, the set
{x e A: y(x) < b} =A e oU. Thus, regarding D as embedded in t, we have
a< y < b for every a e A, be B.
After these preparations we can proceed with the proof of Proposition 18.
Assume first that D is a proper subfield of R. Then, by Lemma 17, R is infinite-
dimensional, considered as a vector space over D. Hence, by the discussion of
Example VI.6, for any positive integer n the n-dimensional projective space over D
can be given the structure of a dense, unending linear geometry.
Thus we now assume that no extension of D is isomorphic to the real field R.
Since D itself is not isomorphic to R, it follows from Proposition 14 that the
interval [0,1] is the union of two nonempty subsets A,B such that a< b for every
a e A, b e B, but there is no c e D such that a < c < b for every a e A, b e B.
Hence, as we have seen, there exists an ordered division ring D' :::> D with an
element c' such that a Sc' Sb for every a e A, be B. Since no extension of Dis
isomorphic to R, this argument can be repeated any finite number of times.
178 VIII. Completeness
Consequently, for any positive integer n, there exists an ordered division ring
D(n) :::> D which has dimension > n as a vector space over D. The conclusion now
(Tl) the whole set X and the empty set 0 are in~'
(Tl) the intersection of two sets in~ is again a set in~'
(T3) the union of any family of sets in ~ is again a set in ~.
The pair (X,d) is a metric space. A subset A of a metric space (X,d) is said to be
bounded if {d(x,y): x,y e A} is a bounded subset of R.
4. Metric and norm 179
A metric space (X ,d) has a natural topology in which the open sets are the
unions of sets of the form
N(x0,e) = {x e X: d(x0 ~) < e},
where x0 e X and e > 0. This topology will always be understood when speaking
of open or closed subsets of a metric space. For a metric space there is a simpler
characterization of compactness: a metric space Xis compact if and only if every
sequence {xn} of elements of X has a convergent subsequence {xntl, i.e. there
exists x e X such that d(xnt~)--+ 0 ask--+ co.
If a subset A of a metric space X is compact, with the topology induced by that
on X, then A is closed and bounded. In general it is not true that, conversely, a
closed and bounded set is compact, but it is true for X =Rd.
A sequence {xn} in a metric space Xis said to be a Cauchy sequence if for
each real e > 0 there is a corresponding integer p > 0 such that d(xm~n) < e for all
m,n > p. A metric space is said to be complete if every Cauchy sequence is
convergent Any compact metric space is complete.
The two uses of the word 'complete' in this chapter correspond to the two
ways, due to Dede.kind and Cantor, of constructing the real numbers from the
rationals. It is hoped that the meaning in each instance will be clear from the
context.
Suppose now that V is a real vector space. We say that a norm is defined on
V, and that Vis a normed vector space, if with each v e V there is associated a real
number llvll with the properties
It follows that Vis a metric space with the metric d(v,w) = llv-wll.
The following two propositions will be proved, since the proofs are less
accessible.
Proof Let d be the dimension of the vector space V and let e 1, ••• ,ed be a (vector
space) basis for V. Then each v e V can be uniquely represented in the form
180 VIII. Completeness
By the induction hypothesis there exists a positive constantµ' such that, for every
v e V, llv'll < µ 'lv'I. Then
llvll < llv'll + llv"ll < µ'jv'I + ladl lledll < µjvj,
whereµ=µ'+ lle~j.
The real-valued function cp{v) = llvll is continuous on S, since
Let p =inf {cp(v): v e S} and let {vn} be a sequence of elements of S such that
cp(vn> ~ p. Since S is compact, by restricting attention to a subsequence we may
suppose that there exists v* e S such that lvn - v*I ~ 0. Then cp{v*) = p, since <p
is continuous, and p > 0, since v* :I: 0. Since llvll > p for lvl = 1, it follows from
(N3) that p lvl < llvll for every v e V. 0
and hence
Proposition 19, there exist a subsequence {vnl of {un} and a vector v e V with
llvll = 1 such that llvn - vii-+ 0. Let Yn'Y be the elements of P(V) corresponding to
vn,v. For n >p we have llvn-vll < 1 and hence llvn +vii> 1. Consequently
182 VIII. Completeness
5 NOTES
The long line is a fruitful source of counterexamples in topology; see Steen and
Seebach (1978). For the roots of Proposition 3 in the investigations by Gauss on
non-Euclidean geometry, see Section 12.6 of Coxeter (1989). Proposition 4 is
proved in Doignon (1976).
