152 - Lectures On Polytopes PDF

Ginter M.
Ziegler
Lectures on
Polytopes
Revised First Edition
Springer
Gunter M. Ziegler
Technische Universta Berlin
Fachbereich Mathematik, MA 6-1
Berlin, D10623
Germany
Editorial Board
S Axler
Department of
Mathematics
San Francisco State University
San Francisco, CA 94132
USA
F.W Gehring
Department of
Mathematics
University of Michigan
Ann Arbor, MI 48109
USA
K.A. Ribet
Department of
Mathematics
University of California
at Berkeley
Berkeley, CA 94720-3840
USA
Mathematics Subject Classification (1991) 52-02, 52B05, 52B11, 52B12

Library of Congress Cataloging-in-Publication Data
Ziegler, Gunter M
Lectures on Polytopes I Gunter M Ziegler
p cm (Graduate texts in mathematics, 152)
Includes bibliographical references and index
ISBN 0-387-94365-X (paper) ISBN 0-387-94329-3 (hard)
1 Polytopes I Title. II Series
94-21784
QA691 Z54 1994
516 3 5dc20
Printed on acid-free paper
1995 Springer-Verlag New York, Inc
All rights reserved This work may not be translated or copied in whole or in part
without the written permission of the publisher (Springer-Verlag New York, Inc , 175 Fifth
Avenue, New York, NY 10010, USA), except for brief excerpts in connection with reviews
or scholarly analysis. Use in connection with any form of information storage and retrieval,
electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed is forbidden
The use of general descriptive names, trade names, trademarks, etc, in this publication,
even if the former are not especially identified, is not to be taken as a sign that such names,
as understood by the Trade Marks and Merchandise Marks Act, may accordingly be used
freely by anyone
Production managed by Karen Phillips, manufacturing supervised by Gail Simon
Photocomposed prepared from the author's LATEX file
Printed and bound by Braun-Brumfield, Inc , Ann Arbor, MI
Printed in the United States of America
9 8 7 6 5 4 3 2 (Corrected second printing, 1998)
ISBN
ISBN
ISBN
ISBN
0-387-94329-3 Springer-Verlag
3-540-94329-3 Springer-Verlag
0-387-94365-X Springer-Verlag
3-540-94365-X Springer-Verlag
New York Berlin Heidelberg (hard cover)

Berlin Heidelberg New York
New York Berlin Heidelberg (soft cover)
Berlin Heidelberg New York SPIN 10644937
Preface
The aim of this book is to introduce the reader to the fascinating world
of convex polytopes. The book developed from a course that I taught at
the Technische Universitat Berlin, as a part of the Graduierten-Kolleg "Algorithmische Diskrete Mathematik." I have tried to preserve some of the
flavor of lecture notes, and I have made absolutely no effort to hide my
enthusiasm for the mathematics presented, hoping that this will be enough
of an excuse for being "informal" at times.
There is no P2C2E in this book.*
Each of the ten lectures (or chapters, if you wish) ends with extra notes
and historical comments, and with exercises of varying difficulty, among
them a number of open problems (marked with an asterisk*), which I hope
many people will find challenging. In addition, there are lots of pointers to
interesting recent work, research problems, and related material that may
sidetrack the reader or lecturer, and are intended to do so.
Although these are notes from a two-hour, one-semester course, they
have been expanded so much that they will easily support a four-hour
course. The lectures (after the basics in Lectures 0 to 3) are essentially
independent from each other. Thus, there is material for quite different twohour courses in this book, such as a course on "duality, oriented matroids,
and zonotopes" (Lectures 6 and 7), or one on "polytopes and polyhedral
complexes" (Lectures 4, 5 and 9), etc.
*P2C2E = "Process too complicated to explain" [434j
vi
Preface
Still, I have to make a disclaimer. Current research on polytopes is very

much alive, treating a great variety of different questions and topics. Therefore, I have made no attempt to be encyclopedic in any sense, although the
notes and references might appear to be closer to this than the text. The
main pointers to current research in the field of polytopes are the book by
Griinbaum (in its new edition [234]) and the handbook chapters by Klee
Sz Kleinschmidt [301] and by Bayer Sz Lee [59].
To illustrate that behind all of this mathematics (some of it spectacularly
beautiful) there are REAL PEOPLE, I have attempted to compile a bibliography with REAL NAMES (Le., including first names). In the few cases where
I couldn't find more than initials, just assume that's all they have (just like
T. S. Garp).
In fact, the masters of polytope theory are really nice and supportive
people, and I want to thank them for all their help and encouragement
with this project. In particular, thanks to Anders Bj6rner, Therese Biedl,
Lou Billera, Jiirgen Eckhoff, Eli Goodman, Martin Henk, Richard Hotzel,
Peter Kleinschmidt, Horst Martini, Peter McMullen, Ricky Pollack, J6rg
Rambau, Jiirgen Richter-Gebert, Hans Scheuermann, Tom Shermer, Andreas Schulz, Oded Schramm, Mechthild Stoer, Bernd Sturmfels, and many
others for their encouragement, comments, hints, corrections, and references. Thanks especially to Gil Kalai, for the possibility of presenting some
of his wonderful mathematics. In particular, in Section 3.4 we reproduce
his paper [272],
GIL KALAI:
A simple way to tell a simple polytope from its graph,

J. Combinatorial Theory Set. A 49 (1988), 381-383;
(D1988 by Academic Press Inc.,
with kind permission of Academic Press.
My typesetting relies on ILYI .X; the drawings were done with xf ig. They
may not be perfect, but I hope they are clear. My goal was to have a
drawing on (nearly) every page, as I would have them on a blackboard, in
order to illustrate that this really is geometry.
Thanks to everybody at ZIB and to Martin Graschel for their continuing
support.
Berlin, July 2, 1994
Giinter M. Ziegler
Preface to the Second Printing
At the occasion of the second printing I took the opportunity to make

some revisions, corrections and updates, to add new references, and to
report about some very recent work.
However, as with the original edition there is no claim or even attempt to
be complete or encyclopedic. I can offer only my own, personal selection. So,
I could include only some highlights from and pointers to Jiirgen RichterGebert's new book [424], which provides substantial new insights about
4-polytopes, and solved a number of open problems from the first version of
this book, including all the problems that I had posed in [537]. A summary
of some recent progress on polytopes is [539].
Also after this revision I will try to update this book in terms of an electronic preprint "Updates, Corrections, and More," the latest and hottest
version of which you should always be able to get at
http://www.math.tu berlin.de/ - ziegler

-
Your contributions to this update are more than welcome.

For the first edition I failed to include thanks to Winnie T. Pooh for
his support during this project. I wish to thank Therese Biedl, Joe Bonin,
Gabor Hetyei, Winfried Hochstdttler, Markus Kiderlen, Victor Klee, Elke
Pose, Jiirgen Pulkus, Jiirgen Richter-Gebert, Raimund Seidel, and in particular Giinter Rote for useful comments and corrections that made it into
this revised version. Thanks to Torsten Heldmann for everything.
Berlin, June 6, 1997
Giinter M. Ziegler
Contents
Preface
Preface to the Second Printing
0
Introduction and Examples

Notes
Problems and Exercises
v
vii
1
22
23
1 Polytopes, Polyhedra, and Cones

1.1 The "Main Theorem"
1.2 Fourier-Motzkin Elimination: An Affine Sketch
1.3 Fourier-Motzkin Elimination for Cones
1.4 The Farkas Lemma
1.5 Recession Cone and Homogenization
1.6 Carathodory's Theorem
Notes
27
27
32
37
39
43
45
47
49
51
51
55
59
64
65
Faces of Polytopes
2.1 Vertices, Faces, and Facets
2.2 The Face Lattice
2.3 Polarity
2.4 The Representation Theorem for Polytopes
2.5 Simplicial and Simple Polytopes
Contents
2.6 Appendix: Projective Transformations
Notes
67
69
70
Graphs of Polytopes
3.1 Lines and Linear Functions in General Position
3.2 Directing the Edges ("Linear Programming
for Geometers")
3.3 The Hirsch Conjecture
3.4 Kalai's Simple Way to Tell a Simple Polytope
from Its Graph
3.5 Balinski's Theorem: The Graph is d-Connected
Notes
77
77
Steinitz' Theorem for 3-Polytopes
80
83
93
95
96
97
4.1 3-Connected Planar Graphs

4.2 Simple AY Transformations Preserve Realizability
4.3 Planar Graphs are AY Reducible
4.4 Extensions of Steinitz' Theorem
Notes
103
104
107
109
113
115
119
Schlegel Diagrams for 4-Polytopes

5.1 Polyhedral Complexes
5.2 Schlegel Diagrams
5.3 d-Diagrams
5.4 Three Examples
Notes
127'
127
132
138
139
143
145
Duality, Gale Diagrams, and Applications

6.1 Circuits and Cocircuits
(a) Affine Dependences
(b) Affine Functions
6.2 Vector Configurations
6.3 Oriented Matroids
(a) Axiomatics
(b) Equivalence
(c) Duality
(d) Deletion and Contraction
6.4 Dual Configurations and Gale Diagrams
6.5 Polytopes with Few Vertices
(a) A Nonrational 8-Polytope
(b) Facets of 4-Polytopes Cannot be Prescribed
149
150
150
153
156
157
159
160
163
163
165
171
172
173
Contents
xi
(c) 2-Faces of 5-Polytopes Cannot be Prescribed

(d) Polytopes Violating the Isotopy Conjecture
6.6 Rigidity and Universality
Notes
175
177
179
183
184
Fans, Arrangements, Zonotopes,

and Tilings
7.1 Fans
7.2 Projections and Minkowski Sums
7.3 Zonotopes
7.4 Nonrealizable Oriented Matroids
7.5 Zonotopal Tilings
Notes
Problems and Exercises .............
191
191
195
198
208
217
224
225
Sheliability and the Upper Bound Theorem

8.1 Shellable and Nonshellable Complexes
8.2 Shelling Polytopes
8.3 h-Vectors and Dehn-Sommerville Equations
8.4 The Upper Bound Theorem
8.5 Some Extremal Set Theory
8.6 The g-Theorem and Its Consequences
Notes
231
232
239
246
254
258
268
275
281
Fiber Polytopes, and Beyond

9.1 Polyhedral Subdivisions and Fiber Polytopes
9.2 Some Examples
9.3 Constructing the Permuto-Associahedron
9.4 Toward a Category of Polytopes ?
Notes
291
292
299
310
319
320
321
References
325
Index
365
o
Convex polytopes are fundamental geometric objects- to a large extent the

geometry of polytopes is just that of tC itself. (In the following, the letter
d usually denotes dimension.)
The "classic text" on convex polytopes by Branko Griinbaum [234] has
recently celebrated its twenty-fifth anniversary and is still inspiring reading. Some more recent books, concentrating on f-vector questions, are
McMullen & Shephard [374], Brondsted [126], and Yemelichev, Kovalev
& Kravtsov [533]. See also Stanley [478] and Hibi [252]. For very recent
developments, some excellent surveys are available, notably the handbook
articles by Klee & Kleinschmidt [301] and by Bayer & Lee [59]. See also
Ewald [189] for a lot of interesting material, and Croft, Falconer Sz Guy [160]
for more research problems.
Our aim is the following: rather than being encyclopedic, we try to
present an introduction to some basic methods and modern tools of polytope theory, together with some highlights (mostly with proofs) of the
theory. The fact that we can start from scratch and soon reach some exciting points is due to recent progress on several aspects of the theory that
is unique in its simplicity. For example, there are several striking papers
by Gil Kalai (see Lecture 3!) that are short, novel, and probably instant
classics. (They are also slightly embarrassing, pointing us to "obvious" (?)
ideas that have long been overlooked.)
For these lectures we concentrate on combinatorial aspects of polytope
theory. Of course, much of our geometric intuition is derived from life in le
(which some of us might mistake for the "real world," with disastrous
results, as everybody should know). However, here is a serious warning:
0. Introduction and Examples
part of the work (and fun) consists in seeing how intuition from life in
three dimensions can lead one (i.e., everyone, but not us) astray: there are
many theorems about 3-dimensional polytopes whose analogues in higher
dimensions fail badly. Thus, one of the main tasks for polytope theory is
to develop tools to analyze and, if possible, "visualize" the geometry of
higher-dimensional polytopes. Schlegel diagrams, Gale diagrams, and the
Lawrence construction are prominent tools in this direction tools for a
more solid analysis of what polytopes in d-space "really look like."
Notation 0.0. We stick to some special notational conventions. They are
designed in such a way that all the expressions we write down are "clearly"
invariant under change of coordinates.
In the following R d represents the vector space of all column vectors of
length d with real entries. Similarly, ( d)* denotes the dual vector space,
that is, the real vector space of all linear functions 1!ì R. These are
given by the real row vectors of length d.
The symbols x, x o ,
. , y, z always denote column vectors in IY (or
in R d1 ) and represent (affine) points. Matrices X, Y, Z,... represent sets
of column vectors; thus they are usually (d x m)- or (d x n)-matrices. The
order of the columns is not important for such a set of column vectors.
Also, we need the unit vectors ei in
which are column vectors, and
the column vectors 0 and 1 = Ei ei of all zeroes, respectively ail ones.
The symbols a, ao , al , . . . , b, c, . . . always denote row vectors in ( d)* ,
and represent linear forms. In fact, the row vector a E (Rd )* represents the
linear form t = fa :R d ---+ R, z
ax. Here ax is the scalar obtained as
the matrix product of a row vector (i.e., a (1 x d)-matrix) with a column
vector (a (d x 1)-matrix). Matrices like A, A', B, . . . represent a set of row
vectors; thus they are usually (n x d)- or (m x d)-matrices. Furthermore,
the order of the rows is not important.
We use 11 = (1, ... ,1) to denote the all-ones row vector in (Rd)*, or
in (Rdi)* Thus, liz is the sum of the coordinates of the column vector x.
Similarly, 0 = (0, . . . , 0) denotes the all-zeroes row vector.
Boldface type is reserved for vectors; scalars appear as italic symbols,
such as a, b, c, d,x, y.... Thus the coordinates of a column vector x will be
, xd E R, and the coordinates of a row vector a will be al,. - , adBasic objects for any discussion of geometry are points, lines, planes and
so forth, which are affine subspaces, also called flats. Among them, the
vector subspaces of r d (which contain the origin 0 E d) are referred to as
linear subspaces Thus the nonempty affine subspaces are the translates of
linear subspaces.
The dimension of an affine subspace is the dimension of the corresponding
linear vector space. Affine subspaces of dimensions 0, 1, 2, and d 1 in Rd
are called points, lines, planes, and hyperplanes, respectively.
For these lectures we need no special mathematical requirements: we just
assume that the listener/reader feels (at least a little bit) at home in the
0 Introduction and Examples
real affine space Rd , with the construction of coordinates, and with affine
maps x 1----+ Ax + x o , which represent an affine change of coordinates if A
is a nonsingular square matrix, or an arbitrary affine map in the general
case.
Most of what we do will, in fact, be invariant under any affine change
of coordinates. In particular, the precise dimension of the ambient space is
usually not really important. If we usually consider "a d-polytope in
then the reason is that this feels more concrete than any description starting
with "Let V be a finite-dimensional affine space over an ordered field,
and ...."
We take for granted the fact that affine subspaces can be described by
affine equations, as the affine image of some real vector space k or as the
set of all affine combinations of a finite set of points,
n
F = {x E
: x = A0x0 + . - + Ax n for Ai G R,
E Ai = 1}.
That is, every affine subspace can be described both as an intersection of

affine hyperplanes, and as the affine hull of a finite point set (i.e., as the
intersection of all affine flats that contain the set). A set of n > 0 points is
affinely independent if its affine hull has dimension n 1, that is, if every
proper subset has a smaller affine hull.
A point set K C Rd is convex if with any two points x, y E K it also
contains the straight line segment [x, y] = Ax + (1 ) )y : 0 < A < 1}
between them. For example, in the drawings below the shaded set on the
right is convex, the set on the left is not. (This is one of very few nonconvex
sets in this book.)
{
Clearly, every intersection of convex sets is convex, and d itself is convex.

Thus for any K C Rd, the "smallest" convex set containing K, called the
convex hull of K, can be constructed as the intersection of all convex sets
that contain K:
conv(K) :=
fl { K'
: K C K', If` convex}.
Our sketch shows a subset K of the plane (in black), and its convex hull
conv(K), a convex 7-gon (including the shaded part).
For any finite set {X i ,. , xk} C K and parameters A 1 ,

Ak > 0 with
Ai
Ak = 1, the convex hull conv(K) must contain the point Aixi +
+ Akxk: this can be seen by induction on k, using
AkXk = ( 1 AO (
AIX'
A1
Ak
4-1
xk-1) + Akxk
1 Ak
for Ak < 1. For example, the following sketch shows the lines spanned by
four points in the plane, and the convex hull (shaded).
Geometrically, this says that with any finite subset K0 C K the convex
hull conv(K) must also contain the projected simplex spanned by K0 . This
proves the inclusion "D" of
conv(K) = {A i x ' + + AkXk {Xi,
,Xk} C K, Ai
= 11.
>
i=1
But the right-hand side of this equation is easily seen to be convex, which
proves the equality.
Now if K
, xn } Rd is itself finite, then we see that its convex
fxl,
hull is
conv(K) = {A i x ' +. + Anxn : n > 1, Ai >

i=1
The following gives two different versions of the definition of a polytope.

(We follow Griinbaum and speak of polytopes without including the word
convex": we do not consider nonconvex polytopes in this book.) The two
versions are mathematically but not algorithmically equivalent The
proof of equivalence between the two concepts is nontrivial, and will occupy
us in Lecture 1.
{{
Definition 0.1. A V-polytope is the convex hull of a finite set of points

in some Rd
An '11-polyhedron is an intersection of finitely many closed halfspaces in
some Rd. An 7-1 -polytope is an 7i-polyhedron that is bounded in the sense
that it does not contain a ray { x + ty : t > 0} for any y $ 0. (This
definition of "bounded" has the advantage over others that it does not rely
on a metric or scalar product, and that it is obviously invariant under affine
change of coordinates.)
.
A polytope is a point set P C a which can be presented either as a

V-polytope or as an 71-polytope.
The dimension of a polytope is the dimension of its affine hull.
A d-polytope is a polytope of dimension d in some I (e > d).
Two polytopes P C Rd and Q C I e are ajfinely isomorphic, denoted
by PL`-=-1 Q, if there is an affine map f : Rd ---* Re that is a bijection
between the points of the two polytopes. (Note that such a map need not
be injective or surjective on the "ambient spaces.")
Our sketches try to illustrate the two concepts: the left figure shows a
pentagon constructed as a V-polytope as the convex hull of five points; the
right figure shows the same pentagon as an 7-1-polytope, constructed by
intersecting five lightly shaded halfspaces (bounded by the five fat lines).
Usually we assume (without loss of generality) that the polytopes we
study are full-dimensional, so that d denotes both the dimension of the
polytope we are studying, and the dimension of the ambient space Rd.
The emphasis of these lectures is on combinatorial properties of the faces
of polytopes: the intersections with hyperplanes for which the polytope is
entirely contained in one of the two halfspaces determined by the hyperplane. We will give precise definitions and characterizations of faces of
polytopes in the next two lectures. For the moment, we rely on intuition
from "life in low dimensions": using the fact that we know quite well what
a 2- or 3-polytope "looks like." We consider the polytope itself as a trivial
face; all other faces are called proper faces. Also the empty set is a face for
every polytope. Less trivially, one has as faces the vertices of the polytope,
which are single points, the edges, which are 1-dimensional line segments,
and the facets, i.e., the maximal proper faces, whose dimension is one less
than that of the polytope itself.
We define two polytopes P,Q to be corribinatorially equivalent (and denote this by Pi_-_, Q) if there is a bijection between their faces that preserves
the inclusion relation. This is the obvious, nonmetric concept of equivalence that only considers the combinatorial structure of a polytope, see
Section 2.2 for a thorough discussion.
Example
0.2. Zero-dimensional polytopes are points, one-dimensional
polytopes are line segments. Thus any two 0-polytopes are affinely isomorphic, as are any two 1-polytopes.
Two-dimensional polytopes are called polygons. A polygon with n vertices is called an n-gon. Convexity here requires that the interior angles (at
the vertices) are all smaller than 7r. The following drawing shows a convex
6-gon, or hexagon.
Two 2-polytopes are combinatorially equivalent if and only if they have

the same number of vertices. Therefore, we can use the term "the convex
n-gon" foF the combinatorial equivalence class of a convex 2-polytope with
exactly n vertices. There is, in fact, a nice representative for this class: the
regular n-gon,
sin( 2"k
)) - 0 < k < n} C
P2(n) := cony {(cos()
n
n
,
The following drawing shows the regular hexagon P2 (6) in R2 . It is cornbinatorially equivalent, but not affinely isomorphic, to the hexagon drawn
above.
Example 0.3. The tetrahedron is a familiar geometric object (a 3-dimensional polytope) in 3

'
Similarly, its d-dimensional generalization forms the first (and simplest)

infinite family of higher-dimensional polytopes we want to consider. We
O. Introduction and Examples
define a d-simplex as the convex hull of any d + 1 affinely independent

points in some In (n > d).
Thus a d-simplex is a polytope of dimension d with d + 1 vertices. Naturally the various possible notations for the d-simplex lead to confusion,
in particular since various authors of books and papers have their own, inconsistent ideas about whether a lower index denotes dimension or number
of vertices. In the following, we consistently use lower indices to denote
dimension of a polytope (which should account for our awkward P2 (n) for
an n-gon... ).
It is easy to see that any two d-simplices are affinely isomorphic. However,
it is often convenient to specify a canonical model. For the d-simplex, we
use the standard d-simplex Ad with d + 1 vertices in d+1 ,
Ad
:=
IX E
d+1 : 11 X = 1, Xi > 0} = COriV{ei, ,ed i i}

- -
Our figures illustrate the construction of 6. 2 in
Example 0.4. The three-dimensional cube C3 and the octahedron C3 4 ' are
familiar objects as well:
Their generalization to d dimensions is straightforward. We arrive at the

d-dimensional hypercube (or the d-cube, for short):
Cd := Ix
E Rd : 1 < xi < 1} = conv{{1, 1}d},
and the d-dimensional crosspolytope:

Cd
Ix E Rd :
lxil < 11 = conv{e i , e l , ... , ed , ed}.
We have chosen our "standard models" in such a way that they are
symmetric with respect to the origin. In this version there is a very close
connection between the two polytopes Cd and Cd: they satisfy
Cd''
Cd
f' -i- fa E ( d )* : ax < 1 for all

____ fa E (Rd )* : ax < 1 for all
E Cd 1
G Cd'},
that is, these two polytopes are polar to each other (see Section 2.3).
Now it is easy to see that the d-dimensional crosspolytope is a simplicial
polyt,ope, all of whose proper faces are simplices, that is, every facet has
the minimal number of d vertices. Similarly, the d-dimensional hypercube
is a simple polytope: every vertex is contained in the minimal number of
only d facets.
These two classes, simple and simplicial polytopes, are very important. In
fact, the convex hull of any set of points that are in general position in Rd
a simplicial polytope. Similarly, if we consider any set of inequalities is
in i d that are generic (i.e., they define hyperplanes in general position)
and whose intersection is bounded, then this defines a simple polytope.
Finally the two concepts are linked by polarity: if P and P are polar,
then one is simple if and only if the other one is simplicial.
(The terms "general position" and "generic" are best handled with some
amount of flexibility you supply a precise definition only when it becomes
clear how much "general position" or "genericity" is really needed. One can
even speak of "sufficiently general position"! For our purposes, it is usually
sufficient to require the following: a set of n> d points in Rd is in general
position if no d of them lie on a common affine hyperplane. Similarly, a set
of n > d inequalities is generic if no point satisfies more than d of them
with equality. More about this in Section 3.1.)
Here is one more aspect that makes the d-cubes and d-crosspolytopes
remarkable: they are regular polytopes polytopes with maximal symmetry. (We will not give a precise definition here.) There is an extensive and
very beautiful theory of regular polytopes, which includes a complete classification of all regular and semi-regular polytopes in all dimensions. A lot
can be learned from the combinatorics and the geometry of these highly
regular configurations ("wayside shrines at which one should worship on
the way to higher things," according to Peter McMullen).
At home (so to speak) in 3-space, the classification of regular polytopes
yields the well-known five platonic solids: the tetrahedron, cube and octahedron, dodecahedron and icosahedron. We do not include here a drawing of the icosahedron or the dodecahedron, but we refer the reader to
Griinbaum's article [239] for an amusing account of how difficult it is to

get a correct drawing (and a "How to" as well).
The classic account of regular polytopes is Coxeter's book [156]; see also
Martini [352, 353], Blind & Blind [99], and McMullen & Schulte [375] for
recent progress. The topic is interesting not only for "aesthetic" reasons,
but also because of its close relationship to other parts of mathematics,
such as crystallography (see Senechal [455]), the theory of finite reflection
groups ("Coxeter groups," see Grove & Benson [231] or Humphreys [265]),
and root systems and buildings (see Brown [128]), among others.
Example 0.5. There are a few simple but very useful recycling operations
that produce "new polytopes from old ones."
If P is a d-polytope and x 0 is a point outside the affine hull of P (for
this we embed P into n for some n> d), then the convex hull
pyr(P) := conv(P U {x 0 })
is a (d + 1)-dimensional polytope called the pyramid over P. Clearly the
affine and combinatorial type of pyr(P) does not depend on the particular
choice of x0 just change the coordinate system. The faces of pyr(P) are
the faces of P itself, and all the pyramids over faces of P.
Especially familiar examples of pyramids are the simplices (the pyramid over Ad is Ad+i ), and the Egyptian pyramid Pyr3 = pyr(P2(4)): the
pyramid over a square.
Similarly we construct the bipyramid bipyr(P) by choosing two points x +

and x_ outside aff(P) such that an interior point of the segment [x + , x_]
is an interior point of P. As examples, we get the bipyramid over a triangle
x+
x_
and the crosspolytopes, which are iterated bipyramids over a point,
bipyr(Cd ) =
Cd+1 A
Especially important, it is quite obvious how to define the product of two

(or more) polytopes: for this we consider polytopes P c IIIP and Q c Rq,
10
and set
PxQ := { (1 : x E P, y E Q}.
Y
We get a polytope of dimension dim(P) + dim(Q), whose nonempty faces
are the products of nonempty faces of P and nonempty faces of Q.
in particular
The prism over a polytope P is the product of P with a segment,
prism(P) := Px A i .
This is polar to the bipyramid:
prism(P) = (bipyr(P A )) .
The smallest interesting prism is the one over a triangle,
also known as the triangular prism.
A2 X Ai ,
The cubes can be interpreted as iterated prisms, starting with a point.

In particular, we get CdX[-1,1] =
Products of simplices are interesting polytopes and more complicated
than one might think (see Problem 5.3(iii)*, an unsolved conjecture).
Just consider P := A2 x A 2 , the product of two triangles. This is a
4-polytope with 9 vertices. It has 6 facets, of the form "edge of one
triangle x the other triangle"- thus they all are triangular prisms.
Furthermore, the intersection of two of them is either "one of the
triangles x a vertex of the other triangle," or it is "an edge x an
edge." In either case the intersection is 2-dimensional. Hence any two
facets of P are adjacent, and PA = (A2 x .6.2 ) A is a 4-polytope with
6 vertices such that any two of them are adjacent. Thus PA is a
2-neighborly 4-polytope that is not a simplex: there is no analogue to
this "phenomenon" in 3-space (Exercise 0.0).
Taking products of several convex polygons, we can construct polytopes "with many vertices." Namely, assuming that d is even, we can
construct a 4fold product of in-gons, which yields a d-dimensional
polytope with "only" V. facets, but with 7/1d/ 2 vertices. If d is odd,
we can use a prism over such a product.
(For fixed dimension d, this simple construction of polytopes with
many vertices is asymptotically optimal, as we will see in Section 8.4.)

Example 0.6. The moment curve in
11
is defined by
/
t i--+ x(t) :=
t2
\ td
ER".
The cyclic polytope Cd(t i ,. . . ,t,i ) is the convex hull
Cd(t i ,...,t,i ) := cony fx(ti),x(t2),...,x(tn)}

of n > d distinct points x(4), with t 1 < t2 ... < t n , on the moment
curve. We will see from "Gale's evenness condition" ahead that the points
x(t i ) are vertices, and the combinatorial equivalence class of the polytope
does not depend on the specific choice of the parameters ti. This justifies
denoting the polytope by Cd(n) and calling it "the" cyclic d-polytope with
n vertices. Our drawing shows C3(6).
t
x(t 3 )
x(t4)
The problem is that in dimension 3 we cannot really see why cyclic

polytopes are so interesting. They are. Before we prove a few things about
them, let's do some "experiments."
We use the program "PORTA" by Thomas Christof [143, 144], which
produces a complete system of facet-defining inequalities from the list of
vertices. Let's do the 4-dimensional cyclic polytope C 4 (8). We use parameters ti = i 1 for 1 <j < 8. The input file for PORTA is
DIM = 4
CONV_SECTION
00
0
0
11
1
1
2 4
8
16
3 9 27
81
4 16 64 256
5 25 125 625
6 36 216 1296
7 49 343 2401
END
12
The output of PORTA yields (after 0.11 seconds of computation time) a

complete minimal system of inequalities for the convex hull of these points,
namely
DIM = 4
VALID
7 49 343 2401
INEQUALITIES_SECTION
( 1) -210x1+107x2-18x3+x4
( 2) -140x1+ 83x2-16x3+x4
( 3) - 84x1+ 61x2-14x3+x4
( 4) - 42x1+ 41x2-12x3+x4
( 5) - 14x1+ 23x2-10x3+x4
( 6) + 6x1- 11x2+ 6x3-x4
( 7) + 12x1- 19x2+ 8x3-x4
( 8) + 20x1- 29x2+10x3-x4
( 9) + 30x1- 41x2+12x3-x4
( 10) + 42x1- 55x2+14x3-x4
( 11) + 50x1- 35x2+10x3-x4
( 12)
+ 78x1- 49x2+12x3-x4
( 13)
+112x1- 65x2+14x3-x4
( 14)
+152x1- 83x2+16x3-x4
( 15)
+154x1- 71x2+14x3-x4
( 16)
+216x1- 91x2+16x3-x4
( 17)
+288x1-113x2+18x3-x4
( 18)
+342x1-119x2+18x3-x4
( 19)
+450x1-145x2+20x3-x4
( 20)
+638x1-179x2+22x3-x4
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
<=
0
0
0
0
0
0
0
0
0
0
24
40
60
84
120
180
252
360
504
840
END
In particular, this polytope has 20 facets.

The "-y" option of the PORTA program produces also the vertex-facet
incidence matrix given on the next page, from which we can derive the
complete combinatorial structure of the polytope.
In this matrix, the vertex-facet incidences are denoted by *'s. From the
matrix we can determine that C4 (8) is simplicial, since every facet has
exactly 4 vertices, corresponding to exactly 4 *'s in every row this is
also recorded in the last column. We also see that every vertex is on exactly
10 facets: there are 10 *'s in every column; see the bottom row of the matrix.
From the rows of the matrix we can observe the following pattern, known
as Gale's evenness condition: every segment of consecutive *'s is of even
length if it is not an initial or a final segment, that is, if it is preceded and
followed by a dot. (For this, the vertices of Cd(n) are labeled 1, ... , n, with
i corresponding to x(ti).)
N\N
Ii
E\T
Q \ S
13
S\
\I
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
#
*...*
*..**
*.**.
***..
****.
**.**
**..*
*4....
**...
.****
.**.*
.**.
.**..
..***
..**.
.**.
..**
..**
..*
***
*.*
..*
..*
*
*..
**,
.**
...
*.,
**,
.**
*..
**.
.**
**.
.**
***
:
:
:
:
:
:
:
:
:
:
:
:
:
:
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
11111 111
00000 000
From this pattern, one can derive that any two vertices of the polytope
are adjacent. We can also check this directly: every pair of vertices is contained in at least 3 facets. So, the edge 12 is contained in the facets (5) =
1238, (6) = 1234, (7) = 1245, (8) = 1256, (9) = 1267, and (10) = 1278.
Similarly, the edge [1,3} is contained in the facets (4) = 1348, (5) = 1238
and (6) = 1234.
Finally, we can note that there is a combinatorial symmetry that sends
vertex i to vertex 9 i; see Exercise 0 1 .
The following theorem and corollary contain a complete description of
the combinatorial structure of the cyclic polytopes as suggested by our
computation. Here we break our promise not to do any proofs in this introduction: mainly because the proofs are fun, and the results are a little
surprising (see Corollary 0.80.
14
0.
Theorem 0.7 (Gale's evenness condition).

(Gale [206])
Let n> d > 2. We will use [n] to denote the set {1, , n} , and choose real
parameters t1 < t2 < < tn
The cyclic polytope
Cd(n) = convfx(t i ), .
is a simplicial d-polytope. A d-subset S C [n] forms a facet of Cd(n) if and
only if the following "evenness condition" is satisfied:
If i <j are not in S, then the number of k E S between i and j is even:
2 I Iffk : k S, < k < jj
for i, j V S.
Proof. Recall the famous Vandermonde determinant identity

det
1
1
x(t o) x( t i)
i
to
...
x(td)
1
td
1
ti
det
b0
tg
4c1 i
0<i<3<d
bi
td
This is easily proved by observing that both sides are polynomials and that
the determinants vanish whenever we have ti = ti for some i
j. From
the identity, one sees that no d +1 points on the moment curve are affinely
dependent. In particular, this shows that Cd (n) is a simplicial d-polytope.
Now let S = {i 1 ,. , id } C [n]. Then the hyperplane Hs through the
corresponding points x(t 8 ) is given by
HS
E Rd : Fs(x) = 01,
where
Fs (x) := det
xi"
1
x(t 1 )
...
x(t)
\
)
In fact, Fs (x) is a linear function in x, and it vanishes on the prescribed

points.
(The reader should check, at least for one or two examples, that the
inequalities that we have computed for C4(8) have the form "Fs(x) < rs"
for some Ts c
Now we let the point x(t) move on the moment curve {x(t) : t c
Note that Fs (x(t)) is a polynomial in t of degree d. It vanishes for t =
15
thus it has d different zeroes, and changes the sign at each of them. The
following sketch is supposed to illustrate this.
Now S forms a facet if and only if Fs(x(t i )) has the same sign for all the
points x(t i ) with i C [n]\S; that is, if Fs(x(t)) has an even number of sign
changes between t = ti and t = ti , for i < j and ti, j E [n]\S
LI
In particular, this criterion shows that the combinatorics of Cd(ti, - , tn)
do not depend on the specific choice of the parameters t i , so Cd (n) is well
defined as a combinatorial equivalence class of polytopes.
It is quite easy to extend the evenness condition to a characterization of
all the faces of Cd (n). This characterization then also shows the following
corollary (Exercise 0.8), for which we give an independent proof.
Corollary 0.8. The cyclic polytope Cd(n) is [ ] -neighborly, that is, any
subset S C [n] of IS! < vertices forms a face.
Proof. Let Cd(n) = Cd(ti,... ,tn ) with t 1 <

< tn , and let T =
Choose some E > 0 small enough
- - , ik} C [n] have cardinality k <
such that t i < t1 + e < ti+1 for all j < n, and some M> t1 + E.
Using x(M+1),x(M+1), ... as dummy points "far out there," we define
a linear function FT(x) as
4.
det (x,x(t ii ), x
+e), . . , x(t), x
+e), x(M +1),
, x(M +d 2k)) .
This is a linear function in x, which vanishes on the points x(t 1 ) for I E T.

If we consider FT (x(t)), then this is a polynomial in t of degree d, and has
d "obvious" distinct zeroes
tZi E7
tjk tjk E7M 1 7.
,M+d-2k.
There is an even number of zeroes between t = t i and t = ti for I j E [71,1\T,

because a zero at t = ti always comes in a pair with a zero at t = t1 + E.
Thus FT (x) has the same sign on all the points x(t i ) : j E [n]\T.
,
For d < 3 Corollary 0.8 just says that the points x(t i ) form vertices of
Cd(n): the points on the moment curve are in convex position. However,
16
for d > 4 Corollary 0.8 yields something "counterintuitive": it describes

a property that does not manifest itself in d < 3 dimensions. Namely, for
d > 4 the polytope Cd(n) has n pairwise adjacent vertices, where n may
be much larger than d.
More generally, one defines a d-polytope to be k neighborly if any subset
of k or less vertices is the vertex set of a face of P. In Exercise 0.10, we see
that, except for simplices, no polytope is more than [C-neighborly. Therefore, polytopes that are [C-neighborly are known as neighborly polytopes.
Thus, by Corollary 0.8, cyclic polytopes are neighborly.
The neighborly polytopes are the solution of various extremal properties.
This is one reason why they are important. For example, the famous upper
bound theorem of McMullen (which we will state and prove in Section 8.4)
implies that among all d-polytopes with n vertices, the neighborly ones
have the greatest number of facets. In particular, no d-polytope with n
vertices has more facets than the cyclic polytope C, (d).
-
Example 0.9. If we apply an affine map 7i to a polytope P, then we get a

new polytope r(P): this is quite obvious from the definition of a V-polytope
in Definition 0.1. If the affine map is injective, then the image polytope
r(P) is (affinely) isomorphic to the original one nothing interesting has
happened.
However, one can also take affine maps that project P to a polytope
f(P) of lower dimension.
In particular, the convex hull

conv{x i , ... , x} c
Rd
can be interpreted as the image of the standard simplex A1 _ 1 C iitn , under

d mapping ei 1--> xi. This is usually inthe linear map 7r : Rn -terpreted geometrically as a projection of polytopes (which suggests some
special choice of coordinates, where R d is embedded as a subspace of Ra ).
We conclude that a (V-)polytope is the same thing as the projection of a
simplex, and that every projection of a polytope is a polytope as well.
,
17
A polytope P C Rd is centrally symmetric if it has a center a point

zo E Illd such that xo + x E P holds if and only if x0
xEP. Every
affine image (projection) of a crosspolytope is centrally symmetric: if P =
{Ax +zo : z E Gd}, then P is centrally symmetric with respect to z o . In
fact, every centrally symmetric polytope is the projection of a crosspolytope
(Exercise 0.2).
The projections of cubes, called zonotopes, form an especially interesting class of polytopes. For example, they encode the structure of linear
hyperplane arrangements; see Lecture 8.
Example 0.10. The perrnutahedron Ild_ i c
d is defined as the convex

hull of all vectors that are obtained by permuting the coordinates of the
1\
(2
vector . . It was apparently first investigated by Schoute [445] in 1911:
d
we have taken the following drawing from his paper [445, Fig. 4].
The permutahedron is a very interesting polytope. In fact, it is a simple

zonotope (Exercise 0.3), which is rare. Its vertices can be identified with the
/xi\
X2
permutations in Sd (namely, by associating with
the permutation
x,:t )
that maps x i) in such a way that two vertices are connected by an
edge if and only if the corresponding permutations differ by an adjacent
transposition. Check this in our drawing of 112:
132
123
312
213
321
231
18
and that of 113 :
2143
4312
2134
3412
3214
3241
There is a simple combinatorial description of all the faces of Ild_ i : its

k-faces correspond to ordered partitions of the set [d] into d k nonempty
parts. Thus the vertices are permutations, and the facets are partitions
of [d] into parts (S,[d]\S) with 0 c S c [d].
The permutahedron is a classical object; see [92, Example 2.2.5] for further references. We'll meet it again as a zonotope in Section 7.3, and as a
fiber polytope (the monotone path polytope of the cube) in Section 9.2.
There is a much more recent counterpart, the associahedron IC,-,_2, first
described as a combinatorial object by Stasheff [485] in 1964, and constructed as a convex polytope by John Milnor (unpublished, unrecorded),
by Mark Haiman [245], and by Carl Lee [327]. The vertices of this (simple)
polytope correspond to all the
) different ways of bracketing a string
of n-letters, that is, of multiplying an expression a 1 a2 ... an when multiplication is not associative. Two vertices are adjacent if they correspond
to a single application of the associative law. Our figure depicts the 5-gon,
which we get as Kn_2 for n = 4:
*( * ( ** ) )
(*(**))*
* ((* *)* )
O. Introduction and Examples
19
Whereas the first constructions of the associahedra were very much "ad
hoc," in Lecture 9 we will get an associahedron from a very natural construction due to Gel'fand, Zelevinsky Si Kapranov [213, 214]; as the "secondary polytope" of the n-gon [213, Rem. 7c)]. More generally, we will construct "fiber polytopes" there, a concept due to Billera Si Sturmfels [74, 75].
Recently, Mikhail M. Kapranov [286] constructed a new combinatorial
object KfI i , the perrnuto-associahedron, which combines the permutahedron and the associahedron. (Kapranov denotes it "KP,i ".) Its vertices
correspond to the different ways of multiplying n terms a l , a2, ... , an, in
arbitrary order, assuming that multiplication is neither commutative nor
associative and again there is a natural way to describe all the faces.
Our drawing shows KII 2 , a 12-gon.
1(3 2)
(1 3)2
2(3 1)
(2 3)1
Kapranov [286] showed that the combinatorially defined object Klin-i

can for every n> 2 be realized by a cell complex that is a topological ball.
The question of whether the permuto-associahedron (or "Kapranotope")
can be realized as a convex polytope was answered in joint work with Vic
Reiner [419] while I was first giving this course; see Section 9.3.
Example 0.11. A class of very interesting polytopes appears in combinatorial optimization: 011-polytopes, all of whose vertex coordinates are 0
or 1 (cf. Schrijver [448]). In other words, a 0/1-polytope is the convex hull
of a subset of the vertices of a (unit) cube.
Note that the d-simplex Ad-1 C d is a 0/1-polytope. Similarly, one can
study the hypersimplex Ad_1(k) in Id, by
d
Adi(k)
..
conv{v E 10,11 d :
E v, , k}
d
-= {X
.0<
xi
< 1 for 1 < i < d, y:xi , 0
i=i
for 1 < k < d - 1.
20
This family includes the standard simplex as A d Ad_ 1 (1). The

hypersimplex di (k) has (dk ) vertices, and 2d facets, if 2 < k < d 2
(but only d facets for k = 1 or k = d 1), For example, the 3-dimensional
hypersimplex A3 (2) C R4 is combinatorially equivalent to an octahedron.
It seems that the hypersimplices first appeared in Gabrilov, Gel'fand
Losik [203, Sect. 1.6] in the theory of characteristic classes. See also
Gel'fand, Goresky, MacPherson Sz Serganova [211], and Exercise 5.3(i).
These interesting polytopes certainly deserve more study!
Example 0.12. A very "classical" class of 0/1-polytopes (introduced by

Birkhoff [79] in 1946) arises from the following construction. Let Sd denote
the set of all permutations of the set [d]. With every permutation 6 in Sd,
we associate the matrix X', given by
1 if a(i) = j,
0 otherwise.
Xf:7
The matrices X' are the 0/1-matrices with exactly one 1 per row and per
column. If we identify Rd2 with the set of all real (dxd)-matrices, then
the matrices X' are 0/1-vectors in Rd ", and their convex hull forms a
0/ 1-polytope
P(d)
cony{ X' : o- E Sd} C Rd2
This is an interesting polytope with many names: the Birkhoff polytope, the
perfect matching polytope of Kn the assignment polytope, the polytope of
doubly stochastic matrices, and so forth.
The polytope P(d) has d! vertices (by construction), d 2 facets, and dimension (d 1) 2 . In fact, a complete linear description is given by
P(d) = {X
dd.
xii > 0, for 1 < j < d,

d
Xik = 1 for 1 <j <
d,
k=1
d
Yj Xki = 1 for 1 < j <

k=1
d.}
This is not hard to prove (just do it!): it is a classical result due to

Birkhoff [79] and von Neumann [391] independently. (See the treatment
by Lovgsz SE Plummer [343].) With this the Birkhoff polytopes are "well
described" that is, we know all the vertices and all the facets. Among
many other interesting properties, we note here that P(d) has a canonical
center point, given by xii = for all i and j.
Brualdi and Gibson undertook a detailed study of the Birkhoff polytopes
in a series of four papers [129]. Still, there are questions left
O Introduction and Examples
21
Example 0.13. For a class of nastier 0/1-polytopes, consider the famous

traveling salesman problem [322], which asks for the shortest possible tour
through a complete graph KT, on n vertices, where every edge has a length
given. For example, in the graph drawn here (n = 6), the length is given
by Euclidean distance, and the shortest tour is shown in thick lines.
Every traveling salesman tour can be considered as a subset of n edges,

T C E(K r,), of the graph. We associate with every tour T its "characteristic vector" X, E 10, 11 (n2) C lik(;), that is, the 0/1-vector whose entries
indicate which edges are in T, and which are not. Now the traveling salesman polytope QT(n) is defined as
n\
QT (n) := convfx, E {0, 1 } ( " : ,c, is a tour through Ifn 1.
It is not hard to see that QT (n) is a polytope of dimension

n=
n(rt 3)/2. We know the vertices of QT(n): they are the (n-1)!/2 different
Hamilton tours through K. Now the question for the shortest tour is
answered if we find a vertex that minimizes a linear function: thus the
traveling salesman problem is a linear programming problem over QT(n).
Similarly, one can define the polytopes Q(n) corresponding to the asymmetric traveling salesman problem, which seeks to find the shortest possible
directed tour through a complete directed graph K'i, on n vertices, where
each of the n(n 1) arcs has a given length. The corresponding polytope
2
Q'T (n) C Er ' has dimension n2 3n+ 1 (for ri > 3), and (n 1)! vertices.
To illustrate that these polytopes are nasty, we just mention the recent
result of Billera Sz Sarangarajan [72] that every 0/1-polytope is isomorphic
to a face of QT,
f (n), for large enough n. A little trick of Karp [290] [269] shows
that (an isomorphic copy of) the asymmetric travelling salesman polytope
Q'T (n) appears as a face of the symmetric travelling salesman polytope
QT (2n). Thus, the result of Billera Sz Sarangarajan [72] also applies to the
symmetric TSP polytope.
Using linear programming, we could solve the traveling salesman problems efficiently, if we could deal with two major obstacles: we do not know
the facet-defining inequalities of QT (n), respectively of C2IT (n), and there
are simply too many of them.
In the next lecture we will describe a general method for finding the
facets of a polytope given in the form Q = conv(V). It is the method that
makes the PORTA program work. It has successfully been applied to get
complete descriptions of the traveling salesman polytopes up to QT (8) and
(n
2)
22
CfT (6); see Exercises 0.14 and 1.1(iv). However, it seems that the method
does not go beyond that: in general the algorithmic determination of all the
facets of Q is certainly much harder and more strenuous than examining
all the vertices of Q.
The problem of finding some of the facets, by using the combinatorial
properties of the traveling salesman problem, is a central problem for a
whole branch of mathematics, called "polyhedral combinatorics" see
Gr6tschel & Padberg [229] and Jnger, Reinelt & Rinaldi [269] for solid
introductions, including detailed information about the structure of the
polytopes QT(n) and Q(n).
Notes
The principal historical "classics" in the theory of polytopes are the 1852
treatment by Schlifli [437] published in 1901, the books by Bruckner [131]
(1900), Schoute [444] (1905), and Sommerville [470] (1929), and the volume by Steinitz & Rademacher [490] (1934) about 3-dimensional polytopes.
(A very helpful bibliography is Sommerville [471].) The modern theory of
polytopes was established by Griinbaum's 1967 book [234]. It should be
stressed that not only did Griinbaum present the major part of what was
known at the time, but his book also contains various pieces of progress
and substantial original contributions, and has been an inspiring source
of problems, ideas, and references to everyone working on polytopes since
then.
There are more recent books and surveys on polytopes. Many of them
concentrate on aspects related to the upper and lower bound theorems and
the g-theorem (among them McMullen & Shephard [374], Brondsted [126],
Stanley [478], and Hibi [252]) and on the various methods of f-vector theory; see Lecture 8. Other aspects are treated in Barnette's exposition on
3-polytopes [41], Schrijver's book on optimization [448], and the handbook
chapters by Kleinschmidt & Klee [301], and Bayer & Lee [591. Also, the
reader might find Pach's volume [401] inspiring.
In our lectures we avoid any larger discussion of general convex sets and
bodies, as well as of most of the convex-geometric aspects of polytopes.
We refer to Bonnesen & Fenchel [117], Schneider [440], and Ewald [189],
the point is that for a convex polytope, we can describe and discuss
everything in terms of vertices, edges, facets, etc. (i.e., a finite collection
of combinatorial data) and bypass the apparatus of support functionals,
nearest point maps, distances, volume, and integration, etc. Correspondingly, in this book we disregard all metric properties of polytopes, such as
volume, surface area, and width, which are part of a very interesting theory
of their own.
23
Also, we disregard all those questions related to integral points in convex

bodies this leads to the beautiful theory that was named the "geometry
of numbers" by its founder, Hermann Minkowski [377]. Modern treatments
are Cassels [136] and Gruber & Lekkerkerker [232]. The algorithmic questions are treated in Kannan [283], Lagarias [318], and Schrijver [448]. See
also ErdeSs, Gruber & Hammer [187] for a nice "problem-oriented" survey.
Furthermore, we do not have the time or space to treat much more of
the aspects of linear and integral optimization related to convex polytopes.
Besides traveling salesman polytopes, many other classes have been studied extensively. It seems that cut polytopes are especially important for
practical applications see Deza & Laurent [175].
The necessity to optimize over polytopes with only partial information
about their facets leads to "cutting plane algorithms": the books by Schrijver [448] and by Griitschel, LovAsz Agi Schrijver [228] explain the powerful
theory behind this. Two recent references that describe the method for
"how to find a good solution for a Traveling Salesman Problem if you
really need one" are Reinelt [418] and Ringer, Reinelt & Thienel [270]. The
"New York Times" and "New Scientist" articles [393] [312], and the survey
by Greitschel & Padberg [230], are references for the spectacular success
of the method on extremely large traveling salesman problems. The latest
success in the race for the "TSP Olympics" (i.e., for "largest traveling
salesman problem ever solved") is reported in [1]: David Applegate, Bob
Bixby, Va:sek ChvAtal and Bill Cook [21] have been able to solve a 7397city instance to optimality, using a polyhedral approach, LP-relaxations, a
branch&cut framework, very clever heuristics, superior programming, and
an array of powerful UNIX workstations.

0.0 Given a 3-dimensional polytope such that every two vertices are adjacent, show that it is a tetrahedron.
0.1 Show that if a polytope is both simple and simplicial, then it is a
simplex or an n-gon.
Similarly, if a d-polytope is simple and cubical (i.e., all its facets are
combinatorially equivalent to (d *cubes), then P is a d-cube or
an n-gon.
0.2 Prove that a polytope can be represented as the affine image of a

crosspolytope if and only if it is centrally symmetric.
Show that if P is a zonotope (an affine image of a d-cube), then every
face of P is centrally symmetric as well. What about the converse?
24
0.3 Show that the permutahedron Ild_ i C Rd (Example 0.10) has dimension d - 1, that it is a zonotope, and that it is simple.
Describe its 2d 2 facets, by constructing inequalities that determine
them.
0.4 Let a l > a2 > ... > ad be real numbers, not all equal. The generalized permutahedron (or orbit polytope) lid- i (ai, . . . , ad ) is the convex
hull of all the vectors given by all the permutations of the multiset
{a l , ... ,ad}.
Investigate the combinatorics of the generalized permutahedra. In
particular, show that their dimension is d - 1. Are they all simple?
(They are not.)
Under what conditions do all the edges of Hd_1 (a l , ... , an ) have the
same length? (Schoute [445, p. 5])
0.5 Let P = Cd C 1. d be the d-cube. Enumerate the 3d 1 faces of Cd,
and show that the nonempty faces are naturally associated with the
sign vectors in {+, -, O } d .
Given a linear function c E (R)*, how can one find a vertex that
maximizes c over P ("optimization problem")?
Given y E d , how do we tell whether y E P? If y V P, how can we
find an inequality that is valid for P but is violated by y ("separation
problem")?
For which other classes of polytopes discussed in Lecture 0 can you
easily solve these problems?
0.6 Describe Cd(d + 2), the cyclic d-polytopes with d + 2 vertices, combinatorially and explicitly.
Is the 2-neighborly polytope (A2 x A2) constructed in Example 0.5
combinatorially equivalent to C4 (6)?
0.7 Consider the cyclic polytope Cd (n) = conv{x(0), x(2), ... , x(n-1)}.
Show that there is an affine symmetry (an affine reflection) which
induces the symmetry i 4----+ n -1- 1 - i (that is, x(i - 1) 4 x(n -i)),
and thus the corresponding combinatorial symmetry of Cd(n).
0.8 From Gale's evenness condition, given in Theorem 0.7, derive a complete combinatorial description of all the faces of Cd(n).
From this, derive that the cyclic polytopes are LC-neighborly (Corollary 0.8).
25
0.9 Show (bijectively) that the number of ways in which 2k elements can
be chosen from [72] in "even blocks of adjacent elements" is (n-k k ).
Thus, derive from Gale's evenness condition that the formula for the
number of facets of Cd (n) is
41\
fd-i(Cd(n))
(n - 1 1- % 1 1
where H is the round-up function, with gi = k

_I. Here the first
term corresponds to the facets for which the first block is even, and
the second term corresponds to the cases where the first block is odd.
Deduce
-
fd-i(Cd(n))
{ nn
(n-k
k -k)
2 (n- kk-1)
for d = 2k even,
for d= 2k + 1 odd.
How many facets do the cyclic polytopes C10(20), C10(100), and

C50(100) have, approximately?
0.10 Show that if a polytope is k-neighborly, then every (2k-1)-face is a
simplex. Conclude that if a d-polytope is
1)-neighborly, then
it is a simplex.
(iti +
0.11* Is there a fast and simple way to decide whether a certain point
G Rd (with rational coordinates, say) is contained in the cyclic
polytope Cd(1, 2, .. . , n)?
(General theory namely the polynomial equivalence of optimization and separation according to Griitschel, Lovisz & Schrij ver [228]
implies that there is a polynomial algorithm for this task, since optimization over Cd(n) is easy, by comparing the vertices. However, we
ask for a simple combinatorial test, not using the ellipsoid method )
0.12 Prove the claims in Example 0.12 about the Birkhoff polytope P(d):
in particular, show that the dimension is (d- 1) 2 , and that the number
of facets is d2 .
The Birkhoff polytope P(d) and the permutahedron 11d_1 are closely
related: show that there is a canonical projection map P(d)
0.13 Draw the 3-dimensional associahedron K3. Justify the general for()
2n-2
mula '
,-2n n-1 for the number of vertices ofK_2.
0.14 Describe the combinatorial structure of the traveling salesman polytopes QT (3), QT(4), and QT(5). How many vertices and facets do
they have? Which vertices are adjacent? Are they simple, or simplicial? Similarly, try to describe Q'T (2), (3), and CA, (4).
26
0.15* What is the maximal number f(d) of facets of a d-dimensional 0/1polytope? How fast does f (d) grow asymptotically?
(It is not hard to see that
2d < f (d) < d! + 2d.

The upper bound was suggested by Imre BAr6ny, via the following
observation: if we add the "missing vertices" of the d-cube to the
polytope one by one, then we add a volume of at least 1/d! for every
facet that is destroyed of the original 0/1-polytope. The process will
stop with the d-cube of volume 1, with only 2d facets.
More recent progress on this problem is recorded in Kortenkamp,
Richter-Gebert, Sarangarajan Sz Ziegler [315]. Still, there is a huge
gap between the lower and the upper bounds. We know
f(1) ---,-- 2, f(2) = 4, f (3) = 8, 1(4) = 16,

121 < f(6) <610,
etc.
f(5) = 40,
for small dimensions. (See Aichholzer [5] for enumeration techniques.)
Asymptotically the best known bounds are
6.4 d!
(3.6) d < f(d)
Nid '
for all large enough d. Here the upper bound is due to Rote [431],
while the lower bound is from explicit computation of "random 0/1polytopes" in low dimensions in combination with a "free sum" construction for 0/1-polytopes from [315]. The value 3.6 was achieved
in March 1997 by Thomas Christof for a random 0/1-polytope (of
dimension 13, with 254 vertices and at least 17,464,356 facets), using
his PORTA code and new ideas described in Christof igi Reinelt [148 ] .
For "current records" in the "Olympic race" for 0/1-polytopes with
many facets see [314] on the Web.)
0.16 Via the construction of "characteristic vectors" from Example 0.13,

show that the vertices of the Birkhoff polytope P(n) correspond to
the perfect matchings in the complete bipartite graph
(
0 1 0 0)
0 0 0 1
1 0 0 0
0 0 1 0
0.17 Show that the Birkhoff polytope P(d) C Rd2 contains the asymmetric
traveling salesman polytope QT (d) C d2 d C d2 . Does every facet
of P(d) yield a facet of QT(d)?
(For a detailed investigation, see Billera & Sarangarajan [72 ] .)
1
Polytopes, Polyhedra, and Cones
In this lecture we prove some fundamental properties, in particular the

equivalence of the two definitions of polytopes in Definition 0.1.
Of course, one could ask whether it is really necessary to go through these
details, since the result is quite obvious anyway, and complete proofs are
in the books [126] [234] [374] [448]. There are several good reasons. One is
that we can give proofs that introduce important machinery (like FourierMotzkin elimination), which is useful for other purposes as well. It also
yields a basic algorithmic tool to deal with polytopes. Additionally, these
proofs provide geometric intuition, which we will need later. We will also see
polarity appear in this context quite naturally, because we do two versions
of Fourier-Motzkin, which are related by polarity The "usual" approach
is to do only one version, and prove the second half using polarity this
saves some work, but avoids the very interesting polar version. Finally, our
proofs are (meant to be) easy and transparent, following simple geometric
ideas through some elementary linear algebra, so they might even be fun.
(There should be no crying in this lecture.)
1.1
The "Main Theorem"
However, to make sure that the pain level does not go below zero, we
start with a few definitions. In the following, we work with two versions
of polyhedra in the course of this lecture we will see that they are
mathematically (but not algorithmically!) equivalent. The two concepts
28
have also proved to be fundamental in a new field called "computational

convexity"; see Gritzmann & Klee [224, 225].
The first concept, an fl-polyhedron, denotes an intersection of closed
halfspaces: a set P C rd presented in the form
P = P(A , z)
Ix E
d : Ax < z1
for some A E
mxd 7 Z
R.
(Here "Ax < z" is the usual shorthand for a system of inequalities, namely
ai x < z1 , . . , am x < zrn, where al . . am are the rows of A, and
Z1 ,... zrn, are the components of z.)
For the second version we need the notion of a cone: a nonempty set of
vectors C C Rd that with any finite set of vectors also contains all their
linear combinations with nonnegative coefficients. In particular, every cone
contains O. For an arbitrary subset Y. C j . d , we define its conical hull (or
positive hula cone(Y) as the intersection of all cones in I d that contain Y.
Clearly C := cone(Y) is a cone for every Y. Similar to the situation for
convex hulls (Lecture 0), one can easily see that
,
cone(Y) =
+ AkYk {Yi, - 'Mc} Y,
>
01.
In the case where Y = {y 1 , ... yn } C i is a finite set this is the only

case we will need here this reduces to
cone(Y)
Itiy i +
+ tny,,
>O } = {Yt : t > 0}.
We define that cone(Y) = {0} if Y is the empty set, i.e., if n = O.

The vector sum (or Minkowski sum) of two sets P,Q C rd is defined to
be
:=-- fx+y:xEP, y EQ1.

The following sketch shows the 2-dimensional Minkowski sum of a cone and
a polytope.
1.1 The "Main Theorem"
29
Now we define a V-polyhedron to denote any finitely generated convexconical combination: a set P C Rd that is given in the form
P = conv(V) + cone(Y)
for some V C Rd ', y E
dxn'
)
as the Minkowski sum of a convex hull of a finite point set and the cone
generated by a finite set of vectors.
Thus, comparing this to Definition 0.1, we get that a V-polytope is a Vpolyhedron that is bounded, that is, contains no ray { u + tv : t > 0} with
y 0. For this we only need to observe that conv(V) is always bounded.
This follows from a trivial computation: if z E conv(V), then
min{vik : 1 <j < n} < xk < max{vik : 1 <j < n},
which encloses conv(V) in a bounded box. Similarly, an 7-1-polytope is the
same thing as a bounded fl-polyhedron.
Now we start with a basic version of the " representation theorem for
polytopes," which will be considerably strengthened and generalized in the
course of the proofs. See Section 2.4 for a definitive version.
Theorem 1.1 (Main theorem for polytopes).
A subset P c Rd is the convex hull of a finite point set (a V-polytope)
P = cony (V)
for some V ER d 'n
if and only if it is a bounded intersection of halfspaces (an 7-l-polytope)
P = P (A , z)
for some A E Rni xd , z c
This result contains two implications, which are equally "geometrically

clear" and nontrivial to prove, and which in a certain sense are equivalent.
Why is this theorem important? It provides two independent characterizations of polytopes that are of different power, depending on the problem
we are studying. For example, consider the following four statements.
Every intersection of a polytope with an affine subspace is a polytope.
Every intersection of a polytope with a polyhedron is a polytope.
The Minkowski sum of two polytopes is a polytope.

Every projection of a polytope is a polytope.
The first two statements are trivial for a polytope presented in the form
P = P (A , z) (where the first is a special case of the second), but both are
nontrivial for the convex hull of a finite set of points. Similarly the last two
statements are easy to see for the convex hull of a finite point set, but axe
nontrivial for bounded intersections of halfspaces.
Theorem 1.1 is the version we really need, a very basic statement about
polytopes; however, it is not the most straightforward version to prove.
Therefore we generalize it to a theorem about polyhedra, due to Mot zkin
[384].
30
1. Polytopes, Polyhedra, and Cones
Theorem 1.2 (Main theorem for polyhedra).

A subset P C I d is a sum of a convex hull of a finite set of points plus a
conical combination of vectors (a V-polyhedron)
P = conv(V) + cone(Y)
for some V
Rd " Y G Rd'
7
if and only if is an intersection of closed halfspaces (an H-polyhedron)

for some A E
P = P(A,z)
mxd
First note that Theorem 1.1 follows from Theorem 1_2 we have already
seen that polytopes are bounded polyhedra, in both the V- and the 7-tversions.
Theorem 1.2 can be proved directly, and the geometric idea for this is
sketched in Section 1.2. However, fighting one's way through the formulas is
quite strenuous, mainly because the points in conv(V) + cone(Y) are hard
to manipulate. It turns out that it is much easier to "homogenize": we
pass from affine d-space to linear (d+ *space; for this, we adjoin an extra
coordinate (which we will take as the zeroeth coordinate in the following),
mapping the point x E llid to the vector ( xl ) E Rd+ 1 .
This reduces Theorem 1.2 to the special case where P is a cone, which
can be proved more easily.
Theorem 1.3 (Main theorem for cones).
A cone C C I d is a finitely generated combination of vectors
C = cone(Y)
for some Y E Rd "
if and only if it is a finite intersection of closed linear halfspaces
C = P(A,O)
for some A E Rn" d
We will prove Theorem 1.3 in Section 1.3. In the following we will usually refer to the polyhedral cones characterized by Theorem 1.3 simply as
"cones," because the objects we consider are clearly polyhedra. Note that
every cone C, by definition, contains the origin O.
Let us see here why Theorem 1.2 follows from Theorem 1.3 by homogenization. For this, we associate with every polyhedron P c I d a cone
C(P) C 110+ 1 , as follows.
If P = P(A, z) is an H-polyhedron, we define
C(P) :__
p( (liz (i)i ) , (g) ).
1 1 The "Main Theorem"
31
That is, if P is defined by the inequalities ax < z then C(P) is defined

by the inequalities zi xo + ax < 0, together with the inequality xo > O.
Clearly, C(P) is again an H-polyhedron in Rd+', and
,
P = Ix R d :
(xl)
E C(P)}.
Also, we see that if P = P(B,u) is an arbitrary fl-polyhedron in Rd+ 1 ,

then
Ix
If P
d:
1 E Pl is an fl-polyhedron as well.
conv(V) + cone(Y) is a V-polyhedron, we define
C(P) := cone ( 11 0
V Y)
Clearly, C(P) is again a V-polyhedron in
d+1 , and
P = fx ERd : ( xl ) E C(P)}.
Conversely, a simple computation shows that if C = cone(W) is any cone
in Rd+ I generated by vectors wi with wio > 0, then Ix c Rd
(1)
E C}
is a V-polyhedron.
Now, given any fl-polyhedron P, we can apply Theorem 1.3 to C(P),
to conclude that C(P) is a V-cone contained in f x E Etd-1-1 . x o > 01, so
P is a V-polyhedron as well. Conversely, if P is a V-polyhedron, then by
Theorem 1.3 the associated cone C(P) is an fl-polyhedron, and hence so
is P.
In both cases C(P) realizes the homogenization of P, which we will discuss in Section 1.4, once we have established the Farkas lemma. The geometric idea is depicted in the sketch above, which shows the cone in 3
associated with an affine polytope in R2 . If P is a polyhedron, then one
has to add the necessary "points at infinity" to P, to make sure that C(P)
is a (closed) polyhedron.
32
1.2 Fourier-Motzkin Elimination: An Affine Sketch

For a direct proof of Theorem 1.2, the idea is the following. We have to
see that every intersection of halfspaces (fl-polyhedron) like P(A, z) is
a convex-conical combination (V-polyhedron) like conv(V) + cone(Y), and
conversely every convex-conical combination is an intersection of halfspaces.
For the forward direction ("every V-polyhedron is an fl-polyhedron") we
note that every V-polyhedron
conv(V) + cone(Y) =
fx E l
d.
n ERn/ :X=Vt+ Yu, t> 0, u> o, nt = 11
itE
can be interpreted as the projection of a set

{
(X,
t, u) E
+nni : x = Vt + Yu, t> 0, u > 0, li t = 1}
that quite clearly is an 7-1-polyhedron. Thus it remains to show that

(I) any projection of an fl-polyhedron is again an fl-polyhedron.
This can be done by the Fourier-Motzkin elimination method: projecting

down one dimension at a time. We will discuss only the case of projections
along coordinate axes, which we need here; the general case can be reduced
to this by an affine coordinate transformation.
In this section we give a geometric sketch for the case of affine polyhedra:
the nice thing about it (as compared to the more elegant --- version
for cones) is that its idea and most of its complications can already be
illustrated in dimension 2, so we can provide pictures, and we will. However,
instead of doing formulas for this version we switch to the version for cones
and do the proofs there.
We start with an fl-polyhedron P = P(A,z) c Rd and assume that
we want to project to fx E Rd Xk = 01 .- Rd-1 along the xk-axis. The
projection of P c I d can be defined in great generality; we will only use
the cases of coordinate directions, where we use the notation
:
fx - xkek : x E PI
= fx E d : Xk = 0, 3yER:x+yekEPI.
proj k (P) :=--
for the projection of P in the direction of ek. The set proj k (P) is contained
in the hyperplane Ilk = fx C Rd : xk = 01. A closely related set is the
elimination
elimk (P)
fx - tek : x
C P, t C 1
{x Elie : 3yER:x+yekEP } .
Thus elimk (P) is the set of all points in Rd which project to proj k (P). In
particular, we get an isomorphism elimk (P) proj k (P) x R.
1.2 Fourier-Motzkin E li mination: An Affine Sketch
33
For an example, we use the following system of inequalities
(1)
xi
2x 1
(2)
4 x2
x2
2x
(3)
(4)
+
+
(5)
(6)
+ 2 xi
2x i
x2
+ 6 x2
(7)
6x 1
x2
xi
xi
<
<
9
4
<
<
<
11
<
17
<
X2
.
.-'
'-'
xi
Now assume that we fix some x1, and ask for the possible values of x2.
Then we see that inequality (4) requires x 1 < 4. All other inequalities can
be rewritten to give either an upper bound on x2 (if the coefficient of x2 is
positive), or lower bound (if the coefficient of x2 is negative). Furthermore,
there is a solution for x2 if and only if every upper bound for x2 derived
this way is larger than every lower bound.
The sketch on the next page shows the projection of the 2-polytope P to
proj 2 (P), by eliminating the x2-variable. Here elim2(P) is the infinite strip
(shaded) of all points that lie above or below proj2 (P).
Observe how the points of P on any vertical line (i.e., with fixed x 1 ) are
bounded from above and below by inequalities with positive, respectively
negative, coefficient ai2 . If there is no solution, then some upper bound is
smaller than some lower bound: that is, the combination of two inequalities,
one with positive and one with negative ai2, leads to a restriction for the
possible values of xl.
Also note that there is one redundant inequality in the original system:
this leads to the effect that the same lower bound on x 1 arises from several
different pairs of inequalities.
34
12
dim )(I ) )
P'12(P)
It is easy to formalize this
..r
2-dimensional description, and to generalize it
to arbitrary dimensions. The algebraic treatment rests on the following fact.

Consider the coefficients of xk in our system of inequalities, and assume
that aik > 0 and ai k < O. Then the respective inequalities can be rewritten
as
ax <
aikxk < aikxk
ax zi
and
ax <zj
( aik)xk ax k + ax zi.
Here the right-hand sides of the rewritten forms do not depend on xk, so the
first one yields an upper bound on xk, the second one a lower bound. The
combination of the two inequalities (multiplied by the positive coefficients
aik respectively ak) yields the condition
aik ai x a ik zi< (aik )ai x + (aik )zi

for "lower bound on xk below the upper bound," which is equivalent to the
"eliminated inequality"
(aka i+ (aik )ai )x < aikzi + (aik )zi .

This is a linear restriction on elimk(P). You can see that it is valid without
computation: it is a positive combination of two valid inequalities of the
original system. However, it is also important to see the geometric content
from which it arises.
Now if z satisfies all these eliminated inequalities, and also those inequalities of the original system which do not involve xk, then we can conversely
find an xk-coordinate that satisfies Ax < z, that is, we have found a complete description of the elimination elimk(P) by linear inequalities, and
thus proved the following theorem.
1.2 Fourier-Motzkin Elimination An Affine Sketch
Theorem 1.4 (Fourier-Motzkin elimination).

Let P =
z) C Rd be a polyhedron, with A E R' d and z E
choose k < d.
Construct the matrix Alk G
"
35
, and
whose rows are
the rows ai of A, for all i with azk = 0, and

the sums aka i + (aj k)ai for all j, j with aik > 0 and aik < O,
and let z/ k E WI' be the corresponding column vector with entries

z for all i with aik = 0, and
aik zi (a ik )zi for all i, j with a,k > 0 and aik < O.
Then elimk(P) = P(A/ k ,z/ k ) and
proj k (P) = P(AP zik ) n fx E Rk : x k =
In particular, this says that for P =
z), the projection proj k (P) is
again an fl-polyhedron. Iterating this, we obtain the forward direction of
Theorem 1.2.
The problem with this is that the formulas are messy, partly because we
have to deal with the right-hand sides zi separately. Those will miraculously
disappear in the homogeneous version; see the next section.
For the backward direction ("every 7-1-polyhedron is a V-polyhedron"),
we observe that every 71-polyhedron
z) = {x E Rd : Ax < z}
can be written as the intersection of a polyhedron (in fact, a cone)
Co(A) := (:) E Rd+in : Ax <w}

with an affine subspace
(z)
E Rd +m
1.7
The cone Co(A) is easily seen to be a V polyhedron: it can be written as

-
Co (A) = cone({
ei
Ae i
: 1 <j < cl} u {(0 : 1

e)
by decomposing
(
X)
E ixi I
(sign(x i ))
ei
Aei
M
i=1
AXii
ei
36
It remains to show that

(II) any intersection of a V-polyhedron with an affine subspace is a
V-polyhedron.
The method to prove this, sometimes called the double description method,
is dual to the method of Fourier-Motzkin elimination (see Motzkin, Raiffa,
Thompson & Thrall [386] and Dantzig & Eaves [167]). We restrict our
discussion to a very special case, namely the intersection of a polytope
conv(V) with a coordinate hyperplane .Hk = fx E d : Xk= 01, which can
then be iterated.
So, we are given a V-polytope P = conv(V) C Y1 , and we want to see
that P n fik is of the same form. In our sketch the black dots denote the
set V, and the bigger white dots denote the two vertices of the intersection
of conv(V) with the hyperplane
From this we get a little geometric intuition, which suggests that we can
write down
P n Hk = conv(V/k),
where ink is the matrix (set) of column vectors constructed as
Vii (Vjk)Vi
: Vik
V ik
t vi . Vik =
>
O, V3k G 0 } .
Vik Vjk
For this it is quite clear that we get P n Hk D conv(V/k), but for the
converse we have to work a little. We omit this ugly little computation
here since you'll see it in the next section: it comes out a little nicer in the
homogeneous form (Lemma 1.6). Anyway, this way "in principle" one can
give an explicit representation of z as a convex combination of vectors in
17 1k .
LI
This finishes the proof of Theorems 1.1-1.3.
One can do a similar argument for P = conv(V) + cone(Y). However,
the corresponding computations become extremely tedious they are too
ugly even to leave them as an exercise. This is why we homogenize and
switch to cones, where all difficulties disappear.
37

The main objective of this section is to prove Theorem 1.3. This again has
two parts.
For the "forward direction" of Theorem 1.3, let C = cone(Y) C Rd be a
V-cone. We can write it as
= {Yt E
=
t > 0}
d 3tER n :t>0, X= Ytl.
The set {(x, t) E R dn : t > 0, x = Yt} clearly is an 7-1-cone. Thus

C = cone(Y) can be written as the projection of this cone to the subspace
{(x,t) E cin : t = 0}. Again, this projection can be formed successively,
by projecting with respect to individual tk-coordinates one by one. Thus it
suffices to prove the following lemma
d then so is the elimiLemma 1.5. If C = P(A , 0) is an fl-cone in I,
nation elimk(C) = {x tek : X E C, t E R}, and thus also the projection
proj k (C) = elimk(C) n Bk. Namely, we get elimk (C) = P(Ai k , 0) for
A/ k := {a i : aik = 0 } U {aka i+ (ai k)ai : aik > 0, ai k <0}.

(Here we interpret A and AP' as sets of row vectors.)
Proof. The row vectors in Al k are positive combinations of row vectors

in A, hence the corresponding inequalities are valid for C, and thus we get
that C C P (AP', 13). Furthermore, the row vectors in Al k all have the kth
component equal to zero (by construction), i.e., the variable x k does not
appear in the system il/kx < 0, which proves that elimk(C) C P(AI k , 0).
For the converse, let x C P(Ai k , 0), and let x k = 0 (without loss of
generality). We claim that x
yek E C for suitable y. In fact, plugging
yek into the system Ax < 0, we find that y has to satisfy
1
(
<
y < min
min{{
a
3k
ma{--- a x : aik > ol
a k < 0}.
This can be satisfied. Namely, if aik > 0 and ail, < O, then we know that
< 1 ( a )x , which is equivalent to (aka i+ (aik)ai )x < 0,
aik
which holds because x E P(A/', 0).
Now we proceed to prove the "backward direction" of Theorem 1.3. For
this let C = P(A, 0) C Rd be an fl-cone. We can write it as
=
{X E
Rd : Ax < 0}
{( x ) E Rd+m : Ax < w}
n{
( x
E Rd-E rn : w = 0 } .
38
Here {(w
x ) E Rd+m : Ax < w} is a V-cone, as we have shown above.
x ) E Rdm : w = 0 } can be formed succesThe intersection with {(w
sively, by setting coordinates to zero one at a time, i.e., intersecting with
coordinate hyperplanes of the form Ilk = fy e yk = 01. Thus it
suffices to prove the following lemma.
Lemma 1.6. If C = cone(Y) is a V-cone in Rd , then so is the intersection

C n Hk. Namely, we get C n Hk =-7 cone(Y/ k ) for
Y ik := { yi: Yik = 0 } U {YikY3 + ( Y3k)Y2 Y2k > 0 , Yik < 0 }(Here we interpret Y and Y/ k as sets of column vectors.)
Proof. First note that the vectors in Yrn all have xk-coordinate 0, so
clearly C n HkD cone(Y/k).
For the reverse inclusion, we consider some y = Yt E cone(Y) > 0)
with vk = O. Now either we have tiYa = 0 for all i, in which case we get
V E cone({y, : yik = 01), or we can expand vk = 0, to get
A:=
t (y 3 k) > O.
tiYik
j
Yik >0
yik<0
With this, we can rewrite y as
tiyi +
iYjk>O
Yik=o
ti yi +
1
A
E
yik=o
ti yl
tiyi +
3 yik<0
t( y3k)
J tiy,
i yik>0 ( j y3 k<0
Y3k<0
vik>0
tiyik) t yi
titi
tiyi +
Y3k)Y2 Yikl/j )A ((
2
j Y3k<C)
This proves the claim, by giving an explicit representation of y as a conical

sum of vectors in Yi k .
LI
This completes the proof of Theorem 1.3. In this proof, we have provided
explicit projections and intersections one could also argue less explicitly
that coefficients exist.
For example, in the last proof the space of possible coefficients turns
out to be a transportation polytope, of which we have implicitly used the
special interior point; see Problem 1.8. An alternative approach is to use a
vertex, and thus to get fewer nonzero coefficients.
39
How efficient is Fourier-Motzkin elimination as a computational tool?

The main problem is that the number of inequalities generated by this
method can go beyond all tractable limits within a few elimination steps.
]
For this observe that if A has m rows, then AR may have as many as [12
rows: the number of inequalities can roughly be squared by every step,
which leads to problems even with a fast implementation on a computer
with a lot of memory.
Nevertheless the computations can be carried out fairly efficiently. See
the notes to this lecture for comments about the available programs.
Let us mention only that Fourier-Motzkin elimination can in principle be
used as an algorithm for linear programming (see Section 3.2). In fact, to
find a point in P(A, z) which maximizes cx, we introduce an extra variable
xo = cx, and eliminate all the other variables. This will tell us the possible
range of xo , and by backtracking one can recover the optimal basis (i.e., a
set of inequalities whose nonnegative combination yields an optimal upper
bound on xo). However, this method for linear programming is exponential,
and there are much better ones available.
1.4
The Farkas Lemma
It was first pointed out by Kuhn [317] that with (termination of) Fourier-
Motzkin elimination we have also done all the work for the Farkas lemma.
This extremely important lemma appears in many different versions all
over the theory of polytopes and polyhedra. It is interesting to note that if
you look into different books and papers, you find quite different lemmas
all called "the Farkas lemma." All of these, however, are easily transformed
into each other.
Essentially, the Farkas lemma yields a characterization for the solvability
of a system of inequalities. There are variations for systems of inequalities
in various standard forms: Farkas lemmas for polyhedra and for cones,
for inequality systems with equalities, inequalities or strict inequalities, in
nonnegative, positive or unrestricted variables, and so on. There are also
quite different ways to formulate theorems "of Farkas type":
as theorems of the alternative (one inequality system has a solution
if and only if a second system has none),
as transposition theorems (because the second system can be derived
by transposing the matrix and vectors of the first),
as duality theorems (the duality theorem for linear programs is of
Farkas type),
as good characterizations (if a system has a solution, then any solution
vector proves this; if it has no solution, then the Farkas lemma yields
a dual vector that encodes this fact),
40
1 . Polytopes, Polyhedra, and Cones
as certificates for validity (if an inequality is valid for the solution set
of a system, then it is a conical combination of the inequalities of the
system),
or, dual to this, as separation theorems (if a point does not lie in
a convex-conical hull, then it can be separated from it by a linear
functional).
We refer to Mangasarian [347] and to Stoer & Witzgall [492] for more
versions, extensions, and generalizations. Separation theorems "of Farkas
type" also hold for convex bodies. Infinite-dimensional versions are fundamental in functional analysis ("Hahn-Banach theorem").
Here is one basic version, characterizing the solvability of a general system of inequalities.
Proposition 1.7 (Farkas lemma I).
Let A E mxd and z c m .
Either there exists a point x E Rd with Ax < z,
or there exists a row vector c E ( m)* with c > 0, cA = 0 and cz < O,
but not both.
Proof. First observe that both conditions cannot hold at the same time:
otherwise there are a column vector x E Rd and a row vector c E
with
0 = Ox = (cA)x = c(Ax) < cz < 0,
which is a contradiction.
Now define P
P(A, z), and Q
P(( z A), o). We note that an
x E Rd exists with Ax < z if and only if Q contains a point with xo > 0.
Here Q is an 7-1-cone. Now we eliminate the variables xl,
xd from Q, to
get the 7-1-cone elim i elim2
elimd(Q)The key observation is that if we do Fourier-Motzkin elimination to get
elimi P(D, 0) = P(Dii, 0), then every inequality in the eliminated system
elimi (D) is a positive combination of at most two rows of D, so D/i can
be written as CID for a matrix Ci with only nonnegative entries, of which
at most two per row are nonzero.
Iterating this idea, we get
elimi elim2 .. elimd(Q) =

= P(C1 C2 Cd (z, A), 0) =:
P((z, A)/ d/d-1 / 2/ 1 , 0)

P(C(z, A), 0),
where C is a nonnegative matrix.

All the inequalities in the system C(z, A) < 0 are of the form Pyiaxo < 0,
since all variables other than xo have been eliminated.
E Rd+ 1 : xo < 01. By eliminaNow assume that P = 0, so that Q C
tion we get
elim1 elim2 elimd (Q) C
E Rd+1 : z0 < 01,
41
and thus the system C(z, A)x < 0 contains an inequality eyioxo <O with
'Yio > O. Let e be the row of C that yields this, then we have c(z, A) =
El
(7i0, 0), that is, ez --=
<O and cA = O.
Now comes another version of the Farkas lemma, for nonnegative solutions of systems of inequalities. Every such system can be rewritten as a
system of inequalities, and this is exactly what we do to prove it. In fact, the
proof nicely illustrates various simple-but-important techniques for rewriting systems of inequalities, such as the introduction of slack variables, and
rewriting unbounded variables as differences of nonnegative ones.
Proposition 1.8 (Farkas lemma II).
Let A E Rmxd and z E Rm.
Either there exists a point x E d with Ax = z, x > 0,
or there exists a row vector e E (Rm)* with cA > 0 and cz < O,
but not both.
Proof. We have the following equivalences:
< >
x Ax < z, (A)x < z, x < 0

( A
z
] x : A x < z
:
Id
FL I
< >
x: Ax = z, x > 0
A c i > 0, C2 > 0, b > 0 :

A )
(c i , c2 , b) A
(
= 0, (c i ,c2 ,b)
z
zo
<O
Id
e, >
<
>
< >
4 c i > 0, c2 > 0, b > O: (e 1 e2)A b =- 0,

A c = c i c2 , b > 0 : cA b = 0, cz < 0
4 c : cA > 0, cz < O.
(el
e2 )z <O
The following is my favorite version of the Farkas lemma. It says that if

an inequality is valid for a polyhedron, then either it can be obtained as a
positive combination of inequalities that define the polyhedron, or the polyhedron is empty, in which case the inequality Ox < 1 can be obtained as
a positive combination. This version of the Farkas lemma includes version I
as a special case: Ox < 1 is valid for all points x : Ax < z if and only if
Ax < z has no solution.
Proposition 1.9 (Farkas lemma III).
Let A E Rmxd, z E Rm, ao E (Rd)*, and zo E R.
Then aox < zo is valid for all x E d with Ax < z, if and only if
(i) there exists a row vector c > 0 such that cA = a() and cz <z0 , or
(ii) there exists a row vector c > 0 such that cA = 0 and cz < O,
or both.
42
Proof. The "if" part is easy to see: the existence of z with Ax <z and
aox > zo contradicts both (ii) (as in Farkas lemma I) and (i) (with a similar
computation).
For the "only if" part, assume that neither (i) nor (ii) is satisfied. Then
we conclude that there is no b> 0 and 0 > 0 with bA = Oa and bz < Ozo:
otherwise, (i) would be satisfied for c := lo b if fi> 0, or (ii) for c := b, if
0 O.
Thus we can apply Farkas lemma I to compute
Il (0, b) (0, 0) : (0, b) ( Aal - O ' (0,11) ( zz ) < 0

( -a()) w < (-zo \
]w
<FL >I
A
z)
]w E Rd :
Aw < z, maw > zo .
< >
Now we reformulate the condition that (i) does not hold, by introducing a
slack variable -y, and then we apply Farkas lemma II to a problem in dual
space:
-, (i) <--> /I (7, c) > (0, 0) :
7 cz = zo,
____> A (7, c) > (o, O): (-y, c)
ra
(1
c( A) = -ao
-
0 )
z -A
= (zo, -ao)
3 (7; ) G Rd-1-1 : ( 1 0 ) ( yo \ > ( 0 )
-0)'
(zo, -ao) (:) <o

4>
A Yo > 0, y E Rd :
Ay
< yoz, aoY >
Yozo-
Now either we have yo > 0, then we put x : = --o- y, and this satisfies Ax < z
and aox > zo , or we have yo = 0, then we use the w constructed before
(remember?) and put x : = w + y. This x satisfies Ax = Aw + Ay <
z + 0 = z and aox = aow + aoy > zo + 0 = zo.
-Ei
-
The following, fourth and last version (but see the exercises) shows that
the Farkas lemma can also be used to separate a point from a V-polyhedron:
if x is not contained in P := conv(V)cone(Y), then there is an inequality
ax < a satisfied by P, but not by x.
Proposition 1.10 (Farkas lemma IV).
Let V E Rdxn , Y E isdxn' and x E d
Either there exist t, u > 0 with 1t= I and x = Vt + Yu,
or there exists a row vector (a, a) E (Rd+ 1 )* with av i < a for all i <
ayi < O for all j < re, while ax > a,
but not both.
,
n,
1.5 Recession Cone and Homogenization
43
Proof. The "either" condition can be stated as

( 17
11
t) (lx )
y
O) ( u
which by version II of the Farkas lemma is equivalent to
ti (a,
a) E (1V+ 1 )* -.
(a ' a)
17- ) .? (0, 0),
1.
(a, a) ( x ) <o
> A (a, a)
E (Rd+ 1 )* : all aV > 0, aY < 0, ax > a,
which is equivalent to the negation of the "or" condition.
1.5
Recession Cone and Homogenization
Using the Farkas lemma, we can give an invariant description of some very
important constructions (notably the recession cone and the homogenization of a convex set) and establish their basic properties. In Proposition 1.14
we will see that the homogenization homog(P) of a polyhedron coincides
with the "associated cone" C(P) that we used in Section 1.1.
Definition 1.11.
of P is defined as
Let P C1 id be a convex set. Then the lineality space
lineal(P) := {y E I : z -I- ty E P for all x E P, t E R},

and the recession cone (or characteristic cone) of P is defined as
rec(P) := { y E Rd : z + ty E P for all x E P, t > 0}.
Directly from the definition we can derive that lineal(P) is a linear subspace of Rd . If we choose a complementary subspace U to lineal(P) (i.e.,
U n lineal(P) = {0} and U lineal(P)
d), then P can be decomposed
as the Minkowski sum
P .
lineal(P) + (P n U)
of a linear subspace L and a convex set Pn whose lineality space is zero:

lineal(P n
=
This reduction usually makes it possible to consider only polyhedra with
lineality space fol, known as pointed polyhedra (if they are nonempty).
For fl-polyhedra we compute lineal(P(A, z)) = fx E Rd : 'ix = 01.
So, except for a trivial linear summand we can usually consider polyhedra
P(A, z) C d for which A has full rank d.
u) fol.
44
Similarly, we see that rec(P) is a convex cone: it contains 0, any positive

multiple of a vector, and any convex combination of any two of its vectors.
Proposition 1.12.
Let P C Rd be a convex set.
(0 If P = P(A, z) is an fl-polyhedron, then so is its recession cone:

rec(P) = P(A,0).
(ii) If P = conv(V) + cone(Y) is a V-polyhedron, then so is its recession
cone:
rec(P) = cone(Y).
Proof. Both parts are "clear," aren't they? Not quite: on close inspection
we see that in part (ii) the direction rec(P) C cone(Y) is not entirely
obvious; it needs the Farkas lemma. Using version IV (for V = 0), we see
that if y tS cone(Y), then there exists a linear functional a with aY < 41)
and ay > O.
Now consider some z = Vt + Yu E conv(V) cone(Y), with t, u >
_ 0,
II t = 1. For this we get
az = aVt + aYu < aVt =
E
1<i<n
tiavi <
max av i =-: K,
1<i<n
where K only depends on a and V. However, we get that a(z + ty) =

az + t(ay) ----* -Foo for t 4 +oo, so we have z + ty (/ P for t large
D
enough, and thus y rec(P).
Definition 1.13.
of P is defined as
Let P C Rd be a convex set. Then the homogenization
homog(P) := {t ( xl ) : x E P, t > 0} + {() : y E rec(P)}.

