Comb Topol

Aspects of Convex Geometry
Polyhedra, Linear Programming,

Shellings, Voronoi Diagrams,
Delaunay Triangulations
Jean Gallier and Jocelyn Quaintance

Department of Computer and Information Science
University of Pennsylvania
Philadelphia, PA 19104, USA
e-mail: jean@cis.upenn.edu
August 8, 2017
2
Aspects of Convex Geometry
Polyhedra, Linear Programming,
Shellings, Voronoi Diagrams,
Jean Gallier and Jocelyn Quaintance
Abstract: Some basic mathematical tools such as convex sets, polytopes and combinatorial
topology, are used quite heavily in applied fields such as geometric modeling, meshing, com-
puter vision, medical imaging and robotics. This report may be viewed as a tutorial and a
set of notes on convex sets, polytopes, polyhedra, combinatorial topology, Voronoi Diagrams
and Delaunay Triangulations. It is intended for a broad audience of mathematically inclined
readers.
One of my (selfish!) motivations in writing these notes was to understand the concept
of shelling and how it is used to prove the famous Euler-Poincaré formula (Poincaré, 1899)
and the more recent Upper Bound Theorem (McMullen, 1970) for polytopes. Another of my
motivations was to give a “correct” account of Delaunay triangulations and Voronoi diagrams
in terms of (direct and inverse) stereographic projections onto a sphere and prove rigorously
that the projective map that sends the (projective) sphere to the (projective) paraboloid
works correctly, that is, maps the Delaunay triangulation and Voronoi diagram w.r.t. the
lifting onto the sphere to the Delaunay diagram and Voronoi diagrams w.r.t. the traditional
lifting onto the paraboloid. Here, the problem is that this map is only well defined (total) in
projective space and we are forced to define the notion of convex polyhedron in projective
space.
It turns out that in order to achieve (even partially) the above goals, I found that it was
necessary to include quite a bit of background material on convex sets, polytopes, polyhedra
and projective spaces. I have included a rather thorough treatment of the equivalence of
V-polytopes and H-polytopes and also of the equivalence of V-polyhedra and H-polyhedra,
which is a bit harder. In particular, the Fourier-Motzkin elimination method (a version of
Gaussian elimination for inequalities) is discussed in some detail. I also had to include some
material on projective spaces, projective maps and polar duality w.r.t. a nondegenerate
quadric in order to define a suitable notion of “projective polyhedron” based on cones. To
the best of our knowledge, this notion of projective polyhedron is new. We also believe that
some of our proofs establishing the equivalence of V-polyhedra and H-polyhedra are new.
Since Chapters 2, 3, 4, and 5 contain all the background (and more) needed to discuss
linear programming (including the simplex algorithm and duality), we have included some
chapters on linear programming.
Key-words: Convex sets, polytopes, polyhedra, linear programming, simplex algorithm,

strong duality, shellings, combinatorial topology, Voronoi diagrams, Delaunay triangulations.
3
4 CONTENTS
Contents
Contents 4
1 Introduction 7
1.1 Motivations and Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Basics of Affine Geometry 11

2.1 Affine Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Examples of Affine Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Chasles’s Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4 Affine Combinations, Barycenters . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Affine Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.6 Affine Independence and Affine Frames . . . . . . . . . . . . . . . . . . . . . 33
2.7 Affine Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.8 Affine Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.9 Affine Geometry: A Glimpse . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.10 Affine Hyperplanes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.11 Intersection of Affine Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3 Basic Properties of Convex Sets 57

3.1 A Review of Basic Topological Concepts . . . . . . . . . . . . . . . . . . . . 57
3.2 Convex Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3 Carathéodory’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4 Vertices, Extremal Points and Krein and Milman’s Theorem . . . . . . . . . 64
3.5 Radon’s, Tverberg’s, Helly’s, Theorems and Centerpoints . . . . . . . . . . . 71
4 Two Main Tools: Separation and Polar Duality 79

4.1 Separation Theorems and Farkas Lemma . . . . . . . . . . . . . . . . . . . . 79
4.2 Supporting Hyperplanes and Minkowski’s Proposition . . . . . . . . . . . . . 95
4.3 Polarity and Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5 Polyhedra and Polytopes 103

5.1 Polyhedra, H-Polytopes and V-Polytopes . . . . . . . . . . . . . . . . . . . . 103
5.2 Polar Duals of V-Polytopes and H-Polyhedra . . . . . . . . . . . . . . . . . . 111
5.3 The Equivalence of H-Polytopes and V-Polytopes . . . . . . . . . . . . . . . 113
CONTENTS 5
5.4 The Equivalence of H-Polyhedra and V-Polyhedra . . . . . . . . . . . . . . . 114

5.5 Fourier–Motzkin Elimination and Cones . . . . . . . . . . . . . . . . . . . . 130
5.6 Lineality Space and Recession Cone . . . . . . . . . . . . . . . . . . . . . . . 139
6 Linear Programs 143

6.1 What is Linear Programming? . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.2 Notational Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.3 Linear Programs, Feasible Solutions, Optimal Solutions . . . . . . . . . . . . 146
6.4 Basic Feasible Solutions and Vertices . . . . . . . . . . . . . . . . . . . . . . 152
7 The Simplex Algorithm 161

7.1 The Idea Behind the Simplex Algorithm . . . . . . . . . . . . . . . . . . . . 161
7.2 The Simplex Algorithm in General . . . . . . . . . . . . . . . . . . . . . . . 170
7.3 How Perform a Pivoting Step Efficiently . . . . . . . . . . . . . . . . . . . . 177
7.4 The Simplex Algorithm Using Tableaux . . . . . . . . . . . . . . . . . . . . . 180
7.5 Computational Efficiency of the Simplex Method . . . . . . . . . . . . . . . 190
8 Linear Programming and Duality 193

8.1 Variants of the Farkas Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.2 The Duality Theorem in Linear Programming . . . . . . . . . . . . . . . . . 198
8.3 Complementary Slackness Conditions . . . . . . . . . . . . . . . . . . . . . . 206
8.4 Duality for Linear Programs in Standard Form . . . . . . . . . . . . . . . . . 208
8.5 The Dual Simplex Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 211
8.6 The Primal-Dual Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
9 Basics of Combinatorial Topology 227

9.1 Simplicial Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
9.2 Nonsingular Faces; Stars and Links . . . . . . . . . . . . . . . . . . . . . . . 232
9.3 Polyhedral Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
9.4 Combinatorial and Topological Manifolds . . . . . . . . . . . . . . . . . . . . 253
10 Shellings and the Euler–Poincaré Formula 257

10.1 Shellings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
10.2 The Euler–Poincaré Formula for Polytopes . . . . . . . . . . . . . . . . . . . 266
10.3 Dehn–Sommerville Equations for Simplicial Polytopes . . . . . . . . . . . . . 269
10.4 The Upper Bound Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
11 Projective Spaces and Polyhedra, Polar Duality 285

11.1 Projective Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
11.2 Projective Polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
11.3 Tangent Spaces of Hypersurfaces . . . . . . . . . . . . . . . . . . . . . . . . 306
11.4 Quadrics (Affine, Projective) and Polar Duality . . . . . . . . . . . . . . . . 312
6 CONTENTS
12 Dirichlet–Voronoi Diagrams 321

12.1 Dirichlet–Voronoi Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
12.2 Triangulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
12.3 Delaunay Triangulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
12.4 Delaunay Triangulations and Convex Hulls . . . . . . . . . . . . . . . . . . . 332
12.5 Stereographic Projection and the Space of Spheres . . . . . . . . . . . . . . . 336
12.6 Relating Lifting to a Paraboloid and Lifting to a Sphere . . . . . . . . . . . 352
12.7 Lifted Delaunay Complexes and Delaunay Complexes . . . . . . . . . . . . . 364
12.8 Lifted Voronoi Complexes and Voronoi Complexes . . . . . . . . . . . . . . . 375
12.9 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Bibliography 387
Chapter 1
Introduction
1.1 Motivations and Goals

For the past eight years or so I have been teaching a graduate course whose main goal is to
expose students to some fundamental concepts of geometry, keeping in mind their applica-
tions to geometric modeling, meshing, computer vision, medical imaging, robotics, etc. The
audience has been primarily computer science students but a fair number of mathematics
students and also students from other engineering disciplines (such as Electrical, Systems,
Mechanical and Bioengineering) have been attending my classes. In the past three years,
I have been focusing more on convexity, polytopes and combinatorial topology, as concepts
and tools from these areas have been used increasingly in meshing and also in computational
biology and medical imaging. One of my (selfish!) motivations was to understand the con-
cept of shelling and how it is used to prove the famous Euler-Poincaré formula (Poincaré,
1899) and the more recent Upper Bound Theorem (McMullen, 1970) for polytopes. Another
of my motivations was to give a “correct” account of Delaunay triangulations and Voronoi
diagrams in terms of (direct and inverse) stereographic projections onto a sphere and prove
rigorously that the projective map that sends the (projective) sphere to the (projective)
paraboloid works correctly, that is, maps the Delaunay triangulation and Voronoi diagram
w.r.t. the lifting onto the sphere to the Delaunay triangulation and Voronoi diagram w.r.t.
the lifting onto the paraboloid. Moreover, the projections of these polyhedra onto the hy-
perplane xd+1 = 0, from the sphere or from the paraboloid, are identical. Here, the problem
is that this map is only well defined (total) in projective space and we are forced to define
the notion of convex polyhedron in projective space.
It turns out that in order to achieve (even partially) the above goals, I found that it was
necessary to include quite a bit of background material on convex sets, polytopes, polyhedra
and projective spaces. I have included a rather thorough treatment of the equivalence of
V-polytopes and H-polytopes and also of the equivalence of V-polyhedra and H-polyhedra,
which is a bit harder. In particular, the Fourier-Motzkin elimination method (a version of
Gaussian elimination for inequalities) is discussed in some detail. I also had to include some
material on projective spaces, projective maps and polar duality w.r.t. a nondegenerate
7
8 CHAPTER 1. INTRODUCTION
quadric, in order to define a suitable notion of “projective polyhedron” based on cones. This
notion turned out to be indispensible to give a correct treatment of the Delaunay and Voronoi
complexes using inverse stereographic projection onto a sphere and to prove rigorously that
the well known projective map between the sphere and the paraboloid maps the Delaunay
triangulation and the Voronoi diagram w.r.t. the sphere to the more traditional Delaunay
triangulation and Voronoi diagram w.r.t. the paraboloid. To the best of our knowledge, this
notion of projective polyhedron is new. We also believe that some of our proofs establishing
the equivalence of V-polyhedra and H-polyhedra are new.
Chapter 9 on combinatorial topology is hardly original. However, most texts covering
this material are either old fashion or too advanced. Yet, this material is used extensively in
meshing and geometric modeling. We tried to give a rather intuitive yet rigorous exposition.
We decided to introduce the terminology combinatorial manifold , a notion usually referred
to as triangulated manifold .
A recurring theme in these notes is the process of “conification” (algebraically, “homoge-
nization”), that is, forming a cone from some geometric object. Indeed, “conification” turns
an object into a set of lines, and since lines play the role of points in projective geome-
try, “conification” (“homogenization”) is the way to “projectivize” geometric affine objects.
Then, these (affine) objects appear as “conic sections” of cones by hyperplanes, just the way
the classical conics (ellipse, hyperbola, parabola) appear as conic sections.
It is worth warning our readers that convexity and polytope theory is deceptively simple.
This is a subject where most intuitive propositions fail as soon as the dimension of the space
is greater than 3 (definitely 4), because our human intuition is not very good in dimension
greater than 3. Furthermore, rigorous proofs of seemingly very simple facts are often quite
complicated and may require sophisticated tools (for example, shellings, for a correct proof
of the Euler-Poincaré formula). Nevertheless, readers are urged to strenghten their geometric
intuition; they should just be very vigilant! This is another case where Tate’s famous saying
is more than pertinent: “Reason geometrically, prove algebraically.”
At first, these notes were meant as a complement to Chapter 3 (Properties of Convex
Sets: A Glimpse) of my book (Geometric Methods and Applications, [30]). However, they
turn out to cover much more material. For the reader’s convenience, I have included Chapter
2 on affine geometry, and Chapter 3 (both from my book [30]) as part of Chapter 3 of these
notes.
Since Chapters 2, 3, 4, and 5 contain all the background (and more) needed to discuss
linear programming (including the simplex algorithm and duality), we have included some
chapters on linear programming.
Most of the material on convex sets is taken from Berger [8] (Geometry II ). Other rel-
evant sources include Ziegler [67], Grünbaum [35] Barvinok [4], Valentine [63], Rockafellar
[49], Bourbaki (Topological Vector Spaces) [13], and Lax [39], the last four dealing with
affine spaces of infinite dimension. As to polytopes and polyhedra, “the” classic reference is
1.1. MOTIVATIONS AND GOALS 9
Grünbaum [35]. Other good references include Ziegler [67], Ewald [26], Cromwell [22], and
Thomas [60].
The recent book by Thomas contains an excellent and easy going presentation of poly-
tope theory. This book also gives an introduction to the theory of triangulations of point
configurations, including the definition of secondary polytopes and state polytopes, which
happen to play a role in certain areas of biology. For this, a quick but very efficient presen-
tation of Gröbner bases is provided. We highly recommend Thomas’s book [60] as further
reading. It is also an excellent preparation for the more advanced book by Sturmfels [59].
However, in our opinion, the “bible” on polytope theory is without any contest, Ziegler [67],
a masterly and beautiful piece of mathematics. In fact, our Chapter 10 is heavily inspired
by Chapter 8 of Ziegler. However, the pace of Ziegler’s book is quite brisk and we hope that
our more pedestrian account will inspire readers to go back and read the masters.
In a not too distant future, I would like to write about constrained Delaunay triangula-
tions, a formidable topic, please be patient!
I wish to thank Marcelo Siqueira for catching many typos and mistakes and for his
many helpful suggestions regarding the presentation. At least a third of this manuscript was
written while I was on sabbatical at INRIA, Sophia Antipolis, in the Asclepios Project. My
deepest thanks to Nicholas Ayache and his colleagues (especially Xavier Pennec and Hervé
Delingette) for inviting me to spend a wonderful and very productive year and for making
me feel perfectly at home within the Asclepios Project.
10 CHAPTER 1. INTRODUCTION
Chapter 2
Basics of Affine Geometry
L’algèbre n’est qu’une géométrie écrite; la géométrie n’est qu’une algèbre figurée.
—Sophie Germain
2.1 Affine Spaces