Separation properties of two convex sets, and in particular the Hahn-Banach
theorem, play a fundamental role in functional analysis; see, for example, Bourbaki
( 1987) and Holmes (197 5). The separation of several convex sets, which is
important in optimization, is treated in Bair and Foumeau (1980) and Boltyanskii
(1975).
Theorem 16 is contained in Doignon (1976), but naturally it had many
forerunners. The theorem throws an interesting light on Examples 11.1, VI.2 and
VI.3. These examples can be given a common formulation in the following way: let
X be a connected finite-dimensional Riemannian manifold such that any two points
a,b of X are joined by a unique geodesic arc, and define the segment [a,b] to be the
set of all points on this arc. A theorem of Beltrami, in conjunction with Theorem
16, shows that the three cited examples are the only complete unending linear
geometries which can be obtained in this way if dim X > 2, or if dim X = 2 and X
has the Desargues property. Beltrami's theorem states that if Xis a connected
Riemannian n-manifold such that every point has a neighbourhood which can be
mapped homeomorphically into JRn, so that geodesic arcs are mapped onto straight
line segments, then X has constant curvature. The theorem is proved in Spivak
(1975), for example. A proof which assumes only continuity of the Riemannian
metric is sketched by Pogorelov (1991).
If dim X = 2 and X does not have the Desargues property, then the possibilities
are by no means so restricted. Busemann (1955), Section 11, shows that in
Example VI.5 a metric d can be defined on JR2 so that, for any points x,y,z e JR2,
5. Notes 183
It follows that [x,y] is the unique geodesic arc joining x and y. Another proof,
using methods derived from integral geometry, is described in Salzmann (1967). It
may be asked: can a metric d with the property (t) be defined on any complete
unending two-dimensional linear geometry?
For an approach which starts from a metric space (X ,d) and defines the
segment [x,y] to be the set of all points z such that d(x,y) = d(x,z) + d(z,y), see
Rinow (1961) (and p. 420, in particular, for the connection with our approach).
Proposition 18 was first proved by Szczerba (1972) for two-dimensional V,
using model theory rather than filters. The main ingredients in the proof of Lemma
17, namely the fundamental theorem of algebra and the fundamental theorem of
Galois theory, are proved in Lang (1984), for example. Lemma 17 is actually a
special case of a more general result of Artin and Schreier: if an algebraically closed
field F has a proper subfield K of which it is a finite extension, then F has
characteristic zero and F = K(i), where i2 = - 1. See, for example, Priess-Crampe
(1983).
IX
Spaces of Convex Sets
This chapter is concerned with spaces of convex sets. We show that a metric may
be defined on the space ~(X) of all nonempty bounded closed subsets of a metric
space X. Furthermore this Hausdorff metric is complete if the metric of X is
complete. These results are then applied to the space '<6 (X) of all nonempty
bounded closed convex subsets of a normed real vector space X. The vector sum
of two elements of '<6(X) need not again be in '<6(X), if Xis infinite-dimensional.
However, the closure of the vector sum is in '<6 (X) and, with this modified
definition of addition, '<6 (X) has the structure of a commutative semigroup.
R4dstrom's cancellation law implies that this semigroup may be embedded in a
group. In fact '<6(X), with the Hausdorff metric and with the partial order induced
by inclusion, may be embedded isometrically and isomorphically in a vector lattice
with two further properties, which guarantee the existence of a norm. Vector
lattices with these two further properties are here called Kakutani spaces.
A basic theorem, the Krein-Kakutani theorem, says that an arbitrary Kakutani
space can be mapped isometrically and isomorphically onto a dense subset of the
Kakutani space C(K) of all continuous real-valued functions on some compact
Hausdorff space K. Thus, as a final result, for a given normed real vector space X
there exists a compact Hausdorff space K such that the space '<6(X) of all nonempty
bounded closed convex subsets of X may be embedded in the space C(K) of all
continuous real-valued functions on K. To make the discussion self-contained, we
also prove the Krein-Kakutani theorem.
Let (X,d) be a metric space and let ~(X) denote the family of all nonempty
bounded closed subsets of X. We are going to show that a metric can be defined
also on ~(X).