Y
45
Again it is quite easy to see that for every convex set P, the homogenization homog(P) is a convex cone in Rd+ 1 . Furthermore, any P can be easily
recovered from its homogenization (if we don't mess with the coordinate
system) as
Proposition 1.14.
Let
PC
be a convex set.
(i ) If P = P(A,z) is an fl-polyhedron, then its homogenization is also

an fl-polyhedron:
hornog(P) --= P
((i zl
434) ' (0)) =
(1) If P = conv(V) cone(Y) is a V-polyhedron, then so is its homogenization:

homog(P) = cone (
1l 0
r y
C(P).
Proof. This now follows from Proposition 1.12.

The following proposition states two versions (linear and affine) of another
basic tool, known as Carathodory's theorem. We want to emphasize that
in contrast to the Farkas lemma this is completely elementary and also
computationally quite trivial. However, it can be successfully applied, for
example, to sharpen the Farkas lemmas, as well as the main theorems and
representation theorems for polytopes, cones, and polyhedra; see the next
lecture.
46
Proposition 1.15 (Carathodory's theorem).

Let X E Rd " and x E Rd .
(i) If z E cone(X), then z E cone(X`) holds for a subset X' c X of at

most rank(X) = dim(cone(X)) vectors in X.
(ii) If x E conv(X), then z E conv(X') holds for a subset X' C X of at
1
most rank ( ) = dim(conv(X)) 1 vectors in X.
X
We first describe the geometric idea (linear version). For this let cone(X)
have dimension k; assume that le > k +1 is the smallest number such that
z E cone(X) can be represented as a positive sum k' vectors in X. (We
obtain k' < n from z E cone(X).)
The cone cone(X`) spanned by such a set X' C X of k' vectors in X
can be interpreted as a projection of the positive orthant in lc' . Since k'
is minimal, the point z lies in the image of the interior of the orthant,
ft E ki : t > 0}. From k' > k we get that the preimage of z under
that projection is at least 1-dimensional. Thus the preimage contains the
intersection of a line with the orthant ft E R k' : t> 0}. Since the orthant
does not contain a whole line, the preimage contains a boundary point of
the orthant, and thus z can be represented as a conical combination of
fewer than k` vectors.
Similarly (affine version), we consider the projection of a simplex to the
polytope conv(X). If the image polytope has smaller dimension than the
simplex, then any point of the polytope has as preimage the intersection
of the simplex with a line. But the simplex does not contain a whole line,
so the line must contain a boundary point of the simplex, which leads to a
representation with fewer nonzero coefficients.
We will now give an algebraic proof.
Proof. For (i), without loss of generality we assume that X has full rank,
rank(X) d, by passing to the linear hull of X. Now let z E cone(X) and
Notes
47
let x = Xt with a vector t> 0 of minimal support supp(t) = : ti > 01,

that is, minimal number Ili : ti > OH of nonzero components. Now if
Isupp(t)i > d, then {xi : ti > 0 } is linearly dependent. This means that
there is a linear dependence of the form 0 = Ein Ai ti x, with all ) L O.
By multiplying this with a nonzero a E IR. we may assume that Ai > 0 for
some j E supp(t), and that max{) i: t i > 01 = 1. But then we get
x =
tix i = E(1 A i )tixi ,
which is a representation with smaller support, contradicting the minimality of t.

Now (ii) follows directly from
x E conv(X)
(1
E cone( X
Notes
The material of this lecture is classical
our discussion is inspired by
Griitschel's treatment in [227].
We recommend Schrijver's book [448, Sect. 12.2] for more historical
comments, as a superb guide to the historical sources, with references
to the original papers by Fourier, Dines, and Motzkin, and also those by
Minkowski, Weyl, Farkas, Carathodory, and others.
The elimination method was developed by Motzkin in his 1936 doctoral
thesis [384]. We quote from Dantzig & Eaves [167]:
For years the method was referred to as the Motzkin Elimination
Method. However, because of the odd grave-digging custom of
looking for artifacts in long forgotten papers, it is now known
as the Fourier-Motzkin Elimination Method and perhaps will
eventually be known as the Fourier-Dines-Motzkin Elimination
Method.
The Fourier-Motzkin elimination method is not only a theoretical tool
with some care it can also be used for computations. "In practice,"
however, one has to deal with an effect known as combinatorial explosion:
every elimination step may transform m inequalities into up to [m2/4] new
ones, which means that after a few steps the number of inequalities in the
system can increase dramatically. However, many of the inequalities we get
by elimination are redundant: they can be deleted without changing the
polyhedron that is described by the system. Thus it is important to eliminate redundant inequalities from the system, and in fact to detect many of
them quickly, in order to keep the problem size and the computation times
down.
48
Now detecting whether an inequality is redundant for an fl-polyhedron is

a nontrivial problem it is equivalent to a linear programming feasibility
problem. However, one can do better if during the elimination process, a
complete description in both the fi- and the V-versions are maintained.
That is, at every stage we assume that the polyhedron P is given in the
form
P = P(A , z) = conv(V),
and from this both types of descriptions are obtained for proji (P) respectively P n H. Given both V and (A, z), there are several different possible
criteria to decide whether an inequality is redundant, plus extra heuristics
that can be used to find some redundant inequalities fast.
This is the key to the double description method of Motzkin, Raiffa,
Thompson & Thrall [386]. The basic redundancy test, in the projection
version and in the intersection version, were discovered and rediscovered by
various authors; the main references we know after Motzkin et al. [384, 386]
are Burger [133], Chernikova [141, 142], Tschernikow [511, Chs. III and V],
Christof [143, 144], Padberg [404, Sect. 7.4], and Le Verge [332]. We give
an account of the basic criteria in Exercises 2.15 and 2.16.
There are several efficient C codes available for experiments, among them
the codes by Le Verge [332, 333], Wilde [528], and Fukuda [198]. A FORTRAN77 implementation is given by Alevras, Cramer & Padberg [6]. In
Example 0.6 we have used the C program PORTA by Christof [144].* With
respect to the basic version of the algorithm described above, PORTA uses
a few extra tricks:
1. rational arithmetic (where the denominator and numerator may be
arbitrarily long integers) is used to guarantee correct results,
2. the new inequalities are checked for irredundancy using criteria
as in Exercises 2.15 and 2.16 before they are generated explicitly,
thus saving time and space,
3. the same routine is used to convert V-polytopes into 71-polytopes
(the "convex hull problem") and conversely using polarity; see
Section 2.3.
The problem of constructing a convex hull has also been investigated
thoroughly in computational geometry Many alternative algorithms have
been suggested and studied. See for example Chazelle [140], Seidel's algorithm based on shelling [453], and the surveys in Mattheiss & Rubin
[356], Christof [143], and Borgwardt [119]. However, it seems that for highdimensional problems, as studied in polytope theory, Fourier-Motzkin elimination is hard to beat.
A different, strikingly elegant, "reverse search" enumeration method was
very recently described by Avis & Fukuda [27]. If we choose any generic
*See ftp://elah.zib.deipub/mathprog/polyth/index.html for these codes.
49
linear function, then the simplex algorithm (cf. Section 3.2) with "Bland's
rule" finds a path from every vertex of the polytope to the maximal vertex
of the polytope. Now if we have a simple polytope, these paths form a tree
that connects the vertices, and which can be searched easily. This yields a
very effective algorithm try Avis [24]. In the nonsimple case, one has to
search a tree on the (huge) set of all feasible bases; with some extra care
one can detect those bases that are lexicographically first at a vertex. (See
also Rote [430].) In the very nonsimple case, this does not seem to be very
efficient at first, though Avis [24] has reported the successful solution of
very large convex hull problems with reverse search; see also [137].
Avis, Bremner & Seidel [25, 26, 124], however, construct and analyze
classes of "bad" test examples for different types of convex hull algorithms.
In particular, products of cyclic polytopes of the form Cd(n) d seem to be
"universally" bad for all known types of convex hull algorithms.

1.0 If we try to restrict the proof of Theorem 1.2 to polytopes, where
does this fail? In other words, where do we use a construction that
necessarily takes us from polytopes to unbounded polyhedra?
1.1 Experiment with the Fourier-Motzkin elimination and the double description method on some examples.
(i) Compute the vertices of the 2-dimensional example of Section
1.2. Check carefully that everything you get actually is a vertex.
(ii) Compute defining inequalities for the cyclic polytope C3 (6), for
example for t = i. For every inequality, find the set of vertices
that satisfy it with equality
(iii) Find the facets of the 4-polytopes of the Exercises 4.8.6 and
4.8 15 from Griinbaum [234, pp. 64,65].
(iv) These were Mickey-Mouse** examples (i.e., very small).
For more realistic ones, find all the facets of the traveling salesman polytopes QT (6), QT (7), QT(8), and of the asymmetric
travelling salesman polytopes Q'T (5) and Q'T (6).
(Better use a computer; QT (8) is a 20-dimensional polytope
with 2520 vertices and 194, 187 facets [146]. The polytope QT (9)
has 42,104,442 facets; for Q T (10) Christof found 51,043,900,866
facets and conjectures that they give a complete description
see [145] and [147]. Similarly, Euler SE Le Verge [188] have derived a description of Q'T (6): a 19-dimensional polytope with
120 vertices and 319,015 facets.)
*C) Walt Disney 1927
50
1.2 Describe a Fourier-Motzkin elimination method to solve strict inequality systems Ix : Ax < zl. Use it to prove a representation
theorem for "open polyhedra."
1.3 Show that if C = cone(W) is any cone in d+1 generated by arbitrary
vectors wi (not necessarily with wio > 0), then Ix E Rd : (xl ) c CI
is a V-polyhedron.
1.4 State and prove a Farkas lemma for systems of the form Ax < z,
x>
O.
1.5 Prove the following Farkas lemma for equality and inequality constraints: for compatible matrices A, B C and vectors u, V w
either there exists a solution vector x for
,
Ax = u, Bx> v, Cx < w,
or there exist row vectors a, b, c with
+ bB +cC = 0, b < 0, c> 0, au +bv + cw < 0.

1.6 Prove the following general Farkas lemma for equality and inequality
constraints: for compatible matrices A, B, C, D and vectors z, w
either there exist solution vectors x,y for
Ax+ By < z, Cx+ Dy =-- w, x> 0,
or there exist row vectors c,d with
cA+dC > 4), cB 1-dD 0, c> 0, cz +
<O.
1.7 State and prove a version of Carathodory's theorem for convexconical combinations.
1.8 Transportation polytopes have the form
P(d : a,b) =
C ER d " : xii > 0
for 1 <
k=1
k=1
xi,
j < d,
ai for 1 < i < d,
xkj bj for 1 < j < d.}
Study transportation polytopes. Determine the dimension. Interpret

vertices and facets. Show that the Birkhoff polytopes are special
transportation polytopes. Show that non-empty transportation polytopes P(d:a,b) have canonical center points, given by xii ft-kg- for
ELI
all i and j, where A :=

ak
1.9 Interpret
Interpret the Farkas lemma IV as a statement about polyhedra, and
observe that it follows "trivially" from the equivalence of V- and 7-tpolyhedra (Theorem 1.2). Derive Farkas lemma II from Farkas lemma
IV, and then Farkas lemma I from Farkas lemma IL
(This alternative route through the jungle of Farkas lemmas was suggested by Joe Bonin.)
2
Faces of Polytopes
In this lecture we will discuss faces, and the face lattice. Here we restrict
our attention entirely to polytopes, although nearly everything can quite
easily be generalized to polyhedra; see the exercises.
I hope that the reader enjoys the ease with which we will get the results
in this lecture. In fact, nearly all results are "geometrically clear," and
as far as we need algebra to verify them, we get by with straightforward
computations and the Farkas lemmas.
2.1
Vertices, Faces, and Facets
Definition 2.1. Let P C Rd be a convex polytope. A linear inequality

ex < co is valid for P if it is satisfied for all points x E P. A face of P is
any set of the form
F -= P riIx E : cx col
where ex < co is a valid inequality for P. The dimension of a face is the
dimension of its affine hull: dim(F) := dim(aff(F)).
For the valid inequality Ox <O, we get that P itself is a face of P. All
other faces of P, satisfying F C P, are called proper faces.
For the inequality Ox < 1, we see that 0 is always a face of P.
The faces of dimensions 0, 1, dim(P) 2, and dim(P) 1 are called vertices, edges, ridges, and facets, respectively. Thus, in particular, the vertices
are the minimal nonempty faces, and the facets are the maximal proper
faces. The set of all vertices of P, the vertex set, will be denoted by vert(P).
52
2. Faces of Polytopes
The following sketches show two valid inequalities for a 2-polytope; they
define a vertex and an edge, respectively.
In the following two propositions we collect some simple but basic facts
about faces.
Proposition 2.2.
Let P C Rd be a polytope.
(i) Every polytope is the convex hull of its vertices: P -= conv(vert(P)).

(ii) If a polytope can be written as the convex hull of a finite point set,
then the set contains all the vertices of the polytope: P =--- conv(V)
implies that vert(P) C V.
Proof. Let P
conv(V). Now if any vector yi E V can be written as a

convex combination of the other vectors in V, then we can clearly substitute
that representation into any convex combination of vectors in V, and thus
get a smaller set of points V' :=--- V \vi, whose convex hull is conv(V 1 ) = P.
Now we claim that if v, cannot be expressed as a convex combination
of V' V\Y,, then it is a vertex of P. Using Farkas lemma II (Proposition 1.8), we get
y, conv(V) < >
t>O:v==V't, llt=1
<
3a : a (v
n ,) 0, a ( v1.2 ) <0
<
A (0, b) = a : by' </31, bvi > 0
< > 3 0,b : bvi < 0 for j i, bY, >

Thus y, is a vertex, defined by the valid inequality bx <
Finally we observe that a vertex v, of P can never be written as a
convex combination of points in P\Yi , which finishes the proof of both (A
and (ii).
2.1 Vertices, Faces, and Facets
Proposition 2.3.
be a face of P.
53
Let P C Rd be a polytope, and V := vert(P). Let F
(0 The face F is a polytope, with vert(F) =- Fn V.

(ii) Every intersection of faces of P is a face of P.
(iii) The faces of F are exactly the faces of P that are contained in F.
(iv) F = Pn aff(F).
Proof. Let F be defined by the valid inequality cx < co.
For the first assertion of part (i), we see that F is a polytope from the
characterization of polytopes as bounded intersections of halfspaces: F is
the intersection of a polytope P with a polyhedron (hyperplane)
H := x E Rd : cx = col.
Furthermore, we find that F C aff(F) C H, which proves (iv).
For the second assertion of (i), note that vert(F) D Fn V =: Vo. For
the converse inclusion, let x E F, so that x can be represented as x = Vt,
with t > 0, 11 t 1. We compute
co
= cx = c(Vt) = (cV)t < collt
thus (cv, co)ti --= 0 for all i. This implies that ti = 0 for all j with
vi Vo , and thus x E conv(Vo ). From this we see F conv(170), and thus
vert(F) C Vo by Proposition 2 2(4 This completes the proof of (i).
For (ii), let
F= Pn { x e Rd : cx
co}
and
G=Pn{xERd :bx=b0}
for inequalities cx < co and bx < 1)0 that are valid for P. Then the inequality (c + b)x < co + bo is valid for P, and
Pn
d : (C b)X --= Co
b0 } --= F n G.
For (iii), if G C F is a face of P, then it is a face of F as well. For the

converse let F Pn
c Rd : cx = col and G = F n Ix c Rd : bx =
bo } C F, where cx < co is valid for P, and bx <b0 is valid for F, but not
necessarily for P.
Let Vo
vert(F) as before, and 1/1 := V\Vo. We can assume that
F P, and thus 171 7k 0. Now (b+ Ac)x < b0 + Aco is valid for F, for every
A E R, and defines G as a face of F.
Now choose A large enough to satisfy A>
bo by
for all v
co cv
E V1. Then
we get that (b-FAc)x < bo -FAQ) is valid with strict inequality for all v c 14.
54
We conclude that G is a face of P. (The sketch below might tell you what
0
was "really going on" during this algebraic manipulation.)
We need another construction: the "vertex figure" obtained by cutting a

polytope by a hyperplane that cuts off a single vertex.
For this, we consider a polytope P with V = vert(P), and a vertex i2 E V.

Let ex < co be a valid inequality with
{v} = P n {x : cx = co }.
Furthermore, we choose some c 1 <c0 with cvi <c 1 for all TY E vert(P)\v.
Then we define a vertex figure of P at v as the polytope
P/v :--- Pn {x :
CX =-- CI }.
Note that the construction of P/v depends on the choice of cl and of the
inequality ex < co ; however, the following result shows that the combinatorial type of P/v is independent of this.
55
Proposition 2.4. There is a bijection between the k-dimensional faces

of P that contain y, and the (k 1)-dimensional faces of Ply, given by
F
u: P n aff({y} U F')
Frqx:cx---- c1},
F'.
7r :
Proof. Denote the "cutting hyperplane" by H :=

: cx = ci}.
: bx = bo b then we get
The map 7r is well defined: let F = Pn
F n H = (P n H) n {x : bx = bob where bx < bo is valid for P and thus
also for Pfv.
To see that u is well defined, let F' = (Pn H) n Ix : bx = bob where
bx <b0 is valid for Ptv. Then (b + )c)x < bo + Aci is valid for Ply, for
all A C R. Now a simple computation shows that for
A :=
bo bv
co c1
the inequality is valid on P, with equality for v. In fact, if we consider

V' E V\v, then we know cv' <c 1 and cv = co > cl , so
(cv c i )vl+ (c 1 cvl)v
cv cvi
P/v.
Now v" is a convex combination of v and v', and from (b + Ao c)v" <
bo + Aoc i and (b + Aoc)v = bo + Aoci we get (b Aoc)v i <b0 + AociNow we check that the maps u and ir are inverses of each other: we
compute
rocr(F1 ) = HnPnaff({v}UF') = Pnlinaff(F') = Pivnaff(P) = F'

where the last equality is from Proposition 2.3(iv). Similarly,
cr o 7r(F) = P n aff({v} U (F n H)) =
* P n aff(F) = F,
where for the equality * we use that every vertex y' of F can be obtained
as an affine combination of v and a point v" E F n H.
Finally we observe that if F' is a face of Pit) of dimension k-1, then
the associated face F of P has dimension k, since its affine hull is aff(F)
aff(F' U {v}), where v aff(F') by construction.
2.2
The Face Lattice
In this section we translate some of our results on polytopes into purely

combinatorial statements. For this we need some terminology about finite
partially ordered sets ("posets," for short). We refer to Stanley's book [480,
Sects. 3.1-3.3] for more information on that subject and its ramifications.
For simplicity, and to unify terminology, we define the key concepts here.
56
Definition 2.5 (Poset terminology).

A poset (S,<) is a finite partially ordered set, that is, a finite set S
equipped with a relation "<" which is reflexive (x < x for all x E S),
transitive (x < y and y < z imply x < z), and antisymmetric (x < y and
y 5 x imply x -= y).
Usually we denote such a poset by S, when the partial order is clear.
Any subset of S is also a poset, with the induced partial order. A chain
in S is a totally ordered subset of S; its length is its number of elements
minus 1.
For elements x, y E S with x < y, we denote by
[x,y] := {wES:x5.w<y}
the interval between x and y. An interval in S is boolean if it is isomorphic
to the poset Bk --= (2[k] , C) of all subsets of a k-element set, for some k.
A poset is bounded if it has a unique minimal element, denoted 6, and a
unique maximal element, denoted . The proper part of a bounded poset S
is 3 :-..-il.
A poset is graded if it is bounded, and every maximal chain has the same
length. In this case the length of a maximal chain in the interval [6, x] is the
rank of x, denoted by r(x). The rank r(S) := r(i) is also called the length
of S. For example, every chain is a graded poset, with r(C) = ICI 1, and
the boolean posets Bk are graded of length r(Bk) = k, for all k > 1.
A poset is a lattice if it is bounded, and every two elements x, y E S have
a unique minimal upper bound in S, called the Join x V y, and every two
elements x, y E S have a unique maximal lower bound in S, called the meet
x A y. (In fact, any two of these three conditions imply the third; also, if
every pair of elements has a join respectively meet, then also every finite
subset has a join respectively meet.)
If S is a graded lattice, then we call the minimal elements of S\ its
atoms, and the maximal elements of SVi its coatoms. Equivalently, the
atoms are the elements of rank 1, and the coatorns are the elements of rank
SVC),
r(S) 1.
A lattice is atomic if every element is a join x = al V ... V ak of k > 0
of atoms, where we obtain x = for k = 0, and an atom x = a l for k = 1.
Similarly, a lattice is coatomic if every element is a meet of coatoms.
We define the opposite poset Se" (or order dual) to have the same underlying set as S, with x < y in 50P if and only if y < x holds in S.
We use the graphical representation of posets by Hasse diagrams, that
is, graphs drawn in the plane so that the elements correspond to vertices,
where x < y holds if and only if there is an increasing path from x to y.
Here we only include the edges corresponding to cover relations, that is, if
x <y and [x, yl = {x, y}.
Of the posets in the following figure, the first one is not bounded, but
all others are. The second one is a lattice, but not graded. The third one
57
is graded (of length 3), but it is not a lattice. The fourth poset is a graded
lattice (of length 3), and the fifth one is even boolean (isomorphic to B3).
The fourth poset is neither atomic nor coatomic, but the fifth one is.
Why is all this interesting for us? Because we want to study the set of
faces of a convex polytope, ordered by inclusion.
Definition 2.6. The face lattice of a convex polytope P is the poset
L := L(P) of all faces of P, partially ordered by inclusion.
The following figure shows, as an example, the face lattice of a convex

pentagon. Here the minimal element corresponds to the empty face, the five
atoms in the layer above correspond to the five vertices, the layer above
this represents the five edges (each containing two vertices), and the top
element represents the pentagon itself.
In Theorem 2.7 we collect the main structural properties of face lattices,

starting with the fact that they are lattices, justifying the terminology of
Definition 2.6.
Theorem 2.7.
Let P be a convex polytope.
(i) For every polyt ope P the face poset L(P) is a graded lattice of length
dim(P) + 1, with rank function r(F) = dim(F) + 1.
(ii) Every interval [G, F] of L(P) is the face lattice of a convex polyt ope
of dimension r(F) r(G) 1.
(iii) ("Diamond property") Every interval of length 2 has exactly four
elements. That is, if G C F with r(F) r(G) = 2, then there are
exactly two faces H with GCHCF, and the interval [G, F] looks
like
(iv) The opposite poset L(P)P is also the face poset of a convex polyt ope.
(N) The face lattice L(P) is both atomic and coatomic.
58
Proof. To see that L(P) is a lattice it suffices to see that it has a unique
maximal element i = P and a unique minimal element = 0, and that
meets exist, with FAG=FnG; this is true because F n G is a face of F
and of G, and thus of P, by Proposition 2.3(h). And clearly every face of P
that is contained in F and in G must be contained in F n G.
We continue with part (ii). For this we can assume that F = P, by
Proposition 2.3(iii). Now if G = 0, then everything is clear. If G 0,
then it has a vertex y E G by Proposition 2.2(i), which is a vertex of P
by Proposition 2.3(iii). Now the face lattice of Pity is isomorphic to the
interval [M, PI of the face lattice L(P), by Proposition 2.4. Thus we are
done by induction on dim(G).
For part (i) it remains to see that the lattice L(P) is graded. If G c F
are faces of P, then from G = Pn aff(G) C Pn aff(F) = F, which holds by
Proposition 2.3(iv), we can conclude that aff(G) c aff(F), and thus that
dim(G) < dim(F). So it suffices to show that if dim(F) dim(G) > 2, then
there is a face H E L(P) with GCHC F. But by part (ii) the interval
[G, 1 ] is the face lattice of a polytope of dimension at least 1, so it has a
vertex, which yields the desired H.
Part (iii) is a special case of (ii): the "diamond" is the face lattice of a
1-dimensional polytope.
We don't prove part (iv) here -- but we will do so in the next section.
Finally, for part (v), the first part is immediate from Proposition 2.2(i),
where the atoms of L(P) correspond to the vertices of P, and the second part follows from this by taking the opposite poset, according to
part (iv).
0
This theorem contains quite restrictive information on the structure of
polytope face lattices (Exercise 2.3). We will get even more precise information later.
We note here that the face lattice is the proper framework to define combinatorial equivalence of polytopes. In fact, our previous definition (before
Example 0.2) can be restated as saying that P and Q are combinatorially
equivalent, P ,--' Q, if and only if L(P) r' L(Q): if their face lattices are
isomorphic.
By Proposition 2.2(i), this is equivalent to a bijection vert(P) 4-* vert(Q)
between the vertices of P and Q, in such a way that the vertex sets of faces
of P correspond (under this bijection) to the vertex sets of faces of Q. A
general observation is that in this context it is enough to deal with vertices
and facets, because the faces are exactly the intersections of facets, and the
vertex sets of faces are exactly the intersections of vertex sets of facets
see Exercise 2.7. (Abstractly, the key properties are that face lattices are
atomic and coatomic.)
Topologically, combinatorial equivalence corresponds to the existence of
a (piecewise linear) homeomorphism between the polytopes P-2'-- Q that
restricts to homeomorphisms between the facets (and hence all the faces)
2.3 Polarity
59
of P and Q. In other words, P212 Q if and only if P and Q define isomorphic

cell complexes (CW-balls) see Munkres [388] or Bji5rner [85, Sect. 12]
[92, Sect. 4.7] for these concepts.
2.3
Polarity
We proceed to construct polar polytopes: this is what (nearly) everybody

else calls the "dual" of a polytope. We will use the term "polar" in order to distinguish polarity from duality in the sense of (oriented) matroid
theory, which we will see later in Lecture 6 in the form of Gale diagrams,
the Lawrence construction, and others. (In this, we follow the conventions
of [92, pp. 44-451 )
A key observation one should not miss is the step into dual space we
take in this section. Equivalently, one could just fix a scalar product on I
and this way we could construct a polar polytope in the same space as the
original. However, there is a lot of choice in any case because the location
of the origin is essential for our construction. If we wanted to avoid this, we
would have to linearize, and develop "cone polarity" using some methods
of Lecture I. We will not do this here (in order to keep the principle of
making this a lecture on polytopes), but see Exercise 2.13.
d ,
Lemma 2.8. Let P be a polytope in R'. Then the following conditions

are equivalent for y E P.
(i) y is not contained in a face of P of dimension smaller than d,
(ii) if ay = ao and a
0, then ax < a0 is not valid for P,
(iii) y can be represented in the form y = Ezd 0 Aixi for d + 1 affluely

independent points x 0 , ... xd E P and for parameters Ai > 0 with
Eid=0= 1 ,
(iv) y can be represented as y =
points x 0 ,
xd c P.
d+1 i=0x
for d+1 affinely independent
Proof. Part (i) implies that P is full-dimensional. From this we get that
part (ii) holds: if ax < a0 were valid for P, then y would be contained in
the face P n {x c d : ax = ao} of smaller dimension. Conversely, if part
(ii) holds, then no inequality can define a facet that contains y.
Part (iv) trivially implies part (iii), and from this we get part (ii) by an
easy calculation:
d
ao = ay
Ai axi
Ai ao = a.
i=0
60
This can hold only if ax i= ao holds for all i, so either a = 0, or the

points xi are not affinely independent.
Now assume that part (ii) holds. We write P in the form P = P(A, z).
No equality ax = ao with a y 0 is valid for P, so P is full-dimensional. We
claim that for every u E rd we have y + au E P, if a > 0 is small enough.
In fact, this is true unless one of the inequalities that define P is satisfied
with equality for y, which is excluded by part (ii). But now we can choose
d+ 1 possible a-values for u = ei,...,ed, (e l + ed), let a' > 0 be
the minimum of those, and write
1
1(Y a`
d +1
ed +
(y +de l ) + +(y+ ded)} ,
which is a representation of the desired form.

If the conditions of Lemma 2.8 are satisfied for y, we say that y is an
interior point of P. Moreover, we use the notation int(P) for the interior
of P, which is the set of all interior points of P. (It is easy to verify that
this coincides with the usual (topological) definition of the interior of the
point set P C d.)
The problem is that the interior of a polytope is not invariant under
affine equivalence of polytopes: the center of a triangle is an interior point
if the triangle is embedded in R2 , but not if it is embedded in R3 . In fact,
int(P) = 0 if P is not full-dimensional in Rd.
Thus, we define the relative interior relint(P) of a polytope, defined
as the interior of P with respect to the embedding of P into its affine
hull aff(P), in which P is full-dimensional. Analogous to Lemma 2.8, the
following lemma characterizes relative interior points of P.
Lemma 2.9. Let P be a polytope of dimension k := dim(P) in some Rd

(k < d). Then the following conditions are equivalent for y E P.
(i) y is not contained in a proper face of P,
(ii) if ax < a0 is valid for P, with equality for y, then ax = ao holds for
all x P,
(iii) y can be represented in the form y =E ik_0 A i x i for k + 1 a.ffine/y
independent points x 0 ,... , xk E P and for parameters Ai > 0,
(iv) y can be represented as y = k _1_ 1
dent points x 0 , .. ,x k E P.
Eik_o xi
for k + 1 aflinely indepen-
Note that if P is nonempty, then its relative interior contains a point,
relint(P) 0. To see this, we can take the barycenter of the vertex set of P,
vi for vert(P) = {v 1 , ... ,vN}, or the barycenter of any
as y :=
set of dim(P) +1 affinely independent points (e.g., vertices) in P according
to Lemma 2.9(iv).
2.3 Polarity
61
Furthermore, we get a decomposition of P into a disjoint union of the

relative interiors of its faces, from the characterization of Lemma 2.9(i):
P=
Lt j
relint(F).
FEL(P)
For many problems, in particular for the polarity construction we want

to do now, it is convenient to assume that 0 c int(P) without loss of
generality ("w.l.o.g."). This can be achieved, if P is nonempty, by an affine
map. In fact, for this we project to aff(P) and then translate any interior
point to 0. Now it is easy to see (with Lemma 2.8) that P satisfies this
condition if and only if it can be represented as P -=- P(4,1).
Definition 2.10.
For any subset P C Rd, the polar set is defined by
P := { c E (Rd ) * : ex < 1 for all x E P} c (Rd)*.

The following shows a convex pentagon P in the plane, determined by
its five vertices, and its polar P, a pentagon given by five inequalities.
A
Clearly, the construction of the polar can be iterated, and thus we get
the polar of the polar, or double-polar, as
PA = {y E Rd : cx < 1 for all x E P implies ey < 1, for c E ( ) }
where we have identified Rd and (Rd)** in the natural way. Now let's examine the nearly obvious (?) basic properties of the polar and double-polar
constructions.
62
Theorem 2.11.
(0 P C Q implies P D QA and PA C QA
(ii) P C PA ,
(iii) P and PA are convex,
(iv) 0 E PA , and 0 E PA ,
(v) if P is a polytope and 0 E P, then P , P ,

(vi) if a polytope P with 0 E int(P) is given by P = conv(V), then
P --,---
fa : aV < 1 1,
(vii) if a polytope P with 0 E int(P) is given by P = P(A,1), then

P = {cA : c > 0, c1 = 1}.
In part (vi), the representation P = conv(V) means that P is a polytope,
and the representation P .--- P(A, 1) in part (vii) implies that 0 E int(P).
Extending the definition of convex hulls and inequality systems to the dual
space (i.e., for row vectors), the statements of (vi) and (vii) can be rewritten
and combined as
P = conv(A) = P(V, 11 ).
Proof. Parts (i) to (iv) we can safely leave to the reader: these are routine
exercises that should not need note paper.
For part (v), we rely on the conscientious reader to get his or her note
pad and do the proof. It also follows from part (vi), for which we compute
P =
=
la : ax < 1 for all x E
PI
la : au < 1 for all v E V},
where for the last equality "C" is trivial, while "D" follows from convexity
(or from a trivial computation).
For part (vii), we compute
PA =
fa : ax < 1 for all x : Ax < 1}
= {cA : c > 0, cl = 1},

where for the last equality "D" follows from a simple computation (cA)x =
c(Ax) < cl = 1, while "C" is a little harder. For this note that Ax < 1 has
a solution. Thus by Farkas lemma III the validity of ax < 1 implies that
there exists a cf > 0 with c'A = a and c'l < 1. Now since P is bounded,
there is no x with Ax < 1, otherwise x L 0 and Ax is in P for all A _?__ O.
2 3 Polarity
63
From this, by Farkas lemma I, there exists a c" > 0 with c"A = 0 and
el > O. With this we can put
i. - c/i_ t
e
c" l
1
which satisfies c > 0, cA = c' A = a, and cl = el + (1 - c'1) = 1.
For illustration of polarity, we also refer the reader to the "classical"

pairs of the (regular) cube and octahedron, as sketched in Example 0.4,
and of the regular dodecahedron and icosahedron assuming that they
are represented in such a way that 0 is the center of symmetry. Furthermore,
the polar of any simplex is a simplex (Exercise 2.12).
In fact, it turns out that the combinatorial structure of P is independent
of the exact embedding in Rd, as long as we have 0 E int(P), as we will see
from the next theorem.
For this, we assume that P C Rd is again a d-polytope with 0 in its
interior. In this situation we study, for all faces F of P, the subsets of P
of the form
F := lc
Theorem 2.12.
in Rd, and that
( )
cx < 1 for all x E P,

cx = 1 for all x E F} C (
Assume that P = conv(V) = P(A, 1) is a polytope
F ..= conv(V') --,--- Ix E
: A''x < 1, A ix =1}
is a face of P, with V = V'EOV" and A = A'WA" .

(Here we need that all the inequalities ax < 1 in the system "Ax < 1"
that satisfy F C Ix E Rd : ax < 1} are included in "Aix <1.")
Then
F = {ek : c' > 0, el = 1} = {a : aV" < 1, a-V 1 = 1} .

Proof. We compute
F -=
.-_-_.
=
,-_-__
{a : ax < 1 for all x E P, ax = 1 for all x E F}

{a : av < 1 for all v E V, av = 1 for all v G VI
fa : av < 1 for all v E V'', av = 1 for all v E VI
}
la : aV" < 1, aV' = 11 1.
For the other half, we use the description of P in parts (vi) and (vii) of
Theorem 2.11, and get
F<> =
la : ax < 1 for all x E P, ax = 1 for all x E F}

= {cA : c > 0, cl = I, cAx = 1 for all x E F}
= { c'A' : e ? 0, el = 1},
64
where for the last equality, "D" is clear, while for "C" we have to work.
In fact, for this we can choose some x E relint(F), which satisfies Aix = 1
and A"x < 1, and rewrite cA ,----- c' A' + c"A" . Then by
1 = (cA)x --, (c` A i +c" Anx .---- ci (A`x)+c" (A" x) < ci l+c"1..----- cl ,--- 1
we have c"(A"x) = c"1, which by A"x < 1 implies c" = O.
Corollary 2.13. Let P be a polytope with 0 E int(P), and let F, G E

L(P) be faces of P. Then
(i) F is a face of P ,
(ii) F" = F, and
(iii) F C G holds if and only if F D G .
Corollary 2.14.
of P:
The face lattice of P is the opposite of the face lattice

1,(13 ) '.' L(P)P.
This, in particular, completes the proof of Theorem 2.7, the last two parts
of which we had deferred (remember?). It means that for every statement
about the combinatorial structure of polytopes, there is a "polar statement," where the translation reverses inclusion of faces, and interchanges
4-9. i = P
4---+ facets
vertices
edges ) ridges
etc.
Note that polarity also identifies the face lattices of facets with (the opposites of) the face lattices of vertex figures. Finally, it says that the "polar"
combinatorial descriptions of polytopes, as V-polytopes in terms of vertices,
and as 7-1-polytopes in terms of facets, are logically equivalent.
Nevertheless, the metric properties of the polarity construction depend
on the location of 0 in P, whereas the combinatorial ones do not. This
motivates to define that two polytopes P and Q are cornbinatorially polar
if L(P) '' L(Q)P. Thus the construction of P establishes the existence
of a combinatorially polar polytope for every polytope P.
2.4
The Representation Theorem for Polytopes
This section has a (by now) simple task: to state and prove the general
representation theorem for polytopes. There is no real work left to do: we
have assembled all the ingredients, notably Fourier-Motzkin elimination,
the Farkas lemmas, Carathodory's theorem, and polarity. One new term
appears in its statement: the k-skeleton of a polytope is the union of its
k-dimensional faces.
65
Theorem 2.15 (Representation theorem for polytopes).

A subset P C rd is a polytope if and only if it can be described in any of
the following (equivalent) ways:
(1) an affine projection of a simplex,

(2) all the convex combinations of a finite point set,
(3) all the convex combinations of the vertex set vert(P),
(4) the union of all simplices spanned by a finite set of points,
(5) the projection of the d-skeleton of a simplex,
(6) a bounded intersection of (closed) halfspaces,
(7) a bounded intersection of facet-defining (closed) halfspaces, one for
each facet, and of the affine hull of P.
Proof. The equivalence of (1) and (2) is from the definitions of a convex hull and a simplex. It is just an example of translation of a geometric
statement (1) into an algebraic one (2). Similarly, (4)< >(5) is such a translation.
The equivalence of (2) and (3) is from Proposition 2.2, while the equivalence of (1) and (5) is from Carathodory's theorem 1.15(ii).
The equivalence of (2) and (6) is the main theorem on polytopes, Theorem 1.1, which we proved by Fourier-Motzkin elimination. Instead, we
could also argue that (2)< >(6) follows from polarity.
Finally, for (3)< >(7) we reduce this to the full-dimensional case, and
then use that the facets of P correspond to the vertices of P under polarity, by Theorem 2.12 and its corollaries. In particular, the facet-defining
inequalities are uniquely determined (if we write them as ai x < 1), and
none of them can be deleted.
0

Proposition 2.16. For any d-dimensional polytope P, the following conditions are equivalent:
(i) every facet of P is a simplex, i.e., P is simplicial,
(ii) every proper face of P is a simplex,

(iii) every facet has d vertices,
(iv) every k-face has k + 1 vertices, for k < d 1,
(N) every lower interval 0, fl c L(P) in the face lattice with F i is a
boolean poset.
66
Faces
of Polytopes
Similarly, the following conditions are equivalent:

(i) every vertex figure of P is a simplex, i.e., P is simpl e,
(ii) every iterated vertex figure of P is a simplex,
(iii) every vertex is in d facets,
(iv) every k-face is contained in d k facets, for k ? 0,
(y) every upper interval [F, ij L(P) in the face lattice with F
boolean po.set.
is a
In particular, a polyt ope is simplicial if and only if any combinatorially

polar polytope is simple, and it is simple if and only if any combinatorially
polar polytope is simplicial.
Proof. This is easy. It only uses the fact that every (d 1)-dimensional
simplex has ( k1 ) k-faces, and the fact that its face lattice is the boolean
poset Bd of all subsets of a d-set.
The first and second parts of the theorem are equivalent, via polarity.
Here we use that the opposite of a boolean poset is isomorphic to the poset
C
itself.
For the following, assume (without loss of generality) that we consider
full-dimensional polytopes. For any simplicial polytope P = conv(V) we
can perturb the vertex coordinates "a little" without changing the combinatorial type. From this we get a combinatorially equivalent polytope with
rational vertex coordinates, which is what one calls a rational polytope.
Similarly, for any simple polytope P = P(A, z) we can perturb the
defining inequalities "a little" to get inequalities with rational coefficients.
Clearing the denominators, we can get representations with integral vertex
coordinates (as a lattice polytope). This proves the following result.
For every simple or simplicial polytope P, there is a

combinatorially equivalent polytope P' '-- P with integral vertex coordinates.
Proposition 2.17.
This answers one question, and opens up two new ones: first, is this true
for all polytopes? We will see later that it holds for polytopes of dimension
d < 3 (the case d < 2 is trivial), but it fails in general (see Lecture 6).
Second, if integral coordinates exist, can we keep them reasonably small?
Again, the answer is yes if we are in low dimension, but in general we
have to cope with coordinates that grow doubly exponential in terms of
the number of vertices; see Goodman, Pollack Si Sturmfels [220]. For fixed
dimension d = 3, however, the problem seems to be open (Problem 4.16*).
We already saw numerous examples of simple and simplicial polytopes
in Lecture 0. The d-simplex, the d-cubes, and the dodecahedron are simple
polytopes. The d-simplex, the octahedron, the icosahedron, and all cyclic
2.6 Appendix: Projective Transformations
67
polytopes are examples of simplicial polytopes. Note that in their "usual"

symmetric version as regular polytopes, the dodecahedron and the icosahedron do not have rational coordinates. (In fact, as regular polytopes they
cannot be represented in rational coordinates!)
Any d-polytope that is both simple and simplicial is a simplex if d> 3
this was covered in Exercise 0.0. To see this, let P be simple, and consider
a vertex y of P. This vertex is in exactly d facets, which are all simplices
Looking at the vertex figure, we see that v is on d edges, and the d vertices
vd adjacent to v are also adjacent to each other: here we use the
condition d > 3. This means, since the same argument could start at vi ,
that there are no other vertices than y, V 1 ,. ,V_ 1 , vi+1 , vd adjacent
to vi . Thus the vertex set of P is { v, v i ,... vd}, and P is the simplex
spanned by this set.
2.6
Appendix: Projective Transformations
Although linear transformations are our main tool to "put polytopes where
we need them," it is sometimes convenient to use more general transformations, which allow us even to "adjust the shape of a given polytope,"
known as projective transformations.
We can describe projective transformations in a very simple way with the
tools of this lecture (in particular, without construction of projective space
and use of projective geometry). For this, we proceed as follows. Given a
polytope P C j d, we embed it into an affine hyperplane H C d + 1 , and
construct homog(P), the homogenization of P. By construction, this is a
pointed cone. Now we cut this cone by a different hyperplane K C d+1,
which is then identified with Rd by an affine map.
-3111*
Here K is required to be an admissible hyperplane: an affine hyperplane that

intersects every ray in homog(P) that starts at O. Under this condition, we
get a new polytope P' c lie, which is affinely isomorphic to Knhomog(P),
and thus combinatorially equivalent to P:
L(P)
L(hornog(P))
L(P`).
68
Briefly, the geometric procedure can be described as homogenization

(embedding into an affine hyperplane) followed by dehomogenization (with
respect to a new affine hyperplane).
We now derive formulas that directly describe the map f: P --> P' in
d-dimensional space. For this, let
x
H = {( ) : x E Rd }
1
and
d+1
X d+1 -77- 1} C R d+1
Xd+1
1(
--= {
Rd+1 :
ax + ad-4-1X d+1
Xd+1
11.
The hyperplane K is admissible if and only if (a, ad+1)(11') > 0 for all
vertices y E vert(P). We map K back to d via an affine map
( X )
Bx +
Xd+1
This map is an isomorphism 7r : K 2--!
det(
if and only if
B z
a ad-Fi
O.
This means that for ad+1 0 we could take B -= Id,
x )
xd+i
+ z'.
0, and thus
x.
Putting these elements together, we get the projective transformation in

formulas as
E P
E H
1
x)
EK
ax + ad+1 ( 1
Bx + z
+ E Rd.
ax + ad-Fi
Thus a projective transformation acts on P as a rational linear map, and

we get for P' the general formula
Bx + z
+ : x E P}
ax + ad+1
under the conditions that
det
(B
a
z
ad+i j
and that ay + ad+1 > 0 for all y E vert(P).

Geometrically, the transformation has the following effect. The map is
defined on the interior of the halfspace k+ = Ix E d : ax + adi > 0}.
It maps subspaces that meet the interior of k+ to subspaces of the same
Notes
69
dimension. The hyperplane K, which bounds the halfspace, is "moved to

infinity" by the projective transformation. Hence, lines that intersect in
the interior of K+ are mapped to intersecting lines, while lines that meet
on the boundary hyperplane K = Ix C Rd : ax ad+1 = 0} of K+ are
mapped to parallel lines.
.
K . ... ,
. ,
\
.....4110,
It is certainly instructive to get a good book on projective geometry you

might try Garner [209] for the basics, or classical treatments like Veblen &
Young [515] and Hodge & Pedoe [254] for more and study its definition
and description of projective transformations, and try to match it with the
one given here.
Understanding projective transformations is absolutely necessary if you
work more with polytopes, even if we will not use them much in the following lectures. Nevertheless, the reader will recognize them at times (used
explicitly and implicitly).
Projective transformations are used for "preprocessing," to get a polytope into the shape to apply certain procedures, without changing the
combinatorial structure. For this, it is (almost) never necessary to go back
to the formulas: it suffices to see geometrically that a projective transformation with certain properties exists. Two such applications are described
in Exercises 2.17 and 2.18 both will be useful later.
Notes
All the basic facts about polarity, the face lattice, and the various parts
of the representation theorem and their proofs are classical, due to Farkas,
Weyl, Minkowski, Carathodory, Motzkin, Kuhn, and others. Again we refer to Griinbaum [234] and Schrijver [448] for the history. Anyway, "history
will teach us nothing" (Sting).
(This was a message from our No Comment department.)
70
2 Faces of Polytopes

2.0 Let P be a polyhedron and let F be any nonempty face of P which is
minimal with respect to inclusion. Taking x 0 E Fo, show that Fo - zo
is a linear subspace, and that F0 - xo = lineal(P).
Thus, if P has lineality space lineal(P) . {0}, then every minimal
nonempty face is a vertex and in particular, P has a vertex.
2.1
Use the previous exercise combined with Theorem 1.2 to formulate

and prove analogues of the Representation Theorem 2.15 for polyhedra. Special attention is needed for the formulation of part (5). In
particular, what do you get in the case where P is a cone?
2.2 Show that every polytope is affinely isomorphic to a bounded intersection of an orthant with an affine subspace.
2.3 Construct a small poset that satisfies the conditions of Theorem 2.7
but does not correspond to a convex polytope. Does your example
correspond to some geometric object?
2.4 Prove directly (i.e., without using polarity) that every face of a polytope P is contained in a facet.
2.5 If two 0/1-polytopes are combinatorially equivalent, does it follow
that they are affinely isomorphic? (The answer is "no.")
2.6 Let f(d) be the number of combinatorial equivalence classes of d-dimensional 0/1-polytope,s. The first values are f(0) = f(1) , 1,
f(2) = 2, and f(3) = 8. Prove that 2 2d-2 < f(d) <22d for d> 5.
(This solves a problem of Billera & Sarangarajan [72, Sect. 3]. For the
lower bound A. Sarangarajan and I suggest that you consider all the
polytopes of the form P = cony(S) for sets S C {0, 1}d that satisfy
x E S for all x E {0, l} d with xd = 1,

0, 1 - ed E S, and
e l , 1 - ed el V S.
-
There are 22d-1-4 such polytopes P(S); show that their combinatorial
equivalence classes are "small.")
2.7 Assume that one is given the vertex-facet incidence matrix
Al(P)
E 10, 11"1 " a-- f.,,,cynxn
of a convex polytope P with m vertices and n facets.

How can the face lattice of the polytope P be uniquely reconstructed
from the knowledge of M(P) alone? How does the dimension of P
appear in the computation? How does your algorithm fail if the matrix you apply it to is not the vertex-facet matrix of a polytope?
What is the relation between the matrices of P and P?
71
2.8 Let P and P` be polytopes with vertex sets V = {v 1 , .. , vr,,1 and

V' =
Assume that for every vertex set F C V of a facet
of P, the corresponding set F' C V' is the vertex set of a facet of P.
Show that for every face of P, the corresponding vertex set of P`
forms a face in P. Deduce that dim(P) < dim(P').
Lemma: if P and P' have the same dimension, then they are
combinatorially equivalent under yi
(Hint: Assume this fails. Then P' has two facets, Ff and
which are adjacent, such that of the corresponding vertex sets
in P, the set F1 forms a facet, but F2 does not. (Here you can
use the fact that the graph of (P') is connected: a proof is in
the next lecture.) Now consider the ridge Fl n II which is a
facet of
and use induction on the dimension, applied to the
polytopes F1 and
(iii) Show that P and P' need not have the same dimension.
(Hint: For this, one can take a cube P = C3, labeled as in the
drawing, and a cyclic polytope
C4(8).)
5
6
(iv) Assume that of a polytope P you are given the dimension, the
vertex set, and a matrix M(P) e 10, 11'" such that every row
of M(P) represents a facet of P.
How can you tell whether this list of facets is complete?
(Remark: this is not too easy; one can use tools from Chapter 8,
or from homology theory.)
(Part (ii) is important: See Klee & Minty [302, p. 167], Amenta &
Ziegler [17], and elsewhere. (iii) points to an error in [302, p. 167].)
2.9 Define the face figure PI F for any face of P by Pf F (F*) , that
is, a polar of the face of P which corresponds to F. (The face figures
F are also known as the quotients of P. Thus a quotient of P is
the same thing as an iterated vertex figure.)
Show that this is a polytope of dimension
dim(P/F) ---= dim(P) dim(F) 1.

Characterize the face lattice of a face figure in terms of the face lattice
of P and of the element F E L(P).
Describe a more direct construction of PIF, generalizing the case of a
vertex figure. How can the face figure PI F be obtained as an iterated
vertex figure?
72
2.10 In conditions (iii) and (iv) of the characterization of the interior of a

polytope (Lemma 2.8), can we assume the xi to be vertices?
2.11 Show that if {v 1 , ... , v k } C vert(P) is a set of vertices of P, then
F = {vi } V ... V {vk } holds in L(P) if and only if lk-. E
vk E
relint(F). Generalize to get a formula for the join of a set of faces
{G I , ,Gk} C L(P).
2.12
Compute directly that every polar of a simplex is a simplex.
2.13 Define the polar of a cone by

C := { c C (Rd )* : cx < 0 for all x E C}.
Show that this definition (with "0" instead of "1") is a special case
of our definition for arbitrary subsets.
Formulate and prove the analogs of our Theorems 2.11 and 2.12.
2.14 Let P be a d-polytope in
given by the system
P = P(A,z).
An inequality in this system is called redundant if deleting it from the
inequality system Xx < z does not change the polyhedron; otherwise
the inequality is called irredundant.
(i) Derive from Farkas lemma III that an inequality is redundant if
and only if it can be written as a positive combination of other
inequalities in the system.
(ii) Derive from a Farkas lemma that if P = P(A,z) (6 and if none
of the inequalities of a system Ax < z is redundant, then each
of them defines a facet. Thus, every polytope is the intersection
of its facet-defining inequalities.
(iii) Show that if x, E relint(F) is a point in the relative interior of a

facet F E MP), then the inequality ax < z defines the facet F
if and only if it is valid for P, and ax, = z.
(iv) Show that for every facet F of P, there is a unique inequality
ax < z which defines F, and for which Ecil , lad = 1.
Together, these statements prove that any irredundant description of
a d-polytope P C Rd as an fl-polytope contains exactly one inequality for each facet of P.
Show that this statement could also be derived by polarization of
Proposition 2.2.
What happens in the situation when dim(P) < d? How much of the
uniqueness statement in part (iii) can you rescue?
73
2.15 Let P = conv(V) C Rd be a convex d-polytope, and assume that

some description of P as an ii-polytope is known,
P =- P(A, z).
With every inequality ax < zi in this system, associate its vertex
set
Vi
E V : at v = zi l.
Show that the following criteria can be used to check whether an
inequality is redundant.
(i) The inequality ax < z is redundant if and only if 3/4 C Vi for
some j i.
(ii) If 14 = Vi, then either the inequalities are multiples of each
other, or they can both be deleted from the system.
(iii) An inequality is irredundant if and only if it defines a facet of P
and no multiple of it is contained in the system.
(iv) If 13/41 < d, then the inequality ax < z -
redundant.
(A) The inequality ax < zi is irredundant if and only if there is no

multiple of it in the system, and the rank of the matrix given
by {(v1 ) : v E Vil is d.
(Parts (i) and (v) yield complete criteria for redundancy, which can be
checked explicitly. Note that there was no assumption that the set V
has to be minimal. Condition (i) is (equivalent to) the main condition
of Chernikova [142], while condition (A) is a rank test that seems not
so efficient. For example, a combination of criteria (i) and (iv) makes
sense in practice.)
2.16 Let P = conv(V) C 1d be a convex d-polytope, and assume that an

irredundant description of P as an 7-t-polytope,
P = P(A, z),
is known, such that the inequalities ax < zi describe all the distinct
facets of P, without duplication. For every inequality ax < zi in the
system, let 3/4 be the set of points in V which satisfy it with equality.
Show that from this, an irredundant description of proj d (P) C
can be obtained from the following criteria:
d-1
(i) if aid = 0, then aix < zi determines a facet of proj d (P),

(ii) if
aid
> 0 and aid < O, then the inequality

aidai + ( ai k)ai
aidzi
( aik)z i
defines a facet of projd (P) if and only if 3/4 n Vi c Vk holds for

no k i,j.
74
Furthermore, show that if I Vi n vi < d 1, then the combined inequality in (ii) is redundant. In particular, this is the case if d > 2 and
n Vi = 0. (Give geometric proofs they are easier than algebraic
ones!)
Explain how, by using Fourier-Motzkin elimination (Theorem 1.4)
together with these redundancy criteria, one can obtain an complete
irredundant description of P
conv(V) C Rd, even if the set V
contains more points than just the vertices of P.
How can the criterion be adapted for the case of polyhedra, where
the input is a polyhedron given as P =-- conv(V) + cone(X)?
What happens in the situation where dim(P)
difficulty there be overcome?
d, and how can the
(The necessary and sufficient criterion can be found, for example, in

Burger's version [133, Thm. 3] of the double description method [386].
The test on "Vi n V. 0" is a heuristic in Chernikova [142].)
2.17 Show that if P C Rd is a polytope with two distinct vertices u, y, then
there is a projective transformation P P` such that the vertices
u' and y' have the smallest, respectively the largest, xd-coordinate
among all vertices of P'.
2.18 Let P C Rd be a polytope, let F be a facet of P, and let r :

Rd' be a projection map (for example, deleting the last coordinate).
By "moving a point beyond F to infinity," show that by a projective
transformation P
P', we can obtain r(P') r(F1) and 7r -1 (G')n
P' = 71--1 (G') fl F' for every proper face G' c r(F').
75
2.19 Using a Farkas lemma, show that for every unbounded pointed polyhedron P there is an inequality ax < 1 such that
:= {x E P : ax <1}
is a polytope with a facet F' :=
E P : ax = 11, such that the
k-faces of P correspond to the unbounded (k+ *faces of P, and the
k-faces of P' that are not faces of F' are in bijection with the k-faces
of P.
Show that a polyhedron combinatorially equivalent to P' can also

be obtained from P by taking the closure of the image of P after
a "nonadmissible" projective transformation that moves the face at
infinity into Rd.
How can this be used to study the combinatorics of pointed unbounded polyhedra in terms of "polytopes with a distinguished face"?
2.20 Let P E Rd be a d-polyhedron with n facets and at least two vertices,
and assume that 0 E int(P). With P, associate the new polyhedron
P :-= (conv(vert(P )\0)) .

Show that if P is a polytope, then 0 is an interior point of P, and
po
= p.
If P is unbounded, then 0 is on the boundary of P. Show that

the polar, with respect to an interior point of conv(vert(P )\0), is
a d-polytope P with n facets and with more vertices than P.
2.21 The Caratheodory curve [135] in led is given by
1/cos( u)
R
It(
y(u) :=
sin( u)
cos(2u)
sin(2u)
cos (du)
sin(du) /
(0
Show that for 0 < u1 <u2 <

hull
qd(ttlIU21 - Un)
 2d, the convex

cony {y(ui),Y(u2),
,Y(ud)}
is combinatorially equivalent to the cyclic polytope C2d(n)(Hint: You will find a useful hint in Griinbaum [234, p. 67, Ex. 23].)
76
(ii) Prove that in the case d ,----- 2 the map:
(n)
2y2 + y4
1
1-4
3 + 4y1 y3
y4
2y2 y4
3 4yi y3
is a projective transformation, which maps (the relevant part of) the

Carathodory curve to the moment curve.
Conclude that for suitable real parameters t 1 < t2 < ... < t n , the
polytope C,14 (u l , ... , un ) is in fact projectively equivalent to the "standard" cyclic polytope C4(t1, . , t.) defined via the moment curve
(Example 0.6).
(In fact, for general d > 1 the polytope qd (u 1 ,u2, ... , un ) is projectively equivalent to some standard cyclic polytope C2d (ti 1 1 ta ).
For this one can explicitly construct a projective transformation that
takes the Carathodory curve to the moment curve, using a substitution of the type
t
:=
1 cos(u)
sin(u)
sin(u)
= 1 + cos(u)
and some elementary trigonometric identities, such as the formulas

sin(2t) = 2 sin(t) cos(t) and cos(2t) ---, 2 cos(t) 2 1.)
2.22
For relatively prime natural numbers p,q E N (i.e., numbers with no

common factor), define the bicyclic polytope P4 (p,q,n) as the convex
hull of the n > 5 points cos(2pr-) \
vi :=
sin(2p/r)
cos(2q71- -7 )
sin(2q7r)
for 1 < j < n.

Describe the geometry and combinatorics of these polytopes.
(i) What symmetries do they have? Show that all the facets of
P4 (PI q, n) are combinatorially equivalent.
(ii) How many facets does P4 (p, q, n) have? Find the conditions on
p, q, and n under which P4(p, q, n) is simplicial.
(iii) Using part (i) or (ii) of the previous exercise, show that the
4-polytopes P(1,2, n) and P(1, V, n) (for odd n) are combinatorially equivalent to cyclic polytopes C4 (n).
(Smilansky [466, 4671)
2.23 Try to estimate the number of combinatorial equivalence classes of
d-dimensional polytopes with n vertices.
(Goodman Si Pollack [217, 218], Alon [11 ])
3
Graphs of Polytopes
The vertices and edges of a d-polytope P form an undirected graph G(P)

that encodes a lot, but not everything, about the combinatorial structure
of the polytope.
In this lecture we discuss three fundamental topics about graphs of polytopes: the monotone Hirsch conjecture (for which we prove validity for
0/1-polytopes, and Kalars recent bound for the general case), Kalai's reconstruction of simple polytopes, and Balinski's d-connectivity theorem.
Before we look into this, we will establish two technical tools: the power
and the glory of half-sentences like "let L be a line in general position," and
the (geometric version of the) simplex algorithm for linear programming,
which is the most fundamental search technique on polytopes.
3.1
Lines and Linear Functions in General Position
We start with a short discussion of the concept of "general position with

respect to P"
this will yield a few useful tools for proofs in subsequent
sections. To illustrate this, we will sketch a quick-and-dirty geometric version of linear programming in the next section.
With all the work we did in Lectures 1 and 2, we can now assume that we
are dealing with a d-dimensional polytope P in I'd, in which 0 is an interior
point. We alternate freely between representations in terms of vertices,
P = conv(V), and in terms of facets, P = P(A,1). Here A G R"d is a
matrix that is considered as a set of rows, A = fai,...,an}-
78
3. Graphs of Polytopes
We also need the hyperplanes in Rd determined by the facets of P, for

which we introduce the notation
:=
Ix E
: ai x = 1}
thus the facets of P are given by Fi = H

Hi determines two halfspaces, where
HI :=
n P. Each of these hyperplanes
E Rd aix <1}
denotes the closed halfspace that contains P, and 11;1- denotes the other
closed halfspace. In particular,
n H2 n n H-
As a start, for every face F E L(P) we know how to find a point

relint(F) in the relative interior of F: for example, we can take the
barycenter of the set of vertices of F.
A very useful thing to get is a point outside P but very close to F
this is used for stellar subdivisions (see Exercise 3.0), Schlegel diagrams
(see Lecture 5), and many other constructions. In particular, z, does not
lie on any of the hyperplanes Hi ; it lies "in general position" with respect
to the hyperplane arrangement (i d, ... ,l/n 1) determined by P.
For this, let F be a proper face of P, and define y E Rd to be a point
beyond F if y and 0 E int(P) lie on different sides of Hi for every facetdefining hyperplane Hi that contains F, but on the same side of H for
every facet-defining hyperplane H that does not contain F.
In other words, y lies beyond F if it satisfies ai y > 1 for every inequality
that is valid with equality for F, but ai y <1 for every other inequality.
Our sketch shows a convex polygon (2-polytope) and indicates a vertex y, a
point yv beyond that vertex, an adjacent edge E, and a point YE beyond E.
Yv
How do we find y F ? Well, we can take it on the ray that emanates

from 0 and goes through a relative interior point x p . In fact, we can take
yF := tx, for any t > 1 such that ai (tx,) < 1 whenever 0 < aix F < 1.
Clearly this t can be found (explicitly).
The next object we sometimes need is a line "in general position."
3.1 Lines and Linear Functions in General Position
79
Definition 3.1. A line through 0 E int(P) is in general position with

respect to P if it is not parallel to any of the hyperplanes Hi and does not
hit the intersection of any two of them.
If the line is written in the form L(u) = {tu : t E } (for some u L 0), then
general position means that au 0 and au au for all 1 < j < n.
The following lemma shows that a direction vector for such a line can be
found arbitrarily close to any given vector.
Let P = P(11,1), and let u E Rd\O. If A > 0 is small
enough, then the line gu(A)) is in general position with respect to P, for
Lemma 3.2.
( A \
u
A:d
Proof. We use that si (A)
au") = Edk=1 at k(uk + Ac) is a nonvanishing polynomial in A of degree at most d, which has at most d positive
zeroes. The polynomials si (A) are distinct, since ai ai for i j.
From this we get that ai u( A) 0 0 for all, except at most d, positive values
of A, and that aiu(A ) ai u(A ) for all positive values with not more than
(3)d exceptions.
LI
A remark about being explicit mathematicians might tend to use either

topological arguments ("a finite set of hyperplanes is nowhere dense in Qd")
or unnecessary algebraic machinery ("let x ij be a set of d n independent
transcendentals over the ground field"). For discrete problems like those
posed by our polytope applications, this is unnecessary.
The construction of Lemma 3.2, which we use to find points, lines, etc.,
in general position, depends on perturbation by some A > 0 that has to be
chosen small enough. It is not hard to be even more precise and completely
explicit: it is (in principle) easy to compute a bound Ao such that every
A with 0 < A < Ao is small enough. This is because for polynomials with
rational coefficients, one can bound the positive zeroes away from O. A good
reference for the ideas used for such explicit bounding is LovAsz' lecture
notes [341, Ch. 1].
(The small positive parameter A that we need would usually be called E l
but that would make this look like a course in analysis. We'll try to avoid
this until Lecture 9, where we start to integrate over polytopes.)
The same method of proof as used in Lemma 3.2 also yields the existence
of a hyperplane in general position, arbitrarily close to a given one. The
way to get this is to perturb the coefficients of the linear function that
defines the hyperplane. For this, we state the result, but skip the proof.
Definition 3.3. A linear function ex on d is in general position (or
generic) with respect to a polytope P C
if it separates the vertices
of P, that is, if cvi cvi for any two distinct vertices vi , vi of P.
80
Graphs of Polytopes
Lemma 3.4. Let P = P(A, 1), and let c E ( )*\ O. If A > 0 is small
enough, then the linear function c(A) x is in general position with respect
to P, for
c(A) := c + (A, ) 2 , ... , Ad ).
3.2 Directing the Edges

("Linear Programming for Geometers")
Definition 3.5. Let P be a convex polytope. The vertices and the edges
of P form an abstract, finite, undirected, simple graph, called the graph
of P and denoted by G(P).
For every face F E L(P), we denote by G(F) the induced subgraph
of G(P) on the subset vert(F) c vert(P) of the vertices of G(P), that
is, the graph of all vertices in F, and all edges of P between them. This
coincides with the graph of F, if F is itself considered as a polytope.
(In this whole course we need very little graph theory, only some terminology. When in doubt, look it up in any graph theory book. For that
purpose, even a mediocre book would do. As for good ones, we recommend
Bondy Sz Murty [116 ] , Bollobils [11.4 ] , or Tutte [5141.)
We will consider orientations of G(P), which assign a direction to every
edge. An orientation is acyclic if there is no directed cycle in it. This implies
(because all our graphs are finite) that there is a sink: a vertex that does
not have an edge directed away from it. (Proof: Start at any vertex, and
keep on walking along directed edges until you close a directed cycle or get
stuck in a sink.)
Linear programming is (in a geometer's version) the task to find a point
xo E P that maximizes a linear function cx, that is, such that czo =
max{cx : x E PI =: co. Now we easily see that the maximum is achieved
in a vertex. In fact, Fo := Ix E P : ex = co l is a face of P, and thus every
vertex of Fo maximizes cx.
FO
3.3. The Hirsch Conjecture
81
Dantzig's simplex algorithm [166] in its "first phase" finds a vertex y of P.

Then it proceeds to find a better vertex w that is a neighbor of v. We use
N(v) to denote the set of neighbors of v, that is, the set of all w E vert(P)
such that conv{v, w} is an edge of P. This improvement step is iterated
until the algorithm stops at an optimal vertex.
I
Now if c is in general position, then this gives us a well-defined way to
direct the graph of P, by directing an edge convfyi , vi I from vi to vi if
cvi < m3 . (Because of the general position assumption, ties cannot occur.)
We call this the orientation of G(P) induced by e.
With this construction monotone paths on P (edge paths for which the
objective function increases strictly in each step) translate into directed
paths in the orientation of G(P) induced by c.
Lemma 3.6. Let y E vert(P) be a vertex, and let N(v) be the set of its
neighbors in G(P). Then the cone (based at y) spanned by the neighbors
of y contains P:
P C + cone{u v : u E N(v)}.
82
3 Graphs of Polytopes
Proof. This follows from our proof of Proposition 2.4: the neighbors of y
are in one-to-one correspondence with the vertices of the vertex figure P/v,
and thus it is equivalent to say that those vertices span a cone that contains
P. But we have also seen there that every ray emanating from ty to any
other point x E P contains a point of the vertex figure. This yields
C {v + t(u - v) : u E Pity, t > 0 }

El y + cone{u - y : u E vert(P/v)}
y + cone{u - y : u E N(v)}.
III
Theorem 3.7. If ex is a linear function in general position for P, then

the orientation of G(P) induced by e is acyclic, with a unique sink. This
sink is the unique point in P where ex achieves its maximum.
Proof. Along any directed path vo, v 1 , ... , vk in G(P), the value of cx
increases strictly. Thus a directed path cannot return to its starting vertex,
and there are no directed cycles. Therefore the induced orientation of G(P)
is acyclic, and it has a sink.
Now assume that y is a sink: then all of its neighbors w E N(v) satisfy
cw < cv. By Lemma 3.6 this implies that cx < cv holds for all x E P
with x 17; that is, y maximizes cx over P, and it is the only point in P
CI
that achieves the maximum.
This proves that for any starting vertex 27 E vert(P), and for any linear
function cx that is in general position with respect to P, every strictly increasing edge path will eventually lead to the unique vertex that maximizes
cx over P.
With this crude description, the problem of linear programming is, of
course, not solved. This starts with the fact that, for efficient treatment,
we have to consider bases and pivots instead of vertices and edges. (See
Exercise 3.10 for a brief sketch.) Here we run into problems of degeneracy
if the polytope is not simple, or if the linear function is not in general
position. One way to treat this is through "perturbation," implicitly or
explicitly. For example, if we know an interior point (this is not a natural
assumption for practical problems!), then we can rewrite P as P(A, 1), and
then (implicitly rather than explicitly) optimize over P(A, 1A ) for small
enough A > 0, which is nondegenerate. This leads to lexicographic pivot
rules; see ChvAtal [151, pp. 34-36]. Furthermore, to construct a simplex
algorithm we have to determine "which edge to take"; this leads to the
question of pivot rules. All this is combinatorial geometry. Later in the
game, numerical questions dominate the picture.
Anyway, this discussion was only meant as a sketch of the geometric
situation a very simple and special picture of the world according to a
discrete geometer.
83

The diameter of a graph G will be denoted by b(G): the smallest number b.
such that any two vertices in G can be connected by a path with at most
b edges.
For n > d > 2, let d(d, n) be the maximal diameter of the graph of
an d-dimensional polytope P with at most n facets. Similarly, let A ii (d, n)
denote this maximal diameter in the unbounded case, for a d-dimensional
pointed polyhedron P with at most n facets (n > d > 2). For example,
7-1 j,
A(2,n) = [ '
Au (2, n) --..: n 2.
Our sketch illustrates the extreme cases for d = 2 and n = 8.
It is a long-standing problem to determine the behavior of the function

A(d,n). The value of A(d, n) is a lower bound for the number of iterations
needed for the simplex algorithm with any pivot rule. Thus the question
of whether .6. (d, n) grows polynomially in n and d is closely related to the
question of whether there is any pivot rule for which the simplex algorithm is a strongly polynomial algorithm for linear programming; see [299,
Sect. 3].
A notorious, very specific, question connected with the graphs of polytopes was first posed by Warren M. Hirsch in 1957 (see Dantzig [166,
pp. 160, 168]) and has become known as the Hirsch conjecture.
Conjecture 3.8 (Hirsch conjecture).
[166, p. 168]
For n> d > 2, let 11(d, n) denote the largest possible diameter of the graph
of a d-polytope with n facets. Then
A (d, n) < n d.
Is this plausible? Here are a few observations, most of them due to Klee
Sz Walkup [303].
84
The Hirsch conjecture is true for d < 3 and all n (even in the monotone and unbounded versions discussed below, by Klee [2931), and for
n d < 5, by Klee & Walkup [3031.
One can show that for Conjecture 3.8 it is sufficient to consider simple
polytopes (see Exercise 3.5).
e If n < 2d, then any two vertices lie on a common facet. From this we
get d, n) < 6(d 1,n 1); iterating this, we get
A(d,
<(n d, 2(n d))
for
n < 2d.
Similarly, we get A u (d, n) < At, (n d, 2(n d)). In both cases these
inequalities hold with equality [303 ] : this is quite obvious in the unbounded case. Thus we restrict our attention to the case n> 2d.
More surprisingly [3031, the Hirsch conjecture for all dimensions would
follow if one could prove it for n = 2d for all dimensions. The special
case n =z 2d has become known as the d step conjecture.
Consider two vertices that do not lie on a common facet. Since each
of them lies on d facets, we see that the d-step conjecture concerns
a very special geometric situation: after a change of coordinates we
can assume that the first vertex y is given by y = 0, where the facets
it lies on are given by x, > 0, which describes the positive orthant
x > O. Then the other vertex u can be assumed to be u = 1, and its
facets describe an affine image of the positive orthant.
-
In this situation there are d edges leaving from u, whose other endpoints are on the hyperplanes {x : x, = 0 } . The claim is that we can
get from u to V in d steps.
From the special case of the d-cube we get that A(d, 2d) > d. Thus
the bound suggested by the d-step conjecture is certainly the best
possible, if it holds. Furthermore, Holt & Klee [2581 have shown that
A(d, n) > n d
for n > d > 13,
that is, the Hirsch conjecture is also best possible for all n, if the
dimension is high enough.
85
If you look for counterexamples, a natural guess would be to consider the polars of cyclic polytopes C d(n) 4 ' , or more generally the
polars of neighborly polytopes since they have the largest numbers of vertices for given n and d (according to the upper bound
theorem; see Section 8.4). However, Klee [297] has shown that the
polars of cyclic polytopes satisfy the Hirsch conjecture. Beyond that,
Kalai [277] could prove that if P is the polar neighborly d-polytope
with n facets, then one has at least a polynomial diameter bound
(5(G) _<,_ & (n d) d log(n).
The nonrevisiting path conjecture, due to Victor Klee and Philip

Wolfe, states the following: for any two vertices u, y of a (simple)
polytope, there is a path from u to y that does not revisit any facet
it has left before.
To illustrate this conjecture, the following drawing shows the graph of
a simple 3-polytope with nine facets (due to Barnette [36]) in which
for two vertices u and y the unique shortest path (of length 3) makes
a revisit:
However, there is a nonrevisiting path: just follow the boundary of

the figure.
It is easy to see that the nonrevisiting path conjecture implies the
Hirsch conjecture. In fact, the starting vertex of a nonrevisiting path
lies on at least d facets, and with every vertex the path reaches at
least one new facet it hasn't visited before. Thus the length of a
nonrevisiting path cannot be more than n d.
The nonrevisiting path conjecture may seem much stronger than the
Hirsch conjecture. However, Klee Sz Walkup [303] proved that the two
conjectures are in fact equivalent.
The convexity assumption is essential: the Hirsch conjecture is false
for some topological cell complexes that are combinatorial spheres,
as Mani Sz Walkup [350] demonstrated. It is also false for simplicial
2-manifolds, see Barnette [45].
86
Klee & Walkup [303] showed that the Hirsch conjecture is also false
for unbounded polyhedra although Hirsch's original conjecture
was asked for unbounded polyhedra. They proved that for n > 2d,
A u (d, n) > n d l_df5j. This is the best lower bound known for
.A.,i (d, n).
Even stronger, the monotone Hirsch conjecture is false, as Todd [507]

demonstrated: it is not true that if ex is a linear function on P and y is a
vertex, then there is a monotone path with at most n d edges from y to
a vertex Yu. of P that maximizes ex.
In fact, consider any d-polyhedron P C Rd with at most n facets, and
let ex be a linear function. To avoid complications, we will assume for the
following that the linear function ex is in general position with respect
to P, that ex is bounded on P, and that the polyhedron is pointed (i.e.,
it has a vertex, and its lineality space is lineal(P) = {0}). From these
assumptions we get that there is a vertex u of P on which cx achieves its
unique maximum.
Now define Hu (d, n) to be the smallest number such that in the situation
above, for every vertex y of P, there is a (strictly) monotone path from y
to the top, that is, a path from y to u along which ex increases in every
single step. Similarly, let H(d, n) be the same number under the additional
assumption that P is a polytope.
The monotone (bounded) Hirsch conjecture would require that
Hu (d, n) <n d,
respectively
H(d, n) <n d.
Disproving that, Todd [507] showed that

n
d min{ H
d]
i ,
In
dl
< H (d, n) < Hu (d, n).
In particular, there is a 4-polytope with n = 8 facets for which every monotone path to the top needs at least five steps. However, in Todd's example
there is a two-step nonmonotone path, which first goes to the bottom, and
then directly to the top! This motivates the following, more restrictive,
version of the monotone Hirsch conjecture, which might as well be true
and which would imply the Hirsch conjecture (via a simple argument using
projective transformations; see Exercise 2.17).
Conjecture 3.9 (Strict monotone Hirsch conjecture).
Let P be a d-dimensional polyt ope with n facets, and let ex be a linear
function that is in general position with respect to P.
Then there is a strictly increasing path with respect to ex, from the
(unique) vertex v min that minimizes cx, to the (unique) vertex vmax that
maximizes ex, of length at most n d.
87
V max
CX
Vmin
To illustrate this for a trivial case, observe that for an n-gon the length
of a shortest monotone path "to the top" can be n 2 =-n), but if
we start "from the bottom," then we need at most 1 121-i
n) steps.
n)?
What about upper bounds on A(d, n) and
In 1967 Barnette [36, 234] proved that A u (d, n) < n3 3 . An improved
bound, Au (d, n) < n2d-3 , was proved in 1970 by Larman [321]. Barnette's
and Larman's bounds are linear in n but exponential in the dimension d.
After that, nothing happened for a long time. In short, we might summarize
the history by saying that the experts thought that the conjecture was
plausible until they tried to prove it and couldn't; therefore now they think
it is false, and can't prove that. However, in the long run Kalai might prove
to be right, when he writes about "the author's guess (which is as good as
the reader's)" [278]. The existence of a polynomial (or even linear) bound
for A(d,n) is still a major open problem...
However, recently Gil Kalai achieved a substantial breakthrough: in a
sequence of papers (each simpler and more striking than the preceding
one) he established the first subexponential bounds for the diameter of a
polytope. In November 1990 he proved flu (d, n) < n2V7L [278, Sect. 3].
In March 1991 he derived a "pseudopolynomial" bound for the diameter
problem [278]:
A v (d,n)
< n2 log2 (d)-1-3 .
A substantial simplification, which also strengthened the result slightly to

Au(d,n) < nig2(d)2 , was subsequently found by Kalai & Kleitman [282].
The proof we give here is (essentially) the modification of this proof given
by Kalai in [279, Sect. 2]. It is equally (surprisingly!) simple but establishes
a stronger result: the existence of a "pseudopolynomial" monotone path to
the top.
Theorem 3.10. (Kalai [279, Sect. 2])

Let P C
be a d-dimensional polyhedron with at most n facets, and let
cx be a generic linear function which achieves its maximum on P in the
vertex w.
88
Then from any starting vertex y G vert(P), there is a monotone path to

the top vertex w, whose length is bounded by
Hu (d, n) <
2n (d -I- [1og2 71]<

1)
2 niog,(d)+1
d1
2 (2d) 10g2 (n)
Proof. The key to this is the notion of an active facet: given any vertex y
of a polyhedron P, and a linear function ex, a facet of P is active (for y)
if it contains a point that is higher than y (that is, either the facet is
unbounded with respect to ex, or it has a top vertex w with cv < cw).
For this proof, we also admit problems for which ex is not bounded
on P, and where the last step "to the top" takes a ray (unbounded 1face) on which ex has no upper bound. (You may think of the top as an
extra vertex u,, in this case, which is adjoined to the directed graph of the
problem.)
Let ft(d, n) be the number of steps that may be required to get to the
top vertex if we start from a vertex y for which the polyhedron has at most
n active facets (and an arbitrary number of nonactive ones!).
Since Hu (d, n) is monotone in n we immediately get
A (d, n) < Au (d, n) < Hu (d, n) < H (d, n).
Thus it suffices to prove the bounds of the theorem for il(d, n). In the
following we require d > 2 and n > O. In the "boundary cases" we get
1(2, n) = n
(all the edges on a monotone path to the top are active facets, and this
may be all of them if the problem is not bounded), and
H (d, 0) = H (d, 1) =---
= fl (d, d 2) = 0,
(if y is not the top vertex, then it has an increasing edge, which lies on
d 1 active facets).
To get a recursion for R(d, n), we verify a sequence of four simple facts:
Given any set T of k active facets of P, we can reach from y either the
top vertex, or a vertex in some facet of T, in at most H(d, nk) monotone
steps.
Let "Ax < z" be a minimal system that defines P (having one inequality
for each facet of P), and let P' := P(A', z') be the polyhedron obtained
by deleting the inactive constraints that don't contain y as well as all the
inequalities that correspond to facets in T. Then y is a vertex of P' (unless
y lies on a facet in T, in which case we have nothing to prove), and it has
at most n k active facets in P'.
Now consider a shortest path from y to the top in P'. This path makes
at most H(d, n k) steps, by definition. If it touches a facet in ,F after at
1.
89
most fi(d, n - k) steps on P, then we are done. If it doesn't, then the top
vertex of P, is also the top vertex of P, and the path to it in P' also yields
a path to the top vertex on P, of length at most 11(d, n - k).
2. If we cannot reach the top in _TI (d , n-k) monotone steps, then the collection g of all active facets that we can reach from y by at most fI(d,n - k)
monotone steps contains at least n - k + 1 active facets.
If there are k facets that cannot be reached, we can delete these facets
together with all the inactive facets, and get a problem where we can reach
the top in at most ft(d,n - k) steps; however, the path in this reduced
problem corresponds to the same path in the original problem, leading to
the same top vertex: Contradiction.
3. Starting at y, we can reach the highest vertex Iv () contained in any facet
F E g within at most H(d,n - k) + f1(d - 1, n - 1) monotone steps.
We need at most 17/(d, n - k) steps to reach any facet of g; this facet (of
dimension d - 1) has at most n - 1 facets, thus in it we can find a path to
its top of length at most 1-1(d - 1,n - 1).
4. From w o we can reach the top in at most H(d, k - 1) steps.
This is because none of the facets in g is active for inch and thus wo has
at most n - (n - k + 1) = k - 1 active facets.
Putting the monotone paths together, we get a bound
H (d, n) < H(d, n - k) + H (d - 1,n - 1) + H(d, k - 1)

for the shortest monotone path from y to the top.
Now we choose k := [3-1. Using the fact that by definition fir(d, n) is a
(weakly) increasing function in n, we get
H(d, n) < 11(d - 1, n - 1) + 2H (d,21

r--]).
This recursion reminds us of the recursion for binomial coefficients and
we make a substitution to transform it into that. For this, we define
f (d, t) := 2- t 11(d, 2t )
for t > 0 and d> 2,
and with this substitution the recursion simplifies to "what we want":
f (d, t) 5 f (d - 1, t) + f (d, t - I).

From the boundary conditions f(2,t) = 2 -t H(2,2t) = 2 -t 2t -,-- 1 = ( tV)
for t > 1 and f (d, 0) = H(d, 1) ----- 0 = ( i) for d > 3, we obtain
f(d,t) <
-
d + t - 3)
d-2)
90
for (d, t) (2,0), by induction on t > 0 and d > 2. From this we derive
H(d,n) < H(d, n) <.
H(d,21+Ll0g2 mi )
-.=
2 1+ Llog2 n -I f(d,
<
2n
<
2n (d - i) 1g2(n) =
1 Llog2 n j )
(d + Llog2 n j 2)
d 2
2 n i-Flog2 (di) ,
for n, d > 2, using the inequality ( t. ()) < (a -I- 1) b , which follows by induction over a > 0 and b> 0.
(In fact, there are various ways to derive bounds on fl(d,n) from the
recursion. This is a standard type of gymnastics for which you should get
training at your "analysis of algorithms" class. Here is another way to proceed, which obtains the original Kalai-Kleitman bound. We use the starting
values H(2, n) = n and ft(d, 0) = 0. Since H(d,n) grows monotonically
in n, we get a simple recursion
H(d, n) < H(d 1, n) 2H(d,[;1)

for n> 0 and d > 3. This we can iterate, to get
d
H(d,n)
<
i=3
<
<
n + 2(d 2) - (2d) l0g(n/2)

2d. (2d) 10g(n)-1 = (2d) b0g(n) ,
using n < 41g2 n < 4 (2d) 1g2 n-1 .)
It would be tricky and probably unnatural to formulate this proof in

such a way that it stays within the family of polytopes: even if P is a
polytope, the polyhedron P' will not, in general, be bounded. This is why
this theorem and proof were done in the generality of polyhedra.
Also the proof does not stay within the class of polyhedra P with only
2 dim(P) facets, as considered by the d-step conjecture (see Exercise 3.7).
However, we can specialize the result to fit this situation, and get A(d, 2d) <
( 2d)l0g2 d+1. In fact, in the special case of n = 2d one can modify/sharpen
the computation of upper bounds to get
A(d,2d) < egg d+2 ,

according to Kalai [279]. Still, this is far away from the conjectured bound
of A(d, 2d) = d.
What's the problem? Why can't we do much better? There is some evidence in Matouek's work [354] that the above analysis is essentially the
3 3 The Hirsch Conjecture
91
best possible, that is, any proof for a substantially better upper bound
has to use more of the specific geometry of the problem. Not much of the
geometry was used in the preceding proof. (In fact, Kalai [278, Sect. 4]
indicates a very general abstract framework, of a "simplicial complex with
a fixed shelling order" (see Lecture 8), in which such upper bounds can be
proved.)
Finally, let us mention that the diameter bounds can indeed (not quite
directly) be used to construct algorithms for linear programming. In his
research, Kalai [279, Sect. 3] found randomized pivot rules for linear programming that roughly require an expected number of n4 `fd arithmetic
operations for every linear programming problem of dimension d with n
facets. See Exercise 3.9(ii) for a simple sketch, and [281] for the latest version.
Very similar results were reached independently and nearly simultaneously (on a completely different path) by Matouek, Sharir & Welz1 [355],
in the setting of a "dual simplex method."
For 0/1-polytopes (Example 0.11), the Hirsch conjecture is quite trivial

however, it took quite a time until Naddef [389] realized this A more
general result that also bounds the diameter of integral polytopes was given
by Kleinschmidt & Onn [308]. We will give a slightly sharpened form of
Naddef's theorem.
Before that, here are some examples of 0/1-polytopes P in Ile: for each
of them we list the space dimension d, the dimension k = dim(P) of the
polytope itself, and the diameter 6(G(P)).
/ /
d=2
d=2
d=2
dim(P) = 1
dim(P) = 2
dim(P) = 2
6(G(P)) =1
b(G(P)) = 1
S(G(P)) -- 2
d=3
d=3
d=3
dim(P) = 2
S(G(P)) = 2
dim(P) = 3
S(G(P)) = 2
dim(P) = 3
6(G(P)) -= 3
92
Theorem 3.11. Let P = conv(V) be a 0/1-polytope, V C {0, 1} d . Then

P satisfies the Hirsch conjecture. In fact, the diameter of G(P) is bounded
by
b(G(P)) < dim(P),
with equality if and only if P is affinely isomorphic to a regular cube.
Proof. Let P have two vertices

symmetry of the cube
y,
u of distance 8(u, y) > d. We use the
Id := [0,11d = conv({0,
to reduce to the case where y = 0 and u E {0, 1}d.