Geometrically, curves and surfaces are usually considered to be sets of points with some
special properties, living in a space consisting of “points.” Typically, one is also interested
in geometric properties invariant under certain transformations, for example, translations,
rotations, projections, etc. One could model the space of points as a vector space, but this is
not very satisfactory for a number of reasons. One reason is that the point corresponding to
the zero vector (0), called the origin, plays a special role, when there is really no reason to have
a privileged origin. Another reason is that certain notions, such as parallelism, are handled
in an awkward manner. But the deeper reason is that vector spaces and affine spaces really
have different geometries. The geometric properties of a vector space are invariant under
the group of bijective linear maps, whereas the geometric properties of an affine space are
invariant under the group of bijective affine maps, and these two groups are not isomorphic.
Roughly speaking, there are more affine maps than linear maps.
Affine spaces provide a better framework for doing geometry. In particular, it is possible
to deal with points, curves, surfaces, etc., in an intrinsic manner, that is, independently
of any specific choice of a coordinate system. As in physics, this is highly desirable to
really understand what is going on. Of course, coordinate systems have to be chosen to
finally carry out computations, but one should learn to resist the temptation to resort to
coordinate systems until it is really necessary.
Affine spaces are the right framework for dealing with motions, trajectories, and physical
forces, among other things. Thus, affine geometry is crucial to a clean presentation of
kinematics, dynamics, and other parts of physics (for example, elasticity). After all, a rigid
motion is an affine map, but not a linear map in general. Also, given an m × n matrix A
11
12 CHAPTER 2. BASICS OF AFFINE GEOMETRY
and a vector b ∈ Rm , the set U = {x ∈ Rn | Ax = b} of solutions of the system Ax = b is an

affine space, but not a vector space (linear space) in general.
Use coordinate systems only when needed!
This chapter proceeds as follows. We take advantage of the fact that almost every affine
concept is the counterpart of some concept in linear algebra. We begin by defining affine
spaces, stressing the physical interpretation of the definition in terms of points (particles)
and vectors (forces). Corresponding to linear combinations of vectors, we define affine com-
binations of points (barycenters), realizing that we are forced to restrict our attention to
families of scalars adding up to 1. Corresponding to linear subspaces, we introduce affine
subspaces as subsets closed under affine combinations. Then, we characterize affine sub-
spaces in terms of certain vector spaces called their directions. This allows us to define a
clean notion of parallelism. Next, corresponding to linear independence and bases, we define
affine independence and affine frames. We also define convexity. Corresponding to linear
maps, we define affine maps as maps preserving affine combinations. We show that every
affine map is completely defined by the image of one point and a linear map. Then, we
investigate briefly some simple affine maps, the translations and the central dilatations. At
this point, we give a glimpse of affine geometry. We prove the theorems of Thales, Pappus,
and Desargues. After this, the definition of affine hyperplanes in terms of affine forms is
reviewed. The section ends with a closer look at the intersection of affine subspaces.
Our presentation of affine geometry is far from being comprehensive, and it is biased
toward the algorithmic geometry of curves and surfaces. For more details, the reader is
referred to Pedoe [46], Snapper and Troyer [53], Berger [7, 8], Coxeter [21], Samuel [50],
Tisseron [62], Fresnel [27], Vienne [65], and Hilbert and Cohn-Vossen [36].
Suppose we have a particle moving in 3D space and that we want to describe the trajectory
of this particle. If one looks up a good textbook on dynamics, such as Greenwood [34], one
finds out that the particle is modeled as a point, and that the position of this point x is
determined with respect to a “frame” in R3 by a vector. Curiously, the notion of a frame is
rarely defined precisely, but it is easy to infer that a frame is a pair (O, (e1 , e2 , e3 )) consisting
of an origin O (which is a point) together with a basis of three vectors (e1 , e2 , e3 ). For
example, the standard frame in R3 has origin O = (0, 0, 0) and the basis of three vectors
e1 = (1, 0, 0), e2 = (0, 1, 0), and e3 = (0, 0, 1). The position of a point x is then defined by
the “unique vector” from O to x.
But wait a minute, this definition seems to be defining frames and the position of a point
without defining what a point is! Well, let us identify points with elements of R3 . If so, given
any two points a = (a1 , a2 , a3 ) and b = (b1 , b2 , b3 ), there is a unique free vector , denoted by
→
− →
−
ab, from a to b, the vector ab = (b1 − a1 , b2 − a2 , b3 − a3 ). Note that
→
−
b = a + ab,
2.1. AFFINE SPACES 13
b
−
→
ab
a
O
Figure 2.1: Points and free vectors.
addition being understood as addition in R3 . Then, in the standard frame, given a point
−→
x = (x1 , x2 , x3 ), the position of x is the vector Ox = (x1 , x2 , x3 ), which coincides with the
point itself. In the standard frame, points and vectors are identified. Points and free vectors
are illustrated in Figure 2.1.
What if we pick a frame with a different origin, say Ω = (ω1 , ω2 , ω3 ), but the same basis
vectors (e1 , e2 , e3 )? This time, the point x = (x1 , x2 , x3 ) is defined by two position vectors:
−→
Ox = (x1 , x2 , x3 )
in the frame (O, (e1 , e2 , e3 )) and
−
→
Ωx = (x1 − ω1 , x2 − ω2 , x3 − ω3 )
in the frame (Ω, (e1 , e2 , e3 )). See Figure 2.2.
This is because
−→ −→ − → −→
Ox = OΩ + Ωx and OΩ = (ω1 , ω2 , ω3 ).
We note that in the second frame (Ω, (e1 , e2 , e3 )), points and position vectors are no longer
identified. This gives us evidence that points are not vectors. It may be computationally
convenient to deal with points using position vectors, but such a treatment is not frame
invariant, which has undesirable effects.
Inspired by physics, we deem it important to define points and properties of points that
are frame invariant. An undesirable side effect of the present approach shows up if we attempt
to define linear combinations of points. First, let us review the notion of linear combination
of vectors. Given two vectors u and v of coordinates (u1 , u2 , u3 ) and (v1 , v2 , v3 ) with respect
e3
e2
Ω
Ωx e1
x
e3
Ox
e1
O e2
Figure 2.2: The two position vectors for the point x.
to the basis (e1 , e2 , e3 ), for any two scalars λ, µ, we can define the linear combination λu + µv
as the vector of coordinates
(λu1 + µv1 , λu2 + µv2 , λu3 + µv3 ).
If we choose a different basis (e01 , e02 , e03 ) and if the matrix P expressing the vectors (e01 , e02 , e03 )
over the basis (e1 , e2 , e3 ) is
 
a1 b 1 c 1
P = a2 b2 c2  ,
a3 b 3 c 3
which means that the columns of P are the coordinates of the e0j over the basis (e1 , e2 , e3 ),
since
u1 e1 + u2 e2 + u3 e3 = u01 e01 + u02 e02 + u03 e03
and
v1 e1 + v2 e2 + v3 e3 = v10 e01 + v20 e02 + v30 e03 ,
it is easy to see that the coordinates (u1 , u2 , u3 ) and (v1 , v2 , v3 ) of u and v with respect to
the basis (e1 , e2 , e3 ) are given in terms of the coordinates (u01 , u02 , u03 ) and (v10 , v20 , v30 ) of u and
v with respect to the basis (e01 , e02 , e03 ) by the matrix equations
   0    0
u1 u1 v1 v1
u2  = P u02  and v2  = P v20  .
u3 u03 v3 v30
From the above, we get
 0    0  
u1 u1 v1 v1
u02  = P −1 u2  and v20  = P −1 v2  ,
u03 u3 v30 v3
and by linearity, the coordinates
(λu01 + µv10 , λu02 + µv20 , λu03 + µv30 )
of λu + µv with respect to the basis (e01 , e02 , e03 ) are given by

 0
λu1 + µv10
      
u1 v1 λu1 + µv1
λu02 + µv20  = λP −1 u2  + µP −1 v2  = P −1 λu2 + µv2  .
λu03 + µv30 u3 v3 λu3 + µv3
Everything worked out because the change of basis does not involve a change of origin. On the
other hand, if we consider the change of frame from the frame (O, (e1 , e2 , e3 )) to the frame
−→
(Ω, (e1 , e2 , e3 )), where OΩ = (ω1 , ω2 , ω3 ), given two points a, b of coordinates (a1 , a2 , a3 )
and (b1 , b2 , b3 ) with respect to the frame (O, (e1 , e2 , e3 )) and of coordinates (a01 , a02 , a03 ) and
(b01 , b02 , b03 ) with respect to the frame (Ω, (e1 , e2 , e3 )), since
(a01 , a02 , a03 ) = (a1 − ω1 , a2 − ω2 , a3 − ω3 )
and
(b01 , b02 , b03 ) = (b1 − ω1 , b2 − ω2 , b3 − ω3 ),
the coordinates of λa + µb with respect to the frame (O, (e1 , e2 , e3 )) are
(λa1 + µb1 , λa2 + µb2 , λa3 + µb3 ),
but the coordinates

(λa01 + µb01 , λa02 + µb02 , λa03 + µb03 )
of λa + µb with respect to the frame (Ω, (e1 , e2 , e3 )) are
(λa1 + µb1 − (λ + µ)ω1 , λa2 + µb2 − (λ + µ)ω2 , λa3 + µb3 − (λ + µ)ω3 ),
which are different from
(λa1 + µb1 − ω1 , λa2 + µb2 − ω2 , λa3 + µb3 − ω3 ),
unless λ + µ = 1. See Figure 2.3.

Thus, we have discovered a major difference between vectors and points: The notion of
linear combination of vectors is basis independent, but the notion of linear combination of
points is frame dependent. In order to salvage the notion of linear combination of points,
some restriction is needed: The scalar coefficients must add up to 1.
e3
e2
Ω = (3,4,5)
e1
a + b = (3,4,2) = (0, 0, -3)

e3
a = (1,1,1)
e1 b = (2,3,1)
O e2
e3
e2
Ω = (3,4,5)
e1
e3
a = (-2,-3,-4)
b = (-1,-1,-4)
e1
O e2
a + b = (-3, -4, -8) = (0, 0, -3)
Figure 2.3: The top figure shows the location of the “point” sum a + b with respect to the
frame (O, (e1 , e2 , e3 )), while the bottom figure shows the location of the “point” sum a + b
with respect to the frame (Ω, (e1 , e2 , e3 )).
A clean way to handle the problem of frame invariance and to deal with points in a more
intrinsic manner is to make a clearer distinction between points and vectors. We duplicate
R3 into two copies, the first copy corresponding to points, where we forget the vector space
structure, and the second copy corresponding to free vectors, where the vector space structure
is important. Furthermore, we make explicit the important fact that the vector space R3 acts
on the set of points R3 : Given any point a = (a1 , a2 , a3 ) and any vector v = (v1 , v2 , v3 ),
we obtain the point
a + v = (a1 + v1 , a2 + v2 , a3 + v3 ),
which can be thought of as the result of translating a to b using the vector v. We can imagine
that v is placed such that its origin coincides with a and that its tip coincides with b. This
action + : R3 × R3 → R3 satisfies some crucial properties. For example,
a + 0 = a,
(a + u) + v = a + (u + v),
→
−
and for any two points a, b, there is a unique free vector ab such that
→
−
b = a + ab.
It turns out that the above properties, although trivial in the case of R3 , are all that is
needed to define the abstract notion of affine space (or affine structure). The basic idea is
→
−
to consider two (distinct) sets E and E , where E is a set of points (with no structure) and
→
−
E is a vector space (of free vectors) acting on the set E.
Did you say “A fine space”?
→
−
Intuitively, we can think of the elements of E as forces moving the points in E, considered
→
−
as physical particles. The effect of applying a force (free vector) u ∈ E to a point a ∈ E is
→
−
a translation. By this, we mean that for every force u ∈ E , the action of the force u is to
“move” every point a ∈ E to the point a + u ∈ E obtained by the translation corresponding
→
−
to u viewed as a vector. Since translations can be composed, it is natural that E is a vector
space.
For simplicity, it is assumed that all vector spaces under consideration are defined over
the field R of real numbers. Most of the definitions and results also hold for an arbitrary
field K, although some care is needed when dealing with fields of characteristic different
from zero. It is also assumed that all families (λi )i∈I of scalars have finite support. Recall
that a family (λi )i∈I of scalars has finite support if λi = 0 for all i ∈ I − J, where J is a finite
subset of I. Obviously, finite families of scalars have finite support, and for simplicity, the
reader may assume that all families of scalars are finite. The formal definition of an affine
space is as follows.
Definition 2.1. An affine space is either the degenerate space reduced to the empty set, or a

→ − →
−
triple E, E , + consisting of a nonempty set E (of points), a vector space E (of translations,
→
−
or free vectors), and an action + : E × E → E, satisfying the following conditions.
(A1) a + 0 = a, for every a ∈ E.
→
−
(A2) (a + u) + v = a + (u + v), for every a ∈ E, and every u, v ∈ E .
→
−
(A3) For any two points a, b ∈ E, there is a unique u ∈ E such that a + u = b.
→
− →
−
The unique vector u ∈ E such that a + u = b is denoted by ab, or sometimes by ab, or
even by b − a. Thus, we also write
→
−
b = a + ab
(or b = a + ab, or even b = a + (b − a)).