186 IX. Spaces of convex sets
Evidently h{A,B) < oo, since A and B are bounded. Moreover h(A,B) > 0, with
equality if and only if A = B. It is obvious from the definition that
h{A,B) = h(B ,A}. It remains to prove the triangle inequality
Proof We wish to show that if {A 11 } is any Cauchy sequence in ~(X), then there
exists A e ~(X) such that An-+ A as n-+ oo. For each n = 1,2, ... let Fn denote
the closure of the set Um~nAm , so that F 1 ~ F 2 ~ · · • • Then A = n;=t F n is a
bounded closed subset of X. We will show that A has the required properties.
Fix any e > 0. For each k = 1,2, · · · there exists a positive integer nk such that
Moreover we may assume that n 1 < n2 < · · · . Choose any point x1 e An 1 and
define inductively a sequence {xk} so that xk e An.t and d(xk,xk_ 1) < e/2k. Then
{xk} is a Cauchy sequence, since d(xj,xk) < e/2k for allj > k. Since Xis complete,
there exists a point y 1 e X such that xk -+ y 1 ask-+ oo. Clearly, d(x 1,y 1) Se.
Moreover y 1 e A, since Xj e F n,t for all j > k and F n,t is closed. Thus A is
nonempty and
S(x,A) < £ for all x e An 1•
On the other hand, let B denote the set of all points ye X such that S(y,An 1) < e.
Then An c B for all n > n 1• Since B is closed, it follows that F ni c B and hence
AcB. Thus
S(y,An 1) < £ for ally e A.
Therefore h(A,An 1) Se and, by the triangle inequality, h(A,An) < 2e for all n > n 1•
Thus we have shown that An -+ A as n -+ oo. 0
Let ':K(X) denote the family of all nonempty compact subsets of X. Then
'X(X) ~ ~(X), since any compact subset of Xis closed and bounded.
Proof We need only show that ':K(X) is a closed subset of the complete metric
space GJ(X),h). Suppose An e ':K(X) and An -+ A in ~(X). We wish to show that
A e ':K(X). Thus we wish to show that any sequence {xn} of elements of A has a
subsequence converging to an element of A. In fact, since X is complete and A is
188 IX. Spaces of convex sets
closed, it is enough to show that the sequence {Xn} has a subsequence which is a
Cauchy sequence.
There exists a positive integer n 1 such that h(A,A,J S 1/2 for all n ~ n 1. Since
An1 is compact, it is covered by finitely many open balls with radius 1/2. It follows
that A is covered by finitely many open balls with the same centres and radius 1.
Thus there is a ball B 1 with radius 1 which contains Xn for infinitely many n. In the
same way there is a ball B 2 with radius 1/2 such that B 1 n B 2 contains Xn for
infinitely many n. In general, for each positive integer k there is a ball Bk with
radius 1/2k-l such that B 1 n ... n Bk contains Xn for infinitely many n. Hence
there exists an increasing sequence {n k} of positive integers such that
Xnk E B 1 n ... n Bk· Evidently the subsequence {Xnkl is a Cauchy sequence. a
The next result shows that the space of nonempty compact subsets of X has the
Bolzano-Weierstrass property if X itself has. The proof uses the diagonal process
which Cantor first used to show that the set of real numbers is uncountable.
To apply the preceding results to convex sets we now impose more structure
on the metric space. Throughout this section we assume that Xis a normed real
vector space. Thus with each x e X there is associated areal number llxll with the
properties
(i) A EB B = B EB A, (A EB B) EB C =A EB (B EB C),
(ii) cx(A EBB) =a.A EB cxB, (cx~)A = cx(~A), IA =A,
(iii) A EB 0 = A, OA = 0,
(iv) if A ~ B, then A EB C ~ B EB C and a.A c: cxB,
(v) if A.,µ~ 0, then (A.+ µ)A =A.A EB µA.
Since the proofs for the other properties are similar but simpler, we prove only the
associativity of addition:
(A EB B) EB C = A EB (B EB C).
where bn' =(bi + ... + bn)/n e B, since B is convex, and Xn' =(xi + ... + Xn)/n e
(1/n)U, by construction. Since C is bounded and Bis closed, it follows that a e B.
Thus A~ B. D
2. The space ~(X) 191
To illustrate the role played by convexity in this result we now show ~at if
X = JR.d, Ac X and C =[A], then
tf-lA + C = tf-lC + C.
Since the left side is obviously contained in the right, and since d-lC + C =
d-l(d + l)C, we need only show that C ~ (d + 1)-lA + d(d + 1)-lC. If x e C then,
by Corollary 111.18, there exists a finite affine independent set F c A such that
x e [F]. Thus x = L ~ 1 A;X;, where X; e A and A; > 0 (i = l ,... ,m ), L ;!1 A; = 1
and m < d + 1. Moreover we may choose the notation so that A. 1 = max; A;. Then
A. 1 > m-1 > (d+l)-1 and
If we choose A bounded and closed, but not convex, then C is a bounded Closed
convex set, but tf-lA ~ d-lC.