Using induction on the dimension d we can assume that P is full-dimensional: otherwise let ax = z be an equation that is valid for P. We get
z = 0 from 0 E P, and thus a
O. By permuting coordinates we may
assume ad O. Then the projection map
it : Rd
d-1
1-0'
(Xxd)
(deleting the last coordinate) maps the 0/1-polytope P C Rd to an affinely

isomorphic 0/1-polytope 7r(P) C r d-1 . Thus we may assume P C rd with
dim(P) = d.
Now assume that ui = 0 for some i. Then 0 and u are both vertices of
the face F(i) := Pn
E Rd : xi = 0} of P, which corresponds to the valid
inequality xi > O. Thus we get
5(0,u) 5_ 6(G(F( ))) < d - 1

by induction on d. Therefore we may assume that u = 1.
Now if any neighbor w E N(1) of 1 has k > 1 components that are 0,
then we get
8(0,1) < 5(0,w)
,1) < (d - k) + 1 < d,
where we use that the face
Fw P n E
= 0 whenever wi 01
has diameter at most d - k, by induction.

Thus if 8(0, 1) > d, then all the neighbors of 1 have exactly one 0component. Since 1 has at least d neighbors (see Lemma 3.6), we find that
N(1) = {1 - e z : 1 < i < ei}.
Also, again considering the faces
P(i) = P n fx E
xi
= 01
3 4 Kalai's Simple Way to Tell a Simple Polytope from Its Graph
93
of P, we get that 0 and 1ei have distance d l in G(F( i)), so by induction

on d we get
F(i) = conv{x E {0, } d :
xi = 0}.
Collecting an the vertices that we now know have to be in P, we get P =

conv({0, 1} d) = Id and 6(G(P)) = d.
LI
The bound 6(G(P)) < dim(P) can also be proved (with the same kind
of argument) in the monotone version, where we ask for the shortest path
"to the top" with respect to a given linear function. If we restrict to the
strictly monotone version of Conjecture 3.9, then the characterization of
the equality case also remains valid (with the same proof).
3.4 Kalai's Simple Way to Tell a Simple Polytope

from Its Graph
In this section we consider simple polytopes and their graphs. Our treatment is based on a striking (and strikingly simple) paper by Gil Kalai.
To be honest the situation is even worse: the following is copied quite
directly from his paper "A simple way to tell a simple polytope from
its graph" [272].
Let P be a simple d-dimensional polytope and let G(P) be the graph
of P. Thus, G(P) is an abstract graph defined on the set of vertices vert(P)
of P. Two vertices y and u in vert(P) are adjacent in G(P) if [y, u] is a
one-dimensional face of P.
Perles [405] conjectured the following result.
Theorem 3.12. (Blind Sz Mani [102])

If P is a simple polytope, then the graph G(P) determines the entire combinatorial structure of P.
In other words, if two simple polytopes have isomorphic graphs, then
their face lattices are isomorphic as well.
Proof. Here is Kalai's [272] simple proof of this result.

We consider the set of all acyclic orientations (i.e., edge orientations with
no oriented cycles) of G(P). We will not distinguish between an orientation
0 of G(P) and the partial order induced by 0 on vert(P), which is defined
by y <0 u whenever there is an 0-directed path from y to u.
Note that if 0 is an acyclic orientation of G(P), then the restriction of
G(P) to any nonempty subset A of vert(P) has a sink (an element with
out-degree zero) with respect to O.
An acyclic orientation 0 of G(P) is called good if for every nonempty
face F of P, the graph G(F) has exactly one sink. Otherwise, 0 is bad.
94
Graphs of Polytopes
The existence of good acyclic orientations of G(P) follows from Theorem 3.7: if cx is in general position for P, then it is also for all faces
of P. Our first goal is to distinguish intrinsically between good and bad
orientations of G(P).
good (one sink)
bad (two sinks)
Let 0 be an acyclic orientation of G(P). Let 14), be the number of vertices

of G(P) with in-degree k in O. Define
ff) := h(03 + 21t? + 4h? + - - - + 2 k h(e + - - - + 2 d 1q.

If x is a vertex of G(P) of in-degree k with respect to 0, then x is a sink in
2k faces of P. (Since P is simple, every i edges incident to x determine an
i-face F of P which includes them.) Let f denote the number of nonempty
faces of P. Since each face has at least one sink, we see that
I.) fo > f, and

II.)
0 is good if and only if f = f.
To distinguish between good and bad orientations from the knowledge of

G(P) only, compute f for every acyclic orientation O. The good acyclic
orientations of G(P) are those having the minimal value of 1 0 .
Now we will show how to identify the faces of P. The graphs of simple kpolytopes are k-regular: they have exactly k edges incident to every vertex,
by Proposition 2.16. With this, the criterion is very simple: an induced
connected k-regular subgraph H of G is the graph of some k-face of P
if and only if its vertices are initial with respect to some good acyclic
orientation 0 of G(P). Indeed, if F is a face of P, it is well known that
vert(F) is an initial set for some good acyclic orientation: a set of vertices
such that no directed edge leads into the set. For this just consider a linear
function with respect to which the vertices of F lie below all other vertices,
which can be obtained by choosing a linear function cx that defines F, and
perturbing it according to Lemma 3.4.
On the other hand, let H be a connected k-regular subgraph of G(P)
and let 0 be a good acyclic orientation with respect to which vert(H)
is an initial set. Let x be a sink of H with respect to O. There are k
edges containing x in H, all oriented toward x. Therefore x is a sink in
the k-face F that contains these k edges. Since the orientation 0 is good,
x is the unique sink of F, and therefore all vertices of F are < x, with
95
respect to O. But vert(H) includes the set of all vertices that are < x with
respect to O. (Remember: vert(P) is an initial set with respect to O.) Thus,
vert(F) c vert(H). Since both H and G(F) are k-regular and connected,
vert(F) = vert(H) and G(F) = H. This completes the proof.
Remarks 3.13.
You could ask: do these parameters h? actually mean anything? They
do
we will come back to this when we study shellability of polytopes, in Lecture 8.
1. We do not have a practical way to distinguish between good and
bad orientations. The algorithm suggested by the preceding proof is
exponential in Ivert(P)I. We do not even know of an efficient way to
compute the number of facets of P from G(P).
2. It is false that general polytopes can be reconstructed from their
graphs this can be seen, for example, from the existence of neighborly (simplicial) polytopes.
Perles [405, 407] proved that simplicial d-polytopes are determined
by their [d/2]--skeleta; see Kalai [280]. General d-polytopes are determined by their (d-2)-skeleta, and this is best possible even for quasisimplicial polytopes (all of whose facets are simplicial polytopes); see
Grfinbaum [234, Ch. 12].
0.

A very fundamental fact about the graphs of d-polytopes is that they are
d-connected, a theorem due to Balinski [33].
Here we use the definition that a simple graph G(P) is d-connected if the
removal of any d 1 or fewer vertices (and all the edges they are incident
with) leaves a connected graph.
The theorem is certainly plausible, since it is easy to see (using Lemma
3.6) that every vertex of G(P) has degree at least d. We have adapted the
following simple proof from Grfinbaum [234]. Two different proofs appear
in Brondsted St Maxwell [127] and in Barnette [46]. Two extensions, which
answer the questions of how many components the graph may have if you
remove k vertices, or if you remove a k-face, appear in Klee [294] and in
Perles St Prabhu [409].
Theorem 3.14 (Balinski's theorem).
[33]
The graph G(P) is d-connected for every d-polytope P.
Proof. Let P = conv(V) C Rd , where the vertex set V of P and of the

graph G(P) has at least d+1 elements. We delete a subset of d-1 of them,
S = {v 1 ,...,vd_ 1 } C V; then we have to show that the graph G(P)\S
induced on the remaining vertices is connected.
96
. -1 vz E P denote the barycenter of the vertex set S.

Let s := 1 d-1E dt=1
We know by Lemma 2.9(i) that s is contained in the relative interior of a
face F. We consider two cases.
Case 1. If s is contained in a proper face F E L(P)\{P}, then all points
V i E S are also contained in this face F this is the usual computation.
Let ex < c0 be a valid inequality that defines F. Then co is the largest value
that ex can achieve on P, while the smallest value is some go < co . In this
case, every vertex in V\S either lies in the face Fo = { x E P: ex = go }, or
it has a neighbor whose ex-value is smaller (this follows from Lemma 3.6),
and which therefore also lies in V\S. Thus every vertex in V\S has a
decreasing path, within V\ S, which connects it to a vertex in Fo . Finally,
the graph of Fo is connected, by induction on d.
Case 2. If s is contained in the interior of P, then we choose a linear
function ex on d such that the hyperplane { x E Rd : ex = co } contains
both S and at least one other vertex vo E V\S. This is possible because
every set of d points is contained in a hyperplane.
Now let cm ax and emir, denote the largest and the smallest value, respectively, that ex takes on P, and let Frna.), and Fmin denote the corresponding
faces. Then the graphs G(Fmax ) and G(FiTiiii ) are again connected, by induction. Every vertex v E V\S is connected either by a strictly cx-increasing
path to Fmax (if it satisfies cv > co), or by a strictly decreasing path to
Fmin (if it satisfies cv < co). Finally, the extra vertex vo is connected to
both Fmax and Frnin , so the whole graph G(V)\S is connected.
C
Notes
The graph of a polytope is treated with care in Griinbaum's book [234,
Chapters 11, 13 and 16].
As for linear programming, this is usually described in a much less geometric way, which is better suited for algorithmic treatment. Also, there is
of course much more to say than our simplified sketch in Section 3.2. We refer to the books by Dantzig [166], Chvg,tal [151], Schrijver [448], Grbtschel,
LovAsz Sz Schrijver [228], Padberg [404], and Borgwardt [118], and to [92,
Ch. 10] for various different aspects of the matter. As for "Dantzig's simplex algorithm" [166], let us just mention that it was already developed by
Kantorovich in the 1920s, but could not published in any reasonable form
for reasons that were equally ideological and stupid [285].
Klee & Kleinschmidt [299] is an inspiring survey on the Hirsch conjecture and its relatives; see also Kleinschmidt [307]. The material in Section
3.3 is derived from the papers by Kalai Sz Kleitman [278, 282, 279]. In
particular, the proof of Theorem 3.10 is from Kalai [279, Sect. 2]. Our discussion of 0/1-polytopes is based on ideas by Naddef [389] and by Kleinschmidt [308]. The observation in Theorem 3.11 that the extreme case here
is only achieved for d-cubes seems to be new (although not deep).
97
In a recent preprint, Lagarias, Prabhu Sz Reeds [319] discuss the configuration space of all d-step configurations for a fixed d, analyze its structure,
and relate the d-step problem to certain factorization problems for matrices.
They also suggested that there might in fact be at least 2c1-1 paths of length
d between the complementary vertices of any d-dimensional Dantzig figure.
However, Klee & Holt [257] have now shown that this is true for d < 4, but
false for all d> 4.
Our treatment of the reconstruction of polytopes from their graphs owes
heavy thanks (thefts) to the paper [272] of Gil Kalai, as indicated there.
We might repeat here that the graph of a polytope carries important information, but by far not all the relevant information about the structure of a
polytope. One aspect of this is the fact that the graph does not determine
the dimension of a general polytope; see also Exercise 3.4. Another one is
that there are far fewer different polytope graphs for various parameters
than there are different polytopes. In fact, Perles proved that the number
of nonisomorphic graphs of d-polytopes with d + k vertices is bounded by a
function of k (independent from d!). The proof for that see Kalai [280]
uses only some lemmas about finite set systems. In contrast to Perles'
result one can easily see (for example with the methods of Section 6.5) that
the number of different d-polytopes with d + 2 vertices (i.e., k = 2) is not
bounded.

3.0
(Stellar subdivisions [192]). Let F be a facet of the d-polytope

p c d and construct a point y F E Rd beyond F. The polytope
st(P, F) := conv(P U
1)
is the stellar subdivision of P at F.

(i) Describe the faces of st(P, F) in terms of faces of P. Conclude
that the combinatorial type of st(P, F) does not depend on the
precise position of yF .
(ii) Show that if P is simplicial, then so is st(P, F). In this case,
count the number of k-faces of st(P, F) in terms of the numbers
(P) of i-faces of P.
(iii) Describe the operation that is "polar" to stellar subdivision,
given by
st (P,
:= (st(P , y )) ,
for any vertex y of P.

3.1
For a vertex y of the d-polytope P, and k > 1, construct the cones

generated by all vertices of P of distance k from V:
Ck := conetw V: tv E vert(P), 6(y,
= kl.
98
Prove the "nested cones theorem" of Hochstattler [253]:

Ci D C2 D C3 D ....
3.2 If P has dimension at least 4, then the graph G(P) is not planar. In
fact, show that it contains a subdivision of the complete graph Kd+1
(Griinbaum [234, pp. 200, 214]).
3.3 If n> d > 4, then the graph of the cyclic polytope Cd(n) is complete,
G(Cd (n)) = K. Give a direct proof: for each edge, construct an
explicit linear function that is maximized by this edge.
3.4 A d-polytope P is called dimensionally ambiguous if there is a polytope Q of a different dimension dim(Q) dim(P) which has an
isomorphic graph, G(P) '='-- G(Q).
Show that the d-simplex is dimensionally ambiguous for d > 5,
but not for d < 4.
(ii) Show that 3-polytopes, and simple 4-polytopes, cannot be dimensionally ambiguous. (Hint: Use Exercise 3.2!)
(iii) Show that if P is a 0/1-polytope whose graph is isomorphic
to G(Cd), then P is affinely isomorphic to Cd. (Compare to
Exercise 2.5!)
(iv)* Are the d-cubes dimensionally ambiguous? In particular, is there
a 4-polytope whose graph is isomorphic to G(C5 )?
(This is a problem of Kalai. Babson, Billera & Chan [29] proved
that for every r > d there exists a cellular d-sphere that has the
Ld 1 J-skeleton of Cr . In particular, there is a cellular 3-sphere
that has the graph of C5. However, it is not clear whether such
an example can be realized as a convex polytope.)
(i)
3.5 If P = P(4,1) is an irredundant description, show that for small

enough A > 0 the polytope
P` -= P(A,1 (A) ),
with 1 (A) , = 1+Ai as in Lemma 3.2, is a simple polytope whose facets
are in natural bijection with the facets of P.
Furthermore, show that then 45(G(11)> S(G(P)): thus it is sufficient
to prove the Hirsch conjecture for simple polytopes.
3.6 If P is a pointed polyhedron in R3 , show that the graph of all bounded
edges is connected. Show that it is not 2-connected in general. What
about higher dimensions?
3.7 Let P c Rd be a d-polytope with 2d facets, such that the facets
containing y = 0 determine the positive orthant u > O. Show that
the facets of P can have 2d 1 facets each, if d > 4. (This is why it
is hard to use inductive arguments for the d-step conjecture.)
99
3.8 Prove the following theorem by Kleinschmidt & Onn [308[: if P is a

d-polytope whose vertex set is contained in 10, 1, ... , kld , then the
diameter of its graph is bounded by
b(G(P) < k d.
Why can't this be used to get effective bounds for the diameters
of d-polytopes? (See Problem 4.16* and Section 6.5(a) for answers.)
3.9 Consider the simplex algorithm, applied to a linear function c on a

simple, d-dimensional polyhedron with at most n facets, such that
there is a unique optimal vertex.
(i)* The EDGE-RANDOM rule moves along random increasing edges,
where at any given vertex the increasing edges leaving it are
taken with equal probability. Can you give any subexponential
upper bound (in n and d) for the expected number of steps of this
rule on any linear programming problem? Is there a polynomial
upper bound?
(ii) Assume that we use the following RANDOM-FACET pivot rule to
choose the increasing edge. That is, at the starting vertex y,
if up-degree(v) = 0, then STOP the current vertex is optimal,
if up-degree(v) = 1, then take the unique increasing edge,
if up-degree(v) > 1, then take a random facet among facets that
contain v, restrict the linear program to that facet, and solve
the restriction by a recursive call to RANDOM-FACET.
(This is what Kalai in [279] calls an "antipivot rule," or the
"bureaucratic" rule.)
Show that the maximal expected running time of this algorithm
can be bounded by a function E(d,n), which for n > d satisfies
the recursion
E(d,n)
<
max { 1 + E(d, n 1),

max{d,nd}
E(d 1, n 1) + ji
E(d,n i)
}-
(These recursions are in fact not hard to see, the derivation of

the asymptotics implied by this recursion, E(n,d) _< nì'V i , is not
that easy (i.e., difficult). See Matouek, Sharir & WelzI [355 ] for
a careful treatment of the asymptotic analysis.)
(Both rules are studied in Gartner & Ziegler [204]. In particular,
for special linear programming problems it is shown that there are
starting vertices for which the expected number of steps is (nearly)
quadratic. See also [279].)
100
3.10
z) C
(Basis version of linear programming). Let P
and let cx be a linear function on Rd . A subset of d of the inequalities
of P, say Aix < is a basis if A' has rank d (equivalently, A'x = z'
has a unique solution x E Rd ). A basis A'x < z' is feasible if the
unique solution of A' x = z' satisfies x E P, and dual feasible if it
maximizes cx over Ix E Rd : < (equivalently, if 0 maximizes
< OD.
cx over Ix E Rd :
(i) If y is a vertex of P C Rd , show (using Carathodory's Theorem
(ii)
(iii)
(iv)
(xi)
1.15) that there is a feasible basis that has y as its solution. If

P is simple, the basis is unique.
Show that all the feasible bases for a vertex of P are connected
by sequences of single-element exchanges; that is, in every such
exchange one inequality from the system A'x < z' is replaced
by a single different one from the big system Ax < z.
If E is an edge of P adjacent to y, show that there is a feasible
basis for y such that all but one of the inequalities A' x < z' are
satisfied on E with equality.
Show that if a basis is both feasible and dual feasible, then it is
an optimal solution for the program max ex, Ax < z.
Use a Farkas lemma to prove that if P is a nonempty polytope,
then there is a basis that is both feasible and dual feasible.
3.11* For a d-dimensional polytope with n facets, what is the maximal

number M(d, n) of vertices in a monotone path?
(This M(d, n) is an upper bound on the largest number of steps in
any simplex algorithm. It is known (see Klee & Minty [302]) that
[d/2]
M(d, 2d) > 2d , and roughly that M(d,n) >
We suggest an answer in Problem 8.41* the upper bound given
We
by the upper bound theorem may be sharp! This would contradict
the suggestion by Klee & Minty [302, p. 175] that there could be a
function 7 such that M (d, n) < -y(n d)d.)
3.12* Does every simple 4-dimensional polytope have a Hamilton cycle?
(This conjecture is due to Barnette; see [191, p. 158]. Some special
cases are in Barnette & Rosenfeld [50].)
3.13* Let P be a simple d-polytope. Is it true that every connected (d 1)regular subgraph that does not separate G(P) is the graph of a facet
of P? (This is a problem of Perles; see [272].)
3.14 A k-path between distinct vertices y and w of a d-polytope P is a
sequence of k faces F1 , . Fm such that V is a vertex of F1 , w is a
vertex of Fm , and Fi and Fi+1 are adjacent for 1 <j <m (i.e., their
intersection is a (k 1)-face). Two k-paths are disjoint if they have
no k-face in common.
-
101
Derive from Balinski's theorem that for k ---= 1 and for k = d 1

there are d disjoint k-paths between any two vertices y and w.
(ii) If P is a simplex, show that there are (dk ) disjoint k-paths between any two vertices y and w.
(iii)* Are there (kd ) disjoint k-paths between any two vertices y and
w of any d-dimensional polytope?
(i)
(The problem is due to Prabhu [412], from 1990; no progress, yet.)
3.15 A complex A is a collection of d-subsets of the set [n] := {1, 2, ... , n}.
(This defines an "abstract simplicial complex" in the sense of Section 8.5.) Two sets F, G E A are adjacent if they differ only in
one element. This defines a graph on the d-sets in A, and A is
called strongly connected if this graph is connected. Even stronger,
A is called ultraconnected if every nonempty subfamily of the form
AK := {FE A.K CFI is strongly connected.
(1) Show that the complex which corresponds to a simplicial dpolytope with vertex set (identified with) [n] is ultraconnected.
(ii) If P is a simple d-dimensional polyhedron whose set of facets is
labeled by [n], then there is a d-set associated with every vertex.
Show that the corresponding complex is ultraconnected.
(iii) Every shellable simplicial complex is ultraconnected. (Shellability is an important combinatorial concept: see Section 8.1 for
the definition.)
More generally, a pure simplicial complex is shellable if and only
if it has an ordering F1 , F2 ,. .. , F8 of its facets such that the subcomplex F1 UF2U...UFs is ultraconnected for all i.
(iv) Let A be the complex of the 4-sets 1234, 2345, 1346, 5678, 2678,
1578, and all the 4-sets that have two elements from 1234 and
two elements from 5678 but do not contain both 1 and 2 nor
both 5 and 6. Show that the distance between 1234 and 5678 in
A is 5. Show that A is ultraconnected.
(N) Describe an ultraconnected complex of triangles (i.e., d = 3) on
n vertices with diameter n 3.
(vi) Let A be an ultraconnected complex of triangles on n vertices.
Show that between any two triangles there is a path of triangles
which visit every vertex at most twice. Deduce that the diameter
is at most 2n.
(vii)* Can you improve 2n to 1.999n (or at least to 2n-1000, 2n-1)?
Can you find an ultraconnected collection of triangles on n vertices with diameter > 1.001n? (or at least n+100, or even n-2?)
(This combinatorial set-up for studying diameter questions is due
to Larman [321] and to Kalai [278]. In particular, in [321] Larman
showed that between every two vertices of an ultraconnected complex of d-sets on n vertices, there is a path that visits every vertex at
102
most 2d-1 times, and this implies a bound of A u (d, n) <2d-mn. (See
also Klee & Kleinschmidt [299, Sect. 7].)
The unbounded 4-polyhedron with 8 facets by Klee & Walkup [303],
which fails the Hirsch bound, yields the complex of part (iv).
Subexponential diameter bounds for ultraconnected complexes were
found by Kalai, see [278, Sect. 4.1]. This exercise is also due to him.
Part (vil)* demonstrates the unbelievable gap between the known
lower and upper bounds. Note that by parts (i) and (ii), every upper bound that one can prove for the diameter of an ultraconnected
complex of d-sets is automatically also valid for Au (d, n).
3.16 Let P be a simple d-polytope such that every k-face of P has at most
2k facets.
(i) Show that the diameter of P is bounded above by d
(ii) Moreover show for such polytopes that for every objective function and any starting point one can "reach the top" in d steps.
(This is from Kalai [278, Thm. 3], where it is proved that for every
fixed r > 2, if every k-face of P has at most rk facets, then the
diameter and the hight of P are bounded by a polynomial in d.)
3.17 Given finite graphs G and H, we define that G is an induced subgraph
of H if we can obtain a graph isomorphic to G by deleting a set S of
vertices (and all edges adjacent to them) from H. We say that H is
a suspension of G if additionally we require that the vertices of S are
connected to all other vertices of H.
Show that every finite graph is an induced subgraph of the graph
of a 4-polytope.
(If G has n > 5 vertices, then start with C4(n), and introduce
an extra vertex beyond every edge that is missing in G.)
(ii) For every finite graph there is some suspension which is the
graph of a d-polytope, for some d.
(If G has n > 5 vertices, then start with C4(n), and introduce
an extra dimension and two new vertices for every edge in G
that is supposed to be missing.)
(ill)* Does every finite graph have a suspension that is the graph of a
4-polytope? What about the case where G is the graph with n
vertices but no edges?
(iv) Give an example of a 4-connected graph on n > 5 vertices which
is not the graph of a 4-polytope. Can you construct a 4-regular
graph with these properties?
(i)
(Perles [4081)
4
Steinitz' Theorem for 3-Polytopes
The combinatorial structure of 2-polytopes is not much of a mystery. For an

illustrated journey into the wonderful world of 3-dimensional convex (and
nonconvex) polyhedra, we direct the reader's attention to the conference
volume [456]. See also Barnette's book [41] for a nice elementary discussion
of combinatorics and graph theory related to convex 3-polytopes.
In this lecture we will establish the basic theory for 3-polytopes, by
proving Steinitz' theorem. The basic version reads as follows.
Theorem 4.1 (Steinitz' theorem).
[487, 490]
G is the graph of a 3-dimensional polytope if and only if it is simple, planar,
and 3-connected.
Polytopal graphs are certainly simple: they have no loops or multiple
edges. The graph G(P) is planar for every 3-polytope (use radial projection to a sphere from an interior point, or a linear projection to a plane
from a point beyond a facet). Also, G(P) is 3-connected by Balinski's Theorem 3.14. Thus, the difficult part is the "if" part of the theorem: it requires
that we show how, given a 3-connected planar graph, one can construct a
3-polytope.
4. Steinitz' Theorem for 3-Polytopes
104
Here are four observations to indicate to this claim is nontrivial:

1. No similar theorem is known, and it seems that no similarly effective
theorem is possible, in higher dimensions.
2. There are lots of interesting consequences and various strengthenings
that follow by the same proof technique (see Section 4.3).
3. Many of the higher-dimensional analogues of these strengthenings are
false (we'll see this in Lectures 5 and 6).
4. There is no extremely simple proof known.
All the combinatorial ("classical") proofs of Steinitz' theorem [487] [490,
54,63] [234, Sect. 13.1 1 [47] basically follow the same pattern: arguing
that every 3-connected planar graph can be "built up" from K4 by some
well-defined operations, which preserve realizability. (See the notes at the
end of this lecture for a different, "nonlinear" line of reasoning.)
The reason why we can give a "nicer than usual" combinatorial proof
here is that we will be more elegant in dealing with the graph theory:
using Truemper's clever treatment [509, 510] of "AY reductions." This will
be done in the next three sections of this lecture, which together imply
Theorem 4.1.
In Section 4.1, we discuss the little graph theory we need, concentrating on 3-connected graphs and AY reductions.
In Section 4.2, we prove that Steinitz' theorem is true for 3-connected
planar graphs which have a "AY reduction."
In Section 4.3, we show that if a 3-connected planar graph G has a
AY reduction, then so does every minor of G.
Finally, we show that every planar graph is a minor of a grid graph,
and every grid graph has a AY reduction.
We close the lecture in Section 4.4 with a list of strengthenings, extensions,
and corollaries of Steinitz' theorem.
4.1
3-Connected Planar Graphs
Again we need some basic graph theory. Let us, for a while, admit nonsimple
graphs as well, that is, graphs that can have loops and parallel edges.
Such a graph G is connected if there is a path between any two distinct
vertices of G. All the graphs we consider will be connected.
A graph G with at least 2 edges is 2-connected if it is connected, has
no loops, and cannot be disconnected by removing one vertex and all the
4 1 3-Connected Planar Graphs
105
edges incident with it. A graph G with at least 4 edges is 3-connected if it

is simple and cannot be disconnected by removing 1 or 2 vertices from G.
Under this definition, the smallest 2-connected graph is the graph C2 with
two parallel edges
and the smallest 3-connected graph is K4,
the complete graph with 4 vertices and 6 edges.

The reason for this version of the definitions is that it is invariant under
duality. That is, if we embed a graph into the sphere S2 and draw the
dual graph G*, then G is k-connected if and only if G* is k-connected (for
k =-- 1, 2, 3). We do not review the construction of a dual graph here, but
trust that the following picture which you may interpret as being drawn
in the plane, or on the 2-sphere explains it all.
-
At this point, observe that the combinatorial structure of a 3-polytope

is completely determined by its graph this is a special case of a fact
we noted in Remark 3.13(2). It follows from Whitney's theorem [527] that
the embedding of a 3-connected planar graph into the sphere is unique.
To prove it, note that for every 3-connected planar graph the regions of
an embedding are bounded by chordless cycles that do not separate the
graph. Any other cycle has more than one region on either of its sides, so
either the cycle has a chord, or it separates two vertices, or both. Note that
Perles' question of Problem 3.13* asks for a higher-dimensional version of
Whitney's theorem.
106
Two very basic "local" operations on graphs are the deletion of edges
-----' >
>
and the contraction of edges,
>
>
for which the two vertices of the edge are identified. Any graph that can
be obtained from G by a sequence of deletions and contractions of edges

is called a minor of G. Note that the edges of a minor can be viewed as a
subset of the edge set of G.
A special case occurs if we contract only edges that are in series with others, being adjacent to a vertex of degree 2 (this is equivalent to "removing
a subdivision point")
>
{ -->. >
or delete edges that are in parallel with others (this is the usual operation
for "making a graph simple").
--->. >------____ ....-----?f
We will refer to any sequence of such operations as series-parallel reductions, or SP reductions.

A Delta-Wye operation, or AY operation, replaces a triangle that bounds
a face (i.e., a nonseparating triangle) by a 3-star that connects the same
vertices, or vice versa. If we want to specify the direction of the transformation, then we will call it a A-to-Y transformation, respectively a Y -to-A
transformation.
f
We refer to the figure for the "natural" correspondence between the edges
of the triangle and the edges of the 3-star. Note that these operations,
replacing a K3 by a K1,3, preserve the number of edges in the graph.
4 2 Simple AY Transformations Preserve Realizability
107
However, a AY transformation might create series or parallel edges, which

can then be SP-reduced.
We note here a simple lemma, characterizing when connectivity is preserved under Y-to-A operations.
Lemma 4.2.
(i) Let G be a 2-connected graph, and let {e, f, g} be the edges at a
vertex y of degree 3 in G.
If none of its edges are parallel (i.e., if y has three different neighbors),
then the result of a Y-to-A operation is again 2-connected.
(ii) Let G be a 3-connected graph (in particular, there are no parallel
edges; all vertex degrees are at least 3) that is not K4. Let { e, f, g}
be the edges at a vertex y in G of degree 3.
If we perform a Y-to-A operation on this 3-star, and then delete all
parallel edges created by this (i.e., all edges that originally connected
neighbors of y), then the resulting graph is again 3-connected.
Proof. This follows directly from the definitions: for this consider a Y-to-A
operation G
G'. Then for any set of one or two separating vertices
in G', the same set is separating in G as well. The only problem is that a
Y-to-A operation can create parallel edges; if we delete them, the operation
preserves 3-connectedness.
El
The nice thing now is that, because of duality, we immediately get a dual
statement, Lemma 4.2*, about connectivity after a A-to-Y transformation.
For this we use that under duality, we have
embedded planar graph G
contracting series edges
k-connected
nonseparating triangle
A-to-Y transformation
dual graph G*,

deleting parallel edges,
k-connected,
3-star,
Y-to-A transformation.
This duality can be carried into our reduction for polytopes, because the
graph of a 3-polytope is exactly the dual graph of the polar polytope:
G(P)* = G(P).
4.2
Simple AY Transformations Preserve

Realizability
By a simple AY reduction we mean any AY operation followed immediately

by all the SP reductions that are then possible. By Lemma 4.2 and its dual,
108
these reductions preserve 3-connectedness, if applied to any 3-connected

graph other than K4We get four different types of simple A-to-Y reductions: for this we
consider a triangle and distinguish whether it has zero, one, two, or three
vertices of degree 3.
Here the dotted lines in our sketch denote edges that may or may not be
present, and are not affected by the simple A-to-Y reduction.
Similarly, we get four types of simple Y-to-A transformations when we
consider a vertex y of degree 3 and distinguish how many of its neighbors
are already connected:
4-
AAAA
These four transformations are exactly the "polar operations" (operations in the dual graph) for the simple A-to-Y reductions.
Lemma 4.3. Let G be a 3-connected planar graph, and let the graph G'
be derived from G by a simple AY transformation.
If G' is the graph of a 3-polytope, then so is G.
109
Proof. By duality, respectively polarity, we have to treat only the four

types of A-to-Y transformations. For these, the transformation from P' to
P just corresponds to "cutting off a vertex" by some suitable plane.
(To visualize this, consider our sketch of the four types of simple A-to-Y
transformations, interpreting them as pictures of 3-polytopes_)
ii
4.3
Planar Graphs are AY Reducible
In this section we show that every 3-connected planar graph (with n > 4
edges) can be reduced to K4 by a sequence of simple AY transformations.
This is a special case of a much more powerful theorem (for 2-connected planar graphs plus a "return edge") that was first established by Epifanov [186]
and has a clever and simple proof by Truemper [509]. In this section, we
follow his expositions in [509] and [510, Sect. 4.3].
For a while, we will only require that the graphs considered are 2connected; in particular, we admit AY operations if they keep our graphs
2-connected.
In the simplicity of Truemper's approach to Epifanov's theorem, the
reader should appreciate the "power of a normal form theorem" (in this
case: the embedding of a planar graph as a minor of a grid graph, which is
also at the heart of Robertson Sz Seymour's work [428] on graph minors).
Here come three lemmas and a corollary, which together prove everything. [Working through this, you can practice "three levels of reading":
first read only the lemmas, and try to understand what they mean and
how much is trivial_ Second glance over the proofs, and try to see whether
you know how to do them yourself. If you know, just do it. If you don't,
try to find counterexamples. In the third step, work your way through the
proofs until you are confident that you found all the errors I made and the
shortcuts I missed. Tell me about them.]
110
For the following, we will call a 2-connected graph G AY -reducible if it

can be transformed into the graph G2 with two parallel edges by a sequence
of AY transformations and SP reductions.
Lemma 4.4. If a planar graph G is AY-reducible, then so is every

2-connected minor H of G.
Proof. We use an induction on the number of reduction steps that are

necessary to reduce G. We can assume that H has no series or parallel
edges, otherwise we can make the corresponding reductions. Now if the
reduction for G starts with a series-parallel reduction step, then H is a
minor of the reduced graph as well, because it does not contain both of the
series or parallel edges.
Therefore, we can assume that the reduction of G starts with a AY
reduction step. Exploiting duality, we may assume that this is a A-to-Y
step G -- G'.
G':
Let e, f, g be the three edges of G that are involved. One possibility is that
all three edges are contained in H: then they form a nonseparating triangle
in H as well, and we can perform the corresponding A-to-Y step H ---> H'.
Then by induction we get that H' is AY-reducible, and hence so is H.
In the other case, some of the three edges e, f, g do not appear in H.
What happened to them? Since H is simple, it is not possible that only
one or two of them were contracted. If all three were contracted, then we
can assume that first one edge was deleted, then the others were contracted.
Using that deletions and contractions commute, and possibly relabeling, we
can thus assume that the first edge that disappears when H is formed from
G, say e, is deleted.
But then we get the same minor H from G', by contracting the corresponding edge from G', because the deletion of e from G and the contraction
of e in G' result in the same graph. Again we are finished by induction. 0
Denote the grid graph with mn vertices and m(n-1) n(m-1) edges by
G(m, n). Clearly the grid graphs G(m, n) and G(n, m) are isomorphic.
111
Our sketch shows the grid graph G(5, 6).
Lemma 4.5.
If G is planar, then it is a minor of a grid graph.
Proof. For this, fix an embedding of G into the plane IR2 ; now split the
vertices of G in order to get a graph G' of which G is a minor, such that
all vertices of G have degree at most 3.
k- k
Then we can construct an embedding of G' (with subdivided edges) into

a finite grid that is combinatorially equivalent to the given embedding (in
the sense that the vertices at a given edge come in the same cyclic order).
For that we draw one edge at a time on the grid. Whenever our grid is too
coarse, we can refine it by taking half the grid size.
11
8
10
10
5
Once a subdivision graph G' is represented as a subgraph of the grid graph

this way, it is clear that G is a minor of the grid graph.
0
This is quite trivial. Much more sophisticated versions of this, restricting
the size of the grid we embed into, are interesting and important for applications in VLSI-layout problems. We refer to Lawler et al. [322, Sect. 3.4.5 ],
1 331, Ch. 5], and Kant [284] for further information as well asLengaur
more precise and sophisticated versions of such "grid layout" and "graph
drawing" results.
112
Lemma 4.6.
All grid graphs G(m,n) with m,n > 3 are AY-reducible
to K4Proof. We use two basic observations. First, if an edge connects two neighbors of a vertex of degree 3, then we can delete it, by performing first a
A-to-Y transformation and then a series reduction.
Second, if an edge connects two neighbors of a vertex of degree 4, then we

can move the edge over to "the other side," using first a A-to-Y transformation and then a Y-to-A transformation.
Using these two observations, we can reduce any grid graph G (m, n) to K4,
as follows.
First perform a series reduction in the upper-right corner. Then, assuming that m > 4, we take the two edges in the lower-left corner and perforn
a series reduction to get a single edge. We can move this single edge across
the whole grid in a sequence of degree-4 moves until we "hit the boundary."
There either the edge is parallel to the diagonal corner edge and we can
parallel-delete, or it can be deleted according to our degree-3 move.
4 4 Extensions of Steinitz' Theorem
113
This way we have deleted the first square in the last row; similarly, we can
delete the second square in this row, and so on; the last square is deleted
by two series reductions and one parallel reduction.
11.
1111111.11
EMILE
We can delete all the squares in the first column of our board symmetrically,
if n > 4.
II
.
.'
A-
At the end, this leaves us with the grid G(3, 3) with one short corner, which
is easily reduced first to the "wheel" graph W4 of a square pyramid, and
then to IC4 .
0
Corollary 4.7. Every 3-connected planar graph G can be reduced to K4
by a sequence of simple AY transformations.
Proof. The three lemmas together show that G is AY-reducible. We follow
this reduction to the first point where parallel or series edges are created.
These can be reduced immediately, which also results in a 3-connected
planar graph G by Lemma 4.2. The graph we now have has fewer edges:
hence we are done by induction on size.
Corollary 4.7 also completes our proof of Steinitz' Theorem 4.1.
4.4
Extensions of Steinitz' Theorem
Corollary 4.8. For every 3-polytope, there is a rational 3-polytope that

is combinatorially equivalent.
Proof. This follows from our proof of Steinitz' theorem: we have shown
in fact that every 3-polytope can be reduced to a simplex by two types of
operations P > P':
constructing a combinatorially polar polytope,

"cutting off" a vertex of degree 3.
In both cases, if we have a rational realization of P', then we can also
construct P with rational vertices and inequalities.
0
114
Corollary 4.9. Every 3-connected planar graph has a representation in

the plane such that all edges are straight, and all the bounded regions
determined by it, as well as the union of all the bounded regions, are convex
polygons.
Proof. This we get by representing G as the graph of a 3-polytope, and
then choosing a point x beyond one of the facets. "Viewing" the polytope
from this perspective (and projecting it to the facet along the "rays of
vision") gives us the required representation. 0
This proof contains the construction of a "Schlegel diagram" for the

3-polytope: see the next lecture. There are various results in graph theory related to this. There are other proofs, like the one by Tutte [512],
which use less geometry and more graph theory. A reduction method by
Thomassen [503] leads to a very short proof [247, Anhang 1]. See also Exercise 4.7
There are various other strengthenings of Steinitz' theorem, which say
that we can prescribe a lot about the polytope to be constructed. For
example, one can prescribe the shape of a facet of P in advance, by Barnette
Sz Grfinbaum [48]. That is, given a 3-polytope P with a k-gon facet K, and
given any k-gon K' C R3 , we can "redraw" P so that we get a polytope PI C
3 which is combinatorially equivalent to P, and such that K' is the facet
of P' which corresponds to K C P.
Similarly, we can prescribe the shadow boundary, by Barnette [37]: for
every cycle in G(P) we can find a realization of P and a projection of P
to the plane that carries the cycle to the boundary of the image polygon.
This is quite amazing: just try to verify this on your favorite 3-polytope; see
Exercise 4.9. However, one shouldn't get too greedy at this point: the shape
of the image polygon cannot be prescribed for this see Exercise 4.12.
Also, for every symmetry of the graph we can construct a polytope that
has this symmetry, see Mani [348]. (See also Theorem 4.13 below.)
The following is perhaps the more important extension, and is what
properly should be taken as a part of the Steinitz theorem. It says that
the space of all ways to find coordinates for a 3-polytope is connected
(modulo reflections). For the stronger statement below, we have to consider
all coordinatizations "modulo rotations and reflections." For this, we need
a few definitions.
Definition 4.10. Let P be a d-polytope on n > d vertices. We can
label the vertices vert(P) = {x 1 , x2 ,. .. , x n } such that {x1, x2, - - - , xd-1.-11
Notes
115
determines a flag, that is, aff({ x i ,... , xk }) n P is a (k-1)-face of P, for

1 < k < d-1-1.
The realization space
R,(P) c dxn
of P is the set of all matrices Y G R d " such that yk = xk for 1 < k < d+1,
and such that P is combinatorially equivalent to Q := cony{ y i , yn }
under the correspondence x i 4 yi .
It is easy to see that the realization space is an elementary semialgebraic set defined over Z, that is, a subset of a real vector space that can
be defined in terms of polynomial equations and strict inequalities with
integer coefficients. Such semialgebraic sets can be arbitrarily complicated
as topological spaces, in general (see Exercise 4.22). The following assumes
that you know what contractible means for a topological space: if not, just
accept that it says that the space has "no holes"; in particular, contractible
spaces are connected.
Theorem 4.11 (Steinitz' theorem).
[490, 4871
For every 3-polytope P, the realization space TZ(P) is contractible, and
thus connected.
All of these theorems on prescribed facets, shadow boundary, symmetry, realization space, and so on were proved with basically the same
technique, and some clever variations. Like Steinitz' original theorem, they
are far from trivial. One way to see this is that they all fail "one dimension
up," for d = 4. We will construct explicit examples in the next two lectures.
Notes
The reason for this version of the definitions for graph connectivity is that
they fit into a larger and very natural pattern. Namely, following Tutte [513,
5141 (see Truemper [510, p. 15]) one defines a k-separation (for k > 1) of
a graph G = (V, E) as a partition into two graphs, G 1 = (V1 , E1 ) and
G2 = (1/2 7 E2), which have exactly k vertices but no edges in common, and
which have at least k edges each. Thus, our sketch
shows a 4-separation, if each side has at least 4 edges (and thus at least
one cycle, or a nonseparating vertex, or both). For any k > 2, a connected
graph is k-connected if it has no t-separation for 1 < 1 < k.
116
With this definition connectivity is preserved under duality, and it readily

extends to the more general theory of matroids that's the reason to set it
up this way. For our discussion, we have specialized to the cases k = 2 and
k = 3, and we have disregarded the cases with few edges (every connected
graph with less than 2 edges is 2-connected, and every 2-connected graph
with less than 4 edges is 3-connected).
The key observation for the proof of Steinitz' theorem, namely the fact
that realizability is preserved under AY transformations (Section 4.2), is
nice and clear in Griinbaum [234, Sect. 13.1]. At the point where Griinbaum
seriously starts to treat graph theory (involving "lens graphs," etc.), our
treatment switches to Truemper's ideas.
For extensions of the Steinitz theorem we have relied on the survey in
Klee Sz Kleinschmidt [301, Sect. 4]. Griinbaum [238] surveys some more research on graphs of 3-polytopes. A deep separation theorem for 3-polytopes
was proved by Lipton & Tarjan [337].
A recent development is the (re-)discovery that Steinitz' theorem can
also be proved along completely different lines, by nonlinear methods. For
this, one first constructs a "correct drawing" of the polytope in the plane
the ideas for that go back to Maxwell, over 100 years ago [357]. Here a
correct drawing is a straight representation (as in Corollary 4.9) in which
one can think of the interior edges as rubber bands of various strengths,
compute the force ("stress") for each of them, and get an equilibrium of
forces at every vertex. (That's what Tutte [512] proves; see Linial, LovAsz
Si Wigdemon [338] for the rubber bands version. The proof that the equilibrium, with positive forces in all the edges, makes the regions convex, is
due to Whiteley [524].)
In the following two drawings, the right one is correct the left one
is not.
Then one proves that a straight line drawing of a 3-connected planar graph
can be lifted to 3-space (to obtain a 3-polytope) if and only if it is correct in
this sense. Of this basic theorem, Maxwell proved one direction, the other
(harder) one was provided by Crapo Sz Whiteley [157, 158] [525, Sect. 1.31;
see also Hoperoft & Kahn [260, Sect. 3], and in particular Richter-Gebert
[424, Sect. 12.2].
One basic observation is that a drawing of a graph is correct if and
only if it has a drawing of the dual (with a vertex "at infinity") such that
Problems and exercises
117
corresponding edges come in orthogonal pairs.
\\A
With proper care this also extends to higher dimensions, in the setting
of d-diagrams and Schlegel diagrams (see Lecture 5), as was shown by
McMullen [370 ] , completing an earlier version of Aurenhammer [23 ] _
Perhaps the nicest version of the approach via correct drawings uses the
circle packing theorem.
The Koebe-Andreev-Thurston Circle Packing Theorem 4.12.

(Koebe [311], Andreev [19 ] , Thurston [504])
Every planar graph can be represented in such a way that its vertices
correspond to disjoint disks, which touch if and only if the corresponding
vertices are adjacent.
Furthermore, if the graph is triangulated, then the representation is

unique up to Mbius transformations of the plane (which map all circles
into circles).
118
Furthermore, if the graph is 3-connected, then there is a simultaneous

representation of the dual graph by disks such that intersecting edges of
the graph and the dual are represented by disks whose boundary circles
intersect orthogonally.
-
(In that case, a good way to view the representation is that one (e.g.,
dual) vertex is represented by the complement of a disk. Then the whole
plane/sphere will be covered by the disks of the representation.)
The history of this result (rather: this circle of results) is involved at

best. The primal version was already known to Koebe and published in
1936. His proof, however, is valid only for the simple or simplicial case.
Thurston rediscovered the theorem, and reduced the proof to a theorem
by Andreev, so the result has become known as the Andreev-Thurston
theorem. Some of this is contained in Thurston's notes [504], but he never
formally published by it. By now there are many different proofs, among
them Schramm's [446, 447], one by Bragger [121], and a global proof by
Colin de Verdire [154]; see also Marden & Rodin [351] and Pach & Agarwal
[402, Chap. 8]. The version with orthogonally intersecting circles is due to
Peter Doyle. It appears independently in Brightwell & Scheinerman [125].
See Mohar [381] for a constructive proof. Oded Schramm has observed (via
personal communication) that Colin de Verdire's proof can be adapted to
yield the orthogonal version as well.
To get a feeling for "how circle packings look like" and "how they behave," I recommend the article by Dubejko & Stephenson [176] and the
(public domain) program they describe.
The primal-dual version can in fact be used to prove Steinitz' theorem
in a very strong form:
Every 3-connected planar graph is the graph of a 3-polytope P whose
edges touch the unit sphere,
There is a "canonical" representation of this form for every polytope.
Theorem 4.13.
(see Schramm [447])

For every planar 3-connected graph, there is a representation as the graph
of a 3-patytope whose edges are all tangent to the unit sphere S2 C R3 ,
and such that 0 is the barycen ter of the contact points.
This representation is unique up to rotations and reflections of the polytope in R3 . In particular, in this representation every combinatorial symmetry of the graph is realized by a symmetry of the polytope.
119
We don't try to do details here: after all, this uses completely "nonlinear"
methods, whereas everything we did is a combination of linear algebra and
combinatorics.
Still a different polytope realization algorithm, for simplicial 3-polytopes,
was given by Das Si Goodrich [170]: it essentially works by doing "many
inverse Steinitz operations on independent vertices in one step," and thus
yields a linear-time realization algorithm that produces realizations with a
singly-exponential bound on the size of the integer coordinates (as does the
Onn-Sturmfels algorithm [398] [424, Sect. 13.2] for general 3-polytopes).
Elementary semialgebraic sets are basic objects. Exercise 4.22 indicates
that they can have very complicated structure. General semialgebraic sets
(for which also nonstrict inequalities are admitted in the defining system)
can be written as finite unions of elementary semialgebraic sets.
In this chapter we only met realization spaces that are quite trivial (contractible); this will change as soon as we have developed the theory to
analyze the realization spaces of some high-dimensional polytopes, in Lecture 6. The study of semialgebraic sets is called real algebraic geometry
a very active and fascinating field of research. We refer to Bochnak, Coste
& Roy [103] and Becker [61] for more information.

4.0 Show that the complete graph K5, and the complete bipartite graphs
K3,, , are AY-reducible. Considering the graphs that are AY-equivalent to K6, show that IC, is not AY-reducible, for n > 6, and that
K,,,, is not AY-reducible for m, n > 4.
4.1
If G is a 3-connected graph with n > 4 edges, show that it contains

a subdivision of K4-
4.2
Show that the definition of k-connectivity given in the notes to this

lecture specialize to our definitions in Section 4.1 of 2-connectivity
(for graphs with more than 1 edge) and of 3-connectivity (for graphs
with more than 3 edges).
4.3 What is the problem with the construction of P from P', if G ---4 G'
is a simple Y-to-A transformation? For this, try to prove the Y-to-A
part of Lemma 4.3.
If you are comfortable with projective transformations (Section 2.7),
explain how to do this without using polarity (as we did).
4.4 Check how to do the reduction theorem for grid graphs entirely within
the framework of 3-connected graphs. Where are the problems? How
do you make grid graphs 3-connected? How many "basic operations"
do you need?
120
4.5 Characterize the graphs of centrally symmetric 3-polytopes.

(Griinbaum [234, Thm. 13.2.5])
4.6 Let G be any finite graph drawn into the plane without crossings, and
let y be its number of vertices, e its number of edges, c its number
of connected components, and f the number of connected regions
determined by it. For example, the graph
has v = 10, e = 11, f = 5, and c = 3.

Show that in general y e + f = 1 + c. Deduce Euler's formula
ve+f=
for the number of vertices v, the number of edges e, and the number
of facets f of a 3-polytope.
4.7 Deduce from Steinitz' theorem (from Corollary 4.9) that every planar
graph has a straight line drawing in the plane without intersections.
(In graph theory literature, this appears as a result of Wagner [517],
rediscovered by Fry [193]. It is also quite easy to prove this directly
by induction on the number of vertices; see, for example, Hartsfield
Sz Ringel [251, p. 167].)
4.8 Conclude from Exercise 4.6 that for every 3-polytope P, either the
polytope P itself or its polar P has a facet that is a simplex.
(This is not true for 4-polytopes for these read about the regular
24 cell, whose facets are octahedra and whose vertex figures are cubes;
see, for example, Coxeter [156, Sect. 8.2].)
-
4.9 The graph of the 3-cube, G(C3 ), has cycles that go through all the
vertices ("Hamiltonian cycles").
121
Take such a cycle, and then construct a realization of a combinatorial

cube in 3-space such that the usual projection ir : 3 --) R2 carries
the cube to an 8-gon, and the cycle to the boundary cycle of the
8-gon. Then do the same for the 3-dimensional permutahedron 11 3 ,
and for the (regular) dodecahedron.
4.10 Show that we cannot prescribe the shape of two facets of a 3-polytope
(even if they have congruent intersection). In fact, in a triangular
prism
if one square face is prescribed such that the edges a and b are parallel,
then the other square faces have to have b and c respectively a and c
parallel.
4.11 One cannot prescribe two disjoint facets of a 3-polytope either. For
this, analyze the prism over an n-gon, and show the following. If
we prescribe the bottom n-gon, then the coordinatizations of the
completed prism are prescribed by n +3 linear parameters ("degrees
of freedom"), while we have 2n 3 choices for the shape of the top
facet. Thus, if we prescribe two generic n-gons, then they cannot be
built into a prism, if n > 7.
122
4.12 Show that we cannot prescribe the shape of the shadow boundary of
a 3-polytope.
(Jurgen Richter-Gebert, who noted that from a triangular prism, you
can get a hexagon as a projection, but the shape of the hexagon cannot be prescribed: see page 141! Earlier, Barnette [42 ] gave a proof for
the case where the 3-polytope is a tetrahedron with a stellar subdivision on every facet, and Q is a regular 8-gon. Also, you can use the
prisms of the previous exercise, and again count degrees of freedom.)
4.13 Show that not all 3-polytopes can be represented in such a way that
all facets touch the unit sphere, or that all vertices are on the unit
sphere. Which 3-polytopes can be represented this way?
(Remark: This is a classical problem that goes back to 1832; see
Steiner [486], Steinitz [488 ] , and Schulte [450 ] . The characterization
problem was solved very recently by Hodgson, Rivin & Smith [255].)
4.14 Show that if an n-gon is represented in r 2 in such a way that all of

its edges touch S1 in their midpoints, then the n-gon is regular.
If the edges just touch S l , but not necessarily with their midpoints,
then the n-gon need not be regular, even if we require that the sum
of the vectors of the contact points is 0.
4.15 Let f2(n) E N be the smallest number such that the convex n-gon
P2(n) can be represented on an f2 (n) x f2 (n) grid, that is, with all
vertices in {0, 1, .. . , h(n)} 2 . For example, we have 12( 3) = f2(4) = 1,
f2( 5 ) = 12(6) = 2, /2(7) = f2(8) . .--- 3, f2 (9) = 4, but f(10) = 5 (see
-
the figure).
o
(i) Show that f2(n) < c n3/2 for some c> O.
(In fact, Thiele [502, Satz 4.1.10] proved that the bound
.f2(n) = 27 ( in ) 3/2 0(n log n)

is sharp.)
(ii) Using part (i), show that the cyclic 3-polytope C3 (n) with n
vertices can be represented on akxkx3 grid, where k --,--
f2(n 1).
123
(iii) Using part (i), show that the polar cyclic 3-polytope C3 (m) A
with n vertices (where n = 2m 4) can be represented on a
le x k' x le grid, where k' = f2( 71- + 1).
4.16* For n > 4 let f3 (n) be the smallest positive integer such that every
3-dimensional polytope with n vertices can be represented with integral vertices in {0, 1, ... , f3(n)} 3 . Let ,a(n) be the same function for
simplicia,1 3-polytopes.
Determine the asymptotic behavior of the functions f3(n) and Mn).
(Work by Goodman, Pollack & Sturmfels [220, Sect. 5 ] shows that
for d-dimensional simplicial polytopes with d 4 vertices one needs
vertex coordinates that grow doubly exponentially in d,
L9, (d + 4) > 22"
for some constant c> O. However, there is no such result that applies
to simplicial polytopes for any constant dimension d.
The case d =... 3 might be special: it is quite possible that there is a
quadratic upper bound on f3 (n); up to now only exponential upper
bounds f (n) < 4 3 and f3 (n) < 2 13122 are known, due to Onn &
Sturmfels [398] and Richter-Gebert [424, p. 143].)
4.17* For n > 4, let g(n) E N be the smallest number such that every 3
connected planar graph can be represented in the plane in such a way
that the vertices are on the g(n) x g(n) grid, the edges are straight,
and the bounded regions determined by the graph embedding, as well
as the union of all these regions, are strictly convex (i.e., all interior
angles are smaller than 7r).
-
How large is g(n)?

(If one just requires straight-line embeddings, without the convexity
assumptions, then one can embed the graph on a (2n-4) x (n-2) grid;
see de Fraysseix, Pach & Pollack [196 ] , and even on an (n-1) x (n-1)
grid, by Schnyder [443]. Chrobak & Kant [150 ] [284, Section 10.2.2]
have is a convex (but not strictly convex) drawing algorithm that runs
in linear time and only needs a grid on (n 1) x (n 1) vertices! Note
that the superlinear lower bounds of Problem 4.15 apply here. With
the strict convexity assumption, the currently best upper bound is
0(n3 ) x 0(n3), due to Chrobak, Goodrich & Tomassia [149].)
4.18* Does every 3-polytope have a realization with rational edge-lengths?

4.19* Can every (simple) convex 3-polytope P be represented in such a way
that a combinatorially polar polytope Q '.J P can be constructed
as the convex hull of vertices that are chosen on the corresponding
facets of P?
124
Can you do this for the 3-polytope obtained by cutting off the vertices
of a tetrahedron?
That is, can you represent this polytope in such a way, with points on
its facets, that adjacent facets of P correspond to adjacent vertices
on the convex hull P of the points?
(For the special polytope P above this can be done, according to
Jrgen Richter-Gebert. The general problem is suggested by the misleading description of polar polytopes in the book by Bartels [51,
p. 74]. Grnbaum & Shephard posed the question as Problem 3
in [241]. Grnbaum sent me the following e-mail message:
As far as I know, the problem is still open. I am inclined
to believe that the answer is negative, and that once some
counterexamples are found, we will all be saying how obvious that is.
I agree.)
4.20 Which subsets of 118 are elementary semialgebraic?
4.21
Show that for every convex d-polytope with n-vertices, the realization
space is an elementary semialgebraic set in Rd ".
(For this, embed the polytope into 1" d+1 , consider the maximal determinants of a realization matrix, and note that the condition that d+1
points have to be on a common facet says that a certain subdeterminant is 0. Similarly, two points being on the same side of the facet
spanned by some basis means that the product of two determinants
is positive.)
4.22 Let S C
d be a polyhedral set: a finite union of convex polytopes.
Show that there is an open semialgebraic set M C Rd that is a

neighborhood for S, such that S is a deformation retract of M.
(Hint: First show that the relative interior, and the exterior, of any
ball is an open elementary semialgebraic set. Also the intersection of
two elementary semialgebraic sets is elementary semialgebraic. Thus
M can be constructed, for example, by deleting a finite number of
125
small closed balls from a larger open ball.
In the drawing, S is a union of a line segment and the boundary of a

triangle; the set M appears shaded.)
4.23 The following problems discuss different ways in which a planar graph
can be represented by objects in the plane.
(i)
Prove that the vertices of a planar bipartite graph can be represented by horizontal and vertical closed line segments in the
plane, such that the segments intersect if and only if the corresponding vertices are adjacent.
(Hartman, Newman & Ziv [250 1 )
(ii)
In fact, every planar bipartite graph can be represented by disjoint horizontal and vertical open line segments in the plane,
which touch if and only if the corresponding vertices are adjacent.
(de Fraysseix, Ossona de Mendez & Pach [399, Sect. 6.3] [195])
(iii)* Can every planar graph be represented by a family of line segments in the plane such that every vertex corresponds to a segment and adjacent vertices correspond to intersecting edges?
4.24 Describe a "fast" algorithm to test whether two 3-polytopes are cornbinatorially equivalent.
For this, assume that the combinatorial structure is given by the
vertex-facet incidence matrices of the polytopes; from these, construct
the graphs. Then use that graph isomorphism can be tested "in linear
time" for planar graphs, according to Hoperoft & Wong [259].
(As a theoretical concept of fast algorithms one has the theory of polynomial algorithms and NP-completeness; see Garey & Johnson [208].
126
4. Steinitz'
Theorem for 3-Polytopes
The question of whether a polynomial algorithm exists for isomorphism of general graphs is one of the prominent open problems in
this theory [208, pp. 155-1581.)
4.25* A simple planar representation of a 3-polytope is obtained if we "cut

it open" along a spanning tree in its graph, lay out the resulting
structure in the plane. Thus we obtain nets of 3-polytopes. The following sketches show nets for a symmetric square pyramid, for a
not-so-symmetric square pyramid, and for a symmetric cube.
Not every way to open a 3-polytope leads to a valid net without

overlaps. For example, the following drawing shows a "bad" way to
unfold a 3-polytope that is combinatorially equivalent to the 3-cube.
Can every 3-polytope be represented by a planar net without overlaps?

(This problem appears in Shephard [460 ] , see also [160, Problem B211.
Nets of 3-polytopes were studied extensively in Alexandrow [101.
Combinatorial ly different 3-polytopes may have the same net thus
we cannot really "represent" 3-polytopes by nets; an example for this
was given by Shephard [460 ] . Another surprising fact in this connection can be found in Namiki, Matsui & Fukuda [3901: even tetrahedra
may have overlapping nets!)
4.26* Characterize the graphs of 4-dimensional polytopes.
5
Schlegel Diagrams for 4-Polytopes
Now that we understand the combinatorics of 3-polytopes (do we?), the

next step is to investigate 4-polytopes. Those are harder to understand,
since we (i.e., most of us) lack a genuine geometric intuition for the geometry of 4-dimensional Euclidean space. Nevertheless, there are various tools
available. The most prominent one is the "Schlegel diagram" of a polytope,
a polytopal complex that represents most of the geometry of a 4-polytope.
We will discuss polytopal complexes in some detail (they will be needed for
other purposes as well), and then we get to discuss Schlegel diagrams, and
some of the traps involved.
5.1
Polyhedral Complexes
Definition 5.1.
in Rd such that
A polyhedral complex C is a finite collection of polyhedra
(i) the empty polyhedron is in C,

(ii) if P E C, then all the faces of P are also in C,
(iii) the intersection P n Q of two polyhedra P, Q E C is a face both of P
and of Q.
The dimension dim(C) is the largest dimension of a polyhedron in C. The
underlying set of C is the point set ICI -': U
PEC P.
if
C is a polytopal complex all the polyhedra in C are bounded (polytopes).
128
5. Schlegel Diagrams for 4-Polytopes
In these lectures we will almost exclusively consider polytopal complexes.

The reader can work out generalizations whenever he or she feels that this
is interesting.
Our sketch shows a polyhedral (in fact, polytopal) complex Co of dimension 2, containing the empty polytope, 9 zero-dimensional polytopes
(vertices), 11 one-dimensional polytopes (edges), and 2 two-dimensional
polytopes (a triangle and a quadrangle).
Sometimes we can identify a polytopal complex C with its underlying

set ICI. For example, this makes sense for a polytope, since we can reconstruct the whole collection of faces from the point set P. However, often (see
below) we are really interested in subdivisions of polytopes and polytopal
complexes, and then C contains decisive extra information that cannot be
recovered from the point set ICI.
The combinatorial structure of a polytopal complex C is captured by
its face poset L(C) := (C, ): the finite set of polytopes in C, ordered by
inclusion. Assuming that we have a polytopal complex, we can read off the
dimension function from the rank function of the face poset, by dim(F) .
r(F)-1 for F G C. We define two polytopal complexes to be combinatorially
equivalent if their face posets are isomorphic as posets.
Our drawing shows the poset L(Co) for the complex Co above. Note that
L(C) does not have a unique maximal element, unless C is the complex
of all faces of one single convex polytope. Thus, L(C) is not a lattice in
general (although it is a finite meet-semilattice: it has a minimal element,
and meets exist). If we adjoin an artificial maximal element i, then we get
a lattice L(C) := L(C) U 01.
The maps f : C ----* D between polyhedral complexes are all the maps
f: ICI -- IDI that are affine maps when restricted to poiytopes in C. Two
complexes are affinely isomorphic if there is such a map f : C p T, which
is a bijection between C and D = {f(F) : F e C} . Equivalently, f has to
129
be a bijection on the underlying sets, such that f (F) is a polytope in D for

every F G C. A subcomplex of a polytopal complex is a subset C' C C that
itself is a polytopal complex.
Examples 5.2. Let P be a polytope.
(i) The complex C(P) of the polytope P is the complex of all faces of P.
The face poset of C(P) is the face lattice L(P).
(ii) The boundary complex c ( p) is the subcomplex of C(P) formed by
all proper faces of P. Thus its underlying set is IC(8P) I = ap =
P\relint(P). Its face poset is " (8 p) := L(P)\{P}.
(ill) A (polytopal) subdivision of a polytope P is a polytopal complex C
with the underlying space ICI = P. The subdivision is a triangulation
if all the polytopes in C are simplices. In particular, one is interested
in subdivisions and triangulations without new vertices, that is, where
the only zero-dimensional polytopes in the complex are the vertices
of P.
Our drawing shows (from left to right, everybody smile, please!) a poly
subdivision of a hexagon, a subdivision without new vertices, a tri--topai
angulation, and a triangulation without new vertices.
The study of subdivisions of polytopes has received an enormous amount

of attention in recent years, motivated by applications that range from the
theory of generalized hypergeometric functions (initiated by I. M. Gerfand;
see [210]) to spline theory and to questions in computational geometry. We
refer to, for example, Billera [65], Edelsbrunner [1801, Pach [401], Gerfand,
Kapranov & Zelevinsky [214], Lee [326, 329], and the references in these
papers.
A basic and central concept is that of "regular" subdivisions.
Definition 5.3. A subdivision C of a polytope Q C Rd is regular if and
only if it arises from a polytope P C Rd +
1 in the following way.
(i)
The polytope Q is the image r(P) = Q of the polytope P, via the

canonical projection map
7r : Ilid+ 1 --+
(X dx-F1)
which "deletes the last coordinate."
130
(ii) C is the set of all lower faces of P, projected down to Q, that is,
C = fir(F) : F is a lower face of PI,

where the lower faces of P are the faces F that satisfy z Aed+1 V P
for each x E F and A > O. Equivalently (by the Farkas lemma!), the
lower faces are the faces of P of the form
F = Ix E P : cx = col, cx < co valid for P, cd+1 < O.

In other words, C is the family of all faces of P that can be "seen" from
oo large enough.
Te d+i , for T
If the polytope projection 7r :P.--Qis given, then we denote the
subdivision C of Q it determines via part (ii) by p(Q).
For d = 1 (subdivision of a line segment Q) this looks as follows, where
the lower faces of P (4 vertices and 3 edges) are indicated by thick lines:
Q
It suggests a reformulation: regular subdivisions arise from piecewise linear
convex functions in the following way. Given the projection 7r : P -----> Q,
the function
f:Q *R,
f (x) = minty E R: (1 E P}
Y
is piecewise linear and convex. (A function f : Q ---) R is piecewise linear

if Q can be written as a finite union of polytopes on which P is a linear
function.) For a converse, note that every piecewise linear convex function
over a polytope Q determines a polytope projection, by setting
P := cony{
x
f (x)
: x E Q}.
The projection of the lower envelope of this P determines a regular subdivision C of Q.

Clearly, every subdivision of a line segment (d = 1) is regular. For d = 2,
this is still true for subdivisions without interior points, that is, all subdivisions of a convex n-gon without new vertices are regular (Exercise 5.0).
131
Example 5.4. For integers z 1 ,..., zd > 1, the pile of cubes Pd(zi,- - , zd)
is the polytopal complex formed by all unit cubes with integer vertices in
the d-box
:= { s E Rd :
zi for 0 < i < d},
that is, the polyhedral complex formed by the set of all cubes
--Axe
C(ki,...,kd)
: ki < xi <ki +11
for integers 0 <k < z, together with all their faces. So, our drawing
represents the piles of cubes P2 (6,4) and P3 (6,4,3).
P2(6. 4)
P3(6, 4, 3
In particular, the pile of cubes Pd(zi, .. , zd) has (zi-1-1)- .. -(zd-F1) vertices, which are given by
vert(Pd (zi ,
, zd)) = B(zi,
, zd) n zd.
The pile of cubes is a regular subdivision of B(zi,

, zd). This is true
because the pile is a "product" (Exercise 5.4) of 1-dimensional regular
subdivisions. In particular, all convex functions of the form
f
:Rd + R,
f (x)
fi (x1)
+ - - + fd(xd)
for convex functions L will produce suitable polytopes

PdA-1(Z1, - - ,Zd)
:=
cony
: V E B(Zi,
Zd)
fl Zd}.
132
A canonical choice for f is the function f (x) = x? + . .. + x= 11x 11 2 . Our

sketch tries to illustrate the construction of the 3-polytope j53(6, 4).
To end this section, we give the following figure as a small example of a

subdivision that is not regular. To see this, note that we can assume that
f (v 4 ) = f (v5) = f(v 6 ) -= 0 for the three interior vertices (by subtracting a
linear function from f). and then we get a cycle Au ' ) > f(v2 ) > f(v3) >
f (v 1 ) of conditions for the three outer vertices.
Note that the subdivision Ep(Q) encodes a lot of information about P,

but essentially we "see only half of P" in Ep(Q). Now, we'll "take a closer
look" in order to "see more."
5.2
Schlegel Diagrams
For a "close look" at polytopes, we choose a point of view k, beyond facet F

(see Section 3.1 for the definition of "beyond") and use the facet F as a
"projection screen" for everything we "see" behind it.
Definition 5.5. Let P be a d-polytope in d and let F G L(P) be a
facet of P, defined by the valid inequality ax < z. We denote by
,
H := aff(F) = ix G Rd : ax = z}
133
the hyperplane spanned by F. Choose a point yE. beyond F. For x E P,

we define
Z ay.
p(x) :, y,
(x y,).
ax ay,
Y,(
The Schlegel diagram of P based at the facet F, denoted as D(P, F), is the
image under p of all proper faces of P other than F; that is, it is the set
system
D(P, F) := fp(G) : G E L(P)\IP, F}},

contained in the hyperplane H.
The map p is nonlinear, and in general we cannot replace "combinatorially equivalent" with "affinely isomorphic" in the following proposition;
see Exercise 5.8. Now comes the reason why this construction makes sense.
Proposition 5.6. The Schlegel diagram of P based at the facet F is a
polytopal subdivision of F that is cornbinatorially equivalent to the complex
C((OP)\{F}) of all proper faces of P other than F.
Proof. We use the notation of Definition 5.5. For every face G of P, the
set
CG := typ, A(x ks x E G, A > 01
)
is a cone with vertex y in fact, ayF > ax for all x E P by construction

of x,
/
/
/
/
Y,
If G is a proper face of P, then it is contained in a hyperplane

Hi = Ix E Rd : ai s = bi l
that does not contain Y. Thus we find that the face lattice of CG
is iso-
morphic to MG). Now if we intersect CG with a hyperplane such as H,

which has a bounded, nonempty intersection with CG, then we have that
134
the face lattice of the intersection is isomorphic to that of CG. Here we

have p(G) = CG fl H (the formula for p(x) is derived from this condition),
and thus L(G) = L(CG) = 1(p(G)).
(Actually, the map p: G ----+ p(G) is a projective transformation, very
much like the version we have in Section 2.6: see Exercise 5.8.) [ii
Let us note here one very important property of Schlegel diagrams: that
G n aF is a face of F=IDI, for all G ED.
Schlegel diagrams are, sure enough, named after a mathematician, Victor
Schlegel [438]. It seems that Sommerville coined the name with his 1929
book [470]. In the next section we will get a more general definition of a
d-diagram and then use "Schlegel" as an adjective for a d-diagram that is
a Schlegel diagram.
If a Schlegel diagram of P is given as a polytopal complex D, then we can
reconstruct the corresponding facet F of P as F = IDI. So with Proposition 5.6 we see that every Schlegel diagram D determines the combinatorial
isomorphism type of P: we can reconstruct the face lattice of P from it.
For this we recover F = MI, and get
L(P) ri (V U {F,i}, <),

where the partial order "<" is by inclusion on D, we have G < F if and
only if G E D is a face of F, and i is an artificial maximal element.
The interesting property of the Schlegel diagram is that it completely
encodes the combinatorial structure of a d-dimensional polytope into a
(d 1)-dimensional object; after all, D(P,F) is a polytopal complex contained in the (d 1)-dimensional hyperplane H = aff(F).
This reduction in dimension makes Schlegel diagrams especially useful
in the case d = 4, where the Schlegel diagram of a 4-dimensional polytope
is a 3-dimensional polytopal complex, for which we have a fighting chance
for a geometric visualization.
To illustrate how Schlegel's construction works, we start, however, at
d = 3. The following drawings show the construction of the Schlegel diagram, and the (2-dimensional) Schlegel diagram itself, for a 3-dimensional
simplex
where the center vertex in the diagram represents the (unique) vertex of
the tetrahedron which is not on the facet we project to. Similarly, here are
135
the drawings for the 3-dimensional cube.
Next we consider the 3-dimensional prism (the product of a triangle with

a 1-polytope): this has nonisomorphic facets, and thus we get different
Schlegel diagrams for points of view beyond different facets (a square facet
and a triangle facet).
And finally, here is an attempt to illustrate how you construct the Schlegel
diagram of a lifted pile of boxes, based on the big square facet on its top.
The drawing is not metrically correct in any sense just an attempt
to understand the geometric situation. Specifically, we go for the Schlegel
136
diagram of j53 (6, 4), from a point of view "above the polytope":
Nokui roil
fall11111114
#0111111 111111114k
Now we go for the 3-dimensional Schlegel diagrams for 4-dimensional

polytopes: these are all the "pictures" of 4-polytopes you will get! For
example, here is (a 2-dimensional drawing of) the (3-dimensional) Schlegel
diagram of the 4-dimensional simplex,
where the center vertex in the diagram represents the (unique) vertex of
the 4-simplex which is not on the tetrahedron we project to; here is the
drawing for the 4-cube,
and here's the Schlegel diagram for a product of two simplices
137
P = A2 X A2/
a decomposition of a triangular prism into five triangular prisms compare

this to our description in Example 0.5.
We close with an attempted sketch of the Schlegel diagram of a 4-dimensional lifted pile of boxes, namely P4(2, 2, 2).
Although this sketch is quite sketchy, the picture might try to tell you about
the combinatorial structure of (the Schlegel diagram of) the 4-polytopes
P4(Z1/ Z2 1 Z3). In fact, from the Schlegel diagram (compare also to the 3dimensional case) we see that P4 (zi , z2 , z3) has exactly z1 z2 z3 + 7 facets.
The seven "big ones" are the big cubical facet on which the diagram is
based and the six 3-dimensional facets that are lifted piles of cubes: two
copies of P2 (Z1 1 Z2) at the bottom and the top of the diagram, two copies
of 'P2(zi, z3) at the front and the back of the diagram, and two copies
of P2 (Z2 ) z3 ) at the left and right sides of the Schlegel diagram. Between
these six big facets, we have the subcomplex formed by z1z2z3 little cubes,
which is isomorphic to the pile 'P3(zi, z2, z3) from which we started the
construction.
Make sure that you "see" this: it will be useful later (in Section 8.2)
when we study the lifted piles of cubes again.
138
5. Schlegel Diagrams
for
4-Polytopes
5.3 d-Diagrams
Schlegel diagrams were studied extensively around the turn of this century;
however, no one realized that there is a problem: not everything that "looks
like" a Schlegel diagram actually is one. (See the notes below.) Now we
define looks like.
Definition 5.7. A d diagram is a polytopal subdivision C of a d-polytope

P= ICI C Rd such that G n OP is a face of P for each GEC.
A d-diagram D is simplicial if IDI and all the polytopes in D are simplices. The diagram is simple if every vertex of IDI is contained in exactly
d different d-polytopes of D, and every other vertex of D is contained in
exactly d + 1 different d-polytopes in D.
-
Check that every Schlegel diagram of a d-polytope is a (d *diagram!

We define the face poset L(C), and combinatorial equivalence, of d-diagrams
in complete analogy to the definitions for Schlegel diagrams. With this, a
Schlegel (d *diagram D(P, F) is simplicial, respectively simple (according to Definition 5.7), if and only if the polytope P is simplicial, respectively
simple. In fact, a diagram D is simplicial if and only if its face lattice L(D)
"looks like" the face lattice of a simplicial polytope, in that all the lower
intervals [0, G] are boolean for G i.; similarly, D is simple if and only if
all upper intervals [G, i] are boolean for G 0.
Here are a few trivial examples: a Schlegel diagram (of a triangular
prism), a 2-diagram that is not a Schlegel diagram, and a polytopal subdivision that is not a 2-diagram.
We note that the question of whether a given d-diagram is a Schlegel

diagram can be reduced to a linear programming problem (Exercise 5.2)
but the question of whether it is combinatori ally equivalent to a Schlegel
diagram is much harder; the task can be split into the enumeration problem
for oriented matroid spheres, and the oriented matroid realizability problem
(see the notes at the end of this lecture). So, the following is a decidedly
nontrivial theorem, accompanied by a trivial proof.
Theorem 5.8.
Every 2-diagram is combinatorially equivalent to a Schlegel diagram.
5.4 Three Examples
139
Proof. A simple combinatorial argument implies that the graph of every

2-diagram is 3-connected (it is simple and planar by construction). Thus the
theorem follows from Steinitz' Theorem 4.1, together with the construction
of Schlegel diagrams of 3-polytopes.
C
Proposition 5.9. If a diagram 73 is Schlegel, then D is a regular subdivision of F = 173I. The converse is true in the case where F is a simplex,
but not in general.
We omit the proof see Exercises 5.2(ii) and 5.9.
There are other properties that separate Schlegel diagrams from some
non-Schlegel diagrams. For example, let 7) be a d-diagram, and let L be
the "reconstructed" face lattice of the (d + 1)-polytope P, if it exists. The
combinatorial information about D is then contained in the pair (L, F),
where F is a distinguished coatom of L. Now we call a diagram invertible
if for every coatom F' of L, there is a d-diagram corresponding to the pair
(L, F'). We know that every Schlegel diagram is invertible. It turns out
that a part of the non-Schlegel diagrams is not invertible.
Similarly, we say that V has a combinatorial polar if there is a diagram
whose combinatorial data are given by the pair (LP, A), where A is a vertex
of D (corresponding to an atom of L, and thus to a coatom of LOP). Again,
every Schlegel diagram has a polar diagram (which is Schlegel), and some
non-Schlegel diagrams have polars, some do not.
5.4
Three Examples
Examples of interesting Schlegel diagrams, and of 3-diagrams that are not

Schlegel diagrams, are not too hard to come by. In the following we will describe three strikingly simple examples, where two are due to Barnette [43],
and one is from Schulz [451]. See also Ewald [189, Sect. IV.4] for the "classical" examples of BrUckner and Barnette (see the notes), and Schulz [452]
for some other interesting constructions.
From general theory (namely the technique of "Gale diagrams," which
we will soon get to) one can see that every d-diagram with at most d + 4
vertices is always Schlegel. Therefore, the minimal counterexamples that
we can hope for are 3-diagrams with 8 vertices.
Here we go: the following construction produces a 3-diagram that is not
combinatorially equivalent to the Schlegel diagram of any 4-polytope.

Example 5.10 (Schulz' 3-diagram). [451]
We start with a 3-polytope Q with 6 vertices, labeled 1, 2, ... , 6 in our
drawing. The 3-polytope can be realized by starting with a tetrahedron,
cutting off the top to get a triangular prism, and then cutting off an extra
triangle [2, 3,5] as in the figure, where 3, 5 were vertices of the truncated
prism, and 2 was on an edge of it.
140
Now we choose extra points: 7 in general position inside Q, but outside the
tetrahedron [2,3, 5,6], and 8 above the top of the original pyramid, so that
all facets of Q, except for the base [4, 5, 6], can be "seen" from 8.
The diagram D I , based on the tetrahedron G = [4, 5, 6, 8], consists of the
following ten 3-polytopes in R3 , and their faces:
A:
B, C:
D, E:
[2, 3, 5, 6, 7], a bipyramid over the triangle 357,

[1, 2, 3, 7] and [4, 5, 6, 7], two tetrahedra, and
[1, 2, 4, 5, 7] and [1, 3, 4, 6, 7], two square pyramids,
where B,C,D,E together cover the interior of CAA, by using 7 as a

cone point,
F, K, L: [1, 2, 3, 8], [2, 3, 5, 8] and [3, 5, 6, 8], three tetrahedra, and

H,I: [1, 2, 4, 5, 8] and [1, 3, 4, 6, 8], two square pyramids,
where F,K,L,H,I together cover the interior of G\Q, by using 8 as
a cone point.
It is easy to see now that this is a valid 3-diagram. Why is it not combinatorially equivalent to a Schlegel diagram? Assuming it is, then there is a
4-polytope P whose vertex set we can label 1,2.. , 8, and whose Schlegel
diagram is equivalent to D I .
Now all vertices of Q are contained in the union of the two 2-faces
[1,2, 4, 5] and [1,3, 4, 6], which share an edge [1, 4]. Thus the affine span
R :-= 0{1, 2, 3, 4, 5, 6} of the vertices of Q has dimension 3 in R4 . Furthermore, the triangles [2,3, 5] and [3, 5, 6] are contained in R. Thus the facet
A has two 2-faces in R, and therefore it must be contained in R. Thus
the vertex '7 and all of its neighbors are contained in R c R4 , which is
a contradiction (to the fact that different facets of P must span different
D
hyperplanes in R4 , or to Lemma 3.6).
5 4 Three Examples
141
Our next example, D2 is a Schlegel diagram for which the total space
F = 17,2 1 cannot be prescribed. This shows that for a 4-polytope, we cannot
prescribe the shape of a facet in contrast to the situation for 3-polytopes,
as discussed in Section 4.4.
Example 5.11 (Barnette's first diagram). {43]
For this, we consider the Schlegel diagram for the prism over the square
pyramid I x Pyr3. This Schlegel diagram can be constructed from a regular
cube, and two vertices on the vertical axis of symmetry.
E4
Now consider any 3-diagram that is combinatorially equivalent to this.

Then the lines determined by the edges E1 , E2 1 and E belong to a pencil of
lines, that is, either they are all parallel, or they have a common point of intersection. To see this, just consider the plane R = aff(Ei UE2 ) determined
by the lines E1 and E2 1 and the intersection point with the line aff(E),
which may be "at infinity." Using that aff(EUE 1 ) is a plane, we get that the
intersection point is contained in the line aff(EUEI )naff(E1 UE2) = aff(Ei
and similarly it is contained in the line afF(E2)-
From symmetric arguments for Ez , Ei+1 , and E, we see that for every
3-diagram that is combinatorially equivalent to the given one, the four
lines E1 , E2 1 E3 1 E4 are parallel, or else they intersect in a common point.
Thus if we start with a "skew" combinatorial cube that does not satisfy
this (such a cube is easy to get), then we cannot even complete it to a
142
3-diagram that is isomorphic to the given one. In particular, such a skew

cube is never a facet of a prism over a square pyramid.
Steinitz' theorem can be stated in the following way: any "reasonable"
cell decomposition of the 2-sphere (technically, we can consider regular
CW-complexes with the intersection property; see [92, Sect. 4.7]) can be
realized as the boundary complex of a convex 3-polytope. It is a weaker
statement that, after deleting the interior of a 2-face, the rest can be realized
by a 2-diagram.
This formulation of Steinitz' theorem has an obvious generalization to
3-spheres and 4-polytopes and Schulz' Example 5.10 showed that this
generalization is false. In fact, it is not hard to see that any d-diagram
defines a "reasonable" cell decomposition of the d-sphere. Thus, diagram
DI represents a 3 sphere that can be realized by a 3-diagram, but not by
a 4-polytope.
Our next example shows that some 3-spheres cannot even be represented
by a 3-diagram (at least not with a specified simplex as its base).
-
Example 5.12 (Barnette's topological diagram). [43]

Barnette's example, D3, is a "curved," topological 3-diagram that cannot
be straightened at all. This suggests that an effective combinatorial characterization of d-diagrams and Schlegel diagrams is too much to hope for.
For this, we start with a tetrahedron T3. Into this tetrahedron, we glue
a subdivision of a quadrangle Q2, as indicated in our drawing, in such a
way that the boundary El , E2 E3 E4 of the quadrangle is identified with a
circuit of four edges E 1 , E2 E3 E4 on the boundary of the tetrahedron, and
the interior of Q2 gets mapped into the interior of T3, along a curved surface
(which you may think of as a soap film bounded by the four edges Et ).
The curved quadrangle partitions the interior of T3 into two "3-cell regions." Into each of them we place a new vertex and then perform a "stellar
subdivision": each vertex is joined to all the faces on the boundary of the
respective 3-cell, so each 3-cell is replaced by the (topological) pyramids
over four triangles and two quadrangles each, and their faces.
But we find that this "topological 3-diagram" is not realizable. In fact,
any two realizations of the tetrahedron are equivalent. Now we try to realize
the quadrangles F2 and F4 in Q2 by planar convex quadrangles. Then the
plane H2
aff(F2) contains E2 and a (unique) point of E4 Similarly,
.
Notes
143
the plane H4 := aff(F4) then contains E4 and a point of E2. Thus the
intersection 112 n H4 connects a point on E2 with a point on E4. This
means that the four points of F2 are not in convex position in T3.
Il
Notes
Schlegel diagrams are the most direct, and probably the most effective, tool
to visualize 4-dimensional objects. Of course, there is a certain amount of
training necessary to develop the geometric intuition. Don't let yourself get
discouraged, not even by statements like the following:
Here, however, a word of warning may be in order: do not try to
visualize n-dimensional objects for n > 4. Such an effort is not
only doomed to failureit may be dangerous to your mental
health. (If you do succeed, then you are in trouble.) To speak
of n-dimensional geometry with n > 4 simply means to speak
(Chvital [151, p. 252])
of a certain part of algebra.
This is wrong, and even Chvital acknowledges the fact that the correspondence between intuitive geometric terms and algebraic machinery can be
used in both ways [151, p. 250].
The technique of Schlegel diagrams was already used extensively in work
of Bruckner [130] early this century, where the distinction between Schlegel
diagrams and 3-diagrams was not made.
There are also simplicial examples known of 3-diagrams that are not
Schlegel diagrams, and not even combinatorially equivalent to such. The
first one was described by Griinbaum's abstract [233], which started the
subject. The first non-Schlegel 3-diagram with 8 vertices was found by
Griinbaum & Sreedharan [242], showing that one of Briickner's 3-diagrams
does not, as Bruckner thought, represent the combinatorial type of a 4polytope. It is now known as the Briickner sphere [234, p. 222]. A second
example of a simplicial 3-diagram that is not equivalent to a Schlegel diagram the Barnette sphere
was found by Barnette [35] a little later.
There are also simplicial 3-spheres that can be represented by topological
diagrams (as in our Example 5.12), but not by straight 3-diagrams. Both
kinds of examples are nicely presented in Ewald [189, Sect. IV.4 and IV.5].
However, there is something special happening in the case of simple diagrams. In fact, every simple d-diagram with d > 3 is the Schlegel diagram
of a (d + 1)-polytope. Thus, there are also things true in 3-space that are
false in 2 dimensions. This was proved by Whiteley [526], and in an even
stronger version by Rybnikov [435]. (They use a quite general setting for
"Mobility"; see Crapo & Whiteley [157, 158, 159].) Earlier results of Davis
[172] and Aurenhammer [23] did not include the "boundary" conditions
that pose extra constraints (as in Problem 5.7).
144
Griinbaum & Sreedharan's [242] complete enumeration of all simplicial

4-polytopes with 8 vertices (correcting Briickner's [130] earlier attempt)
also produced the first example of a neighborly polytope that is not cyclic,
disproving a conjecture by Motzkin [385] [205]. (In Exercise 6.15 you'll
construct an example yourself!) See Griinbaum & Sreedharan [242], Barnette [39], and Altshuler, Bokowski & Steinberg [15] for the beginnings of
a classification of the polytopal and nonpolytopal simplicial 3-spheres on
a "small" number of vertices. A general framework for the (difficult) algorithmic questions that arise in this context was developed by Bokowski
& Sturmfels [110, 111], using the theory of oriented matroids that we will
encounter soon.
The first 4-polytopes for which a facet cannot be prescribed were constructed by Barnette & Griinbaum [48] (an 8-polytope with 12 vertices;
see also [234, p. 96]) and then by Barnette [39] (a 4-polytope with 13 vertices). The minimal number of vertices was achieved by Kleinschrnidt [304];
a 4-polytope with 8 vertices and 15 facets. It has the extra feature that all
other facets are simplicial (tetrahedra). We will get to this in Lecture 6 (it
requires different methods), where we will also give a different approach
to Example 5.11. It turned out recently that one cannot prescribe a 2-face
for 4-polytopes: see Exercise 5.11. We still do not know whether one can
prescribe a facet for simple 4-polytopes (Problem 6.17(ii)*).
The structure of d-diagrams (even in the special case d = 3) is far from
completely understood. A great deal of interesting combinatorial and algebraic problems arise from the general study of "Which structures have
straight embeddings into real space?". Via some basic lemmas by Bing [76]
and Whitehead [523] that we will see in action in Example 8.9, this is in
fact equivalent to the question of "Which structures can be substructures
of d-diagrams?" We wish to point the reader to the handbook article by
Brehm & Wills [123].
As an aside, observe that every Schlegel diagram 73(C4 (n) ,F) of a
polar of a cyclic polytope has the following properties: it yields a simple
configuration of n different 3-dimensional polytopes that pairwise meet in a
common facet. Such configurations have been constructed "by hand" again
and again, by Tietze [505, 506], Besicovitch [62], Rado [415], Dewdney &
Vranch [174], myself [536], and others, because they kill a 3-dimensional
version of the notorious "four color theorem." In fact, the Schlegel diagram
D(C4 (n) ) shows that arbitrarily many regular convex regions can be pairwise adjacent, so that no finite number of colors suffices to color the regions
of simple 3-diagrams. See also the discussion in Griinbaum [234, Sect. 7.4],
in Danzer, Griinbaum & Klee [169], and in [160, Problem E7].
Related to this, the problem of how many simplices can be pairwise
adjacent in Rd has received a lot of attention; see Perles [406] and Zaks [534]
and the references therein. The titles of these two papers tell you what's
known about upper bounds, and Exercise 5.12 tells you the lower bound,
which is also conjectured to be best possible.
145

5.0 Prove that every polytopal subdivision of a polygon without new
vertices is regular.
5.1
Show that the following 3-polytope (the capped prism) has a nonregular triangulation without new vertices.
(Kleinschmidt and Lee, see Lee [329, Sect. 6])
5.2 Let C be any polytopal subdivision of a d-polytope in Rd.