→− →
−
The dimension of the affine space E, E , + is the dimension dim( E ) of the vector space
→
−
E . For simplicity, it is denoted by dim(E).
−
→
E E
b=a+u
u
a c=a+w w
Figure 2.4: Intuitive picture of an affine space.
→
−
Conditions (A1) and (A2) say that the (abelian) group E acts on E, and Condition (A3)
→
−
says that E acts transitively and faithfully on E. Note that
−−−−−→
a(a + v) = v
→
− −−−−−→ −−−−−→
for all a ∈ E and all v ∈ E , since a(a + v) is the unique vector such that a+v = a+ a(a + v).
→
−
Thus, b = a + v is equivalent to ab = v. Figure 2.4 gives an intuitive picture of an affine
space. It is natural to think of all vectors as having the same origin, the null vector.

→ −
The axioms defining an affine space E, E , + can be interpreted intuitively as saying
→
−
that E and E are two different ways of looking at the same object, but wearing different
sets of glasses, the second set of glasses depending on the choice of an “origin” in E. Indeed,
we can choose to look at the points in E, forgetting that every pair (a, b) of points defines a
→
− →
− →
−
unique vector ab in E , or we can choose to look at the vectors u in E , forgetting the points
in E. Furthermore, if we also pick any point a in E, a point that can be viewed as an origin
→
−
in E, then we can recover all the points in E as the translated points a + u for all u ∈ E .
→
−
This can be formalized by defining two maps between E and E .
→
−
For every a ∈ E, consider the mapping from E to E given by
u 7→ a + u,
→
− →
−
where u ∈ E , and consider the mapping from E to E given by
→
−
b 7→ ab,
where b ∈ E. The composition of the first mapping with the second is

−−−−−→
u 7→ a + u 7→ a(a + u),
which, in view of (A3), yields u. The composition of the second with the first mapping is
→
− →
−
b 7→ ab 7→ a + ab,
→
− →
−
which, in view of (A3), yields b. Thus, these compositions are the identity from E to E
and the identity from E to E, and the mappings are both bijections.
→
− →
−
When we identify E with E via the mapping b 7→ ab, we say that we consider E as the
vector space obtained by taking a as the origin in E, and we denote it by Ea . Because Ea is
a vector space, to be consistent with our notational conventions we should use the notation
−
→
Ea (using an arrow), instead of Ea . However, for simplicity, we stick to the notation Ea .

→ −
Thus, an affine space E, E , + is a way of defining a vector space structure on a set of
points E, without making a commitment to a fixed origin in E. Nevertheless, as soon as
we commit to an origin a in E, we can view E as the vector space Ea . However, we urge
→
−
the reader to think of E as a physical set of points and of E as a set of forces acting on E,
rather than reducing E to some isomorphic copy of Rn . After all, points are points, and not

→ − →
−
vectors! For notational simplicity, we will often denote an affine space E, E , + by (E, E ),
→
−
or even by E. The vector space E is called the vector space associated with E.
One should be careful about the overloading of the addition symbol +. Addition
is well-defined on vectors, as in u + v; the translate a + u of a point a ∈ E by a
→
−
vector u ∈ E is also well-defined, but addition of points a + b does not make sense. In
this respect, the notation b − a for the unique vector u such that b = a + u is somewhat
confusing, since it suggests that points can be subtracted (but not added!).
→
− →
−
Any vector space E has an affine space structure specified by choosing E = E , and
→
−
→− →−
letting + be addition in the vector space E . We will refer to the affine structure E , E , +
→
− →
−
on a vector space E as the canonical (or natural) affine
structure
on E . In particular, the
vector space Rn can be viewed as
the affine space Rn , Rn , + , denoted by An . In general,
if K is any field, the affine space K n , K n , + is denoted by AnK . In order to distinguish
between the double role played by members of Rn , points and vectors, we will denote points
by row vectors, and vectors by column vectors. Thus, the action of the vector space Rn over
the set Rn simply viewed as a set of points is given by
 
u1
 .. 
(a1 , . . . , an ) +  .  = (a1 + u1 , . . . , an + un ).
un
We will also use the convention that if x = (x1 , . . . , xn ) ∈ Rn , then the column vector
associated with x is denoted by x (in boldface notation). Abusing the notation slightly, if
a ∈ Rn is a point, we also write a ∈ An . The affine space An is called the real affine space of
dimension n. In most cases, we will consider n = 1, 2, 3.
Figure 2.5: An affine space: the line of equation x + y − 1 = 0
2.2 Examples of Affine Spaces

Let us now give an example of an affine space that is not given as a vector space (at least, not
in an obvious fashion). Consider the subset L of A2 consisting of all points (x, y) satisfying
the equation
x + y − 1 = 0.
The set L is the line of slope −1 passing through the points (1, 0) and (0, 1) shown in Figure
2.5.
The line L can be made into an official affine space by defining the action + : L × R → L
of R on L defined such that for every point (x, 1 − x) on L and any u ∈ R,
(x, 1 − x) + u = (x + u, 1 − x − u).
It is immediately verified that this action makes L into an affine space. For example, for any
two points a = (a1 , 1 − a1 ) and b = (b1 , 1 − b1 ) on L, the unique (vector) u ∈ R such that
b = a + u is u = b1 − a1 . Note that the vector space R is isomorphic to the line of equation
x + y = 0 passing through the origin.
Similarly, consider the subset H of A3 consisting of all points (x, y, z) satisfying the
equation
x + y + z − 1 = 0.
The set H is the plane passing through the points (1, 0, 0), (0, 1, 0), and (0, 0, 1). The plane
H can be made into an official affine space by defining the action + : H × R2 → H of R2 on
2.3. CHASLES’S IDENTITY 21
(0,0,1)
Figure 2.6: An affine space: the plane x + y + z − 1 = 0.

u
H defined such that for every point (x, y, 1 − x − y) on H and any ∈ R2 ,
v

u
(x, y, 1 − x − y) + = (x + u, y + v, 1 − x − u − y − v).
v
For a slightly wilder example, consider the subset P of A3 consisting of all points (x, y, z)
satisfying the equation
x2 + y 2 − z = 0.
The set P is a paraboloid of revolution, with axis Oz. The surface P can be made into an
official affine space by defining the action+: P × R2 → P of R2 on P defined such that for
u
every point (x, y, x2 + y 2 ) on P and any ∈ R2 ,
v

2 2 u
(x, y, x + y ) + = (x + u, y + v, (x + u)2 + (y + v)2 ).
v
This should dispel any idea that affine spaces are dull. Affine spaces not already equipped
with an obvious vector space structure arise in projective geometry.
2.3 Chasles’s Identity

→
− →
−
Given any three points a, b, c ∈ E, since c = a + →
−
ac, b = a + ab, and c = b + bc, we get
→
− →
− →
− →
− → −
c = b + bc = (a + ab) + bc = a + ( ab + bc)
2
(x + u, y + v, (x + u)2 + (y + v) )
P
(x, y, x 2 + y 2 )
(u,v
)
(x, y)
(x + u, y + v)
Figure 2.7: The paraboloid of revolution P viewed as a two-dimensional affine space.
by (A2), and thus, by (A3),

→
− → −
ab + bc = →
−
ac,
which is known as Chasles’s identity, and illustrated in Figure 2.8.
Since a = a + −
→ and by (A1) a = a + 0, by (A3) we get
aa
−
→ = 0.
aa
Thus, letting a = c in Chasles’s identity, we get

→
− →
−
ba = − ab.
Given any four points a, b, c, d ∈ E, since by Chasles’s identity

→
− → − −
→ → −
ab + bc = ad + dc = →
−
ac,
we have the parallelogram law

→
− →
− →
− −
→
ab = dc iff bc = ad.
2.4 Affine Combinations, Barycenters

A fundamental concept in linear algebra is that of a linear combination. The corresponding
concept in affine geometry is that of an affine combination, also called a barycenter . However,
there is a problem with the naive approach involving a coordinate system, as we saw in
Section 2.1. Since this problem is the reason for introducing affine combinations, at the risk
2.4. AFFINE COMBINATIONS, BARYCENTERS 23
−
→
E E
b
−
→
ab
−
→
ac
a c
−
→
bc
Figure 2.8: Points and corresponding vectors in affine geometry.
of boring certain readers, we give another example showing what goes wrong if we are not
careful in defining linear combinations of points.
Consider R2 as anaffine space,
under its natural coordinate system with origin O = (0, 0)
1 0
and basis vectors and . Given any two points a = (a1 , a2 ) and b = (b1 , b2 ), it is
0 1
natural to define the affine combination λa + µb as the point of coordinates
(λa1 + µb1 , λa2 + µb2 ).
Thus, when a = (−1, −1) and b = (2, 2), the point a + b is the point c = (1, 1).
Let us now consider the new coordinate system with respect to the origin c = (1, 1) (and
the same basis vectors). This time, the coordinates of a are (−2, −2), the coordinates of b
are (1, 1), and the point a + b is the point d of coordinates (−1, −1). However, it is clear that
the point d is identical to the origin O = (0, 0) of the first coordinate system. This situation
is illustrated in Figure 2.9.
Thus, a + b corresponds to two different points depending on which coordinate system is
used for its computation!
This shows that some extra condition is needed in order for affine combinations to make
sense. It turns out that if the scalars sum up to 1, the definition is intrinsic, as the following
proposition shows.
Proposition 2.1. Given an affine space E, let (ai )i∈I be a family of points in E, and let
(λi )i∈I be a family of scalars. For any two points a, b ∈ E, the following properties hold:
P
(1) If i∈I λi = 1, then
X − →
λi −→=b+
X
a+ aa i λi bai .
i∈I i∈I
b = (2,2) b = (1,1)
c = a + b = (1,1) c
O = (0,0) d = a + b = (-1,-1)
a = (-1,-1) a = (-2, -2)
Figure 2.9: The example from the beginning of Section 2.4.

P
(2) If i∈I λi = 0, then
−→
λi −→=
X X
aa i λi bai .
i∈I i∈I
Proof. (1) By Chasles’s identity (see Section 2.3), we have

X → − − →
λi −→=a+
X
a+ aa i λi ( ab + bai )
i∈I i∈I
→− X − →
X
=a+ λi ab + λi bai
i∈I i∈I
− X −
→ → P
= a + ab + λi bai since i∈I λi = 1
i∈I
X −→ →
−
=b+ λi bai since b = a + ab.
i∈I
An illustration of this calculation in A2 is provided by Figure 2.10.

(2) We also have
→
− − →
λi −→ =
X X
aa i λi ( ab + bai )
i∈I i∈I
→
− X − →
X
= λi ab + λi bai
i∈I i∈I
X −→
= λi bai ,
i∈I
P
since i∈I λi = 0.
2.4. AFFINE COMBINATIONS, BARYCENTERS 25
a2
a2
aa2 b a2 a3
a a3 a3 a
a1 a a1 a a1
aa
i ab
ai
ai b a3
ab ba1
b b
Figure 2.10: Part (1) of Proposition 2.1.
Thus, by Proposition
P 2.1, for any family of points (ai )i∈I in E, for any family (λi )i∈I of
scalars such that i∈I λi = 1, the point
λi −→
X
x=a+ aa i
i∈I
is independent of the choice of the origin a ∈ E. This property motivates the following
definition.
Definition
P 2.2. For any family of points (ai )i∈I in E, for any family (λi )i∈I of scalars such
that i∈I λi = 1, and for any a ∈ E, the point
λi −→
X
a+ aa i
i∈I
(which is independent of a ∈ E, by Proposition 2.1) is called the barycenter (or barycentric

combination, or affine combination) of the points ai assigned the weights λi , and it is denoted
by X
λi ai .
i∈I
In dealing with barycenters, it is convenient to introduce the notion of a weighted point,

which is just a pair (a, λ), where a ∈ E Pis a point, and λ ∈ R is a scalar. Then, given
P a family
of weighted points ((ai , λi ))i∈I , where i∈I λi = 1, we also say that the point i∈I λi ai is
the barycenter of the family of weighted points ((ai , λi ))i∈I .
Note that the barycenter x of the family of weighted points ((ai , λi ))i∈I is the unique
point such that
−
→= λi −→ for every a ∈ E,
X
ax aa i
i∈I
and setting a = x, the point x is the unique point such that
λi −→ = 0.
X
xa i
i∈I
In physical terms, the barycenter is the center of mass of theP family of weighted points
((ai , λi ))i∈I (where the masses have been normalized, so that i∈I λi = 1, and negative
masses are allowed).
Remarks:
(1) Since the barycenter of a family ((ai , λi ))i∈I of weighted

P points is defined for families
(λi )i∈I of scalars with finite support (and such that i∈I λi = 1), we might as well
assume that I is finite. Then, for all m ≥ 2, it is easy to prove that the barycenter
of m weighted points can be obtained by repeated computations of barycenters of two
weighted points.
(2) This result still holds, provided that the field K has at least three distinct elements,
but the proof is trickier!
P
(3) When i∈I λP
P −→
i = 0, the vector i∈I λi aai does not depend on the point a, and we may
denote it by i∈I λi ai . This observation will be used to define a vector space in which
linear
P combinations of both points and vectors make sense, regardless of the value of
i∈I λi .
Figure 2.11 illustrates the

geometric construction of the barycenters g1 and g2 of the
weighted points a, 14 , b, 41 , and c, 21 , and (a, −1), (b, 1), and (c, 1).