For a normed vector space the definition of the Hausdorff metric can be
reformulated in the following way:
LEMMA 6 If X is a normed real vector space then, for any sets A,B e ?1(X),
Proof Since A is nonempty, bounded and closed, we need only show that it is
convex. Let z = A.x + (1 -A.)y, where x,y e A and 0 <A.< 1. By Lemma 6, for
any £ > 0 there is a positive integer m such that
z e An + eU ~ A + 2eU.
3 EMBEDDINGS OF '€(X)
Since '€ has the structure of a commutative semigroup under addition, with a
zero element and with multiplication by non-negative real numbers as a semigroup
of operators, and since the cancellation law for addition holds, it is possible to
embed~ in a vector space. We now sketch the construction, which is similar to
that by which the semigroup of non-negative integers is embedded in the group of
all integers (and to that used at the end of Section 1 in Chapter Vill).
3. Embeddings of ~(X) 193
If (A,B) and (C ,D) are ordered pairs of elements of'€, we write (A,B) - (C ,D)
if A El1 D = B El1 C. It follows directly from Corollary 5 that - is an equivalence
relation. Let {A,B} denote the equivalence class containing the pair (A,B) and let
;;£ = fi(X) denote the set of all equivalence classes {A,B}.
If {A,B} and {C,D} are elements of;;£, we define their sum by
It is easily seen that the sum is uniquely defined, i.e. if {A',B'} = {A,B} and
{C',D'} = { C,D} then
It is not difficult to verify that with these definitions;;£ is a vector space with {0,0}
as the zero vector. Evidently any {A,B} e ;;£has the form
It follows from Proposition 4 that this order relation is uniquely defined and is
indeed a partial order. Also, if {A,B} < {C,D} then, for any {E,F} e ff, and any
A.> 0,
{A,B} + {E,F} < {C,D} + {E,F},
A.{A,B} < A.{C,D}.
Furthermore any two elements {A,B} and {C,D} of;;£ have a least upper
bound, namely {G,B El1 D }, where G = (A El1 D) v (B El1 C). For we certainly
have
{ A,B} = {A e D' B e D}
< {G ,B e D}'
{C,D} = {B El1 C,B EBD} < {G,B EBD}.
On the other hand, if {A,B} < {E,F} and {C,D} < {E,F}, then
194 TX. Spaces of convex sets
A EB D EB F ~ B EB D EBE, B EB C EB F c B EB D EB E,
and hence
G EB F = (A Ea D EB F) v (B Ea C EB F) ~ B EB D Ea E.
Thus~ is a vector lattice.
We draw attention also to two other properties of ~. First, for any
{A,B} e ~there is a positive integer n such that
Indeed we can choose the positive integer m so large that A,B ~ mU and then, if
n > 2m, we will have A~ B + nU andB ~A+ nU =A+ (-n)U.
Secondly, if {A,B},{C,D} e ~and n{A,B}< {C,D} for every positive
integer n, then {A,B} < {0,0}. Equivalently, if nA Ea D ~ nB EB C for every
positive integer n, thenA ~ B. Indeed if a e A and de D, then there exist bn e B,
en e C and un e U such that
THEOREM 10 Let X be a normed real vector space and, with the Hausdorff
metric h derived from the norm, let ("6(X),h) be the metric space of nonempty
bounded closed convex subsets of X.
Then there exists an injective map <p o/"6 = "6(X) into a Kakutani space:£
with the following properties:
A.'€' + µ'€' ~ <(6 'for all A.,µ > O and:£ ='€' - '€ '. a
By identifying A with <p(A), we may regard'€ as embedded in:£. From
Theorem 10 we immediately obtain the translation invariance of the Hausdorff
metric (which can also be proved directly):
COROLLARY 11 Let X be a normed real vector space and, with the Hausdorff
metric h derived from the norm, let ("6(X),h) be the metric space of nonempty
bounded closed convex subsets of X. If A,B,C e <(6(X) and A.> 0, then
and/< g is defined by
f(t) < g(t) for every t e K.
Moreover, for any f,g e C(K), if nf < g for every positive integer n, then/< 0.