(i) Show that the decision of whether C is regular can be reduced
to a linear programming feasibility problem. How do you get rid
of the strict inequalities that come up?
(ii) If C is a Schlegel diagram, show that it can be obtained from a
"lower faces" construction as in Definition 5.3.
"to infinity";
(Use a projective transformation that moves
compare to Exercise 2.18.)
(iii) If C is a d-diagram, show that the question of whether it is a
Schlegel diagram can also be reduced to a linear programming
feasibility problem.
x,
5.3
Produce examples of polytopes that have nonregular triangulations

without new vertices.
(i)
(ii)
The 4-cube C4 has nonregular triangulations without new vertices.

(It was an open problem for a long time whether Cd has nonregular triagulations for any d. Now de Loera [339] has shown
that there are even such triangulations with 24 maximal simplices, all of which have volume 1/4!.)
The second hypersimplices Ad _ 1 (2) have nonregular triangulations without new vertices, for d > 9.
(These were constructed with combinatorial tools by de Loera,
Sturmfels & Thomas [340]; you are allowed to use the computer,
via part (i) of the previous exercise. This might in particular be
useful in the next two parts, which are unsolved problems.)
146

(iii)* Can products d m- 1 x An_ i of two simplices have nonregular
triangulations?
(iv)* A lattice polytope P C d (i.e., a d-polytope with vertex set is
contained in Zd ) is totally unimodular if all the d-simplices that
are spanned by vertices of P have volume lid!.
Are all triangulations of a totally unimodular lattice polytope
regular?
(This is a conjecture of Sturmfels. It would, in particular, imply
a negative answer to part, Pr.)
5.4 Define the product of two polyhedral complexes, in such a way that
the product of subdivisions of two polytopes P and Q is a subdivision
of the product P x Q.
Prove that the product subdivision is regular if and only if the original
subdivisions of P and Q were regular.
5.5 Compute the numbers fk of k-faces for the polytope P4(zi , z2, z3).
5.6 A "default" convex function is given by the paraboloid function
f:
Let V C d be a finite set of points (vertices). Show that the regular
subdivision of Q := conv(V) associated with this f has the following
property: for every facet of the subdivision there is a sphere that
contains all its vertices, but no other vertices from V.
(This subdivision is known as the Delaunay triangulation of the point
set V. It is of great importance for many aspects of computational
geometry; see for example Edelsbrunner [184)
5.7 Show that the following 2-diagram is regular, but not Schlegel:
5.8 Show that in general the polytopal complex C(aP)\IFI and the
Schlegel diagram D(P, F) are not affinely isomorphic. (For this, it
suffices to consider the 3-cube!)
However, show that p: G --) p(G) is a projective transformation,
for all proper faces G of P.
Problems
and Exercises
147
5.9 Construct a Schlegel diagram for the polar of a product
P , (A2 x A2) A Construct Schlegel diagrams for the cyclic polytopes C3 ( 6 ) and C4 ( 7 ) .
5.10* What is the smallest number f (d) of d-simplices that is sufficient to

triangulate a d-cube?
Combining results by Hughes and Anderson [261] [263] [262] [264]
and Haiman [246], we know that
f(2) =-- 2
f(6) =-.- 308
f (3) ---, 5
1(7) = 1493
f(4) =-- 16
f(5) ---, 67
and 5522 < f (8) < 11944.
For large d a method due to Smith [469], using volume estimates in

hyperbolic geometry, yields the best lower bounds on f(d) so far.
Is it true that the smallest number is always/only achieved by a
triangulation without new vertices?
What is the maximal number of simplices that may be needed to
triangulate any 0/1-polytope: Is this the same number you get in
case of the d-cube?
("Efficient" triangulations of d-cubes, with few facets, have been studied extensively, for example for use in finite element methods for
solving differential equations. Thus, a lot is known about such triangulations. For example, one knows that asymptotically, at most pdd!
simplices are needed for large d, for some constant p < 1. We refer to
Haiman [246] and Todd & Tunel [508] for more information.)
5.11 For a 4-polytope, one cannot prescribe the shape of a 2-face! Namely,
Richter-Gebert [424, p. 91] [426] provides the following diagram of a
4-dimensional polytope X*:
Show that this represents the Schlegel diagram of a 4-polytope with

8 facets and 12 vertices. Show that the shape of the hexagon at the
base cannot be prescribed arbitrarily: three lines, determined by two
opposite edges and by the diagonal between them, must (projectively)
meet in a point. (See also Exercise 6.11)
148
5.12 Show that 2d simplices can be arranged in r d in such a way that any
two are adjacent (that is, the intersections are (d 1)-dimensional).
5.13 The smallest triangulation of the torus surface has 7 vertices, 21 edges
and 14 triangles. Construct it, and show that it can be embedded as a
subcomplex into C(C7(4)), and thus as a simplicial complex into r 3
p. 253]; see also Altshuler [13] and(CsAzdr[16],Ginbaum234
Bokowski Sz Eggert [106])
5.14* Can every triangulation of the torus be realized by a simplicial complex C in r 3 ?

(This is a classical problem of Griinbaum [234, p. 253], which is still
wide open. See Ewald, Kleinschmidt, Pachner & Schulz [191, p. 153 1 ,
Altshuler, Bokowski & Schuchert [14] and their references.)
5.15 Show that every d-dimensional simplicial complex can be realized as
a subcomplex of a simplicial (2d+2)-polytope, and thus has a straight
realization as a simplicial complex in r 2d+ 1 .
(Hint: Take a suitable cyclic polytope and its Schlegel diagram.)
5.16 Show that there are polyhedral complexes that are not subcomplexes
of polytopes. Namely, show that the subdivision of the Mbius band
drawn below with 6 vertices, 10 edges, and 4 facets can be realized as
a polyhedral complex in r 3 but not as a subcomplex of any polytope.
(This is due to Betke, Schulz & Wills [64]; see Barnette [44] for a
similar, but simplicial, Mbius strip that serves as an "impediment
for polyhedrality." It is not even true that every triangulated Mbius
strip has a straight embedding into 3 : see Brehm [122]. However,
the Schlegel diagram construction shows that no such Mbius strip
can appear in the boundary of a 4-polytope.)
5.17 Modify Schulz' 3-diagram D I from Example 5.10 by subdividing the

bipyramid A into three tetrahedra, [2, 3, 6], [2, 5, 6], and [2, 6, 7]. Show
that this yields a new 3-diagram DI with 8 vertices and 12 facets,
which contains the Mbius band of the previous exercise as a subcomplex.
Derive that VI is not polytopal, either.
6
Duality, Gale Diagrams,
and Applications
More about life in high dimensions: after "successfully" dealing with polytopes in four dimensions, we now study polytopes with few vertices, that is,
d-polytopes with only d-plus-a-few vertices. For this, we develop a duality
theory, which describes them in terms of structures in low dimensions.
This duality theory (developed by Perles in the 1960s, and recorded
by Griinbaum [234]) is classically known as Gale diagrams. Later it was
realized (apparently first by McMullen [366]) that Gale diagrams are a
manifestation of "oriented matroid duality."
Thus behind the construction of Gale diagrams one finds (barely hidden)
oriented matroids. Their theory was initiated in the 1970s by at least four
independent authors, Jon Folkman, Jim Lawrence, Robert G. Bland, and
Michel Las Vergnas; see [194] and [96].
By now oriented matroids form a theory with many facets, extensive
enough to fill thick books [92]. One aim of our lectures is to give a simple introduction to a few topics that help understand polytopes. Keys to
this include the description of the relative position of vertices in terms of
"circuits" and "cocircuits," as well as the simple duality between these two
descriptions, oriented mat roid duality, which in polytope theory manifests
itself in (linear and affine) Gale diagrams.
Thus this lecture includes a brief crash (?) course on oriented matroids.
(This will be continued in Lecture 7, when we discuss hyperplane arrangements and zonotopes.) Although oriented matroids need some amount of
new notation and terminology, there is little magic involved: just don't be
scared of names. For further reading, we refer to the "Orientation Session"
in Bjiirner et al. [92, Ch. 11.
150
6. Duality, Gale Diagrams, and Applications
In this lecture, we give several striking applications of Gale diagrams and

oriented matroid duality, among them
the construction of an 8-polytope with 12 vertices that is nonrational,
meaning that it cannot be realized with rational vertex coordinates
(due to Perles),
the construction of a 5-polytope for which the shape of a 2-dimensional face cannot be prescribed (new!), and
the construction of a 24-polytope with 28 vertices for which the realization space is not connected.
These examples are "easily" derived from special low-dimensional point
configurations via Gale diagrams.
Finally, we develop the "Lawrence construction," a systematic method
by which properties of arbitrary point configurations can be encoded into
polytopes. This makes it possible to provide the corresponding "universality
theorems," which say that the realization spaces of convex polytopes can
be arbitrarily bad, in a sense that we make precise ahead.
6.1
Circuits and Cocircuits
C 1Y1 be a finite set of n points

For this section, let X = fx 1 ,
in affine space Rd for example, the vertices of a d-polytope. There is
no problem with multiple points, but we always assume that the points
in X affinely span Rd . We continue our habit to interpret X as a matrix
X rdx n when this is convenient.
We now explore two "dual" ways to derive combinatorial data from such
a point configuration.
(a) Affine Dependences

The affine dependences of the point configuration X are the vectors z E
with 11 z = 0 such that Xz = O. These vectors z form a vector subspace
of 18:
a-Dep(X) := {z Rn : Xz = 0, 11z = 0}.
Geometrically, let z 0 be an affine dependence. Then we can look at the
sets of negative and positive coefficients of z, N(z) := fi : z < 0} and
Z = EiEN(z) z > 0. By
P(z) := fi
> 0}. We get that A := E
multiplying with we can rewrite the affine dependence z as
E
ieP(Z)
_zixi =
A
E _zi
iEN(Z)
y,
151
so that
y
conv{x i : i E P(z)}
conv{z i : i C N(z)}
represents a point that lies both in the convex hull of the points with
positive coefficients and in the convex hull of the points with negative
coefficients.
( 0 3 5 5 2 0\
For example, let X =
be the vertex set of a
0
01221
hexagon
Then the affine dependence
4
6
z ,-----4
1
\ 0 )
represents y =
4
1
c conv{xi, x3, xs}
n convfx 2 , x41.
In particular, we are interested in the special case of affine dependences

involving a minimal set of points from X, that is, dependences of nonempty
point sets such that every proper subset is affinely independent.
We define the support of a vector as the set of components that are not
zero. Thus the minimal affinely dependent point sets correspond to the
dependences with inclusion-minimal supports.
For the minimal dependences, we see that conv{x i : i E P(z)} and
conv{x i : i E N(z)} are simplices whose relative interiors intersect in a
unique point y. Such configurations are known as "minimal Radon partifions" because of Radon's theorem, a quite trivial but basic lemma from
convexity theory; see Exercise 6.0.
It is not hard to show that every affine dependence z is a finite sum of
such minimal dependences, z' + z" +... + z(k ) . By Carathodory's Theorem 1.15 (see Lemma 6.7 below) the minimal dependences z(i) can be
chosen consistent with z, that is, such that the jth component of each z(i)
either has the same sign as zi or else it vanishes. Finally, the minimal affine
dependences determine the point configuration X up to affine coordinate
152
6_ Duality, Gale Diagrams, and Applications
change (Exercise 6.1). Together these three facts mean that the minimal
affine dependences completely determine the structure of the point configuration.
For the hexagon we considered before, the linear dependence can be
written as a sum
/ 1\
( 1\
l o\
5
_3
4
2
6
3
3
=
+
__
4
2
1
1
0
o)
\ o)
o)
\
which writes it as a sum of the two minimal dependences depicted here:
4
/
6
1
To distill the "combinatorial essence," we use the sign function

+ if z > 0,
z 1> sign(z) =- 0 if z = 0,
sign : R > I+, , 01,
if z < O.
We will apply the sign function componentwise to vectors, so for a column
vector z E IV we get a column sign vector sign(z) E I+, , Or, and for a
row vector C E (Rnr we get a row sign vector sign(c).
Definition 6.1. Let X = {x 1 , . . . ,x,,} C Rd be a set of n points in affine
space Rd .
The signed vectors of X are the column sign vectors sign(z) corresponding to affine dependences of the points in X,
-
V(X) :=-- {sign(z) : z E
, Xz =- 0, 11 z = 0} = SIGN(a-Dep(X)),
where SIGN(U) denotes {sign(x) : x E U} for any subset U C n .

The signed circuits of X are the column sign vectors sign(x) corresponding to minimal affine dependences of the points in X. The set of signed
circuits of X is denoted by C = C(X).
For example, the 6-point configuration discussed above has
/ + \
the vector
\ 0 j
[ + \
,
and
and the circuits
0
0 i
153
corresponding to the nornninimal dependence ("the segment [2, 4] intersects

the triangle [1, 3, 5]") and to the two minimal dependences ("the segment
[2,4] intersects the segment [1, 3], respectively the segment [3, 5]") that we
have calculated above.
The sets C(X) and V(X) of signed circuits and vectors thus are combinatorial data associated with any affine point configuration. We will now
proceed to extract the natural "dual data," called signed cocircuits and
signed covectors. In Section 6.2 we will derive the theoretical framework
that shows that these four types of data are all "equivalent," and that
makes the "duality" precise.
(13) Affine Functions

f (x) = ex z, for
From every affine function f on Rd , of the form x
c E ( ) , z E R, we get a row vector, called an affine value vector,
(f
(xi),
,f
(zn)) = (ex i z,
, cx r, z) = cX zll ,
which records the values of cx z on the points xi of X. For the hexagon

from our example, the affine function f(x) = x1 + 2x2 3 generates the
affine value vector (-3, 0, 4, 6, 3, 1):
where our drawing lists the values of f in brackets, and illustrates the
hyperplane Hf := {x c Rd : f (x) = 0} as a dashed line.
We note that the set
a-Val(X) := {cX zl G(
:cG ORd r , z E R}
Mar.
of affine value vectors is a vector subspace of

Geometrically, the row vector f(X) := cX
zit records "signed distances" of the points x i from the oriented hyperplane Hf . Thus the row
sign vector sign(f(X)) records which points of X are on the positive side
HIAFI
f{x ERn : f(x) > 0} of the hyperplane Hf, on Hf itself, or on
f
the negative side of Hf.
For example, for the hexagon and the affine function illustrated above, we
have 1(X) = (-3, 0, 4, 6, 3, 1), and thus sign( f (X)) = ( , 0, +, +, +, ).
154
Definition 6.2. Let X = {x 1 , .. . , x} c i be a finite set of n points in

affine space Rd .
The signed covectors of X are the row sign vectors cXz11 corresponding
to affine functions of the points in X,
V*(X) := {sign(cX zll ) : c E (Rd)*, z E RI = SIGN(a-Val(X)).

The signed cocireuits of X are the signed covectors of minimal support, for
which the hyperplane H = Ix c r n : cx d = 01 is spanned by points
in X. The set of signed cocircuits of X is denoted by C* -= C* (X).
In particular, it is easy to read off the faces of the polytope conv(X)
from the signed covectors. For this we identify each face of P with the
set of vertices not on it, that is, the face F is associated with the coface
vert(P)\vert(F). Similarly, we define a eofacet as the set of all vertices not
on one facet. Thus the cofacets are the minimal (under inclusion) nonempty
cofaces, and the cofaces are exactly all the unions of cofacets.
By slight abuse of language we will call a sign vector positive if it is nonnegative and nonzero, that is, if it lies in {0, +}\O, and similarly for row
sign vectors. So we talk about "positive signed covectors," which correspond
to the nonvanishing nonnegative affine functions on a point configuration.
So, if X C Rd , then the cofaces of conv(X) are the supports of the
positive covectors in V* (X). Moreover, the cofacets of conv(X) are the
supports of the positive cocircuits in C*(X). In particular, the face lattice
of conv(X) can be read off from V* (X) as the set of all supports of positive
covectors, ordered by inclusion. It can similarly be determined from C* (X),
since the cofaces are exactly the unions of cofacets.
For example, the drawing
illustrates affine functions x l + 5 and x l + x2 + 4 that determine the

positive covectors
(+, +, 0,0, +, +)
and
corresponding to the cofaces {1, 2, 5, 6 } and {1, 2, 4, 5, 6 } , and thus to the

vertex 131 and to the edge conv{3, 4 } . In principle, one can also read this off
from the circuits (they are determined by the cocircuits, see next section),
but this is not so straightforward.
155
Example 6.3 (Two octahedra).

Let Pi = C3 4 = conv{e i , e i , e2 , e2 , e3 , e3 } be the regular octahedron
in i
and let P2 be obtained by perturbing the vertex e l to e l + e2. P2 is
a nonregular octahedron, that is, Pi and P2 are combinatorially equivalent,
but not affinely isomorphic.
In matrices, we get Pi. = conv(Xi ) and P2 = conv(X2 ), for
,
1 1 0 0 0 0 )
Xi = ( 0 0 1
1 0 0
0 0 0 0 1-1 /
= (1 1 0 0 0 0 )
X2
bl 0 1 1 0 0
0 0 0 0 1 -1
We are not patient enough to list all the vectors and covectors. Here,
however, are complete lists of all the cocircuits: the reader is not expected to
check their details, but just to convince him/herself how this is constructed
"from the picture," and that it "seems correct." Does it?
(0
(0
(0
(0
+(0
C*(X1): (0
(+
(+
(+
(+
+ (+
0
0
+
+
+
+
0
0
0
0
0 0 + ---.),
+
00),
0 + 0 +),
0 + + 0),
+ 0 0 +),
+ 0 + 0),
0 + 0 +),
0 + + 0),
+ 0 0 +),
+ 0 + 0),
0 0 0 0),
( 0 0 0 0 + ),
(+ 0 + 0 0),
+ ( 0 + + 0 0),
+(0 0 + 0),
+ ( 0 0 + 0 ),
+(0 + 0 + 0 + ),
C*(X2):
(0
(0
(0
(+
(+
(+
(+
(+
+ 0
+
+ +
0 0
0 0
0 +
0+
0
+ + 0 ),
0 0 + ), +
0 + 0),
+ 0 +),
+ 0),
0 0 + ),
0 + 0
0 0 0),
),
156
Similarly, the circuits are the columns of the matrices we describe next,
and their negatives:
+ + 0 \
++0
0 +
0+
0
0
/
6.2
C X2 )
(
+ + + 0 \
+++0
-0+
0++
0
\ 0
Vector Configurations
While our discussion in Section 6.1 was on affine point configurations in Rd ,

we now proceed to linear configurations of vectors in r d+ 1 . The transition
is the obvious one: with any configuration of points xi in d (such as the
(x12 ) in Rd+'. To
vertices of a d-polytope), we associate the vectors vi
get our notation for dimensions straight, we introduce a new parameter,
called rank, as r := d + 1. As so often when dealing with a transition from
affine to linear, it is convenient to have an extra letter r for affine rank (i.e.,
linear dimension), which is one more than the affine dimension d. Thus we
have vectors v i E W.
In fact, what we get this way is an acyclic vector configuration V =
{v 1 ,. , v.,2 } C rr, characterized by the following two properties (which
are equivalent by a simple application of Farkas lemma II):
(i) There is no nonnegative dependence, i.e., no y > 0, y
Vy = 0.
0, such that
(ii) There is a linear function c E (W)* such that cV > O (i.e., cv, > 0
for all i).
The construction of the sets of signed circuits, vectors, cocircuits, and

covectors works for general vector configurations, acyclic or not. In fact,
the linear algebra becomes slightly simpler in the linear case.
There
For this, let V E Rr " be the matrix of a set of n vectors in
may be multiple vectors, but we assume that the vectors span Rr , so that
rank(V) ---= r.
157
The space of linear dependences of the vector configuration V is
Dep(V) := Iv E Rn : Vv = 0}
c W1.
This is a linear subspace of n of dimension n r. The signed vectors of V

are given by
V(V)
{sign(v) E {, , Or
:V E
W1, Vv = 0} = SIGN(Dep(V)),
and the signed circuits are the signed vectors of minimal nonempty support.
Dually, the space of value vectors on a vector configuration, which correspond to linear functions c E (W)*, is constructed as
Val(V) := {cV : c c Orr} C
).
This is a linear subspace of ( )* of dimension r. From it, we derive the

set of signed covectors of the configuration V:
V*(V) := {sign(cV) : e E (Rr ) * } = SIGN(Val(V)).

Here we get the signed cocircuits as the signed covectors of minimal nonempty support: they correspond to the linear functions such that the vec.
tors y c V which have value 0 linearly span a hyperplane in
Note that these definitions are consistent with our conventions for the
affine case; see Exercise 6.3.
Proposition 6.4. Let V E R" n represent a spanning configuration of n
vectors in Rr
Then Val(V) is the set of all linear functions that vanish on all the vectors
in Dep(V), and Dep(V) is the set of all vectors on which all the functions
in Val(V) vanish.
Proof. This follows from the dimension counts dim(Dep(V)) = n r,
dim(Val(V)) = r, with the computation (cV)v = c(Vv) = c0 = O.
(If we identify Rn and ( n ) * via the standard scalar product, then by
the same argument we get that Dep(V) is the orthogonal complement of
Val(V) in r .)
6.3
Oriented Matroids
In this section, we meet oriented matroids for the first time. Just to keep
things more exciting (and to make sure that the timid reader isn't overwhelmed by the first sight), we won't lift all the veils on the first encounter.
So we won't even define oriented matroids this time, but we promise more
for later.
158
For application to Gale diagrams, it is sufficient to know some basic facts

in the case of realizable oriented matroids; and we will only do proofs for
those below, with the Farkas lemma as a basic ingredient. The general case
is similar, but it relies on more notation and on the precise version of the
axioms to work with. So the main gain we get from oriented matroid theory
for the moment is some notation and terminology (that's where "circuit,"
"cocircuit," "covector," and so on are from) and the right intuition. Don't
underestimate the value of that.
Definition 6.5. Let V E W xn be a set of vectors that spans

The oriented matroid .A4(V) of V is the combinatorial structure encoded
by the following four collections of sign vectors:
the set of circuits of C(V),
the set of vectors of V(V),
the set of cocircuits of C*(V),
the set of covectors of V (V ).
The families of sign vectors arising from a vector configuration V in this
way are called a realizable oriented mat roid.
In the following, we will use small capitals, like x, u, v E {-1-, , Or and
C E ({+, , }n)*, to denote (column or row) sign vectors.
For a simple 2-dimensional configuration V we read off as follows:
circuits: C(V) =
{ ()'(--)}'
vectors: V(V) =
f() , -
o0 } '
(--) , ( 0)
cocircuits: C*(V) =
{(0++),
(0---),
(+0), (-0+), (++0),
(- -0)}
covetrs:
V*(V) =
{ (+++), (----), (++-), (--+), (+--), (-++),

=
(0++), (0--), (+0), (-0+), (++0), (--0),

(000)
In this section, we want to make four main points:

The data of an oriented rnatroid are highly structured.
All four sets of data are equivalent.
Duality is built into the system.
The basic constructions of "deletion" and "contraction" are oriented
matroid operations that are dual to each other.
159
(a) Axiomaties
The collections of sign vectors that make up an oriented matroid are highly
structured, and not just random collections. In fact, they arise in the following way.
For any linear subspace U C R , we define the operator SIGN as
SIGN(U) := {sign(x) : x G U} c
Or.
There is a natural partial order on the set of signs {-1-, 0}: we set 0 <
and 0 < , while + and are incomparable:
0
This corresponds to the fact that a number that is slightly perturbed either
keeps its sign, unless it is zero, in which case its sign can change to -1-, or
to , or remain O.
On sets of sign vectors S C { +, ,0}n, we use componentwise partial
ordering: u < u' if and only if ui < u't holds for all positions i. Thus we
get, for example,
(o+o-k000--+ ) < (0+++O--+)
but
(0+ 0 +000--+)
because of the second position, where +
O.
We use the componentwise partial order to define the operator MIN,
which takes all the minimal nonzero sign vectors in S:
MIN(S) :=
{I; E S\O : there is no u` < u with tif E S\0}.
We will apply the operator SIGN equally to sets of row vectors and of
column vectors. Similarly, we apply MIN both to sets of row sign vectors
and of column sign vectors.
With these conventions, we can describe the oriented matroid data for V
as follows:
V(V) = SIGN(Dep(V)) C(V) = MIN(SIGN(Dep(V))) = MIN(V(V))

V*(V) = SIGN(Val(V)) C*(V) = MIN(SIGN(Val(17))) = MIN(V*(V))
Furthermore, the spaces Dep(V) and Val(V) determine each other by Proposition 6.4, so all four sets of data are determined by U := Val(V). In this
sense, we talk about "the oriented matroid M M(U) of the subspace
U C Rn." The dimension r := dim(U) is called the rank of M. So what we
160
were studying in Sections 6.1 and 6.2 were the vectors and circuits of two
realizable oriented matroids of rank r respectively ri r,
M = M(Val(X))
M* = M(Dep(X)).
and
It is easy to write down extensive lists of axioms that are satisfied by any
collection SIGN(U). So one gets to axiom systems for oriented matroids.
We'll get to that in Lecture 7, but without much detail. In fact, the proofs
relating various axiom systems for oriented matroids tend to involve hard
work, something we try to avoid on this show. (This is, however, at the
basis of oriented matroid theory; we refer to [92, Ch. 3]).
The oriented matroid of a point configuration is a delicate model for its
geometry. It provides a much finer model than what the matroid encodes
about a vector configuration. (If you want to know what a matroid is, see
Welsh [518], White [520], or Oxley [400].) One can prove that in fact the approximation of the combinatorial model to "geometric reality" is extremely
good this is made precise in Lawrence's "topological representation theorem." We'll get back to this in Lecture 7.
This means also that all the main features of the geometry of vector
configurations can be derived from formal properties (the axioms). In fact
there are very few geometric statements that would be true for vector
configurations but fail for oriented matroids so whatever we find in that
direction is even more exciting. (See Theorem 7.20 for an example.)
(13) Equivalence
Different sets of data "A" and "B" about a geometric situation are equivalent if any two configurations with the same data A also have the same
data B, and conversely. This means that (at least in principle) one can
construct the data A from the data B. Any set of data that is equivalent
to the set of circuits is referred to as the oriented matroid of V.
We now want to show that the four sets of data for an oriented matroid given by Definition 6.5 are equivalent. For that we need to define the
combinatorial analogue of the condition cx = 0.
Definition 6.6. Let x G {,,O} n and C E ({,
vectors. Then we define that "c-x = 0" if
for each i, we have
Orr be two sign
= 0 or x, = 0,
or there are indices i, j with ci =

For a family of sign vectors S C
:= fc E ({+,
}
n
0 and ci = x i 0.
Or, we define
) * C-U = 0 for all u E SI,
and analogously for collections of row vectors.
6 3 Oriented Matroids
161
In fact, the condition c-x = 0 holds if and only if there are real vectors
E Rn and c E (Rn)* such that sign(x) = x, sign(c) = c, and cx =0. For
example, we have
+)
(+00) ( = 0,
but
(+00+) )
0.
The reader might want to check the next two statements for the small
3-vector configuration after Definition 6.5. They are both solid theorems
in the setting of oriented matroids [92, Ch. 3]. For the realizable case, we
won't have a lot of problems with them.
Lemma 6.7. Let U C n be a vector subspace of dimension r, and let
U E U be a vector with sign(u) = u E SIGN(U) C {+, , O}.
The vector u can be written as a finite sum u = u1 + .. . + uk of k < r
vectors Ui G U whose sign vectors ui := sign(ui) are below u and minimal,
that is, such that ui < u, and Ui G MINSIGN(U)).
Proof. The vectors y E U whose sign vector sign(v) is componentwise
smaller than or equal to sign(u) form a polyhedral cone:
C(u) :=
E U : sign(v) < sign(u)} C
U.
This cone is in fact pointed: all vectors X G C(u) satisfy Ein_ i uixi > 0,
with equality only for z = O.
Thus P(u) :=
E C(u) : Etn_ i uixi = 1} is a polytope of dimension
at most r 1. By the results in Lecture 1 (with Carathodory's Theorem
1.15) every point in P(u) can be written as a convex combination of at
most r vertices. By linearizing we get that u is the sum of k < r vectors on
extreme rays (1-faces) of C(u). Finally, we observe that the sign vectors on
the proper faces of C(u) are strictly smaller than u, and thus the minimal
El
nonzero sign vectors are precisely found on the extreme rays.
Proposition 6.8.
For any vector subspace U C ilkn we have
(MIN(SIGN(U))) -L = (SIGN(U))" = SIGN(U -1- ).

Proof. Let u E MIN(SIGN(U)) and C E SIGN(U-L ). Then we can find
u E U and c E U-L with sign(u) = u and sign(c) = c. Since Cu = 0, we
get c-u = O by definition. This implies
(MIN(SIGN(U))) -L
(SIGN(U))
SIGN(U ).
For the converse, let c c I+, , Oln \SIGN(U -L ). Then the conditions
cE U-1- ,
sign(c) = c
162
a system consisting of linear equations and strict inequalities

have
no solution. Now we use a Farkas lemma, for example as follows.
Write U in the form U =
lin(V), and define index sets
Z := fi : c = 01, P := : c > 01, and N :=
: c < 01. With this,
c SIGN(U-L ) says that the system of inequalities and equalities
chik = 0
di > 0
di < 0
di = 0
for
for
for
for
1< k<r
iE P
icN
iE Z
has no solution d E (Er )* . Since every positive multiple of a solution d

for this system would be a solution as well, we get that equivalently the
following system has no solution:
dv k = 0
di > +1
< 1
= 0
for
for
for
for
1< k<r
iE P
N
i E Z.
Now we apply the Farkas lemma (Proposition 1.7 adapted for systems with
inequalities and equations, see Exercise 1.6) to get existence of
eixì` = 0 and
for i E N
O for i C P <
EiEp Xj
EiEN <O.
Now letting u := Ve, we get that u E U with c-sign(u) 0 in fact, we

get "c.sign(u) > 0" in the obvious sense.
This proves that C
(SIGN(U)) . Furthermore, if we decompose u =
u1 . . . uk into minimal vectors according to Lemma 6.7, then we find
that ui -c L 0 has to hold for some i, and this yields a certificate to see that
c (MIN(SIGN(U))) .
El
Corollary 6.9. For any vector configuration V E Rrxn , the four sets of
data given by Definition 6.5 determine each other (denoted by it> ,7 j,1 as
follows:
Dep(V)
Val(V)
SIGN
SIGN
vectors V (V)
covectors V* (V)
MIN
MIN
C(V) circuits
C* (V) cocircuits
Thus any of the four sets of data determines the other three, and thus also
the oriented matroid M(V).
163
Proof. We get this by applying Proposition 6.8 both to Dep(V) and to

Val(V), which are dual to each other by Proposition 6.4.
D
Therefore two (labeled) configurations of vectors or points have the same
oriented matroid if they have the same set of circuits, or (equivalently) the
same set of cocircuits, the same set of vectors, or the same set of covectors.
(c) Duality
A concept of duality is built into the whole structure of oriented matroids.
In fact, since the vectors and the covectors of an oriented matroid M(V)
arise in the same way as the sign vectors of a subspace, they also satisfy
the same axioms (ignoring a switch from row vectors to column vectors).
Definition 6.10. The dual of an oriented matroid M is the oriented
matroid M* with the following properties:
The vectors of M* are the covectors of M, and thus
the circuits of M* are the cocircuits of M.
The covectors of M* are the vectors of M, and thus
the cocircuits of M* are the circuits of M.
From the way we have defined (realizable) oriented matroids, it is clear

that for every oriented matroid M, there is a unique dual oriented matroid
M*, whose dual is
(M * ) * = M.
In fact, M =-- .A4(U) is realizable with U C
the orthogonal space
then M* -=-- .A4(U"), for
U := {c E ( lin)* : c.x = 0 for all x E U}.

So existence and uniqueness of the dual follow from Proposition 6.4 (in the
realizable case):
M* = M(Dep(V))
for
M := M(Val(V)).
Furthermore, if M has rank r, then M* has rank n r, and conversely.
(d) Deletion and Contraction

There are two very natural and fundamental operations on point configurations: deletion and contraction. They are dual to each other, and they
directly translate into oriented matroid language and terminology.
164
V. We can
Consider a vector configuration V E r ", and let u
certainly delete ui from V, to get the new configuration V\ui .
-
We get value vectors on V\ui from those on V by simply deleting the

components corresponding to u while dependences and circuits of V\u,
are those dependences, respectively circuits, of V which do not involve ui
(i.e., have zero i-component). This proves the following result, where we
label the vectors in V (the "ground set" of M(V)) by 11,..., n1, so we
write .A4V for M(V\ui ).
,
Proposition 6.11.
The oriented rnatroid M(V)\i of V\u i is given as
follows:
V(V\ui) = {Y\i : y E V(V), = 0} V* (V\ui) ={ \TV : y E 1,* (V)}
C(V\ui) = fc\i : c E C(V), ci = 01 C*(V\ui ) = MIN{cV : C E C*(V)}.
for this we project V parallel
The dual operation is the contraction of
to ui to some hyperplane that does not contain ui . If u = 0, then we just
delete ui .
Algebraically, we can do this by choosing a linear function c E

such that cu i 0 (for example, c := ul will do), and map
17.L3
:=
CIL 3
cui
tLi
This yields the new configuration
:=
, fti-vi, -
165
in the hyperplane { v c W : cv = 01.

We get dependences on Viui from those on V by simply deleting the
components corresponding to u while value vectors of Viui are those
value vectors V which are zero on ui (i.e., have zero i-component).
,
Proposition 6.12.
follows:
The oriented matroid M(V)/i of V/u i is given as
V(V/ui)= {Y\i : V c V(V)}
V*(V/ui) = fv\i : y E V*(V), y = 01
C(Viu i ) = MIN{c\i : C E C(V)}
C* (V/ui) = {C\i : C E C*(V), e = 0}.
Let us mention two examples that show how deletion and contraction
appear in connection with polytopes.
Examples 6.13. Let P C Rd be a polytope, let X := vert(P) be its
vertex set, and let V E R"n be the corresponding vector configuration
in
(r = d+1).
If F is a face of P, then the vector configuration for F is obtained by
deleting from V all the vertices that do not lie on F.
If x E vert(P) C Rd is a vertex of P and vEVC Rr is the corresponding
vector, then the vector configuration of the vertex figure Plx is the contraction V/v. (In this case the projection hyperplane for the contraction
can be taken parallel to the hyperplane spanned by P.)
Note that by contracting we get a vector configuration that may also represent a lot of interior points of the vertex figure, corresponding to vertices
w E vert(P) such that [w, y] is not an edge of P.
6.4
Dual Configurations and Gale Diagrams
Now observe that Dep(V) C Rn determines the configuration V E Rr "

of column vectors uniquely up to coordinate transformations in r, which
correspond to row operations on the matrix V. Thus the dual space Val(V)
determines a configuration of row vectors in (Rn)*, which completely
encodes the vector configuration V.
166
Theorem 6.14 (Dual configuration).

Let V E R"n be a configuration of n column vectors in W. .
Then there is a matrix G E Rnx(n-r) of n row vectors in (W.')*, such
that
Val(V) = lc C ( r ) : cG = 01
and
Dep(V) = {Gx : x E Ile -r }.
The configuration of row vectors G is uniquely determined by either of the

two conditions, up to linear coordinate transformations in (1R71-7 ) * , which
correspond to column operations on the matrix G.
Proof. The matrix G E Rnx (n-r) has to satisfy
rank(G) = n r
and
VG = 0,
where 0 is the zero-matrix in R(-T).

For computation, this means that we have to find a basis for the orthogonal complement of the space spanned by the rows of V in (in)*.
This is computationally easy: it only requires to get V into a normal
form like V -= (iriM), so that the dual configuration can be obtained
as G :=(J/I )
n-r
Existence of G, and uniqueness up to column operations, follows from
this.
0
If we define the spaces of dependences and of value vectors for configurations of row vectors in exact analogy to the case of column vectors, then
we can read Theorem 6.1 as saying that there is a dual configuration G,
essentially unique, such that
Dep(V) = Val(G)
and
Val(V) --= Dep(G).

The following corollary has the additional information that the combinatorics (in particular, the circuits and cocircuits) of the vector configuration V can be read off not only from G, but in fact from the combinatorics
of G.
Corollary 6.15. The circuits of a vector configuration V E R""- are the
cocircuits of the dual configuration G E l':nx(n -r) , and vice versa.
In particular, the oriented matroid of the dual configuration is determined by the oriented matroid itself; it is the dual oriented matroid
M(G) = (M(V))*.

Proof. From Corollary 6.9 and Section 6.3(c).
167
0
In particular, this implies that the dual of an acyclic vector configuration

is a totally cyclic configuration G = ig1 , ... , gn } of row vectors: this property is characterized by the following two properties (which are equivalent
by the Farkas lemma):
(i) There is no nonnegative value vector, i.e., no x E Dr' with Gx > 0
and Gx
O.
(ii) There is a positive dependence, i.e., some c> 0 with cG -= O.
Comparison between these descriptions of "totally cyclic" and the corresponding ones that we have given for the dual concept "acyclic" on page 156
might give a feel for how the translation between dual concepts works on
the linear algebra level.
On the combinatorial side, we derive the following.
Corollary 6.16. A vector configuration V is acyclic if and only if the
following equivalent conditions hold:
(i) M(V) has no positive signed circuit.

(ii) (++. . . +) is a signed covector of ..A4 (V).
(iii) Every i is contained in a nonnegative cocircuit.

Dually, a row vector configuration G is totally cyclic if and only if the
following equivalent conditions hold:
(i) M(G) has no positive signed cocircuit.

(+
(ii)
is a signed vector of M(G).
+
(iii) Every i is contained in a nonnegative circuit.
In particular, a vector configuration V is acyclic if and only if its dual
configuration G is totally cyclic, and conversely.
Now we'll put the pieces together.
168
Definition 6.17 (Linear and affine Gale diagrams).

Let P = conv{xi, , x n } be a d-polytope in d with n vertices. A Gale
diagram and an affine Gale diagram of P are obtained by the following
sequence of operations.
d-polytope with n vertices x i (1 < i < n)
n vectors in
(polytope)
d+1 , vi := (xt )
* (oriented matroid duality)
n vectors g i
nr
-d-1
n (signed) points ai in affine (n d
Gale diagram
2)-space affine Gale diagram
Here the passage from Rd to Rd+1 is the usual embedding, used to linearize the situation. The dual configuration of this vector configuration is
a Gale diagram for P, determined uniquely up to a change of coordinates.
For the reduction of (Rn()* to (i n-d-2 ) * , we find a suitable vector
y E Rn-d-1 such that g i y 0 unless gi = 0, for all i. Then we associate
with gi the point
ai :=
92
fc c (Rn-d-1)*
cy 11
9iY
which we call a positive point in the affine space 11 -d-2 if gi y > 0, and a
negative point if gi y < O.
This yields the affine Gale diagram, a labeled point configuration
{a i ,... an } C Rn-d-2
where the point ai is labeled i if it is a positive point, labeled if it represents a negative point, and not represented by a point (or, represented by
a "special" point) in the case where g i = O.
-d-2 ) * does not lose combinatorial information: the
The reduction to (
circuits and cocircuits of this affine point configuration still represent the
cocircuits and circuits of P. This is extremely useful for polytopes with
"few vertices," where n d is small, as we will see in the following section.
Let us consider one example, to illustrate the technique.
169
Example 6.18. For the octahedra of Example 6.3, we have the matrices
vi =
1 1 1 1 1 1
1 1 0 0 0 0
0 0 1-1 0 0
00001-1)
1-1
I6 0
0 0
v2=
1 1 1 1
0 0 0 0
1-1 0 0
00 1-1
and compute Gale transforms

0
1
0
1
1
1
\ 1
/
G2=
GI =
1 \
1
0
i
_ i_
1
_ 13
12
From this we can directly draw Gale diagrams (they are 2-dimensional),
and derive 1-dimensional affine Gale diagrams, for y = (1). Here they are,
linear and affine.
We use the convention for affine Gale diagrams that black dots denote
positive points, while white dots denote negative points.
It is really important that the reader figure out how to read off the circuits
and the cocircuits of the octahedra from their affine Gale diagrams: he or
170
she might appreciate that this is a very compact (1-dimensional!) encoding

of the combinatorics of 3-dimensional geometric figures.
The basic pattern is as before. Namely, from affine linear functions we
read off sign vectors like
\ - /
which are cocircuits for both Gale diagrams, and thus circuits of the octahedra. Similarly, from minimal affine dependences we read off sign vectors
like
(+ 00 + + 0),
which form circuits for the Gale diagram, and thus cocircuits for the octahedra. The only new feature is that the sign is reversed for any negative
point in the diagram.
Every spanning set G = fg 1 , . . . , gii } of n row vectors in (I
r can
be interpreted as the Gale diagram of a vector configuration of n vectors that span Rd . However, these vectors need not come from a (d - 1)polytope: the vector configuration might not be "affine" (acyclic), and even
if it is, the vectors need not be in convex position. However, there is a
simple combinatorial condition that characterizes Gale diagrams, see the
following theorem. It is important because it allows us to conclude the
existence of a (high-dimensional) polytope with specific properties from a
(low-dimensional) configuration of signed points.
) *
Theorem 6.19 (Characteristic property of Gale diagrams).

A matrix G G Rnx(n-r) of row vectors (of full rank n-r) is a Gale diagram
of a (r - 1)-polytope with n vertices if and only if every cocircuit has at
/east two positive elements.
Proof. Every spanning configuration G of row vectors is the dual configuration of some spanning vector configuration V in I. The configuration V
is acyclic if and only if G is totally cyclic, that is, if G has no negative
cocircuit: every cocircuit of G has at least one positive element.
With this, we can scale the vectors of V, without changing the combinatorics, such that V comes from a point configuration in some affine
,.:' I! r -1 in T r. The points in H are in convex position unless
hyperplane He_.`
one point is in the convex hull of the others, that is, unless V has a circuit
with exactly one positive element, and thus G has a cocircuit with exactly
D
one positive element.
It is easy to translate this condition to affine Gale diagrams. Check it for

the affine Gale diagrams of the two octahedra given earlier in Example 6.18!
171
It yields a criterion that is very easy to check "by inspection" for the
interesting case n d = 4; see below.
Corollary 6.20 (Characterization of Gale diagrams of polytopes).
A configuration A = {a 1 ,.
an } points in (Rn-d-2 ) * , each of them declared to be either "positive" or "negative," that affinely spans ( n-d-2)* 1
is the Gale diagram of a d-polytope with n vertices if and only if the fol.
lowing condition is satisfied: for every oriented hyperplane H in (IV -d-2 )*

spanned by points of A, the number of positive A-points on the positive
side of H, plus the number of negative A-points on the negative side of H,
is at least 2.
In our descriptions we have disregarded the case of "special" points:
they are just the cone points, so adding k special points to the diagram
G corresponds to taking the k-fold pyramid over the polytope represented
by G.

Any d-polytope with d + 1 vertices is a d-simplex: this we know and have
seen before. In this case the Gale diagram is in 0-dimensional space, so all
vectors are 0-vectors trivially.
Next consider the case of d-polytopes with d + 2 vertices. The result is
that there are 142 /4] combinatorial types of d-polytopes with d+2 vertices.
Of those, 14/2J represent simplicial polytopes, and the others are (multiple)
pyramids over simplicial polytopes of this type. The case of d + 2 vertices
is classic and can be found in Schoute [4441 and Sommerville [470]; see also
Gri_inbaum [234, Sect. 6.11 or Ewald [189, Sect. 2.61.
The affine Gale diagrams representing d-polytopes with d + 2 vertices
are 0-dimensional and may be represented by a "cloud" of positive points
(black), negative points (white), and special points (grey). The condition of
Corollary 6.20 requires that there are at least 2 black and at least 2 white
points. Furthermore, interchanging black and white points does not change
the combinatorial type of the polytope.
Thus we get the following complete enumeration for d = 3, n = 5:
e
represents the bipyramid over a triangle (this is a simplicial polytope with
6 facets), and
yields the square pyramid (a nonsimplicial polytope with 5 facets).

Similarly, you should analyze and classify the 4-polytopes with 6 vertices
in Exercises 6.8.
172
For polytopes with d + 3 vertices, the complete classification becomes

difficult, but not out of reach. Their affine Gale diagrams are configurations
of signed points on a line. The main problem is to decide which different
diagrams represent combinatorially equivalent polytopes. First results were
due to Gale, while Perles developed the (Gale diagram) techniques necessary to analyze polytopes with d + 3 vertices. The special case of simplicial
polytopes was done in Griinbaum [234, Sect. 6.2]; the work was completed
by Mani [349] and Kleinschmidt [305]. Explicit formulas for the number
of d-polytopes with d + 3 vertices were obtained by Perles [234] for the
simplicial case and apparently by Lloyd for the general case [344] (though
there is some doubt whether his formula is completely correct). Note that
the octahedra that we considered before have d + 3 vertices, for d = 3.
The polytopes with d+3 vertices still do not have any unusual properties.
Finally, we arrive at the case of polytopes with d + 4 vertices: here is
the threshold for counterexamples, as Sturmfels [4951 calls it. They can be
analyzed in terms of planar point configurations which can be arbitrarily
complicated. In particular, for high enough d there are
d-polytopes with d + 4 vertices that do not have a realization with
rational coordinates,
d-polytopes with d + 4 vertices for which the shape of a facet cannot
be prescribed, and
d-polytopes with d + 4 vertices that have disconnected realization
spaces.
We will now describe Gale diagram approaches to these three phenomena.
(a) A Nonrational 8-Polytope

Using Gale diagrams, Perles has shown that there are nonrational polytopes, that is, polytopes for which there are no combinatorially equivalent
polytopes with rational coordinates.
Example 6.21 (Perles). [234, p. 95] [92, Fig. 8.4.1]

There is a nonrational 8 polytope with 12 vertices.,
To see this, one verifies that the configuration G in the figure on the next
page has three properties:
-
1. G cannot be realized with rational coordinates "as a matroid": there

is no rational planar configuration of 12 points such that the same sets
of points as in G coincide, respectively are collinear. (Essentially, there
is a golden ratio involved in the construction of a regular pentagon,
so it can only be realized with coordinates in a field containing 1/6.)
6 5 Polytopes with Few Vertices
173
2. G is the Gale diagram of an 8-dimensional polytope with 12 vertices

(check this!); a polytope represented by G cannot have all coordinates
rational.
\%11;W
irapi
69
mmimp-
10
3. Consider any spanning configuration G' of 12 signed points in the

plane (an affine Gale diagram in R2). If G' has the the same positive
circuits as G, then the three pairs of points that coincide in G have
to coincide in G' as well, and the triples and quadruples that are
collinear in G have to be collinear in G' as well, because they are all
positive vectors (unions of positive circuits).
Thus also G' cannot be realized with rational coordinates.
Thus G is the Gale diagram of a nonrational polytope P. If P' is a polytope
that is combinatorially equivalent to P, then its Gale diagram G' has the
same positive circuits as the Gale diagram G, hence with part 3. above G'
and P' cannot be rational either.
(b) Facets of 4-Polytopes Cannot be Prescribed

Perles apparently first observed that the shape of a facet of a d-polytope
cannot in general prescribed; see Griinbaum [234, p. 96, Ex. 3]. Kleinschmidt [304] finally constructed a 4-polytope with 8 vertices for which the
shape of a facet cannot be prescribed -- this is the smallest dimension and
the minimal number of vertices for such an example, because all facets can
be prescribed for 3-polytopes, and for all d-polytope with at most d+ 3 ver4 and n = 8, Kleinschmidt's polytope can be constructed
tices. With d
174
as a 2-dimensional affine Gale diagram (Exercise 6.18). It has the special

property that all facets except the "bad" one are simplicial.
In Example 5.11 we saw "Barnette's example," which has a minimal
number of facets, namely 7. In fact, if P is a prism over a square pyramid,
then the shape of its cubical facet cannot be prescribed. The square pyramid Pyr3 is isomorphic to its polar. Thus we get that PA , a bipyramid
over a square pyramid, is a 4-polytope with 7 vertices for which a vertex
figure cannot be prescribed.
The following constructs an (equivalent) Gale diagram description of
Barnette's example [43], which SturmfeLs [495] found independently of Barnette's work.
Example 6.22. [495, Prop. 5.1]

There is a 4-polytope P with 7 facets for which the shape of a facet cannot
be prescribed.
For this, let PA be the bipyramid over a square pyramid, as given by
V=
/1 1 1 1 1 1
1\
1 0-1 0 0 0
0
0 1 0 1 0 0 0
0
0 0 0 0 1 1
\ 0 0 0 0 0 1 1 /
/0 0 2 2
1 0-1 0
0 1 0 1
0 0 2 2
\O 0 2 2
0
0
0
1
0
0
0
0
0
1
1 \
0
0
1
0
0 /
from which we read off a Gale diagram

o\
f'
0
1
1
0
0
1
G =
2
2
2 2
2 2 /
Now we observe that the vertex figure at the vertex 5 is an octahedron.

Its Gale diagram is obtained by deleting the point 5 from the diagram: it
corresponds to a regular octahedron (compare to Example 6.18).
Now for PA we see that 5 6 and 57 are positive cofacets, which requires
that the points 6 and 7 coincide on the affine Gale diagram of the vertex
figure PA /5. Thus, if we start with the nonregular octahedron with Gale
diagram
6 5 Polytopes with Few Vertices
175
then this is not the vertex figure of a 4-polytope that is combinatorially

equivalent to P.
(c) 2-Faces of 5-Polytopes Cannot be Prescribed

We have now seen the second proof that the shape of a facet cannot be
prescribed for 4-polytopes, so we'll have to start and accept it.
But perhaps this was the wrong generalization of the 3-dimensional theorem that a facet can be prescribed. What about prescribing the shape of
a 2-face for a d-polytope? The case d = 4 was an open problem in the first
edition of this book (Problem 6.11*), now you can find it as Exercise 6.11.
For d = 5 we have the following counterexample from the first edition. It is
the type of analysis that Gale diagrams "were made for": however, the construction of this example may have been the first time that a 3-dimensional
affine Gale diagram was seriously used.
Example 6.23. There is a 5-polytope P with 10 facets and 12 vertices,

for which the shape of a 2-face cannot be prescribed.
To prove this, we construct the polar polytope Q := P , a 5-polytope
with 10 vertices and 12 facets, and verify that for one contraction (face
figure, see Exercise 2.9) of a 2-face, the shape cannot be prescribed.
For this consider the signed point configuration given by the vertices of
a triangular prism as positive points, labeled 2, 3, ... , 7, the centers of the
square facets of the prism as negative points 8, 9, 10, and the center of the
whole prism as another negative point 1. See the figure for how we label
this.
It is easy to give coordinates. In fact, the corresponding vector configuration

in R4 could be taken to be
(
1 1 1 1
1 2 2 2 0 0 0
2 0 0 6 0 0 6
2 0 6 0 0 6 0
1
1
0
3
1 )
3
3
3
0
176
Now we check the following facts, which together imply all we need to know.
1. This signed point configuration is the affine Gaie diagram of a 5
poiytope Q with 10 vertices.
For this we check that every cocircuit of this configuration has at least
two negative and two positive elements.
2. The triple (8, 9, 10) describes a triangle F --= [8, 9, 10] which is a 2 face
of the 5-polytope Q.
The point 1 is in the interior of the prism, so the points different from
8,9, 10 support a positive circuit (+++++++000).
3. The face figure Q / F is a hexagon whose diagonals cross.
In fact, we get the affine Gale diagram of Q / F by deleting the points
8, 9, 10 from the diagram for Q. But what is left then is the affine Gale
diagram of a hexagon [2, 6, 4, 5, 3, 7] with 1 as the intersection point of the
long diagonals:
7
4. Q is a poiytope with 12 facets.

The facets are the (convex hulls of the points corresponding to the)
complements of the positive circuits in the diagram, which are easily enumerated as
286
23157
385
23167
2107
24156
4105
24167
397
34156
496
34157
5. Every Gale diagram G' with the same positive circuits contains the
diagram of the hexagon with crossing diagonals.
Consider any other (linear) Gale diagram G' on 10 points with the same
positive circuits. From the 3-point circuits we see that the sets 2'4'5'7'10',
2'3'5'6'8', and 3'4'6'7'9' have to be planar. However, they cannot collapse to
be on a line, because then the whole diagram would collapse to a plane and
couldn't have 5-point circuits. From this one can show (using projective
uniqueness of the triangular prism here we are skipping the detailed
arguments) that in suitable coordinates G1\1' coincides with G\1: the Gale
diagram G' consists of a triangular prism and its facet centers as well.
Now the 5-point circuits imply that the point l' has to be in the interior
of this triangular prism. Hence, if we consider the diagram G'\{8', 9', 10' } ,
then this has the 5-point circuits listed above (so it describes the right
hexagon), and it has 3-point cocircuits 1', 2', 5', 1', 3', 6', and 1', 4',7' (so
iii
the long diagonals of the hexagon cross in 1', as required).
6_5 Polytopes with Few Vertices
177
An explicit geometric description of the polytope P = Q is given in

Exercise 6.27.
(d) Polytopes Violating the Isotopy Conjecture

Recall from Section 4.4 that the "realization space" of a polytope is an
elementary semialgebraic set, and that elementary semialgebraic sets can
be arbitrarily complicated spaces: they can be disconnected, with holes,
and so forth (Exercise 4.22).
However, Steinitz' Theorem 4.11 states that for every 3-dimensional polytope the realization space 1R(P) is contractible, and thus connected. (To
get this right, we had to fix an affine basis in our Definition 4.10 of realization spaces, to make sure that the "reflection" doesn't create a second
component of the realization space.)
In other words, any two 3-polytopes of the same combinatorial type and
orientation can always be continuously deformed into each other, such that
each intermediate object is a 3-polytope of the same combinatorial type.
The same is not hard to show for d-polytopes with at most d + 3 vertices,
using Gale diagrams.
However, this "isotopy property" fails even for 4-polytopes: Kleinschmidt
constructed a 4-dimensional example with 10 vertices [107]; its combinatorial type is obtained by glueing two copies of the 8-vertex Kleinschmidt
polytope of Exercise 6.18 in their octahedron facets in an "incompatible
way." See Mil& [378, p. 530] and Bokowski & Guedes de Oliveira [108].
A systematic construction method for 4-dimensional counterexamples is
provided by Richter-Gebert's Universality Theorem for 4-polytopes [424].
Here we start a construction with a planar point configuration, where the
isotopy conjecture [427] [216] fails and transfer this result to polytopes.
Example 6.24 (A nonisotopic 24-polytope with 28 vertices).
The "isotopy property" fails for d-polytopes with d+4 vertices: for that, we
can start from a planar point configuration that does not satisfy isotopy.
The smallest nonisotopic planar point configurations that are currently
known have 14 points. The first such configuration was found by Suvorov
[500] [92, p. 363]. Here we present the nicest and newest one, found by
Richter-Gebert [423].
For this we refer to the figure on the next page: it represents RichterGebert's example, with two points drawn "at infinity" (which we interpret
as "very far away," to get an affine configuration).
The key property of the figure can be seen as follows. Try to construct a
new configuration of 14 points in the (projective) plane, with the condition
that point sets that are collinear in the old figure also have to be collinear in
the new figure. After a projective transformation, we may assume that the
points 1, 2, 3, and 4 (a "projective basis") are positioned as in the old figure.
178
Then we choose the point 5 on the diagonal through 3 and 4. Thus we have
five points which then successively determine the points 6, 7, ... , 13. The
fourteenth point 14 is then placed at the intersection of the lines 1,3 and
2, 4, but not on the line 12, 13: this is possible if the point 5 had been drawn
slightly Southeast of the center, as in our figure, or if it is taken slightly
Northwest of the center, which results in a figure that is a reflection of our
figure; however, it is impossible (it results in a different configuration, with
12, 13, 14 collinear) to get a realization of the same configuration which is
y diagonal: for that we'd have to
itself symmetric with respect to the z
choose 5 in the center, and would get 12, 13 and 14 collinear.
----
1'
*1
'41!illP
5 8*
13
10
.
3
14
12
A configuration with this effect is not too hard to construct, but this
example has a stronger property: If the point 5 is chosen close enough
to the midpoint of the segment [3, 41, then the "Southeast" and "Northwest" realizations not only have the same collinearities (and thus the same
unsigned cocircuits, the same matroid), but they have the same signed cocircuits
both realizations yield the same oriented matroid. We do not
know whether one really needs 14 points for this effect: can you do with
less (Problem 6.26*)?
From this, one can easily construct an affine planar Gale diagram that
has a disconnected realization space. However, we must also make sure
that every polytope that is (only!) combinatorially equivalent to the one we
construct has the same diagram. A very "aggressive" method is to replace
every point of the diagram by a pair of positive and negative points: this
yields the affine Gale diagram of a 24-polytope with only 28 vertices that
has two isotopy classes of realizations. Implicitly, this is what the "Lawrence
construction" does, which we will discuss below.
179
Another nonisotopic point configuration, of 21 points, is explicitly constructed in White [522]. A general construction method is due to Mnv [378].
In particular, Mnv's universality theorem (see below) shows that the "realization space" for planar point configuration can be arbitrarily complicated.
Suvorov [5001 and Jaggi et al. [266] furthermore construct non-isotopic
configurations whose points are in general position. From these, one can
using a technique of Sturmfels [110, Thm. 6.5] get simplicial polytopes
that violate the isotopy conjecture. This yields examples for a much more
general "universality theorem for polytopes," described in the next section.
6.6
Rigidity and Universality
We have characterized Gale diagrams of polytopes in Corollary 6.20. To

make statements about polytopes of a fixed combinatorial type, however,
it is not sufficient to look at a specific Gale diagram: we have to make
sure that the statement we make holds for all polytopes combinatorially
equivalent to the given polytope.
Here we have to deal with two different notions of equivalence. We have
noted that two polytopes may be combinatorially equivalent but have different oriented matroids. (See, for example, the octahedra of Example 6.3.)
However, any two polytopes with the same oriented matroid are combina
equivalent: they have the same covectors, hence the same positive-tonalyequivalent:
covectors, and hence the same cofaces, and thus the same faces.
To make Gale diagrams useful for high-dimensional polytopes, we have
to get a hold on all Gale diagrams representing a combinatorial equivalence
class of polytopes. This is not simple, and it would lead into a discussion of
"partial oriented matroids" that we want to avoid (there is not much theory
for that, either). Instead, we restrict ourselves to an important special case:
when only one oriented matroid is possible for a given convex polytope.
Definition 6.25. The oriented matroid of a d-polytope P C Rd is rigid
if every polytope P' that is combinatorially equivalent to P has the same
oriented matroid.
Here our convention is to identify the vertex sets of P and of P' with [n]
in a way that is compatible with the combinatorial equivalence. With this,
combinatorial equivalence means that the oriented matroids M(P) and
M(P') have the same set of positive cocircuits, and rigidity means that
this implies that all cocircuits of P and of I' coincide:
C*(P) n
0-, or --,-- c* (PI) n {, Or
>
C(P) ----- C*(P).
For example, triangular prisms are rigid, but octahedra are not. The concept of rigidity is only interesting because there are some rigid polytopes
around, although "most" polytopes are not rigid. Here is one construction
to get rigid polytopes.
180
Theorem and Definition 6.26 (The "Lawrence construction").

Let V E Ekr" be a vector configuration in r r , possibly obtained from a
point configuration X c mdxn in d with r = d 1. (To avoid trouble, we
assume V has no coloops: even if we delete one of its vectors, the others
still span W'.)
If G E nx(nr) is a Gale diagram of V, then adding the opposite to
every vector of G we get the Gale diagram
2nx (rtr)
GG)
of a polytope with 2n vertices in R2n-(n- r) -1 = 118.n+d ; this polytope is
denoted by A(V) C Rn+d and called the Lawrence polytope of V.
Equivalently, we get the Lawrence polytope A(V) by successive Lawrence
extensions V o -z (V): for this we replace each vector '"in V by two
new vectors
Vi+ := Vi er+i
and
:= v i
to get an acyclic vector configuration in I r+n, from which we pass to affine
space
If we start with an affine point configuration X, we can perform the
Lawrence liftings directly on the point configuration, without linearization.
Our figures illustrate both the "linear picture" of a single Lawrence lifting
applied to a vector configuration
and the "affine picture," where a Lawrence lifting is performed on a single

point in a finite point configuration, thus increasing the dimension of the
configuration by 1:
ii+
0 i
181
Proof. It is easy to check that the Lawrence construction in fact yields a

D
convex polytope and that the two descriptions are equivalent.
As a very trivial example, consider three nonzero vectors in RI (r = 1,
n = 3). The Lawrence construction applied to them yields a triangular
prism in R3 : we show the affine picture for this.
Lawrence polytopes are rigid, that is, if P' is combinatorially equivalent to A(V), then the oriented matroid of P' is also isomorphic
to that of P. In particular, the Gale diagrams of P' and of A(V) are iso-
Theorem 6.27.
morphic.
Proof. In the Gale diagram G of the Lawrence polytope A(V), and in
every Gale diagram G' with the same positive circuits, the points come
in pairs of positive and negative points, since those pairs form positive
circuits.
Furthermore, if we take any other circuit, then it contains at most one
point from every such pair. Hence we get all the circuits from the positive
ones by replacing a positive point by the negative "other point" of its pair.
Thus all circuits of the diagram are determined by the positive ones, and
thus the configuration V is rigid.
0
The Lawrence construction has numerous applications. Perhaps the most
striking one is the "universality theorem" for polytopes, which we want to
describe now.
For this, one needs a suitable equivalence relation for semialgebraic sets;
we use Richter-Gebert's version from [424]. Two semialgebraic sets S and
T are stably equivalent if they can be related by a sequence of "rational
changes of coordinates" (such that f and f-1 are both rational functions
with Q-coefficients, and induce homeomorphisms of the sets that we consider) and "stable projections" (whose fibers are the relative interiors of
rational polyhedra) see Richter-Gebert [424, Sect. 2.5] for the precise
definitions and more details. Stable equivalence is a very "restrictive" concept. In fact, stably equivalent sets S and T
have the same homotopy type (in particular, S is connected if and only
if T is connected)
182
have the same algebraic complexity (in particular, S contains rational

points if and only if T has rational points),
have comparable singularity structure (in particular, S is a manifold if
and only if T is a manifold.
Universality Theorem for Polytopes 6.28.
(Mnv [378])
Every elementary semialgebraic set defined over Z is stably equivalent to
the realization space of some polytope.
Every open elementary semialgebraic set defined over Z is stably equivalent to the realization space of some simplicial polytope.
Essentially, this means that the realization space of a polytope can be
"arbitrarily complicated" - it can be disconnected with many components,
it can consist of circles and spheres (can have "homology" in all dimensions), and can have all kinds of complicated singularities in general it
is certainly not a manifold, as claimed in [429, p. 18].
The theorem on which all of this is based is Mnar's universality theorem: the realization space for planar point configurations (i.e., for oriented
matroids of rank 3) can be any semialgebraic set, up to stable equivalence.
For a long time, there was no detailed proof available for this theorem.
Mnv's paper [378] only provides a sketch of the basic ideas for the "local"
version of the theorem; two further sketches are in Shor [463, Sect. 4] and
in Goodman & Pollack [219, Sect. 7]; see also Bj6rner et al. [92, Sect. 8.6].
Finally, a complete, detailed proof was provided by Giinzel [243]. His proof
also covers the far-reaching extension announced in Mn6v [379], the "universal partition theorem" for oriented matroids.
On the other hand, it is much easier to see (using the "van Staudt constructions" for addition and multiplication of points, of classical projective
geometry [254, Sect. VI.7] [111, Sect. 2.1] [521, Sect. 7] that the smallest
subfield of Ill over which all planar point configurations can be realized is
the field of all algebraic numbers A C 11'. This means with Theorem 6.28
that A is also the smallest field over which all polytopes can be realized.
There is a great new development: Richter-Gebert's Universality Theorem for 4-Polytopes, and the (even stronger) Universal Partition Theorem
for 4-Polytopes, with all their corollaries and extensions.
Universality Theorem for 4-Polytopes 6.29. (Richter-Gebert [4241)
Every elementary semialgebraic set defined over Z is stably equivalent to
the realization space of some 4-dimensional polytope.
In the course of this work done after the first version of this book
appeared Richter-Gebert solved quite a number of basic open problems
(see Problems 5.11*, 6.10*, and 6.11*). There is neither time nor space
to explain Richter-Gebert's work here: See Richter-Gebert [424], and the
announcement in [426]. See also Giinzel [244].
Not es
183
Notes
For all information about oriented matroids, we rely on the monograph by
Bjiirner et al. [92]. Other expositions that include surveys on oriented matroids are Bachem [30], Bachem Sz Kern [31], Bokowski & Sturmfels [111],
and Bokowski [105].
As you may have noticed, we have deliberately tried to keep linear algebra
concepts low-key. You may reformulate all of the basic constructions in
more advanced language. For that, the vector configuration V E rXn is
r r , the space Dep(V) is the kernel
considered as a linear map V :
of this map, Val(V) is the image of the dual map, and so forth.
It seems that deletion and contraction of an "element" are fundamental
operations in many areas: for graphs (see Section 4.1), for vector configurations and oriented matroids (see Section 6.3(d)), and for arrangements
and zonotopes (see the next lecture). In fact, there is a tremendous power
in proofs "by deletion and contraction," which proceed by induction on the
number of elements, and by putting a structure together from the information given by deletion and contraction of the same element. Zaslavsky's
work on hyperplane arrangements [535] is the classic source for that approach.
Gale diagrams are a tool that emerged from work of Gale [205] and
were developed to their full power and beauty by Perles, as documented in
Griinbaum's book [234]. Additional sources are the book by McMullen &
Shephard [374, Ch. 3], McMullen's survey [366], and the treatment (with
nice illustrations and examples) in Ewald's book [189]. It seems that the
close connection between the Gale diagram technique and oriented matroid
duality was first mentioned in [366], and the explicit identification was
worked out by Sturmfels [494].
The reduction to affine Gale diagrams is implicit in Perles' work (see
Griinbaum [234, p. 59]) and also used by Bokowski [109]; it appears as a tool
of its own standing in Sturmfels' work [495]. We mention for completeness
that any two affine Gale diagrams of the same polytope are connected by
a projective transformation and the corresponding reorientation.
Many interesting properties of polytopes can be profitably studied from
the oriented matroid point of view, not only via Gale diagrams. Surveys of
applications are in Griinbaum's book [234], in Bokowski & Sturmfels [111],
and in Bayer & Lee [59, Sect. 4].
Some authors distinguish between "Gale diagrams" and "Gale transforms." We did not make this distinction, but essentially what we constructed here were Gale transforms, while any configuration that represents
the dual oriented matroid is a Gale diagram. We also just note that there
are several useful reformulations and variations of the Gale diagram construction, among them a "coordinate-free" formulation [190] [366], which
was useful in the investigation of infinite-dimensional polytopes by Kleinschmidt & Wood [310, 532].
184
For issues related to the isotopy conjecture, we refer to [92, Sect. 8.6].
The Lawrence construction is due to Jim Lawrence (surprise), but he
never published it. It appears in Billera Sz Munson [73, Sect. 2] and is also
explained in detail (in oriented matroid terms) in [92, Sect. 9 31
A "cr-construction" to produce rigid 6-polytopes from planar configurations was given by Sturmfels in [493]. However, we found the arguments
in [493] to be incorrect. (Specifically, the claim that the orientation of all
the "outer simplices" of a polytope is determined by the combinatorics of
the face lattice is only true for simplicial polytopes: for this, consider the
cone over a nonrigid polytope, where all simplices are outer; however, the
polytopes produced by the a-construction are not simplicial, and they turn
out not to be rigid in general.)

6.0 Prove Radon's theorem: given any set V of d + 2 points in Rd, we
can find disjoint nonempty subsets V1 , V2 C V such that relint(Vi ) n
relint(V2) 0. Why can we assume that conv(V,) are simplices?
6.1
Show that if two configurations of n points in Rd have the same set

of minimal affine dependences, then they are affinely isomorphic.
6.2 List all the circuits and cocircuits for the hexagon discussed in Section 6.1. How many vectors and covectors are there? (Don't list them
all, there are many.)
6.3 Show that the definitions of vectors, circuits, and so on, for the affine
( 11 )
and linear ces
as are consistent: if V -then
a-Dep(X) -= Dep(V)
and
'
a-Val(X) = Val(V),
and thus we get the same oriented matroid (the same circuits, cocircuits, etc.) for X and for V.
6.4 Let D = (V, A) be a directed graph with arc set A = {1, 2, ... , n}.
Define the signed circuits of D to be the sign vectors u E 10, +1 In
that correspond to circuits in D together with a chosen orientation,
0;
as follows: if the arc i is not contained in the circuit, then u
if it is in the circuit and directed according to the orientation, then
ui ,--- +; and if it is directed opposite to the orientation, then ui = .
-
185
For example, for the digraph drawn here and the oriented circuit
marked in it we read off the signed circuit u next to the drawing.
Show that the signed circuits we get that way are from a realizable
oriented matroid (as in Definition 6.5), whose cocircuits correspond
to the minimal directed cuts in the graph. Interpret the vectors and
the covectors of this oriented matroid in terms of the graph.
(Hint: Associate a vector configuration with D. A canonical choice is
= ei e3 for an arc from the node j to the node i.)
6.5 Prove that our two characterizations of acyclic vector configurations
are equivalent. Prove that the dual of an acyclic configuration is totally cyclic (Corollary 6.16).
Describe a small vector configuration that is neither acyclic nor totally cyclic.
6.6 A d-polytope with n vertices is simplicial if and only if every nonempty coface has at least n d elements. Derive from this a characterization of the (affine) Gale diagrams that represent simplicial
polytopes.
6.7 Given a Gale diagram, how can one (computationally) enumerate the
facets of the corresponding polytope?
6.8 Show that the following diagrams represent the four different combinatorial types of 4-polytopes with 6 vertices.
Describe the polytopes.

How many facets, and how many edges do they have? Which polytopes are simple or simplicial? Which is bipyr(A3 )? Which is C4 (6)?
6.9 Describe all the 4-polytopes with 7 vertices. For this, use all the
"visualization tools" that we have developed so far:
Schlegel diagrams
Gale diagrams
combinatorial descriptions (vertex-facet matrix)
and show how the various types of data correspond to each other.
188
6 Duality, Gale Diagrams, and Applications
6.17 Show that the figure
is an affine Gale diagram of a 4-polytope P with 8 vertices, and verify

the following facts. The polytope has 9 facets, four tetrahedra, four
square pyramids, and an octahedron 235678. Every Gale diagram G'
with the same set of positive circuits has 7 and 8 on the same point:
so for every combinatorially equivalent polytope to P the vertices
2356 of the octahedron facet 235678 are coplanar. Thus the shape of
the octahedron facet cannot be prescribed. However, show that the
oriented matroid of P is not rigid.
6.18 Show that the figure below is the a ffine Gale diagram of a 4-polytope
with 8 vertices and 11 facets, the Kleinschmidt polytope K4 (8), and
verify the following facts.
Except for the facet 235678, which is an octahedron, the facets of
K4 (8) are tetrahedra. The octahedron facet is not regular_ No Gale
diagram G' with the same set of positive circuits can have 136, 125,
and 178 on lines: so there is no combinatorially equivalent polytope
to K4(8) such that the octahedron facet is regular.
Show that the oriented matroid of K4 (8) is not rigid.

(Kleinschmidt [304} [1911, Sturmfels [495, Fig. 6(a)1)
6.19 (i)
Construct a simple polytope for which the shape of a facet cannot be prescribed. For example, you might examine the polar of
189
the simplicial 6-polytope with 10 vertices, given by
(ii)* Can you prescribe the shape of a facet for simple 4-polytopes?
6.20 Let a centrally symmetric polytope with 2n vertices in
as P conv(X) for
be given
X=
Show that the dependences and the value vectors of X can be reconstructed from those of the following set of only n + 1 points in Rd:
X0
= fu,u
v2, ,u +
Thus the combinatorics of P can be read off from the dual configuration Go C (Rn-d ) * to Vo, the central (Gale) diagram of P, due to
McMullen & Shephard [373].
Use central diagrams to classify the centrally symmetric polytopes

with at most 2d + 2 vertices.
Show that the metric properties of this diagram are important, and
that no further reduction to affine diagrams is possible. Similarly,
show that the signed circuits of X0 do not determine those of X, so
we cannot simply reduce to oriented matroid data [373].
6.21
Apply the Lawrence construction to three points on a line, either all

three distinct, or two coinciding. What polytopes do you obtain? List
all circuits and cocircuits both for the original configuration and for
the Lawrence lifting.
6.22 Consider the prisms over simplices, prism(Ad), and construct their
Gale diagrams. Show that they all arise as Lawrence polytopes.
6.23 Show that the oriented matroid given as an example for a 5-polytope
with a nonprescribable 2-face is not rigid. (Use the fact that one can
perturb the point 1 without changing any positive circuit.)
Is Perles' example of a nonrational 8-polytope rigid?
186
6.10* What is the smallest number of vertices for a nonrational 4-polytopes?

(Non-rational 4-polytopes exist by Richter-Gebert's universality theorem, see the Notes for this chapter. The smallest (explicit) example
by Richter-Gebert has 33 vertices and ?? facets [424, p. 80]. The
minimal number of vertices and facets is not known.)
6.11 Analyze the 4-dimensional polytope X* with 8 facets and 12 vertices
whose polar is given by the affine Gale diagram
Show that X* has a hexagon 2-face whose shape cannot be prescribed.

Verify that X* is combinatorially equivalent to the polytope whose
Schlegel diagram is given by Exercise 5.11.
6.12 Draw an arbitrary "nice" configuration of black and white points into
the plane, and analyze:
(i) Is this the Gale diagram of a polytope? (If not, add points to
get one.)
(ii) What is its dimension and its number of facets?
(iii) Is it simple or simplicial? Can you describe its facets?
(iv) Are there vertices that are not adjacent? Can you compute the
graph?
6.13 For n > d > 2, consider the moment curve in Rn -d-2 , and place on
it n points, alternating between negative and positive points.
Show that the oriented matroid of this is dual to the oriented matroid of the cyclic polytope Cd(rt): so it is the Gale diagram of a
polytope that is combinatorially equivalent to Cd(n). Is this really a
Gale transform of Cd(n)?
6.14 Show that the following conditions are equivalent for a 2d-dimensional
polytope P with n vertices:
(i) P is neighborly.
(ii) Every set of n - d points forms is a positive covector.
(Hi) Every circuit of P contains exactly d + 1 positive and

ative elements.
d+ 1
187
neg-
Derive the corresponding criteria to detect whether a Gale diagram

represents an (even-dimensional) neighborly polytope.
6.15 Show that of the following two figures, the left is a Gale diagram
for C4(8), while the second is the Gale diagram of a 2-neighborly
4-polytope with 8 vertices that is not cyclic (due to [494, p. 5431).
7:
___----.
1
_____-----c)
8
0-------_________46, .70________---
"Many" noncyclic neighborly polytopes can be constructed this way,

by "small" modification of the Gale diagram of a cyclic polytope.
(More were constructed by Shemer [458].)
6.16* Perles conjectured that every simplicial polytope is combinatorially
equivalent to a face figure (iterated vertex figure) of an even-dimensional neighborly polytope. (Equivalently, every simple polytope is a
face of a polar of a neighborly polytope.)
(i) Define that a (finite) configuration of vectors in r 3 is uniform if

any 3 of the vectors span R3 . Say that it is balanced if for every
plane spanned by two vectors, the number of vectors on the two
sides are equal.
Show that for simplicial d-polytopes with d+4 vertices, the Gale
diagram construction reduces Per les' problem to the "embedding
problem" of whether every uniform configuration of d+ 4 vectors
in R3 can be extended to a uniform balanced configuration.
(ii)
Using part (0, solve Per les' problem for simplicial d-polytopes
with n = d + 4 vertices: every simplicial d-polytope with d +
4 vertices is a quotient (i.e., combinatorially equivalent to an
iterated vertex figure) of a neighborly (2d+4)-polytope with 2d+
8 vertices.
(Hint: For part (i) you will need the characterization of Exercise 6.14.
The Gale diagram formulation of Perles' problem is due to Sturmfels. See Sturmfels [495, Sect. 7], where some partial results are also
derived. Part (ii) was done by Kortenkamp [313].)
190
6.24 A polytope is projectively unique if any combinatorially equivalent

polyt ope can be obtained by a projective transformation.
(i) Show that in the plane, triangles and quadrilaterals are projectively unique, but n-gons with n > 5 are not projectively unique.
(ii) Show that, more generally, d-polytopes with n < d 2 vertices
are projectively unique.
(iii) Show that if P is projectively unique, then so is P. Conclude
that d-polytopes with n < d + 2 facets are projectively unique.
(iv) Derive from part (iii) that 3-polytopes with h < 9 edges are
projectively unique. (Use Euler's formula y e + f =-- 2 from
Exercise 4.6 or Lecture 8.)
Prove the converse, too.
(v) Show that if P is projectively unique, and F is a face of P, then
the face need not be projectively unique. In this situation, the
face F cannot be arbitrarily prescribed.
6.25 How many (positive and negative) points do you need to create the
Gale diagram of a rigid polytope from Suvorov's configuration?
6.26* What is the smallest number of points in a planar point configuration
that violates the isotopy conjecture?
(At the moment the smallest known configurations have 14 points; see
Example 6.24. In contrast, isotopy is known for n < 9 points, but only
in the case of general position configurations, by Richter [420] [92,
Sect. 8.21.
6.27 Show that the polytope P of Example 6.23 can be constructed, via
three Lawrence extensions, from the configuration of Pascal's Theorem ("the vertices of a hexagon lie on an ellipse if and only if the three
intersection points of opposite sides are colinear"). This is indicated
in the following picture.
Deduce from this construction method that this is a 5-polytope for

which the shape of a 2-face cannot be prescribed.
(See Richter-Gebert [424, Example 3.4.3] for a detailed discussion.)
7
Fans, Arrangements, Zonotopes,
and Tilings
Zonotopes are the images of n-cubes under affine projection maps. Since
for most aspects of polytope theory n-cubes are not very complicated, this
definition may hide the complexity and richness of this concept. Zonotopes
are interesting from various points of view. Their combinatorial structure
is closely linked to (and in a precise sense equivalent to) that of real linear
hyperplane arrangements.
The aim of this lecture is to provide basic geometric intuition and the
tools for a combinatorial description of zonotopes. We will see how zonotopes again are modeled by oriented matroids, and discuss the surprising
appearance of general ("nonrealizable") oriented matroids in the study of
zonotopal tilings, and of hyperplane arrangements.
7.1
Fans
Definition 7.1.
A fan in Rd is a family
.T. = {Ci , c2 , 7 CN}
of nonempty polyhedral cones, with the following two properties:

(0 Every nonempty face of a cone in F is also a cone in F.
(ii) The intersection of any two cones in .f. is a face of both.
The fan ..T. is complete if the union U.F := C1 u ... U CN of F is Rd, that
is, if UiN i Ci = Rd . We will consider only complete fans here, and thus we
will omit the word "complete" most of the time.
192
7. Fans, Arrangements, Zonotopes, and Tilings
.7- is pointed if {0} is a cone in .F (and thus is a face of every cone in .F).
It is sirnplicial if all its cones are simplicial cones, that is, cones spanned by
linearly independent vectors. Simplicial cones and fans are automatically
pointed.
The following figure shows three complete fans in R2 , with N = 13,
N = 11, and N = 3 cones, of which 6, 5, and 2, respectively, are fulldimensional. The first two fans are pointed (for d = 2 this implies they are
simplicial); the third one is not.
There are various equivalent or similar ways to define fans (see also the
notes to this lecture). What we have given here essentially just describes
a polyhedral complex of cones (as in Definition 5.1) whose union is d .
In particular, the definition implies that the relative interiors of the cones
in Y form a partition of space:
N
Ed- relint(Ci ) = Rd
Now comes the first reason why we look at fans in the theory of polytope,s.
Example 7.2. Let P be a polytope in Rd with 0 E relint(P). We define
the face fan of P as the set of all the cones spanned by proper faces of P:
.F(P) :=--- {cone(F) : F E L(P)\P}.