The point g1 can be constructed geometrically as the middle of the segment joining c to
the middle 12 a + 12 b of the segment (a, b), since

1 1 1 1
g1 = a + b + c.
2 2 2 2
The point g2 can be constructed geometrically as the point such that the middle 12 b + 12 c of
the segment (b, c) is the middle of the segment (a, g2 ), since

1 1
g2 = −a + 2 b + c .
2 2
Later on, we will see that a polynomial curve can be defined as a set of barycenters of a
fixed number of points. For example, let (a, b, c, d) be a sequence of points in A2 . Observe
that
(1 − t)3 + 3t(1 − t)2 + 3t2 (1 − t) + t3 = 1,
2.5. AFFINE SUBSPACES 27
g1
a b
c
g2
a b
Figure 2.11: Barycenters, g1 = 41 a + 14 b + 12 c, g2 = −a + b + c
since the sum on the left-hand side is obtained by expanding (t + (1 − t))3 = 1 using the
binomial formula. Thus,
(1 − t)3 a + 3t(1 − t)2 b + 3t2 (1 − t) c + t3 d
is a well-defined affine combination. Then, we can define the curve F : A → A2 such that
F (t) = (1 − t)3 a + 3t(1 − t)2 b + 3t2 (1 − t) c + t3 d.
Such a curve is called a Bézier curve, and (a, b, c, d) are called its control points. Note that
the curve passes through a and d, but generally not through b and c. It can be sbown
that any point F (t) on the curve can be constructed using an algorithm performing affine
interpolation steps (the de Casteljau algorithm).
2.5 Affine Subspaces

In linear algebra, a (linear) subspace can be characterized as a nonempty subset of a vector
space closed under linear combinations. In affine spaces, the notion corresponding to the
notion of (linear) subspace is the notion of affine subspace. It is natural to define an affine
subspace as a subset of an affine space closed under affine combinations.

→ −
Definition 2.3. Given an affine space E, E , + , a subset V of E is an affine subspace (of
→ − P
E, E , + ) if for
P every family of weighted points ((a i , λi ))i∈I in V such that i∈I λi = 1,
the barycenter i∈I λi ai belongs to V .
An affine subspace is also called a flat by some authors. According to Definition 2.3,
the empty set is trivially an affine subspace, and every intersection of affine subspaces is an
affine subspace.
As an example, consider the subset U of R2 defined by
U = (x, y) ∈ R2 | ax + by = c ,

i.e., the set of solutions of the equation
ax + by = c,
where it is assumed that a 6= 0 or b 6= 0. Given any m points (xi , yi ) ∈ U and any m scalars
λi such that λ1 + · · · + λm = 1, we claim that
m
X
λi (xi , yi ) ∈ U.
i=1
Indeed, (xi , yi ) ∈ U means that

axi + byi = c,
and if we multiply both sides of this equation by λi and add up the resulting m equations,
we get
X m m
X
(λi axi + λi byi ) = λi c,
i=1 i=1
and since λ1 + · · · + λm = 1, we get

m
! m
! m
!
X X X
a λ i xi + b λi y i = λi c = c,
i=1 i=1 i=1
which shows that !

m
X m
X m
X
λi xi , λi yi = λi (xi , yi ) ∈ U.
i=1 i=1 i=1
Thus, U is an affine subspace of A2 . In fact, it is just a usual line in A2 .

It turns out that U is closely related to the subset of R2 defined by
→
−
U = (x, y) ∈ R2 | ax + by = 0 ,

i.e., the set of solutions of the homogeneous equation
ax + by = 0
U
−
→
U
Figure 2.12: An affine line U and its direction.
obtained by setting the right-hand side of ax + by = c to zero. Indeed, for any m scalars λi ,
the same calculation as above yields that
m
X →
−
λi (xi , yi ) ∈ U ,
i=1
this time without any restriction on the λi , since the right-hand side of the equation is
→
− →
−
null. Thus, U is a subspace of R2 . In fact, U is one-dimensional, and it is just a usual line
in R2 . This line can be identified with a line passing through the origin of A2 , a line that is
parallel to the line U of equation ax + by = c, as illustrated in Figure 2.12.
Now, if (x0 , y0 ) is any point in U , we claim that
→
−
U = (x0 , y0 ) + U ,
where
→
− n →
−o
(x0 , y0 ) + U = (x0 + u1 , y0 + u2 ) | (u1 , u2 ) ∈ U .
→
− →
−
First, (x0 , y0 ) + U ⊆ U , since ax0 + by0 = c and au1 + bu2 = 0 for all (u1 , u2 ) ∈ U . Second,
if (x, y) ∈ U , then ax + by = c, and since we also have ax0 + by0 = c, by subtraction, we get
a(x − x0 ) + b(y − y0 ) = 0,
→
− →
−
which shows that (x − x0 , y − y0 ) ∈ U , and thus (x, y) ∈ (x0 , y0 ) + U . Hence, we also have
→
− →
−
U ⊆ (x0 , y0 ) + U , and U = (x0 , y0 ) + U .
The above example shows that the affine line U defined by the equation
ax + by = c
→
−
is obtained by “translating” the parallel line U of equation
ax + by = 0
passing through the origin. In fact, given any point (x0 , y0 ) ∈ U ,

→
−
U = (x0 , y0 ) + U .
More generally, it is easy to prove the following fact. Given any m × n matrix A and any
vector b ∈ Rm , the subset U of Rn defined by
U = {x ∈ Rn | Ax = b}
is an affine subspace of An .
Actually, observe that Ax = b should really be written as Ax> = b, to be consistent with
our convention that points are represented by row vectors. We can also use the boldface
notation for column vectors, in which case the equation is written as Ax = b. For the sake of
minimizing the amount of notation, we stick to the simpler (yet incorrect) notation Ax = b.
If we consider the corresponding homogeneous equation Ax = 0, the set
→
−
U = {x ∈ Rn | Ax = 0}
is a subspace of Rn , and for any x0 ∈ U , we have

→
−
U = x0 + U .
This is a general situation. Affine subspaces can be characterized in terms of subspaces of

→
−
E . Let V be a nonempty subset of E. For every family (a1 , . . . , an ) in V , for any family
(λ1 , . . . , λn ) of scalars, and for every point a ∈ V , observe that for every x ∈ E,
n
λi −→
X
x=a+ aa i
i=1
is the barycenter of the family of weighted points

n
X
(a1 , λ1 ), . . . , (an , λn ), a, 1 − λi ,
i=1
since n n
X X
λi + 1 − λi = 1.
i=1 i=1
→
− →
− →
−
Given any point a ∈ E and any subset V of E , let a + V denote the following subset of E:
→
− n →
−o
a+ V = a+v |v ∈ V .
−
→
E E
−
→
a V
−
→
V =a+ V
→
−
Figure 2.13: An affine subspace V and its direction V .

→ −
Proposition 2.2. Let E, E , + be an affine space.
(1) A nonempty subset V of E is an affine subspace iff for every point a ∈ V , the set
→
−
Va = {−
→|x∈V}
ax
→
− →
−
is a subspace of E . Consequently, V = a + Va . Furthermore,
→
−
V = {−
→ | x, y ∈ V }
xy
→
− →
− →
− →
−
is a subspace of E and Va = V for all a ∈ E. Thus, V = a + V .
→
− →
− →
−
(2) For any subspace V of E and for any a ∈ E, the set V = a + V is an affine subspace.
Proof. The proof is straightforward, and is omitted. It is also given in Gallier [29].
→
−
In particular, when E is the natural affine space associated with a vector space E ,
→
−
Proposition 2.2 shows that every affine subspace of E is of the form u + U , for a subspace
→
− →
− →
−
U of E . The subspaces of E are the affine subspaces of E that contain 0.
→
−
The subspace V associated with an affine subspace V is called the direction of V . It is
→
− →
−
→−
also clear that the map + : V × V → V induced by + : E × E → E confers to V, V , + an
affine structure. Figure 2.13 illustrates the notion of affine subspace.
→
−
By the dimension of the subspace V , we mean the dimension of V .
An affine subspace of dimension 1 is called a line, and an affine subspace of dimension 2
is called a plane.
An affine subspace of codimension 1 is called a hyperplane (recall that a subspace F of

a vector space E has codimension 1 iff there is some subspace G of dimension 1 such that
E = F ⊕ G, the direct sum of F and G, see Strang [58] or Lang [37]).
We say that two affine subspaces U and V are parallel if their directions are identical.
→
− →
− →
− →
−
Equivalently, since U = V , we have U = a + U and V = b + U for any a ∈ U and any
→
−
b ∈ V , and thus V is obtained from U by the translation ab.
In general, when we talk about n points a1 , . . . , an , we mean the sequence (a1 , . . . , an ),
and not the set {a1 , . . . , an } (the ai ’s need not be distinct).
→
−
By Proposition 2.2, a line is specified by a point a ∈ E and a nonzero vector v ∈ E , i.e.,
a line is the set of all points of the form a + λv, for λ ∈ R.
→
−
We say that three points a, b, c are collinear if the vectors ab and → −
ac are linearly depen-
dent. If two of the points a, b, c are distinct, say a 6= b, then there is a unique λ ∈ R such
→
−
that → − −
→
ac
ac = λ ab, and we define the ratio − → = λ.
ab
→
−
A plane is specified by a point a ∈ E and two linearly independent vectors u, v ∈ E , i.e.,
a plane is the set of all points of the form a + λu + µv, for λ, µ ∈ R.
→
− − −
→
We say that four points a, b, c, d are coplanar if the vectors ab, → ac, and ad are linearly
dependent. Hyperplanes will be characterized a little later.

→ −
Proposition 2.3. Given P an affine space
P E, E , + , for any family (ai )i∈I of points in E, the
set V of barycenters i∈I λi ai (where i∈I λi = 1) is the smallest affine subspace containing
(ai )i∈I .
P
Proof. If (ai )i∈I is empty, then V = ∅, because of the condition i∈I λi = 1. If (ai )i∈I is
nonempty, then
P the smallest affine subspace containing (ai )i∈I must contain the set V of
barycenters i∈I λi ai , and thus, it is enough to show that V is closed under affine combina-
tions, which is immediately verified.
Given a nonempty subset S of E, the smallest affine subspace of E generated by S is

often denoted by hSi. For example, a line specified by two distinct points a and b is denoted
by ha, bi, or even (a, b), and similarly for planes, etc.
Remarks:
(1) Since it can be shown that the barycenter of n weighted points can be obtained by
repeated computations of barycenters of two weighted points, a nonempty subset V
of E is an affine subspace iff for every two points a, b ∈ V , the set V contains all
barycentric combinations of a and b. If V contains at least two points, then V is an
affine subspace iff for any two distinct points a, b ∈ V , the set V contains the line
determined by a and b, that is, the set of all points (1 − λ)a + λb, λ ∈ R.
(2) This result still holds if the field K has at least three distinct elements, but the proof
is trickier!
2.6. AFFINE INDEPENDENCE AND AFFINE FRAMES 33
2.6 Affine Independence and Affine Frames

Corresponding to the notion of linear independence in vector spaces, we have the notion of
affine independence. Given a family (ai )i∈I of points in an affine space E, we will reduce the
notion of (affine) independence of these points to the (linear) independence of the families
(−
a−→
i aj )j∈(I−{i}) of vectors obtained by choosing any ai as an origin. First, the following
proposition shows that it is sufficient to consider only one of these families.

→−
Proposition 2.4. Given an affine space E, E , + , let (ai )i∈I be a family of points in E. If
the family (−
a−→ −−→
i aj )j∈(I−{i}) is linearly independent for some i ∈ I, then (ai aj )j∈(I−{i}) is linearly
independent for every i ∈ I.
Proof. Assume that the family (−a−→

i aj )j∈(I−{i}) is linearly independent for some specific i ∈ I.
Let k ∈ I with k 6= i, and assume that there are some scalars (λj )j∈(I−{k}) such that
λj −
a−→
X
k aj = 0.
j∈(I−{k})
Since
−
a−→ −−→ −−→
k aj = ak ai + ai aj ,
we have
λj −
a−→ λj −
a−→ λj −
a−→
X X X
k aj = k ai + i aj ,
j∈(I−{k}) j∈(I−{k}) j∈(I−{k})
λj −
a−→ λj −
a−→
X X
= k ai + i aj ,
j∈(I−{k}) j∈(I−{i,k})
X
λj −
a−→ λj −
a−→
X
= i aj − i ak ,
j∈(I−{i,k}) j∈(I−{k})
and thus X
λj −
a−→ λj −
a−→
X
i aj − i ak = 0.
j∈(I−{i,k}) j∈(I−{k})
Since the familyP(−

a−→
i aj )j∈(I−{i}) is linearly independent, we must have λj = 0 for all j ∈
(I − {i, k}) and j∈(I−{k}) λj = 0, which implies that λj = 0 for all j ∈ (I − {k}).
We define affine independence as follows.

→ −
Definition 2.4. Given an affine space E, E , + , a family (ai )i∈I of points in E is affinely
independent if the family (−
a−→
i aj )j∈(I−{i}) is linearly independent for some i ∈ I.
−
→
E E
a2
−
a−→
0 a2
a0 a1 −
a−→
0 a1
Figure 2.14: Affine independence and linear independence.
Definition 2.4 is reasonable, because by Proposition 2.4, the independence of the family
−−→
(ai aj )j∈(I−{i}) does not depend on the choice of ai . A crucial property of linearly independent
vectors (u1 , . . . , um ) is that if a vector v is a linear combination
m
X
v= λi ui
i=1
of the ui , then the λi are unique. A similar result holds for affinely independent points.