Finally, for any f e C(.K) there is a positive integer n such that- ne <f < ne, where
e e C(.K) is the constant function defined by
The associated norm for the Kakutani space C(K) is the supremum norm:
The Krein-Kakutani theorem, which will be proved in the next section, says
that an arbitrary Kakutani space can be mapped isomorphically and isometrically
onto a dense subset of the Kakutani space C(.K) for some compact Hausdorff space
K. By composing the injective map <p: <€(X) ~ fl of Theorem 10, with the linear
isometry and lattice isomorphism J of fl into C(K) guaranteed by the Krein-
Kakutani theorem, we immediately obtain
THEOREM 12 Let X be a normed real vector space and, with the Hausdorff
metric h derived from the norm, let (<€(X),h) be the metric space of nonempty
bounded closed convex subsets of X.
Then there exist a compact Hausdorff space K and an injective map 'I' of
<€ = <€(X) into the Kakutani space C(K) of all continuous real-valued/unctions
on K with the following properties:
In this section we prove the Krein-Kakutani theorem, not only to make our
account self-contained but also because the existence in our case of an order unit
makes possible some simplifications. At the same time the proof may serve as a
minicourse on convexity in functional analysis.
A linear functional on a normed real vector space Vis a map/: V ~ R such
that
J(ax + (3y) = aj(x) + (3/(y) for all x,y e Vandall a,(3 e R..
4. The Krein-Kakutani theorem 197
The set V' of all bounded linear functionals on Vis also a normed vector space if we
define
(a/+ pg)(x) = aft.x) + pg(x),
llf 11 = sup {lf(x)I: llxll = 1}.
For any non-zero vector x0 e V, there is a linear functional f e V' such that
llf 11 = 1 and/(xo) = llxoll. This is a special case, with p(x) =llxll, w = {a.xo: a E R}
and g(ax0) = allxoll for a e R, of the analytic form of the Hahn-Banach theorem:
(HB) Let W be a vector subspace of the real vector space V and let p: V ~ R
be a sublinear functional on V, i.e.
p(x + y) < p(x) + p(y), p(/...x) = 'Ap(x) for all x,y e Vandall real A. > 0.
then there exists a linear functional f on V such that f(y) = g(y) for every y e W
and
f(x) < p(x) for every x e V.
Proof Let
which implies p(x) <a. On the other hand, suppose p(x) <a and choose a' so
that p(x) <a'< a. For any (y,p) e V x JR, if 0 < 0 < 1 and 0 is sufficiently close
to 1, then
p(0x + (1 - 0)y) < 0a' + (1 - 0)p(y) < ea+ (1 - 0)p.
Since g(y) ~p(y), it follows that h(O,l) > 0. For any (x,cx.) e Vx JR,
Since V' is also a normed vector space, we can in the same way define the
normed vector space V" =(V1' of all bounded linear functionals on V'. For any
x e V, the map lx: V' ~ R. defined by lif) = f(x) is a bounded linear functional on
V'. Moreover IVJI = l~I, since
(BA) The closed unit ball U' = {/ e V': llf 11 S 1} of V' is compact in the weak*
topology.
Proof lf/e U', then lf(x)I Sllxll and so/(x) lies in the compact interval
Ix= C- llxll,llxlll. If P = Ilxevlx then U' ~ P, since Pis the set of all functions
g: V -> R such that g(x) e Ix for every x e V. Moreover the topology U' inherits as
a subspace of Pis the weak* topology. ButP is a product of compact spaces and
hence itself compact, by Tychonoffs theorem. Hence, to show that U' is compact,
we need only show that it is a closed subset of P.
Let g e P be a point in the closure of U'. Then g maps V into R and
lg(x)I <II.xii for each x e V. Suppose z =ax+ py, where x,y e V and a,p e R.
For each e > 0, the set N = {h e P: lh(t) - g(t)I < e for all t e A}, where
A = {x,y,z}, is an open subset of P which contains g. Since g is in the closure of
U', there exists/ e U' r.N, and since/is linear,/(z) = af<.x) + pj(y). Hence
Since this inequality holds for each e > 0, we inust actually have
(KM) If C is a nonempty weak* compact convex subset of V', then the set K of
all extreme points of C is nonempty and C is the weak* closure of the convex
hull of K.