.F(P) is a pointed fan in lin(P): its union is the linear hull lin(P). It is a
complete fan in d if P is a d-polytope, with 0 E int(P).
7.1 Fans
193
Our figure indicates the construction of the face fan, for a 2-polytope with
a given origin in it. Note that the geometry of the face fan does depend on
the position of the origin in P.
Example 7.3. Let P be a nonempty polytope in Rd. For the normal fan
of P we take the cones of those linear functions which are maximal on a
fixed face of P. That is, for every nonempty face F of P we define
NF :=
(Rd) * : F C fx E P : cx = rnax cy : y E
and with this we define

{NF: F E
411\0}.
N(P) is a complete fan in ( d . If P is d-dimensional, then the fan is

pointed, since then {O}
Np is in the fan.
The above figure illustrates the construction of the normal fan of a 2polytope: for this we have identified 2 with (R2 )* via the usual scalar
product, which accounts for the right angles in the figure.
The face fan and the normal fans are very natural objects associated
with a polytope. In particular, they come up in the theory of optimization.
For this note that the question "Which cone of Ar(P) does c lie in?" is
the linear programming problem max cx : E P. Similarly, the question
"Which cone of .F(P) does y lie in?" is the separation problem of finding
one single valid inequality that determines whether av E P, for various
a > O.
Example 7.4. Let A
{HI , . . . Hp } be a finite set of linear hyperplanes
in Rd, where each Hi is of the form Hi =
e Rd : cix = 01 for some
ci ( d)*.
Clearly, the arrangement A decomposes I d into a complete fan TA. The
cones of the fan are also referred to as the faces of the (linear) hyperplane
194
arrangement A. The combinatorics of the fan encodes a lot about the configuration of row vectors C = {c1 , - . . , cp} it is quite easy to see that the
cones in .7' are in natural correspondence with the covectors of C. We'll see
details about this in Theorem 7.16 and Corollary 7.18.
In the figure after Definition 7.1, the first and the third fan are given by
hyperplane arrangements, the second one is not.
Example 7.5. There are also complete simplicial fans .F0 in R3 that are
nonpolytopal, that is, not of the form ,T(P) for any polytope P.
For one possible construction, we start from a simplicial 2-diagram D
that is not a Schlegel diagram, for example the one we constructed before
Theorem 5.7. This we place into an affine plane, and take all the cones
over faces of the diagram. We complete .F0 by adding one extra ray and
the simplicial cones that are spanned by this ray together with the cones
generated by the boundary of the diagram D. Our figure illustrates the
construction:
This fan .7 0 is not polytopal. In fact, assume that P is a polytope with

Fa = F(P). Now consider the polytope P` given by the convex hull of
all vertices of P except for the one on the "extra ray." Then the origin is
beyond a triangular facet of P', and the corresponding Schlegel diagram
would be (projectively) equivalent to the one we started with. This yields
a contradiction.
This example shows by far not the worst that can happen: for example,
there exists a (nonsimplicial) fan that is not even the face fan of a starshaped
sphere (with flat facets); see Eikelberg [183 ] . We also refer to Ewald [189,
Sect. 111.5] and the examples and references he gives.
We continue with three more "trivial" operations on fans, which will turn
out to be very valuable soon.
Definition 7.6.
sum is the fan
If Y is a fan in
Y0 g
and
g is a fan in Rq, then their direct
:= {C xC': CE .r, C' E g}-
195
If .F and g are both fans in the same space Rd, then we define their
common refinement as
IC n c' : C E T, C`
0.
If .F is a fan in Rd and V C Rd is a vector subspace, then the restriction

of .F to V is the fan
.FIV :---
IC nV:
CE .F}.
It is easy to check that all three constructions are again fans in the
sense of Definition 7.1. If you want, you can consider gv := {V} as a
(noncomplete) fan in Rd: then the restriction to V is the intersection with
gv, that is, .FIV =
F A gv-
Lemma 7.7. Let P C il,I P and Q C Rq be two polytopes. Then the normal
fan of the product PxQC P+41 is the direct sum
Af(PxQ) = A/(P) E B Ar(Q)-
Example 7.8. The normal fan of the cube Cd = { x E d : 1 < xi < 1}

coincides with the face fan of its polar, the d-crosspolytope (cf. Exercise 7.1
for the general case of this).
We find that the normal fan is given by the arrangement of coordinate
hyperplanes in (Rd)*. This is quite trivial (since it is trivial to optimize
over a cube), and it also confirms Lemma 7.7: the fan of the arrangement
of coordinate hyperplanes is the direct sum of 1-dimensional fans.
ED
Also, the cones in the fan are characterized by the sign function: each cone
in the fan can be identified with a vector in {, , O}d, and the orthants
correspond to the sign vectors in {, } d
7.2
Projections and Minkowski Sums
Let P C RP be a p-polytope, and let 7r :

7r(x) = Ax
Rd be an affine map,
z,
196
dxP and z E Rd .
with A E (Rd)P
If 7r is injective (that is, A has rank p), then we refer to it as an affine
transformation, and 7r(P) is affinely isomorphic to P.
If 7r is not required to be injective, then we refer to it as an affine projection or an affine map. In this case Q := 7r(P) is again a polytope, whose
dimension is dim(Q) = dim(7r(RP)) = rank(A), where we had assumed
dim(P) = p.
We usually assume that Ir is surjective we may do this, after restricting
the image of 7 to 7r( r'P) C R d
and soir maps the p-polytope P C RP to
a d-polytope Q c Rd. Also, if we are only interested in properties that are
invariant under translation, then we may assume that z = 0 and that the
map 7r is actually linear.
Definition 7.9.
A projection of polytopes
z, such that p C I
is an affine map 7r : RP ---> Rd, x i--- Ax
p-polytope, Q C Rd is a d-polytope, and 7r(P) = Q.
-
is a
Here is a very simple, but basic, fact about projections.
Lemma 7.10. Let ir : P f Q be a projection of polytopes. Then for

every face of Q, F E L(Q), the preimage r -1 (F) =fy E P: r(y) E Q} is
a face of P.
Furthermore, if F,G are faces of Q, then F C G holds if and only if
ir- 1 ( F) g
Proof. If c E (Rd )* defines F, then c o 7r defines 7r -1 (F) Instead of the
corresponding trivial calculation, we give a picture.
a
The linear algebra behind this construction is quite simple. From the
surjective map 7r : RP --4 Rd we get a dual map 7r* : (R()* --- (RP)*,
mapping c 1---+ c o 7r, which is injective. This distinguishes a certain set of
functions on P (and thus on Rd ): those that are constant on the fibers
of 7r. The embedding 7r* is used in the following lemma.
197
The normal fan Ar(Q) of a projected polytope Q is isomorphic, via 7r*, to the restriction of Ar(p) to the image of 7r*, the linear
subspace 7r* (Rd )*:
Lemma 7.11.
Ar(Q)
-`--. Ar(P)I7r*(Rd )
We had considered very special projection maps in Lecture 1: projections

along a coordinate axis (with d = p 1), which correspond to FourierMotzkin elimination. The general case is certainly interesting. In fact, if we
try to view polytopes as a category in natural terms, then probably isomorphism should be affine isomorphism, and surjective maps should be affine
projection maps. The geometry of projections is not really understood: we
will talk about this later.
In terms of polarity, which links the two versions of Fourier-Motzkin
elimination given in Lecture 1, note that the polar operation to projection
is intersection, taking Q --= P n V, where 0 E relint(P) and V C Rd is a
vector subspace. In this case we see that .F(Q) = T(P)i V.
A simple application of projection is the construction of Minkowski sums,
which we already met in Section 1.1. Here we work out only the case of
two summands, as the extension to more summands is then obvious.
The Minkowski sum (or vector sum) of two polytopes P and P' in
1
P + Pi :, {z + x' : x E P, x' E P } .
P+7
198
We use the projection map 7r : R2d , 1' d given by ir(2f, ) := X + X i , whose

dual map ir* : ( d)* , (R2d)* is the diagonal map r*(c) = (c, c), by
CI
7r
(;)
c(x
+ x`) = (c, c) (;) .
Now we can write the Minkowski sum as a projection of the product:
P + 13' := ir(P x P`).

Thus, putting together very simple observations, we get the normal fan
of a Minkowski sum. (Careful: it is not the direct sum of the fans of the
factors!)
Proposition 7.12. The normal fan of a Minkowski sum is the common
refinement of the individual fans:
Ar(P + 13') = Ar(P) A Af(P').

Proof.
g*
1V(13 + Pi ) =-- _N-(7r(P x PT ,
=
.Al(P x P`)17r*(R d)* =

)* r.J- Ar(P) A N(P')
using Lemma 7.7 for the direct sums, Lemma 7.11 for the projection, and
the dual map 7r*(c) .--- (e, c).
0
7.3 Zonotopes
Zonotopes are special polytopes that can be viewed in various ways: for
example, as projections of cubes, as Minkowski sums of line segments, and
as sets of bounded linear combinations of vector configurations. Each of
these descriptions gives different insight into the combinatorics of zonotopes. The following includes several such descriptions, all of which lead us
to the same "associated" system of sign vectors that describes the combinatorics of a zonotope. The main goal will be to see in what sense zonotopes
and arrangements can be considered equivalent, and how the combinatorial
structure of a zonotope is given by an oriented matroid.
After the general discussion of projections in the last section we now
consider a very special (but interesting) case: projections of cubes, that is,
ir : P --+ Q, x i--> Vx + z is an arbitrary (surjective) affine map, but P
is the d-cube,
P = Cp = fx E RP : --1 < xi < 1 for all il.
7.3 Zonotopes
199
Definition 7.13. A zonotope is the image of a cube under an affine projection, that is, a d-polytope Z C d of the form
Z -= Z(V)
:=
VCp + z = {Vy + z : y C Cp }
P
Ix E Rd : x = z + Exivi , 1 < xi < 1}
i=1
dxp .
for some matrix (vector configuration) V = (v 1 , ... , vi,) E
Z
----------
Equivalently, since every d-cube Cd is a product of line segments Cd =

C1 x... X C1 , we get that every zonotope is the Minkowski sum of a set of
line segments. In fact, if R- is linear we get
Z(V)
= 7r(C1 x ... x C1)

=
=
and thus Z(V) = Hui, vi] + ... + {vp ,vpl + z for an affine map given by
r(y) = V y + z.
In the following we will usually assume that Z -- Z is centrally symmetric with respect to the origin 0, corresponding to a linear map 7F: C --* Z.
Example 7.14. By definition, the cubes Cd are zonotopes, where the projection map can be taken to be the identity.
Also, every centrally symmetric, 2-dimensional 2p-gon P2(2p) arises as
the projection of a p-cube to the plane. In fact, if the vertices of P2(2p) are
x i,
Xp,Xp+17 7 X2p
in cyclic order, with xp+i = xi, then we get

P2(2p)
x2 xi x2
I
2
'
2
r
I .-.[
2 p+1 Xp+1 Xp Xp+1
'
XP 1
1-
200
One way to prove this is by induction on p, by taking any pair of opposite (parallel, of same length) edges, and showing that it corresponds to a
Minkowski summand of P2 (2p).
7
8
We invite the reader to provide his or her own proof.
Example 7.15. The permutahedron Iln_ i (Example 0.10) is a zonotope

of dimension d -- = - n - 1, arising from an affine projection of the cube of
dimension p --=-- ( 2 ):
n+1
11n- 1 =
1 + [-
e2 - e i e2 - e
2
'
e3 - e i e3 - e l
l1 + [
en
2
'
en_ 1 en
2
'
1+-
2
en 1 ,
l
2
Perhaps the easiest way to see that this Minkowski sum yields the right
polyhedron is first to observe that it is invariant under permutation of
coordinates, and then compute the points of the sum that maximize a
linear function C E (Rn)* with e 1 < e2 < . . . < Co : this is easily seen to be
the vertex
v
--=--
n+1
2
e2 - el
'+
n +1
=
2 1
n-1
el
2
en - ea,_ 1
2
e3 - e l
2
n-3
2
e2
--.
n + 1 - 2n
en
2
1)
=(2
;.1
There are a few more "obvious" properties of zonotopes: for example,

all zonotopes are centrally symmetric. Also, since every face of a cube
is a (translated) cube, we get that every face of a zonotope is again a
zonotope, and thus centrally symmetric with respect to its barycenter. This
property characterizes zonotopes. In fact, any polytope all whose 2-faces are
centrally symmetric is a zonotope; see Bolker [113] and the references given
there, Schneider [439], or [92, Prop. 2.2.14]. Even stronger: any polytope
whose k-faces are all centrally symmetric, for some k with 2 < k < d - 2,
7.3 Zonotopes
201
is a zonotope, by McMullen [360]. The same fails if only all the facets
(k = d -1) are centrally symmetric: in this case we have a counterexample in
the regular 24-cell (a sporadic 4-dimensional regular polytope, described for
example in Coxeter [156, Sect. 8.2]), whose facets are all regular octahedra;
counterexamples for all d > 4 can be found in McMullen's paper [365].
In particular, we see that being a zonotope is a geometric property, not
a combinatorial one. For example, the quadrilateral CQi is not a zonotope,
but the quadrilateral Q2 is.
Thus being a zonotope is preserved under affine equivalence (in fact, under
affine projections), but not in general under combinatorial equivalence.
It may require a second of thought to figure out that in general zonotopes
are not simple polytopes (though the permutahedra are). Our next picture
shows a zonotope generated by four line segments in R3 , no three of them
coplanar. The resulting zonotope (d = 3, p = 4) has 8 simple vertices of
degree 3, and 6 nonsimple vertices of degree 4. The figure indicates (by a
dotted line) that the vertex figure of the top vertex is a square.
For the combinatorial structure of zonotopes, we have Lemma 7.10: the

faces of Z can be uniquely associated with the faces of the cube it is projected from.
Now every nonempty face F of the p-cube can be associated with a
sign vector. Here the natural construction associates with F a row vector
202
7. Fans, Arrangements, Zonotopes, and 'Filings
cr(F) E ({,,0}P) * , for example by

u(F) = sign(int(F )).
This is well defined, since the polar of a nonempty face of the cube is a
proper face of the crosspolytope, which has constant signs on the interior.
There are other, equivalent, ways to relate a face F with the sign vector
o- = a(F) E ({,,0} 1)) * E.--_- (1+1, 1,01P) * , for example by
F =
Aiei :
Ai = +1 for ui = +,
Ai = 1 for ori
Cp :
1 < Ai < +1 for a, = 0}

xi = a(F) ifor all i with o- (F)i 01.
Recall from Section 6.3(a) the componentwise partial order on sign vectors,
which is induced by
We find that the smaller a face F of Cp , the larger its sign vector u(F)
will be in this partial order "<." Also, in order to obtain the whole face
lattice we have to add an extra minimal element, since the empty face does
not contain points, and it does not have an associated sign vector, either.
Thus, we get
(L(Cp), g)
{6} u
The following sketch shows the signs associated with the faces of C2, and
the face lattice L(C2) with the signs corresponding to its elements.
(00)
(0+)
(-
(00)
(0-)
7.3 Zonotopes
203
With this we get sign vectors not only for the faces of cubes, but also for
the faces of zonotopes: if ir : Cp + Z is the projection that defines the
zonotope Z, then for every nonempty face G E L(Z) of the zonotope we
get the nonempty face 71 --1 (G) E L(Cd), and thus we put
o(G) := cy(r-1 (G)) = sign(x) E (I+, ,0}P)*,

where z is an arbitrary point of ii- '(G) for example, the center.
From this we get a sign for every face of the zonotope, and we have
u(G) < G ( G') if and only if G D G': thus the face lattice of the zonotope
is entirely determined by the system of sign vectors, and antiisomorphic to
it as a poset:
(L(Z), g)
i--,-t
({0-(c): G E L(Z)\0}, ? ),
where is defined to be smaller than any sign vector u(G). Caution: the
partial order on the face lattices is opposite to the order on sign vectors.
The larger a face of Z, the smaller a sign vector we associate with it, in
the partial order on sign vectors induced by 0 < and 0 < -.
This assignment of sign vectors to the faces of a zonotope may look a
little mysterious. In particular, the description that we have given (from
the projection of a cube) does not tell us much about the structure of
the collections of sign vectors that we get, and how they are related to the
matrix V that defines the projection. Thus, we take a "fresh start" here, and
obtain the same sign vectors from a different approach, via optimization.
The following theorem is so basic that we give two proofs. The first proof
shows how a sign vector erc, E {,,0}P is associated with every cone in
the normal fan Ar(Z) of the zonotope. Similarly, the second proof obtains
a sign vector u(G) E {, -, 0}P for every face G E L(Z) the same sign
vector that we have just "pulled down" from the p-cube, of course. Thus
after this we have three different constructions of the sign vector system
associated with a zonotope.
Theorem 7.16. Let Z = Z(V) C I' d be a zonotope. Then the normal

fan of Af(Z) of Z is the fan TA of the hyperplane arrangement
A = A v := {H1 , . . . , Hp }
in Rd given by
:=
lc E (Rd)* :
cv, = 01.
First Proof. For each of these hyperplanes we define the positive halfspace
by
Ilii-
1c E (Rd
) *
: cvi _> 01,
204
7. Fans, Arrangements, Zonotopes, and Things
and the negative one similarly. Thus the normal fan of the single line segment [v i , /Ji], is the set {Hi, Ht, a hyperplane and the two halfspaces determined by it.
HT
Now we use Proposition 7.12 to see that the normal fan of the Minkowski
sum of the line segments [-vi, Ili] is the hyperplane arrangement Av , that
is, the common refinement of the fans of the individual hyperplanes.
The position of c with respect to the fan {Hi, Ht, HT} is determined by
the sign of cvi . In fact, if sign(evi ) = 0, then c lies in Hi ; if this sign is +,
then e is in the interior of Hz ; and if it is , then c lies in the interior
of HT.
Thus in the common refinement A v , the position of c is given by the
sign vector
sign(eV) E I+, MP,
whose first coordinate records the relative position with respect to HI , the
second coordinate refers to H2, and so on. In particular, we get distinct sign
vectors for the distinct cones in Ar(Z), and inclusion of cones corresponds
to the usual partial order on sign vectors.
Our sketch shows a zonotope, its normal cones (drawn into the same
figure with right angles), the normal fan assembled from them (which is a
hyperplane arrangement), and the signs that we associate with each of the
cones in the normal fan.
(-++)
(- 0+)
(0--
(0++)
(+0-)
(+--)
7.3 Zonotopes
205
Second Proof. Assume that we want to maximize a function x

over the set of all bounded linear combinations
cx
We can maximize this sum by maximizing each of the summands separately,

and thus achieving the maximum on the following face of Z:
Zc =
{yEZ:Cy ISEWICZ}
Ein=1
A i = 1
1 < Ai < +1
if cvi <0,
Ai = +1
if cv i > 0 1.
if cvi = 0,
Thus the decision "which face of Z maximizes c" is equivalent to the decision, for each i, of whether c lies on the hyperplane Hi itself, on its negative
side, or on its positive side, that is, by the position of c in the fan of the
arrangement A.
The family of hyperplanes A thus gives a combinatorial interpretation for
the covectors of the configuration V = {v 1 ,. vn }. Here the interesting
case is the one where the configuration V spans Rd , such that the zonotope
Z(V) has dimension d, and the hyperplane arrangement Av is essential: the
intersection of all the hyperplanes is the origin, 1/1 n H2 n. n H = 101,
and thus the cones in Av are all pointed cones.
Corollary 7.17. Let V E Rdx P be a vector configuration in Rd . Then
there is a natural bijection between the following three families:
(the sign vectors of) the nonempty faces of the zonotope Z (V) C
(the sign vectors of) the faces of the hyperplane arrangement Ay,
the signed covectors of the configuration V.
Thus we have identifications

L(Z(V) A )\{.}
L(Av)
V * (V) g (I+,
0 113 ) *
Similarly (assuming that V has full rank), there is a natural bijection between the following three families:
(the sign vectors of) the facets of the zonotope Z(V)
(the sign vectors of) the one-dimensional rays of the hyperplane arrangement A v , and
the signed cocircuits of the configuration V.
206
In notation:
facets(Z(V))
vert(Z(V)) -->
rays (Av)
C* (V).
The ith zone of Z(V) is the collection of all faces that have [vi , vi] as
a Minkowski summand. The zones geometrically form "belts" around the
surface of a zonotope, and in fact completely cover it.
These zones may also be held responsible for the name "zonotope." Under the bijection between faces of a zonotope and cones in its hyperplane
arrangement, the ith zone corresponds to the ith hyperplane Hi in the
arrangement.
Note that vi and Ili determine the same zone, and the same hyperplane,
exactly if they are parallel vectors. This is a degenerate case that one usually excludes from the discussion. In fact, there is an even more degenerate
case, if vi = 0 for some i. In this case v i does not contribute to the geometry, we can just ignore it when constructing the zonotope, H = (Rd)* is
the full dual space, and i is a one element circuit (known as a loop) in the
oriented matroid, which can safely be deleted.
The number of zones is the principal complexity measure for zonotopes:
it coincides with the number of different hyperplanes in the associated
arrangement, and with the number of equivalence classes of elements in
the oriented matroid. If we assume that V is simple, that is, there are no
zero or parallel vectors in V E Rdx P, then the number of zones is p. The
key observation is that this parameter can be read off directly from the
zonotope Z, and does not depend on the choice of V. However, for every
zonotope Z there is a simple vector configuration that defines it, and the
vector configuration is unique up to permutations and sign changes.
Under the translation from zonotopes to arrangements and back, simple zonotopes correspond to simplicial hyperplane arrangements a very
7.3 Zonotopes
207
classical topic of geometric study because of its relation to the theory of

reflection groups and Lie algebras. See the notes at the end of the lecture.
There is also a metrical correspondence, which we get by using polarity.
In fact, from the simple observation that the normal fan of a polytope is the
face fan of its polar (Exercise 7.1), we get a polytope for every hyperplane
arrangement that "spans" the arrangement as a fan.
Corollary 7.18. Let Av be a hyperplane arrangement in ( d)* . Then
the face fan of the polar of the associated zonotope is given by Av:
.FAv
In particular, if V spans Rd, then the arrangement is essential, the zonot ope
is full-dimensional, and its polar is a palytope.
Thus the fan of an essential hyperplane arrangement is always polytopal.
The following constructs an explicit example for this. Note that from our
set-up the proof is simple but the geometric fact that we can construct
a polytope that "fits" any given hyperplane arrangement is not obvious at
first sight.
Example 7.19. For the matrix
= (
-1 1
1
we construct
Corollary 7.18 shows that the combinatorics of zonotopes is equivalent to

the combinatorics of hyperplane arrangements in a very strong sense. Let us
close this section with one nontrivial application of this equivalence. For this
we take a classical theorem from the theory of hyperplane arrangements,
Shannon's theorem [4571:
208
Every essential arrangement A of n hyperplanes in Rd has at least 2n

simplicial regions.
More precisely, adjacent to every hyperplane there are at least 2d
simplicial regions, and nonadjacent to any given hyperplane there
are at least 2(n d) simplicial regions.
(Note that the simplicial regions come in pairs, since the opposite of a
simplicial cone is simplicial as well). From Shannon's theorem, we derive
the following theorem about zonotopes.
Theorem 7.20 (Shannon's theorem for zonotopes).
Every d-zonotope with n zones has at least 2n simple vertices.
More precisely, on every zone there are at least 2d simple vertices, and
disjoint from any given zone there are at /east 2(n d) simple vertices.
The proof for Shannon's theorem for arrangements is not difficult: we
refer to Roudneff & Sturmfels [432], where several different proofs are presented, and to [92, Thm. 2.1.5]. The interesting thing is that there is no
"entirely combinatorial" proof. In fact, the corresponding statement for
oriented matroids is false. As we will see below, this translates into very
interesting effects for "zonotopal tilings."

We have by now seen so many oriented matroids around that you shouldn't
be scared any more if you hit their axiomatic definition a few lines down.
In fact, the axioms below just describe abstractly the "most important"
properties shared by the sign vector systems of equivalently
hyperplane arrangements
zonotopes
vector configurations
affine point configurations (vertices of polytopes!).

We have seen that in each of these cases we get a sign vector system of the
form V* = SIGN(U), for a linear subspace U C Rn .
For the following, choose V* to arise from any one of the above models:
117,} to be an oriented, essential
perhaps it is best to take A = {H1
arrangement of hyperplanes, and let V* be the set of sign vectors of all its
cones. (For example, take the one sketched in the proof of Theorem 7.16.)
We need the following operations on sign vectors. The zero vector 0 and
the negative u of a sign vector u have their obvious meanings. The support
,
..
209
: u, 74 Of. The composition of two vectors

of a sign vector u is supp(U) :=
u, y is defined componentwise by
(u o v), :=
{ i
if u,
0,
y-
The separation set for u, V is defined by

S(u, y)
: u = --vi 74 0}.
Finally, if j E S(u,y), we say that w eliminates j between IT and Ai if
w3 = 0 and wi = (u o Y)i for all i S(u, v).

This may look like a lot of definitions, but they all have very concrete
meanings for (sets) of sign vectors, and a concrete geometric interpretation,
as follows.
Definition 7.21 (Oriented matroids).

A collection V* C {-,,O}n is the set of covectors of an oriented matroid
if it satisfies the following covector axioms:
(VO) 0 E V*
("The zero covector is always a covector")
> -U E V *
(V1) u E V*
("The negative of a covector is always a covector")
(V2) u,v E V*
> uovE V*
("The set of covectors is closed under composition")
(V3) u, V E V * , j E S(U,y) == NtNi E V*: w eliminates j between u and v
("The set of covectors admits elimination")
The rank of the oriented matroid V* is the largest number r such that V*
contains a chain of covectors of length r:
0 < X 1 < X2 < - - - < X r

In this case, we write r(V*)
with X i E V * .
r for the rank of V.
Proposition 7.22. Consider a linear subspace U C

Then the set of column sign vectors
SIGN(U) = {sign(u) : u E U} C
Ip
of dimension r.
{, --, Or
is the covector set of an oriented matroid of rank r.

(This proves that the realizable oriented matroids M(V), as given by
Definition 6.5, are indeed oriented matroids of rank r in the sense of Definition 7.21, if we use U := Val(V) to derive the covector system V*.)
210
Proof. This is easy, but important, because it provides the "geometric

meaning" of the axioms and of the operations on sign vectors.
The hyperplanes {xi = 0} induce an essential hyperplane arrangement
in U, via Hi := fx e U : xi = 01. For these hyperplanes within U,
positive sides are uniquely determined via I-1;1- = Ix e U: xi > 01. With
this construction, we see that (V*, <) is the face poset of the essential,
oriented hyperplane arrangement n = {H 1 , , Hn } in U. Thus this poset
has length r.
For axiom (VO), we note that 0 E U has the sign 0 = sign(0) E SIGN(U).
For (V1), take u sign(u) for some u E U. From u E U u E U we
get u = sign(u) = sign(u) e SIGN(U).
The idea for the composition operation in (V2) is that if u, y E U, then
for any E E R we have u + ev E U.
Now, for any u,

, if e> 0 is small enough, then we get sign(u) o
sign(v) = sign(u ev)
this is easy to see by looking at the sign vectors
componentwise. From this we obtain sign(u) o sign(v) E SIGN(U).
For elimination as in (V3), we use that with u, v E U, general linear
combinations of u and y are contained in U, in particular the combination
w := uiv vu. Now if ui > 0 and vi < O, then this is a positive linear
combination of the two vectors u and v.
40.
40.
Considering the individual components of the sign vectors, we see that the
5th coordinate of w is zero. Similarly, if the ith coordinates of u and y
don't have opposite sign, then the ith coordinate of any positive linear
combination has the sign sign(wi ) = (u o v) i , as required.
211
The axioms of Definition 7.21 provide an entirely combinatorial model

for the geometry of hyperplane arrangements, vector configurations, point
configurations, or zonotopes. There are two (closely related) questions that
we will not avoid here:
how closely do the combinatorics of ori How good is this model
ented matroids represent a situation of an actual geometric object in
real space?
What is this model good for?
In the following, we try to answer both questions.
Remark 7.23:
"How good is the oriented matroid model?"
The model is excellent. All the basic structural properties that we

have proved in the realizable case in Section 6.3 extend to the case
of general oriented matroids. In particular, we have
duality (as in Definition 6.10)
equivalence of various types of data (as in Corollary 6.9)
deletion and contraction as basic operations (as in Proposition 6.11).
The topological representation theorem of Lawrence [194] [92, Ch. 5]
shows that every oriented matroid is "nearly" realizable: it may not
correspond to a real hyperplane arrangement, but it does correspond
to an arrangement of pseudohyperplanes, which need not be straight
but may be topologically deformed in some mild way.
The rest of this section will sketch this in the case of r = 3, which
corresponds to "arrangements of pseudolines" in the plane, as investigated by Griinbaum [237]; see also [92, Ch. 6].
For certain ranges of parameters, in particular for r < 2, for r > n-2,
and for n < 7, every oriented matroid is realizable, so the model is
perfect.
Even the nonrealizable oriented matroids come up "in practice." We
will present one example for this in the next section, where nonrealizable oriented matroids appear in the study of zonotopal tilings.
Remark 7.24:
"What are oriented matroids good for?"
Oriented matroids explain constructions like Gale diagrams and the

Lawrence construction (Lecture 6), which are clearly useful and have
various applications.
212
Oriented matroids provide a unifying framework, consistent terminology, and widely applicable tools for several areas of geometry.
Within the last twenty years, an extensive theory for oriented matroids was developed, with many nontrivial results. Results about
oriented matroids are easily transported from one field of applications
(e.g., hyperplane arrangements) to another one (like polytopes). For
the existing body of theory, we refer to the book by Bjiirner et al. [92],
and the handbook chapter by Bokowski [105].
The theory of oriented matroids allows us to handle, in a precise
sense, the combinatorics of objects that are geometric (like certain
simplicial spheres) but that cannot be represented in real "Euclidean
space" (or at least not as polytopes).
Thus, oriented matroids appear as a natural intermediate step in the

classification of (simplicial) spheres into polytopes and nonpolytopes:
this approach was pioneered by Bokowski; see [20] [15] [110] and the
monograph by Bokowski Sz Sturmfels [111].
We will now start to discuss one topic where nonrealizable oriented matroids come up the study of pseudoline arrangements. To make the
connection, consider an arrangement A of n hyperplanes in I d (all hyperplanes through the origin!). To draw and represent this, we consider the
intersection with an affine hyperplane: a hyperplane not through the origin,
but parallel to one of the hyperplanes in the arrangement. Thus we get an
affine hyperplane arrangement Aaff,, consisting of n 1 affine hyperplanes.
(Thus, an affine hyperplane need not contain the origin 0, but if we talk
only about a hyperplane, then we mean a linear hyperplane, which contains
the origin.)
The entire geometric combinatorial structure of the hyperplane arrangement A can easily be reconstructed from its "affine picture" Aaff,, up to
linear isomorphism, so nothing is lost by this reduction to affine space. In
particular, one can read off the oriented matroid of A from the arrangement
./laff, assuming that the hyperplanes in Aaff are labeled and that positive
sides are determined.
So in our drawing, for the case d ----- 3, we obtain an affine arrangement Aaff
of n 1 = 4 lines in the affine (dotted) plane, from an original arrangement
213
of 5 hyperplanes, namely the horizontal one, and the 4 planes determined

as the affine hull of the 4 lines together with the origin.
This is how in the case d = 3 affine arrangements of lines in 2
represent 3-dimensional (hyper)plane arrangements. Every arrangement of
lines in the plane determines an oriented matroid of rank 3.
However, it turns out that if we have an arrangement of "nonstraight"
lines in the plane, then we can also, in the same way, read off an oriented
matroid of rank 3. There are various quite general ways to define such
"nonstraight" lines Basically, any type of two-way unbounded topological
curves will do. See Griinbaum {237] or {92, Ch. 6] for general versions. Here,
for simplicity, we will use a simpler version, with the same combinatorial
results.
i
Definition 7.25. A pseudoline is a polygonal curve without self-intersections, with finitely many break points in R2 , and whose ends "head off to
infinity" in opposite directions.
An arrangement of pseudolines is a finite set of pseudolines in the plane
such that
(i) any two pseudolines either are disjoint (then we call them parallel), or
they meet in a single point and cross in this point, and
(ii) being parallel is transitive (that is, if a pseudoline intersects one pseudoline of a parallel pair, then it also has to intersect the other one).
ti
Now let P be an arrangement of n 1 pseudolines, which have been

labeled L i,...
and where a positive side has been chosen for each of
them. For every X E {-r, -, O n-1 , let Fx be the set of all those points
x E R2 which lie on the positive side of Li if xr , +, on the negative side
of ti if xr . _, and on the pseudoline Li if xr . O. This set Fx may be
empty: if not, it is called the face of P associated with X.
}
214
For example, in our next drawing small arrows are used to indicate the
positive side for each pseudoline. The shaded region (without its boundary) is the face associated with X = (-1---), the bold edge (without
endpoints) is the face associated with (-0+--), whereas (---0--) does
not correspond to a face.
3
.---->.
, -e4
Similarly, we assign labels to the faces at infinity: namely, we get a "face at

infinity" for every unbounded face in P, where we have to take into account
that "parallel lines meet at infinity." Thus, in the following drawing, the
two bold edges F1 , F2 and the shaded unbounded face F, with XF1 =
(--0--), X F2 = (---F0), and X F = (---1---), all determine the
same face GY at infinity, with Y = (--00).
Theorem 7.26. Let P = 11 ,... ,tn _ i l be a labeled arrangement of

pseudolines, for which positive sides have been chosen. Then the family of
sign vectors
V* (P) :=
U
U
{ (X',+) : F is a face of P}
{ (VG, 0 ) : G is a face at infinity for P} U {(O, 0)}
g f+,,01n
{(--X F, ) : F is a face of P}
is an oriented matroid of rank 3.
215
Proof. From an arrangement p of n 1 pseudolines, the previous procedure reconstructs a linear arrangement of n "pseudoplanes" through the
origin in R3 , where the nth pseudoplane is a straight plane, corresponding
to the line at infinity in P.
The following drawing shows two pseudoplanes of the resulting arrangement

in R3 : the fiat one corresponds to the line at infinity for P, and the nonflat
one corresponds to a pseudoline in P.
The oriented matroid v*(p) c {+, , 0}n arises from this arrangement
in the same way as in the case of a straight arrangement of planes in R3 .
Similarly, the proof that the system V* (P) is an oriented matroid is analogous to the realizable case in Proposition 7.22 except that the "linear
algebra arguments" of that proof have to be replaced by "combinatorial
arguments" that remain valid in the nonlinear case.
We leave the details to the reader, and refer to [92, Sect. 5.1], where the
111
proof is nicely done in even greater generality.
In fact, there is a surprisingly strong theorem available here: Lawrence's
topological representation theorem states that
every linear arrangement of n pseudohyperplanes in Rd yields an
oriented matroid V* C {,,O } of rank d.
Theorem 7.26 just presents the case d = 3 of this statement. (We do not
intend to give a precise definition of arrangements of pseudohyperplanes
here. Intuitively this should be clear; see [92, Ch. 5] for a careful explanation.) The topological representation theorem, however, also includes a
converse:
every oriented matroid on n elements of rank d can be represented
by an arrangement of n linear pseudohyperplanes in Rd , which is
essentially unique.
216
Thus we have a bijection

oriented matroids <----> (equivalence classes of) pseudoarra,ngements.
The second half of the theorem, constructing a pseudoarrangement for a
given oriented matroid, is by far the harder part to prove: it is not easy
even in the special case of d = 3.
Complete proofs of the topological representation theorem have been
given by Folkman & Lawrence [1941, by Edmonds & Mandel [181] (who first
proved the stronger piecewise linear version corresponding to our version
of polygonal pseudolines), in Bachem & Kern [31], and in Bj6rner et al.
[92, Chs. 4 and 5].
Definition 7.27. Two arrangements of pseudolines are combinatorially

possibly after relabeling, and after change of the positive
equivalent if
sides they have the same oriented matroid.
An arrangement of pseudolines P is realizable (or stretchable) if it is
combinatorially equivalent to a (straight) arrangement of lines, that is, if
the oriented matroid V* (P) is realizable.
Are pseudoline arrangements really more general than line arrangements?
Or is every pseudoline arrangement stretchable? Well, nonrealizable pseudoline arrangements do exist, and they are not even that hard to construct.
The following construction of a nonrealizable pseudoline arrangement already appeared in Levi's 1926 paper [334], the first paper on pseudoline
arrangements ever written (as far as I know).
Example 7.28. (Levi [334]) Consider the following arrangement of 8 lines

in the plane, the well-known Pappus configuration.
Pappus' theorem states that the three black dots are collinear in every line
arrangement that is combinatorially equivalent to this arrangement. This
implies that there is no straight representation of the following pseudoline
217
arrangement, the non-Pappus configuration.
However, this pseudoline arrangement does give an oriented matroid via

Theorem 7.26 (uniquely, if we number the pseudolines and specify a positive
side for each of them), the non Pappus oriented matroid Map. This is an
oriented matroid of rank 3 on 10 elements. However, if we delete the element
that corresponds to the line at infinity, then we get an oriented matroid
Mi p on 9 elements which is nonrealizable as well, since the line at infinity
was not needed for the nonrealizability argument.
-
There are even nonrealizable pseudoline arrangements that are simple (no
three pseudolines cross in a point or are parallel) see Ringel [427] [237,
p. 42] and Exercise 7.16 for an example with only 9 pseudolines in the
plane In fact, all (simple or nonsimple) arrangements of pseudolines with
at most 8 pseudolines are realizable (for this count the line at infinity, if
it is there), and Ringers example is essentially the unique simple one with
9 pseudolines, according to Richter [420].

What do you "see" if you "look at" a d-dimensional zonotope? You see its
"front facets," which are (d 1)-dimensional facets and which fill a shape
that is a projection of a zonotope, and thus is a zonotope itself.
If d = 3 (where this is most likely to happen to you anyway), and you
look from a point very far away, the picture might look like the drawing on
the next page. Here the shape you see is a 10-gon, filled "face to face" by
quadrilaterals and hexagons.
In general, the shape you see is a (parallel) projection of a zonotope,
which is a zonotope itself. It is covered by the images of the front facets of
the zonotope, which are also zonotopes.
218
Thus, "looking at zonotopes" leads to zonotopal tilings, which one can

formally define as follows.
Definition 7.29. A zonotopal tiling of rank d is a (d
1)-dimensional
polyhedral complex C such that both the union ICI and the faces F E C are
zonotopes.
The zonotopal tiling is regular if it arises from "viewing" a d-dimensional
zonotope from a point at infinity, that is, if it arises from a projection
: Z ---> ICI of a zonotope Z via the construction of Definition 5.3.
Now if this zonotopal tiling is regular, then it encodes a d-dimensional
zonotope, and thus it corresponds to an oriented matroid of rank d.
However, not all zonotopal tilings are regular. To see this, first observe
that in the case d = 3, there is an obvious way to "draw" an arrangement
of pseudolines into the zonotopal tiling, as on the next page.
Note that there is no similar (systematic) way to draw an arrangement
of straight lines into the picture although this is the picture of an actual 3-dimensional zonotope! So we see arrangements of pseudolines, in the
polygonal version of Definition 7.25, coming up quite naturally.
Do you recognize the pseudoline arrangement we have just drawn? It
is combinatorially equivalent to the Pappus line arrangement that we had
constructed before, with an extra horizontal line added through the three
special points that have to be collinear anyway, because of Pappus' theorem.
Thus this pseudoline arrangement is certainly realizable!
Lemma 7.30. All the pseudoline arrangements that come from a regular
zonotopal tiling of rank 3 are realizable.
219
Proof. Assume (without loss of generality) that the projected 3-zonotope

has a "special" summand in the direction of projection. Consider the associated hyperplane arrangement, which contains a dual to the special summand.
Now construct the affine arrangement by intersecting with an affine plane
that is parallel to the special one. The resulting arrangement of straight
lines is combinatorially equivalent to our pseudoline arrangement; they have
the same oriented matroid by construction. 0
For the zonotopal tiling above, the construction in the proof of Lemma 7.30 produces the line arrangement we had used to illustrate Pappus'
theorem, with the horizontal line through the three dots added.
Now let's modify the zonotopal tiling just a little, with all the worst
intentions, so this will also modify the pseudoline arrangement. We obtain
a zonotopal tiling, depicted on the next page, which "realizes" the nonPappus pseudoline arrangement, which is nonrealizable, inside a zonotopal
tiling! What happened?
We are about to hit a quite remarkable result, the Bohne-Dress theorem.
It claims that every zonotopal tiling of rank d represents an oriented matroid of rank d, and conversely it characterizes the oriented matroids that
can be represented by tilings of a given zonotope Z.
Now it is difficult here to supply the details for the two-way path from
zonotopal filings to pseudoarrangements, which the pictures suggest, since
we have not even defined pseudoarrangements and do not intend to do this
here (in order to avoid undue topological subtleties). However, everything
220
we discuss here has a higher-dimensional version, and we want to supply

at least the basic tools to describe them.
Thus, we work in a different direction: in view of the topological representation theorem it is sufficient to construct the oriented matroid associated
with a zonotopal tiling and this is just a system of sign vectors, without
any topology! Here we go.
The following provides the basic construction and shows how the faces in
a zonotopal tiling get sign vectors associated with them, almost canonically.
Construction (with Definitions) 7.31.
([104], [425, Sect. 11)
Let Z be a zonotopal tiling in d
Two edges e, e' E Z are defined to be equivalent if there is a sequence
e =----- eo , e l , ... , et = e' of edges in Z such that ei_ I and ei are opposite
edges in a 2-face of Z, for 1 < i < t.
If this divides the edges in Z into n equivalence classes, then n is the
number of zones of Z. Let El , 2 , ... , En enumerate the corresponding equivalence classes of edges. The ith zone of Z, denoted A, is the collection of
all those F E Z which have a face in E.
.
The edges of an equivalence class are all translates of each other, so we

can choose vectors v i E Rd such that the edges in are translates of the
edge [v i , vij c Rd. In this situation, we say that the vector configura-
ei
221
E Rnxd corresponds to the tiling Z. This vector

tion V :=
configuration is unique, up to relabeling and to reversal of signs.
The vector configuration V is a multiset: it may contain parallel or antiparallel vectors. It is not hard to see that the zonotope Z(V) it generates
is (a translate of) 121.
The choice of a vector v i also determines a positive side and a negative
side of the zone A. Thus we can associate a sign vector X F E {,-,Or
with every face F E 2, via
xr
I+ if F is on the positive side of the zone A,

0 if F E Zi,
if F is on the negative side of the zone A.
The set
0(2) := 1XF : F E
is the family of affine sign vectors of Z.

The following sketch shows one zone ZIL in the Pappus zonotopal tiling.
The zone consists of the bold edges (they form the set el ) and the shaded
2-faces. One possible vector y is indicated, and both the positive and the
negative sides depend on this choice: if we replace y by y1, then the
sides of Zi are exchanged.
positive side - - - of the zone Z1
,
_ _negative side
of the zone Z1
In the next sketch, all zones have been labeled and directed. The associated sign vector is indicated for one vertex, one edge, and one 2-face of the
tiling.
222
The Bohne-Dress Theorem 7.32.
[104, Thms. 4.1, 4.2] [425, Thm. 1.7]

Let V E be a vector configuration of rank d, let Z := Z (V) be its
zonotope, and let V* :--=-- V* (V) be its oriented matroid. If Z is a zonotopal
tiling of Z for which V corresponds to Z, then the family of sign vectors
V*
:----
{ (X, +) : X E 0(Z)}
U
(Y, 0) : E V*(V)}
{(X,---) : X E 0(Z)}.
is an oriented matroid of rank d+1. Furthermore, the construction

induces a canonical bijection between
the zonot opal Wings Z of Z(V) with associated vector configuration V, and
the oriented matroids V* C
{X e
{+, ,0}n :
,0}n 1 with
(X, 0) 1-1} = V*(V).
Proof. This is the correspondence suggested by the pseudoarrangement of

hyperplanes one can draw into every zonotopal tiling. We have certainly
seen that this is plausible, at least in rank 3.
The actual proof of Theorem 7.32, together with the proof that Construction 7.31 works correctly as claimed, is surprisingly difficult. We refer to Bohne's thesis [104], and to the new proof by Richter-Gebert
LI
Ziegler [425].
223
With the Bohne-Dress theorem, we can translate results about oriented

matroids into facts about zonotopal tilings, and back. In the following we
describe one (striking) instance.
Definition 7.33. A vertex in a zonotopal tiling 2 of rank d is simple if
it has degree d 1 and is a vertex of Z = 121, or if it has degree d and is
not a vertex of Z.
From Shannon's Theorem 7.20 we get the following estimate.

Corollary 7.34. Every regular zonotopal tiling of rank d with n 1 zones
has at least d simple vertices on the boundary of Z :=-- 121, and at least
n d simple vertices in the interior of Z.
One can see by elementary arguments that every zonotopal tiling of

rank 3 satisfies these estimates, even if it is not regular (cf. Exercise 7.15).
For example, the nonsimple vertices of the non-Pappus zonotopal tiling are
marked in the following figure all the other vertices are simple.
However, in the framework of pseudoarrangements one can construct oriented matroids (in terms of arrangements of pseudohyperplanes) that have
fewer than 2n simplicial regions. The first example of this kind was presented by Roudneff & Sturmfels [432].
The current "world records" about simple vertices are due to RichterGebert [421, Thm. 2.21, who constructed oriented matroids of rank 4 on 4n
elements that have only 6n simplicial regions. Furthermore, from RichterGebert's example [421, Thm. 2.3] we get an oriented matroid R(20) of rank
4 which has a pseudohyperplane "8" that is not adjacent to any simplicial
region. Furthermore, the restriction to the pseudohyperplane 8 (the contraction R(20)/8 of the oriented matroid) is realizable. Via the Bohne-Dress
theorem, these results translate into the following theorem.
224
Theorem 7.35.
There is a zonotopal tiling of rank 4 (in i 3 ), with 7 zones

(n = 8), which has only 7 simple vertices, and only 3 of them are on
the boundary.
(j ) [432]
(ii) [421, Thm. 2.2] For k > 2 there are zonotopal Wings of rank 4 (in 1R3 ),
with 4k 1 zones (n = 4k), which have only 3k + 1 = ln + 1 simple
vertices.
There is a 3-dimensional zonotopal tiling with 19
zones, which has no simple vertex on the boundary.
(iii) [421, Thm. 2.3]
You should try to visualize these in view of the "geometric" description of the pseudoplane arrangements in Richter-Gebert's paper with many
drawings this is not out of reach. A photo of a geometric model for the oriented matroid of part (0, built by Bokowski and Richter-Gebert, can be
found in [105, p. 562].
Notes
The permutahedron was first written about by Schoute in 1911 [445], it
seems General zonotopes were known to Blaschke [97, p. 250]. The first
systematic investigation of zonotopes was in Bolker [112], followed immediately by Schneider [439], and then by McMullen [364], who developed
zonal diagrams a version of Gale diagrams suitable for "zonotopes with
few summands" (see Exercise 7.7). There is a revived interest now, due to
the connection to oriented matroids, hyperplane arrangements, aspects of
optimization, computational geometry and convexity, and so on. We refer
to the surveys by McMullen [366] and by Schneider & Weil [442], to [92,
Sect. 2.2], and to the paper by Gritzmann & Sturmfels [226] and the references therein. More on Minkowski sums can also be found in [226].
The subject of hyperplane arrangements has a lot of different aspects,
and we do not even try to give an introduction here. We refer to [92] for the
case of real hyperplane arrangements and their oriented matroids, and for
further references. Arrangements of lines and pseudolines (corresponding
to arrangements of rank 3) are beautifully discussed by Griinbaum in [237].
Fans, polytopal or not, are of great interest for algebraic geometry. In
particular they represent toric varieties. In this case, the interest is restricted to fans that are pointed (i.e., {0} e .r) and rational (every cone
is generated by rational vectors). We refer to books by Fulton [2011 and
Oda [396], and in particular to the combinatorial treatment in Ewald [189].
Simple zonotopes exist, but they are rare. As we have seen, they correspond to simplicial arrangements of hyperplanes. Examples of such arrangements arise naturally in the theory of Coxeter groups, root systems,
225
and Lie algebras [1201 [128] [265] [92, Sect. 2.3]. There is a conjecture that
except for a few "obvious" infinite families, most of which come from these
theories, there are only finitely many "sporadic" examples: but currently
no one seems to have the faintest idea how to prove this. We refer to work
by Granbaum [236] for the case r = 3, and to Griinbaum & Shephard [240]
for the case r = 4. The enumeration of all "known" arrangements of rank 3
attempted in [236] had only one addition and one correction up to now
(Griinbaum [237], and Barthel, Hirzebruch & Iltifer [52]) and might be essentially complete, while the enumeration of [2401 for r > 4 is probably far
from complete; see Alexanderson & Wetzel [7, 8, 9]. Up-to-date references
can be found in Wetzel [519].
The Bohne-Dress theorem was announced by Andreas Dress at the 1989
Symposium on Combinatorics and Geometry in Stockholm. It is a strikingly simple geometric observation that had previously eluded people. A
complete proof, however, is surprisingly difficult, and it took some time
until the complete written version by Bohne [104] was available. A simpler, more geometric proof is given by Richter-Gebert & Ziegler [425]. Our
sketch in Section 7.5 follows that paper.
The Bohne-Dress theorem relates the set of all zonotopal tilings on a
given zonotope with an extension space problem ("Is the space of all extensions of an oriented matroid homotopy equivalent to a sphere?," see
Sturmfels & Ziegler [499]). Thus zonotopal tilings allow one to study a special case of two very basic, general, and apparently very difficult problems,
the generalized Baues problem" of Billera, Kapranov & Sturtnfels [69 ], and
the problem of "Combinatorial Grassmannians" by MacPherson [346], see
also in Mnv & Ziegler [380]. We will discuss the setting of the Generalized Baues Problem in Lecture 9. For the problems themselves and their
ramifications we refer to the original sources.

7.0 In Definition 7.1, show that if the cones are convex, then they are
automatically polyhedral, so the condition "polyhedral" could be
dropped from the definition.
7.1 Let P C Rd be a polytope with 0 in its interior. Show that the face
fan of P is the normal fan of the polar .13 , and the normal fan of P
is the face fan of P.
7.2 Enumerate all the 3-zonotopes generated by 5 vectors in R3 , draw
them, and count the vertices, facets, and simple vertices.
What about 6 vectors in R3? What about 5 or 6 vectors in R4?
7.3 Give examples of simple zonotopes and of simple zonotopal tilings.
226
7.4 Show that the following are equivalent for a polytope Z:

every 2-face of Z has an even number of edges, and opposite
edges are parallel.
for every edge, Z has some multiple or part of it as a Minkowski
summand.
the normal fan of Z is a hyperplane arrangement.
(Such polytopes, generalized zonotopes, were introduced by the Russian crystallographer Fedorov; Coxeter apparently misunderstood the
definition and assumed that Fedorov was considering zonotopes
see
Taylor [501]. Bolker [112] studies them under the name of planets,
Baladze [32] calls them belt polytopes.)
Give examples of generalized zonotopes that are not zonotopes. The
above shows, however, that every generalized zonotope is combinatorially equivalent (in fact, normally equivalent) to a zonotope.
7.5 If every projection of a polytope to

tope is a zonotope itself.
3 is a zonotope, then the poly-
Show that the projections to R2 are not good enough for this.
(Witsenhausen [531])
7.6 Interpret the deletion and the contraction of a vector in the configuration V in terms of zonotopes. That is, describe how the zonotopes
Z(V\v) and Z(V/v) can be constructed geometrically.
7.7 Let Z(V) be the zonotope generated by a configuration V E I. d x n
G E (R*( n-d / Xn be the dual vector configurawhich spans I.
tion.
).
(i) How can the combinatorics of the zonotope Z(V) be read off
from the configuration G?
(ii) Use this to describe the zonotopes with n < d + 2 zones.
(iii) Describe the relation between the zonotope Z(V) and its asso-
ciated zonotope Z (G) C
(Rn-d )* .
(This was developed by McMullen [364)
7.8 Assume that x, y E Rd are given. Give an explicit formula for some
small enough E > 0 such that sign(x Ey) = sign(x) o sign(y).
7.9 Let V* C {-1-, --,O }n be a system of sign vectors.
(j ) Assuming that (V0): 0 E V* holds, show that the axioms (V1)
and (V2) together are equivalent to the axiom
(V2'): u, V E V
==U 0 (V)
V.
227
(ii) Consider any affine arrangement of n hyperplanes in which a

positive side has been chosen for each of the hyperplanes.
Show how a sign vector is associated with every face of the
arrangement, and the resulting collection of sign vectors satisfies
(V2`) and (V3), but not in general (VO) and (V1).
(iii) Show that not every sign vector system satisfying (V2') and
(V3) corresponds to an affine arrangement, even if we admit
topologically deformed arrangements ("pseudoarrangements").
(The characterization of the sign vector systems of affine oriented
matroids was a difficult combinatorial problem, recently solved by
Karlander [2891.)
7.10 Let D = (V, A) be a directed graph with n arcs, A -= {al, , an }.
With every subset U C V of its vertex set, associate a sign vector
b(U) E I+, ,O}', the "directed cut" of U, where b(U) i = + if ai is
an arc that leaves the set U; (U) i = if ai enters the set U; and
b(U) i = 0 otherwise:
if tail(ai ) E U and head(ai ) U,

if tail(ai ) U and head(a i ) E U,
b(U) i =
0 if head(a i ) and head(a i ) are both in U,
or both not in U.
For example, in the graph we have drawn
o
y3 , v4 1)
\0 j
Show that the family V* := { 6(U): U C V} is an oriented matroid.
What is its rank?
What is the relation to the oriented matroid associated with such a
digraph according to Exercise 6.3?
7.11 Show directly from the axioms in Definition 7.21 that every oriented
matroid of rank r < 2 (that is, an oriented matroid V* C {, , Or
that does not contain a chain 0 < X < < X") is realizable.
7.12 Show the following theorem.

Let V*, C* C
,01n be sign vector systems such that C* is the
collection of sign vectors of minimal nonempty support in V*,
C* =MIN(V*),
228
while V* is the collection of all conformal products of sign vectors in C*,
V* = fOoy l o...ovk:k>0, vi E C* for i = 1, 2, . . . , kl.

Then V* is the covector set of an oriented matroid (that is, it satisfies
the axioms of Definition 7.21) if and only if C* is the set of cocircuits
of an oriented matroid, that is, if C* satisfies the following cocircuit
axioms:
(CO) 0 f/ C* ("The zero vector is not a cocircuit" )
(Cl) u E C* = U E C*
("The negative of a cocircuit is always a cocircuit")
(C2) u, v E C*, supp(u) C supp(v) = u = v
("Co circuits have noncomparable supports")
(C3) u, v E C* , U
V, j E S(U, V) == w E C*, w' E I+, , Or:
W < w', and w' eliminates j between u and v ("The set of
cocircuits admits elimination")
7.13 Show that for any vector subspace U C Rn , the set of minimal
nonempty supports in II = SIGN(U), given by MIN(SIGN(U)) =
fu E sign(U)\{0} : v E sign(U), supp(v) C supp(u) implies v = 01,
is the set of cocircuits of an oriented matroid, that is, it satisfies the
axioms of Exercise 7.12.
7.14 How can you test whether a given zonotopal tiling is the picture of an
actual zonotope? Show that, essentially, one has to decide whether a
certain polyhedron has nonempty interior, which can be solved as a
linear programming problem, as in Exercise 5.2 (i ).
So, is it true that the first figure in Section 7.5 represents the picture
of a 3-dimensional zonotope?
7.15 Show that every nontrivial zonotopal tiling in R2 (a tiling of a centrally symmetric polygon by centrally symmetric polygons) has a simple vertex on the boundary, and also a simple vertex in the interior.
7.16 Consider the following zonotopal tiling.
229
Show that it corresponds to an arrangement of 9 pseudolines which

is not stretchable, because it violates Desargues' theorem.
Show that the pseudoline arrangement of the following tiling is also
not stretchable:
(The second configuration is closely related to Ringel's simple configuration of 9 pseudolines, one only has to delete the line at infinity
and perturb the arrangement there.
These drawings were produced by Jiirgen Richter-Gebert, using his
postscript program described and listed in [422], which produces exceptionally nice pictures of zonotopal Wings.)
7.17 Consider the following sirnplicial arrangement of 16 pseudolines from

Griinbaum [237, p. 441.
(i) Show that it is not realizable.
(ii) Use it to construct a simple zonotopal tiling of a 12-gon whose

oriented matroid is not realizable.
230
7_ Fans, Arrangements, Zonotopes, and Tilings
7.18 For every d-zonotope, the numbers fk of k-faces satisfy the relations
>
dk+l
fk for 1 < k 5_ d, and thus fk <
d
) fo .
k
(This is given in terms of hyperplane arrangements and oriented matroids in Fukuda, Tamura Sc Tokuyama [200, 1991.)
7.19 For any vector configuration V = {v 1 , , vn } c lY, prove the volume formula for its zonotope:
vol(Z(V)) = 2d -
E
1<ii<
Idet(v i
, vid )I .
<td<n
For this, decompose the zonotopes into parallelepipeds, whose volumes are given by determinants.
(McMullen, see Shephard [459, Sect. 51, from where we have also
taken the illustration.)
7.20* You can use the formula in the preceding exercise to compute the
volume of a zonotope, but that is not very effective: the formula
has
terms, which may all be nonzero.
Is there a fast (polynomial) way to compute the volume of a zonotope?
(Answer: most probably not this is #P-hard according to Dyer,
Gritzmann & Hufnagel [177 ] .)
(n
d)
Sheliability and
the Upper Bound Theorem
Perhaps the most famous result about convex polytopes is the EulerPoincar formula:
fi
fo
+12 + - - + (-1)d-1 fd, + (-1) d fd == 0,
where fi denotes the number of i-dimensional faces of a d-polytope P. Here

fi = 1 and h--= 1 correspond to the trivial faces (the empty face and
the polytope itself), and b, fi , fd- 2 , and Id-1 are the numbers of vertices,
edges, ridges, and facets, respectively. So for 2-polytopes we obtain that
0, the number of vertices equals the number of edges (not much
fo
of a surprise). For 3-polytopes we get "Euler's formula"
ve+ f =2
for a 3-polytope with v = fo vertices, e = fi edges, and f f 2 facets.
For d < 3 the Euler-Poincar formula is easy to prove, but for higher
dimensions care is needed. As Griinbaum [234] observed, all the classical
inductive proofs (starting with [437] proof from 1852; see also
Sommerville [470, p. 1471, Schoute [444, p. 611, and the references in [234,
p. 1411) assume that the boundary of every polytope can be built up inductively in a nice way, that is, it can be "shelled." That this is in fact possible
was only proved by Bruggesser & Mani in 1970. A striking application of
shellability was McMullen's proof of the "upper bound theorem" in the
same year, 1970.
In this lecture we have several big goals. We start with shellability for
polyhedral complexes, a concept that is both useful and nontrivial. We will
show that
232
8. Sheliability and the Upper Bound Theorem

polytopes are shellable,
subdivisions of polytopes are not shellable in general, and
while shelling polytopes, one can get stuck (that's a new result).
Then, we'll present McMullen's proof for the upper bound theorem, give
a glimpse of extremal set theory, and end with the famous 9-theorem, and
derive some of its surprising consequences. So, there's a lot to do: let's get
going.
... wait, here is one more remark. This lecture has a distinctive "topological" flavor. In fact, already the first correct and complete proof of the
Euler-Poincar formula, by Poincar [410, 411], was done using tools of algebraic topology that Poincar had developed himself. Here we will avoid
most topological subtleties, for example by restricting our attention to polyhedral subdivisions of polytopes and their boundaries, instead of subdivided
topological balls and spheres. Thus, for this lecture no knowledge is needed
of the wonderful subtleties of piecewise linear topology, nor of the powerful
machinery of algebraic topology. Nevertheless, it is helpful and desirable,
and good for your intuition, if you take, at least, an excursion into these
worlds. I recommend Stillwell [491], Munkres [388], Daverman [171], and
Bjeirner [85] as guides to different points of view.

A polytopal complex (see Definition 5.1) is a finite, nonempty collection C
of polytopes (called the faces of C) in r d that contains all the faces of its
polytopes, and such that the intersection of two of its polytopes is a face
of each of them.
The dimension dim(C) of a polytopal complex is the largest dimension of
a polytope in C. A polytopal complex is pure if each of its faces is contained
in a face of dimension dim(C), that is, if all the inclusion-maximal faces of C,
called the facets of C, have the same dimension. The underlying set of C is
F
the union of its faces, ICI := U
For example, a graph represents a polytopal complex if it is drawn in the
plane or in R3 with straight lines that do not cross. It has dimension 1 if
the graph has at least one edge, and then it is pure if the graph does not
have an isolated vertex.
In Lecture 5 we met five important classes of polytopal complexes:
.
'
(i) Every polytope P together with all its faces forms the polytopal complex C(P). The only maximal face ("facet") of this complex is P itself.
(ii) All the proper faces of P form the boundary complex C ( P), whose
facets are just the facets of P. This is a pure simplicial complex of
dimension dim(P) 1.
8.1 Sheliable and Nonshellable Complexes
233
(iii) Any Schlegel diagram of P with respect to a facet F (remember

Lecture 5?) forms a pure polytopal complex D(P, F), and the facets
of this complex correspond to the facets of P that are different from F.
(iv) Every d-diagram D is a polytopal complex.
(A) The pile of cubes Pd(zi,... , zd), as defined in Example 5.4, is a pure
polytopal complex, with z1-z2.... -zd facets.
All of these complexes are pure. In the following picture gallery, the first
complex is pure 1-dimensional (a graph), the second one is not pure, and
the third one is pure 2-dimensional.
<
We now proceed to define shellability, in a version that is slightly more

restrictive than the original one used by Bruggesser & Mani [132]. Several
variations are discussed in Danaraj & Klee [163]. It turned out in the work
of Bjerner & Wachs [94] [81] [92, Sect. 4.7] that the following one is the
one that "works" in very general geometric and combinatorial contexts.
Definition 8.1. Let C be a pure k-dimensional polytopal complex. A

shelling of C is a linear ordering F1 , F2, .. , F, of the facets of C such that
either C is 0-dimensional (and thus the facets are points), or it satisfies the
following conditions:
(i) The boundary complex
C(3F1 )
of the first facet F1 has a shelling.
(ii) For 1 < j < s the intersection of the facet Fi with the previous
facets is nonempty and is a beginning segment of a shelling of the
(k 1)-dimensional boundary complex of Fi , that is,
j-1
Fi
FT)
G1
G2 U U
z=1
for some shelling GI, G2 , .. , Gr , . , Gt of c(aFj ), and 1 <r < t.

(In particular, this requires that Fi n (UL-11. Ft ) has a shelling, so it
has to be pure (k 1)-dimensional, and connected for k> 1.)
A polytopal complex is shellable if it is pure and has a shelling.
Thus shellability is not defined for complexes that are not pure, although
it turned out in recent work by Bj6rner & Wachs [95] that an extension of
the concept to nonpure complexes is possible and useful.
234
Examples 8.2.
(i) Every 0-dimensional complex is shellable, by definition. A 1-dimensional complex (a graph) is shellabIe if and only if it is connected. In
particular, this means that it is pure (i.e., has no isolated vertices). A
shelling order is an ordering of the edges e l , e2 , ... , es in such a way
that the set { e l , ... , ei } describes a connected subgraph for every j:
this comes from the condition that the intersection of the edge e3
with the earlier edges has to be 0-dimensional, and thus nonempty.
(ii) The following are three 2-complexes in the plane R2 .
The first two are not shellable, but the third one is. (Check this!) In
each of them, there is a beginning of a shelling indicated, that is, the
complex given by the numbered facets together with their faces is
shellable. However, if you try to add the last facet in any of the first
two examples, you violate condition (ii) of Definition 8.1.
(iii) Every simplex is shellable, and every ordering of its facets is a shelling
order. This immediately follows by induction on the dimension, since
the intersection of Fi with Fi (i < j) is always a facet of F3 in this
case.
(iv) The d-cubes are shellable: by induction on dimension one can show
that every ordering of the 2d facets F1 , F2 , ... , F2d such that the
first and the last facet are opposite, F1 = F2d, is a shelling order. (The condition F1 = F2d is sufficient, but not necessary, see
Exercise 8.1(i)!)
(N) The pile of cubes Pd(a l , ... , ad), see Example 5.4, is shellable for
arbitrary finite ai > 1. For this, we use part (iv) to see that the
lexicographic order on the little cubes in the pile is a shelling order.
Remarks 8.3.
(i) We will see in the next section that condition 8.1(i) is in fact redundant: the boundary complex C(8F1 ) of every polytope is shellable.
However, if one defines shellability more generally for cell complexes
rather than polytopes, as in Bjiirner [81], then this is necessary.
(ii) For simplicial complexes condition 8.1(i) is redundant because of Example 8.2(iii). In this situation, condition 8.1(ii) can also be simplified
considerably: it can be replaced by
235
8.1(W) For I < j < s the intersection of the facet Fi with the previous
facets is nonempty and pure (k 1)-dimensional.
In other words, for every i < j there exists some I <j such that
the intersection Fi n Fi is contained in Fi n Fi , and such that
Ft n Fi is a facet of Fi .
(iii)
One might be tempted to weaken condition 8.1(ii), and only require

the following:
8.1(ii11) For 1 < j < s the intersection of the facet Fi with the previous
facets
i--1
Fn (U Fi) = Gi u ... U Gr
i=1
is nonempty, pure (k 1)-dimensional, and shellable.

This can be done. It yields the original definition by Bruggesser &
Mani [132], which is weaker than Definition 8.2. Although the main
conclusions of shellability one wants remain valid, this weaker version
does not have the nice combinatorial characterization that is possible
for the stronger version [94].
We continue, with more examples.
Examples 8.4.
(i) Every polytopal subdivision of a 2-polytope is shellable see Exercise 8.0.
(ii) The boundary of every 3-polytope is shellable. This follows from (i):
for this we first shell a Schlegel diagram D(P,F) of P, which is a subdivision of a 2-polytope. This corresponds to a shelling of the whole
boundary OP except for the facet F. The shelling can be completed
by taking F as its last facet.
Subdivisions of 2-polytopes and boundaries of 3-polytopes are easy to
shell. One reason is that no matter how we start the shelling, we can't get
stuck. Let's introduce some fancy terminology for this.
Definition 8.5. (Danaraj & Klee [164, p. 37D

A polyhedral complex is extendably she liable if every partial shelling can be
continued, that is, if for every shellable subcomplex of the same dimension
there is a shelling of the whole complex that shells the subcomplex first.
In this sense, we see from Examples 8.4 that subdivisions of 2-polytopes
as well as the boundaries of 3-polytopes are extendably shellable. Similarly,
one can show that the d-cubes are extendably shellable (Exercise 8.1(0).
236
Lemma 8.6.
The pile of cubes P3(9, 9, 4) is not extendably shellable.
Proof. For this, consider the picture below. It depicts a subcomplex of

the pile, broken into layers (so you can see what happens inside). The
subcomplex is easily seen to be shellable (for example, first shell the bottom
layer, then the next layer and the walls, then add the central axis, then the
top layer of the interior part, then the six remaining cubes).
ro
Two
ff
r
O
Af
mfAr.149 0
aria*
:1410
"Iir WIMP
lOr aI I I
41111/41ffilala
AIPerarawfa
10
kil150
PALP
ATIZI1
rillIAWAW0
41111IAIVAIWAWANIMIIII,AIMIA1171
idwarderAwderAwarAn
ArAiramorAmormrAlisimr,
ArrAmordwrArramorararraro
411/411101111P2111/4111FAININFAIIII
AVAIIPTAFAIVAIIIIIIPTAIINIVA/1
Adur,mrarr,mprdurdrPrar.arav
111111 111111111
A different way to draw this type of complex (more abstractly) shows

the four layers, top to bottom, from left to right.
111111

MM.

MINER
MEN
N.=
If we try to add any new cube of the pile P3(9, 9,4) to the subcomplex,
then the intersection with what is already there is not pure 2-dimensional,
or not connected, or both. Thus, this is a shellable part of P3(9, 9, 4), and
we are stuck: so P3 (9 7 9, 4) is not extendably shellable.
D
237
For a polyhedral complex C, define the star star(v,C) of the vertex v to

be the polytopal subcomplex of all faces that contain y, and their faces.
Let the link be the subcomplex link(?), C) of all faces G E star(v,C) of the
star that do not have v as a vertex.
star(v, C)
If C is pure of dimension d, then so is star(v,C), and link(Y,C) is then pure

of dimension d - 1.
The following lemma is a quite trivial, but important, piece of information about the "local structure" of shellable simplicial complexes. For
nonsimplicial complexes, it becomes nontrivial (if it is true: Problem 8.4*).
Lemma 8.7. Let C be a shellable simplicial complex, with shelling order F1 , F2 ,. , Fs . Then the restriction of this order to star(?), C) yields a
shelling order for the star, and also for link(v, C).
Proof. We directly verify condition 8.1 (u i1 ): let Fi be a facet in the star
(so y G Fi ), and let Fi be an earlier facet that also lies in the star (with
< j and y E Fi ). Since we have a shelling of C, there is a facet F1 with
1 <j such that Fi n F3 C F1 n Fi . But this implies that y E F1, so F1 is in
the star of v.
The same proof also shows that we get a shelling of the link, since (in
the simplicial case) we have a bijection between k-faces G E link(?), C) and
LI
(k+1)-faces a = conv(G U y) E star(v, C).
Theorem 8.8. (Rudin [433 ] )
The 3-simplex 6.3 can be triangulated in a nonshellable way.
The first nonshellable triangulation of a tetrahedron (with 14 vertices,
41 facets, all vertices on the boundary) was constructed by Rudin [433]
in 1958. Her construction is subtle and hard to visualize, and it seems to
be the only one in the literature. So, instead of reproducing it, here is a
different construction that shows that the tetrahedron and the 3-cube have
nonshellable triangulations.
Example 8.9 (The Danzer cube). Let C3 be our standard cube in R3 ,

and let V 12 be the set of 12 midpoints of the edges of C3, that is, the set
of all points in R3 with one coordinate 0 and the others equal to 1. The
twelve points lie on the boundary of the cube C3, but also on the boundary
238
of the tetrahedron
T3 :=--
IX
E R3
+x+y+z<2
+x y z < 2
x+yz <2
x y + z < 2
2\
2 ) ( +2 ) ( +2 )
2 j , ( +2 ,
2 , +2
2
+2
+2
2
Next we construct a set E' 2 of 12 edges between the points in V 12 . For

this, take one edge that connects the midpoints of two skew edges of the
cube, as well as all the images of that edge under the 12 symmetries of the
cube which correspond to orientation-preserving symmetries of the tetrahedron T3.
It turns out that the 12 edges we get (see the drawing) are grouped into
four disjoint triangles. The key property is that every edge is surrounded
by a triangle. By symmetry, it is in fact sufficient to verify this for one
single edge e E E12 .
The next step is to construct a triangulation of the cube C3, respectively
of the tetrahedron T3, which contains the edges in E12 as faces (not subdivided!). While this can be done explicitly, here we resort to a powerful
tool: Whitehead's completion lemma [523, Thm. 5], according to which every partial triangulation of R3 can be completed. (In fact, the same is true
in R d , according to Bing [76, Lemma 6]. See Bing [78, Sect. 1.2] for a textbook version.) Hence, we can take any triangulation of the boundary of C3,
respectively T3, that uses the vertices in V12 , plus the twelve edges in E12 ,
and complete this simplicial complex to a triangulation of C3, respectively
T3. There will be additional interior vertices necessary for this.
239
Now assume that the resulting simpliciai complex C is shellable. We start

with one tetrahedron F1 , and then add new ones, F2 F3 . . . At every step,
except possibly the first, not more than one new edge from the set E 12 can
be added, since the edges in one of the triangles cannot be in a common
simplex, and if two skew edges were new, then this would contradict a
characteristic property of shellings of simplicial complexes: there has to be
a unique minimal new face at each step, which is nonempty after the first
step (see Exercise 8.2).
Now let F3+1 be the tetrahedron by which the last edge e from E 12 is
added. At this point of the shelling the complex C3 C(FI , , Fi ) already
contains the three edges e l , e2, e3 that surround e. In fact, the circle formed
in C3 by C := el U e2 U e3 can be contracted within the complex C3 : this is
a property of shellable complexes of dimension at least 2, which is easy to
verify by induction (they are "simply connected").
Since the circle C surrounds the edge e, we have to "pass over" a vertex
C e when we contract C in C3 . Also, the link of this vertex y in the
complex C3 , link(Y, C3), is a 2-dimensional shellable complex, by Lemma 8.7.
However, we can contract our circle C within IC3 I until it lies in link(v, C3 ),
but it cannot be contracted within the link, because then it would not pass
over v. This shows that the link is not shellable: contradiction.
El
8.2
Shelling Polytopes
The complex of a polytope C(P) is shellable if and only if the boundary

complex C(OP) is shellable. Thus, "shelling a polytope" means finding a
shelling order for the facets of P. How do we do this? Is there an obvious
240
8. Shellability and the Upper Bound Theorem
way? If we polarize, then the problem is to find a good ordering on the

vertices of the polar polytope P. For this an obvious thing to do is to
take a linear function in general position (see Sections 3.1 and 3.4), and
to order the vertices according to that linear function. This works: it does
yield a shelling (in fact, many shellings) for the boundary of P.
Below we describe these shellings directly on P, without polarization. In
Exercise 8.10 you are asked to check that the constructions on P and P
really are equivalent.
Lemma 8.10. If F1 , F2 ,... , Fs is a shelling order for the boundary of a

polytope P, then so is the reverse order Fs , F8 _ 1 ,... , F1 .
Proof. Let Fi be one of the facets in the shelling, then for every facet G of
Fi there is a unique other facet Fi of P such that G = Fi n F. This other
facet Fi can be either earlier (i < j) or later (i > j) than Fi . These roles
are interchanged if, while reversing the shelling of UP, we also reverse the
shellings of the boundaries of its facets: and this we may do, by induction
D
on the dimension.
Theorem 8.11.
(Bruggesser & Mani [132 ] ) Polytopes are shellable.
Well, this is the essence. However, what one actually needs is that there
are shellings with very specific properties. These are obtained from the
Bruggesser-Mani construction, which yields the much more specific theorem
below (which includes later refinements by McMullen [3611, Danaraj & Klee
[163 ] , and by Bj6rner & Wachs [94 ] ). Also, this is the technical statement
that has an easy proof by induction on the dimension.
Theorem 8.12. [132] Let P C 1W1 be a d-polytope, and let z c id be a

point outside P. If z lies in general position (that is, not in the affine hull
of a facet of P), then the boundary complex c (U p) has a shelling in which
the facets of P that are visible from z come first.
241
Here we can use our intuition to understand what visible means: a facet
F C P is visible from x if for every y E F the closed line segment [x, y]
intersects P only in the point y. Equivalently, F is visible from x if and
only if x and int(P) are on different sides of the hyperplane aff(F) spanned
by F. For example, if xG is beyond the face G (in the sense of Section 3.1),
then the facets that contain G are exactly those that are visible from xG.
Proof. Given x, we choose a line i through x and through a point in

general position in P. The properties we need are that i contains x, hits
the interior of P, and the intersection points with the facet hyperplanes
in aff(F) are distinct. For simplicity, assume that the line is not parallel to
any of the facet hyperplanes, so we have no intersection point "at infinity."
We orient the line i from P to x.
A
aff(F2)
A
Now imagine P to be a little polyhedral planet, and have a rocket start

on its surface at the point where the oriented line i leaves the planet. This
point lies on a unique facet F1 , and for the first few minutes of the flight
only this one facet F1 is visible from the rocket.
A
242
After a while, a new facet will appear on the horizon: the rocket passes
through a hyperplane aff(F2 ), and we label the corresponding facet F2Continuing this, we label the facets F3, F4, . . . in the order in which the
rocket passes through their hyperplanes, that is, in the order in which the
facets appear on the horizon, becoming visible from the rocket. Now we pass
through infinity, and imagine that the rocket comes back to the planet from
the opposite side. We continue the shelling by taking the facets in the order
in which we pass through the hyperplanes aff(Fi ), that is, in the order in
which the corresponding facets disappear on the horizon.
This "rocket flight" clearly gives us a well-defined ordering on the whole
set of facets. Also, the facets that are visible from x form a beginning
segment, since we see exactly those at the point where the rocket passes
through x.
To see that the ordering is a shelling, we consider the intersection OFi n
(F1 u u Fi _ i ). If Fi is added before we pass through infinity, then this
intersection is exactly the set of those facets of Fi that are visible from the
point t n aff(Fi ), at which F3 appears on the horizon. Thus, we know by
induction on the dimension that this collection of facets of Fi is shellable,
and can be continued to a shelling of the whole boundary aFi .
aff(F7 )
I
After passing through infinity, the intersection is the family of nonvisible

facets. This is shellable, too, because reversing the orientation of the line t
LI
yields the shelling with the reversed ordering of the facets.
Shellings that arise by Bruggesser & Mani's construction are also known
as line shellings. Note that in this construction, reversing the orientation
243
of the line i also reverses the line shelling. Thus the reverse of every line
shelling is not only a shelling (by Lemma 8.10), but a line shelling as well.
The Bruggesser-Mani construction has a lot of flexibility: we can get
shellings with special properties by suitable choice of the shelling line L.
Corollary 8.13.
For any two facets F and F' of a polytope P, there is a

shelling of OP in which F is the first facet and F' is the last one.
For every vertex y of P, there is a shelling such that the facets that
contain y form a beginning segment of the shelling, that is, the star of the
vertex y is shelled first.
Proof. For the first claim, choose any shelling line 1 which intersects the
boundary of P in the facets F and F'. (For example, choose x beyond F,
choose x' beyond F', and let f be the line determined by x and z'. Perturb
f to general position, if necessary.)
For the second claim, let x v be a point beyond the vertex y, and choose
the shelling line to contain this x v .
111
Corollary 8.14. Every Schlegel diagram is shellable.
More generally, every regular subdivision of a polytope is shellable.
Proof. For any Schlegel diagram D(P, F), choose a shelling of the polytope
P such that the facet F comes last. Thus the shelling of P also induces a
shelling of the Schlegel diagram D(P, F).
Every regular subdivision Ep(Q) of a d-polytope Q is, by Definition 5.3,
isomorphic to the complex of faces of a (d+1)-polytope P that are visible from a certain point z = Ted +i: and thus we can apply Theorem
8.12.
0
This corollary applies to piles of cubes (Example 8.2(v)), for example. It

suggests that every d-diagram is shellable. Is that true? (Problem 8.3*)
Knowing that polytopes are shellable, one might ask whether they are
extendably shellable. Here we answer this question in the negative, and
give in fact several proofs.
Theorem 8.15. Not all 4-polytopes are extendably shellable.
In particular, 04
- (7, 7,3) and 734(7, 5,5) are not extendably shellable.
244
Proof. (1)
Consider the 4-polytope P4 := P4(9, 9,4), which contains

a combinatorially equivalent copy of the pile of cubes P3(9, 9,4) on its
boundary. Now we start a shelling of OP4 with the facets that correspond
to the nonextendable partial shelling of P3(9, 9,4) given in Lemma 8.6. This
partial shelling has two connected parts in common with the boundary of
the box B :=---- 1P3(9, 9, 4)1. Any facet of P4 that we can add to continue
the partial shelling of ap4 will be connected to one of these parts of the
boundary of the box, but not to both. Thus, the "missing cubes" from the
box cannot be added to the partial shelling of ap4 at a later time, either.
This proves that f,4(9, 9,4) is not extendably shellable. In fact, it also
proves that any 4-polytope that contains an isomorphic copy of P3(9, 9,4)
in its boundary complex is not extendably shellable. So, fj'4(z1, z2, z3) is
not extendably shellable for z1 > 9, z2 > 9, and z3 > 4.
(2) We delete the "bottom layer" and the "walls" of the pile P3(9, 9, 4),
to obtain P3(7, 7, 3). Now consider the Schlegel diagram of P4(7, 7, 3), and
start shelling it by first shelling the 5 facets below and next to the pile
of cubes. (This replaces the little cubes in the bottom and the walls of
our previous example.) We continue the shelling by the 49 little cubes
that correspond to the interior cubes in the not-extendable configuration
of Lemma 8.6. If we lift this partial shelling of the Schlegel diagram to the
boundary of P4(7, 7, 3), then we can also add the facet on which the diagram was based: after that we are stuck. Thus P4(7, 7, 3) is not extendably
shellable.
(3) Start a shelling of P4 := f/54(7, 5,5) at the facet F1 that is completely
disjoint from the pile of cubes. Then the remaining facets are the ones in
the Schlegel diagram 730(P4, F1). Of this Schlegel diagram we next add the
bottom and the top facet to the shelling, as F2 and F3. Then, from the
pile of cubes isomorphic to P3(7, 5, 5), we take little cubes F4, F5 , . . . along
a knotted curve that connects the bottom facet to the top facet of the
Schlegel diagram, as suggested by the drawing on the next page. All this
yields legal shelling steps, until one reaches the last (white) cube, which is
in the top layer of the pile of boxes.
However, now the complex contains a knotted curve, for which all edges
except for one are on the boundary of the subcomplex that we have shelled
at this point. This knotted curve would be completed by the little white
cube from the top layer. Now if we continue the shelling elsewhere, then
there remains to be a knotted curve of this type: so the little white cube
in the drawing cannot be added to the shelling in a later step, either. Thus
D
this partial shelling of /34(7,5,5) cannot be completed.
The third proof for Theorem 8.15 can also be adapted to see that neither simple polytopes, nor simplicial polytopes, are extendably shellable in
general. Also, it can be generalized to see that "most" 4-polytopes are not
extendably shellable. We refer to [538].
245
moiffo
Shellings of polytopes allow us to "build up" the boundary of a polytope step by step, adding one facet at a time. Thus one can do proofs by
induction on the number j of facets in the complex
C3 := C(111 U F2 U U F3 ),
which represents a shellable part of the boundary of a polytope.

Our first, simple but classical, application of this technique will prove
the Euler-Poincar formula.
Definition 8.16.
is the vector
The f -vector of a d-dimensional polyhedral complex C
f (C)
(fi,
fa,
- , fd)
c Nci+2 ,
where fk = fk(C) denotes the number of k-dimensional faces in C.

By the f -vector of a d-polytope we mean the f-vector of its boundary
complex:
f (P)
f (c(aP)) =- (f -1, fa, h,- -
, f_)
Note that all the f-vectors we consider start with the entry /I = 1,
corresponding to the empty face. The f-vectors of polytopes satisfy only
one nontrivial linear equation: the Euler-Poincar formula.
246
Corollary 8.17 (Euler-Poincar formula).
For every d-dimensional polytope,
fo fi + - - + (- 1 )d ifd_ i
=1
Proof. In order to do induction, we need to consider the alternating sum

of face numbers for general polytopal complexes. For this we define the
(reduced) Euler characteristic x(1)) of a polyhedral complex of dimension
at most d, by
X(D) :=
- f--1 lo - fi +"- - + (-1) dfd.
Now if V and TY are polytopal complexes such that the union is a polytopal
complex too (that is, if F n F' e D n V' for F E 1), F' E D'), then the
Euler characteristic is additive:
x(D) + x(D') --= x(D U V') + x(D n D').

We prove by induction on d that the complex of a polytope C(P) always
has Euler characteristic 0, and thus its boundary has Euler characteristic
( 1 )" -1 :
x(C(aP)) = (_1)d-1 .
x(C(P)) --= 0,
This is clear for d < 1. Now if P is a d-polytope with shelling order
F1 , F2 , ..., then we have more precisely that
W(F1 U F2 U ' ' ' U
Fj))
fo
for 1 < j < fd-i

for j = Id-i
which follows by induction on j and dimension, since the facets Fi

that we add in are (d - 1)-polytopes, the Euler characteristic is additive,
and the intersection Fi n (Iji<i Fi ) is a shellable part of, but not the whole
boundary of Fi, for j < fd_i . This last fact is immediate from Lemma 8 10,
0
or geometrically in the special case of a line shelling.

From now on, to the end of this lecture, all polytopes are simplicial.
For simplicial polytopes, the combinatorics of shellings can be described
even more concretely. For the following let P be a simplicial d-polytope, so
of dimension d - 1. Let
its boundary is a simplicial complex C :=
V := vert(P) be the vertex set of P of size n := fo(P)We identify the proper faces of P with their vertex sets that is, we
identify the "geometric simplicial complex C" with the "abstract simplicial
c(ap)
247
complex" on the finite set V. In particular, the facets of P are (d-1)simplices, and thus correspond to d-subsets of V. The complex C is pure
(d-1)-dimensional, so it is completely determined by the family of facets
C (V1
d ). All the other faces in C correspond to the subsets of facets in Y.
Now fix a shelling order F1, F2, . . . on the facets in T. We define the
restriction R of the face Fj as the set of all vertices y E Fj such that Fi \y
is contained in one of the earlier facets:
Ri
Fi : Fi\y
C Fi for some 1 < i < j}.
The main observation here is that when we build up C according to the

shelling, the new faces at the jth step are exactly the vertex sets G with
Rj c C C
In fact, a face G that is new is necessarily a subset of Fj : if it misses a vertex
y E R3, then it was already contained in a previous facet, by construction.
Finally, if G satisfies Rj C G C Fj but is not new, with G C Fi for some
< j, then by the definition of shellings G is contained in some F1 (1 < j)
such that Fi n Fi = FAw is a facet of Fi . From FM C F1 we get w E R3 ,
and from R3 C G C F1 n Fj = Fj \w we get w Rj: a contradiction.
Thus every shelling gives us a partition I W
/8 of the set of faces of
the simplicial complex into intervals of the form
:= {G : Ri C G c
A pure simplicial (d-1)-complex that has such a partition (with exactly

one part for each facet of C) is called partitionable, a concept developed
independently by Provan [413, App. 4], Stanley [476, p.149], and Garsia [90,
p. 6071. Shellable simplicial complexes are partitionable, as we have seen.
Our drawing tries to illustrate how the face poset of a partitionable complex
decomposes into intervals.
For a partitionable simplicial complex the f-vector can be read off from
the partition. Namely, if 1R31 = i, then there are exactly (dk :ii ) (k-1)-faces
248
contained in the part Ii , and thus
=
Let
hi
---
E
j= 1
IRA)
hi (C) denote the number of parts in the partition such that the
corresponding restriction set has size i:

hi (C)
:= I{i : IRA -= ti, 1 5 j 5 s}
The h-vector of a partitionable simplicial (d 1) complex C is this sequence

-
h(C) --= (ho, hi,

For example, the following graph (1-dimensional complex) C on 6 vertices
has f --= (1, 6, 7).
NZ
12
13
34 35
45
36 56
It is connected, hence shellable, a shelling order being given by the facet

ordering
12, 13, 34, 35, 45, 36, 56.
The bold edges in the face poset indicate the corresponding partition. Its
"minimal new faces" are
0, 3, 4, 5, 45, 6, 56,
and thus we get a contribution of "1" in this order to
ho , h1 , h1 , h l , h2, h1, h2,
and thus
h(C) =
(1,4,2).
All partitionable complexes have nonnegative h-vectors, since hi is the

number of minimal faces in the partition that have i vertices.
We have seen that shellable simplicial complexes are partitionable. The
converse is not true: here is a partitionable but nonshellable complex, with
8.3 h Vectors and Dehn Sommerville Equations

-
d= 2, n = 5,
f = (1,5,4) and h = (1,3, 0).
249
12 34
45 35
The f-vector of the shellable simplicial complex C can be computed from

the h-vector. In fact, adding up the contributions at the individual steps
of the shelling gives
d1
hk + (dk+l)hk_ i + + ( k i )h i + (d)ho.
However, the f-vector also determines the h-vector: from this formula we
can recursively compute hk from fk_ i together with (h0,... , hk _ i ). Here
is one way to do the bookkeeping. We consider the f polynomial
-
f (x) := fdi+ fd_2x + - +
fox' + fix d =
xd -i
and the h-polynomial

d
h(x) := lid + hd_ix
+ h 1 xd-1 + Itoxd
E hi
i=0
From the above derivation, we see that a shelling step with IRA = j contributes a summand of (z 1)d i to the f-polynomial. Thus, we get a
formula
d
f(x) =
E hi (x + 1)d i It(z + 1).
If we compare the coefficients of X dk in this formula, then we get the

above expression of fk_i in terms of the hi. However, f(x) = h(x + 1), so
we certainly also have
h(x) = f(x
1).
Now if we compare the coefficients of x dk on both sides of this equation,

we get a formula for hk in terms of the L. We take the result, and make it
into a definition.
250
Let C be a simplicial complex of dimension d-1. The
Definition 8.18.
h-vector of C is
h(C) = (ho , h i ,
, hd)
fd
iN
given by the formula
E")"
i=o
hk
that is,
d- i
i.o
fk-1 -
- k + 2)
ne -3
2
k 1)fk-2 +
i)ife
In particular, we have ho = 1, h1
hd
fd- 1
Id-2
fd- 3
ifo(kd
1.),k(kti).
- d, and
( -1 ) d i f
( - 1) d
Also, it is easy to verify ho h 1 +

+ hd fd- 1 , which for partitionable
complexes holds by construction this is just f(0) = h(1)!
The advantage of the definition of h-vectors in 8.18 is that it makes
sense even if the simplicial complex is not partitionable, and in the case of
a partitionable complex it shows that the numbers hi are independent of
the particular partition that we have chosen. Taking it as the definition,
we have proved the following theorem.
Theorem 8.19. Let C be a pure, d-dimensional simplicial complex. If C is

partitionable, then the h-vector is nonnegative. If C is even shellable, then
the entry hi counts the facets in a shelling whose restriction has size i, and
this number is independent of the particular shelling chosen.
Instead of explicit evaluation of the formulas, the h-vector can also be
computed by a difference table, a variant of Pascal's triangle, known as
Stanley's trick [479, p. 213] (see also [328, p. 5]). For this we write the
numbers L to the last entries of the rows of Pascal's triangle (to the place
where ordinarily we would put ( li ) = 1), and then compute the other
entries as
upper right neighbor - upper left neighbor.
251
Examples 8.20. We compute the h-vector for three different complexes.