→ −
Proposition 2.5. Given an affine space E, E , + P , let (a0 , . . . , am ) P
be a family of m + 1
m m
points in E. Let x ∈ E, and assume that x = λ a
i=0 i i , where i=0 λi = 1. Then,
the family (λ0 , . . . , λm ) such that x = i=0 λi ai is unique iff the family (− a−→ −−→
Pm
0 a1 , . . . , a0 am ) is
linearly independent.
Proof. The proof is straightforward and is omitted. It is also given in Gallier [29].
Proposition 2.5 suggests the notion of affine frame. Affine frames are the affine analogues

→−
of bases in vector spaces. Let E, E , + be a nonempty affine space, and let (a0 , . . . , am )
be a family of m + 1 points in E. The family (a0 , . . . , am ) determines the family of m
→
−
vectors (− a−→ −−→
0 a1 , . . . , a0 am ) in E . Conversely, given a point a0 in E and a family of m vectors
→
−
(u1 , . . . , um ) in E , we obtain the family of m + 1 points (a0 , . . . , am ) in E, where ai = a0 + ui ,
1 ≤ i ≤ m.
Thus, for any m ≥ 1, it is equivalent to consider a family of m + 1 points (a0 , . . . , am ) in
→
−
E, and a pair (a0 , (u1 , . . . , um )), where the ui are vectors in E . Figure 2.14 illustrates the
notion of affine independence.
Remark: The above observation also applies to infinite families (ai )i∈I of points in E and
→
−
families (→
−
ui )i∈I−{0} of vectors in E , provided that the index set I contains 0.
→
−
When (−a− → −−→ −→
0 a1 , . . . , a0 am ) is a basis of E then, for every x ∈ E, since x = a0 + a0 x, there
is a unique family (x1 , . . . , xm ) of scalars such that
x = a0 + x 1 −
a−→ −−→
0 a1 + · · · + xm a0 am .
The scalars (x1 , . . . , xm ) may be considered as coordinates with respect to

(a0 , (−
a−→ −−→
0 a1 , . . . , a0 am )). Since
m m
! m
−−
→
X X X
x=a +
0 x a a iff x = 1 −
i 0 i x a + xa,i 0 i i
i=1 i=1 i=1
x ∈ E can also be expressed uniquely as

m
X
x= λi ai
i=0
Pm Pm
with i=0 λi = 1, and where λ0 = 1 − i=1 xi , and λi = xi for 1 ≤ i ≤ m. The scalars
(λ0 , . . . , λm ) are also certain kinds of coordinates with respect to (a0 , . . . , am ). All this is
summarized in the following definition.

→ −
Definition 2.5. Given an affine space E, E , + , an affine frame with origin a0 is a family
(a0 , . . . , am ) of m + 1 points in E such that the list of vectors (− a−→ −−→
0 a1 , . . . , a0 am ) is a basis of
→
−
E . The pair (a0 , (− a− → −−→
0 a1 , . . . , a0 am )) is also called an affine frame with origin a0 . Then, every
x ∈ E can be expressed as
x = a0 + x 1 −
a−→ −−→
0 a1 + · · · + xm a0 am
for a unique family (x1 , . . . , xm ) of scalars, called the coordinates of x w.r.t. the affine frame
(a0 , (−
a−→ −−→
0 a1 , . . ., a0 am )). Furthermore, every x ∈ E can be written as
x = λ0 a0 + · · · + λm am
for some unique family (λ0 , . . . , λm ) of scalars such that λ0 +· · ·+λm = 1 called the barycentric
coordinates of x with respect to the affine frame (a0 , . . . , am ). See Figure 2.15.
The coordinates (x1 , .P

. . , xm ) and the barycentric coordinates (λ0 , . . ., λm ) are related by
the equations λ0 = 1 − m i=1 xi and λi = xi , for 1 ≤ i ≤ m. An affine frame is called an
affine basis by some authors. A family (ai )i∈I of points in E is affinely dependent if it is not
affinely independent. We can also characterize affinely dependent families as follows.
x = (-1, 0,2)
a 3 = (1,3,2)
a2 = (-1,3,1)
a0 = (1,2,1)
a1 = (2,3,1)
O
x = (-1, 0,2)
a 3 = (1,3,2)
a2 = (-1,3,1)
a0
a1 = (2,3,1)
O
Figure 2.15: The affine frame (a0 , a1 , a2 , a3 ) for A3 . The coordinates for x = (−1, 0, 2)
are x1 = −8/3, x2 = −1/3, x3 = 1, while the barycentric coordinates for x are λ0 = 3,
λ1 = −8/3, λ2 = −1/3, λ3 = 1.

→ −
Proposition 2.6. Given an affine space E, E , + , let (ai )i∈I be a family of points in E.
P (ai )i∈I is affinely
The family P dependent iff there is a family (λi )i∈I such that λj 6= 0 for some
−→
j ∈ I, i∈I λi = 0, and i∈I λi xai = 0 for every x ∈ E.
Proof. By Proposition 2.5, the family (ai )i∈I is affinely dependent iff the family of vectors
(−
a− → −−→
i aj )j∈(I−{i}) is linearly dependent for some i ∈ I. For any i ∈ I, the family (ai aj )j∈(I−{i})
is linearly dependent iff there is a family (λj )j∈(I−{i}) such that λj 6= 0 for some j, and such
that
λj −
a−→
X
i aj = 0.
j∈(I−{i})
Then, for any x ∈ E, we have

λj −
a−→ λj (−→−− →)
X X
i aj = xa j xa i
j∈(I−{i}) j∈(I−{i})
X
λj −→− λj −→,
X
= xa j xa i
j∈(I−{i}) j∈(I−{i})
−→
P P P
and letting λi = − j∈(I−{i}) λj , we get i∈I λi xai = 0, with i∈I λi = 0 and λj 6= 0 for
some j ∈ I. The converse is obvious by setting x = ai for some i such that λi 6= 0, since
P
i∈I λi = 0 implies that λj 6= 0, for some j 6= i.
a2
a0 a0 a1
a3
a0 a1 a0 a2
a1
Figure 2.16: Examples of affine frames and their convex hulls.
Even though Proposition 2.6 is rather dull, it is one of the key ingredients in the proof
of beautiful and deep theorems about convex sets, such as Carathéodory’s theorem, Radon’s
theorem, and Helly’s theorem.
→
−
A family of two points (a, b) in E is affinely independent iff ab 6= 0, iff a 6= b. If a 6= b, the
affine subspace generated by a and b is the set of all points (1 − λ)a + λb, which is the unique
line passing through a and b. A family of three points (a, b, c) in E is affinely independent
→
−
iff ab and →
−
ac are linearly independent, which means that a, b, and c are not on the same line
(they are not collinear). In this case, the affine subspace generated by (a, b, c) is the set of all
points (1 − λ − µ)a + λb + µc, which is the unique plane containing a, b, and c. A family of
→
− − −
→
four points (a, b, c, d) in E is affinely independent iff ab, →
ac, and ad are linearly independent,
which means that a, b, c, and d are not in the same plane (they are not coplanar). In this
case, a, b, c, and d are the vertices of a tetrahedron. Figure 2.16 shows affine frames and
their convex hulls for |I| = 0, 1, 2, 3.
Given n+1 affinely independent points (a0 , . . . , an ) in E, we can consider the set of points
λ0 a0 + · · · + λn an , where λ0 + · · · + λn = 1 and λi ≥ 0 (λi ∈ R). Such affine combinations are
called convex combinations. This set is called the convex hull of (a0 , . . . , an ) (or n-simplex
spanned by (a0 , . . . , an )). When n = 1, we get the segment between a0 and a1 , including
a0 and a1 . When n = 2, we get the interior of the triangle whose vertices are a0 , a1 , a2 ,
including boundary points (the edges). When n = 3, we get the interior of the tetrahedron
whose vertices are a0 , a1 , a2 , a3 , including boundary points (faces and edges). The set
{a0 + λ1 −
a−→ −−→
0 a1 + · · · + λn a0 an | where 0 ≤ λi ≤ 1 (λi ∈ R)}
is called the parallelotope spanned by (a0 , . . . , an ). When E has dimension 2, a parallelotope

is also called a parallelogram, and when E has dimension 3, a parallelepiped . Figure 2.17
shows the convex hulls and associated parallelotopes for |I| = 0, 1, 2, 3.
a0
a0 a
1
a
2
a0 a
1
a
3
a0
a
3
a
1
a
2
a0
a
1
Figure 2.17: Examples of affine frames, convex hulls, and their associated parallelotopes.
More generally, we say that a subset V of E is convex if for any two points a, b ∈ V , we
have c ∈ V for every point c = (1 − λ)a + λb, with 0 ≤ λ ≤ 1 (λ ∈ R).
Points are not vectors! The following example illustrates why treating points as
vectors may cause problems. Let a, b, c be three affinely independent points in A3 .
Any point x in the plane (a, b, c) can be expressed as
x = λ0 a + λ1 b + λ2 c,
where λ0 + λ1 + λ2 = 1. How can we compute λ0 , λ1 , λ2 ? Letting a = (a1 , a2 , a3 ), b =

(b1 , b2 , b3 ), c = (c1 , c2 , c3 ), and x = (x1 , x2 , x3 ) be the coordinates of a, b, c, x in the standard
frame of A3 , it is tempting to solve the system of equations
2.7. AFFINE MAPS 39
    
a1 b 1 c 1 λ0 x1
a2 b2 c2  λ1  = x2  .
a3 b 3 c 3 λ2 x3
However, there is a problem when the origin of the coordinate system belongs to the plane
(a, b, c), since in this case, the matrix is not invertible! What we should really be doing is to
solve the system
−→ −
→ −
→ −→
λ0 Oa + λ1 Ob + λ2 Oc = Ox,
where O is any point not in the plane (a, b, c). An alternative is to use certain well-chosen
cross products.
It can be shown that barycentric coordinates correspond to various ratios of areas and
volumes; see the problems.
2.7 Affine Maps

Corresponding to linear maps we have the notion of an affine map. An affine map is defined
as a map preserving affine combinations.

→ −
− →
Definition 2.6. Given two affine spaces E, E , + and E 0 , E 0 , +0 , a function fP : E → E 0 is
an affine map iff for every family ((ai , λi ))i∈I of weighted points in E such that i∈I λi = 1,
we have !
X X
f λi ai = λi f (ai ).
i∈I i∈I
In other words, f preserves barycenters.
Affine maps can be obtained from linear maps as follows. For simplicity of notation, the
same symbol + is used for both affine spaces (instead of using both + and +0 ).
→
− −
→
Given any point a ∈ E, any point b ∈ E 0 , and any linear map h : E → E 0 , we claim that
the map f : E → E 0 defined such that
f (a + v) = b + h(v)
P
is an affine map. Indeed, for any family (λi )i∈I of scalars with i∈I λi = 1 and any family
(→
−vi )i∈I , since
X X −−−−−→ X
λi (a + vi ) = a + λi a(a + vi ) = a + λi vi
i∈I i∈I i∈I
and X X −−−−−−−→ X
λi (b + h(vi )) = b + λi b(b + h(vi )) = b + λi h(vi ),
i∈I i∈I i∈I
we have
! !
X X
f λi (a + vi ) = f a+ λi vi
i∈I i∈I
!
X
= b+h λi vi
i∈I
X
= b+ λi h(vi )
i∈I
X
= λi (b + h(vi ))
i∈I
X
= λi f (a + vi ).
i∈I
P
Note that the condition i∈I λi = 1 was implicitly used (in a hidden call to Proposition
2.1) in deriving that X X
λi (a + vi ) = a + λi vi
i∈I i∈I
and X X
λi (b + h(vi )) = b + λi h(vi ).
i∈I i∈I
As a more concrete example, the map

x1 1 2 x1 3
7→ +
x2 0 1 x2 1
defines an affine map in A2 . It is a “shear” followed by a translation. The effect of this shear
on the square (a, b, c, d) is shown in Figure 2.18. The image of the square (a, b, c, d) is the
parallelogram (a0 , b0 , c0 , d0 ).
d = (5,2) c = (6,2)
d = (0,1) c = (1,1)
a = (3,1) b = (4,1)
a = (0,0) b = (1,0)
Figure 2.18: The effect of a shear.

2.7. AFFINE MAPS 41
c = (5,4)
d‘ = (4,3)
d = (0,1) c = (1,1)
b = (4,1)
a = (0,0) b = (1,0) a = (3,0)
Figure 2.19: The effect of an affine map.
Let us consider one more example. The map

x1 1 1 x1 3
7→ +
x2 1 3 x2 0
is an affine map. Since we can write
√
√
√
1 1 2/2 −√ 2/2 1 2
= 2 ,
1 3 2/2 2/2 0 1
this affine map is the composition

√ of a shear, followed by a rotation of angle π/4, followed by
a magnification of ratio 2, followed by a translation. The effect of this map on the square
(a, b, c, d) is shown in Figure 2.19. The image of the square (a, b, c, d) is the parallelogram
(a0 , b0 , c0 , d0 ).
The following proposition shows the converse of what we just showed. Every affine map
is determined by the image of any point and a linear map.
→
− → − −
→
Proposition 2.7. Given an affine map f : E → E 0 , there is a unique linear map f : E → E 0
such that
→
−
f (a + v) = f (a) + f (v),
→
−
for every a ∈ E and every v ∈ E .
Proof. Let a ∈ E be any point in E. We claim that the map defined such that
→
− −−−−−−−−−→
f (v) = f (a)f (a + v)
→
− →
− → − −
→
for every v ∈ E is a linear map f : E → E 0 . Indeed, we can write
a + λv = λ(a + v) + (1 − λ)a,
−−−−−→
since a + λv = a + λa(a + v) + (1 − λ)−
→ and also
aa,
a + u + v = (a + u) + (a + v) − a,
−−−−−→ −−−−−→ →
since a + u + v = a + a(a + u) + a(a + v) − −
aa. Since f preserves barycenters, we get
f (a + λv) = λf (a + v) + (1 − λ)f (a).