Proof The family ~ of all nonempty weak* compact faces of C is nonempty, since
it contains C itself. If we regard ~ as partially ordered by inclusion then, for any
200 IX. Spaces of convex sets
and let D be the subset of all points of E at which the supremum is attained. Since
E is weak* compact, D is nonempty and weak* closed. Hence D is also weak*
compact. Moreover D is convex, since if h1th 2 e D and h = 9h 1 + (1 - 9)h 2 ,
where 0 < 9 < 1, then
h(x) = 9µ + (1 - 9)µ = µ.
[y- cl>(f)]/p < cl>(u') < - [y- cl>(/)]/p for all u' e U',
4. The Krein-Kakutani theorem 201
the linear functional ct> is actually bounded. Hence ct> attains its supremum on C at
an extreme point e of C. Since e e A, cl>(e) < y< ci><IJ. Hence/ e C, which is a
contradiction. a
A real vector space V is said to be a vector lattice if a partial order < is defined
on V with the properties
(i) if x,y e V and x < y, then x + z < y + z for every z e V and A.x < A.y for
every real A. > 0,
(ii) any two vectors x,y e V have a least upper bound x v y.
It follows that any two vectors x,y e V also have a greatest lower bound
x " y, namely
x "y = - ((-x) v (- y)).
We also have the following simple properties: if x,y,z e V and A.> 0, then
[Proof Put
u = x v y, v = (x + z) v (y + z), w =(Ax) v (A.y).
From x,y < u we obtain x + z < u + z, y + z < u + z and hence v < u + z. On the
other hand from x + z < v, y + z < v we obtain x < v - z, y < v - z. Hence
u < v - z and u + z < v. This proves that v = u + z, and similarly we can prove that
w = AU if A > 0.
The corresponding relations with v replaced by " follow immediately. Also,
by what we first proved,
[Proof Put
u =(x " z) v (y " z), v =x v y.
Since x "z < v "z and y A z <v A z, we certainly have u < v A z. On the other
hand, since
x A z =x + z - (xv z), y A z =y + z -(y v z),
we have
x + z - (x v z) < u, y + z - (y v z) < u,
In particular, if we put
x+ =x v 0, x- = - (x A 0) = (- x) v 0,
then
x+ ~ 0, x- ~ 0, x+ A x- = 0,
x+-x- =x, x+ +x- =xv (-x).
V+={xe V:x~O}.
We will say that a vector lattice V is a Kakutani space if it has the additional
properties
(iii) if x,y e V and nx < y for every positive integer n, then x < 0,
(iv) there exists e e V such that, for each x e V, there is a positive integer n
for which-ne <x < ne.
It is easily verified that with this definition the Kakutani space Vis a normed vector
space. Moreover norm and order are connected by the properties
and hence
h(x) < l(x) for every x e V+·
It is clear from the definition of h that h('A.x) = A.h(x) for every x e V + and
every A. e R+, and lh(x)I S <IYll +Ilg ll>llxll. We will show that also
Hence
h(Xi +X2) = sup utY1) + /(y2) + g(x1-Y1) + g(x2-Y2): 0 <Yi < Xtt 0 < Y2 < X2}
= h(x1) + h(x2 ).
k(x + x') =k(x) + k(x1, k(-x) =- k(x), k(A.x) =A.k(x) for A. e R+,
If/ e v; then llf 11 = f(e), by what we have just proved, and hence
(NS)' llf + gll = lltll + llgll iff.g e v;.
This property replaces the property (NS) in V. It will now be shown that the
property (N4) continues to hold in V', i.e.
We have
If we put w = y - z, then
w+ - w- =w, w+ + w- =w v (- w).
Hence
IJf+ +f-11 = sup lf(w): w+ + w- Se}
or, since llw+ + w-11 = llwll,
Proof Suppose first that/is a lattice homomorphism. If/is not an extreme point
of C, then/= 8/1 + (1 - 8)/2 for some/1/ 2 e C with/1 ~ / 2 and some 8 e (0,1).
206 IX. Spaces of convex sets
Iff(x+) = 0 for some x e V, then / 1(x+) = / 2(x+) = 0, and similarly iff(r) = 0 then
/ 1 (x-) = / 2 (x-) = 0. If /(x) = 0 for some x e V, then /(x+) = f(x-) = 0, since
/(0) = 0 and/ is a lattice homomorphism. Since /;(x) = /;(x+) - /;(r) (i = 1;2), it
follows that/(x) =0 implies/1(x) =/ 2(x) =0.