(i) For the first graph considered above, with f-vector f = (1,6, 7), we
get a table
1
1
6
7
1
5
2
4
h=( 1
)
where the entries of the f-vector in the table appear in boldface.

(ii) Similarly, for the boundary of the octahedron
(1,6, 12, 8), and the table reads
we have f =
1
1
1
h=(
12
5
4
7
3
to test this, draw an octahedron, and compute the h-vector from

a shelling!
(iii) For a quite pathological example, consider the 5-dimensional simplicial complex on 12 vertices that consists of a single 5-simplex plus
6 isolated vertices. For this we have
f --= (1,6 + 6, (D, (), ( 64 ), (D, 1) = (1,12,15,20,15,6,1),

and Stanley's difference table yields the following:
1
12
1
1
1
15
11
10
20
16
15
9
6
1
6
8
15
22
7
7 23
23
37
30
60
60
30
6
1
1
1
= (
1
)
In particular, the h-vector has large negative components. This cannot happen for shellable complexes, but this one is not shellable: it
is not even pure, and it is disconnected.
Why do we study h-vectors? For various problems about simplicial polytopes, h-vectors are a much more convenient and concise way to encode the
information about the face numbers than f-vectors.
252
A first striking instance for this are the Dehn-Sommerville equations.

Their history starts with the observation that for simplicial polytopes we
get extra equations on the f-vectors from double-counting. In fact, for
simplicial 3-polytopes we see that every edge is in two facets, while every
facet has three edges. Thus we get 2fi = 3f2 from two ways of counting
the edge-facet incidences. Similarly, for simplicial d-polytopes we derive
2 fd- 2 =
dfd
l.
This is the only new equation for d < 4, but in higher dimensions there are
more (and more complicated) ones. Dehn [173] did the case d = 5, and the
general case was done by Sommerville [472]. Here is the version in terms of
the h-vector.
Theorem 8.21 (Dehn-Sommerville equations).
The h-vector of the boundary of a simplicial d-polytope satisfies
hk = hd_ k
for k = 0,1, . , d.
Proof. (by McMullen [3611) We use Lemma 8.10 and the observations in
its proof. Namely, if F1 ,
, Fs is a shelling, then its reverse Fs ,
, F1 is
a shelling as well. Furthermore, if F, comes earlier than Fi (that is, i < j)
in the first shelling, then it comes later in the reversed shelling. From this
we see that the restriction set for Fi in the reversed shelling is exactly
F3 \R: the complement of the restriction set for the original shelling. Thus
if Fi contributes "1" to hk in the original shelling (where k = IR3 I), then
it contributes "1" to hd_ k in the reversed shelling (where dk
Thus the value of hk computed by the original shelling is the same as the
value of hd_k computed by the reversed shelling.
However, by Theorem 8.19 the h-vector is independent of the shelling
LI
chosen to compute it, and hence hk = hd_k.
Example 8.22. If the vertices of the octahedron are numbered as in the
sketch,
8,3 h-Vectors and Dehn-Sommerville Equations
253
then a shelling order is given by

123, 126, 135, 156, 234, 246, 345, 456,
where the corresponding minimal new faces are
0,
6, 5, 56, 4, 46, 45, 456,

ho, h 1 , hi, h2, h1, h2, h2, ha,
and thus
h(C) = (1,3,3,1),
as we had computed in Example 8.20(H). Now we reverse the shelling, to
get the shelling order
456, 345, 246, 234, 156, 135, 126, 123,
where the corresponding minimal new faces are
VI, 3, 2, 23, 1, 13, 12, 123,

ho, h1, h1, h2, h1, h2, h2,
ha,
and from this we derive the same h-vector as before, of course.

There are various ways to write the Dehn-Sommerville equations in terms
of the f-vector, ranging from the "obvious" one,
k
E(-1)k ---t
E ( ir _k_i(d - i) Ji--i,
.
(I
)fi'
k
d- k
o
d-k
obtained by expanding the equation hk = hd _ k in terms of the face numbers

fi, to the perhaps most elegant and simple version,
d
fk -1 =
E(-1)d-i
i=k
( k)fi_i.
Still, hk = hd_k is hard to beat in its simplicity.

In all of these versions, the equation obtained for k = 0 is the EulerPoincar formula.
The Dehn-Sommerville equations for 0 < k < g are linearly independent conditions on the h-vector (this is obvious) and thus on the f-vector
(because f and h-vectors are linearly equivalent). For the proof that these
equations give a complete list of all linear relations, we refer to Griinbaum's
book [234, Sect. 9.2]. There you also find a direct proof of the various versions of the Dehn-Sommerville equations, by double-counting incidences
(generalizing our argument for 2fd-2 = dfd_ i above). However, Griinbaum
did not yet have shelling available as a technique, which yields the most
elegant argument.
-
254

"What is the maximal number of k-faces for a d-polytope with n vertices?"
The answer to this question is the upper bound theorem: "The cyclic polytope Cd(n) has the maximal number of k-faces for all k." This claim made
by Motzkin [385] in 1957 became known as the upper-bound conjecture.
During a long and involved history (see also Griinbaum [234]), including
premature announcements and many partial results, it was proved for polytopes with "few" vertices (i.e., n < d + 3, by Gale [2071) and for polytopes
with "many" vertices by Klee [296] (see the next section!), and in "low"
dimensions (for d < 8); see Griinbaum [234, p.175]. You may observe that
the result is quite easily derived from the Dehn-Sommerville equations for
d < 5.
Finally, in 1970 McMullen gave a complete proof of the upper-bound
conjecture since then it has been known as the upper bound theorem. McMullen's proof is amazingly simple and elegant, combining two
key tools: shellability and h-vectors. This section presents this proof of the
upper bound theorem, following McMullen's original paper [361]. Here is
the theorem.
Theorem 8.23 (Upper bound theorem). (McMullen [361])
If P is a d-polytope with n = k vertices, then for every k it has at most
as many k-faces as the corresponding cyclic polytope (cf. Example 0.6):
fk-i(P) fk-i(Cd(n))Here equality for some k with [4] <k < d implies that P is neighborly.
The first fact to note is that we can restrict our attention to simplicial
polytopes.
Lemma 8.24. (Klee [296] and McMullen [359]) The vertices of a dpolytope P can be perturbed in such a way that the resulting polytope P`
(with the same number of vertices) is simplicial, and
fk-i(P)
fk-i(P)
for 0 < k < d. Here equality for some k >
Ad
can occur only if P is
simplicial.
This is not the hard part, so we avoid the distraction of a proof. Thus from
now on we only consider simplicial d-polytopes: this is essential, because it
allows us to use the Dehn-Sommerville equations! What do they get us?
First, we note that we always have
f k - 1 < (n
k)
with equality if and only if P is k-neighborly. This bound is achieved with
equality for k < Lti in the case of neighborly polytopes like the cyclic
8.4 The Upper Bound Theorem 255
polytopes of Example 0.6. For k > [4 .1 this bound cannot be achieved,

except in the case of a simplex, by Exercise 0.10.
However, the face numbers fi,
- fw-1 already determine the
complete f-vector. Namely, they determine h o , , g by Definition 8.18,
and the Dehn-Sommerville equations give us the rest of the h-vector. In particular, all neighborly simplicial d-polytopes with n vertices have the same
f-vector as the cyclic polytope Cd(n). In this context recall that for odd
d > 3, neighborly polytopes need not be simplicial, however, by Lemma 8.24
the nonsimplicial ones have smaller fk _ i for k > 14j, and thus we don't
worry about them here.
Let's look at the expression of the f-vector in terms of Ito , , h . To
get the prettiest possible formulas, we will use the notation
d
To +
To
i=0
+ - - - + 714 1
pl c
if d is odd,
if d is even.
That is, the asterisk means that we take only half of the last term for i -=
if d is even, and take the whole last term for i =
=c-Y if d is odd.
Similarly, we will use
to denote a sum where only half of the first term
is taken if the starting index of the summation is integral.
For k > j, we have with this notation
E.
fk-1
E (d
i=o
i)
- (d i\
k i hi
i=0
i hi
di
i.0
(where the terms vanish for i > k)
_ i)
(8.25)
(k d +
where for the last equality we have substituted d i for i, and used the
Dehn-Sommerville equations.
Looking at that, we see that what we "really have to prove" is that for
k<
the neighborly simplicial polytopes not only maximize fk_ i (as we
know), but they also maximize hk. That is, the following lemma is "more
than enough."
Lemma 8.26.
[361, Lemma 2] Let P be a simplicial d-polytope on
fo = n vertices. Then for 0 < k < d,
hk(P)
(n d
1 +k ).
Equality holds for all k with 0 < k < t if and only if t < Lc4 j and P is
1-neighborly.
256
The statement and proof of this lemma are the key steps in McMullen's
(d) ifor
solution of the upper-bound conjecture [361] (with the notation gk_
what we now call hk)-
Proof. The proof is done by induction on k. The lemma is clearly true for
k = 0, since we have defined ho to be 1. Thus it suffices to verify
(nd+k\
k k+1
nd-1+k \ 7
k
that is,
(k + 1)hk+1 < (n
d + k)hk
(8.27)
for k > O.
We get this by putting together two parts. The first one is the formula
ha/v)
(k + 1)hki (C) + (d k)hk(C),
(8.27a)
vEvert(C)
where we use Cfy as a convenient abbreviation for the link of y in the

simplicial complex C, that is, C/v := link(v,C).
Equation 8.27a is easy to prove, because it is valid also during a shelling,
when instead of C = C(P) we consider Ci C(Fi U- -UFi ), and because a
shelling on ap also induces a shelling order for all the links, by Lemma 8.7.
The formula is valid at the beginning, for the empty complex Co (when no
facet is present, and all terms vanish). Now assume a new facet Fi is added,
and consider its contribution to
hail)). Clearly this only affects the
terms for vertices y C F3 . There are two cases.
If y itR then there is a new face of size IRA in the link of v. This
affects hk(C/v) only if IRi I = k, and in this case we get a contribution
of "1" to IFARi I = d k different summands. (The left drawing in our
sketch indicates this case, for k = 1 and d = 3.) In this case, we also get
that hk(C) increases by 1, and thus the right-hand side increases by d k,
as it should.
Et,
257
If y E Ri , then we get a minimal new face of size IRj I-1 in the link of v.
So we get a contribution to hk(C/v) only if IRA = k+1, and in this case
we get a contribution of "1" to k+1 different terms on the left-hand side
of the equation. At the same time, we get that hk +1(C) increases by 1, so
the right-hand side increases by k+1, and we are even. (The right drawing
in our sketch depicts this case.)
This proves equation (8.27a).
The second part we need is an inequality,
hk (C/v)
5_
n hk (C).
(8.27b)
vEvert(C)
For this, we prove that hk(C/y) < hk(C) holds for all n vertices y E vert(C).
To see this, take a shelling that shells the star of y first. This means that
the minimal new face in C and in the link Civ coincide at every step while
we are shelling the star. Later, after the shelling has left the star, we may
get new contributions to hk(C), but not any more to hk(C/v). With this we
get the inequality (8.27b), and putting it together with equation (8.27a)
we derive the inequality (8.27).
What about the equality case? To get hk(C/v) = hk(C), it is necessary
that in a shelling that starts with the star of y, there is no "new" face of
size at most k outside the star of v. Thus we get that, for 1 > 1, the equality
hk(C/v) = hk(C) holds for all k <1 if and only if in a shelling that starts
with the star of y, there is no minimal new face of size at most 1 outside
the star of y. Equivalently, this says that every face G with at most IGI < 1
vertices is contained in the star of y, so that Cu {v} is a face, too. Equality
in (8.27b) holds only if we have equality for all vertices y. From this we get
that equality in (8.27b), and thus in (8.27), holds for all k < 1 if and only
if C is (1-1-1)-neighborly. 0
On the way, we have also computed the f-vector of the neighborly polytopes: for this we only have to put the equality case of Lemma 8.26 into
the formula 8.25.
Corollary 8.28.
vertices, then
If P is a simplicial neighborly d-polytope with fo = n

d
di
ndl+i
i
)
i=aqk i) + (lc id + i)) (
for 0 < k < d. For every k this gives the maximal number of (k-1)-faces
for a d-polytope with n vertices.
For k = d, this reduces to a formula for the number of facets of the cyclic
polytope Cd (n):
fk-1 =
.1- 2 (n d 1 + i) .
E
i
i--=o
d
258
8. Shellability and the Upper Bound Theorem
(Compare this to Exercise 0.9!) Note that in fixed dimension, fd_ i (Cd (n))
grows like a polynomial of degree VD in the number of vertices.
Here is a brief asymptotic argument, due to Seidel [454] (see also Mulmuley [387]), for this corollary to the upper bound theorem. Namely, consider
any shelling of OP. For every facet we get that either the restriction Ri or
its complement Fi \Ri has size at most
So, either in the shelling or in
its reverse we have that Fi has a restriction of size at most Ilk and the
restriction sets in a shelling are distinct by construction. Thus the number
From
of facets is at most twice the number of k-faces of P with k <
this we get
[g].
Lc.
and this rough estimate bounds fd_ i by a polynomial of degree
Lti
in n.
Some Extremal Set Theory
8.5
We have used already that the simplicial complex C with n vertices can be
identified with a set system, the collection of subsets S(C) of an n-set,
S(C) := {vert(G) : G E L(C)}.
For the following we identify the vertex set of C with the set
[n] := {1, 2, . . .
and the k-faces of the complex with the (k+1)-subsets of the ground set [n],
for -1 < k < dim(C). Thus, if C is a pure (d-1)-dimensional simplicial
complex, then it is determined by its family of facets, which is a subset
of
the collection of d-subsets of [n].
The construction behind this identifies geometric simplicial complexes,
as we get them for example as boundary complexes of simplicial polytopes,
with abstract simplicial complexes, where we only retain the information
on the vertex set, and the information about "which vertex sets correspond
to faces, respectively facets." This approach is useful for all problems that
are not concerned with the geometry of a complex, but only with its combinatorial structure. The combinatorial structure of a complex, however, is
faithfully represented by the abstract set system: from the set system data,
it is easy to reconstruct the simplicial complex (this is a process known as
geometric realization of the abstract simplicial complex).
On the next page is a "generic picture" of a simplicial complex, viewed as
a set system. (Of course, this need not be your way of viewing set systems
supply your own sketch!) The left version just shows you the "shape"
of a simplicial complex within the lattice of all subsets of [n], while the
n
ci l ),
259
right side shows you the bipartite graph of all k-faces and (k-1)-faces,
that is, of all subsets of size k+1 and k within the complex. In both cases
the minimal element is 0, which is always supposed to be contained in the
complex, corresponding to fi = 1.
Lemma 8.29.
(Sperner [473]) If C is a simplicial complex of dimension
d on fo <n vertices, then for 1 < k < d one has
n-k
k 1)
k+1
with equality if and only if C is k-neighborly, such that fk = (k:1) and
Proof. We double-count the edges in the bipartite graph above. Every

(k+1)-set contains k+1 different k-sets, thus there are (k+l)fk edges. Similarly, every k-set in the complex is contained in at most n - k different
(k+1)-sets of the complex, so there are not more than (n-k)fk_ i edges,
and we get the inequality (k+1)fk < (n- k) f k _ 1 .
If we have equality, then with every k-set the complex contains all the
(k+1)-sets that contain it. However, since we can get from every (k+1)set to every other one by throwing out one element, adding a new one,
throwing one out, and so on, this means that in the equality case the
complex contains all the (k+1)-sets, or none.
For k = 1, this is trivial. For k = 2, it just says that on n vertices there
cannot be more than ( 2) edges. For higher k it gets more interesting. Here
is one thing we can easily derive from it.
Lemma 8.30. Let C be a simplicial complex of dimension d - 1, with
n = fo (C) vertices. If n > dk - (k - 1) 2 , then the h-vector of C satisfies
h,.
< (n - d - 1 + k)
with equality if and only if C is k-neighborly, that is, if fk-i = (z).
260

i) i (d- ik+t) fk i
in pairs, as
Proof. We group the terms of hk
hk
-=
(d-k+ 0).4_1 (d-k+1\ ,

J k-2
(cl-k+ 2
(d- k + 3)
J k-3
Each of the terms is of the form ( k-j- II)
3
_ t -i\ f.d
lk j1 J 3 -1
-
k-4
J
for 1 < j <k - 1.
(If k is odd we'll have an extra term of (kd )f__1 at the end, but that's a
constant anyway.) By Sperner's lemma, we can bound this by
d - - 1
(d- j.-1\f3 J f
Vc -3 - 1)
1 f_ k-jj
(d- j. -1\ f3 in-j
- 3 - 1)
lj+1 k-jj
-1
(n.') for all j, provided that we are sure

3
This is maximized exactly if
' =2- - k-- ) are nonnegative. From this we get that hk

that the differences 1
is maximized by the simplicial complexes that are k-neighborly, if we have
d-j
for j
1, ... k - 1.
Since this lower bound is monotone in j, we only need to consider this for
j = k -1, and get the condition n > (k -1) + k(d - k + 1), which we had
required to hold.
LI
Note that this lemma is false without the assumption that n is large
enough: for the complex in Example 8.20(iii) we have d = 6, fa -= n -= 12,
and
(8) (n d 1 + 3)
h3 = 60 > 56 =
3)
From this elementary lemma we get McMullen's Lemma 8.26, and thus
a proof of the upper bound theorem, for polytopes with a sufficiently large
number of vertices. This simple proof not only works for polytopes: the
argument equally applies to all kinds of simplicial complexes that satisfy
the Dehn-Sommerville equations. This includes all spherical polytopes (corresponding to simplicial fans, see Kleinschmidt & Smilansky [309]), and
even more generally, all simplicial Eulerian pseudomanifolds (see Klee [295],
Bayer & Billera [58], Chan, Jungreis & Stong [139], and Stanley [480,
Sect. 3.14]).
Corollary 8.31 (The upper bound theorem for complexes with
many vertices).
Let C be a (d-1)-dimensional simplicial complex that
satisfies the Dehn-Sommerville equations hk = hd-k-
8.5 Some Extrema! Set Theory
261
If the number n = fo of vertices satisfies

d
d
n > cl[] ([ ] 1 )2 ,
then for 0 < k < d the complex cannot have more k-faces than the boundary of the cyclic polytope Cd(n):
fk(e) 5_ fk(Cd(n)).
k
Proof. We get hk < ( n-d-1+10
) = hk(Cd(n)) for k < Lli from Lemma
8.30, and for k > Lg-j from the Dehn-Sommerville equations. The rest
follows from the fact that the fks are positive combinations of the his with
0
i < k + I.
In this proof, you can see some of the power of the translation of geometric simplicial complexes into finite set systems (abstract simplicial
complexes). In this setting, extremal problems about simplicial complexes
are a principal topic of "extremal set theory." We will review in the rest of
this section some basic concepts, constructions, and results from this field
some of them without proofs, to save time and space. You might want
to look at the wonderful survey by Greene & Kleitman [2231 for some of
the missing details.
A basic tool of extremal set theory is the use of various partial and
linear orderings on the k-subsets of an n-set (i.e., on the (k-1)-faces of a
complex). Since we assume that the vertex set is [n], that is, the vertices are
labeled 1, 2, ..., there is a natural linear ordering ("well-ordering") on the
vertex set. With this, we can in particular talk about the largest element
rnax(G), if G is nonempty.
Using this, we define the r-lex order (or reverse lexicographic ordering)
on the k-subsets of vertices. For this we write
G - H
if and only if G H and the largest element in which G and H differ is
in H, that is, if
max(G\H) < max(H\G).
Equivalently, this means that either max(G) < max(H), or max(H) =
max(G) =: p and G\p - H\p.
In the definition of the r-lex order, the number of elements n is not
specified. Thus, we can take "-<" as a linear order on the set of all the ksubsets of N. Furthermore, for every k-subset G c N, there is only a finite
number of k-subsets of N that are smaller than G, because G >-- H implies
that H C (ink]) for n :---,- max(G). This means that we can use the r-lex order
to enumerate all the k-subsets of N, as Fi (k), F2 (k),. . .. So, we define Fi (k)
to be the jth subset in this increasing listing according to r-lex order.
262
For example, the r-lex order on the 3-subsets of N begins
123 - 124 - 134 - 234
125 - 135 - 235 -
145 -
and with the above notation, this list shows
Fi (3) - F2(3)
F3(3)
F4(3) - F5(3)
F6(3) - F7(3)
F8(3)
For n, k > 0, there is a unique binomial expansion of n of the form
(ak)
with ak > ak_i > - > a2 > a i > 0.
In fact, existence and uniqueness of this expansion are easy to verify, by
choosing ak first, ak_i after that, and so on. A more systematic explanation
may be the following. Define the integers ak > ak_i > - - - > a2 > al > 0
by setting
Fn+i (k)
=: {a i + 1, a2 + 1,...,ak_i + 1,ak +1} <.
(Here the subscript "<" indicates that the elements are listed in increasing
order.) Then there are exactly n different k-subsets GcN that are smaller
than fal +1, , a,+1} in r-lex order. Namely,
of them have a maximal
element smaller than ak + 1;
have maximal element ak -1- 1 but the
next smallest element smaller than ak_ i + 1; and so on.
One more thing is easy to see: the (k-1)--subsets of N that are contained
in some Fi (k) with j < n +1 also have maximal element smaller than ak +1,
or they have maximal element ak + 1 but the next element is smaller than
ak_i 1, etc. so there are exactly
(a
k
i
(a
kk -11 )
ak
k 1
()
k 2
(a2)
1
(al)
(k 1)-subsets contained in the k-sets Fi(k),

, F +i (k).
For an example, let k = 3 and n = 7. We expand
(7)
and from this we see that F8(3) = {5,4,1}, consistent with our listing
above. There are 7 smaller sets in r-lex order, where 4 = (34 ) have largest
element smaller than 5; 3 = (32 ) have largest element 5 but the next element
smaller than 4; and 0 = CD have the two largest elements 5 and 4 but the
smallest element smaller than 1 (impossible). Also, there are
(93(8)
()
263
2-subsets contained in the first eight 3-sets, namely 6 = (42 ) with largest
element smaller than 5; 3 = ( I ) with largest element 5 but the next one
smaller than 4; and 1 = (g) with the elements 5 and 4. Note that this last
one is contained in F8(3), but not in a smaller 3-set.
The r-lex order is very natural in various respects. For example, it yields
a shelling order for the (k-1)-skeleton of the simplex Ad 1 for k < n = d+ 1
(Exercise 8.24(0). In fact, many other linear orders work as well. However,
it is a fascinating open problem whether skeleta of simplices are extendably
shellable; see Problem 8.24(iii)*.
Perhaps the most basic result of extremal set theory, and a principal
application of r-lex order, is the characterization of the f-vectors of simplicial complexes. It is known as the Kruskal-Katona theorem [316] [291],
although Schiitzenberger [449 ] was earlier, and even before this Harper got
close: his paper [249] does not explicitly state the theorem, but the result
is easy to derive, and I was told that Harper was aware of it at the time.
Theorem 8.32 (Krusical-Katona theorem).
Let f = (fi, fo, fi , - - - , fdi) e 14+1 be a sequence of nonnegative integers. Then the following conditions are equivalent.
(0 The sequence f is the f-vector of a simplicial complex of dimension

at most d-1.
(ii) The family .9f) := {Fi(k) : 0 < k < d, 1 < j _< fk _ 1 } is a simplicial
complex (that is, with every set it contains all subsets).
(iii) f_ i = 1, and fk_ i > ak (fk ) for 0 < k < d-1.
Proof. The implication (ii) >.(i) is trivial, and the equivalence (ii) 4-4 (iii)
is clear with our construction of the "boundary operator" ak(n) above.
The remaining nontrivial part is (i) >-(ii): see Greene & Kleitman [223,
ii
p. 73] for a nice and simple proof by "compression."
A simplicial complex (on a vertex set V C N) is compressed if its k-faces
form an initial segment with respect to r-lex order, for all k, that is, if it is
a complex as given by Theorem 8.32(4
The compression technique mentioned for the last proof takes as an input a simplicial complex, and outputs a compressed simplicial complex
with the same f-vector. The technique stems from a paper by Lindstriim
& Zetterstriim [336]. It works quite the same way for multicomplexes (see
Macaulay's theorem 8.34 below), and also for a generalization of both theorems, due to Clements & Lindstr6m [153} [18, Sect. 9.1].
What we really need for the following is not this theorem for simplicial
complexes, but a version for "multicomplexes." For this, we introduce some
new terminology I guess you've seen some of this before, but perhaps
with different names.
A multiset is a finite sequence of elements that may contain repeated
elements. The order of the elements is irrelevant, but their multiplicities
264
OSA
are part of the structure. So, a multiset F with elements in N can be

uniquely written in the form
fbi,b2,
where the subscript "<" indicates that we have arranged the elements in
weakly increasing order, bi < b2 < ... < bk. The size of a multiset is the
number of elements, counting multiplicities. So the multiset IT' above has
size IFI = k, and we would call it a k-multiset. Also, a submultiset G C F
is a multiset in which every element has smaller or equal multiplicity than
in F. Finally, a multicomplex is a finite collection of multisets that is closed
under taking submultisets.
Multicomplexes can be interpreted in a variety of different ways (see
Exercise 8.22). For example, they are equivalent to order ideals in Ng and
to systems of monomials that are closed under taking factors. Our sketch
shows the "generic" drawing of what a multicomplex might look like. Note
the small diamond shape at the bottom, which denotes all the sets in the
multiset system.
One can attempt a (quite technical) topological interpretation of multicomplexes, leading to the extensive apparatus of semisimplicial sets [358]
which we avoid. There are only a few pieces of topological terminology
we use. So, the dimension of a multiset is defined to be one less than its
size, dim(F) := 1 1.11 1; the dimension of a multicomplex is the greatest
dimension of a multiset it contains; and the f -vector of a multicomplex is
U-47 fo, fi, , f4, where fi is the number of multisets of dimension un
the multicomplex.
Here is a basic bijection, which takes the k-rnultisets with elements
from [n] to the k-subsets of [n+k-11:
4)
{b1 7
bk } <
1--
ibi,b2 +1, ,bk + k 1 1<-
In particular, 4) proves the basic identity

((n
k))
(n + k 1) ,
(8.33)
8 5 Some Extrema! Set Theory
265
where the symbol on the left side denotes the number of k-multisets with
elements from [n] the multiset analogue of the binomial coefficient (z).
See Exercise 8.22 for three other proofs of this.
Many set concepts are easily generalized to multiset concepts, if we just
replace binomial coefficients by their multiset counterparts*. In particular,
we need the r lex order on k-multisets. For this we write
-
P - 6
if max(P) < max(d), or if max(P) = max(d) =: p and F\73 \p, where
"P\p" means that we remove exactly one copy of the largest element from
F. So, r-lex is a linear order on the set of all k-multisubsets of N. All the
nice properties of r-lex order on sets generalize to multisets. The reason
is that under the bijection 8.33, r-lex order on k-multisets is equivalent to
r-lex order on k-sets,
-
q5 ( P
--
Thus, for every k-multisubset, there is only a finite number of smaller ones,
and thus we can use r-lex order to enumerate and label the k-multisubsets
of N, as F1(k),F2(k), .... Thus we define F3 (k) to be the jth multiset in
the listing according to r-lex order, and find that in fact it is the 0-image
of the jth subset:
15(Pi(k)) = Fi(k)For example, the r-lex order on the 3-multisubsets of N begins
P1(3) - P2( 3) .-< P3( 3) -- P4( 3) - P5( 3) - P6( 3) - P7( 3) -- PO) - - - - ,
that is,
111 - 112
122
222
113 - 123 -- 223 - 133 --
Now, for n, k > 0 there is a unique expansion of n of the form

n -=
22)) ((b11))
((bkk)) ((kbk--11 )) ... r
with bk > bk_i > ... > b2 > bi > 0
we get this the same way as before, by defining
Pn-Fl(k) =:
{b1 + 1, b2 + 1, ... , bk _ i + 1, bk +1}<,
or by setting bi := ai ti + 1 in the expansion we had before.

*Check the following carefully! Do not take my word for it! Go ahead! And don't ask
Helga!
266
8.
Sheliability and the Upper Bound Theorem
There are exactly n different k-multisets that are smaller in r-lex order
than fbi + 1, ... , bk + 11. Namely, for ea of them the maximal element
is smaller than bk 1; for ekk -3 the maximal element is bk + 1 but the
next element is smaller than bk_i + 1; and so on.
One more thing is easy to see:* the (k-1)-multisets that are contained in
some Fi (n) with j <n 1 also have maximal element smaller than bk 1,
or they have maximal element bk +1 but the next smallest element smaller
than bk_i -I- 1, and so forth so there are exactly
k
(Gb 1)) (Gbk--- 12
))
... r21)) + C))
(bk + k - 2) + (bk_i + k - 3) + ... i (b2\ i (bi - 1\

1- 1 ) 1k - 1 j
I\ k - 2 )
0 )
(ak_ i
1)
(a2 - 1)
(a l - 1 \
+ ... +
+
k -1 +
k 2
1
0 )
-
of them.
For an example, again let k = 3 and n = 7. We can expand
7
((g))
0+M,
and from this we see that F8(3) = {1, 3, 3}, as in the listing above. There
are 7 smaller 3-multisets in r-lex order, where (() = 4 have largest element
smaller than 3 1 22))
( = 3 have largest element 3 but the next element smaller
than 3, and M = 0 have the two largest elements equal to 3 but the
smallest element smaller than 1. Also, there are
493 (8)
= (g) + ((i )) +
((c)) = g + () + ( -01) =
3+2+1 = 6
2-multisets contained in the first eight 3-sets, namely 3 = (( 2)) with largest
element smaller than 3, 2 = ((i )) with largest element 3 but the next one
smaller than 3, and 1----- ( -01) with the elements 3 and 3. This last one is
contained in F8(3), but not in a smaller 3-set.
Our main reason of doing multisets and their r-lex ordering is to get
some intuition for what multicomplexes are, how they behave to be able
to make sense out of the following theorem. It uses a "relative" (1)d of the
0-map 8.33, which takes a naultiset {bk, ... , b 1 }>, adds 1 to each of the
elements, adjoins d-k zeroes to the multiset, and then applies the 0-map
to get a set:
'Ild({bk, ,b11>)
dk
Offbk+1, .. ,b1+1,0,0, .. ;OW

-= Ibk-F1-Ed,bk-i-F1-1-d - 1 , - ,b1-1-1-1-d-k+1, d-lc, 7 2, 1 }>
=
*Do you get a dj-vu feeling? Of course, what we are doing here for multisets is
exactly the same as we did for sets before!
8.5 Some Extrema! Set Theory
267
Theorem 8.34 (Macaulay's theorem).

Let h = (h0 , h1 ,... , hd) E Ng-Fi be a sequence of nonnegative integers.
Then the following are equivalent.
(i) The sequence h is the f -vector of a multicomplex.
(ii) The sequence h is the f -vector of a compressed multicomplex, that
is, F := {F3 (k) : 0 < k < d, 1 <j < hk} is a multicomplex.
(iii) 110 = 1, and hk_i > /Y (h) for 1 <k < d.

(iv) The sequence h is the h-vector of a shellable simplicial complex of
dimension d-1.
(y) The family {(1)d(i 3 (k)) : 0 < k < d, 1 < j < hk } is the set of facets
of a shellable simplicial complex with h-vector h.
Proof. Again part (ii) >(i) is trivial, while (ii)-4=P.(iii) follows from our
previous discussion.
The part (i)(ii), from multicomplexes to compressed multicomplexes,
is originally due to Macaulay. It can be proved by the "compression" technique that we have mentioned in the proof of the Kruskal-Katona Theorem 8.32.
For (ii)(v), from multicomplexes to shellable complexes, this is the
special case "s = 1" of a construction in Bjeirner, Frank! & Stanley [89],
which takes a multicomplex and produces a pure complex from it:
3 c1
-4
C tbd(P) for some 1' E
If is the compressed multicomplex from (ii), then the pure complex i'd(- - -)"
is shellable. In fact, in this case r-lex order defines a shelling, and the restriction set is R(4)d(Fi(k))) = 4)d(Fi (k))\{dk,... , 2, 1 } , of cardinality k.
Thus every k-multiset in the multicomplex contributes "1" to hk in the hvector of the corresponding simplicial complex. For the proof with details
we refer to [89, Sect. 5].
The implication (v)(iv) is trivial, thus we are left with proving the
direction (iv)(i), from shellable complexes to multicomplexes. For this,
Stanley [475, 476] has given an algebraic argument: the multicomplex arises
in this case from a monomial basis for "the Stanley-Reisner ring modulo
a system of parameters." Is there a simple combinatorial argument? Note
that this innocent-looking claim in particular implies that
hk < ((h i )) = (hi + k _1)
for shellable complexes, and thus this reproves the upper bound theorem
(McMullen's Lemma 8.26)! In fact, this is the key to Stanley's proof of the
upper-bound conjecture for spheres [474] [478, Sect. 11.3].
268
Macaulay's contribution [345] was, essentially, the equivalence (i)< >(ii)

of Theorem 8.34. We have combined it with important work by Stanley
[475, 476] and by Bjeirner, Frankl & Stanley [89].
The sequences characterized in Theorem 8.34 are called M-sequences
(dial "M" for "Macaulay") or "0-sequences" ("0" for whatever). They
are of fundamental importance, as we will also see in the next section.
Here are a few examples of M-sequences, for d , 3. The sequence
(1, 3, 3, h3) is an M-sequence for 0 < h3 < 4. In fact, part (ii) of Macaulay's
theorem suggests that for this we can take the multicomplex
{ 0 , 1,2,3 , 11,12,22 }
together with the first h3 sets from the list
F1 (3) = 111, F2(3) = 112, F3 (3) = 122, F4 (3) = 222.

Since F5 (3) = 113 contains the subinultiset 13 that is not among the 1faces we listed, we get that (1, 3,3, 5) is no longer an M-sequence. Note
that among these, for h3 = 1 we get (1,3, 3,1) as an M-sequence: this is
the h-vector of the boundary of an octahedron; see Example 8.22.
8.6
The g-Theorem and Its Consequences
From the last section, I hope we gathered some intuition for "what an
M-sequence is." All kinds of interpretations are useful: so, the best is to
alternate between various explanations, between
the f-vector of a multicomplex,
the h-vector of a shellable complex, and
a sequence of nonnegative integers satisfying ak(hk
Here comes one big reason why M-sequences are useful. It yields a complete characterization of the f-vectors of simplicial d-polytopes P. What
do we know about them so far? Forgetting about Id = 1, we know they can
be encoded by their h-vectors h(P) = (ho, h 1 , ... , hd), which satisfy the
Dehn-Sommerville equations hk = hd_ k for 0 < k < d. Also, we know that
h(P) is an M-sequence from Macaulay's Theorem 8.34(iv), which implies
the upper bound inequality
hk
<
(h i +kk 1) .
In quite a daring step, McMullen in 1970 combined all the then available
information (including the lower bound theorem by Barnette, see below)
8 6 The g-Theorem and Its Consequences
269
into a conjectured complete characterization [363]. It became known as the

g-conjecture, because it referred to the g-vector of the polytope, defined as
g(P) :=
with go := ho = 1, and gk := hk
go , gil . - - )gL4J)
hk_ i for 1 < k <
Lgi-
(Billera & Lee [70, 71] and Stanley [474])
The g-Theorem 8.35.
A sequence g = (go, gi , . . . , g piii ) E
NO-1+1 is the g-vector of a simplicial
d-polytope if and only if it is an M-sequence.

We will not even attempt to prove this (see the notes below): we'll be
content with deriving some of its most striking consequences. For this,
we use a matrix formulation of the "McMullen correspondence," due to
Bj6rner [80, 82, 86].
In the following, we use the convention hd+1 := 0 and again define gk :=
hk hk_ i for 0 < k < d+1. With this, we find
for 1 < k < d+1.

These "Dehn-Sommerville equations for the g-vector" explain why we restrict our attention to gk for k < LC: we reconstruct gk --=
k > (d +1) Ai. Careful: there might be one more term in thegd
sequence,
nco
f er,
namely gk for k = 111+1 = d LC in the case when d is odd. However, we
can ignore this case since this gk vanishes.
With this, we express the f-vector in terms of the g-vector as follows:
fk-1
d
E
(dk
: i) hi
i
i=o \
=
2.0
d+1 d
=
EE
(d i
j=0 i=j
d+1
(d+1j\
= V
z-d cl+1k) g3i=e)

1-fl
2_,
.i=-0
d+1k)
(d + il k)) .
We take this result and interpret it as a matrix correspondence. For this,

we define a coefficient matrix
Md =( Mjk) :=
jk
( (d + 1 j ) (j
Vl +1 k)
VI +1 k))o<i<LC,
and use it to restate the g-theorem, as follows.
0<k<d
270
Theorem 8.36 (The "McMullen correspondence").
[86]
g 1--* g- Md
is a bijection between the M-sequences
g E NP
I +1 with gi -= nd-1, and
the f-vectors Nri of simplicial d-polytopes with n ----- g1 +d+1 vertices.

For example, we compute
M1 = (1 2),
( 1 4 6 4
M3=0
0 1 3 2
M4 =
)
10
"
1 i )7
5
1
0
(
I 0
0
10
4
1
10
6
2
5
3
1
From this we get trivialities for d < 2. For d ,---- 3 we get the f-vectors of
simplicial 3-polytopes, which are also easy to get by elementary arguments
(Exercise 8.28). However, starting at d = 4, we get nontrivial characterizations: so the f-vectors of simplicial 4-polytopes are all the row vectors f of
the following form:
f(P4)
= (1,5,10,10,5) -I- gi(0, 1, 4, 6, 3) + g2(0, 0, 1,2,1)
with gi, g2 _?_. 0, 492 (g2) ,...giThe matrices Md are given explicitly and are not hard to analyze, which
allows us to study the f-vectors of simplicial d-polytopes. In particular,
one can easily (using well-known recursions, monotonicity properties, and
so forth, of binomial coefficients) verify the following simple properties.
The entries of the matrix Md are nonnegative integers,

with zeroes below the diagonal (m3 k ,----- 0 for j > k), ones on the diagonal
(m33 ,----- 1 for all j), and larger values above the diagonal (rnik > 1 for
j < k, except for m
1 in the case when d is even).
Lemma 8.37.
Instead of a proof, here is the computation of Md for d = 7: we get a

(4 x 7)-matrix, M7 =
/ (:)- (Li) M - G) (:) -CD (: )--M (1)- C4) (1) - M (D - (D (7 )- (?) ]

(41) ( 3
7 ) (31) (27) (21) (D G)
(86) __ (28 ) (67) (27) (66) A82) (65) _ (25) (64) _(42) (1) (D (62) _ (22) (61)(21)
(87) ( 81 ) ( 7 ) (71 ) (67) (61) (57) ( 51 ) (47)
\ (:) (D G) M (:) a) (:) C) C51) CO C) (D (D (D CO ()
-,
( 1
0
0
0
0
0
0
0
8 0
1 0
0 0
0 0
28-0
70
10
00
56-0
21-0
60
1 0
70-0
35-0
15-0
56-0
35-0
20-0
28-0
21-0
15-1
50
10-1
10-3
80
71
6 2
53
8 6 The g-Theorem and Its Consequences
271
and this means that the f-vectors of simplicial 7-polytopes are exactly the
vectors of the form
(go
1 8 28 56 70 56 28 8
0 1 7 21 35 35 21 6
0 0 1 6 15 20 14 4
000 1 5972
gl 92 93)
for an M-sequence (go , 91 , g2 , g3 ) E N.

From the McMullen correspondence, one gets the upper bound theorem
as an immediate consequence, but also the lower bound theorem, which
was first proved by Barnette in 1970 [38, 401.
Corollary 8.38 (Upper and lower bound theorem).
We consider sirnplicial d-polytopes P of fixed dimension d and fixed number
of vertices n = gi + d + 1.
(UBT) The f-vector f (P) = gMd has its componentwise maximum if and
only if all the components of g are maximal, with
9k = ((91))
k
(n - d 2)
( Y i + k -1 )
1
Also, fk-i is maximal if and only if g i is maximal for all i with

i < minfk,
(LBT) The f -vector f(P) = gMd takes its componentwise minimum if
and only if all the components of g are minimal, that is, if gi = 0 for
i > 1.
Also, fk_i is minimal if and only if gi = 0 for 2 <j < min{ k,

An analysis of the matrices Md can also be applied to the unimodality
conjecture for convex polytopes: the question of whether for every polytope
it is true that the f-vector satisfies
1 = f_ i < fo <
<
fd .1,
fp fp+1?
for some p, that is, the f-vector has to be unimodal. It seems that this
question was first asked by Motzkin in the late 1950s; see [801.
For this, it is not hard (but a little tedious, perhaps) to check that the
rows of Md are unimodal: they first increase, until they reach a maximum, and then they decrease again. Furthermore, the maximum occurs in
columns with indices j between j = 1. 4j and j = 3
This type of
analysis yields the crucial part of the following theorem.
u-i-A.
Theorem 8.39.
(Byirner [80, 86])

The f- vectorsof simplicial d-polytopes with d > 3 satisfy
f-i < fo < fi <...<fL41-1 <f
and
43(v) > - > fd-i.
272
Lc
The bounds
and [(31)j are best possible in the sense that for every
p and d with LI] fp+i > - - - > fd-i-
Thus, the "shape" of the face lattice of simplicial convex polytopes looks
roughly as follows (taking into account also that they are "top heavy"; see
Exercise 8.34):
k=d
k = 3d/4
,*
k =--- d/2
k=0
Eiji5rner's theorem implies the unimodality conjecture for simplicial polytopes of dimension d < 10. With more work, one can get it up to dimension 15 (Bj8rner [80, 86]), and even to dimension 19 (Eckhoff [178]). Surprisingly enough, the unimodality conjecture for simplicial polytopes is false
in dimension 20, as was first discovered by Bjbrner [80] and Lee [324, 70].
Examples 8.40. The unimodality conjecture fails for a simplicial polytope
of dimension d --.-- 20 with the following f-vector, for which fii > fi2 <113.
1
f-i =
4203045807626
fo =
,-84060916163336
fi
798578704207074
4791472253296106
h = 20363758019368323
=
65164051780016980
15
= 162910744316489788
/6
= 325834059588060117
17
= 529707205213463823
18
19
= 709935971390166248
805494832051588614
ho
Ili
= 821976324224631043
/2 =
,_-13
112 --- 821976324224611712

113 = 822000129478641948
747383755288236256
114
115
116
117
118
119
=
=
,
=
=
546761228419958342
293715859557026466
106920718330384544
23458617733909980
2345861773390998
/
\
8.6 The g-Theorem and Its Consequences
273
To construct such f-vectors we use g-vectors of the form

91
n d 1+ r,
(n d 2 + k)
gk :=
for k 1.
Now take d = 20, n = 169, and r = 4203045807457, and compute (i.e., let
MAPLE compute).
The existence of the corresponding polytope follows from the (necessity
part of the) g-theorem. However, the corresponding polytopes Cd (n)<r>
are also easy to construct "by hand": see Exercise 8.32.
If we go a little higher in dimension, then the same construction produces
nonunimodal f-vectors for simplicial polytopes with much fewer vertices:
so, for d = 30, n = 47, and r = 65555 one obtains a simplicial f-vector
with only fo = 65602 vertices. However, Eckhoff [178] observed that with a
more complicated f-vector one can do even better. The simplicial f-vector
with the smallest number of vertices he found is
1
f-i =
1320
fo =
869619
24650747
342491792
h =
3070918789
h =
19918328394
f5 =
99465082767
.f6 =
h = 397591643442
= 1306188319799
.f8
3593770140180
=
h
b. = 8397239870111
fi i = 16843753477928
f12 = 29259588507633
f13 = 44370698483306
f14 = 59263421467414
f15 = 70604148959649
f16 = 76609321169592
L7 ---= 78245589858777
fis = 78245589349944
fig = 78245589350797
ho = 76598891788386
f21 = 69592677861523
f22 = 55485099387534
f23 = 37137014371927
f24 = 20144065902012
f25 = 8558343705069
f26 .--- 2730558787586
613985498319
f27 =
86678396880
f28 =
f29--,--5778559792
fi =
f2 =
274
This is obtained from the h vector

-
ho =
h1 =
h2 =
1
1289
830484
18) (i;fir)
1252344784
hi
h15 =
for 3 < j < 14
By the (sufficiency part of) the g-Theorem 8.32, the existence of the corresponding simplicial polytope follows, if we verify that these hi form an
M-sequence, and this is easy do it!
From these f-vectors of "large" polytopes, you can perhaps get a more
realistic feeling for what f-vectors look like "in practice." Observe how the
monotonicity statements of Bjetrner's Theorem 8.39 still hold. I guess the
more general moral is that you shouldn't rely too much on intuition from
3- and 4-polytopes when you want to get a feel for the behavior of "typical"
simplicial polytopes.
Furthermore, in contrast to all the detailed (essentially complete, by the
g-Theorem 8.35!) information known about f-vectors of simplicial polytopes, we do not know much about nonsimplicial polytopes. Our knowledge
is not even complete for 4-polytopes (Problem 8.29*). In Problems 8.33*
and 8.35* we ask basic questions about the f-vectors of general polytopes.
Here we end the chapter with a construction due to Eckhoff [178]
that "easily" produces nonunimodal f-vectors in low dimensions.
Example 8.41. (Eckhoff [178])
Let P be a simplicial polytope, and P' a simple polytope, both of dimension d. Now we "cut off" one vertex from P'; then, after a projective
transformation, we can "glue" the rest of P' onto a facet of P, to obtain
the connected sum P#P'. Instead of formal details for this construction,
we just provide a sketch of a 3-dimensional connected sum.
Thus, the sum A3 #6.3 is combinatorially equivalent to the capped prism.

Now if P has the f-vector
fo,
..., fd-2 d-1)
= (1 , f,
A-1)
f (P) =
and P` has the f-vector

J '
;
Notes
then
275
Pli-P` will have the f-vector
f(P#11
= (1 ) fo + .1.4 - 1,f 1 + f1, ----, fd-2 + f:1-2, fd-1 +f_1 -1)
- this is just the sum of the f-vectors, except that 1 has been subtracted
in the dimensions -1, 0 and d - 1, corresponding to the vertex of P' and
the facet of P that were deleted in the construction.
Now if P = Cd (n) is a cyclic d-polytope with many vertices, then its
f-vector peaks in dimension 1_ 3(d4-1) _], and the f-vector of its polar peaks
at in dimension VT]. This suggests that, if d and n are large enough,
then the f-vector of Cd(n)#Cd(n) A cannot be unimodal. For example,
straightforward computations, for d = 8 and n = 25, yield
f(C8(25)#C8(25) ) =
= (1, 7149, 28800, 46800, 46400, 46400, 46800, 28800, 7149).
Similarly, for d
= 9 and n = 18 one gets
f(C9 (18)#C9(18) )
= (1, 1447, 6588, 12984, 15618, 15552, 15618, 12984, 6588, 1447).
These polytopes, in dimension 8, and with less than 1500 vertices, you
might even consider as "small" (if you compare them to our previous,
simplicial examples).
Notes
Sections 8.1 and 8.2. Sch1511i [437] had made a shellability assumption for
his 1852 proof of the d-dimensional Euler-Poincar formula, but did not
specify the exact condition he needed. Thus the theory of shellability got
its basis with the paper by Bruggesser & Mani [132], published in 1971, in
which they first defined the concept. Bruggesser and Mani write in their
introduction: "We were surprised to find that Schlafli's assumption can be
justified in an almost trivial manner" [132, p. 197].
Since then, shelling has become a very basic and useful technique with
many (geometric and combinatorial) applications. We refer to Danaraj &
Klee [164], Bjtirner & Wachs [94], Bji5rner et al. [92, Sect. 4.7], and in
particular to Bji5rner [83] for further reading and references.
A nonshellable triangulation of a tetrahedron (with 14 vertices, 41 facets,
all vertices on the boundary) was constructed by Rudin [433] in 1958. This
discouraged geometers from trying to prove that the boundary complex of
every polytope is shellable. Rudin's ball can even be brought into convex
position [155, p. 305], so it can be considered as a nonshellable triangulation
of a 3-polytope without new vertices.
276
Rudin's construction is pretty subtle, and hard to visualize. I spent long

and rainy days at the Majestic Caf in Paris (Rue vieille du Temple, zle arr.)
trying to "understand" it. It is a challenging exercise to see that it actually
works see also Problem 8.7*. Our alternative construction in Example 8.9
is based on an ingenious construction by Danzer [168] for a completely
different problem. Danzer's construction yields a nonshellable subdivision
of a 3-polytope into only 13 convex polytopes. His construction is, however,
not easy to visualize either.
In this context let us mention that nonshellable topological (nonstraight,
as in Example 5.12) subdivisions of 3-polytopes are not that hard to construct. The earliest example of a nonshellable 3-ball (with 4.7 + 2 = 30
vertices and 4.16 + 4.8 4.2 104 facets) was given by Newman [392] in
1929. His construction is in fact very simple and geometric. The smallest
known nonshellable triangulated 3-ball has 12 vertices and 28 facets [538].
There is also a reversed way of viewing a shelling, by deleting facets instead of adding them. Since this version has more topological subtleties,
we have avoided it completely. However, these reversed steps lead to the
extremely beautiful constructions for nonshellable topological balls as subcomplexes of piles of cubes by Bing [77], including the "two-room house"
and the "knotted hole ball," which is actually due to Furch [202]. (Stillwell,
in his wonderful book [491], says this "shows knots again causing trouble.")
The knotted hole ball is what you get if you start with a pile of cubes, and
drill a knotted hole through it (that is, remove the cubes along the hole in
reverse shelling steps), until there is only one cube left before one would
reach the opposite wall. Such a nonshellable ball is in fact what is left over
as the "unshellable part" of our third construction for Theorem 8.15. See
[538] for a survey and more about non-shellable balls and spheres.
Related to this, there is *Tinos Pach's (still unsolved) "animal problem."
according to Pach is any topological 3-ball in R3 , consistAn animal
ing of unit cubes (a subcomplex of a pile of cubes). The question is whether
every animal can be reduced to a single unit cube by adding and deleting
cubes, while maintaining the animal property throughout. Just deleting
cubes is not sufficient for this: for example, Bing's nonshellable "knotted
hole" balls provide counterexamples. Specific small counterexamples were
constructed by Ke 8z O'Rourke [292] and by Shermer [461, 462]. Our first
two constructions for Theorem 8.15 are based on Shermer's smallest irreducible animal, the "Z-animal" from [462].
Explicit nonshellable piecewise linear 3-spheres (nonstraight subdivisions
of 4-polytope boundaries) arise from the same circle of ideas as Bing's nonshellable balls; see Lickorish's paper [335]. For this consider a triangulated
Bing ball corresponding to an arbitrary knot K, and complete it to a triangulation of S3 with one new vertex, by adding a pyramid over the boundary
complex. Lickorish shows (using the Alexander invariant of classical knot
theory, which has the special feature of having "a meaning" in homotopy
theory) that if the knot is complicated enough, then the sphere we get is
Notes
277
not shellable. In particular, he argues that if we take K to be one trefoil

knot, as above, this might not be good enough, but if we take K to consist
of three trefoil knots in a row, then this produces a nonshellable sphere.
(For an alternative construction see Armentrout [22], and Vince [516] for
a specific (nonsimplicial, cellular) example.)
Surveys of the topic of nonshellable balls and spheres appear in Bing {771,
Danaraj & Klee [164], and Ziegler [538].
On extendable shellability, I have said most of what I know in the text.
Theorem 8.15 solves an old problem of Helge Tverberg; see the 1978 paper
of Danaraj & Klee [164, p. 37], and also Ewald et al. [191, p. 141111. The key
observation is that Bing's nonshellable balls are easily embedded into the
boundary of a 4-polytope. Kleinschmidt [306] has verified that d-polytopes
with d + 2 vertices are extendably shellable.
The Bruggesser-Mani method of shelling polytopes is usually described
as a rocket flight; see for example, [92, Sect. 4.7(c)]. Bruggesser and Mani
themselves were much less aggressive at the time (and not influenced by the
NASA-Saturn craze): they thought of this as a balloon trip. The proof of the
Euler-Poincar equation we get from shelling is close to Schldfli's original
proof from 1852 filling the gap where Schlafli assumed the existence of
a shelling order, without defining or proving it.
Here is another problem of Tverberg: is every complete (simplicial) fan
shellable? This situation is strictly more general than shelling polytopes,
since face fans of polytopes are a special case. In this situation, the Euler
and Dehn-Sommerville equations still hold, but our proofs do not: see Kleinschmidt & Smilansky [309], Eikelberg [183], and their references. Shellability of fans indeed follows from a lemma of Ehlers [182, Lemma 3], which
turned out to be faulty [183, S. 20]. However, fans are partitionable (see
Kleinschmidt & Smilansky [309]), and this suffices to prove the upper bound
theorem for fans.
Section 8.3. With the f- and h-polynomials for simplicial polytopes you
have seen a glimpse of the method of generating functions. We have no
need or time for more on this elegant and powerful method, but if you got
interested you might want to study Graham, Knuth & Patashnik [222] or
Stanley [480] to learn more about it.
Section 8.4. Here we have closely followed McMullen's original proof [361]
for the upper-bound conjecture. The idea for the upper bound theorem for
complexes with many vertices in Section 8.5 is from Klee [296]. Our version
combines McMullen's reduction to hi < (n-d-i-l+iN) with Sperner's lemma
(which was rediscovered by Klee [296]; see also [362] and [234, p. 182]).
As McMullen added "in proof" at the end of his paper, part of his proof
becomes simpler if one switches to the polar version, for simple polytopes;
see Exercise 8.11. Some parts, however, like the characterization of the
equality case, become more involved, see the full-length versions of the polar proof in [115], [126] and [387]. A different proof is by Alon & Kalai [12],
278
also presented in Fiiredi [197] and in Ewald [189, Sect. 111.7]; their proof
is based on "shifting," a linear algebra method by Kalai which you may
find explained in more detail in Bj8rner & Kalai [91]. Shifting also leads to
farreaching extensions of the upper bound theorem (for subcomplexes of
polytopes) by Kalai [277, Sect. 9], which in turn can be applied to the diameter problem [277]. A proof of the upper-bound conjecture that is valid for
general triangulated spheres, not necessarily shellable, was found by Stanley [474, 478] (see also Hibi [252]), using the commutative algebra methods
we mentioned before. Very nice surveys are Stanley [479] and Bjiirner [82].
See Clarkson [152] for a different, combinatorial proof. Considerable very
recent progress is provided by Novik [395].
Section 8.5. Extremal set theory is an extremely interesting and very widely applicable part of combinatorics, of which we have only "scratched the
surface." We recommend the paper by Greene & Kleitman [223] and the
book by Anderson [18] for more material. See also Fiiredi [197], Engel &
Gronau [184] and Engel [185]. A spectacular recent success of extremal set
theory methods applied to a polytope is due to Kahn & Kalai [271], with
a lovely one page version by Nilli [394]).
Section 8.6. Both parts of McMullen's g-conjecture were established in
1979. In that year Billera & Lee [324, 70, 71] proved the sufficiency of McMullen's conditions (they describe an ingenious combinatorial-geometric
construction of a simplicial polytope with any prescribed M-sequence as
its g-vector). The paper [71] is highly recommended for study: it has motivated some spectacular research, notably Kalai's construction of "many
nonpolytopal spheres" [274] [330].
The necessity part of McMullen's g-conjecture (i.e., that the g-vector has
to be an M-sequence in all cases) was in the same year proved by Stanley [477, 479]. This relied on heavy machinery from algebraic geometry: the
hard Lefschetz theorem for the cohomology of projective toric varieties. (It
may be noted that the algebraic geometry tools were not complete at the
time: the only available proof turned out to be faulty. A new and even
more technical one was eventually done by Saito [436], see Stanley [481,
p. 64], Fulton [201, Sect. 5.2], and also Oda [397].) A more elementary proof
of this half of the g-theorem was long searched for, and recently given
by McMullen [369]. The new proof also uses developments (McMullen's
polytope algebra, see McMullen [367, 368] and Morelli [382, 383]) outside
the scope of this book; however, it keeps getting simpler. McMullen's paper [371] explains that it is not necessary to study the polytope algebra for
this purpose: the (much simpler) "weight algebra" will do the job. McMullen's abstract concludes: "Thus a yet easier proof of the g-theorem is now
available." See [372] for the latest "state of the art."
Note that there is still no proof that would establish McMullen's conditions for simplicial spheres, like for simplicial fans (where fi counts the
(i + 1)-dimensional cones).
Notes
279
The matrix formulation of the "McMullen correspondence" is due to

Bj6rner [82, 86]. It seems to be the nicest (though still complicated) way
of analyzing the f-vectors of simplicial polytopes, and deriving various
consequences of the g-theorem. Very recent work is Bj6rner & Linusson [93].
His matrix formulation also led Bj6rner to disprove the unimodality conjecture for simplicial polytopes [80]. Bj6rner's first counterexample was in
dimension d = 24 with roughly 2.6 x 10 11 vertices, while counterexamples
with d = 20 were soon after found by Bji5rner [80] and Lee [80, 71, 324]
(all these counterexamples correspond to stellar subdivisions of cyclic polytopes). In Lee's thesis [324, p. 1111 one finds (6 digits of) an f-vector of 20dimensional convex polytopes on about 4.2 x 10 12 vertices with > fi2 <
113 , as in Examples 8.40. Independently, Eckhoff disproved the unimodality conjecture in dimension 21 for simplicial polytopes, and in dimension 8
for general polytopes (see Example 8.41 and Problem 8.33*). However, rumour has it (and Eckhoff knew) that already in 1964 Danzer lectured in
Graz (Austria) about the construction of (very high-dimensional, nonsimplicial) polytopes with nonunimodal f-vectors, based on the join operation
P * P on polytopes (Exercise 9.9) which corresponds to a convolution of
fifk _i_ i . By 1973, Danzer knew that nonunif-vectors, fk (P*P)
modal f-vectors of simplicial polytopes in dimension d = 54 can be obtained by repeated stellar subdivisions of cross polytopes. None of this was
published.... Is this (nearly) forgotten mathematics?
For the computations of f-vectors, I have used and enjoyed the
computer algebra system MAPLE, in particular its facility to efficiently
handle extremely large integers and binomial coefficients.
The lower bound theorem for simplicial polytopes was proved by Barnette [38, 40], roughly at the same time when McMullen proved the upper
bound theorem. Extensions appear in McMullen & Walkup [376], and in
Klee [298]. The extremal polytopes were characterized by Barnette and by
Billera & Lee [71].
Ei
Extensions. There has been enormous work and progress on f-vectors of

polytopes in recent years, so much that we could not even mention all the
main directions here. We refer to the excellent surveys by Bayer & Lee [59],
Lee [328], and Klee & Kleinschmidt [301] for up-to-date information and
references. Let us just mention a few points here.
(1) In [49], Barnette, Kleinschmidt & Lee derive an upper bound theorem
for polytope pairs polytope pairs are important because they correspond
to the case of (unbounded) polyhedra, capturing also their "combinatorial
structure at infinity" (cf. Exercise 2.19). Similarly, there is a lower bound
theorem for polytope pairs by Lee [325].
(2) It is an open problem to formulate and prove an upper bound theorem
for centrally symmetric polytopes. Here one would call a polytope centrally k-neighborly if every set of k vertices, among them no two opposite
ones, form a face. Surprisingly enough, there seems to be no straightfor-
280
ward generalization of cyclic polytopes that would do the job. By results of

McMullen & Shephard [373], Schneider [441] and Burton [134], the maximal neighborliness is severely limited. In particular, Griinbaum [234, p. 116]
showed that there is no centrally symmetric 4-polytope with 12 vertices that
has (122\ _
)
= 60 edges. Even more surprisingly, a combinatorial sphere
with such parameters exists [235]! In fact, Jockush [268] has recently constructed centrally symmetric, 2-neighborly 3-spheres with 2m vertices for
all n > 4. Thus, there is a considerable gap between the upper bound
theorems for centrally symmetric polytopes, and for centrally symmetric
spheres. This suggests interesting problems. For example, can one construct
centrally symmetric fans that are centrally Ili-neighborly, or at least 2neighborly in this sense, with a large number n = h of one-dimensional
rays?
Also, it seems to be extremely difficult to get good lower bounds for the
face numbers of centrally symmetric polytopes. The first nontrivial step was
taken by Stanley [481], who proved h i > (di) (.d1) for simplicial
centrally symmetric polytopes, verifying by this a conjecture by Bj5rner.
See also Problem 8.36*.
Also, polytopes with other types of symmetries have been studied. So,
an interesting approach of treating the Dehn-Sommerville equations in an
equivariant setting (for example, the centrally symmetric case) can be found
in Barvinok's [53] work. Adin [2, 4] has lower bound theorems for polytopes
with higher-order symmetries.
(3) There has been some progress in understanding cubical polytopes, all of
whose proper faces are combinatorial cubes. Adin [3] developed h-vectors of
cubical polytopes, and used them to derive the [4] equations for their face
numbers, the cubical analogues of the Dehn-Sommerville equations. See
also Jockush [267], Blind & Blind [98, 100, 101], Billera, Chan & Liu [66],
and Babson, Billera & Chan [29].
(4) Perhaps the most striking problem is to understand the f-vectors of
general (nonsimplicial) polytopes. In this case, it is not clear that the fvector itself is sufficient information to deal with. The flag vector, counting
chains of faces in specified dimensions, is a much more informative object. It was studied in some detail by Bayer & Billera [58], Bayer [55, 57],
Kalai [273], and others.
Stanley [4821 has defined a generalized h vector for general polytopes,
which was suggested by related concepts from algebraic geometry. This
might be the right way to view the combinatorial information. However,
the recursive definition makes this object hard to study, and some of its
most basic properties are still not proved.
An alternative way to encode the information, which is equivalent to
the generalized h-vector, is provided by the cd-index of Jonathan Fine; see
Bayer & Klapper [60]. There is a lot of activity in this field, with interesting
new work by Purtill [414], Stanley [484], and others.
-
281
(5) The structure of polytopal complexes is far from understood. Even

polytopal subdivisions of 3-polytopes pose more problems than there are
answers in the moment. The reader will find some open problems among
the exercises; we refer also to Bayer [56, 57] and to Lee [326, 329]. The
space of all regular subdivisions of a polytope will reappear in the next
lecture, in a completely different setting. There are challenging questions
in the study of f-vectors of subdivisions and their properties as well. We
just mention Stanley's local h-vector as a new tool; see Stanley [483] and
Chan [138].

8.0 Show that every polyhedral subdivision of a 2-polytope is extendably
shellable we can start with an arbitrary 2-face, and never get stuck.
(This is classical, see Newman [392], but also [165] and [221].)
Show that one cannot, however, prescribe the last 2-face of a shelling
in general.
Show that a set of facets of the d-cube determines a shellable
subcomplex of 8Cd if and only if it contains no facets (is empty),
or all facets, or if it contains at least one facet such that the
opposite facet is not in the complex.
Deduce that the boundary complexes of the d-cubes are extendably shellable.
(ii) Describe a shelling of the d-dimensional crosspolytope. Use it
to compute the f-vector and the h-vector of the d-dimensional
crosspolytope.
(iii) Given the h-vector of a simplicial polytope P, how can one derive
the h-vector of the bipyramid bipyr(P)?
(iv)* Are the d-dimensional crosspolytopes Cd extendably shellable?
8.1 (i)
8.2
Show that an ordering F1, F2, . . . ,F8 of the facets of a pure simplicial
complex is a shelling order if and only if for every i > 1, the facet Fi
contains a unique minimal face which is not contained in an earlier
facet Fi with j < i.
Show that the permutation F1 , ... , Ft is a shelling of A if and only if
A is partitionable with a partition such that
Ri C F3
', .
i < j.
8.3* Is every d-diagram shellable? What can you say about the case d = 3?
(For d = 2 this is true, by Exercise 8.0. If you want a guess for d > 3,
I'd vote for "no," because of the rule of thumb, "if Bruggesser-Mani
doesn't shell it, then it isn't sheilable." )
282
8.4* If a polytopal complex C is shellable, but not necessarily simplicial,

is it still true that its stars and links are shellable?
(Be careful: results of Pachner [403] show that the boundary of a
shellable ball need not be shellable. So the answer could be "yes" for
stars and "no" for the links.)
8.5 Show that for every vertex y of a d-polytope P, there is a polytopal
complex C that subdivides the boundary complex of the vertex figure,
ICI =- Piz), and that is combinatorially equivalent to link(v,C).
(The face fan of the complex C is a flattening of the boundary complex
of P at the vertex v; see MacPherson [346].)
8.6 If C = c ( p) is the boundary complex of a (nonsimplicial) polytope
and v is a vertex of P, then link(v,C) is shellable. To prove this, show
that the links are isomorphic to the boundary complexes of polytopes.
(Hint: Use a point beyond v.)
8.7* What is the smallest possible number of vertices for a nonshellable
triangulation of a 3-polytope?
(Rudin [433] claims that one needs 14 vertices if the complex is realizable as a simplicial geometric subdivision of a simplex. Is that
true?)
How many vertices are needed for a simplicial, nonshellable 3-sphere?
8.8* Beat Lemma 8.6: is this the smallest pile of cubes that is not extendably shellable?
8.9 Every shelling F1 , F2 ,. , F8 of the facets of a polytope P also induces, for every facet Fi , an ordering of the facets of F. Namely, one
can take facets of Fi in the order in which they appear in the list
1 n Fi, , F8 n F. A shelling is perfect if
n
- Fi-i n F, F
this ordering is a shelling order of the boundary of Fi , for all i.
For example, shellings of simplicial polytopes are always perfect.
(i)
Show that Bruggesser-Mani shellings are not perfect in general.
(ii)
Show that the d-cubes Cd have perfect shellings, for all d > 1.
(iii)
Show that all 3-polytopes have perfect shellings.
(iv)
Show that the polars of cyclic polytopes Cd(n) have perfect

shellings.
(v)* Does every polytope have a perfect shelling? (Kalai)

8.10 For a d-polytope P, show that the linear functions on P in general
position really correspond to the Bruggesser-Mani line shellings of P.
If P is simplicial, verify that under this correspondence the vertices vj
of in-degree k on P correspond to the facets with restriction set of
size IRj I = d - k.
283
In particular, the formula

f = 14? + 2h? + 4h? + ... + 2 k h(e + ... + 2dh,
which we used for the total number of faces of P in Section 3.4,
amounts to the evaluation of f (1) = h(2) for the polytope P.
Using this, show that Kalai's "good orientations" on the graph G(P)
are exactly the shelling orders for OP.
8.11 Prove the upper bound theorem for simple d-polytopes with n facets.
For this, consider a linear function C E (Rd ) in general position on a
simple d-polytope P C Rd with n facets. For t E R define hk (P, t)
to be the number of vertices in Ix E P: cx < t} that are the highest
point for k different edges (as suggested by the previous exercise).
By letting t increase, show that for all t E
(dk)hk (P,t) + (k+1)hk+1 (t) Ehk (F,t) < nhk+MP,t),
F
where the sum is over all the n facets of P, and the second inequality
follows from consideration of a linear function c for which the vertices
of F are smaller than all other vertices of P. Deduce from this that
hk (P)
(it k 1\
d k )'
for d > k > d [g], and from this the upper bound theorem.
(This is the "dual proof" of the upper bound theorem 8.23, from
McMullen [361, Note added in proof].)
8.12 For a simple d-polytope P C Rd with n vertices, a numbering of the

vertices by 1, 2, ... , n is called completely unimodal if every k-face
(2 < k < d) has a unique local minimum, that is, every face F has
only one vertex such that all its neighbors on F get a larger number.
(i) Show that the completely unimodal numberings of P exactly
correspond to the shelling orders of the (simplicial) polar polytope P. (Williamson Hoke [529, Prop. 1]; see also [2561)
(ii) Show that if P is the d-dimensional cube, then it suffices to
assume that the numbering has a unique local minimum on every
2-face. (Due to Hammer, Simeone, Liebling & de Werra [248];
see [529, Prop. 2 ] .)
(iii) Show that the 4-cube has a numbering that is not completely
unimodal, but for which every k-face with k
2 has a unique
local minimum.
284

(iv) Construct a simple 3-polytope P with a numbering such that
every 2-face has a unique local minimum, but there are two
local minima on P.
(Hint: First construct an acyclic orientation of a 3-polytopal
graph such that there are two sinks, but only one sink if one
restricts to one of the facets.)
8.13 Let f = (1,23,47,52,38,12).

(i) Is f the f-vector of a simplicial complex?
(ii) Is f the f-vector of a shellable complex?
(iii) Is f the f-vector of a simplicial polytope?
8.14 Show that '9k(n) is a monotonically increasing function of n (for
fixed k).
Characterize the values n for which ak(n) -= ak(n -I- 1).
8.15 Characterize the f-vectors of connected simplicial complexes, as follows. The sequence f = (1, fo , f i , f2 ,... , fd) is the f-vector of a
connected simplicial complex if and only if it satisfies the conditions of the Kruskal-Katona Theorem 8.32 and the additional relation
a3(f2) h fo + 1.(This is due to Bji5rner [87].)
8.16* Characterize the f-vectors of pure simplicial complexes.
(i) Start in low dimensions, with 2-dimensional and 3-dimensional
simplicial complexes. Note that there are gaps: for example, a
2-dimensional pure simplicial complex with f2 = 4 facets has at
least fi > 6 edges, and at most fl < 12 edges but fi = 7 is
impossible.
(See Leck [324)
(ii) In general, this is probably an intractable problem, since a complete answer would solve virtually all basic problems in design
theory this observation may originally be due to Singhi &
Shrikhande [465, p. 67] : it is called "trivial" there. (See [63] for
more on design theory.)
As an example, show that the existence of a projective plane of
order d is equivalent to the existence of a pure simplicial complex (of dimension d) with fo = d2 d +1 = fd and fi = (.D.
(It is a notorious problem to show that such an object can exist
only if d is a power of a prime. The nonexistence is classical for
d = 6. It was only recently proved for d = 10, by Lam, Thiel &
Swiercz and a CRAY lA [320]. It is completely open for d = 12.
Don't try! Try!)
Problems arid Exercises
285
8.17 For the standard octahedron C3'6' C R3 , find a shelling line i E R3

thagenrs
the shellings of Example 8.22.
(Hint: Use Exercise 8.10).
Show that the facet ordering of the octahedron (labeled as in Example 8.22)
123, 234, 345, 135, 246, 126, 156, 456,
is a shelling order for the octahedron. Prove that, however, it is not
a Bruggesser-Mani shelling for any realization of the octahedron.
(Smilansky [4681)
8.18 Derive
d
fk
'
( i
i=k
for k= 0, 1, . .. , d
for the f-vector of a simplicial polytopes from the Dehn-Sommerville

equations of Theorem 8.21.
In particular, how does the "obvious" equation 2fd- 2 = dfd_ i follow
from the Dehn-Sommerville equations?
8.19 Prove Stanley's trick: give a formula for the (1,3)-entry of Stanley's
difference table, and show that the last row correctly computes the
h-vector.
Why are all the entries of the table nonnegative for a shellable complex?
8.20 (i) Prove that for 0 < j < k<d, one has
d( di)
i.,
0+1 j\
V1+1 kj .
(ii) Prove directly that for 0 < j <d < n, we have

i
= (n
d 1+j) .
i=o
For this, verify that the equation is true for d = j (by induction)
and for j = 0, and then use induction on d, where the left-hand
side and the right-hand side satisfy the same simple recursion.
(See [234, p. 149], and also [222, p. 169].)
Use this to compute the h-vector for simplicial neighborly polytopes, directly from Definition 8.18.
286
8.21 Use Exercise 2.20 to show that one need not consider unbounded
polyhedra for the upper bound theorem, as follows.
For every unbounded d-polyhedron P with at least two vertices, there
exists a d-polytope with the same number of facets, but with more
vertices than P.
What happens in the cases where P has at most one vertex?
8.22 Give natural bijections between
the k-multisets with elements from [n],
the monomials of degree k in the variables x 1 , x2,. x n , and
vectors z E Ng with liz = k.
Show that under these bijections, inclusion of multisets corresponds
to divisibility of monomials and to the componentwise ordering of
vectors.
_ (+k-1)
n k . For example,
Give three more proofs of ((z)
(i) Show that every k-multiset with elements in [n] corresponds to
a sequence like**I*11***1*1 with n stars and k-1 bars, where
the number of stars between the ith bar and the (il)st bar
is the multiplicity of i in the multiset. Then count the star-bar
strings.
(ii) Use induction on n, and a basic identity for binomial coefficients.
For inspiration, see also Stanley [480, Sect. 1.2].
8.23 On the vertex set [3n] = {1, 2, , 374, consider the pure complex of
dimension d = 3n 4 generated by the n facets [3n]\{3i-2, 3i-1, 3i}
for j = 1, ... n. There are 3n minimal nonfaces: the sets of cardinality
n that contain exactly one of 3i-2, 3i-1, 3i for each j.
By comparing this complex to an (n+1)-neighborly one, show that
we have
(3:1
h,
(i-3) for i < n
fn-1 = ( 3:z) 3n h n ( nI3) 3n

fn = (3in) n-3n hn +1 =--
(nI4 ) + (n 4)3n
so that for n > 6 we have hn <0, and the upper bound condition
of Lemma 8.26 is violated for h n+1 (where n + 1 < [1 .1) for a pure
complex. (Wistuba & Ziegler [530])
8.24 Show that the k-skeleta of d-polytopes are shellable polyhedral complexes.
287
(i) Show that r-lex order defines a shelling order F1 (k), F2(k) - - for the (k-1)-skeleton of the d-simplex, by directly verifying
condition 8.1(W).
(ii) Show that in fact the k-skeleta of all shellable polyhedral complexes are shellable.
Pr Is the (k-1)-skeleton of every d-simplex extendably shellable?
(This is the shelling extension conjecture, due to Simon [464,
Ch. 5]; the conjecture is known to be true for k < 3, by Bjiirner
& Eriksson [88 ] , and for k > d 1, by Kalai.)
8.25 Show the following weaker version of Macaulay's theorem, which estimates the M-sequence h = (110 , h1, . . . hd) without using the subtle
operator 8k . If hk = (1) for some real x E IR and if k > 1, then
hk_i
(k ii)(Byirner, Frankl & Stanley [89, Thm. 3])
8.26* What combinatorial conditions on a simplicial complex imply that
<
(The result is known for shellable complexes, but only with algebraic
tools, like Kalai's "algebraic shifting." Is there a fully combinatorial
rule that with every shellable complex would associate a multicomplex whose f-vector is the h-vector of the complex? Can one prove it
for pure complexes satisfying the Dehn-Sommerville equations, like
Eulerian complexes [296 ] [58 ] ?)
8.27 Show that the maximal number of vertices of a d-polytope with 2d

facets is larger than the number of vertices of the d-cube, for d > 4.
For large d, the maximal number is roughly
L
11]
\ L4J
considerably more than 2d
8.28 Show that the f-vectors of general 3-polytopes are exactly the vectors
of the form
f(P3)
= (1, 4, 6, 4) -I- a(0, 1, 1,0) -I- b(0, 0,1, 1),

with 2a > b > 0 , > a > 0,
and the f-vectors of simplicial 3-polytopes are given by b = 2a, i.e.,

f(P3 ) = (1,4,6, 4) -I- gi (0, 1, 3, 2),
with gi > O.
8.29* Characterize the f-vectors of d-polytopes.

(The f-vectors for 3-polytopes were characterized by Steinitz [489 ] :
288

see the previous exercise. This is a big unsolved research problem for
every d > 4. However, for d = 4 quite a number of necessary conditions is known, and the possible pairs (fi , fi) have been characterized
for all i < j. See Bayer [55] and Bayer & Lee [59, Sect. 3.8] for further
information.
8.30 Prove Byirner's Theorem 8.39.
For the first half, you need a lemma that verifies the monotonicity of
the rows of Md, and shows that the peak lies between j =
and
1 3(d-1) 1
Lc
For the second half, you can use g-vectors of the form
g
=
such that the kth row of Md peaks at mkp . Now let gl get "very
large."
8.31 Let P be a simplicial d-polytope P, and let P' be obtained by a

stellar subdivision (as defined in Exercise 3.0: erecting a "pyramidal
cap" over a facet). Show that
f(8P) f(8P) +
MAO
+ f(aAd i)
-
2/(Ad 1).
-
8.32 Consider the polytopes Cd(Th)<r> obtained by r stellar subdivisions

from cyclic ones. Using the previous exercise, show that the g-vector
of Cd(n)<r> is given by (1, g i +r,g2 , ... , gq i ), where gi = (n+d-21
represents the g-vector of the cyclic polytope Cd(n).
8.33* Are the f-vectors of (general) polytopes unimodal for d < 7?
(Connected sums of the form P#PA have unimodal f-vectors in dimension d < 7, see Bjarner [86, Sect. 3].)
Similarly, what is the smallest number of vertices for a d-polytope
with nonunimodal f-vector?
(Eckhoff [178] has nonunimodal f-vectors for 8-polytopes with 6375
vertices and for 9-polytopes with only 1393 vertices (Example 8.41).
Can you do with less? In the case of simplicial polytopes, Eckhoff
proved do = 19, but the smallest number of vertices is not known,
either: here fo = 1320 is the current record, see Example 8.40.)
8.34 For simplicial d-polytopes, show that fk < fd-2k and fk < Li-1-k,
for 0 < k < d; 3) ]. (Byirner [82, 86])
[
8.35* Does
fo < fi < f2 < < fLdiiii

hold for all d-polytopes?
289
8.36* Every centrally symmetric d-polytope has at least 3 d proper faces.

(This is known in the case of simplicial and of simple polytopes,
proved by Stanley [481]. However, even to show that every simplicial centrally symmetric polytope has at least 2d facets a fact first
proved by B6r6ny & Lovisz [34] one knows of no simple, "elementary" argument. Centrally symmetric d-polytopes with exactly
3d proper faces exist: take for example the d-cubes, and all the polytopes which one can construct from k-cubes by taking products and
polars. Kalai [275] conjectures that this yields all the polytopes with
exactly 3d proper faces.)
8.37* For every integer k > 1, is there an integer f(k) such that every
d-polytope with d > f(k) has a k-face that is either a simplex, or
combinatorially equivalent to a k-cube?
For every integer k > 1, is there an integer g(k) such that every
d-polytope with d > g(k) has a quotient that is a k-simplex, that is,
it has faces G 1 C G2 such that [G 1 , G2] '' /(.6, k ) = Bk (that is, the
k-simplex arises as an iterated vertex figure of a face; cf. Exercise 2.9)?
(The first question is due to Gil Kalai [276], who even conjectures that
"in some sense" a "typical" k-face of a "typical" simple d-polytopes
with n facets will be combinatorially equivalent to the k-cube, if d
and n d are both large enough compared with k.
In [276], Kalai proves that f(2) is finite in fact, f(2) = 5.
The second question is due to Micha Perles [408], who remarks that
g(0) = 0, g(1) = 1, g(2) = 3 (by Euler's equation), and g(3) ? 5
(from the 24-cell).)
8.38 Investigate the face numbers of cubical d-polytopes.
In particular, show the following:
(i) Every 3-dimensional cubical polytope has more vertices than
facets (in fact, f2 = h 2).
(ii) If P is a cubical zonotope with n zones, then

/n 1\
fo 2 n 1\ /n 1\
(/ 0 )
1 ) ... + c:1 1)) '
n \
fi = 2/ 1) .
In particular, P has more vertices than facets for d> 3.

(iii) Show that if P is a cubical d-polytope, then fk (P) > fk (Cd)
holds for all k.
(Blind & Blind [98])
(iv)
Study the possible f-vectors of 4-dimensional cubical polytopes.

In particular, they can have more facets than vertices.
(Jockusch [2671)
290
(v)* Does every cubical dpolytope with d> 4 have an even number
of vertices?
(For even d this was shown by Blind tS6 Blind ]100].)
8.39* Is every cubical polytope rational?
8.40 The twisted lexicographic ordering on the facets of Cd(n) is defined as
follows. Consider facets F = {i n . . . ,id} and G = {j 1 , ... , jd }, and let
k be the smallest index where they differ, ik ik- Setting io = jo = 0,
define that F e G holds either if ik <2k and ik-1 = ik-1 is even, or
if ik > ik and ik-1 = .1k -1 is odd. Otherwise we set F > ' G.
-
(i)
Show that in this ordering, every facet is adjacent to the previous

one.
(ii)
Show that -e is a shelling order for Cd(n), if d < 4.