P
If we recall
P that x = i∈I λi ai is the barycenter of a family ((ai , λi ))i∈I of weighted points
(with i∈I λi = 1) iff
→
− X − →
bx = λi bai for every b ∈ E,
i∈I
we get
−−−−−−−−−−→ −−−−−−−−−→ −−−−−→ −−−−−−−−−→
f (a)f (a + λv) = λf (a)f (a + v) + (1 − λ)f (a)f (a) = λf (a)f (a + v),
→
− →
−
showing that f (λv) = λ f (v). We also have
f (a + u + v) = f (a + u) + f (a + v) − f (a),
from which we get

−−−−−−−−−−−−→ −−−−−−−−−→ −−−−−−−−−→
f (a)f (a + u + v) = f (a)f (a + u) + f (a)f (a + v),
→
− →
− →
− →
−
showing that f (u + v) = f (u) + f (v). Consequently, f is a linear map. For any other
point b ∈ E, since
→
− −−−−−→ → → −
b + v = a + ab + v = a + a(a + v) − −
aa + ab,
b + v = (a + v) − a + b, and since f preserves barycenters, we get
f (b + v) = f (a + v) − f (a) + f (b),
which implies that

−−−−−−−−→ −−−−−−−−→ −−−−−→ −−−−−→
f (b)f (b + v) = f (b)f (a + v) − f (b)f (a) + f (b)f (b),
−−−−−→ −−−−−−−−→
= f (a)f (b) + f (b)f (a + v),
−−−−−−−−−→
= f (a)f (a + v).
−−−−−−−−→ −−−−−−−−−→ →
−
Thus, f (b)f (b + v) = f (a)f (a + v), which shows that the definition of f does not depend
→
− →
−
on the choice of a ∈ E. The fact that f is unique is obvious: We must have f (v) =
−−−−−−−−−→
f (a)f (a + v).
2.7. AFFINE MAPS 43
→
− → − −
→
The unique linear map f : E → E 0 given by Proposition 2.7 is called the linear map
associated with the affine map f .
Note that the condition
→
−
f (a + v) = f (a) + f (v),
→
−
for every a ∈ E and every v ∈ E , can be stated equivalently as
→
− → −−−−−→ → − →
f (x) = f (a) + f (−
ax), or f (a)f (x) = f (−
ax),
for all a, x ∈ E. Proposition 2.7 shows that for any affine map f : E → E 0 , there are points
→
− → − −
→
a ∈ E, b ∈ E 0 , and a unique linear map f : E → E 0 , such that
→
−
f (a + v) = b + f (v),
→
− →
−
for all v ∈ E (just let b = f (a), for any a ∈ E). Affine maps for which f is the identity
→
−
map are called translations. Indeed, if f = id,
→
− →
f (x) = f (a) + f (−
ax) = f (a) + −
→=x+−
ax →+−
xa
−−→ →
af (a) + −
ax
−−−→ −−−→
= x+− → + af (a) − −
xa → = x + af (a),
xa
and so
−−−→ −−−→
xf (x) = af (a),
−−−→
which shows that f is the translation induced by the vector af (a) (which does not depend
on a).
Since an affine map preserves barycenters, and since an affine subspace V is closed under
barycentric combinations, the image f (V ) of V is an affine subspace in E 0 . So, for example,
the image of a line is a point or a line, and the image of a plane is either a point, a line, or
a plane.
It is easily verified that the composition of two affine maps is an affine map. Also, given
affine maps f : E → E 0 and g : E 0 → E 00 , we have
→
− →
−
→−
g(f (a + v)) = g f (a) + f (v) = g(f (a)) + g f (v) ,
−−→ − → −
which shows that g ◦ f = → g ◦ f . It is easy to show that an affine map f : E → E 0 is injective
→
− → − −
→ →
− → − −
→
iff f : E → E 0 is injective, and that f : E → E 0 is surjective iff f : E → E 0 is surjective.
→
− → − −
→
An affine map f : E → E 0 is constant iff f : E → E 0 is the null (constant) linear map equal
→
−
to 0 for all v ∈ E .
If E is an affine space of dimension m and (a0 , a1 , . . . , am ) is an affine frame for E, then
for any other affine space F and for any sequence (b0 , b1 , . . . , bm ) of m + 1 points in F , there
is a unique affine map f : E → F such that f (ai ) = bi , for 0 ≤ i ≤ m. Indeed, f must be

such that
f (λ0 a0 + · · · + λm am ) = λ0 b0 + · · · + λm bm ,
where λ0 +· · ·+λm = 1, and this defines a unique affine map on all of E, since (a0 , a1 , . . . , am )
is an affine frame for E.
Using affine frames, affine maps can be represented in terms of matrices. We explain how
an affine map f : E → E is represented with respect to a frame (a0 , . . . , an ) in E, the more
general case where an affine map f : E → F is represented with respect to two affine frames
(a0 , . . . , an ) in E and (b0 , . . . , bm ) in F being analogous. Since
→
−
f (a0 + x) = f (a0 ) + f (x)
→
−
for all x ∈ E , we have
−−−−−−−−→ −−−−→ → −
a0 f (a0 + x) = a0 f (a0 ) + f (x).
−−−−→ −−−−−−−−→
Since x, a0 f (a0 ), and a0 f (a0 + x), can be expressed as
x = x1 −a− → −−→
0 a1 + · · · + x n a0 an ,
−−−−→
a0 f (a0 ) = b1 −
a−→ −−→
0 a1 + · · · + b n a0 an ,
−−−−−−−−→
a0 f (a0 + x) = y1 − a−→ −−→
0 a1 + · · · + y n a0 an ,
→
−
if A = (ai j ) is the n × n matrix of the linear map f over the basis (− a− → −−→
0 a1 , . . . , a0 an ), letting x,
y, and b denote the column vectors of components (x1 , . . . , xn ), (y1 , . . . , yn ), and (b1 , . . . , bn ),
−−−−−−−−→ −−−−→ → −
a0 f (a0 + x) = a0 f (a0 ) + f (x)
is equivalent to
y = Ax + b.
Note that b 6= 0 unless f (a0 ) = a0 . Thus, f is generally not a linear transformation, unless it
has a fixed point, i.e., there is a point a0 such that f (a0 ) = a0 . The vector b is the “translation
part” of the affine map. Affine maps do not always have a fixed point. Obviously, nonnull
translations have no fixed point. A less trivial example is given by the affine map

x1 1 0 x1 1
7→ + .
x2 0 −1 x2 0
This map is a reflection about the x-axis followed by a translation along the x-axis. The
affine map √
x1 1 − 3 x1 1
7→ √ +
x2 3/4 1/4 x2 1
2.7. AFFINE MAPS 45
can also be written as

√
x1 2 0 √1/2 − 3/2 x1 1
7→ +
x2 0 1/2 3/2 1/2 x 2 1
which shows that it is the composition of a rotation of angle π/3, followed by a stretch (by a
factor of 2 along the x-axis, and by a factor of 12 along the y-axis), followed by a translation.
It is easy to show that this affine map has a unique fixed point. On the other hand, the
affine map
x1 8/5 −6/5 x1 1
7→ +
x2 3/10 2/5 x2 1
has no fixed point, even though

8/5 −6/5 2 0 4/5 −3/5
= ,
3/10 2/5 0 1/2 3/5 4/5
4
and the second matrix is a rotation of angle θ such that cos θ = 5
and sin θ = 35 .
There is a useful trick to convert the equation y = Ax + b into what looks like a linear
equation. The trick is to consider an (n + 1) × (n + 1) matrix. We add 1 as the (n + 1)th
component to the vectors x, y, and b, and form the (n + 1) × (n + 1) matrix

A b
0 1
so that y = Ax + b is equivalent to

y A b x
= .
1 0 1 1
This trick is very useful in kinematics and dynamics, where A is a rotation matrix. Such
affine maps are called rigid motions.
If f : E → E 0 is a bijective affine map, given any three collinear points a, b, c in E,
with a 6= b, where, say, c = (1 − λ)a + λb, since f preserves barycenters, we have f (c) =
(1 − λ)f (a) + λf (b), which shows that f (a), f (b), f (c) are collinear in E 0 . There is a converse
to this property, which is simpler to state when the ground field is K = R. The converse
states that given any bijective function f : E → E 0 between two real affine spaces of the
same dimension n ≥ 2, if f maps any three collinear points to collinear points, then f is
affine. The proof is rather long (see Berger [7] or Samuel [50]).
Given three collinear points a, b, c, where a 6= c, we have b = (1 − β)a + βc for some
unique β, and we define the ratio of the sequence a, b, c, as
→
−
β ab
ratio(a, b, c) = =→ − ,
(1 − β) bc
provided that β 6= 1, i.e., b 6= c. When b = c, we agree that ratio(a, b, c) = ∞. We warn

−
→
ba
our readers that other authors define the ratio of a, b, c as −ratio(a, b, c) = −
→ . Since affine
bc
maps preserve barycenters, it is clear that affine maps preserve the ratio of three points.
2.8 Affine Groups

We now take a quick look at the bijective affine maps. Given an affine space E, the set of
affine bijections f : E → E is clearly a group, called the affine group of E, and denoted by
→
−
GA(E). Recall that the group of bijective linear maps of the vector space E is denoted by
→
− →
− →
−
GL( E ). Then, the map f 7→ f defines a group homomorphism L : GA(E) → GL( E ).
The kernel of this map is the set of translations on E.
→
The subset of all linear maps of the form λ id−E
, where λ ∈ R − {0}, is a subgroup of
→
− ∗
GL( E ), and is denoted by R id−→
E
→
(where λ id−
E
(u) = λu, and R∗ = R−{0}). The subgroup
DIL(E) = L−1 (R∗ id− →
E
) of GA(E) is particularly interesting. It turns out that it is the
disjoint union of the translations and of the dilatations of ratio λ 6= 1. The elements of
DIL(E) are called affine dilatations.
Given any point a ∈ E, and any scalar λ ∈ R, a dilatation or central dilatation (or
homothety) of center a and ratio λ is a map Ha,λ defined such that
Ha,λ (x) = a + λ−
→
ax,
for every x ∈ E.
Remark: The terminology does not seem to be universally agreed upon. The terms affine
dilatation and central dilatation are used by Pedoe [46]. Snapper and Troyer use the term
dilation for an affine dilatation and magnification for a central dilatation [53]. Samuel uses
homothety for a central dilatation, a direct translation of the French “homothétie” [50]. Since
dilation is shorter than dilatation and somewhat easier to pronounce, perhaps we should use
that!
Observe that Ha,λ (a) = a, and when λ 6= 0 and x 6= a, Ha,λ (x) is on the line defined by
a and x, and is obtained by “scaling” −
→ by λ.
ax
Figure 2.20 shows the effect of a central dilatation of center d. The triangle (a, b, c) is
magnified to the triangle (a0 , b0 , c0 ). Note how every line is mapped to a parallel line.
−−→
When λ = 1, Ha,1 is the identity. Note that Ha,λ = λ id− →E
. When λ 6= 0, it is clear that
Ha,λ is an affine bijection. It is immediately verified that
Ha,λ ◦ Ha,µ = Ha,λµ .
We have the following useful result.

2.8. AFFINE GROUPS 47
b b
d
c
Figure 2.20: The effect of a central dilatation Hd,λ (x).
→
−
Proposition 2.8. Given any affine space E, for any affine bijection f ∈ GA(E), if f =
→
λ id−
E
, for some λ ∈ R∗ with λ 6= 1, then there is a unique point c ∈ E such that f = Hc,λ .
→
−
Clearly, if f = id− →E
, the affine map f is a translation. Thus, the group of affine
dilatations DIL(E) is the disjoint union of the translations and of the dilatations of ratio
λ 6= 0, 1. Affine dilatations can be given a purely geometric characterization.
Another point worth mentioning is that affine bijections preserve the ratio of volumes of
→
−
parallelotopes. Indeed, given any basis B = (u1 , . . . , um ) of the vector space E associated
with the affine space E, given any m + 1 affinely independent points (a0 , . . . , am ), we can
compute the determinant detB (− a− → −−→
0 a1 , . . . , a0 am ) w.r.t. the basis B. For any bijective affine
map f : E → E, since
→
− −−→ →
− −−→ →
−
detB f (a0 a1 ), . . . , f (a0 am ) = det f detB (− a− → −−→
0 a1 , . . . , a0 am )
→
−
and the determinant of a linear map is intrinsic (i.e., depends only on f , and not on the
particular basis B), we conclude that the ratio
→
− −→ →
− −→
detB f (−a0 a1 ), . . . , f (−
a0 am ) →
−
−− → − −→ = det f
det (a a , . . . , a a )
B 0 1 0 m
is independent of the basis B. Since detB (− a−→ −−→

0 a1 , . . . , a0 am ) is the volume of the parallelotope
spanned by (a0 , . . . , am ), where the parallelotope spanned by any point a and the vectors
b1
H1 a1
H2 b2
a
2
H3 b3
a3
A B
Figure 2.21: The theorem of Thales.
(u1 , . . . , um ) has unit volume (see Berger [7], Section 9.12), we see that affine bijections
preserve the ratio of volumes of parallelotopes. In fact, this ratio is independent of the
choice of the parallelotopes of unit volume. In particular, the affine bijections f ∈ GA(E)
→
−
such that det f = 1 preserve volumes. These affine maps form a subgroup SA(E) of
GA(E) called the special affine group of E. We now take a glimpse at affine geometry.
2.9 Affine Geometry: A Glimpse