The set H = {z e V:/(z) = 0} is a hyperplane of V which passes through the
origin. We have shown that / 1 (z) = / 2 (z) = 0 for every z e H. Since
/ 1(e) =/ 2(e) = 1 and every x e V has the form x = z + ae for some z e H, a e R,
it follows that/1(x) =a= / 2(x). Thus/1 = / 2, which is a contradiction.
Suppose next that f is an extreme point of C. Assume first that
f(x+) "/(r) = 0 for every x e V. Since x = x+ -r, we have f(x) = /(x+) - /(x-).
Since f(x+) > O,/(r) ~ 0 and min {f(x+)J(.r)} = 0, by hypothesis, it follows that
/(x+) = /(x)+,f(r) =f(x)-. From x v y = y + (x -y)+ we now obtain
(SW) Let K be a compact Hausdorff space containing more than one point, let
C(K) be the Kakutani space of all continuous real-valued functions on K, and
let E be a vector sublattice of C(K), i.e. E is a subset of C(K) such that if
f,g e E and a,p e R then a/+ pg e E and f v g e E.
Suppose also that E contains the constant function 1 and that E separates
the points of K, i.e. if s,t e K and s '¢ t then there exists f e E such that
f(s) ~ f(t).
Then Eis dense in C(K), i.e. for any he C(K) and any e > 0, there exists
f e E such that llf- hll < e.
Proof We first observe that, for any distinct points s,t e Kand any a,p e JR, there
exists/ e E such that/(s) = a,f(t) = p. Indeed if g e Eis such that g(s) "¢ g(t), we
can t.ake
f = [g(s) -g(t)]-1( (a- ~)g + ~g(s)- a.g(t) }.
Lethe C(K) and, for any distinct points s,t e K, letfst e E be such that
fs 1(s) = h(s) andfs,(t) = h(t). For given e > 0, define open subsets Ost of K by
Then, for each t e K, K =U seK 0 st' since s e 0 st· Since K is compact, there is a
finite subcover:
If ft= /s 11 v ... v fsmt' then ft e E,ft(t) = h(t) andft(u) > h(u) - e for all u e K.
Now define an open subset 0 1 of Kby
K = 0 11 u ... u 0 1n
Proof Let
c = lf e v;: llfll = 1} = lf e v;:Jte) = 1 },
and let K be the set of all extreme points of C. The theorem will be established in a
very explicit manner by proving
We have already seen that the weak* topology on V' is a Hausdorff topology,
that C is a compact convex subset of V', and that K is the set of all/ e C which are
lattice homomorphisms. It follows that K is closed, and hence compact. Since the
map J~: V' ---+ R defined by J~ (j) = /(x) is a bounded linear functional, its
restriction lx to K is certainly continuous and
where K' is the set of extreme points of the unit ball U' = {/ e V', lltll ~ 1 }. It is
evident that if/ e K', then lltll = 1 and hence, by (N4)'-(N5)',
would not be an extreme point of U'. Since x ~ 0, in evaluating llxll we can restrict
attention to those I e K' for which I= r' and these I are in K. Hence J is an
isometry:
llxll = sup f.ftx):/ e K} = l~(x)ll- D
5 NOTES
distinguish between -A and (- l)A. In fact the two are distinct for any A e '€with
IAI > 1.
The Krein-K.akutani theorem is usually formulated for complete Kakutani
spaces, or 'AM-spaces with unit' in the official terminology. It was independently
proved by M.G. and S.G. Krein, and by K.akutani, in the years 1940-1941. The
theorem is discussed in a more general context in the books of Day (1973),
Schaefer (1974) and Meyer-Nieberg (1991). Tychonoffs theorem is proved in
Hewitt and Stromberg (1975), for example.
References
L. Danzer, B. Grilnbaum and V. Klee (1963), Belly's theorem and its relatives,
Convexity (ed. V. Klee), pp. 101-180, Proc. Symp. Pure Math. 7, Amer.
Math. Soc., Providence, R.I.
M.M. Day (1973), Normed linear spaces, 3rd ed., Springer-Verlag, Berlin.
G. Debreu (1967), Integration of correspondences, Proc. Fifth Berkeley Symp.
Math. Statist. Probability (ed. L.M. Le Cam and J. Neyman), Vol. II,
Part l, pp. 351-372, University of California Press, Berkeley.
J.-P. Doignon (1973), Convexity in cristallographic lattices, J. Geom. 3, 71--85.
J.-P. Doignon (1976), Caracterisations d'espaces de Pasch-Peano, Acad. Roy.
Belg. Bull. Cl. Sci. (5) 62, 679-699.