Show that -e is not a shelling order in general.
(Hint: list the first 8 facets in the ordering for C7 (10), and determine the minimal new faces for each of them.)
(iii)
(This linear ordering is from Gartner 86 Ziegler [204, Sect. 4], motivated by Klee's construction [297, Thm. 1.1] of a Hamilton cycle
(a simple closed cycle in the graph which passes through all the vertices) in the graph of Cd (n)X. Part (iii) was observed by Robert
Hebble.)
8.41* Can the cyclic polytope Cd(n) be realized in such a way that it has
a Bruggesser-Mani shelling for which every facet is adjacent to the
previous one?
Equivalently, can the polars Cd(n) A of cyclic polytopes be realized in
such a way that they have a monotone path through all the vertices?
(For example, you could try to modify the ordering of the previous
exercise so that you obtain a shelling for all d, and maintain the
adjacency property (i). A positive answer to this problem would determine M (d, n) (see Problem 3.11*) to be the number of facets of
Cd(n), by the upper bound theorem.
It has been shown in Amenta Si Ziegler [16, 17] that a 2-dimensional
projection of a polar C4 (n) of a cyclic 4-polytope has no more than
3n vertices. This is equivalent to the conclusion that for large n a
monotone path through all vertices if it exists cannot be found
by a "shadow vertex algorithm," for any realization of C4(n) A . Equivalently, a Bruggesser-Mani shelling of C4(n) cannot in general be
generated by a 2-dimensional section of C4 (n) that would cut all the
facets.)
8.42* It seems to be an open problem to show that all f-vectors of cyclic
polytopes are unimodal. Are they?
9
Fiber Polytopes, and Beyond
The secondary polytopes of Gel'fand, Kapranov & Zelevinsky [213] are a

quite recent development that grew out of their theory of "A-hypergeometric functions." After Izrail M. Gerfand and Andrei Zelevinsky presented
the miraculous construction at the 1989 Symposium on Combinatorics and
Geometry in Stockholm, a lot of effort was put into understanding what is
going on geometrically. It seems that the definitive answer unexpectedly
simple is the one supplied by Louis Billera and Bernd Sturmfels [74,
75, 497], who introduced the fiber polytope of a projection of polytopes and
showed that secondary polytopes arise in the special case where we project
the (n 1)-simplex to a given polytope with n vertices.
The main goal in this lecture is to develop geometric intuition for the fiber
polytope construction. Many interesting examples have been studied, and
among them we will concentrate on the construction of the permutahedron
and the associahedron as fiber polytopes. With "fiber polytopes intuition"
we then construct the permuto-associahedra: nice new polytopes proposed
by Kapranov [286] as combinatorial objects and realized as polytopes in
Reiner & Ziegler [419].
So fiber polytopes help to solve special cases of a difficult general problem: the construction of polytopes with specified combinatorics. Using the
Lawrence construction (Lecture 6), one can see that the solution for any
specified face lattice is difficult: this "algorithmic Steinitz problem" is as
difficult as the solution of general polynomial systems over the reals; see [92,
p. 4071 . (The inverse problem, the complete description of the face lattice
of a given polytope, is not trivial either: see Lecture 1 and Exercise 9.0).
292
9.1
9. Fiber Polytopes, and Beyond
Polyhedral Subdivisions and Fiber Polytopes
The basic object of study is a projection of polytopes : P -4 Q, that is,

an affine map It : RP --+ Rq such that 77 -(P) = Q, for polytopes P C RP
and Q C R. We may assume that P is a p-dimensional polytope, and Q
is a q-polytope. A simple example, for p = 2, q = 1, is drawn here.
Definition 9.1. Let Ir : RP

Rq, 7r(P) = Q be a projection of polytopes.
A 7r-induced subdivision 7r(F) of Q is a polyhedral complex that subdivides Q, with the following two conditions:
(j)
The subdivision is of the form {7r(F) : F E F}, for some specified

collection .T C L(P) of faces of P.
(ii) 7r(F) C 7r(F1) implies F =F'n 77-1 (7r(F)), and thus, in particular,
F C Ft.
Every pol3rtope in a 7r-induced subdivision 7r(F) arises from a unique face
F G F, and the collection .T is part of the definition of 7r(F). Thus, we
usually abuse notation and call the family of polytopes F C L(P) itself the
7r-induced subdivision.
We define a partial order on these subdivisions by setting
F1 < F2
if and only if
So, F1 is "smaller" than .F2 if the union of the polytopes in F1 is contained

in the union of the polytopes in F2. This means that the subdivision {TV):
FE
of Q is a refinement of the subdivision induced by F2.
The resulting partially ordered set, containing all subdivisions of Q that
Q, will be denoted by
are induced by IT: P
Q).
293
For example, consider the projection sketched before, and label the vertices and edges of P:
V2
E2
173
P
E3
V5
'i
E4
V4
44,
/
Q
In this situation, there are three subdivisions of Q C R induced by the

projection from P C R2 7 given by
F0= {VI, 13,E3},

Fit = {v1, E17V27E27113},
F2= {v1, E5, y5, E4, V4}
Note that the actual tilings 7r(..T 1 ), 7r(F2) of Q coincide, but they are distinguished since they correspond to distinct collections ..7-1,.7.2 C L(P).
Condition (ii) in Definition 9.1 excludes "noncontinuous" sections like
{v 1 , El , v2, E4, v4 }. Thus, a 7r-induced subdivision of Q is given by a family .7- of faces of P
note that the projected set fr(F) : F E .71 is not
sufficient to determine the subdivision as defined in Definition 9.1.
Condition (II) is stronger than just requiring that 7r(F) C 71- ( P) implies
F C F': for example, families like {v1, P, v3 } are also excluded. Condition
(ii) also implies that every collection .7 - C L(P) is completely determined
by its inclusion-maximal members.
The partial order that we get for our example is
-7-0
w (P, (2)
/\
1-1
Y-2
which corresponds to the nonempty faces of a 1-dimensional polytope: we

will see why ahead.
Note that dim(P) > dim(Q) holds for every surjective map 7r : P --- Q
of polytopes. Thus, if ..7. C L(P) describes a 7r-induced subdivision of Q,
then we necessarily have dim(F) > dim(7r(F)) for all F E F. If equality
dim(F) = dim(7r(F)) holds for all F E .7-, then the subdivision is called
tight. The condition is equivalent to requiring dim(F) = dim(7r(F)) = q for
all inclusion-maximal faces F E F. So, in the above example .F
F2
are tight, but ..T0 is not.
294
We leave it as an exercise to prove that the tight subdivisions exactly

correspond to the minimal elements of w(P, Q), that is, to the minimal
elements of the poset of 7-induced subdivisions. (See Exercise 9.3, and also
Lemma 9.5).
Recall from Section 5.1 that a polyhedral subdivision of Q is regular
if it arises from all the "lower faces" of a polytope Q C iliq xR under the
projection map 7r : Rq xR . Rq that forgets the last coordinate. Formally,
the lower faces of Q are those faces that minimize some linear function
(c, co ) E (Rq x R)* with co >0 over Q.
For example, in the preceding sketch the faces t7 1 , E5 1 tl5 ) E4 1 and ti4
are the lower faces, and taking Q := P defines the regular subdivision
7r-induced by .F2.
The following construction yields regular 7r-induced subdivisions. (A different one is described in [75].)
Definition 9.2.
C
Let 7r : P 4 Q be a projection of polytopes, and let
E (RP) * . Then wc :
x i ( r(x)
cx ) is a linear map from RP to Rq x so
c determines a polytope
71.0
f (
7r (x) )
.). ---.c
tre := i.
: x E P} C illq+1
cx
which projects down to Q via the map p that deletes the last coordinate.
Let Li (QC) C L(Qc) be the family of lower faces of Q c. Then
TC .=
(R .c) _i
L4.(Qc) .
IP n(R-c) -1 (F) : F E L (Q c)}
C L(P)
induces a subdivision of Q. Such subdivisions will be called r coherent. By

-
wcoh(P1Q)
we will denote the subposet of 7-coherent subdivisions of Q, in the (usually
larger) poset of all 7-induced subdivisions.
Note that the subdivisions {ir(F) : F E .F} that arise this way are regular
by construction. Also 7r-coherent subdivisions are 7-induced, because 7r =
p o ire : p ---+ Q. It is not true that all the regular 7-induced subdivisions
are 7r-coherent, as we will see now. However, this is true if P is a simplex
(see Exercise 9.5).
Example 9.3. Let P = conv(V)-1 - bipyr(A2 ) be the bipyramid over a

triangle given by
( 0 1 2 3 4
v = -1 0 0 0 1 ) ,
1 0 2 0 1
and let 7r : P -----> Q := [0,4] be the projection to the first coordinate. Then
the tight subdivision
.F = {v i [vi,v2],v2,[7-72,v3]7v3,[v3,
V411 1) 41 [V47 275], 275}
295
illustrated in the drawing is 7r-induced,
P
V5
Q
and it is regular (all subdivisions of a 1-dimensional polytope are regular).
Q
However, this subdivision of Q is not it-coherent. In fact, if Q is any such
polytope whose lower faces yield the subdivision of Q, then there cannot
be a linear map rc : P 4 QC -f2---'
as required by Definition 9.2. In fact,
such a map would have to take yi to wi in the following sketch.
W5
Q
'
./
2 ---"*",.....-- W4
W3
However, this is impossible for an affine map, since the line [y 1 , /75] intersects the triangle [2,2 , v3, v4] in P, and there is no such intersection of
the triangle [w 2 , 10 3 ,104] with the line [w 1 , 105 ] in Q.
296
For a general projection of polytopes, the structure of the poset of all subdivisions is not clear. The "generalized Baues problem" of Billera, Kapranov
& Sturmfels [69] asked whether this poset always has the "homotopy type
of a (p q 1) sphere." For this Rambau & Ziegler [417] gave explicit counterexamples: For example, there is a projection of a 5-dimensional polytope
(simplicial, 2-neighborly, 10 vertices, 42 facets) to a hexagon such that the
poset of all subdivisions is disconnected. Nevertheless, Edelman & Reiner
[179] have proved that the generalized Baues conjecture is true in the case of
a (general position) projection of a simplex into the plane. The important
cases of general projections of d-simplices (related to spaces of triangulations) and the projections of d-cubes (related to zonotopal tilings and
oriented matroids, see Chapter 7) are still wide open and very tantalizing.
(See also Bj6rner [84], Sturmfels [497], and Mnv & Ziegler [384)
The figure on the next page shows the poset of all subdivisions for the
projection of Example 9.3.
We met a special case of this before: when P = Cp is a p-cube and Q is
thus a zonotope, then the set of all zonotopal tilings (those are the tilings of
Z = Q by faces of P = P) is the poset of all one-element oriented matroid
extensions, which appears in the Bohne-Dress Theorem 7.32 (Section 7.5).
In contrast to the set of all subdivisions, the poset of all 7r coherent
subdivisions is the face poset of a polytope: of the "fiber polytope" of the
projection . In the drawing on the next page, this is the part of the poset
drawn with solid lines the face poset of a hexagon (Exercise 9.1)!
-
Definition 9.4. Let it : P ----+ (2 be a projection of polytopes. A section

is a (continuous) map -y : Q t P that satisfies 7ro-y = ide, that is,
7r(-y(x)) --= x for all x E Q.
The fiber polytope E(P,Q) is the set of all average values of the sections
of it, that is,
E(p, (2)
f 1
'y(z)da : -y is a section of irl.
1 vol(Q) 14
Without loss of generality we can restrict our discussion to those sections

that are piecewise linear over a polyhedral subdivision of Q. We can integrate such sections componentwise, using classical Riemann integrals. Here
we use that for a linear function f on a polytope R, one has the formula
R f (x) dx = vol(RH(ro),
where ro is the barycenter of R, ro =

vo 1(R) fR z dz.
Any convex combination of sections is a section as well, and from this we
get that the fiber polytope is a convex set. Furthermore, a simple calculation
shows that it is contained in the fiber of the barycenter of Q,
E(P,Q)
7r-1 (ro ) n.P.
00
297
298
The scaling factor 1/vol(Q) in the definition is only needed for this inclusion, but is irrelevant for the geometry of the fiber polytope.
Lemma 9.5. A subdivision given by .F C L(P) induces a tight 7r-coherent

subdivision of Q if and only if it is a minimal element in the partial order
wcoh(P, Q) of 7-coherent subdivisions.
Proof. From the definition of the partial order it is clear that every tight
subdivision is minimal. For the converse, observe that if we have c E ( Pr
which induces a certain 7r-coherent subdivision .re , then we can perturb
this c to general position e, and get .Fe', and then the resulting subdivision
will be tight by construction and smaller (or equal) to the one we started
with.
D
The following is the key result from Billera Si Sturmfels [74].
Theorem 9.6. E(P, Q) is a polytope of dimension dim(P) dim(Q),

whose nonempty faces correspond to the 7r-coherent subdivisions of Q, that
is, the face lattice of E(P, Q) is
L(E(P, Q)) ------ {6} U wcoh(P, (2)
Here the vertices of E(P,Q) correspond to the finest 7r-coherent subdivisions, which are the tight ones, while the facets correspond to the coarsest
proper subdivisions.
In particular, in the special case where P = Ap is a simplex we get that
the vertices of E(A p , Q) correspond to triangulations of Q this is the
case of secondary polytopes as considered in [213, 214]; see the next section.
Proof. (Sketch)
Any convex combination of two sections is a section

again. Linearity of the integral yields from this that the set E(P, Q) is
convex. Its dimension cannot be larger than dim(P) dim(Q), because
E(P, Q) is contained in the fiber 71-1 (7-0), which has this dimension.
Every piecewise linear section that is not tight can be changed locally
in two opposite directions; thus it can be written as a convex combination
of two other sections that have a different integral. Furthermore, there are
only finitely many different tight sections. Thus we get that the set E(P, Q)
is the convex hull of the integrals 1 f 7(x)dx for which 7 is a tight
vo (Q) Q
(piecewise linear, continuous) section of 7r : P 3 Q. From this we conclude
that E(P, Q) is a polytope.
To detect the vertices of E(P, Q), we use that they arise as the unique
maxima for generic linear functions c E (RP) * . However, if c is generic, then
every fiber 7r' (r) for r E C2 has a unique maximal element with respect
to c. This shows that c determines a unique, tight, coherent section -ye
of Definition 9.2, and the integral over this section is viatheconsru
the only point of E(P, Q) which maximizes c. Thus the vertices of E(P, Q)
coincide with the tight coherent subdivisions 7r : P ---- Q.
9.2 Some Examples
299
Every face of >(P, Q) C P is defined by some linear function on P

Thus it defines a map 7rc : QC, and thus a 7r-coherent subdivision
.Fc of Q. Again a simple computation shows that for a continuous section
: Q
P, the point vo(Q) f -ydx E (P, Q) lies in the face defined by c
if and only if the image of the section is entirely contained in the collection
of faces .7"c C L(P).
This yields a bijection between the faces of E(P, Q) and the coherent
subdivisions of Q, and thus in particular, between the coarsest such subdivisions and the facets of E (P, Q).
The correspondence between faces of E(P, Q) arid 7r-coherent subdivisions of Q also yields an explicit method to construct the vertices and the
facets of E(P, Q), which we will use heavily.
Namely, for every vertex there is a unique section, which we only have to
integrate. For the facets we always proceed as follows. For every coarsest
subdivision of Q C R" that has a chance to be 7r-coherent, we construct a
"lifting" Q C Ilq-171 such that the subdivision comes from the lower faces
of Q. If we can find the affine map 7rc : P Q, then the function c that
defines the facet can be reconstructed via c.x = (ire (x)) q+i-
9.2
Some Examples
Some special cases of the fiber polytope construction are easily analyzed.
If Q = tql is a point (q = 0), then E(P, Q) = E(P,{q}) = P.
More generally, if P=RxQ is a product, and 71" : R x Q Q is the
canonical projection, then it is easy to see that E(P, Q) = E(RxQ,Q) R:
the fiber polytope is a translate of R. In fact, we get
E(R x Q,Q) = R x {q0 },

where /10 denotes the barycenter of Q.
If p = q, then we have P=Q and E(Q ,Q) = {q0 }.
If p = q + 1, then E(P, Q) is an edge {s1 , st], which ir maps to the
barycenter of Q.
Ir
:10
The lower end 51 of the interval arises as the integral of the collection
1
of "lower faces" of P, while st is the integral of the collection of "upper
faces," divided by vol(Q) in both cases.
300
These were trivial cases we will now do two more interesting ones.
We start with the permutahedron, and the (more general) "monotone path
polytopes." Then we will construct the associahedron, as the "secondary
polytope" of an n-gon. So both examples are special cases of important
constructions.
Let P C RP be a p-dimensional polytope, and consider
a nonzero linear function a E (RP) * on P. This defines a projection
Definition 9.7.
P --> Q := fax : x E P} C
to the 1-dimensional polytope Q [a min , am}, where
amin = min ax,
-= max ax.
XEP
XEP
The fiber polytope of this projection,
II(P, a)
E(P, fax : x E
is the monotone path polytope of P and a.

By Theorem 9.6, the vertices of II(P, a) are in bijection to certain paths
on the boundary of P that are monotone (strictly increasing) with respect
to the function ax. In fact, every path
VO
V1
Vn- 1
1.7n)
for vertices v i E P with

Cmin = CV()
<
CV'
< . . . < CV y _1 < CUn, emmo
defines a section
P,
t-v i_ i +
for a: =
74' Q
+ (1t).avi (0 < t < 1).
Every such section 745 defines a point in the monotone path polytope
II(P, a), namely the integral
vol(Q)
70(x)dx =
(avi avo)
amax amin
vo
+ vi
2
(aVn aVn- 1 )
Vn- Vn
a vertex of the monotone path polytope. For

Not every such point v
this it is necessary that all the segments tv i_ i , vii are edges of P, and
9.2 Some Examples
301
has to be a monotone edge path on P which is "selected" by a secondary

objective function c. Now if you disentangle definitions, this means that the
path defines a coherent section (that is, v45 is a vertex of E(P, a)) if and
only if q5 is a path that could occur under Borgwardt's [118] shadow vertex
algorithm. This is a very natural pivot rule for linear programming that is
also used under the name "Gass-Saaty rule" in parametric optimization;
see Klee & Kleinschmidt [300).
The following example describes a very special case, in which we will
again meet our friend from Lecture 0 (Example 0.10): the permutahedron.
-
Example 9.8 (Permutahedron).

[74, Ex. 5.4]
Let P = [0, 1 ] C r n be the unit cube in Rn, and let Q = [0, n] be a
segment in Im Then we get a projection
x =
xi .
i=1
e l e2 e3 = 1
The vertices of H([0, 1 , Il) = E([0 7 1] n 01 7* correspond to increasing

edge paths. Here we have the very special situation that all those paths
have their vertices at the same values of the linear function, Il vi = i, and
thus they all induce the same subdivision of Q, which breaks [0, n] into the
segments [i 1, i]. These finest subdivisions arise from maps [0,
--)
7r(x)
x
x ) such that is a convex 2n-gon. We may choose the lower
vertices of
to lie on the curve f(k) = k2 . Now ej is a projection of the
cube [0, 1 ] , so we get 0
- to be the centrally symmetric, convex 2n gon
,
] n
cony ({
( z2
: j = 0, 1, 2,
, n} u { (
n2 (i n)2
) : j = 0, 1, 2, .
n}.
302
ej correspond to permutations: the

The possible projection maps ir : P
permutation a = a(1)a(2) . a(n) corresponds to the map
(;)
( 1
ecr(i)
()
(:::11) 2 )
2i 1)
which maps ea ( i) +
i2
(i
(p ). From this we derive ce()

+ ea (i)
1)2 = 2i 1, and thus recover the linear function ca E (Rn)* as
ea
E(2k 1)xa(k) = E(2(7 -1 (k)
1)x k
:
n1
[0, 1[n that minimizes the integral
The (unique) section
fon ccry(x)dx describes a path in the 1-skeleton of [0, 1 ] :
"Ya : 0 --)
Ca
ecr (i) +ea (2)
..
ecr (1)--1-.. -f-e 0 ( n)
1-
el + e2
el
e2 + e3 = 1
For example, our drawing illustrates the path on P = [0,11 3 corresponding

to a = 231, which is selected by ex = x2 + 3x3 + 5x1.
9.2 Some Examples
303
Note that this is a special case of our discussion and computation after
Definition 9.7. Here the integral of 7' is given by the sum
ion
7' (x) dx
+ WM + 7' (2)) + - - -
1
1(2nl)e,( 1 ) + (2n-3)ecr(2) + ... + (1)e(,))
1 1.-.
/
= 2(2n 1 2i)e,( i)
2n + 1
1(
2
u ;(n)
We have to divide this by vol[0,n] = n to get the vertices of the fiber

polytope. Thus the fiber polytope of the projection Ir is an affine image,
under
1 ,..
1
x t--+
2n
of the "usual" representation of the permutahedron, which represents the
permutation a by the column vector whose entries are given by a -1 :
1
11([0, l ] n , Il) =
ED, 11n [07 n}) = (1 + )1

2n
3
To derive inequalities for the facets of this fiber polytope, we consider

the coarsest subdivisions of {0, n], which are generated by functions like
fk(x) -= max{0,x Ic} for k --= 1, 2, ... , n-1. These correspond to the
parallelogram
&k
-
cony
{ (01 , ( 0k) , (n n lc) , (nn k0} .
n-k
n-k
The corresponding maps Fr P > ejk are indexed by the subsets A C

{0, ... ,n} with 0 < IAI -= n k < n, and are given by
:
ir (x)
( E in= 1 x i )
EiEil X i
=
(
7(x)
cAx).
304
Our figure illustrates the position of the 2-dimensional permutahedron

(a hexagon) as a fiber polytope contained in the 3-cube [0, 1] 3 . You might
notice that the drawing is metrically incorrect it is just supposed to
sketch how vertices of the hexagon and the corresponding monotone paths
on the cube are selected by the same linear function on 3
el + e2 + e3 = 1
Thus we get a complete description of the fiber polytope in terms of

equations (because the fiber polytope lies over the barycenter
of Q) and
inequalities (from coarsest subdivisions) as follows:
E([0, l] n , [ 0, nD = fx E
n
: Ilx =
2,
E x, <
iEA
IAI(2n IA1)
for 0 c A c [n]}.
2n
With this we have obtained a complete description of the permutahedron

as a fiber polytope.
D
We now turn our attention to the secondary polytopes. Our main example here will be Stasheff's associahedron [485], which was first constructed
see Example 0.10.
as a polytope by Milnor, Haiman [245] and Lee [327]
It turns out that the associahedron can be realized as a fiber polytope
E(An , C2(n+1)). For that, we use the existence of canonical maps from an
n-simplex to any polytope with n 1 vertices.
For the following, we use the special n-simplex
convlei : 0 < i < n1 C
where we set e0 := O.
9.2 Some Examples

Definition 9.9 (Secondary polytopes).
[213, 214]
Let Q C Rd be a d-polytope with n + 1 vertices, vert(Q) = {vo,
The secondary polytope of Q is
E(Q) := (d + 1) vol(Q)
v}.
Q),
Q) arises from the affine map
where the fiber polytope

7r
that maps Rn 3 e i
305
Ain
v, E Q, for 0 < i < n.
Q
For every n-simplex P C
there is a projection map 7r :
that maps the vertices of Ain to the vertices of Q. Furthermore, P and
the map 7r are unique up to affine coordinate changes in n . So, for every
projection of an n-simplex P to Q, the fiber polytope E(P, Q) is affinely isomorphic to the secondary polytope E(Q). (Equivalently, one could also use
our "standard" n-simplex An c Rn+1.) Thus the secondary polytope is a
canonical object associated with any polytope Q. We refer to the papers by
Gel'fand, Zelevinsky & Kapranov [214], Billera, Gerfand & Sturmfels [68],
and Billera, Filliman & Sturmfels [67] for extensive discussions. The key
observation is that, by Theorem 9.6 together with Exercise 9.4, once the
map 7r : Ain Q is fixed, every regular subdivision of Q (without new
vertices) is 7r-coherent.
Corollary 9.10. The vertices of E(Q) are in bijection with the regular
triangulations of Q, via
(ei. +
+ eid ),
[v 0 ,... y id ] E T
where every regular triangulation of Q is represented by its collection of
d-dimensional simplices.
To get the associahedron from this construction, we use a "well-known"
bijective correspondence between the complete bracketings of a string of
n letters and the triangulations T without new vertices of the (n + 1)-gon
C2 (n + 1). So, to every complete bracketing a of the string 123 ... n of
length n we associate the corresponding triangulation of C2 (n 1), written
as a set of triples T(a) C ( 10' 3 1 ). We denote by En the set of all these
triangulations. For example, we get (using square brackets for the triples)
T2 =
=
1[013], [123]}, 1[023], [C112 ]} },

1[014], [124], [23411, 1[014], [134], [12311, 1[024], [012], [234[1,
[0341, [013] , [123}}, [034], [0231, [o12]} 1.
306
Instead of a formal definition, we give a pictorial example that explains the

correspondence (in a special case for n = 4):
T =-- 1[014],
[123],
[1341}
4(2.3)4)
Example 9.11 (Associahedron).

The associahedron (Example 0.10) was first constructed as the secondary
polytope of a convex (n+1)-gon by Gel'fand, Zelevinsky & Kapranov [213,
Rem. 7c)] [212, Example 7.3.B]. Here we get especially nice coordinates by
taking the "cyclic" (n 1)-gon, namely
C2(n + 1)
for vi
convIvi : 0 < <
.
2) (z
In this case the projection map is linear: it maps

v
for 0 < i < n,
and thus in particular 0 = eo

vo =
The area of a typical triangle spanned by vertices of C2(n + 1) is
vol[vi,vi ,vkl
-21- (j
(k
i)(k
j)
for i < j < k, which is an integer. Thus the triangulations T of C2(n + 1)

without new vertices are represented by the points
- (ei + ei
ek).
[i,j,kJ E T
Here the sum is over all triples i < j < k such that [71- (ei),Ir(ei),71 - (ek)1 is
a triangle in the triangulation T, of area 1(j (k i)(k j).
Taking this together with the above bijection, this defines a point in Rn
for every complete bracketing of 12 ... n. For example, with the bracketing
9.2 Some Examples
307
1((2.3)4) we associate the triangulation T = 1[014], [123], [134 ]} , compute

the volumes vol[014] = 6, vol[123] = 1, vol[1341 = 3, and thus we get for
this T the point
V
6(e0 + el + e4) 1(ei e2 e3) + 3(ei + e3 + e4)

(1)
13)
/ 10 )
1
0 1
1
4 I.
3
0
\ 3
\ 9
( 66g
Our figure shows the corresponding section .1T, whose integral yields vT .
e1
V4
Vo
Vi
Let us derive a complete description of the polytope E(C2(n + 1)) in

terms of equations and inequalities. For this we first need the volume and
the barycenter of C2 (n + 1). For this we can use the triangulation T
1[012], [023], ... , [0 nln ] }, and from this compute the volume
Vn := vol(C2 (n 1 ))
nEi(i 1)
2
i=2
(n + 1)
3 P
and (with an extra computation, see Exercise 9.7) the barycenter
(dgi
cn) :,_ q 0 (C2(n-l- 1 )) (
3
)
(6n2 + 1)/
for n > 2.
308
From this we get that E(C2 (n + 1)) is contained in the affine subspace
of Rn given by the equations
n
ix,
3 vol(C2(n 1)) cn =
E i2xi .
3 vol(C2 (n + 1)) dn
n4
n2
4
6 n5 5n3 n
30
How do we get the facet-defining inequalities? We use the method outlined after Theorem 9.6. The facets correspond to the diagonals of C2(n+1),
which we interpret as going from yi to yi, for 0 < i < j < n, with
2 G j i < n. For every such "admissible" pair (i, j), we construct a
regular function
fi3
x
) := max{0, y + (i Ax ij}.
This formula can be derived from the condition that the linear function
and for (x) = ( ). The comey + (i + j)x ij vanishes for ()
Y
Y =()
/2
sponding "lifted polytope" ej is given by
45 :,
conv{fii (kk2)) : 0 < k < nl.
Since P is a simplex here, we get a canonical map

7re : P --). (:),
ek 1--->
f i3
(kk2 ) .
Using this, the corresponding facet-defining linear function is ci e (R

with
*,
=
L-d fii (kk2) Xk
,1
kn
=E
max{0, k2 + (i + Ak ii} xk
k=1
n
=E
max10,(k i)(j k)} xk
k=1
=E
(k i)(j k)Xle-
k=i
The function ciix will be minimized, by construction, by those vertices vT

whose triangulation T contains the diagonal (i, j). One can work out that
9.2 Some Examples
309
the minimum then is
fj
min{eivT : T E 7-;,} =
i +1\3(j i) 2 -2
3
10
Thus we actually get the associahedron,
E(C2 (n + 1))
Kn_2 ,
and a complete linear description for it, as follows:

n4 n2
4
xi
E(C2 (n+ 1)) = {x E Rn :

j=.1
-2
i= 1
Xi
6n5 5n3 n
30
i 1) 3(j
3
i)2 2
10
for 0 < i < j < n, 1 < j i <

We can use the PORTA program to check the validity of this description
for small n, by inputting the set of vertices, or the inequality system, and
checking whether we get a polytope with the correct combinatorics.
Here is an example. For n = 4, the linear system above has the form (a
file ass2. i.eq in PORTA input format)
DIM = 4
VALID
10 1 4 9
INEQUAL ITIES_SECTI ON
1x1 + 2x2 + 3x3 + 4x4 == 60
1x1 + 4x2 + 9x3 +16x4 == 194
xl
x2
x3
2x1+2x2
2x2 +2x3
>= 1
>= 1
>= 1
>= 10
>= 10
END
where the inequalities correspond to the diagonals [02], [13], [24], [031, and
[14] (in this order). PORTA requires knowing a valid point for this system,
so we give it the point that we had computed before.
Now the PORTA command traf v ass2. le q produces from this the
list of vertices and the incidence matrix for a pentagon, in a file named
ass2. ieq. poi:
-
310
DIM = 4
CONV_SECTION
( 1) 1 4 9 6
( 2) 4 1 10 6
( 3) 9 4 1 10
( 4) 10 1 4 9
( 5) 1 10 1 9
END
strong validity table :

\I
\N
P \E
0\Q
I \S
Il
N\
T\
S\
1
*..*. :
.*.* . :
..*.* :
.*..* :
*.*.. :
2
3
4
5
2
2
2
2
2
I 22222
So the polytope is in fact a pentagon, as K2 should be....
9.3
Constructing the Permuto-Associahedron
We will now describe the construction of the permuto-associahedron of

Kapranov [286], as recently achieved in [419] (Example 0.10). There are
analogous objects constructed for signed, bracketed permutations in [419],
but we do not discuss those here. Also our discussion skips some details:
we refer to [419] for the "missing pieces."
The construction depends on a version of the associahedron in especially
nice coordinates. For this, we define
conv{fo ,f 1 ,
...
for
fi := el
.- ei,
f 0 = 0,
as a reference simplex. Again we use the special "cyclic" (n 1)-gon

C2(n 1) ---= conv{ (:2 ) : 0 1. Its scaled fiber polytope
K4_ 2
:=3
(n-F 1)
3
f
.E(A n , C2(n 1 ))
has integral vertices, given by
1(j i)(k i)(k j) -
v T :=
[i,j, ic]
(fi -I- fi
fk)
E Zn,
for all triangulations T of C2 (71 1) without new vertices. Here the sum
is over all triples i <j < k such that (71-(fi ),7r(fi ),r(f k )) is a triangle in
the triangulation T, of area 1(j i)(k j) (k j).
Furthermore, all vertices lie on a sphere around the origin:
n
E(VT) 2
(n 1) 30n4 33n2 2
3
i=1
70
A linear description of Krfi_2 is given by the equations
n
E
xi
(2i 1) - xi
i=1
n
i=1
(n 1) n
\3)22
-,. =
n2 (n2 1)
4
3 (n 1) 6n2 + 1
6n5 5n3 n
1
=
30
3 ) 15
which describe the (n-2)-subspace of Rn that contains K_27 and facetdefining inequalities
((-2k + 1) i
:=
k=i+1
>
( j i + 1) 31j
3 )
02
10
for0<i<j<nand2 <ji<n-1.
This is just what we worked out in Example 9.11, after the linear transformation in Rn that sends ei 1-4 fi for 0 < i < n (that is, e0 = 0 to
La = 0)The magical little thing is that the vertices lie on a sphere. One can
prove this by analyzing the situation along an edge, corresponding to a
single rebracketingichange-of-diagonal, but there is no really good (that is,
geometric) explanation. Do you have any ideas? The strange thing is that
312
the whole construction of fiber polytopes is certainly affinely invariant:

but suddenly here we get an effect that is decidedly nonlinear, since affine
transformations distort the unit sphere.
Now the associahedron, in the coordinates of Proposition 9.12, is used
to construct the permuto-associahedron.
Definition 9.13. (Kapranov [286])
The face lattice of the permuto-associahedron Kiln_ i is a partially ordered
set, defined as follows.
The elements of laln_ i are ordered partitions of {1, 2, ... , n } into at
least two parts, partially bracketed: this means that the blocks are treated
as if they were being multiplied together, and some of them are grouped
together by brackets to indicate order of multiplication. In particular, every
pair of brackets encloses at least two blocks.
The order relation on these bracketed partitions is as follows: A G B if
and only if B is obtained from A by removing pairs of brackets and possibly
combining all the blocks within it into one block (if there are no brackets
inside the pair we are considering).
Finally, an extra minimal element is included in Klln-i.
This yields a large, combinatorially defined poset. Typical elements (for
n = 7) that are comparable in KII6 are
((4.3)((5.7)1))(6.2) < (3 4.1 5 7)6.2 <
3 4.1 5 7.6.2 .
One can show quite easily that KlIn_i is in fact a graded, atomic, and
coatomic lattice of length n. Thus it "looks like" the face lattice of an
(n - 1)-polytope. The coatoms ("facets") of KII... 1 are the ordered partitions of {1, ... , n } , without brackets. The atoms ("vertices") correspond to
complete parenthesizations of permutations of the letters 1, 2, ... , n. The
edges are of two types: they correspond either to a single reparenthesization, or to a transposition of two adjacent letters that are grouped together.
For n = 3 we get the face lattice of a 12-gon, as follows:
1.2.3
1.23
1.3.2
13.2
K II2 :
3.1.2
3.12
313
Kapranov [286] used quite heavy machinery that to show that

is the face poset of a "cellular ball"; a simpler argument is also in [419,
Sect. 2]. In the following we want to show the stronger result that this is
the face lattice of a convex (n-1)-polytope, which we will also denote by
Kll 1 .
Our large figure (on the next page) shows a drawing of Kr13 , as a polytope. A few vertices have been labeled by the corresponding completely
bracketed permutations you might continue this a little: for fun or to
figure out how the combinatorial description matches the geometry. (However, don't scribble in the book if it is not yours!)
Example 9.14 (Permuto-associahedron). [419]
How do we get a vertex for every bracketed permutation a? For this we
rewrite a as a pair a = (a, T), where a is a permutation of {1, ... ,n}, and
T = T (a) C ( 1 ' 3 'n} ) is the set of triples of the triangulation of C2(n+1)
that is given by the bracketing of a. With this we interpret the bracketed
permutation as a triangulation of the (n+1)-gon C2(n+1), whose lower
edges are labeled by cr(1), , (n) for example, 2(3.1) is represented by
2(3.1)
Now every permutation a = o- (1)o- (2) .. . a- (n) determines a simplex in
namely
An (a)
:=
cony-CO, ea (i), e0.(1) -Fecr(2), {x G
n:1
X0 (1)
e,.(1)+. - 1-ea(n) = 111

. . .> X,( n )
X012)
1.
Thus A(u) is just the convex hull of the section

:
[0, n]
Rn
that we have associated to u in Example 9.8. For example, the permutation

= 12 ... n determines the "standard simplex"
L171:, = convff 1 ,
,f
The description of the simplices dn (u) in terms of their inequality systems also shows that they fit nicely together to form a triangulation of the
unit cube [0, 11n.
314
The permuto-associahedron K11 3

(in wonderful postscript graphics by Jiirgen Richter-Gebert,
generated from PORTA output).
315
As an example, the simplex corresponding to the permutation 231 for

n = 3 is shown in the following figure:
e2 + e3
el
=1
231
Each of these simplices has a natural map down to R2 , where we map
--{ ec (i )
- -
1--*
(! 7
z2
for 0 <j < n.
Since the simplices fit together so nicely to form a triangulation of the cube
[0,1r, and the projection maps are defined consistently on the vertices, we
obtain a continuous, but nonlinear "folding map"
H: [0, 1r --> C2 (n 1),
which is linear on the simplices
(o-). Furthermore, for every bracketed
permutation, there is an obvious section to this folding map! For this we
define the section on the vertices by
i
Vi = (.2 ) t--+ ecr(1) -F + ecr(i)
Z
and then extend linearly on the triangles of the triangulation of C2 (n+1),

to get a section
^/* :
C2 (n+1) ---- [0, lr
associated with the string a = (a, T).

The integral over this section defines a point in Rn for the completely
bracketed permutation a, and this creates the vertices of the permuto-
associahedron.
316
Our figure illustrates the section -y" : C2(4) > [O, 11 3 that is associated
to the bracketed permutation a = 2(3.1) by this method.
e2 + e 3 + e l = 1
2(3.1)
Another way to view this construction is the following. The vertices of
the "special associahedron" IC4_2 in Proposition 9.12 satisfy y 1 > y2 >
... > un : in fact, the little associahedron
1
n-2
3 1)Kf
3 (n+
= E(A4 ,C2(n + 1))
is a fiber polytope, and thus it lies in the corresponding simplex A.,(z of

the triangulation of the cube [0, 1 ] we considered. Now by just permuting coordinates, we get n! copies of the little associahedron in the various
simplices, and the convex hull of those n! little associahedra is the permutoassociahedron.
Our figure tries to sketch this for n = 3, where the associahedra are six
little line segments, whose convex hull is a 12-gon. Again the drawing is not
metrically correct it is a mere sketch of the geometric situation, trying
to illustrate the position of the "little associahedra" within the n-cube, and
how their convex hull forms the (n 1)-dimensional associahedron. ("The
idea is the important thing," as Mr. Lehrer would say.) In contrast to 'this,
9_3 Constructing the Perrnuto-Associahedron
317
our big picture on page 314 of the 3-dimensional perniuto-associahedron

was computer generated from the actual coordinates we gave for its position
in R4 thus it represents the actual geometry of the polytope, not only
its cornbinatorics.
Ci
+ e2 e3 = 1
The only change we do for the formulas is that instead of the average
integral we take three times the integral, that is, we blow the polytope up
by a factor
) = 3vol(C2(n -I- 1)), in order to get integral coordinates.
WI 1
Theorem 9.15.
va
The formula
3-y'dx dy
Jc2(n+i)
1
--.
0 -0(k i)(k
:7) - (f
-()
f cr(i) f cr(k))
(z,j,k)ET(a)
associates a point va e Zn to every completely bracketed permutation a.

The patytope
cony fe : a = (o-,T(a)) a completely bracketed permutation of ITO
is the permuto-associahedron, that is, its face lattice is isomorphic to the
poset KI1 7,_ 1 of Definition 9.13, under the correspondence a 1-4 va . It is
(n 1)-dimensional, contained in the hyperplane
fx E Rn
Ex, =
i
n4 n2 ,
4
/.
318
Furthermore the vertices of the permuto-associahedron are integral in this

coordinatization, and they all lie on the sphere around the origin:
= (n + 1) 30n4 33n2 + 2
k 3)
70
Proof. A proof with details is in [419], and we refer to the treatment there.
What is the idea? First the associahedron in Proposition 9.12 lies in the
n : 11 x ....= n 4 _4 n 2 , .
hyperplane H =-- Ix E
} Since the vertices of KlIn_ i can
be generated by permuting the coordinates of vertices of Krfi _ 2 , we get that
our polytope KlIn_ i is also contained in H.
Then we obtain the facet-defining inequalities. That is, to every ordered
partition 0 of [n] we associate a linear function c0 . For this let 0 have p
blocks, and write it as
(15 = a(ii)
- - (7 (ii) Gr(i2)
(ip) ' '
u(ip),
where the numbers ir and jr just tell us about the "block structure" of 0,
with
1 =-- il < i i, il + 1 = i2 < :721

-
)ip-1
+1 = ip
ip =
n.
Then the linear function cb we need is given by
ct = ir + jr
if ir < o--1 (k) < jr,
that is, if the letter k lies in the rth block of 0.

Now it is not too hard to show (if you use the explicit description of the
associahedra in Proposition 9.12, and the symmetry of the situation) that
the function di' is minimized exactly by those vertices V with a < 475 (in
the lattice KlIn_ i ), that is, it defines a facet with exactly the right vertices
on it.
Now we have to argue that we have found all the facet-defining inequalities. One way to do this is to use a lot of combinatorics of the poset KlIn_ i ,
such as that every element of rank n-1 lies on exactly two coatorns (facets),
together with Exercise 2.8(iv). Such an argument for the associahedron
is in [245 ] . The alternative is to use that the convex hull of the vertices that we have constructed is contained in a larger polytope given by
the inequalities that we have found. Now one can show that for every linear function, one of "our" vertices maximizes the linear function over the
facet-defining inequalities we have found. This shows that our description
of KIIii_i by inequalities is complete.
The final fact one then needs is that the vertex-facet incidences already
determine the polytope see Exercise 2.7. Thus, we have constructed a
polytope with the "right" face lattice Klln-r. 0
94 Toward a Category of Polytopes ?
9.4
319
Toward a Category of Polytopes ?
Fiber polytopes form an important first step in investigating a "category

of polytopes," a program suggested by Louis Billera. In fact, this category
should have interesting properties that are fundamental to many geometric
questions. It is surprising that the basic "universal constructions" for such
a category have hardly been studied. Among them are the fiber polytopes,
which form "kernel objects"; the mapping polytopes discussed below, which
are "spaces of maps"; and eofiber polytopes that should play the role of "cokernel objects" for which we lack even a good definition (Problem 9.161.
Definition 9.16. Let P C P and Q C Rq be full-dimensional polytopes
(of dimensions p and q). Then the set of affine maps f A 'Z : RP ---4 Rq ,
Ax + z can be identified with Rqx 13 x Rq R(P+ 1 )q The subset
(I) (P, Q) := { (A, z) G R(p-1-1)q
fA,Z (p) c Q}
is a polytope in R (P+ 1)q of dimension (p + 1)q: the mapping polytope of the

pair (P, Q).
We omit the (easy) proof that (I)(P, Q) actually is a full-dimensional
polytope (see [416]). Also there is a natural generalization to polyhedra.
Here we note a few important examples.
Examples 9.17.
(0
When P is a point (p = 0), then (I)(P,Q) is isomorphic to Q. More

generally, if P = Ap is a simplex with p + 1 vertices, then the affine
image of the vertices of P can be chosen independently in Q, which
proves an affine equivalence
QP+'.
(ii) In particular, take the (d- *simplex 2d1 = conv{e i ,..., ed} c Rd
on the hyperplane H := Ix E Rd lix = 11, as before. Then we
H with linear maps Rd 4 Rd
can identify affine maps with H
that fix O. Such maps are given by a matrix V = (v 1 , , vd) E Rdxd ,
where ei Thus the mapping polytope 4)(A d_ 1 , Ad- 1 ) turns
out to be
1) (Ad-1,Ad-1)
{(Vi, , Vd) E R dxd : Vi E Ad_i
f (vii ) E Rctxd
for 0 for 1 <
E = 1 for 1 < i <

(Ad-i) d.
< d,
320
(iii) Now we will consider the subset of all maps that preserve barycenters,
that is, map the barycenter 1 of 'd-i to itself. But the map takes
this to ti (vi + + vi), so we get extra conditions
d
= 1,
for 1 _o,
Ei vii =
1 for all
Ei vii = 1 for all j

and this is the Birkhoff polytope of all doubly stochastic matrices,
as in Definition 0.11.
The full version of a theory of "Universal Constructions for Polytopes"
will, I suspect, need two important extensions (both of which correspond
to fundamental features in the modern development of algebraic topology):
1. There should be an equivariant set-up, which takes into account the
study of group actions and symmetries on the polytopes. So, for example, our construction of the permuto-associahedra relies on a subtle interaction of fiber polytopes and a symmetry group action, and
this should be collected in a general construction of equivariant fiber
polytopes.
2. It should admit polytope pairs rather than polytopes as the primary
objects: for example, unbounded polyhedra can often be treated in
terms of pairs formed by a polytope together with a facet. Also, the
maps of simplices Ap
A q that preserve the barycenter of Example 9.17(iii) fit this pattern, as do those of Exercise 9.15.
Notes
The original motivation for the constructions of secondary polytopes and of
fiber polytopes did not come from polytope theory, but from the theory of
A-hypergeometric functions [210, 212], and from state polytopes in commutative algebra [54] [498]. It turns out that also there are strong connections
to constructions in algebraic geometry [287] and elimination theory [288]. A
survey was given by Loeser [342] in the "Bourbaki seminar." We especially
321
recommend the recent book by Gelfand, Kapranov & Zelevinsky [212] for
study.
Our definitions (starting with the Definition 9.1 of 7r-induced subdivisions) are different in appearance from, but equivalent to, the original setup by Billera & Sturmfels [74], which works with vertex sets rather than
polytopes. Only the explicit version of 7r-coherent subdivisions in Definition 9.2 might be new here. We also refer to the survey of fiber polytopes
in [497], and to the alternative set-up (via normal fans) in [75].
Similarly, the original definition of secondary polytopes by GePfand,
Zelevinsky & Kapranov [213, 214] (see [212, Ch. 7]!) was quite different
from the one, due to Billera & Sturmfels [74], which we have presented in
Definition 9.9. Our presentation also reverses the historical order of things:
the ingenious construction of [213] now appears as a very special case of
the fiber polytope construction; let us just say that the secondary polytopes are in many respects the most fundamental case. In particular, every
fiber polytope can be written as a projection of a secondary polytope; see
Exercise 9.6.
The "permuto-associahedron" Kri n_ i is a combinatorial object introduced by Kapranov [286] (he denotes it as K/3,). The construction as
a polytope, and the generalization to "Coxeter-associahedra," by Reiner
& Ziegler [419], were born in December 1992, so to speak (and not baptized). The nonlinear effects appearing in this, like the sphericity in Proposition 9.12, suggest that there is much more to be discovered and that the
constructions are not yet well understood.

9.0 For which of the examples of polytopes discussed in Lecture 0 can
we, by now, give complete combinatorial descriptions?
Which of them are related by projections?
Which can we represent as fiber polytopes associated with a projection of simpler polytopes?
9.1
Compute the fiber polytope for the projection in Example 9.3. For
this, compute both the vertex coordinates, corresponding to the six
coherent tight sections.
(According to our discussion, you should get a hexagon!)
Also, compute the point in R3 which corresponds to the noncoherent
tight section of Example 9.3. Does it lie in the relative interior of the
fiber polytope?
Finally, use the methods described in this chapter to derive a description of the fiber polytope by equations and inequalities.
322
9.2 For the projection
enumerate all the 7r-induced subdivisions, and identify the 7r-coherent

ones among them. Draw the whole poset, and show how it "retracts"
to the subposet of 7r-coherent subdivisions.
Compute the fiber polytope, and describe the position in the fiber
polytope of the points that correspond to noncoherent subdivisions.
9.3 For a polytope projection 7r : P
Q, let ,F be any 7r-induced
subdivision. Show that for every linear function c E (RP)* there is a
relative coherent subdivision Ge with Ge < F.
(The construction can be done analogously to Definition 9.2.)
Show that if c is generic, then gc is tight. Conclude that the minimal
elements of co(P,Q) are exactly the tight subdivisions.
9.4 Given a monomial m = x x2t2 -...-x dtd, define its degree by
ti
deg(m) :=
td
For a polynomial in d variables,
f
for a,
Xdl,
and monomials m , define its Newton polytope by

Newton(f) := convIdeg(m i ) : ai 01.
(i) Show that Newton(f.g) = Newton(f) + Newton(g). From this,

describe the Newton polytope Newton(f k). What can you say
about Newton( f +g), and about Newton( f FEg) for small enough
0?
E
-
(ii) Describe Newton(det(xii)), for the determinant of an (n x n)matrix, with d = n2 different variables as entries.
9.5 If 7r : P
Q is a projection of polytopes, then the 7r-induced
subdivisions of Q only use vertices in the finite set 7r(vert(P)), which
need not all be vertices of Q.
Show that if P = Ap is a simplex, then all the regular 7r-induced
subdivisions of Q are 7r-coherent.
323
9.6 Let It : P +Q be a projection of polytopes, where P C RP is a

polytope on n vertices. Show that the fiber polytope E(P, Q) can be
constructed by projecting the secondary polytope of P:
E(P,Q) =
(Billera & Sturmfels [74 ])

9.7 Verify the formulas for the volume (area) and the barycenter of the
(n + 1)-gon
C2 (n+1) =
cony { (i2) : i = 0, 1, 2, ... , n}
What about the volume and the barycenter of the general cyclic polytopes Cd(n+1) ?
9.8 Let P and Q be polygons in the plane. Show that the fiber polytope
of the projection
PxQ ---* PQ
from the product to the Minkowski sum has a fiber polytope that is
isomorphic to P + (--Q).
What goes wrong here if Q degenerates to a line segment?
9.9 Define the join P * Q of two polytopes to be the convex hull of P

and Q, if they are placed into affine subspaces of some Rd such that
their affine hulls aff(P) and aff(Q) are skew.
Show that P *Q is a polytope of dimension dim(P) +dim(Q) +1, and
that up to affine equivalence the join does not depend on the choice
of affine subspaces.
Show that the secondary polytope E(P * Q) of the join is isomorphic
to E(P) x E(Q).
(Dalbec [162])
9.10 Consider the projection it
R, An_ i --- [1, n] given by
ei 1--+ i.
Show that the fiber polytope E(A, [1, nn of this map is combina
a (n 2)-cube. Is it in fact affinely isomorphic-tonalyequivalnto
to C_ 2?
(Gelfand, Kapranov & Zelevinsky [212, Example 7.3.A].)
:
9.11 Compute the secondary polytope of An_ i x L.

For that first determine the dimension of the secondary polytope,
then determine its set of vertices. Before you start to actually compute
vertices, you should figure out the combinatorics of the polytope you
get. (It is a good old friend!)
324
Also, study the secondary polytopes of E, x

(See Gelfand, Kapranov & Zelevinsky [212, Examples 7.3.C,D]; The
secondary polytopes of products of simplices are really complicated:
see Babson Sz Billera [28 ] !)
9.12 Show that if Z is a d-zonotope with n zones, then the fiber polytope
of the canonical projection
is a zonotope as well.
(Billera & Sturmfels [74, Thm. 6.11)
Compute the fiber polytopes for the projections
and
C6
P12
2
'
where p22i denotes a centrally symmetric 2i-gon. For the projection

C6 4 P212 , the answer depends on the specific choice of a 12-gon.
Determine the (five) different f-vectors that occur in this case.
(Sturmfels [4961)
9.13 If P and Q are disjoint polytopes in d, show that the region between
them can be triangulated without new vertices. That is, there exists
a simplicial subdivision of conv(P UQ) whose vertex set is vert(P) U
vert(Q), and such that there are subcomplexes that triangulate P
and Q.
(Hint: Start with a convex function that is linear on P and constant
on Q, and use it to "lift" Q. Then take the convex hull, and perturb
all the vertices. The result is due to Goodman Sz Pach [215 ] .)
9.14 For arbitrary polytopes P C P and (21 1 Q2 C
4)(13) Qi
+ Q2
) (P1Q 1)
show that
+ (PS22)-
Show that equality does not hold in general.

(Rambau Sz Ziegler [416 ] )
9.15 Compute the mapping polytopes
731),
1
(Aq,
and describe them combinatorially.

9.16* What is a cofi ber polytope?
(This should be a polytope that is naturally associated to any inclusion of polytopes P Q )
References
7397-city traveling salesman instance solved, Optima No. 45, March 1995, 7 -8. (23)
[1] KAREN AARDAL:
[2] RON M. ADIN: On face numbers of rational simplicial polytopes with

symmetry, Advances in Math. 115 (1995), 269- 285. (280)
[3] RON M. ADIN: A new cubical h-vector, Discrete Math. 157 (1996),
3-14. (280)
[4] RON M. ADIN: On h vectors and symmetry, in: "Jerusalem Combinatorics '93" (H. Barcelo and G. Kalai, eds.), Contemporary Mathematics 178, Amer. Math. Soc. 1994, 1-20. (280)
-
[5] OSWIN AICHHOLZER
: Combinatorial & Computational Properties
of the Hypercube. New Results on Covering, Slicing, Clustering, and

Searching on the Hypercube. (Ph.D. thesis, TU Graz (1997), 170
pages.) 26
[6] DIMITRIS ALEVRAS, GARY CRAMER Sz MANFRED PADBERG: DemoSofts for Optimization I: DODEAL, ENTSP, ENZERO, preprint, New
York University 1993, 41 pages. (48)
[7] GERALD L. ALEXANDERSON
Si
JOHN E. WETZEL:
A simphcial
3-arrangement of 21 planes, Discrete Math. 60 (1986), 67 - 73. (225)

The numbers at the end of each reference indicate the pages on which the reference
occurs
326
References
Sz JOHN E. WETZEL: A simplicial

3-arrangement of 22 planes, Geometriae Dedicata 21 (1986), 257264. (225)
[8] GERALD L. ALEXANDERSON
simplicial
4-arrangement of 33 planes, Geometriae Dedicata 24 (1987), 245254. (225)
[9] GERALD L. ALEXANDERSON Si JOHN E. WETZEL: A
[10]
ALEXANDER
DANILOWITSCH ALEXANDROW:
Konvexe Po/yeder,
Akademie-Verlag, Berlin 1955. (126)

The number of polytopes, configurations and real matroids, Mathematika 33 (1986), 62-71. (76)
[11] NOGA ALON:
A simple proof of the upper bound theorem, European J. Math. 6 (1985), 211-214. (277)
[12] NOGA ALON Si GIL KALAI:
[13] Amos AursHuLER: Polyhedral realizations in I of triangulations

of the torus and 2-manifolds in cyclic 4-polytopes, Discrete Math. 1
(1971), 211-238. (148)
[14] Amos ALTSHULER, JRGEN BOKOWSKI Sz PETER SCHUCHERT:
Spatial polyhedra without diagonals, Israel J. Math. 86 (1994), 373396. (148)
[15] Amos ALTSHULER, JRGEN BOKOWSKI & LEON STEINBERG: The
classification of simplicial 3-spheres with nine vertices into polytopes
and nonpolytopes, Discrete Math. 31 (1980), 115-124. (144, 212)
M.
Shadows and slices of polytopes, in: Proceedings of the 12th Annual ACM Symposium on Computational Geometry, ACM Press 1996, 10-19. (290)
[16] NINA AMENTA Sz GNTER
ZIEGLER:
[17] NINA AMENTA & GUNTER M. ZIEGLER: Deformed products and
maximal shadows, Preprint 502/1996, TU Berlin, March 1996,

32 pages; in: Proc. "Discrete and Computational Geometry: Ten
Years Later," Mt. Holyoke, July 1996 (B. Chazelle, J.E. Goodman,
R. Pollack, eds.), Contemporary Mathematics, Amer. Math. Soc., to
appear. (71, 290)
Combinatorics of Finite Sets, Oxford Science Publications, Clarendon Press, Oxford University Press, Oxford 1987.
(263, 278)
[18] IAN ANDERSON:
[19] E. M. ANDREEV: On convex polyhedra in Lobadevslai spaces, Math.

of the USSR Sbornik 10 (1970), 413-440. (117)
Ein Algorithmusansatz fuir Realisierungsfragen im Ed getestet an kombinatorisehen 3-Sphdren, Staatsexamensarbeit, Universita Bochum 1982. (212)
[20] CHRISTOPH ANTONIN:
References
[21]
327
DAVID APPLEGATE, ROBERT BIXBY, VAEK CHVATAL & WILLIAM
Finding cuts in the TSP (a preliminary report), preprint,

August 1994, 57 pages. (23)
COOK:
Links and nonshellable cell partitionings

of S 3 , Proc. Amer. Math. Soc. 118 (1993), 635-639. (277)
[22] STEVE ARMENTROUT:
A criterion for the affine equivalence of cell

complexes in R d and convex polyhedra in d+1 Discrete Comput.
Geometry 2 (1987), 49-64. (117)
[23] FRANZ AURENHAMMER:
[24] DAVID AVIs: A C implementation of the reverse search vertex enumeration algorithm, preprint, McGill University 1992; report and
code available by anonymous ftp from mutt .cs.mcgill ca, directory "pub/C". (49)
.
[25] DAVID AVIS & DAVID BREMNER: How good are convex hull algo-
rithms? in: "11th Annual ACM Symp. Comput. Geometry" 1995,

20-28. (49)
How good are
convex hull algorithms? Computational Geometry: Theory and Applications 7 (1997), 265-301. (49)
[26] DAVID Avis, DAVID BREMNER & RAIMUND SEIDEL:
[27]
A pivoting algorithm for convex

hulls and vertex enumeration of arrangements and polyhedra, Discrete
Comput. Geometry 8 (1992), 295-313. (48)
DAVID AVIS & KOMEI FUKUDA:
[28] ERIC K. BABsON &
Lows J.
BILLERA: The geometry of products
of minors, Preprint 24 pages; Discrete Comput. Geometry (1998), to

appear. (324)
[29] ERIC K. BABSON, Louis J. BILLERA & CLARA S. CHAN: Neigh-
borly cubical spheres and a cubical lower bound conjecture, Preprint

1996; Israel J. Math. (1998), to appear. (98, 280)
Convexity and optimization in discrete structures,
in: "Convexity and Its Applications" (P. M. Gruber and J. Wills,
eds.), Birkhauser, Basel 1983, pp. 9-29. (183)
[30] ACHIM BACHEM:
&
Linear Programming Duality.

An Introduction to Oriented Matroids, Universitext, Springer-Verlag,
Berlin 1992. (183, 216)
[31] ACHIM BACHEM
WALTER KERN:
[32] EMZARDI BALADZE: Solution of the Szaefalvi-Nagy problem for a
class of convex polytopes, Geometriae Dedicata 49 (1994), 25-38.

(226)
[33] MICHEL L. BALINsKI: On the graph structure of convex polyhedra in
n-space, Pacific J. Math. 11 (1961), 431-434. (95)
328
References
[34] IMRE BARANY 86 I.ASZIA5 LOVSZ: Borsuk's theorem and the number
of facets of centrally symmetric polytopes, Acta Mathematica Acad.

Sci. Hung. 40 (1982), 323 329. (289)
-
Diagrams and Schlegel diagrams, in: "Combinatorial Structures and Their Applications" (Proc. Calgary Internat. Conference, Calgary 1969), Gordon and Breach, New York 1970,
pp. 1-4. (143)
[35] DAVID W. BARNETTE:
[36] DAVID W. BARNETTE: Wu paths on 3-polytopes, J. Combinatorial

Theory 7 (1969), 62-70 (85, 87)
[37] DAVID W. BARNETTE: Projections of 3-polytopes, Israel J. Math. 8
(1970), 304-308. (114)
[38] DAVID W. BARNETTE: The minimum number of vertices of a simple
polytope, Israel J. Math. 10 (1971), 121-125. (271, 279)
The triangulations of the 3-sphere with up to

8 vertices, J. Combinatorial Theory 14 (1973), 37-42. (144)
[40] DAVID W. BARNETTE: A proof of the lower bound conjecture for

convex polytopes, Pacific J. Math. 46 (1973), 349-354. (271, 279)
[41] DAVID W. BARNETTE: Map Coloring, Polyhedra, and the Four Color
Theorem, Dolciani Mathematical Expositions No. 8, Mathematical

Association of America, Washington DC 1983. (22, 103)
[42] DAVID W. BARNETTE: Preassigning the shape of projections of convex polytopes, J. Combinatorial Theory, Ser. A 42 (1986), 293-295.
(122)
[43] DAVID W. BARNETTE: Two "simple" 3-spheres, Discrete Math. 67
(1987), 97 99. (139, 141, 142, 174)
-
An impediment to polyhedrality, J. Combinatorial Theory Ser. A 48 (1988), 259 265. (148)
A 2-manifold of genus 8 without the Wvproperty, Geometriae Dedicata 46 (1993), 211 214. (85)
A short proof for the d-connectedness of

d-polytopes, Discrete Math. 137 (1995), 351-352. (95)
[47] DAVID W. BARNETTE
& BRANKO GRNBAUM: On Steinitz' theorem
concerning convex 3-polytopes and on some properties of 3-connected

graphs, in: "The Many Facets of Graph Theory," Lecture Notes in
Mathematics 110, Springer-Verlag, Berlin Heidelberg 1969, pp. 2740. (104)
[48] DAVID W. BARNETTE & BRANKO GRNBAUM: Preassigning the
shape of a face, Pacific J. Math. 32 (1970), 299-302. (114, 144)
References
329
[49] DAVID W. BARNETTE, PETER KLEINSCHMIDT & CARL W. LEE:
An upper bound theorem for polytope pairs, Math. Operations Research 11 (1986), 451-464. (279)
Hamiltonian circuits
in certain prisms, Discrete Math. 5 (1973), 389-394. (100)
[50] DAVID W. BARNETTE & MOSHE ROSENFELD:
[51] HANS G. BARTELS: A priori Informationen zur Linearen Programmierung. Ober Ecken und Hyperfliichen auf Polyedern, Beitrage zur
Datenverarbeitung und Unternehmensforschung, Band 4, Verlag Anton Hain, Meisenheim am Glan 1973. (124)
[52] GOTTFRIED BARTHEL, FRIEDRICH HIRZEBRUCH & THOMAS HoFER, Geradenkonfigurationen und Algebraische Fleichen, Aspekte der
Mathematik D4, Vieweg, Wiesbaden 1987. (225)
[53] ALEXANDER
I. BARVINOK: On equivariant generalization of Dehn-
Sommerville equations, European J. Combinatorics 13 (1992), 419

428. (280)
[54] DAVID BAYER & IAN MOFtRISON: Standard bases and geometric in-
variant theory, I. Initial ideals and state polytopes, J. Symbolic Computation 6 (1988), 209 217. (320)
-
[55] MARGARET M. BAYER: The extended f vectors of 4 polytopes, J.

Combinatorial Theory, Ser. A 44 (1987), 141 151. (280, 288)
-
[56] MARGARET M. BAYER: Equidecomposable and weakly neighborly

polytopes, Israel J. Math. 81 (1993), 301 320. (281)
-
[57] MARGARET
M. BAYER: Face numbers and subdivisions of poly-
topes, in: "Polytopes: Abstract, Convex and Computational"

(T. Bisztriczky, P. McMullen, and A. Weiss, eds.), Proc. NATO Advanced Study Institute, Toronto 1993, Kluwer Academic Publishers
1994, pp. 155-172. (280, 281)
M. BAYER & LOUIS J. BILLERA: Generalized DehnSommerville relations for polytopes, spheres and Eulerian partially
ordered sets, Inventiones Math. 79 (1985), 143-157. (260, 280, 287)
[58] MARGARET
M. BAYER & CARL W. LEE: Combinatorial aspects of

convex polytopes, in: "Handbook of Convex Geometry" (P. Gruber
[59] MARGARET
and J. Wills, eds.), North-Holland, Amsterdam 1993, pp. 485-534.

(vi, 1, 22, 183, 279, 288)
[60] MARGARET M. BAYER & ANDREW KLAPPER: A new index for
polytopes, Discrete Comput. Geometry 6 (1991), 33 47. (280)
-
[61] EBERHARD BECKER: On the real spectrum of a ring and its application to semialgebraic geometry, Bulletin Amer. Math. Soc. 15 (1986),
19-60. (119)
330
References
On Crum 's problem, J. London Math. Soc. 22 (1947), 285-287. (144)
[62]
ABRAM SAMOILOVICH BESICOVITCH:
[63]
Si HANFRIED LENZ: Design Theory, B.I. Wissenschaftsverlag, Bibliographisches Institut,

Mannheim 1985 and Cambridge University Press, Cambridge 1986.
(284)
THOMAS BETH, DIETER JUNGNICKEL
M.
Bander
und Maiusbander in konvexen Polytopen, Abh. Math. Sem. Univ.
Hamburg 44 (1975), 249-262. (148)
[64] ULRICH BETKE, CHRISTOPH SCHULZ & J6RG
WILLS:
[65] Louis J. BILLERA: Homology of smooth splines: generic triangulations and a conjecture of Strang, Transactions Amer. Math. Soc. 310
(1988), 325-340. (129)
[66] Louis J. BILLERA, CLARA CHAN Az NIANDONG Liu: Flag complexes, labelled rooted trees, and star shellings, preprint 1996. (280)
[67] Louis J. BILLERA, PAUL FILLIMAN Sz BERND STURMFELS: Constructions and complexity of secondary polytopes, Advances in Math.
83 (1990), 155-179. (305)
[68] Louis J. BILLERA, IZRAIL M. GEL'FAND & BERND STURMFELS:
Duality and minors of secondary polyhedra, J. Combinatorial Theory,
Ser. B 57 (1993), 258-268. (305)
[69] Louis J. BILLERA, MIKHAIL M. KAPRANOV & BERND STURMFELS:
Cellular strings on polytopes, Proc. Amer. Math. Soc. 122 (1994),
549-555. (225, 296)
[70] Louis J. BILLERA & CARL W. LEE: Sufficiency of McMullen's conditions for f -vectors of simplicial polytopes, Bulletin Amer. Math.
Soc. 2 (1980), 181-185. (269, 272, 278)
[71] Louis J. BILLERA & CARL W. LEE: A proof of the sufficiency of
McMullen's conditions for f -vectors of simplicial polytopes, J. Combinatorial Theory, Ser. A 31 (1981), 237-255. (269, 278, 279)
[72] Louis J. BILLERA & A. SARANGARAJAN: All 0-1 polytopes are traveling salesman polytopes, Combinatorica 16 (1996), 175-188. (21, 26,
70)
[73] Louis J. BILLERA Si BETH SPELLMAN MUNSON: Polarity and inner
products in oriented matroids, European J. Combinatorics 5 (1984),
293-308. (184)
[74] Louis J. BILLERA Si BERND STURMFELS: Fiber polytopes, Annals
of Math. 135 (1992), 527-549. (19, 291, 298, 301, 321, 323, 324)
[75] Louis J. BILLERA Si BERND STURMFELS: Iterated fiber polytopes,
Mathematika 41 (1994), 348-363. (19, 291, 294, 321)
References
331
[76] R H BING: An alternate proof that 3-manifolds can be triangulated,

Annals of Math. 69 (1959), 37-65. (144, 238)
[77] R H BING: Some aspects of the topology of 3-manifolds related to
the Poincar conjecture, in: "Lectures on Modern Mathematics II"
(T. L. Saaty, ed.), Wiley, New York 1964, pp. 93-128. (276, 277)
[78] R H BING: The Geometric Topology of 3-Manifolds, AMS Colloquium Publications, Vol. 40, Amer. Math. Soc., Providence RI 1983.
(238)
[79] GARRETT BIRKHOFF: Tres observaciones sobre el algebra lineal, Re-
vista Facultad de Ciencias Exactas, Puras y Applicadas Uni versidad

Nacional de Tucuman, Serie A (Matematicas y Fisica Teoretica) 5
(1946), 147-151. (20)
[80] ANDERS BJ6RNER: The unimodality conjecture for convex polytopes,
Bulletin Amer. Math. Soc. 4 (1981), 187-188. (269-272, 279)

[81] ANDERS BJ6RNER: Posets, regular CW complexes and Bruhat order,
European J. Combinatorics 5 (1984), 7-16. (233, 234)

[82] ANDERS B.K5RNER: Face numbers of complexes and polytopes, Pro-
ceedings of the International Congress of Mathematicians, Berkeley

CA, 1986, 1408-1418. (269, 278, 279, 288)
[83] ANDERS BJ8RNER: Homology and shellability of mat voids and geo-
metric lattices, in: Matroid Applications (N. White, ed.), Cambridge

University Press, Cambridge 1992, pp. 226-283. (275)
[84] ANDERS BJ6RNER: Essential chains and hornotopy type of posets,
Proc. Amer. Math. Soc. 402 (1992), 1179-1181. (296)

[85] ANDERS BJaRNER: Topological methods, in: "Handbook of Combi-
natorics" (R. Graham, M. Griitschel, and L. LovAsz, eds.), NorthHolland/Elsevier, Amsterdam 1995, pp. 1819-1872. (59, 232)
[86] ANDERS B.JaRNER: Partial unimodality for f -vectors of simplicial
polytopes and spheres, in: "Jerusalem Combinatorics '93" (H. Barcelo

and G. Kalai, eds.), Contemporary Mathematics 178, Amer. Math.
Soc. 1994, 45-54. (269-272, 279, 288)
Face numbers, Betti numbers and depth, in
preparation, 1994. (284)
[87] ANDERS BJ6RNER:
[88] ANDERS BJ6RNER & KIMMO ERIKSSON: Extended shellability for
rank 3 matroid complexes, Discrete Math. 132 (1994), 373-376. (287)

[89] ANDERS BJ6RNER, PETER FRANKL Sz RICHARD P. STANLEY: The
number of faces of balanced Cohen-Macaulay complexes and a generalized Macaulay theorem, Combinatorica 7 (1987), 23-34. (267, 268,
287)
332
References
[90] ANDERS BJ45RNER, ADRIANO M. GARSIA Sc RICHARD P. STANLEY: An introduction to Cohen-Macaulay partially ordered sets, in:
"Ordered Sets" (I. Rival, ed.), D. Reidel, Dordrecht 1982, pp. 583615. (247)
[91] ANDERS B.16RNER Sc GIL KALAI:
An extended Euler-Poincar the-
orem, Acta Math. 161 (1988), 279-303. (278)

[92] ANDERS 1116RNER, MICHEL LAS VERGNAS, BERND STURMFELS,
NEIL WHITE Sz GNTER M. ZIEGLER: Oriented Matroids, Ency-
clopedia of Mathematics, Vol. 46, Cambridge University Press, Cambridge 1993. (18, 59, 96, 142, Lectures 6 and 7, 233, 275, 277, 291)
[93] ANDERS BJ6RNER Si SVANTE LINUSSON: The number of
k-faces of
a simple d-polytope, Preprint 1996; Discrete Comput. Geometry, to

appear. (279)
Sc MICHELLE L. WACHS: On lexicographically

shellable posets, Transactions Amer. Math. Soc. 277 (1983), 323-341.
[94] ANDERS BJ6RNER
(233, 235, 240, 275)

[951 ANDERS BJ6RNER Sc MICHELLE L. WACHS: Shellable nonpure complexes and posets, I, Transactions Amer. Math. Soc. 348 (1996),
1299-1327. (233)
[96] ROBERT G. BLAND Sc MICHEL LAS VERGNAS: Orientability of matroids, J. Combinatorial Theory, Ser. B 24 (1978), 94-123. (149)
[971 WILHELM BLASCHKE: Voriesungen liber Differentialgeometrie. II:
Affine Differentialgeometrie, 1. und 2. Auflage, Berlin 1923. (224)
[98] GERD BLIND
Sc
ROSWITHA BLIND: Convex
polytopes without trian-
gular faces, Israel J. math. 71 (1990), 129-134. (280, 290)

[99] GERD BLIND Sc ROSWITHA BLIND: The
semiregular polytopes, Cam-
mentarii Math. Helveticae 66 (1991), 150-154. (9)

[100] GERD BLIND Si ROSWITHA BLIND: Gaps
in the numbers of vertices
of cubical polytopes I, Discrete Comput. Geometry 11 (1994), 351356. (280, 289)

[101] GERD BLIND Sc ROSWITHA BLIND: The cubical
d-polytopes with less
than 2d+1 faces, in: "The Litsz16 Fejes T6th Festschrift" (I. BdrAny,
J. Pach, eds.), Discrete Comput. Geometry 13 (1995), 321-345. (280)
[102] ROSWITHA BLIND Sc PETER MANI-LEVITSKA: On puzzles
and poly-
tope isomorphisms, Aequationes Math. 34 (1987), 287-297. (93)

[103] JACEK BOCHNAK, MICHEL COSTE 48, 6 MARIE-FRANOISE ROY:
Gomtrie Algbrique Relle, Ergebnisse der Mathematik und ihrer

Grenzgebiete, 3. Folge, Band 12, Springer-Verlag, Berlin Heidelberg
1987. (119)
References
[104]
333
Eine kombinatorische Analyse zonotopaler Raumaufteilungen, Ph.D. dissertation, Bielefeld 1992; Preprint 92-041, Sonderforschungsbereich 343 "Diskrete Strukturen in der Mathematik,"
Universitat Bielefeld 1992, 100 pages. (220, 222, 225)
JOCHEN BOHNE:
Oriented matroids, in: "Handbook of Convex

Geometry" (P. Gruber and J. Wills, eds.), North-Holland, Amster-
[105] JURGEN BOKOWSKI:
dam 1993, pp. 555-602. (183, 212, 224)

[106]
JRGEN BOKOWSKI & ANSELM EGGERT:
All realization of Mbius'
torus with 7 vertices, Structural Topology 17 (1991), 59-78. (148)

[107] JURGEN BOKOWSKI, GUNTER EWALD
Si
PETER KLEINSCHMIDT:
On combinatorial and affine automorphisms of polytopes, Israel J.

Math. 47 (1984), 123-130. (177)
[108]
Simplicial convex 4-polytopes do not have the isotopy property, Port ugaliae
Math. 47 (1990), 309-318. (177)
JRGEN BOKOWSKI & ANTONIO GUEDES DE OLIVEIRA:
& IDO SHEMER: Neighborly 6-polytopes with 10

vertices, Israel J. Math. 58 (1987), 103-124. (183)
[109] JURGEN BOKOWSKI
Polytopal and nonpolytopal spheres An algorithmic approach, Israel J. Math. 57 (1987),

257-271. (144, 212)
[110]
JRGEN BOKOWSKI & BERND STURMFELS:
[111]
JRGEN BOKOWSKI &
BERND STURMFELS:
Computational Syn-
thetic Geometry, Lecture Notes in Mathematics 1355, SpringerVerlag, Berlin Heidelberg 1989. (144, 179, 182, 183, 212)
[112] ETHAN D. BOLKER:
A class of convex bodies, Transactions Amer.
Math. Soc. 145 (1969), 323-345. (224, 226)

[113] ETHAN D. BOLKER:
Centrally symmetric polytopes, in: "Proc. of the
Twelfth Biannual Intern. Seminar of the Canadian Math. Congress"

(Vancouver 1969), Time Series and Stochastic Processes; Convexity
and Combinatorics (R. Pyke, ed.), Canadian Math. Congress, Montreal 1970, pp. 255-263. (200)
[114]
BLA BOLLOBS:
Graph Theory, Graduate Texts in Mathematics
63, Springer-Verlag, New York 1979. (80)

[115]
& ARNE BRONDSTED: A dual proof of the upper

bound conjecture for convex polytopes, Math_ Scand. 46 (1980), 95-
AAGE BONDESEN
102. (277)
[116] JOHN ADRIAN BONDY
& U. S. R.
MURTY:
Graph Theory with
Applications, MacMillan and American Elsevier, New York, 1976.

(80)
[117] TOMMY BONNESEN & WERNER FENCHEL: Theorie der konvexen
Kiirper, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 3,
334
References
Springer-Verlag, Berlin 1934; berichtigter Reprint, Springer-Verlag,
Berlin 1974; English Translation: Theory of Convex Bodies, BCS Associates Pub., Moscow, Idaho, 1987 (22)
[118] KARL HEINZ BORGWARDT:
The Simplex Method. A Probabilistic

Analysis, Algorithms and Combinatorics 1, Springer-Verlag, Berlin
Heidelberg 1987. (96, 301)
[119] KARL HEINZ BORGWARDT: Average complexity of a gift-wrapping
algorithm for determining the convex hull of randomly given points,

Discrete Comput. Geometry 17 (1997), 79-109. (48)
[120]
Groupes et Algbres de Lie, Chap. 4,5, et 6,

Hermann, Paris 1968. (225)
[121]
Kreispackungen und Triangulierungen, Enseignement Math. 38 (1992), 201-217. (118)
NICOLAS BOURBAKI:
WALTER
BRAGGER:
[122] ULRICH BREHM: A nonpolyhedral triangulated Mbius strip, Pro-
ceedings Amer. Math. Soc. 89 (1983), 519-522. (148)

[123] ULRICH BREHM & J6RG M. WILLS:
Polyhedral manifolds, in:

"Handbook of Convex Geometry" (P. Gruber and J. Wills, eds.),
North-Holland, Amsterdam 1993, pp. 535-554. (144)
[124]
Incremental convex hulls are not output sensitive,

Preprint 1996; Discrete Comput. Geometry, to appear. (49)
[125]
R. BRIGHTWELL & EDWARD R. SCHEINERMAN: Representations of planar graphs, SIAM J. Discrete Math. 6 (1993), 214-229.
(118)
DAVID BREMNER:
GRAHAM
[126] ARNE BRONDSTED: An Introduction to Convex Polytopes, Graduate
Texts in Mathematics 90, Springer-Verlag, New York-Berlin 1983.

(1, 22, 27, 277)
[127]
A new proof of the

d-connectedness of d-polytopes, Canadian Math. Bull. 32 (1989), 252254. (95)
ARNE BRONDSTED &
GEORGE MAXWELL:
[128] KENNETH S. BROWN: Buildings, Springer-Verlag, New York 1989.
(9, 225)
[129]
A. BRUALDI & PETER M. GIBSON: Convex polyhedra of

doubly stochastic matrices, I. Applications of the permanent function, J. Combinatorial Theory, Ser. A 22 (1977), 194-230; II. Graph
of S-2 71 , J. Combinatorial Theory, Ser. B 22 (1977), 175-198; III. Affine
and combinatorial properties of nn , J. Combinatorial Theory, Ser. A
22 (1977), 338-351, W. Linear Algebra Appl. 15 (1976), 153-172.
(20)
[130]
Ober die Ableitung der allgemeinen

Polytope und die nach Isomorphismus verschiedenen Typen der allge-
RICHARD
(JOHANNES) MAX BRUCKNER:
335
References
meinen Achtzelle, Verhand. Konink. Akad. Wetenschap,

10, No. 1 (1910). (143, 144)
Erste Sec tie
[131] (JOHANNES) MAX BRCKNER: Vielecke und Vielflache. Ihre Theorie

und Geschichte, Leipzig 1900. (22)
[132] HEINZ BRUGGESSER 86 PETER MANI: Sheliable decompositions of
cells and spheres, Math. Scand. 29 (1971), 197-205. (233, 235, 240,
275)
[133] EWALD BURGER: Ober homogene lineare Ungleichungssysteme, Zeit-
schrift Angewandte Math. Mechanik 36 (1956), 135 139. (48, 74)

-
[134] GEOFFREY R. BURTON: The non-neighbourliness of centrally sym-
metric convex polytopes having many vertices, J. Combinatorial Theory, Ser. A 58 (1991), 321 322. (280)
-
Ober den Variabilitiitsbereich der

Fourier'schen Konstanten von positiven harmonischen Funktionen,
Rendiconto del Circolo Matematico di Palermo 32 (1911), 193 217;
reprinted in: "Constantin Carathodory, Gesammelte Mathematische
Schriften, Band III" (H. Tietze, ed.), C. H. Beck'sche Verlagsbuchhandlung, Miinchen 1955, pp. 78-110. (75)
[135] CONSTANTIN CARATHODORY:
[136] JOHN WILLIAM SCOTT CASSELS: An Introduction to the Geome-
try of Numbers, Grundlehren der mathematischen Wissenschaften,

Band 99, Springer-Verlag, Berlin 1959. (23)
[137] GERBRAND CEDER, GERARDO D. GARBULSKY, DAVID AVIS Si
KOMEI FUKUDA: Ground states of a ternary fcc lattice model with
nearest- and next-nearest-neighbor interactions, Physical Review B.
Condensed Matter 49 (1994), 1 7. (49)
-
[138] CLARA CHAN: On subdivisions of simplicial complexes: Characteriz-
ing local h-vectors, Discrete Comput. Geometry 11 (1994), 465 476.

(281)
-
[139] CLARA CHAN, DOUGLAS JUNGREIS & RICHARD STONG: Buchsbaum and Eulerian complexes, J. Pure Appl. Algebra, to appear.
(260)
[140] BERNARD CHAZELLE: An optimal convex hull algorithm in any fixed
dimension, Discrete Comput. Geometry 10 (1993), 377 409. (48)

-
[141] N. V. CHERNIKOVA: Algorithm for finding a general formula for the
non-negative solutions of a system of linear equations, U.S.S.R. Computational Mathematics and Mathematical Physics 4 (1964), 151
158. (48)
[142] N. V. CHERNIKOVA: Algorithm for finding a general formula for
the non-negative solutions of a system of linear inequalities, U.S.S.R.

Computational Mathematics and Mathematical Physics 5 (1965),
228-233. (48, 73, 74)
336
References
[143] THOMAS CHRISTOF: Ein Verfahren zur Transformation zwischen
Polyederdarstellungen, Diplomarbeit, Universitat Augsburg 1991, 79

pages, plus manual and listing. (11, 48)
[144] THOMAS CHRISTOF: PORTA - A Polyhedron Representation Transformation Algorithm, version 1.3.1, available at
http : //www. iwr . uni-he idelberg .de/iwr/comopt/
soft/PORTA/readme . html and at
ftp: //elib .zib .de/pub/mathprog/polyth/index . html (11, 48)
[145]
THOMAS CHRISTOF:
SMAPO Library of linear descriptions of
small problem instances of polytopes in combinatorial optimization,

http://www.iwr.uni-heidelberg.de/iwr/comopt/
soft/SMAPO/SMAPO.html (49)
[146] THOMAS CHRISTOF, MICHAEL JNGER & GERD REINELT:
A com-
plete description of the traveling salesman polytope on 8 nodes, Operations Research Letters 10 (1991), 497-500. (49)
[147]
THOMAS CHRISTOF & GERHARD REINELT: Combinatorial optimiza-
tion and small polytopes, Top (Spanish Statistical and Operations

Research Society) 4 (1996), 1-64. (49)
[148] THOMAS CHRISTOF & GERHARD REINELT: Efficient parallel facet
enumeration for 011-polytopes, Preprint, IWR Heidelberg, March
1997, 16 pages. (26)
[149]
MAREK CHROBAK, MICHAEL T. GOODRICH & ROBERTO TAMASSIA: Convex drawings of graphs in two and three dimensions, in: "12th
ACM Symp. Computational Geometry", ACM Press 1996, pp. 319328. (123)
[150]
Convex grid drawings of 3connected planar graphs, Preprint 1993; Int. J. Comput. Geometry
and Applications, to appear. (123)
[151]
VAEK CHVTAL: Linear Programming, W. H. Freeman, New York
MAREK CHROBAK & GOOS KANT:
1983. (82, 96, 143)

[152] KENNETH L. CLARKSON: A bound on local minima of arrangements
that implies the upper bound theorem, Discrete Comput. Geometry
10 (1993), 427-433. (278)
[153] GEORGE F. CLEMENTS & BERNT LINDSTRI5M: A generalization of
a combinatorial theorem of Macaulay, J. Combinatorial Theory 7
(1969), 230-238. (263)
pour les empilements de cercles, Inventiones Math. 104 (1991), 655-669. (118)
[155] ROBERT CONNELLY & DAVID W. HENDERSON: A convex 3-complex
not simplicially isomorphic to a strictly convex complex, Math. Proc.
Cambridge Phil. Soc. 88 (1980), 299-306. (275)
[154]
YVES COLIN DE VERDIRE: Un principe variationnel
References
337
[156] HAROLD SCOTT MACDONALD COXETER: Regular Polytopes, second

edition, Macmillan, New York 1963; corrected reprint, Dover, New
York 1973. (9, 120, 201)
[157] HENRY CRAPO Si WALTER WHITELEY: Statics of frameworks and
motions of panel structures, Structural Topology 6 (1982), 42-82.

(116, 143)
[158] HENRY CRAPO Si WALTER WHITELEY: Plane self stresses and pro-
jected polyhedra I: The basic pattern, Structural Topology 20 (1993),

55-78. (116, 143)
[159] HENRY CRAPO Si WALTER WHITELEY: 3-Stresses in 3-space and
projections of polyhedral 3-surfaces: Reciprocals, liftings and parallel

configurations, preprint, York University 1993, 40 pages. (143)
[160] HALLARD T. CROFT, KENNETH T. FALCONER Sz RICHARD K.
GUY: Unsolved Problems in Geometry, Springer-Verlag, New York
1991. (1, 126, 144)
[161]
CSSZikR: A polyhedron without diagonals, Acta Sci. Math.

(Szeged) 13 (1949/50), 140. (148)
AKOS
[162] JOHN DALBEC: Geometry and Combinatorics of Chow Forms,

Ph.D. thesis, Cornell University 1995. (323)
[163] GOPAL DANARAJ Si VICTOR KLEE: Shellings of spheres and polytopes, Duke Math. J. 41 (1974), 443-451. (233, 240)
[164] GOPAL DANARAJ Si VICTOR KLEE: Which spheres are shellable?
in: "Algorithmic Aspects of Combinatorics" (B. Alspach et al., eds.),
Annals of Discrete Math. 2, 1978, pp. 33-52. (235, 275, 277)
[165] GOPAL DANARAJ Si VICTOR KLEE: A representation of 2-dimensional manifolds and its use in the design of a linear-time shelling
algorithm, Annals of Discrete Math. 2 (1978), 33-52. (281)
[166] GEORGE B. DANTZIG: Linear Programming and Extensions, Princeton University Press, Princeton 1963. (81, 83, 96)
[167] GEORGE B. DANTZIG Si B. CURTIS EAVES: Fourier-Motzkin elimination and its dual, J. Combinatorial Theory, Ser. A 14 (1973),
288-297. (36, 47)

[168] LUDWIG DANZER: To problem 8, in: "Proc. Colloquium on Convexity" (W. Fenchel, ed.), Copenhagen 1965, Kobenhavns Universitets
Matematiske Institut, Copenhagen 1967, pp. 312-313. (276)
[169] LUDWIG DANZER, BRANKO GRONBAUM Si VICTOR KLEE: Helly's
theorem and its relatives, in: "Convexity" (V. Klee, ed.), Proc. Symposia in Pure Mathematics, Vol. VII, Amer. Math. Soc., Providence
RI 1963, pp. 101-180. (144)
338
References
[1701 GAUTAM DAs 8,6 MICHAEL
T. GOODRICH: On the complexity of op-
timization problems for 3-dimensional convex polyhedra and decision

trees, Preprint, Dept. Computer Science, Johns Hopkins University
1996, 15 pages. (119)
[171] ROBERT J. DAVERMAN: Decompositions of Manifolds,
Press, Orlando 1986. (232)
Academic
[172] CHANDLER DAVIS: The set of non-linearity of a convex piecewiselinear function, Scripta Math. 24 (1959), 219-228. (143)
[173] MAx DEHN: Die Eulersche Formel im Zusammenhang mit dem Inhalt in der nicht-Euklidischen Geometrie, Mathematische Annalen
61 (1905), 561-568. (252)
[174] A. K. DEWDNEY 86 JOHN K. VRANCH: A convex partition of
with application to Crum 's problem and Knuth 's post-office problem,
Utilitas Math. 12 (1972), 193-199. (144)
[175) MICHEL MARIE DEZA AL MONIQUE LAURENT: Geometry of Cuts and
Metrics, Algorithms and Combinatorics 15, Springer-Verlag, Berlin
Heidelberg 1997. (23)
[176] TOMASZ DUBEJKO & KENNETH STEPHENSON: Circle packing: Experiments in discrete analytic function theory, Experimental Math.
4 (1995), 307-348. (118)
[171
MARTIN DYER, PETER GRITZMANN 8,6 ALEXANDER HUFNAGEL, On

the complexity of computing mixed volumes, preprint 1994; J. London
Math. Soc., to appear. (230)
[178] JRGEN ECKHOFF: Combinatorial properties of f -vectors of convex

polytopes, Unpublished manuscript, 1985; preprint 1994, in preparation. (272-274, 288)
[179] PAUL H. EDELMAN & VICTOR REINER: Visibility complexes and the
Banes problem for triangulations in the plane, preprint 1995; Discrete
Comput. Geometry, to appear. (296)
[180] HERBERT EDELSBRUNNER: Algorithms in Computational Geometry,
EATCS Monographs in Theoretical Computer Science 10, SpringerVerlag, Berlin 1987. (129, 146)
[181] JACK EDMONDS Sz ARNALDO MANDEL: Topology of oriented matroids, Ph.D. thesis of A. Mandel, University of Waterloo 1982, 333
pages. (216)
[182] FRITZ EHLERS: Eine Klasse komplexer Mannigfaltigkeiten und die
Aufliisung einiger isolierter Singularitiiten, Mathematische Annalen

218 (1975), 127-156. (277)
References
339
[183] MARKUS EIKELBERG: Zur Homologie torischer Varieteiten, Ph.D.
dissertation, Ruhr-Universitat Bochum, Bochum 1989, 76 Seiten.