In this section we state and prove three fundamental results of affine geometry. Roughly
speaking, affine geometry is the study of properties invariant under affine bijections. We now
prove one of the oldest and most basic results of affine geometry, the theorem of Thales.
Proposition 2.9. Given any affine space E, if H1 , H2 , H3 are any three distinct parallel
hyperplanes, and A and B are any two lines not parallel to Hi , letting ai = Hi ∩ A and
bi = Hi ∩ B, then the following ratios are equal:
− −−→
a−→
1 a3 b1 b3
−−→ = −−→ = ρ.
a1 a2 b1 b2
−→
a1 d
Conversely, for any point d on the line A, if −−a
a →
1 2
= ρ, then d = a3 .
Proof. Figure 2.21 illustrates the theorem of Thales. We sketch a proof, leaving the details
→
−
as an exercise. Since H1 , H2 , H3 are parallel, they have the same direction H , a hyperplane
2.9. AFFINE GEOMETRY: A GLIMPSE 49
→
− →
− → −
in E . Let u ∈ E − H be any nonnull vector such that A = a1 +Ru. Since A is not parallel to
→
− →
− →
−
H, we have E = H ⊕ Ru, and thus we can define the linear map p : E → Ru, the projection
→
−
on Ru parallel to H . This linear map induces an affine map f : E → A, by defining f such
that
f (b1 + w) = a1 + p(w),
→
− →
−
for all w ∈ E . Clearly, f (b1 ) = a1 , and since H1 , H2 , H3 all have direction H , we also have
f (b2 ) = a2 and f (b3 ) = a3 . Since f is affine, it preserves ratios, and thus
− −−→
a−→
1 a3 b1 b3
−−→ = −−→ .
a1 a2 b1 b2
The converse is immediate.
We also have the following simple proposition, whose proof is left as an easy exercise.
Proposition 2.10. Given any affine space E, given any two distinct points a, b ∈ E, and
for any affine dilatation f different from the identity, if a0 = f (a), D = ha, bi is the line
passing through a and b, and D0 is the line parallel to D and passing through a0 , the following
are equivalent:
(i) b0 = f (b);
(ii) If f is a translation, then b0 is the intersection of D0 with the line parallel to ha, a0 i
passing through b;
If f is a dilatation of center c, then b0 = D0 ∩ hc, bi.
The first case is the parallelogram law, and the second case follows easily from Thales’
theorem. For an illustration, see Figure 2.22.
We are now ready to prove two classical results of affine geometry, Pappus’s theorem and
Desargues’s theorem. Actually, these results are theorems of projective geometry, and we
are stating affine versions of these important results. There are stronger versions that are
best proved using projective geometry.
Proposition 2.11. Given any affine plane E, any two distinct lines D and D0 , then for
any distinct points a, b, c on D and a0 , b0 , c0 on D0 , if a, b, c, a0 , b0 , c0 are distinct from the
intersection of D and D0 (if D and D0 intersect) and if the lines ha, b0 i and ha0 , bi are parallel,
and the lines hb, c0 i and hb0 , ci are parallel, then the lines ha, c0 i and ha0 , ci are parallel.
a’ = f(a)
b
D’
D
a’ = f(a) a’ = f(a)
b’ = f(b)
a a
b’ = f(b)
b b
D’ c D’
D D
(i) (ii)
Figure 2.22: An illustration of Proposition 2.10. The bottom left diagram illustrates a
translation, while the bottom right illustrates a central dilation through c.
Proof. Pappus’s theorem is illustrated in Figure 2.23. If D and D0 are not parallel, let d be
their intersection. Let f be the dilatation of center d such that f (a) = b, and let g be the
dilatation of center d such that g(b) = c. Since the lines ha, b0 i and ha0 , bi are parallel, and
the lines hb, c0 i and hb0 , ci are parallel, by Proposition 2.10 we have a0 = f (b0 ) and b0 = g(c0 ).
However, we observed that dilatations with the same center commute, and thus f ◦ g = g ◦ f ,
and thus, letting h = g ◦ f , we get c = h(a) and a0 = h(c0 ). Again, by Proposition 2.10, the
lines ha, c0 i and ha0 , ci are parallel. If D and D0 are parallel, we use translations instead of
dilatations.
There is a converse to Pappus’s theorem, which yields a fancier version of Pappus’s

theorem, but it is easier to prove it using projective geometry. It should be noted that
in axiomatic presentations of projective geometry, Pappus’s theorem is equivalent to the
commutativity of the ground field K (in the present case, K = R). We now prove an affine
version of Desargues’s theorem.
Proposition 2.12. Given any affine space E, and given any two triangles (a, b, c) and
(a0 , b0 , c0 ), where a, b, c, a0 , b0 , c0 are all distinct, if ha, bi and ha0 , b0 i are parallel and hb, ci and
hb0 , c0 i are parallel, then ha, ci and ha0 , c0 i are parallel iff the lines ha, a0 i, hb, b0 i, and hc, c0 i
are either parallel or concurrent (i.e., intersect in a common point).
Proof. We prove half of the proposition, the direction in which it is assumed that ha, ci and
ha0 , c0 i are parallel, leaving the converse as an exercise. Since the lines ha, bi and ha0 , b0 i are
2.9. AFFINE GEOMETRY: A GLIMPSE 51
c
D
b
a
c
b
D
a
Figure 2.23: Pappus’s theorem (affine version).
parallel, the points a, b, a0 , b0 are coplanar. Thus, either ha, a0 i and hb, b0 i are parallel, or
they have some intersection d. We consider the second case where they intersect, leaving
the other case as an easy exercise. Let f be the dilatation of center d such that f (a) = a0 .
By Proposition 2.10, we get f (b) = b0 . If f (c) = c00 , again by Proposition 2.10 twice, the
lines hb, ci and hb0 , c00 i are parallel, and the lines ha, ci and ha0 , c00 i are parallel. From this it
follows that c00 = c0 . Indeed, recall that hb, ci and hb0 , c0 i are parallel, and similarly ha, ci and
ha0 , c0 i are parallel. Thus, the lines hb0 , c00 i and hb0 , c0 i are identical, and similarly the lines
−→ −→
ha0 , c00 i and ha0 , c0 i are identical. Since a0 c0 and b0 c0 are linearly independent, these lines have
a unique intersection, which must be c00 = c0 .
The direction where it is assumed that the lines ha, a0 i, hb, b0 i and hc, c0 i, are either parallel
or concurrent is left as an exercise (in fact, the proof is quite similar).
Desargues’s theorem is illustrated in Figure 2.24.

There is a fancier version of Desargues’s theorem, but it is easier to prove it using pro-
jective geometry. It should be noted that in axiomatic presentations of projective geometry,
Desargues’s theorem is related to the associativity of the ground field K (in the present
case, K = R). Also, Desargues’s theorem yields a geometric characterization of the affine
dilatations. An affine dilatation f on an affine space E is a bijection that maps every line
D to a line f (D) parallel to D. We leave the proof as an exercise.
b b
d
c
Figure 2.24: Desargues’s theorem (affine version).
2.10 Affine Hyperplanes

We now consider affine forms and affine hyperplanes. In Section 2.5 we observed that the
set L of solutions of an equation
ax + by = c
is an affine subspace of A2 of dimension 1, in fact, a line (provided that a and b are not both
null). It would be equally easy to show that the set P of solutions of an equation
ax + by + cz = d
is an affine subspace of A3 of dimension 2, in fact, a plane (provided that a, b, c are not all
null). More generally, the set H of solutions of an equation
λ1 x1 + · · · + λm xm = µ
is an affine subspace of Am , and if λ1 , . . . , λm are not all null, it turns out that it is a subspace
of dimension m − 1 called a hyperplane.
We can interpret the equation
λ1 x1 + · · · + λm xm = µ
in terms of the map f : Rm → R defined such that
f (x1 , . . . , xm ) = λ1 x1 + · · · + λm xm − µ
for all (x1 , . . . , xm ) ∈ Rm . It is immediately verified that this map is affine, and the set H of
solutions of the equation
λ1 x1 + · · · + λm xm = µ
2.10. AFFINE HYPERPLANES 53
is the null set, or kernel, of the affine map f : Am → R, in the sense that
H = f −1 (0) = {x ∈ Am | f (x) = 0},
where x = (x1 , . . . , xm ).
Thus, it is interesting to consider affine forms, which are just affine maps f : E → R
from an affine space to R. Unlike linear forms f ∗ , for which Ker f ∗ is never empty (since it
always contains the vector 0), it is possible that f −1 (0) = ∅ for an affine form f . Given an
affine map f : E → R, we also denote f −1 (0) by Ker f , and we call it the kernel of f . Recall
that an (affine) hyperplane is an affine subspace of codimension 1. The relationship between
affine hyperplanes and affine forms is given by the following proposition.
Proposition 2.13. Let E be an affine space. The following properties hold:
(a) Given any nonconstant affine form f : E → R, its kernel H = Ker f is a hyperplane.
(b) For any hyperplane H in E, there is a nonconstant affine form f : E → R such that
H = Ker f . For any other affine form g : E → R such that H = Ker g, there is some
λ ∈ R such that g = λf (with λ 6= 0).
(c) Given any hyperplane H in E and any (nonconstant) affine form f : E → R such that
H = Ker f , every hyperplane H 0 parallel to H is defined by a nonconstant affine form
g such that g(a) = f (a) − λ, for all a ∈ E and some λ ∈ R.
When E is of dimension n, given an affine frame (a0 , (u1 , . . . , un )) of E with origin a0 ,

recall from Definition 2.5 that every point of E can be expressed uniquely as x = a0 +
x1 u1 + · · · + xn un , where (x1 , . . . , xn ) are the coordinates of x with respect to the affine frame
(a0 , (u1 , . . . , un )).
Also recall that every linear form f ∗ is such that f ∗ (x) = λ1 x1 + · · · + λn xn , for every
x = x1 u1 + · · · + xn un and some λ1 , . . . , λn ∈ R. Since an affine form f : E → R satisfies the
→
−
property f (a0 + x) = f (a0 ) + f (x), denoting f (a0 + x) by f (x1 , . . . , xn ), we see that we have
f (x1 , . . . , xn ) = λ1 x1 + · · · + λn xn + µ,
where µ = f (a0 ) ∈ R and λ1 , . . . , λn ∈ R. Thus, a hyperplane is the set of points whose

coordinates (x1 , . . . , xn ) satisfy the (affine) equation
λ1 x1 + · · · + λn xn + µ = 0.
2.11 Intersection of Affine Spaces

In this section we take a closer look at the intersection of affine subspaces. This subsection
can be omitted at first reading.
First, we need a result of linear algebra. Given a vector space E and any two subspaces M
and N , there are several interesting linear maps. We have the canonical injections i : M →
M +N and j : N → M +N , the canonical injections in1 : M → M ⊕N and in2 : N → M ⊕N ,
and thus, injections f : M ∩N → M ⊕N and g : M ∩N → M ⊕N , where f is the composition
of the inclusion map from M ∩ N to M with in1 , and g is the composition of the inclusion
map from M ∩ N to N with in2 . Then, we have the maps f + g : M ∩ N → M ⊕ N , and
i − j : M ⊕ N → M + N.
Proposition 2.14. Given a vector space E and any two subspaces M and N , with the
definitions above,
f +g i−j
0 −→ M ∩ N −→ M ⊕ N −→ M + N −→ 0
is a short exact sequence, which means that f + g is injective, i − j is surjective, and that
Im (f + g) = Ker (i − j). As a consequence, we have the Grassmann relation
dim(M ) + dim(N ) = dim(M + N ) + dim (M ∩ N ).
Proof. It is obvious that i − j is surjective and that f + g is injective. Assume that (i −

j)(u + v) = 0, where u ∈ M , and v ∈ N . Then, i(u) = j(v), and thus, by definition of i and
j, there is some w ∈ M ∩ N , such that i(u) = j(v) = w ∈ M ∩ N . By definition of f and g,
u = f (w) and v = g(w), and thus Im (f + g) = Ker (i − j), as desired. The second part of
the proposition follows from standard results of linear algebra (see Artin [3], Strang [58], or
Lang [37]).
We now prove a simple proposition about the intersection of affine subspaces.
Proposition 2.15. Given any affine space E, for any two nonempty affine subspaces M
and N , the following facts hold:
→
− − → → −
(1) M ∩ N 6= ∅ iff ab ∈ M + N for some a ∈ M and some b ∈ N .
→
− −
→ → −
(2) M ∩ N consists of a single point iff ab ∈ M + N for some a ∈ M and some b ∈ N ,
−
→ → −
and M ∩ N = {0}.
→
− −
→ → − →
−
(3) If S is the least affine subspace containing M and N , then S = M + N + K ab (the
→
−
vector space E is defined over the field K).
2.11. INTERSECTION OF AFFINE SPACES 55
Proof. (1) Pick any a ∈ M and any b ∈ N , which is possible, since M and N are nonempty.
−
→ → | x ∈ M } and →− →
−
Since M = {− ax N = { by | y ∈ N }, if M ∩ N 6= ∅, for any c ∈ M ∩ N we have
→
− →
− −
→ →
− →
− →
− −
→ → −
ab = → −
ac − bc, with →
−
ac ∈ M and bc ∈ N , and thus, ab ∈ M + N . Conversely, assume that
→
− −
→ → − →
− →+→ −
ab ∈ M + N for some a ∈ M and some b ∈ N . Then ab = − ax by, for some x ∈ M and
some y ∈ N . But we also have
→
− →+→ −
ab = −→+−
ax xy yb,
and thus we get 0 = − →+→
xy
− → − → = 2→
yb − by, that is, −
xy
−
by. Thus, b is the middle of the segment
−
→ →
−
[x, y], and since yx = 2 yb, x = 2b − y is the barycenter of the weighted points (b, 2) and
(y, −1). Thus x also belongs to N , since N being an affine subspace, it is closed under
barycenters. Thus, x ∈ M ∩ N , and M ∩ N 6= ∅.
(2) Note that in general, if M ∩ N 6= ∅, then
−−−−→ − → → −
M ∩ N = M ∩ N,
because
−−−−→ →
− →
− →
− −
→ → −
M ∩ N = { ab | a, b ∈ M ∩ N } = { ab | a, b ∈ M } ∩ { ab | a, b ∈ N } = M ∩ N .
−−−−→
Since M ∩ N = c + M ∩ N for any c ∈ M ∩ N , we have
−
→ → −
M ∩ N = c + M ∩ N for any c ∈ M ∩ N .
−
→ → −
From this it follows that if M ∩N 6= ∅, then M ∩N consists of a single point iff M ∩ N = {0}.
This fact together with what we proved in (1) proves (2).
(3) This is left as an easy exercise.
Remarks:
→
− −
→ → −
(1) The proof of Proposition 2.15 shows that if M ∩ N 6= ∅, then ab ∈ M + N for all
a ∈ M and all b ∈ N .
(2) Proposition 2.15 implies that for any two nonempty affine subspaces M and N , if
→
− −
→ → − →
− −
→ → −
E = M ⊕ N , then M ∩ N consists of a single point. Indeed, if E = M ⊕ N , then
→
− → − −
→ → −
ab ∈ E for all a ∈ M and all b ∈ N , and since M ∩ N = {0}, the result follows from
part (2) of the proposition.
We can now state the following proposition.
Proposition 2.16. Given an affine space E and any two nonempty affine subspaces M and
N , if S is the least affine subspace containing M and N , then the following properties hold:
(1) If M ∩ N = ∅, then
−
→ → −
dim(M ) + dim(N ) < dim(E) + dim(M + N )
and
−
→ → −
dim(S) = dim(M ) + dim(N ) + 1 − dim(M ∩ N ).
(2) If M ∩ N 6= ∅, then
dim(S) = dim(M ) + dim(N ) − dim(M ∩ N ).
Proof. The proof is not difficult, using Proposition 2.15 and Proposition 2.14, but we leave
it as an exercise.
Chapter 3
Basic Properties of Convex Sets
Convex sets play a very important role in geometry. In this chapter we state and prove some
of the “classics” of convex affine geometry: Carathéodory’s theorem, Radon’s theorem, and
Helly’s theorem. These theorems share the property that they are easy to state, but they
are deep, and their proof, although rather short, requires a lot of creativity.
3.1 A Review of Basic Topological Concepts