P. Duchet (1987), Convexity in combinatorial structures, Rend. Circ. Mat.
Palermo (2) Suppl. No. 14, 261-293.
J. Eckhoff (1979), Radon's theorem revisited, Contributions to geometry (ed. J.
Tolke and J.M. Wills), pp. 164-185, Birkhauser, Basel.
J. Eckhoff (1993), Belly, Radon, and Caratheodory type theorems, Handbook of
convex geometry (ed. P.M. Gruber and J.M. Wills), Volume A, pp. 389-
448, North-Holland, Amsterdam.
P.H. Edelman and R.E. Jamison (1985), The theory of convex geometries, Geom.
Dedicata 19, 247-270.
J.W. Ellis (1952), A general set-separation theorem, Duke Math. J. 19, 417-421.
M. Erne (1984), Chains, directed sets and continuity, Institut fur Mathematik,
Universitit Hannover, Preprint Nr. 175.
Euclid (1956), The thirteen books of Euclid's elements, English translation by
T.L. Heath, 2nd ed., reprinted in 3 vols. by Dover, New York.
R. Frank (1992), Ein lokaler Fundamentalsatz fiir Projektionen, Geom. Dedicata
44, 53-66.
H. Grassmann (1844), Die Iineale Ausdehnungslehre, Leipzig. [Reprinted in
Gesammelte Werke 11, Chelsea, New York, 1969]
B. Griinbaum (1967), Convex polytopes, Interscience, New York.
R. Hammer (1977), Beziehungen zwischen den Satzen von Radon, Belly und
Caratheodory bei axiomatischen Konvexit!ten, Abh. Math. Sem. Univ.
Hamburg 46, 3-24.
F. Hausdorff (1927), Mengenlehre, 2nd ed., de Gruyter, Berlin. [English
translation of 3rd ed., Set theory, Chelsea, New York, 1962]
E. Belly (1923), Ober Mengen konvexer Korper mit gemeinschaftlichen Punkten,
Jahresber. Deutsch. Math.-Verein. 32, 175-176.
G. Hessenberg (1905), Beweis des Desarguesschen Satzes aus dem Pascalschen,
Math. Ann. 61, 161-172.
E. Hewitt and K. Stromberg (1975), Real and abstract analysis, 3rd printing,
Springer-Verlag, New York.
A. Heyting (1980), Axiomatic projective geometry, 2nd ed., North-Holland,
Amsterdam.
References 213
Propositions
(P)' 54 (C)",(P)" 92
11.2', 11.17', 11.19' 54 11.2", 11.17", 11.19" 93
Examples
11.1,2 28 Vl.2,3,4 116
Il.3 29 VI.5 117
11.4,5,6,7 ,8 30 Vl.6 122
11.9 34 VI.7 123
11.10 46 Vlll.1 158
Vl.1 115
Index
convex face
hull 6, 31 of polyhedron 110, 111
partition 42, 97, 164 of polytope 74, 76, 78
set 6, 28, 31 of sum 152, 154
cover 21 proper 10
open 178 facet
covering property 21 of convex set 73
of polyhedron 110
d-connected graph 80 of polytope 76, 79
dense linear geometry 85 factor geometry 81, 132
Desargues property 124, 127 filter 175
converse of 127 fundamental theorem
Desargues theorem 121, 136 of ordered geometry 173
converse of 122 of projective geometry 155
deviation 186
dimension 23, 57 generating set 9
directe.d family 25
division ring 28 Hahn-Banach theorem 167, 197
ordere.d 28, 142, 172 half-space 61
Hausdorff
Eckhoff's conjecture 70 distance 186
e.dge 73 metric 186
Euler-Poincare relation 84 space 178, 97
exchange Bausdorffs maximality theorem 5
alignment 20, 56 Belly
property 20 number 25
extension theorem 167 set 11, 38, 66
extreme Belly's theorem 67
point 8, 32, 153, 199 hemispace 42
subset 25 hull operator 7
algebraic 7
face hyperbolic geometry 116
of convex set 9, 33, 71, 84 hyperplane 22, 57
of intersection 10, 90 'hypeiplane' 120
220 Index
.. norm 179
JOin
supremum 4
symmetric convex cone 108
ultrafilter 176
unending linear geometry 95
upper bound 4
vector
lattice 201, 194
sum 150
vertex
of a cone 104
ofagraph 80
vertex set
of a cone 104
of a convex cone 104
of a polyhedron 113
visible point 35