(194, 277)
Sperner
Theory in Partially Ordered Sets, Teubner-Texte zur Mathematik,
Bd. 78, BSB B. G. Teubner Verlagsgesellschaft, Leipzig 1985. (278)
[184]
KONRAD ENGEL Si HANS DIETRICH O. F.
[185]
Sperner Theory, Encyclopedia of Mathematics 65,

Cambridge University Press 1997. (278)
GRONAU:
KONRAD ENGEL:
[186] G. V. EPIFANOV: Reduction of a plane graph to an edge by a startriangle transformation, Soviet Math. Doklady 7 (1966), 13-17. (109)
[187]
Lattice Point
Problems, Pitman Monographs and Surveys in Pure and Applied
Mathematics 39, Longman, Essex, and John Wiley Si Sons, New
York 1989. (23)
[188]
HERVA LE VERGE: Complete linear descriptions of small asymmetric traveling salesman polytopes, Discrete Applied Math. 62 (1995), 193-208. (49)
[189]
GUNTER EWALD:
[190]
GUNTER EWALD Sz KONRAD Voss:
[191]
GUNTER
PAUL ERDCIS, PETER GRUBER Si JOSEPH HAMMER:
REINHARDT EULER Si
Combinatorial Convexity and Algebraic Geometry,

Graduate Texts in Mathematics 168, Springer-Verlag New York 1996.
(1, 22, 139, 143, 171, 183, 194, 224, 278)
Konvexe Polyeder mit Symmetriegruppe, Commentarii Math. Helvetii 48 (1973), 137-150. (183)
EWALD,
PETER KLEINSCHMIDT, UDO PACHNER Si
Neuere Entwicklungen in der kombinatorischen

Konvexgeometrie, in: "Contributions to Geometry," Proc. Geometry
Symposium, Siegen 1978 (J. Tiilke and J. Wills, eds.), Birkhduser,
Basel 1979, pp. 131-163. (100, 148, 188, 277)
CHRISTOPH SCHULZ:
[192]
Stellar subdivisions
of boundary complexes of convex polytopes, Mathematische Annalen
210 (1974), 7-16. (97)
GUNTER EWALD Si GEOFFREY C. SHEPHARD:
[193] IsTvAN FRY: On straight line representations of planar graphs,

Acta Sci. Math. (Szeged) 11 (1948), 229-233. (120)
[194]
JIM LAWRENCE: Oriented matroids, J. Combinatorial Theory, Ser. B 25 (1978), 199-236. (149, 211, 216)
JON FOLKMAN Si
[195] HUBERT DE FRAYSSEIX, PATRICE OSSONA DE MENDEZ St JANOS

PACH: Representations of planar graphs by segments, ALCOM Technical Report ALCOM II-031, 1993; in: "Intuitive Geometry," Collo-
quia Mathematica Societatis Jdnos Bolyai, to appear. (125)

[196] HUBERT DE FRAYSSEIX, JANOS PACH Si RICHARD POLLACK:
How
to draw a graph on a grid, Combinatorica 10 (1990), 41-51. (123)
340
References
[197] ZOLTAN FREDI: Matchings and covers in hypergraphs, Graphs and

Combinat orics 4 (1988), 115 206. (278)
-
[198] KOMEI FUKUDA: CDD A C implementation of the double description method, available by anonymous ftp from ftp . epf 1 . ch, directory "incoming/dma". (48)
-
[199] KOMEI FUKUDA, SHIGEMASA SAITO, AKIHISA TAMURA &

TAKESHI TOKUYAMA: Bounding the number of k-faces in arrangements of hyperplanes, Discrete Appl. Math. 31 (1991), 151 165.
(230)
-
[200] KOMEI FUKUDA, AKIHISA TAMURA & TAKESHI TOKUYAMA: A theorem on the average number of subfaces in arrangements and oriented
matroids, Geometriae Dedicata 47 (1993), 129 142. (230)
-
[201] WILLIAM F. FULTON: Introduction to Toric Varieties, Annals of

Math. Studies 131, Princeton University Press, Princeton 1993. (224,
278)
[202] ROBERT FURCH: Zur Grundlegung der kombinatorischen Topologie,
Abh. Math. Sem. Hamb. Univ. 3 (1924), 69 88. (276)
-
[203] ANDREI M. GABRILOV, IZRAIL M. GEL'FAND & MARK V. LOSIK:

Combinatorial computation of characteristic classes, Functional
Analysis Appl. 9 (1975), 103-115. (20)
[204] BERND GRTNER & GNTER M. ZIEGLER: Randomized simplex algorithms on Klee Minty cubes, in: Proc. 35th Annual "Symposium on
Foundations of Computer Science" (FOCS), IEEE Computer Society
Press, Los Alamitos CA, 1994, pp. 502-510; also:
BERND GRTNER, MARTIN HENK & GNTER M. ZIEGLER: Randomized simplex algorithms on Klee-Minty cubes, preprint, TU
Berlin, May 1996, 21 pages. (99, 290)
-
[205] DAVID GALE: Neighboring vertices on a convex polyhedron, in:

"Linear Inequalities and Related Systems" (H. W. Kuhn and
A. W. Tucker, eds.), Annals of Math. Studies 38, Princeton University Press, Princeton 1956, pp. 255-263. (144, 183)
[206] DAVID GALE: Neighborly and cyclic polytopes, in: "Convexity"
(V. Klee, ed.), Proc. Symposia in Pure Mathematics, Vol. VII Amer.
Math. Soc., Providence RI 1963, pp. 225-232. (14)
[207] DAVID GALE: On the number of faces of a convex polytope, Canadian
J. Math. 16 (1964), 12-17. (254)
[208] MICHAEL R. GAREY Si DAVID S. JOHNSON: Computers and Intractability. A Guide to the Theory of NP-Completeness, W. H. Freeman, San Francisco 1979. (125, 126)
[209] LYNN E. GARNER: An Outline of Projective Geometry, North Holland, New York 1981. (69)
References
341
[210] IZRAIL M. GEL'FAND: General theory of hypergeometric functions,
Soviet Math. Doklady 33 (1986), 573-577. (129, 320)

[211] IZRAIL M. GELTAND, MARK GORESKY, ROBERT D. MACPHERSON
& VERA SERGANOVA: Combinatorial geometries, convex polyhedra
and Schubert cells, Advances in Math. 63 (1987), 301-316. (20)

[212] ISRAEL M. GELFAND, MIKHAIL M. KAPRANOV & ANDREI V.
ZELEVINSKY: Discriminants, Resultants, and Multidimensional De-
terminants, Birkhauser, Boston 1994. (306, 320-324)

[213] IZRAIL M. GELTAND, ANDREI V. ZELEVINSKIT & MIKHAIL M.
KAPRANOV: Newton polytopes of principal A-determinants, Soviet
Math. Doklady 40 (1990), 278-281. (19, 291, 298, 305, 306, 321)
[214] IZRAIL M. GELTAND, ANDREI V. ZELEVINSKIT Sz MIKHAIL M.
KAPRANOV: Discriminants of polynomials in several variables and
triangulations of Newton polyhedra, Leningrad Math. J. 2 (1991),

449-505. (19, 129, 298, 305, 321)
[215] JACOB E. GOODMAN & JANOS PACH: Cell decomposition of poly-
topes by bending, Israel J. Math. 64 (1988), 129-138. (324)

[216] JACOB E. GOODMAN Sz RICHARD POLLACK: A combinatorial ver-
sion of the isotopy conjecture, in: Proc. Conf. "Discrete Geometry and
Convexity," New York 1982, (J.E. Goodman, E. Lutwak, J. Malkevitch, and R. Pollack, eds.), Annals of the New York Academy of
Sciences 440 (1985), 12-19. (177)
[217] JACOB E. GOODMAN SE RICHARD POLLACK:
Upper bounds for

rd,
Discrete Comput. Geometry 1
configurations and polytopes in
(1986), 219-227. (76)
[218] JACOB E. GOODMAN & RICHARD POLLACK: New bounds on higher
dimensional configurations and polytopes, in: "Proc. Third Int. Conf.

Combinatorial Mathematics," (G.S. Bloom, R.L. Graham, and J.
Malkevitch, eds.), Annals of the New York Academy of Sciences 555
(1989), 205-212. (76)
[219] JACOB E. GOODMAN & RICHARD POLLACK: Allowable sequences
and order types in discrete and computational geometry, in: "New

Trends in Discrete and Computational Geometry" (J. Pach, ed.), Algorithms and Combinatorics 10, Springer-Verlag, Berlin Heidelberg
1993, pp. 103-134. (182)
[220] JACOB E. GOODMAN, RICHARD POLLACK
&
BERND STURMFELS:
The intrinsic spread of a configuration in R", Journal Amer. Math.

Soc. 3 (1990), 639-651. (66, 123)
[221] PIERRE GOOSSENS: Shelling pseudopolyhedra, Discrete Comput. Ge-
ometry 7 (1992), 207-215. (281)
342
References
[222] RONALD L. GRAHAM, DONALD E. KNUTH & OREN PATASHNIK:
Concrete Mathematics. A Foundation for Computer Science, Addison-Wesley, Reading, MA 1989; second edition 1994, in preparation.
(277, 285)
J.
Proof techniques in the

theory of finite sets, in: "Studies in Combinatorics" (0.-C. Rota, ed.),
MAA Studies in Math. 17, Mathematical Association of America,
Washington DC 1978, pp. 22-79. (261, 263, 278)
[223] CURTIS GREENE & DANIEL
KLEITMAN:
[224] PETER GRITZMANN & VICTOR KLEE: On the complexity of some
basic problems in computational convexity: I. Containment problems,

Discrete Math. 136 (1994), 129-174. (28)
[225] PETER GRITZMANN & VICTOR KLEE: Computational complexity
of inner and outer j-radii of polytopes in finite dimensional normed

spaces, Math. Programming 59 (1993), 163-213. (28)
[226] PETER GRITZMANN & BERND STURMFELS: Minkowski
addition
of polytopes: Computational complexity and applications to Griibner
bases, SIAM J. Discrete Math. 6 (1993), 246-269. (224)
[227]
Optimierungsmethoden I, Vorlesungsskripturn, Universitat Augsburg 1985. (47)
[228]
LovAsz & ALEXANDER SCHRIJVER:

Geometric Algorithms and Combinatorial Optimization, Algorithms
and Combinatorics 2, Springer-Verlag, Berlin Heidelberg 1988; second edition 1994. (23, 25, 96)
[229]
Polyhedral Theory,
Chapter 8 in: "The Traveling Salesman Problem" (E.L. Lawler, J.K.
Lenstra, A.H.G. Rinnooy Kan, and D.B. Schmoys, eds.), Wiley-Interscience Series in Discrete Mathematics and Optimization, John Wiley
Si Sons, Chichester New York 1985, 251-360. (22)
[230]
Ulysses 2000: In
search of optimal solutions to hard combinatorial problems, Preprint
SC 93-34, ZIB Berlin 1993, 27 pages. (23)
MARTIN GR6TSCHEL:
MARTIN GR6TSCHEL, LASZIA5
MARTIN GR8TSCHEL & MANFRED PADBERG:
MARTIN GIVSTSCHEL & MANFRED PADBERG:
[231] LARRY T. GROVE & CLARK T. BENSON: Finite Reflection Groups,
Graduate Texts in Mathematics 99, Springer-Verlag, New York 1971;

second edition 1985. (9)
[232] PETER M. GRUBER & C. GERRIT LEKKERKERKER: Geometry of
Numbers, North-Holland, Amsterdam 1987. (23)
[2331 BRANKO GRNBAUM: Diagrams and Schlegel diagrams, Abstract
625-112, Notices Amer. Math. Soc. 12 (1965), 578. (143)
[234] BRANKO GRONBAUM: Convex Polytopes, Interscience, London 1967;
revised edition (V. Klee and P. Kleinschmidt, eds.), Graduate Texts
References
343
in Mathematics, Springer-Verlag, in preparation. (vi, 1, 22, 27, 49,

69, 75, 87, 95, 96, 98, and in Lectures 4, 5, 6 and 8)
[235] BRANKO GRNBAUM:
The importance of being straight (?), in:
"Proc. of the Twelfth Biannual Intern. Seminar of the Canadian

Math. Congress," Time Series and Stochastic Processes; Convexity and Combinatorics (R. Pyke, ed.), (Vancouver 1969), Canadian
Math. Congress, Montreal 1970, pp. 243-254. (280)
hyperplanes, in: "Proc. Second Lousiana Conference on Combinatorics, Graph Theory and Computing" (R. C. Mullin et al., eds.), Louisiana State University, Baton
Rouge 1971, pp. 41-106. (225)
[236] BRANKO GRNBAUM: Arrangements of
[237] BRANKO GRNBAUM: Arrangements
and Spreads, Regional Confer-
ence Series in Mathematics No. 10, Amer. Math. Soc., Providence RI

1972. (211, 213, 217, 224, 225, 229)
Polytopal graphs, in: "Studies in Graph Theory, Part II" (D. R. Fulkerson, ed.), MAA Studies in Math. 12, Mathematical Association of America, Washington DC 1975, pp. 201-224.
(116)
[238] BRANKO GRNBAUM:
[239] BRANKO GRONBAUM: Geometry strikes again, Mathematics Maga-
zine 58 (1985), 12-17. (9)
Simplicial arrangements in projective 3-space, in: "Coxeter-Festschrift, Teil IV,"

Mitteilungen aus dem Mathem. Sem. Giessen 166 (1984), pp. 49-101.
(225)
[240] BRANKO GRNBAUM 8z GEOFFREY C. SHEPHARD:
[241] BRANKO GRNBAUM & GEOFFREY C. SHEPHARD: Some problems
on polyhedra, J. Geometry 29 (1987), 182-190. (124)

[242] BRANKO GRNBAUM & V. P. SREEDHARAN:
An enumeration of
simplicial 4-polytopes with 8 vertices, J. Combinatorial Theory 2

(1967), 437-465. (143, 144)
[243] HARALD GNZEL: The universal partition theorem
for oriented ma-
troids, Discrete Comput. Geometry 15 (1996), 121-145. (182)

[244] HARALD GNZEL: On
the universal partition theorem for 4-polytopes,
Preprint no. 64, RWTH Aachen, November 1995, 26 pages. (182)

[245]
Constructing the associahedron, unpublished manuscript, MIT 1984, 11 pages. (18, 304, 318)
[246]
A simple and relatively efficient triangulation of the

n-cube, Discrete Comput. Geometry 6 (1991), 287-289. (147)
MARK HAIMAN:
MARK HAIMAN:
Graphentheorie, 2. Auflage, Wissenschaftliche Buchgesellschaft, Darmstadt 1989. (114)
[247] RUDOLF HALIN:
344
References
[248] PETER L. HAMMER, BRUNO SIMEONE, THOMAS M. LIEBLING &

DOMINIQUE DE WERRA: From linear separability to unimodality: A
hierarchy of pseudo-Boolean functions, SIAM J. Discrete Math. 1

(1988), 174 184. (283)
-
[249] LARRY H. HARPER: Optimal numberings and isoperimetric problems

on graphs, J. Combinatorial Theory 1 (1966), 385 393. (263)
-
[250] 1. BEN-ARROYO HARTMAN, ILAN NEWMAN & RAN ZIV: On grid

intersection graphs, Discrete Math. 87 (1991), 41 52. (125)
-
[251] NORA HARTSFIELD & GERHARD RINGEL: Pearls in Graph Theory: A Comprehensive Introduction, Academic Press, San Diego 1990.
(120)
[252] TAKAYUKI HIBI: Algebraic Combinatorics on Convex Polytopes,

Carslaw Publications, Glebe, Australia 1992. (1, 22, 278)
[253] WINFRIED HOCHSTTTLER: Nested cones and onion skins, Applied
Math. Letters 6 (1993), 67-69. (98)
[254] WILLIAM V. D. HODGE & DANIEL PEDOE: Methods of Algebraic
Geometry, Cambridge University Press, Cambridge 1947/1952, paperback reprint 1968. (69, 182)
[255] CRAIG D. HODGSON, IGOR RIVIN & WARREN D. SMITH: A char-
acterization of convex hyperbolic polyhedra and of convex polyhedra

inscibed in a sphere, Bulletin Amer. Math. Soc. 27 (1992), 246 251.
-
(122)
[256] KATHY HOKE: Extending shelling orders and a hierarchy of functions
of unimodal simple polytopes, Discrete Applied Math. 60 (1995), 211
217. (283)
& VICTOR
KLEE: Counterexamples to the strong d-step

conjecture for d > 5, Preprint 1995; Discrete Comput. Geometry, to
appear. (97)
[257] FRED HOLT
[258] FRED HOLT & VICTOR KLEE: Many polytopes meeting the conjectured Hirsch bound, Preprint 1996; Discrete Comput. Geometry, to
appear. (84)
[259] JOHN E. HOPCROFT &
J. K. WONG: Linear time algorithm for
isomorphism of planar graphs (preliminary report), in: "Proc. Sixth

Annual ACM Symp. Theory Computing" (Seattle 1974), ACM Press
1974, pp. 172-184. (125)
KAHN: A paradigm for robust
geometric algorithms, Algorithmica 7 (1992), 339 380. (116)
[260] JOHN E. HOPCROFT
& PETER J.
[261] ROBERT B. HUGHES: Minimum-cardinality triangulations of the

d-cube for d --, 5 and d ,------ 6, Discrete Math. 118 (1993), 75 118.
(147)
-
References
345
[262]
B. HUGHES: Lower bounds on cube simplexity, Discrete

Math. 133 (1994), 123-138. (147)
[263]
ROBERT
ROBERT
B.
HUGHES
&
MICHAEL
R.
ANDERSON: A triangulation of
the 6-cube with 308 simplices, Discrete Math. 117 (1993), 253-256.
(147)
B. HUGHES & MICHAEL R. ANDERSON: Simplexity of the

cube, Discrete Math. 158 (1996), 99-150. (147)
[264]
ROBERT
[265]
Reflection Groups and Coxeter Groups,

Cambridge Studies in Advanced Mathematics 29, Cambridge University Press, Cambridge 1990. (9, 225)
JAMES E. HUMPHREYS:
[266] BEAT JAGGI, PETER MANI-LEVITSKA, BERND STURMFELS &

NEIL WHITE: Uniform oriented matroids without the isotopy property, Discrete Comput. Geometry 4 (1989), 97-100. (179)
[267] WILLIAM JOCKUSCH: The lower and upper bound problems for cubi-
cal polytopes, Discrete Comput. Geometry 9 (1993), 159-163. (280,

289)
[268]
An infinite family of nearly neighborly centrally

symmetric 3-spheres, J. Combinatorial Theory Ser. A 72 (1995), 318321. (280)
WILLIAM JOCKUSCH:
[269] MICHAEL JNGER, GERHARD REINELT & GIOVANNI RINALDI: The
traveling salesman problem, in: "Network Models" (M.O. Ball, T.L.

Magnanti, C.L. Monma, G.L. Nemhauser, eds.), Handbooks in Operations Research and Management Science, Vol. 7, North Holland
1995, pp. 225-330. (21, 22)
&
Provably good solutions for the traveling salesman problem, Z. Operations

Research 40 (1994), 183-217. (23)
[270] MICHAEL JNGER, GERHARD REINELT
STEFAN THIENEL:
[271] JEFF KAHN & GIL KALAI: A counterexample to Borsules conlecture,

Bulletin Amer. Math. Soc. 29 (1993), 60-62. (278)
[272] GIL KALAI: A simple way to tell a simple polytope from its graph,
J. Combinatorial Theory, Ser. A 49 (1988), 381-383. (vi, 93-95, 97,
100)
[273] GIL KALAI: A new basis of polytopes, J. Combinatorial Theory, Ser. A
49 (1988), 191-209. (280)
[274] GIL KALAI: Many triangulated spheres, Discrete Comput. Geometry
3 (1988), 1-14. (278)
[275] GIL KALAI: The number of faces of centrally-symmetric polytopes,
(Research Problem), Graphs and Combinatorics 5 (1989), 389-391.
(289)
346
[276]
References
GIL KALA1: On low-dimensional faces that high-dimensional
poly-
topes must have, Combinatorica 10 (1990), 271-280. (289)

[277]
The diameter of graphs of convex polytopes and f -vector

theory, in: "Applied Geometry and Discrete Mathematics -- The Victor Klee Festschrift" (P. Gritzmann and B. Sturmfels, eds.), DIMACS
Series in Discrete Mathematics and Theoretical Computer Science 4,
Amer. Math. Soc. 1991, pp. 387-411. (85, 278)
GIL KALAI:
[278] GIL KALAI: Upper bounds for the diameter and height of graphs of
convex polyhedra, Discrete Comput. Geometry 8 (1992), 363-372.
(87, 91, 96, 101, 102)
[279]
A subexponential randomized simplex algorithm, in:

"Proc. 24th ACM Symposium on the Theory of Computing (STOC),"
ACM Press 1992, pp. 475-482. (87-91, 96, 99)
[280]
Some aspects of the combinatorial theory of convex

polytopes, in: "Polytopes: Abstract, Convex and Computational"
1994, pp. 205-230. (95, 97)
GIL KALAI:
GIL KALAI:
[281] GIL KALAI: Linear programming, the simplex algorithm and simple
polytopes, Preprint 1997; in: "Proc. Int. Symp. Mathematical Programming" (Lausanne 1997), to appear. (91)
[282] GIL KALAI Sz DANIEL J. KLEITMAN: A quasi-polynomial bound for
the diameter of graphs of polyhedra, Bulletin Amer. Math. Soc. 26
(1992), 315-316. (87, 96)
[283] RAVI KANNAN: Algorithmic geometry of numbers, in: "Annual Review of Computer Science", Vol. 3, Annual Reviews Inc., New York
1987, pp. 231-267. (23)
[284] Goos
for Drawing Planar Graphs, Dissertation,

Rijksuniversiteit te Utrecht 1993, 218 pages. (111, 123)
KANT: Algorithms
[285] LEONID V. KANTOROVICH: My journey in science (proposed report to the Moscow Mathematical Society), Russ. Math. Surveys 42
(1987), 233-270. (96)
[286] MIKHAIL M. KAPRANOV: Pei muto-associahedron, MacLane coherence theorem and asymptotic zones for the KZ equation, J. Pure and
Applied Algebra 85 (1993), 119-142. (19, 291, 310, 312, 313, 321)
[287] MIKHAIL M. KAPRANOV, BERND STURMFELS & ANDREI V.
ZELEVINSKY: Quotients of toric varieties, Mathematische Annalen
290 (1991), 643-655. (320)
[288] MIKHAIL M. KAPRANOV, BERND STURMFELS Sz ANDREI V.
ZELEVINSKY: Chow polytopes and general resultants, Duke Math. J.
67 (1992), 189-218. (320)
References
347
[289] JOHAN KARLANDER: A characterization of affine sign vector systems, preprint, KTH Stockholm 1992. (227)
[290] RICHARD M. KARP: Reducibility between combinatorial problems, in: "Complexity of Computer Computations" (R.E. Miller,
J.W. Thatcher, eds.), Plenum Press, New York 1972, pp. 85-103.
(21)
[291] GYULA KATONA: A theorem on finite sets, in: "Theory of Graphs,"
Proc. Colloquium at Tihany (Sept. 1966), Academic Press, and
Akadmiai Kiado, Budapest 1968, pp. 187-207. (263)
[292] YAN KE S.L JOSEPH O'ROURKE: Comment on Pach's animal problem, preprint 1988, 4 pages. (276)
[293] VICTOR KLEE: Paths on polyhedra I, J. Soc. Indust. Math. 13 (1965),
946-956. (84)
[294] VICTOR KLEE: A property of d-polyhedral graphs, J. Math. Mechanics 13 (1964), 1039-1042. (95)
[295] VICTOR KLEE: A combinatorial analogue of Poincares duality theorem, Canadian J. Math. 16 (1964), 517-531. (260)
[296] VICTOR KLEE: The number of vertices of a convex polytope, Canadian J. Math. 16 (1964), 701-720. (254, 277, 287)
[297] VICTOR KLEE: Paths on polyhedra II, Pacific J. Math. 17 (1966),
249-262. (85, 290)
[298] VICTOR KLEE: A d-pseudomanifold with fo vertices has at least dfo (d 1)(d + 2) d-simplices, Houston Math. J. 1 (1975), 81-86. (279)
[299] VICTOR KLEE & PETER KLEINSCHMIDT: The d-step conjecture and
its relatives, Math. Operations Research 12 (1987), 718-755. (83,
96, 102)
[300] VICTOR KLEE & PETER KLEINSCHMIDT: Geometry of the Gass-
Saaty parametric cost LP algorithm, Discrete Comput. Geometry 5

(1990), 13-26. (301)
[301] VICTOR KLEE Sz PETER KLEINSCHMIDT: Polyhedral complexes and
their relatives, in: "Handbook of Combinatorics" (R. Graham, M.
Greltschel, and L. LovAsz, eds.), North-Holland/Elsevier, Amsterdam
1995, pp. 875-917. (vi, 1, 22, 116, 279)
[302] VICTOR KLEE 8.E GEORGE J. MINTY: How good is the simplex algorithm? in: "Inequalitites, III" (O. Shisha, ed.), Academic Press, New
York 1972, pp. 159-175. (71, 100)
[303] VICTOR KLEE & DAVID W. WALKUP: The d-step conjecture for
polyhedra of dimension d < 6, Acta Math. 117 (1967), 53-78. (83
86, 102)
348
References
(304] PETER KLEINSCHMIDT:
On facets with non-arbitrary shapes, Pacific
J. Math. 65 (1976), 97-101. (144, 173, 188)

[305]
PETER KLEINSCHMIDT: Spheiren
mit wenigen Ecken, Geometriae
Dedicata 5 (1976), 307-320. (172)

von
Komplexen und Polytopen, J. Geometry 11 (1978), 161-176. (277)
[306]
PETER KLEINSCHMIDT: Stellare Abeinderungen und Scheilbarkeit
[307]
PETER KLEINSCHMIDT:
The diameter of polytopes and related ap-
plications, in: "Polytopes: Abstract, Convex and Computational"

1994, pp. 467-492. (96)
[308]
On the diameter of convex

polytopes, Discrete Mathematics 102 (1992), 75-77. (91, 96, 99)
[309]
New results for simplicial spherical polytopes, in: "Discrete and Computational Geometry: Papers from the DIMACS Special Year" (eds. J. E. Goodman,
R. Pollack, W. Steiger), DIMACS Series in Discrete Mathematics
and Theoretical Computer Science, Vol. 6, Amer. Math. Soc. 1991,
pp. 187-197. (260, 277)
[310]
PETER KLEINSCHMIDT & GRAHAM
PETER KLEINSCHMIDT ik SHMUEL ONN:
PETER KLEINSCHMIDT Sz ZEEV SMILANSKY:
R. WOOD: Gale transforms

and closed faces of infinite dimensional polytopes, Mathematika 31
(1984), 291-304. (183)
[311]
Kontaktprobleme der konfornien Abbildung, Ber. Verh.

S Achs. Akademie der Wissenschaften Leipzig, Math.-Phys. Klasse 88
(1936), 141-164. (117)
[312]
GINA KOLATA:
PAUL KOEBE:
Math problem, long baffling, slowly yields. The traveling salesman problem still isn't solved, but computers can now get
most of the answers, New York Times, Tuesday March 12 (1991),
pp. Cl, C9. (23)

[313]
Every sirnplicial polytope with at most

d+4 vertices is a quotient of a neighborly polytope, Preprint 1995;
Discrete & Comput. Geometry, to appear. (187)
[314]
Small 0/1-polytopes with many facets,

http : //www . math . tu-berl in . de/ --'hund/Ol -Olympics . html (26)
ULRICH H. KORTENKAMP:
ULRICH H. KORTENKAMP:
JfIRGEN RICHTER-GEBERT, A.
SARANGARAJAN & GNTER M. ZIEGLER: Extrernal properties of
0/1-polytopes, Discrete & Comput. Geometry 17 (1997), 439-448.
(26)
[316] JOSEPH B. KRUSKAL: The number of simplices in a complex, in:
"Mathematical Optimization Techniques" (R. Bellman, ed.), University of California Press, Berkeley CA, 1963, pp. 251-278. (263)
[315]
ULRICH
H.
KORTENKAMP,
References
349
[317] HAROLD W. KUHN: The solvability and consistency for linear equa-
tions and inequalities, Amer. Math. Monthly 63 (1956), 217-232.

(39)
[318] JEFFREY C. LAGARIAS: Point lattices, in: "Handbook of Combinatorics" (R. Graham, M. Gr8tschel, and L. LovAsz, eds.), NorthHolland/Elsevier, Amsterdam 1995, pp. 919-966. (23)
[319] JEFFREY C. LAGARIAS, NAGABHUSHANA PRABHU & JAMES A.
REEDS: The d-step conjecture and Gaussian elimination, Discrete
Comput. Geometry 18 (1997), 53-82. (97)
[320] CLEMENT W. H. LAM, LARRY HENRY THIEL SL S. SWIERCZ:
The nonexistence of finite projective planes of order 10, Canadian
J. Math. 41 (1989), 1117-1123. (284)
[321] DAVID G. LARMAN: Paths on polytopes, Proc. London Math. Soc.
20 (1970), 161-178. (87, 101)
[322] EUGENE L. LAWLER, JAN KAREL LENSTRA, ALEXANDER H. G.
RINNOOY KAN & DAVID B. SHMOYS, EDS.: The Traveling Salesman
Problem. A Guided Tour of Combinatorial Optimization, John Wiley
& Sons, Chichester New York, 1985; reprinted with an index, 1990.
(21, 111)
[323] UWE LECK: The shadow spectra of uniform set systems, preprint,
FU Berlin 1994, in preparation. (284)
[324] CARL W. LEE: Counting the faces of simplicial polytopes, Ph.D. thesis, Cornell University 1981, 171 pages. (272, 278, 279)
[325] CARL W. LEE: Bounding the numbers of faces of polytope pairs
and simple polyhedra, in: "Convexity and Graph Theory" (Jerusalem,
1981), North-Holland Math. Stud. 87, North-Holland, AmsterdamNew York 1984, pp. 215-232. (279)
[3261 CARL W. LEE: Some notes on triangulating polytopes, in: Proc.
"3. Kolloquium ber Diskrete Geometrie," Institut fiir Mathematik,
Universitdt Salzburg, Salzburg, May 1985, 173-185. (129, 281)
[327] CARL W. LEE: The associahedron and triangulations of the n-gon,
European J. Combinatorics 10 (1989), 551-560. (18, 304)
[328] CARL W. LEE: Some recent results on convex polytopes, in: "Mathe-
matical Developments arising from Linear Programming" (J. C. Lagarias and M. J. Todd, eds.), Contemporary Mathematics 114 (1990),
pp. 3-19. (250, 279)
[329] CARL W. LEE: Regular triangulations of convex polytopes, in: "Applied Geometry and Discrete Mathematics The Victor Klee
Festschrift" (P. Gritzmann and Bernd Sturmfels, eds.), DIMACS Series in Discrete Mathematics and Theoretical Computer Science 4,
Amer. Math. Soc. 1991, pp. 443-456. (129, 145, 281)
350
References
[3301 CARL W. LEE: Kalai's squeezed spheres
are shellable, preprint 1994.
(278)
[331] THOMAS LENGAUER: Combinatorial Algorithms
for Integrated Cir-
cuit Layout, John Wiley & Sons, Chichester New York, and Teubner,
Stuttgart 1990. (111)
[332] HERV LE VERGE: A note on Chernikova's algorithm, Preprint

No. 635, INRIA, Univerity of Rennes 1992,28 pages. (48)
A C code for computing the

[333] HERV LE VERGE: CHERNIKOVA
dual representation of a cone, code listed in [332], and implemented
in [528]. (48)
der proyektiven Ebene durch Gerade

oder Pseudogerade, Berichte der Math. - Phys. Klasse Shchs. Akad.
Wiss., 78 (1926), 256 - 267. (216)
[334] FRIEDRICH LEVI: Die Teilung
[335] W. B. RAYMOND LICKORISH: Unshellable triangulations of spheres,

European J. Combinatorics 12 (1991), 527 - 530. (276)
A combinatorial
problem in the k-adic number system, Proc. Amer. Math. Soc. 18
[336] BERNT LINDSTR6M & HANS-OLOV ZETTERSTR6M:
(1967), 166-170. (263)

[337]
J. LIPTON & ROBERT E. TARJAN: A separator theorem

for planar graphs, SIAM J. Applied Math. 36 (1979), 177 - 189. (116)
[338]
NATHAN LINIAL,
RICHARD
LaiszL6 Lovitsz & AVI WIGDERSON: Rubber

bands, convex ernbeddings and graph connectivity, Combinatorica 8
(1988), 91 - 102. (116)
[339] JESUS DE LOERA: Computing regular triangulations for point configurations, Discrete Comput. Geometry 15 (1996), 253 - 264. (145)
[340] JESUS DE LOERA, BERND STURMFELS & REKHA R. THOMAS:
Grabner bases and triangulations of the second hypersimplex, Cornbinatorica 15 (1995), 409-424. (145)
[341] LisszL6 Loviisz: An Algorithmic Theory of Numbers, Graphs and
Convexity, CMBS-NSF Regional Conference Series in Applied Mathematics 50, Society for Industrial and Applied Mathematics (SIAM),
Philadelphia 1986. (79)
Polytopes secondaires et discriminants, in:
"Sminaire Bourbaki", 43me anne, 1990-91, no. 742, Juin 1991,
Astrisque 201 202 203 (1991), 387-420. (320)
[342] FRANOIS LOESER:

-
[343] LiiszL6 Lovitsz & MICHAEL D. PLUMMER: Matching Theory,

Akadrniai Kiad6, Budapest, and North-Holland, Amsterdam 1986.
(20)
[344] E. KEITH LLOYD: The number of d polytopes with d + 3 vertices,
Mathematika 17 (1970), 120-132. (172)
-
References
351
[345] FRANCIS SOWERBY MACAULAY: Some properties of enumeration in

the theory of modular systems, Proc. London Math. Soc. 26 (1927),
531-555. (268)
[346] ROBERT D. MACPHERSON: Combinatorial differential manifolds, in:
"Topological Methods in Modern Mathematics: A Symposium in
Honor of John Milnor's Sixtieth Birthday" Stony Brook NY 1991
(L. R. Goldberg and A. V. Phillips, eds.), Publish or Perish, Houston
TX 1993, pp. 203-221. (225, 282)
[347] OLVI L. MANGASARIAN: Nonlinear Programming, McGraw-Hill,
New York 1969. (40)
[348] PETER MANI: Automorphismen von polyedrischen Graphen, Mathematische Annalen 192 (1971), 279-303. (114)
[349] PETER MANI: Spheres with few vertices, J. Combinatorial Theory,
Ser. A 13 (1972), 346-352. (172)

[350] PETER MANI & DAVID W. WALKUP: A 3-sphere conterexample to
the W v -path conjecture, Math. Operations Research 5 (1980), 595598. (85)
[351] AL MARDEN 86 BURT RODIN: On Thurston's formulation and proof
of Andreev's theorem, in: "Computational Methods and Function
Theory" (St. Ruscheweyk et al., eds.), Lecture Notes in Mathematics
1435, Springer Verlag, Berlin Heidelberg 1990, pp. 103-115. (118)
[352] HORST MARTINI: Reguliire Polytope und Verallgemeinerungen, in:
"Geometrie und ihre Anwendungen" (Hrsg. O. Giering, J. Hoschek),
Hanser Verlag, Miinchen 1994, pp. 247-281. (9)
[353] HORST MARTINI: A hierarchical classification of Euclidean polytopes
with regularity properties, in: "Polytopes: Abstract, Convex and Computational", (T. Bisztriczky, P. McMullen, and A. Weiss, eds.), Proc.
NATO Advanced Study Institute, Toronto 1993, Kluwer Academic
Publishers 1994, pp. 71-98. (9)
[354] JIfli MATOUEK: Lower bounds for a subexponential optimization algorithm, Random Structures & Algorithms 5 (1994), 591-607. (90)
[355] hill MATOUEK, MICHA SHARIR & Emo WELZL: A subexponential

bound for linear programming, in: "Proc. Eighth Annual ACM Symp.
Computational Geometry" (Berlin 1992), ACM Press 1992, pp. 1-8.
(91, 99)
[356] THEODORE H. MAT[T]HEISS Sz DAVID S. RUBIN: A survey and comparison of methods for finding all vertices of a polyhedral set, Math.
Operations Research 5 (1980), 167-185. (48)
[357] JAMES CLERK MAXWELL: On reciprocal figures and diagrams of
forces. Philosophical Magazine Ser. 4 27 (1864), 250-261. (116)
352
References
[358] J. PETER MAY: Simplicial Objects in Algebraic Topology, Van Nostrand Mathematical Studies, No. 11, D. Van Nostrand, Princeton
1967. (264)
[359] PETER MCMULLEN: On
a problem of Klee concerning convex poly-
topes, Israel J. Math. 8 (1970), 1-4. (254)

[360] PETER MCMULLEN: Polytopes with centrally symmetric faces, Israel
J. Math. 8 (1970), 194-196. (201)
The maximum numbers of faces of a convex
[361] PETER MCMULLEN:
polytope, Mathematika 17 (1970), 179-184. (240, 252, 254-257, 277,

283)
[362] PETER MCMULLEN: On
the upper-bound conjecture for convex poly-
topes, J. Combinatorial Theory, Ser. B 10 (1971), 187-200. (277)

[363] PETER MCMULLEN: The numbers of faces of simplicial polytopes,
Israel J. Math. 9 (1971), 559-570. (269)

[364] PETER MCMULLEN: On zonotopes,
Transactions Amer. Math. Soc.
159 (1971), 91-109. (224, 226)

[365] PETER MCMULLEN: Polytopes with centrally symmetric facets, Israel
J. Math. 23 (1976), 337-338. (201)

[366] PETER MCMULLEN: Transforms, diagrams and representations, in:
"Contributions to Geometry," Proc. Geometry Symposium, Siegen

1978 (J. 'Hike and J. Wills, eds.), Birkhduser, Basel 1979, pp. 92
130. (149, 183, 224)
[367] PETER MCMULLEN:
The polytope algebra, Advances in Math. 78
(1989), 76-130. (278)

[368] PETER MCMULLEN: Separation in the polytope algebra, Beitrfige zur
Algebra und Geometrie/Contributions to Algebra and Geometry 34

(1993), 15-30. (278)
[369] PETER MCMULLEN: On simple polytopes, Inventiones Math. 113
(1993), 419-444. (278)

[370] PETER MCMULLEN: Duality, sections and proyections
of certain eu-
clidean Wings, Geometriae Dedicata 49 (1994), 183-202. (117)

[371] PETER MCMULLEN: Weights on polytopes, Discrete
Comput. Geom-
etry 15 (1996), 363-388. (278)

[372] PETER MCMULLEN: Polytope algebras, tensor weights and piecewise
polynomials, preprint 1995, 15 pages; in: Proc. "International Conference on Intuitive Geometry" (Budapest, Sept. 1995), to appear.
(278)
[373] PETER MCMULLEN & GEOFFREY C. SHEPHARD: Diagrams for centrally symmetric polytopes, Mathematika 15 (1968), 123-138. (189,
280)
References
[374] PETER MCMULLEN & GEOFFREY C. SHEPHARD:
353
Convex Polytopes
and the Upper Bound Conjecture, London Math. Soc. Lecture Notes
Series 3, Cambridge University Press, Cambridge 1971. (1, 22, 27,
183)
[375] PETER MCMULLEN & EGON SCHULTE:
Abstract Regular Polytopes,
book in preparation. (9)

[376] PETER MCMULLEN & DAVID W. WALKUP: A
generalized lower-
bound conjecture for simplicial polytopes, Mathematika 18 (1971),

264-273. (279)
Geomet rie der Zahlen, Teubner Verlag,
Leipzig 1896 and 1910; reprinted by Chelsea, New York 1953, and
by Johnson, New York 1968. (23)
[377] HERMANN MINKOWSKI:
[378] NICOLAI E. MNEv: The universality theorems on the classification

problem of configuration varieties and convex polytopes varieties, in:
"Topology and Geometry Rohlin Seminar" (O. Ya. Viro, ed.), Lecture Notes in Mathematics 1346, Springer-Verlag, Berlin Heidelberg
1988, pp. 527-544. (177, 179, 182)
The universality theorem on the oriented mafroid stratification of the space of real matrices, in: "Discrete and
[379] NICOLA! E. MNEy:
Computational Geometry: Papers from the DIMACS Special Year"

(eds. J. E. Goodman, R. Pollack, W. Steiger), DIMACS Series in Discrete Mathematics and Theoretical Computer Science, Vol. 6, Amer.
Math. Soc. 1991, pp. 237-243. (182)
[380] NICOLAI E. MNV & GNTER M. ZIEGLER: Combinatorial models
for the finite-dimensional Crassmannians, Special issue on "Oriented
Matroids" (J. Richter-Gebert and G. M. Ziegler, eds.), Discrete Com-
put. Geometry 10 (1993), 241-250. (225, 296)

[381] BOJAN MOHAR: A polynomial time circle packing algorithm, Discrete
Math. 117 (1993), 257-263. (118)
[382] ROBERT MORELLI: A theory of polyhedra, Advances in Math. 97
(1993), 1-73. (278)
[383] ROBERT MORELLI:
Translation scissors congruence, Advances in
Math. 100 (1993), 1-27. (278)

[384] THEODORE S. MOTZKIN: Beitreige zur Theorie der lineal-en Ungleichungen, Doktorarbeit, Univeristdt Basel 1936; translation: Contribution to the theory of linear inequalities, RAND Corporation Translations 22, Santa Monica 1952, 86 pages; also in: T. S. Motzkin:
Selected Papers (D. Cantor, B. Gordon, and B. Rothschild, eds.),
Birkhduser, Boston 1983, 1-80. (29, 47, 48)
[385] THEODORE S. MOTZKIN: Comonotone curves and polyhedra, Abstract, Bulletin Amer. Math. Soc. 63 (1957), 35. (144, 254)
354
References
[386] THEODORE S. MOTZKIN, HOWARD RAIFFA, GERALD L. THOMPSON & ROBERT M. THRALL: The double description method, in:
"Contributions to the Theory of Games, Vol. II" (H. W. Kuhn and
A. W. Tucker, eds.), Annals of Math. Studies 28, Princeton University Press, Princeton 1953, pp. 81-103. (36, 48, 74)
[387] KETAN MULMULEY: Computational Geometry: An Introduction
Through Randomized Algorithms, Prentice Hall, Englewood Cliffs,
NJ 1994. (258, 277)
[388] JAMES R. MUNKRES: Elements of Algebraic Topology,
Addison-
Wesley, Menlo Park CA, 1984. (59, 232)

[389] DENIS NADDEF: The Hirsch conjecture is true for (0, 1)-polytopes,
Math. Programming 45 (1989), 109-110. (91, 96)

[390] MAKOTO NAMIKI, TOMOMI MATSUI & KOMEI FUKUDA: 3-polytopes
with bad unfoldings, electronic Mathematica notebook document,
available at http: //maw. if or. math. ethz . ch/
if oristaf f ifukuda/fukuda . html (126)
[391] JOHN VON NEUMANN: A certain zero-sum two-person game equivalent to the optimal assignment problem, in: "Contributions to the
Theory of Games, Vol. II" (H. W. Kuhn and A. W. Tucker, eds.),
Annals of Math. Studies 28, Princeton University Press, Princeton
1953, pp. 5-12. (20)
[392] MAXWELL H. A. NEWMAN: A property of 2-dimensional elements,
Koninklijke Nederlandse Akademie van Wetenschappen, Amsterdam, Afdeling voor de wis- en natuurkundl. Wet enschappen (Royal
Academy of Sciences, Proceedings of the Section of Sciences), Series A
29 (1926), 1401-1405. (276, 281)
[393] Short-circuiting the traveling salesman problem, New Scientist, June

27, 1992, p. 18. (23)
[394] ALON NILLI: On Borsuk's problem, in: "Jerusalem Combinatorics
'93" (H. Barcelo and G. Kalai, eds.), Contemporary Mathematics
178, Amer. Math. Soc. 1994, 209-210. (278)
[395] ISABELLA NOVIK: Upper bound theorems for simplicial manifolds,
Preprint 1995. (278)

[396] TADAO ODA: Convex Bodies and Algebraic Geometry: An Introduction to the Theory of Toric Varieties, Ergebnisse der Mathematik
und ihrer Grenzgebiete, 3. Folge, Band 15, Springer-Verlag, Berlin
Heidelberg 1988. (224)
[397] TADAO ODA: Simple convex polytopes and the strong Lefschetz theorem, J. Pure Applied Algebra 71 (1991), 265-286. (278)
References
[398]
355
A quantitative Steinitz' theorem, Beitrage zur Algebra und Geometrie/Contributions to Algebra

and Geometry 35 (1994), 125-129. (119, 123)
SHMUEL ONN Si BERND STURMFELS:
[399] PATRICE OSSONA DE MENDEZ: Orientations Bipolaires, doctoral
thesis, cole des Hautes tudes en Sciences Sociales, Paris 1994,

113 pages. (125)
[400]
JAMES G. OXLEY:
Matroid Theory, Oxford University Press, Oxford
1992. (160)
[401] JiiNos PACH, ED.: New Trends in Discrete and Computational Geometry, Algorithms and Combinatorics 10, Springer-Verlag, Berlin
Heidelberg 1993. (22, 129)
[402] JNOS PACH Si PANKAJ K. AGARWAL: Combinatorial Geometry,
J. Wiley & Sons, New York 1995. (118)
[403] UDO PACHNER: Konstruktionsmethoden und das kombinatorische

Homijomozphieproblem fur Trzangulationen semilinearer Mannigfaltigkeiten, Abh. Math. Sem. Univ. Hamburg 57 (1986), 69-86.
(282)
[404] MANFRED PADBERG: Linear Programming, Algorithms and Combinatorics 12, Springer-Verlag, Heidelberg 1995. (48, 96)
[405] MICHA A. PERLES: Results and problems on reconstruction of polytopes, unpublished, Jerusalem 1970. (93, 95)
[406] MICHA A. PERLES: At most 2d+ 1 neighborly simplices in Ed, Annals
of Discrete Math. 20 (1984), 253-254. (144)
[407]
Reconstruction of polytopes A survey, in: "DIMACS workshop on polytopes and convex sets," abstracts, DIMACS,
Rutgers University, January 1990, 1 page. (95)
MICHA A. PERLES:
[408] MICHA A. PERLES: Problems 3.17 and 3.18, in: "DIMACS workshop
on polytopes and convex sets," list of problems, DIMACS, Rutgers
University, January 1990, 8 pages. (102, 289)
[409] MICHA A. PERLES Si NAGABHUSHANA PRABHU: A property of
graphs of convex polytopes, Note, J. Combinatorial Theory, Ser. A
62 (1993), 155-157. (95)
[410] HENRI POINCAR: Sur la gnralisation d'un theorem d'Euler relatif
aux polydres, Comptes Rend. Acad. Sel. Paris 117 (1893), 144-145.
(232)
[411] HENRI POINCAR: Complment a l'analysis situs, Rendiconti Circolo
Mat. Palermo 13 (1899), 285-343. (232)
[412] NAGABHUSHANA PRABHU: Problem 3.11, in: "DIMACS workshop

on polytopes and convex sets," list of problems, DIMACS, Rutgers
University, January 1990, 8 pages. (101)
356
References
[413] J. SCOTT PROVAN: Decompositions, shellings, and diameters of simplicial complexes and convex polyhedra, Ph.D. thesis, Cornell University 1977. (247)
Andr permutations, lexicographic shellability, and

the cd-index of a convex polytope, Transactions Amer. Math. Soc.
[414] MARK PURTILL:
338 (1993), 77 104. (280)

-
[415] RICHARD RADO: A sequence of polyhedra having intersections of
specified dimensions, J. London Math. Soc. 22 (1947), 287 289. (144)

-
[416] JiiRG RAMBAU
& GNTER
polytopes and
M. ZIEGLER: Mapping
cofi ber polytopes, Notes, ZIB Berlin 1994. (324)

[417] J611G RAMBAU & GNTER M. ZIEGLER: Projections of
polytopes
and the Generalized Baues Conjecture, Discrete Comput. Geometry

16 (1996), 215 237. (296)
-
[418] GERHARD REINELT: The Traveling Salesman. Computational Solu-
tions for TSP Applications, Lecture Notes in Computer Science 840,

Springer-Verlag, Berlin 1994. (23)
[419] VICTOR REINER & GNTER M. ZIEGLER:
Coxeter associahedra,
-
Mathematika 41, 364-393. (19, 291, 310, 313, 318, 321)

[420] JRGEN RICHTER:
Kombinatorische Realisierbarkeitskriterien fiir
orientierte Matroide, Diplomarbeit, TH Darmstadt 1988, 113 pages;

Mitteilungen Math. Seminar Giefien 194 (1989). (190, 217)
Oriented matroids with few mutations, Special issue on "Oriented Matroids" (J. Richter-Gebert and
G. M. Ziegler, eds.), Discrete Comput. Geometry 10 (1993), 251-269.
(223, 224)
[421] JRGEN RICHTER-GEBERT:
[422] JRGEN RICHTER-GEBERT: Line arrangements and
zonotopal W-
ings: A little printer exercise, HyperSpace 2 (1993), 8 17. (229)

-
[423] JURGEN RICHTER-GEBERT: Two interesting oriented
matroids, Doc-
umenta Mathematica 1 (1996), 137-148. (177)

[424] JRGEN RICHTER-GEBERT: Realization Spaces of
Polytopes, Lecture
Notes in Mathematics 1643, Springer-Verlag, Berlin Heidelberg 1996.

(vii, 116, 119, 123, 147, 177, 181, 182, 186, 190)
[425] JURGEN RICHTER-GEBERT
& GNTER
M. ZIEGLER:
Zonotopal
Wings and the Bohne-Dress theorem, in: "Jerusalem Combinatorics

'93" (H. Barcelo and G. Kalai, eds.), Contemporary Mathematics
178, Amer. Math. Soc. 1994, 211-232. (220, 222, 225)
[426] JURGEN RICHTER-GEBERT & GNTER M. ZIEGLER: Realization
spaces of 4-polytopes are universal, Bulletin Amer. Math. Soc. 32

(1995), 403-412. (147, 182)
References
357
[427] GERHARD RINGEL: Teilungen der Ebene durch Geraden oder topolo-
gische Geraden, Math. Zeitschrift 64 (1956), 79-102. (177, 217)
&
D. SEYMOUR: Graph minors A survey,

in: "Surveys in Combinatorics" 1985, Proc. Tenth British Combinatorial Conference (I. Anderson, ed.), London Math. Society Lecture
Notes Series 103, Cambridge 1985, pp. 153-171. (109)
[428] NEIL ROBERTSON
PAUL
[429] STEWART A. ROBERTSON: Polytopes and Symmetry, London Math.
Soc. Lecture Note Series 90, Cambridge University Press, Cambridge

1984. (182)
[430] GNTER ROTE: Degenerate convex hulls in high dimensions without
extra storage, in: "Proc. Eighth Annual ACM Symp. Computational

Geometry" (Berlin 1992), ACM Press 1992, pp. 26-32. (49)
[431] GUNTER ROTE: An upper bound on the number of facets of a 0-1-
poiytope, Preprint, December 1996, 2 pages. (26)

[432] JEAN-PIERRE ROUDNEFF & BERND STURMFELS: Simplicial cells in
arrangements and mutations of oriented mat roids, Geometriae Dedicata 27 (1988), 153-170. (208, 223, 224)
[433] MARY ELLEN RUDIN: An unshellable triangulation of a tetrahedron,

Bulletin Amer. Math. Soc. 64 (1958), 90-91. (237, 275, 282)
[434] SALMAN RUSHDIE: Haroun and the Sea of Stories, Granta Books,
London 1990. (v)

Stresses and liftings of cell-complexes,
Preprint 1997, 35 pages. (143)
[435] KONSTANTIN RYBNIKOV:
[436] MORIHIKO SAITO: Hodge structure via filtered D-modules, Asterisque
130 (1985), 342-351. (278)

[437] LUDWIG SCHLFLI: Theorie der vielfachen Kontinuitdt, written
1850-1852; Zrcher und Furrer, Zrich 1901; Denkschriften der
Schweizerischen naturforschenden Gesellschaft 38 (1901), 1-237;
reprinted in: "Ludwig Schlgfli, 1814-1895, Gesammelte Mathematische Abhandlungen," Vol. I, Birkhauser, Basel 1950, pp. 167-387.
(22, 231, 275)
[438] VICTOR SCHLEGEL: Theorie der homogen zusammengesetzten
Raumgebilde, Engelmann, Leipzig, Halle 1883: Nova Acta Leop.
Carol. (Verhandlungen der Kaiserlichen Leopoldinisch-Carolinischen
Deutschen Akademie der Naturforscher) 44 (1883), 343-359. (134)
[439] ROLF G. SCHNEIDER: Ober eine Integralgleichung in der Theorie der
konvexen larper, Math. Nachrichten 44 (1970), 55-75. (200, 224)

Theory, Encyclopedia of Mathematics, Vol. 44, Cambridge University
Press, Cambridge 1993. (22)
[440] ROLF G. SCHNEIDER: Convex Bodies: The Brunn-Minkowski
358
References
G. SCHNEIDER: Neighbourliness of centrally symmetric convex polytopes in high dimensions, Mathernatika 22 (1975), 176-181.
(280)
[441]
ROLF
[442]
G. SCHNEIDER & WOLFGANG WEIL: Zonoids and related topics, in: "Convexity and Its Applications" (P. M. Gruber and J. Wills,
eds.), Birkhauser, Basel 1983, pp. 296-317. (224)
[443]
Embedding planar graphs on the grid, in: "Proc.

First ACM-SIAM Symposium on Discrete Algorithms," ACM Press
1990, pp. 138-147. (123)
ROLF
WALTER SCHNYDER:
[444] PIETER HENDRIK SCHOUTE: Mehrdirnensionale Geometrie. Vol. 2
(Die Polytope), Leipzig 1905. (22, 171, 231)

[445]
Analytic treatment of the polytopes regularly derived from the regular polytopes, Verhandelingen der Koninklijke Akademie van Wetenschappen te Amsterdam, Deel 11, No. 3,
Johannes Mller, Amsterdam 1911, 87 pages. (17, 24, 224)
PIETER HENDRIK SCHOUTE:
[446] ODED SCHRAMM: Existence and uniqueness of packings with specified
combinatorics, Israel J. Math. 73 (1991), 321-341. (118)

[447] ODED SCHRAMM: How to cage an egg, Inventiones Math. 107 (1992),
543-560. (118)
[448]
Theory of Linear and Integer Programming, Wiley-Interscience Series in Discrete Mathematics and Optimization, John Wiley & Sons, Chichester New York 1986. (19, 22,
23, 27, 47, 69, 96)
ALEXANDER SCHRIJVER:
[449] MARCEL-PAUL SCHTZENBERGER: A characteristic property of cer-
tain polynomials of E. F. Moore and C. E. Shannon, in: "RLE Quarterly Progress Report No. 55," Research Lab. of Electronics, MIT
1959, pp. 117-131 (263)
Analogues of Steinitz's theorem about non-inscribable polytopes, in: "Intuitive Geometry" (Si6fok 1985), Colloquia
Soc. JAnos Bolyai 48, North-Holland, Amsterdam 1987, pp. 503-516.
(122)
[450] EGON SCHULTE:
[451] CHRISTOPH SCHULZ: An invertible 3-diagram with 8 vertices, Dis-
crete Math. 28 (1979), 201-205. (139)
Dual pairs of non-polytopal diagrams and

spheres, Discrete Math. 55 (1985), 65-72. (139)
[452] CHRISTOPH SCHULZ:
[453] RAIMUND SEIDEL: Small-dimensional linear programming and con-
vex hulls made easy, Discrete Comput. Geometry 6 (1991), 423-434.

(48)
References
359
[454] RAIMUND SEIDEL: The upper bound theorem for polytopes: an easy
proof of its asymptotic version, Comput. Geometry: Theory and Applications 5 (1995), 115 - 116. (258)
[455] MAJORIE SENECHAL: Crystalline Symmetries. An Informal Mathematical Introduction, Adam Hilger, Bristol, Philadelphia and New
York 1993. (9)
[456] MAJORIE SENECHAL & GEORGE FLECK, EDS.: Shaping Space. A
Polyhedral Approach, Design Science Collection, Birkhhuser Verlag,
Boston 1988. (103)
[457] ROBERT WILLIAM SHANNON: Simplicial cells in arrangements of hyperplanes, Geometriae Dedicata 8 (1979), 179-187. (207)
[458] IDO SHEMER: Neighborly polytopes, Israel J. Math. 43 (1982), 291314. (187)
[459] GEOFFREY C. SHEPHARD: Combinatorial properties of associated
zonotopes, Canadian J. Math. 26 (1974), 302 - 321. (230)
[460] GEOFFREY C. SHEPHARD: Convex polytopes with convex nets, Math.

Proceedings Cambridge Math. Soc. 78 (1975), 389 -403. (126)
[461] THOMAS C. SHERMER: A smaller irreducible animal; A very small

irreducible animal, in: "Snapshots of Computational and Discrete
Geometry" (G. Toussaint, ed.), McGill University TR SOCS-88.11,
Montreal 1988, pp. 139-141 and pp. 142-143. (276)
[462]
Three more animals, preprint, Dept. Computer Science, McGill University, Montral 1988, 5 pages. (276)
THOMAS C. SHERMER:
[463] PETER W. SHOR: Stretchability of pseudolines is NP hard, in:

"Applied Geometry and Discrete Mathematics The Victor Klee
Festschrift" (P. Gritzmann and B. Sturmfels, eds.), DIMACS Series
in Discrete Mathematics and Theoretical Computer Science 4, Amer.
Math. Soc. 1991, pp. 531-554. (182)
-
[464] ROBERT S. SIMON: The combinatorial properties of "cleanness,"

Ph.D. dissertation, Bielefeld 1992; preprint 92-077, Sonderforschungsbereich 343 "Diskrete Strukturen in der Mathematik," Universitat
Bielefeld 1992, 101 pages. (287)
[465] NAVIN SINGHI & S. S. SHRIKHANDE: A reciprocity relation for tdesigns, European J. Combinatorics 8 (1987), 59 - 68. (284)
[466] ZEEV SMILANSKY: Convex hulls of generalized moment curves, Israel
J. Math. 52 (1985), 115 - 128. (76)
[467] ZEEv SMILANSKY: Bi cyclic 4 polytopes, Israel J. Math. 70 (1990),
82 -92. (76)
-
[468] ZEEV SMILANSKY: A non geometric shelling of a 3 polytope, Israel

J. Math. 71 (1990), 29-32. (285)
-
360
References
[4691 WARREN SMITH: Lower bounds for triangulations of the N -cube,

preprint 1994, 8 pages. (147)
[470] DUNCAN MIAREN YOUNG SOMMERVILLE: An Introduction to the
Geometry of n Dimensions, London 1929; reprint, Dover Publications, New York 1958. (22, 134, 171, 231)
[471] DUNCAN M 'LAREN YOUNG SOMMERVILLE: A Bibliography of NonEuclidean Geometry, London 1911; supplemented reprint, Chelsea
Publishing Company, New York 1970. (22)
[472] DUNCAN MTAREN YOUNG SOMMERVILLE: The relations connecting
the angle-sums and volume of a polytope in space of n dimensions,
Proc. Royal Society London, Ser. A 115 (1927), 103-119. (252)
[473] EMANUEL SPERNER: Ein Satz ber Untermengen einer endlichen
Menge, Mathematisehe Zeitschrift 27 (1928), 544-548. (259)
[474] RICHARD P. STANLEY: The upper bound conjecture and CohenMacaulay rings, Studies in Applied Math. 54 (1975), 135-142. (267,
269, 278)
[475] RICHARD P. STANLEY: Cohen-Macaulay complexes, in: "Higher
Combinatorics" (M. Aigner, ed.), NATO Advanced Study Institute
Series, Reidel, Dordrecht/Boston 1977, pp. 51-62. (267, 268)
[476] RICHARD P. STANLEY: Balanced Cohen-Macaulay complexes, Transactions Amer. Math. Soc. 249 (1979), 139-157. (247, 267, 268)
[477] RICHARD P. STANLEY: The number of faces of simplicial convex
polytopes, Advances in Math. 35 (1980), 236-238. (278)
[478] RICHARD P. STANLEY: Combinatorics and Commutative Algebra,
Progress in Mathematics 41, Birkhauser, Boston 1983. (1, 22, 267,
278)
[479] RICHARD P. STANLEY: The number of faces of simplicial polytopes
and spheres, in: "Discrete Geometry and Convexity," New York 1982,
(J. E. Goodman, E. Lutwak, J. Malkevitch, and R. Pollack, eds.),
Annals New York Academy of Sciences 440, New York 1985, pp. 212223. (250, 278)
[480] RICHARD P. STANLEY: Enumerative Combinatorics, Volume I,
Wadsworth & Brooks/Cole, Monterey 1986. (55, 260, 277, 286)
[481] RICHARD P. STANLEY: On the number of faces of centrally-symmetric simplicial polytopes, Graphs and Combinatorics 3 (1987), 55-66.
(278, 280, 289)
[482] RICHARD P. STANLEY: Generalized h-vectors, intersection cohomology of toric varieties, and related results, in: "Commutative Algebra
and Combinatorics" (M. Nagata and H. Matsumura, eds.), Advanced
References
361
Studies in Pure Mathematics 11, Kinokuniya, Tokyo, and NorthHolland, Amsterdam/New York 1987, pp. 187-213. (280)
[483] RICHARD P. STANLEY: Subdivisions and local h-vectors, J. Amer.
Math. Soc. 5 (1992), 805-851. (281)

[484] RICHARD P. STANLEY: Flag f -vectors and the cd-index, Math. Zeit-
schrift 216 (1994), 483-499. (280)

[485] JAMES DILLON STASHEFF: Homotopy associativity of H -spaces I,
Transactions Amer. Math. Soc. 108 (1963), 275-292. (18, 304)

der Abhdngigkeit geometrischer Gestalten von einander, Fincke, Berlin 1832; also in
"Gesammelte Werke, Vol. 1," Reimer, Berlin 1881, pp. 229-458.
(122)
[486] JACOB STEINER: Systematische Entwicklung
[487] ERNST STEINITZ: Polyeder und Raumeinteilungen, Encyclophdie der

mathematischen Wissenschaften, Band 3 (Geometrie), Teil 3AB12,
1922, 1-139. (103, 104, 115)
Ober isoperimetrische Pro bleme bei konvexen

Polyedern, J. reine angewandte Math. 159 (1928), 133-143. (122)
[488] ERNST STEINITZ:
[489] ERNST STEINITZ: Ober die Eulerschen Polyederrelationen, Archiv fiir
Mathematik und Physik 11 (1906), 86-88. (287)

[490] ERNST STEINITZ 8,/ HANS RADEMACHER: Vorlesungen ber die Theone der derPolyeder, Springer-Verlag, Berlin 1934; reprint, Springer-
Verlag 1976. (22, 103, 104, 115)

[491] JAMES STILLWELL: Classical Topology and Combinatorial Group
Theory, Graduate Texts in Mathematics 72, Springer-Verlag, New
York 1982; second edition 1993. (232, 276)
[492] JOSEF STOER 8z CHRISTOPH WITZGALL: Convexity and Optimization in Finite Dimensions I, Grundlehren der mathematischen Wissensc,haften, Band 163, Springer-Verlag, Berlin Heidelberg 1970. (40)
[493] BERND STURMFELS: Boundary complexes of convex polytopes cannot
be characterized locally, J. London Math. Soc. 35 (1987), 314-326.
(184)
[494] BERND STURMFELS: Neighborly polytopes and oriented matroids, European J. Combinatorics 9 (1988), 537-546. (183, 187)
[495] BERND STURMFELS: Some applications of affine Gale diagrams to
polytopes with few vertices, SIAM J. Discrete Math. 1 (1988), 121133. (172, 174, 183, 187, 188)
[496] BERND STURMFELS: Pretty pictures, or How to subdivide a 12-gon
into parallelograms, Letter to Jan Stienstra (Note), November 1991,
10 pages. (324)
362
References
[497] BERND STURMFELS: Fiber polytopes: A brief overview, in: "Special Differential Equations" (M. Yoshida, ed.), Kyushu University,
Fukuoka, 1991, pp. 117-124. (291, 296, 321)
[498] BERND STURMFELS: Asymptotic analysis of toric ideals, Memoirs of
the Faculty of Science, Kyushu University, Ser. A Mathematics 46
(1992), 217-228. (320)

[499] BERND STURMFELS Sz GUNTER M. ZIEGLER: Extension spaces of
oriented mat roids, Discrete Comput. Geometry 10 (1993), 23-45.

(225)
[500] P. Y. SUVOROV: Isotopic but not rigidly isotopic plane systems
of straight lines, in: "Topology and Geometry Rohlin Seminar"
(O. Ya. Viro, ed.), Lecture Notes in Mathematics 1346, SpringerVerlag, Berlin Heidelberg 1988, pp. 545-556. (177, 179)
[501] JEAN E. TAYLOR: Zonohedra and generalized zonohedra, Amer.
Math. Monthly 99 (1992), 108-111. (226)
[502] TORSTEN THIELE: Extremalprobleme fiir Punktmengen, Diplomarbeit, Freie UniversitEd Berlin 1991, 80 pages. (122)
[503] CARSTEN THOMASSEN: Planarity and duality of finite and infinite
graphs, J. Combinatorial Theory, Ser. B 29 (1980), 244-271. (114)
[504] WILLIAM P. THURSTON: Geometry and Topology of 3-Manifolds,
Lecture Notes, Princeton University, Princeton 1977-1978. (117, 118)

[505] HEINRICH TIETZE: Ober das Problem der Nachbargebiete im Baum,
Zeitschrift fr Mathematik und Physik 16 (1905), 211-216. (144)
[506] HEINRICH TIETZE: Famous Problems of Mathematics, New York
1965. (144)
The monotonic bounded Hirsch conjecture is
false for dimension at least 4, Math. Operations Research 5 (1980),
599-601. (86)
[507] MICHAEL J. TODD:
[508] MICHAEL J. TODD & LEVENT TUNEL: A new triangulation for

simplicial algorithms, SIAM J. Discrete Math. 6 (1993), 167-180.
(147)
[509] KLAUS TRUEMPER: On the delta-wye reduction for planar graphs,
J. Graph Theory 13 (1989), 141-148. (104, 109)
[510] KLAUS TRUEMPER: Mat roid Decomposition, Academic Press, San
Diego 1992. (104, 109, 115)
[511] SERGEI NIKOLAJEWITSCH TSCHERNIKOW: Lineare Ungleichungen,

Deutscher Verlag der Wissenschaften, Berlin 1971. (48)
[512] WILLIAM T. TUTTE: Convex representations of graphs, Proc. London Math_ Soc. 10 (1960), 304-320. (114, 116)
363
References
[513] WILLIAM T. TUTTE: Connectivity in Graphs, University of Toronto

Press, Toronto 1966. (115)
[514] WILLIAM T. TUTTE: Graph Theory, Encyclopedia of Mathematics,
Vol. 21, Cambridge University Press, Cambridge 1984. (80, 115)
[515] OSWALD VEBLEN Sz JOHN WESLEY YOUNG: Projective Geometry,
Ginn and Co., Boston, Vol. 11910; Vol. 11 1917. (69)
[516] ANDREW VINCE: A non-shellable 3-sphere, European J. Combinatorics 6 (1985), 91-100. (277)
[517] KLAUS WAGNER: Bemerkungen zurn Vierfarbenproblem, Jahresbericht der Deutschen Mathematiker-Vereinigung 46 (1936), 26-32.
(120)
[518] DOMINIC J. A. WELSH: Matroid Theory, Academic Press, London
1976. (160)
[519] JOHN E. WETZEL: Which Griinbaum arrangements are simplicial?
Expositiones Mathematicae 11 (1993), 109-122. (225)
[520] NEIL WHITE, ED.: Theory of Matroids, Encyclopedia of Mathematics, Vol. 26, Cambridge University Press, Cambridge 1986. (160)
[521] NEIL WHITE: Coordinatizations, in: "Combinatorial Geometries"
(N. White, ed.), Encyclopedia of Mathematics, Vol. 29, Cambridge
University Press, Cambridge 1987, pp. 1-27. (182)
[522] NEIL WHITE: A non-uniform oriented mat roid which violates the
isotopy property, Discrete Comput. Geometry 4 (1989), 1-2. (179)
[523] JOHN HENRY CONSTANTINE WHITEHEAD: On subdivisions of complexes, Proc. Cambridge Phil. Soc. 31 (1935), 69-75. (144, 238)
[524] WALTER WHITELEy: Motions and stresses of projected polyhedra,
Structural Topology 7 (1982), 13-38. (116)
[525] WALTER WHITELEY: Matroids and rigid structures, in: "Matroid Applications" (N. White, ed.), Encyclopedia of Mathematics, Vol. 40,
Cambridge University Press, Cambridge 1992, pp. 1-53. (116)
[526] WALTER WHITELEY: 3-diagrams and Schlegel diagrams of simple 4polytopes, preprint 1994. (143)
[527] HASSLER WHITNEY: Congruent graphs and the connectivity of
graphs, American J. Math. 54 (1932), 150-168. (105)
[528] DORAN K. WILDE: A library for doing polyhedral operations, Internal
Publication No. 785, IRISA, Rennes, France 1993, 45 pages; report
and code library available by anonymous ftp from ftp. irisa fr,
directory Vlocal/API/polylib". (48)
,
[529] KATHY WILLIAmSON HOKE: Completely unimodal numberings of a

simple polytope, Discrete Applied Math. 20 (1988), 69-81. (283)
364
[530]
References
HARALD WISTUBA
&
GNTER M. ZIEGLER: h-vectors of pure sim-
plicial complexes, Notes, Institut Mittag-Leffler, Djursholm 1991, unpublished. (286)

[531]
S. WITSENHAUSEN: A support characterization of zonotopes,

Mathematika 25 (1978), 13-16. (226)
[532]
R. WOOD, STEPHEN R. McDowELL Sz PETER KLEINSCHMIDT: Constructing infinite dimensional polytopes using Gale
transforms, preprint 1992; SEAMS Bull. Math., in print. (183)
[533]
A. YEMELICHEV, MIKHAIL M. KOVALEV & MIKHAIL

K. KRAVTSOV: Polytopes, Graphs and Optimization, Cambridge University Press, Cambridge 1984. (1)
[534]
ZAKS: No nine neighborly tetrahedra exist, Memoirs Amer.

Math. Soc. 447 (1991). (144)
HANS
GRAHAM
VLADIMIR
JOSEPH
[535] THOMAS ZASLAVSKY: Facing up to arrangements: Face count formulas for partitions of space by hyperplanes, Memoirs Amer. Math. Soc.
No. 154, 1 (1975). (183)
[536] GNTER M. ZIEGLER: Vier farbige Probleme Nachbargebiete und
Kartenfiirbung in drei und mehr Dimensionen, Jugend forscht (German National Science Fair) 1982. (144)
[537] GNTER M. ZIEGLER: Three problems about 4-polytopes, in: "Poly-
topes: Abstract, Convex and Computational" (T. Bisztriczky,

P. McMullen, and A. Weiss, eds.), Proc. NATO Advanced Study
Institute, Toronto 1993, Kluwer Academic Publishers 1994, pp. 499502. (vii)
[538] GNTER M. ZIEGLER: Shelling polyhedral 3-balls and 4-polytopes,
Preprint 1995; Discrete Comput. Geometry 18 (1997), in print. (244,

276, 277)
Recent progress on polytopes, Preprint
537/1996, TU Berlin, September 1996, 10 pages; in: Proc. "Discrete
and Computational Geometry: Ten Years Later," Mt. Holyoke, July
1996 (B. Chazelle, J.E. Goodman, R. Pollack, eds.), Contemporary
Mathematics, Amer. Math. Soc., to appear. (vii)
[539] GNTER M. ZIEGLER:
Index
active facet, 88
acyclic orientation, 80
acyclic vector configuration, 156
admissible hyperplane, 67
affine dependences, 150
affine Gale diagram, 168
affine hull, 3
affine map, 3, 195
affine subspaces, 2
affine transformation, 3, 196
affinely independent, 3
affinely isomorphic, 5, 128
animal problem, 276
arrangement, 193
affine, 212
of hyperplanes, 193
of pseudohyperplanes, 211
of pseudolines, 213
realizable, 216
assignment polytope, 20
associahedron, 18, 306
axiom systems, 160
Balinski's theorem, 95
Barnette sphere, 143
being explicit, 79
belt polytopes, 226
bicyclic polytope, 76
binomial expansion, 262
bipyramid, 9
Birkhoff polytope, 20, 320
Bohne-Dress theorem, 220, 225
boundary complex, 129, 232
bounded, 4, 29
Brfickner sphere, 143
Bruggesser-Mani shellings, 240
capped prism, 145
Carathodory curve, 75
Carathodory's theorem, 46
category of polytopes, 319
centrally symmetric, 17
characteristic cone, 43
circle packing theorem, 117
coface, 154
cofacet, 154
cofiber polytope, 319, 324
column vectors, 2
combinatorial explosion, 47
366
Index
combinatorially equivalent, 5, 58,

128, 216
combinatorially polar, 64
completion lemma, 238
complex
abstract simplicial, 101, 258
boundary complex, 129, 232
compressed, 263
geometric simplicial, 258
h-vector, 248
link, 237
multicomplex, 264
of a polytope, 129
partit ionable, 247
polyhedral, 127
polytopal, 127, 232
pure, 232
shellable, 233
star, 237
computational convexity, 28
computational geometry, 48
cone, 28
polyhedral, 30
conical hull, 28
connected sum, 274
contractible, 115
contraction, 106, 163
convex, 3
convex hull, 3
convex hull problem, 48
convexity theory, 22
correct drawing, 116
covector axioms, 209
crosspolytope, 8
cube, 7
cubical polytopes, 23, 280
cut polytopes, 23
cutting plane algorithms, 23
cyclic polytope, 11
Gale diagram, 186
Schlegel diagram, 144
d-cube, 7
d-diagram, 138
examples, 139
invertible, 139
non-Schlegel, 139
simple, 138
simplicial, 138
topological, 142
Dehn-Sommerville equations, 252
Delaunay triangulation, 146
deletion, 106, 163
deletion and contraction, 183
Delta-Wye operation, 106
AY operation, 106
diagram, see d-diagramdiameter,
83
diamond property, 57
dimension, 2, 5, 51, 127, 232
double description method, 36, 48
d-polytope, 5
d-simplex, 7
d-step conjecture, 84
dual graph, 105
dual vector space, 2
duality, 163
duality theorems, 39
edges, 5, 51
Egyptian pyramid, 9
elimination, 32
equivalent, see
combinatorially equivalent
equivalent data, 160
Euler's formula, 120
Euler-Poincar formula, 231
extendably shellable, 235
extremal set theory, 261
compression, 263
face, 5, 51, 232
proper, 5
face fan, 192
face figure, 71, 187
face lattice, 57
face poset, 128
facet, 5, 51, 232
fan, 191
common refinement, 194
Index
direct sum, 194
face fan, 192
nonpolytopal, 194
normal fan, 193
restriction, 195
simplicial, 192
Farkas lemma, 39, 50
fast algorithms, 126
fiber polytope, 291, 296
flag vector, 280
flats, 2
Fourier-Motzkin elimination, 32,
35, 47
for cones, 37
4-polytopes, 127
f-vector, 245
nonunimodal, 272
Gale diagram, 149, 168
affine, 168
central, 189
zonal, 224
Gale's evenness condition, 12, 14
g-conjecture, 269
general position, 8, 79
generating functions, 277
generic, 8, 79
geometric realization, 258
geometry of numbers, 23
graph, 80
d-connected, 95
dimensionally ambiguous, 98
dual, 105
good orientation, 93
k-connected, 104, 116
k-regular, 94
layout, 111
of 3-polytopes, 103
of 4-polytopes, 102, 126
planar, 103
simple, 103
straight line drawing, 120
graph theory, 80
grid graph, 110
group action, 320
367
g-theorem, 269
g-vector, 269
Hamilton cycle, 290
hexagon, 6
Hirsch conjecture, 83
for 0/1-polytopes, 91
monotone, 86
strict monotone, 86
upper bounds, 87
history, 69
homogenization, 31, 44
7-1-polyhedron, 28
7-1-polytope, 29
h-vector, 248
generalized, 280
local, 281
hypercube, 7
hyperplane, 2, 212
linear, 2, 193
hyperplane arrangement, 193
affine, 212
essential, 205
linear, 193
hypersimplex, 19
induced subgraph, 80, 102
integer coordinates, 122
integral points, 23
interior, 60
interval, 56
isomorphic, see affinely isomorphic
isotopy conjecture, 177
join, 323
Kleinschmidt polytope, 188

Kruskal-Katona theorem, 263
lattice, 56
atomic, coatornic, 56
join and meet, 56
lattice polytope, 66
totally unimodular, 146
Lawrence extensions, 180
Lawrence polytope, 180
368
Index
length, 56
line shellings, 242
lineality space, 43
linear dependences, 157
linear forms, 2
linear programming, 80, 193
basis version, 100
linear subspaces, 2
lines, 2
link, 237
lower faces, 130, 294
Macaulay's theorem, 267
main theorem, 27
for cones, 30
for polyhedra, 30
for polytopes, 29
MAPLE, 279
mapping polytopes, 319
matching polytope, 320
matroid, 160
McMullen correspondence, 270
Minkowski sum, 28, 197
minor, 106
Mbius band, 148
moment curve, 11
monotone Hirsch conjecture, 86
monotone path polytope, 300
monotone paths, 81
M-sequences, 268
multicomplex, 264
multiset, 263
negative point, 168
neighborly polytopes, 16, 186, 187,
254
nets, 126
Newton polytope, 322
n-gon, 6
non-Pappus configuration, 217
nonrational 8-polytope, 172
nonrevisiting path conjecture, 85
normal fan, 193
notation, 2
octahedron, 7
orientation of a graph, 80
acyclic, 80, 93
oriented matroid, 149, 158, 160
axiom systems, 159
cocircuit axioms, 228
covector axioms, 209
deletion and contraction, 163
duality, 149, 163
equivalent data, 160
non-Pappus, 217
nonrealizable, 208, 216
rank, 159, 209
realizable, 158
what are they good for? 211
Pappus configuration, 216
partially ordered sets, 55
pencil of lines, 141
perfect matching polytope, 20
permutahedron, 17, 200, 301
generalized, 24
permuto-associahedron, 19, 310
piecewise linear, 130
pile of cubes, 131
planes, 2
point beyond F, 78
pointed polyhedra, 43
points, 2
polar polytope, 8, 59
combinatorially, 64
polar set, 61
polygons, 6
polyhedral complex, 127
polyhedral cones, 30
polyhedron
fl-polyhedron, 4, 28
pointed, 43
V-polyhedron, 29
polytopal complex, 127, 232
shellable, 233
polytope, 4, 5
bicyclic, 76
cd-index, 280
centrally symmetric, 17
Index
connected sum, 274
cubical, 23, 280
d-polytope, 5
edges, 51
equivalent, 5
face lattice, 57
faces, 51
facets, 51
flag vector, 280
4-polytopes, 127
f -vector, 245
graph, 80
g-vector, 269
7-i-polytope, 4, 29
h-vector, 246, 250
integer coordinates, 122
isomorphic, 5
join, 323
Minkowski sum, 28
neighborly, 16, 186, 187, 254
nonrational, 172, 186
polar, 59
product, 9
projection, 196, 292
proper face, 51
quotient, 71, 289
rational, 66
reconstruction problems, 95
representation theorem, 65
ridges, 51
rigid, 179
simple, 8, 66
simplicial, 8, 65
spherical, 260
3-polytopes, 103
vertices, 51
V-polytope, 4, 29
with "few vertices", 171
polytope algebra, 278
polytope pairs, 320
PORTA, 11, 48, 309
poset, 56
bounded, 56
graded, 56
interval, 56
369
rank, 56
positive halfspace, 203
positive hull, 28
positive point, 168
positive sign vector, 154
prescribing
shadow boundary, 114
shape of a facet, 114, 141, 174
2-face, 175
prism, 10
product, 9
projection, 16, 32, 292
projection of polytopes, 196
projective transformations, 67
applications, 74
projectively unique, 190
proper face, 5, 51
pseudoline, 213
pyramid, 9
quotients, 71, 289
Radon's theorem, 151, 184
rank, 56, 156
rational polytope, 66
realizable oriented matroid, 158
realization space, 115
recession cone, 43
redundancy criteria, 73
redundant inequality, 47, 72
criteria, 73
regular n-gon, 6
regular subdivision, 129
relative interior, 60
representation theorem, 65
reverse lexicographic ordering, 261
reverse search, 49
ridges, 51
row vectors, 2
Schlegel diagram, 133, 134
of cyclic polytope, 144
secondary polytope, 291, 305
section, 296
semialgebraic set, 115, 119
370
Index
semisimplicial sets, 264

separation problem, 193
separation theorems, 40
series-parallel reductions, 106
set system, 258
shadow-vertex algorithm, 301
Shannon's theorem, 207
shellable, 233
extendably, 235
shelling, 233
perfect, 282
shelling extension conjecture, 287
sign function, 152
sign vectors, 208
composition, 209
elimination, 209
signed circuits, 152, 157
signed cocircuits, 154, 157
signed covectors, 154, 157
signed vectors, 152, 157
simple, 8, 66, 138
simplex, 7
standard, 7
simplex algorithm, 81
simplicial, 8, 65, 138
skeleton, 64
Sperner's lemma, 259
spherical polytopes, 260
stable equivalence, 181
standard d-simplex, 7
Stanley's trick, 250
star, 237
Steinitz' theorem, 103
classical proofs, 104
new proofs, 116
stellar subdivision, 78, 97
strongly connected, 101
subcomplex, 129
subdivision, 129
coherent, 294
induced, 292
regular, 129, 294
support, 151
tetrahedron, 6
3-polytopes, 103
topological representation
theorem, 211, 215
transportation polytope, 38, 50
traveling salesman polytope, 21
traveling salesman problem, 21
triangulation, 129
2-neighborly, 10
ultraconnected, 101
underlying set, 127, 232
unimodality conjecture, 271
universality theorem, 182
upper bound theorem, 16, 254
for centrally symmetric
polytopes? 279
for polytope pairs, 279
valid inequality, 51
value vector, 153, 157
vector configuration
acyclic, 156
dual, 165
simple, 206
totally cyclic, 167
vector space, 2
vector sum, 28, 197
vectors
column vectors, 2
row vectors, 2
vertex figure, 54
vertex set, 51
vertices, 5, 51
visible, 240
V-polyhedron, 29
V-polytope, 29
0/1-polytopes, 19, 25, 70
zonal diagrams, 224
zone, 206
zonotopal tiling, 218
zonotope, 17, 191, 199
associated, 226
generalized, 226
volume, 230

152 - Lectures On Polytopes PDF

Uploaded by

Copyright:

Available Formats

152 - Lectures On Polytopes PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

152 - Lectures On Polytopes PDF

Uploaded by

Copyright:

Available Formats

Ginter M.

Mathematics Subject Classification (1991) 52-02, 52B05, 52B11, 52B12

New York Berlin Heidelberg (hard cover)

Still, I have to make a disclaimer. Current research on polytopes is very

A simple way to tell a simple polytope from its graph,

Preface to the Second Printing

At the occasion of the second printing I took the opportunity to make

http://www.math.tu berlin.de/ - ziegler

Your contributions to this update are more than welcome.

Introduction and Examples

1 Polytopes, Polyhedra, and Cones

Steinitz' Theorem for 3-Polytopes

4.1 3-Connected Planar Graphs

Schlegel Diagrams for 4-Polytopes

Duality, Gale Diagrams, and Applications

(c) 2-Faces of 5-Polytopes Cannot be Prescribed

Fans, Arrangements, Zonotopes,

Sheliability and the Upper Bound Theorem

Fiber Polytopes, and Beyond

Convex polytopes are fundamental geometric objects- to a large extent the

0. Introduction and Examples

0 Introduction and Examples

That is, every affine subspace can be described both as an intersection of

Clearly, every intersection of convex sets is convex, and d itself is convex.

: K C K', If` convex}.

0 Introduction and Examples

For any finite set {X i ,. , xk} C K and parameters A 1 ,

conv(K) = {A i x ' + + AkXk {Xi,

conv(K) = {A i x ' +. + Anxn : n > 1, Ai >

The following gives two different versions of the definition of a polytope.

Definition 0.1. A V-polytope is the convex hull of a finite set of points

Introduction and Examples

A polytope is a point set P C a which can be presented either as a

0.2. Zero-dimensional polytopes are points, one-dimensional

0. Introduction and Examples

Two 2-polytopes are combinatorially equivalent if and only if they have

Example 0.3. The tetrahedron is a familiar geometric object (a 3-dimensional polytope) in 3

Similarly, its d-dimensional generalization forms the first (and simplest)

O. Introduction and Examples

define a d-simplex as the convex hull of any d + 1 affinely independent

d+1 : 11 X = 1, Xi > 0} = COriV{ei, ,ed i i}

Our figures illustrate the construction of 6. 2 in

Their generalization to d dimensions is straightforward. We arrive at the

E Rd : 1 < xi < 1} = conv{{1, 1}d},

0. Introduction and Examples

and the d-dimensional crosspolytope:

lxil < 11 = conv{e i , e l , ... , ed , ed}.

f' -i- fa E ( d )* : ax < 1 for all

Introduction and Examples

Griinbaum's article [239] for an amusing account of how difficult it is to

Similarly we construct the bipyramid bipyr(P) by choosing two points x +

Especially important, it is quite obvious how to define the product of two

Introduction and Examples

The cubes can be interpreted as iterated prisms, starting with a point.

0 Introduction and Examples

The cyclic polytope Cd(t i ,. . . ,t,i ) is the convex hull

Cd(t i ,...,t,i ) := cony fx(ti),x(t2),...,x(tn)}

The problem is that in dimension 3 we cannot really see why cyclic

0. Introduction and Examples