→
−
Given an affine space E with associated vector space E , for any two points a, b ∈ E, the
→
−
unique vector from a to b is denoted ab rather than ab, so that b = a + ab.
The vector space Rd viewed as an affine space is denoted by Ad . In addition, if Rd is
equipped with the standard Euclidean inner product and Rd is viewed a an affine space, then
it is denoted by Ed .
Now, Ad is a topological space under the usual topology on Rd (in fact, Ad is a metric
space). Recall that if a = (a1 , . . . , ad ) and b = (b1 , . . . , bd ) are any two points in Ad , their
Euclidean distance d(a, b) is given by
p
d(a, b) = (b1 − a1 )2 + · · · + (bd − ad )2 ,
which is also the norm kabk of the vector ab, and that for any > 0, the open ball of center
a and radius , B(a, ), is given by
B(a, ) = {b ∈ Ad | d(a, b) < }.
A subset U ⊆ Ad is open (in the norm topology) if either U is empty or for every point
a ∈ U , there is some (small) open ball B(a, ) contained in U .
A subset C ⊆ Ad is closed iff Ad − C is open. For example, the closed balls B(a, ) where
B(a, ) = {b ∈ Ad | d(a, b) ≤ }
57
58 CHAPTER 3. BASIC PROPERTIES OF CONVEX SETS
are closed.
A subset W ⊆ Ad is bounded iff there is some ball (open or closed) B so that W ⊆ B.
A subset W ⊆SAd is compact iff every family {Ui }i∈I that is an open cover of W (which
means that W = i∈I (W ∩ Ui ), with each Ui an open set) Spossesses a finite subcover (which
means that there is a finite subset F ⊆ I so that W = i∈F (W ∩ Ui )). In Ad , it can be
shown that a subset W is compact iff W is closed and bounded.
Given a function f : Am → An , we say that f is continuous if f −1 (V ) is open in Am
whenever V is open in An . If f : Am → An is a continuous function, although it is generally
false that f (U ) is open if U ⊆ Am is open, it is easily checked that f (K) is compact if
K ⊆ Am is compact.
An affine space X of dimension d becomes a topological space if we give it the topology
for which the open subsets are of the form f −1 (U ), where U is any open subset of Ad and
f : X → Ad is an affine bijection.
Given any subset A of a topological space X, the smallest closed set containing A is
denoted by A, and is called the closure or adherence of A. A subset A of X is dense in X
◦
if A = X. The largest open set contained in A is denoted by A, and is called the interior of
A. The set Fr A = A ∩ X − A is called the boundary (or frontier ) of A. We also denote the
boundary of A by ∂A.
3.2 Convex Sets

Convex sets are defined as follows.
Definition 3.1. A subset V of a real affine space E is convex if for any two points a, b ∈ V ,
we have c ∈ V for every point c = (1 − λ)a + λb, with 0 ≤ λ ≤ 1 (λ ∈ R). Given any two
points a, b, the notation [a, b] is often used to denote the line segment between a and b, that
is,
[a, b] = {c ∈ E | c = (1 − λ)a + λb, 0 ≤ λ ≤ 1},
and thus a set V is convex if [a, b] ⊆ V for any two points a, b ∈ V (a = b is allowed).
The empty set is trivially convex, every one-point set {a} is convex, and the entire affine
space E is, of course, convex.
It is obvious that the intersection of any family (finite or infinite) of convex sets is
convex. Then, given any (nonempty) subset S of E, there is a smallest convex set containing
S denoted by conv(S) or C(S), and called the convex hull of S (namely, the intersection of
all convex sets containing S). The affine hull of a subset S of E is the smallest affine set
containing S and it will be denoted by hSi or aff(S).
Definition 3.2. Given any affine space E, the dimension of a nonempty convex subset S
of E, denoted by dim S, is the dimension of the smallest affine subset aff(S) containing S.
3.2. CONVEX SETS 59 1
(a) (b)
Figure 3.1: (a) A convex set; (b) A nonconvex set
A good understanding of what conv(S) is, and good methods for computing it, are
essential. First we have the following simple but crucial lemma:

→ −
Lemma 3.1. Given an affine P space E, E , +P, for any family (ai )i∈I of points in E, the set
V of convex combinations i∈I λi ai (where i∈I λi = 1 and λi ≥ 0) is the convex hull of
(ai )i∈I .
P
Proof. If (ai )i∈I is empty, then V = ∅, because of the condition i∈I λi = 1. As in the case
of affine combinations, it is easily shown by induction that any convex combination can be
obtained by computing convex combinations of two points at a time. As a consequence, if
(ai )i∈I is nonempty, then the smallest
P convex subspace containing (ai )i∈I must contain the
set V of all convex combinations i∈I λi ai . Thus, it is enough to show that V is closed
under convex combinations, which is immediately verified.
In view of Lemma 3.1, it is obvious that any affine subspace of E is convex.

Convex sets also arise in terms of hyperplanes. Given a hyperplane H, if f : E → R is
→
−
any nonconstant affine form defining H (i.e., H = Ker f , with f (a + u) = f (a) + f (u) for
→
− →
− → −
all a ∈ E and all u ∈ E , where f (a) ∈ R and f : E → R is a nonzero linear form), we can
define the two subsets
H+ (f ) = {a ∈ E | f (a) ≥ 0} and H− (f ) = {a ∈ E | f (a) ≤ 0},
called (closed) half-spaces associated with f .

Observe that if λ > 0, then H+ (λf ) = H+ (f ), but if λ < 0, then H+ (λf ) = H− (f ), and
similarly for H− (λf ). However, the set
{H+ (f ), H− (f )}
depends only on the hyperplane H, and the choice of a specific f defining H amounts
to the choice of one of the two half-spaces. For this reason, we will also say that H+ (f )
H+ (f )
H− (f )
Figure 3.2: The two half-spaces determined by a hyperplane, H
and H− (f ) are the closed half-spaces associated with H. Clearly, H+ (f ) ∪ H− (f ) = E

and H+ (f ) ∩ H− (f ) = H. It is immediately verified that H+ (f ) and H− (f ) are convex.
Bounded convex sets arising as the intersection of a finite family of half-spaces associated
with hyperplanes play a major role in convex geometry and topology (they are called convex
polytopes).
P
The convex combinations i∈I λi ai arising in computing Pthe convex hull of a family of
points (ai )i∈I have finite support, so they can be written as j∈J λj aj for some finite subset
J of I, but a priori there is no bound on the size such finite sets J. Thus it is natural to
wonder whether Lemma 3.1 can be sharpened in two directions:
(1) Is it possible to
P have a fixed bound on the number of points involved in the convex
combinations j∈J λj aj (that is, on the size of the index sets J)?
(2) Is it necessary to consider convex combinations of all points, or is it possible to consider

only a subset of points with special properties?
The answer is yes in both cases. In Case (1), assuming that the affine space E has dimen-
sion m, Carathéodory’s theorem asserts that it is enough to consider convex combinations
of m + 1 points. For example, in the plane A2 , the convex hull of a set S of points is the
union of all triangles (interior points included) with vertices in S. In Case (2), the theorem
of Krein and Milman asserts that a convex set that is also compact is the convex hull of its
extremal points (given a convex set S, a point a ∈ S is extremal if S − {a} is also convex,
see Berger [8] or Lang [38]). Next, we prove Carathéodory’s theorem.
3.3 Carathéodory’s Theorem

The proof of Carathéodory’s theorem is really beautiful. It proceeds by contradiction and
uses a minimality argument.
3.3. CARATHÉODORY’S THEOREM 61
Theorem 3.2. (Carathéodory, 1907) Given any affine space E of dimension m, for any
(nonvoid) family S = (ai )i∈L in E, the convex hull conv(S) of S is equal to the set of convex
combinations of families of m + 1 points of S.
Proof. By Lemma 3.1,

X X
conv(S) = λi ai | ai ∈ S, λi = 1, λi ≥ 0, I ⊆ L, I finite .
i∈I i∈I
We would like to prove that

X X
conv(S) = λi ai | ai ∈ S, λi = 1, λi ≥ 0, I ⊆ L, |I| = m + 1 .
i∈I i∈I
We proceed by contradiction. If the theorem is false, Pthere is some point b ∈ conv(S) such
that b can be expressed as a convex combination b = i∈I λi ai , where I ⊆ L is a finite set
of cardinality
P |I| = q with q ≥ m + 2, and b cannot be expressed as any convex combination
b = j∈J µj aj of strictly fewer than q points in S, that is, where |J| < q. Such a point
b ∈ conv(S) is a convex combination
b = λ 1 a1 + · · · + λ q aq ,
where λ1 + · · · + λq = 1 and λi > 0 (1 ≤ i ≤ q). We shall prove that b can be written as a

convex combination of q − 1 of the ai . Pick any origin O in E. Since there are q > m + 1
points a1 , . . . , aq , these points are affinely dependent, and by Lemma 2.6.5 from Gallier [30],
there is a family (µ1 , . . . , µq ) all scalars not all null, such that µ1 + · · · + µq = 0 and
q
X
µi Oai = 0.
i=1
Consider the set T ⊆ R defined by
T = {t ∈ R | λi + tµi ≥ 0, µi 6= 0, 1 ≤ i ≤ q}.
The set T is nonempty, since it contains 0. Since qi=1 µi = 0 and the µi are not all null,
P
there are some µh , µk such that µh < 0 and µk > 0, which implies that T = [α, β], where
α = max {−λi /µi | µi > 0} and β = min {−λi /µi | µi < 0}

1≤i≤q 1≤i≤q
(T is the intersection of the closed half-spaces {t ∈ R | λi + tµi ≥ 0, µi 6= 0}). Observe that

α < 0 < β, since λi > 0 for all i = 1, . . . , q.
We claim that there is some j (1 ≤ j ≤ q) such that
λj + αµj = 0.
Indeed, since
α = max {−λi /µi | µi > 0},
1≤i≤q
as the set on the right hand side is finite, the maximum is achieved and P
there is some index
j so that α = −λj /µj . If j is some index such that λj + αµj = 0, since qi=1 µi Oai = 0, we
have
q q
X X
b= λi ai = O + λi Oai + 0,
i=1 i=1
q q
X
X
=O+ λi Oai + α µi Oai ,
i=1 i=1
q
X
=O+ (λi + αµi )Oai ,
i=1
q
X
= (λi + αµi )ai ,
i=1
q
X
= (λi + αµi )ai ,
i=1, i6=j
Pq Pq
since λj + αµj = 0. Since i=1 µi = 0, i=1 λi = 1, and λj + αµj = 0, we have
q
X
λi + αµi = 1,
i=1, i6=j
and since λi + αµi ≥ 0 for i = 1, . . . , q, the above shows that b can be expressed as a convex
combination of q − 1 points from S. However, this contradicts the assumption that b cannot
be expressed as a convex combination of strictly fewer than q points from S, and the theorem
is proved.
If S is a finite (of infinite) set of points in the affine plane A2 , Theorem 3.2 confirms
our intuition that conv(S) is the union of triangles (including interior points) whose vertices
belong to S. Similarly, the convex hull of a set S of points in A3 is the union of tetrahedra
(including interior points) whose vertices belong to S. We get the feeling that triangulations
play a crucial role, which is of course true!
An interesting consequence of Carathéodory’s theorem is the following result:
Proposition 3.3. If K is any compact subset of Am , then the convex hull, conv(K), of K
is also compact. In particular, the convex hull conv(a1 , . . . , ap ) of a finite set of points is
compact, and thus closed (and bounded).
3.3. CARATHÉODORY’S THEOREM 63
Proof. Let C be the subset of Am+1 given by
C = {(λ0 , . . . , λm ) ∈ Rm+1 | λ0 + · · · + λm = 1, λi ≥ 0, i = 0, . . . , m}.
Clearly C is cut out by hyperplanes and is bounded (since 0 ≤ λi ≤ 1) so C is compact

(and also convex), and as K is compact, C × K m+1 is also compact. Then, consider the map
f : C × (Am )m+1 → Am given by
f (λ0 , . . . , λm , a0 , . . . , am ) = λ0 a0 + · · · + λm am .
This is obviously a continuous map. Furthermore, by Carathéodory’s theorem,
f (C × K m+1 ) = conv(K),
and since the image of a compact set by a continuous function is compact, we conclude that
conv(K) is compact.
A closer examination of the proof of Theorem 3.2 reveals that the fact that the µi ’s add
up to zero ensures that T is a closed interval, but all we need is that T be bounded from
below, and this only requires that some µj be strictly positive. As a consequence, we can
prove a version of Theorem 3.2 for convex cones. This is a useful result since cones play such
an important role in convex optimization. let us recall some basic definitions about cones.
Definition 3.3. Given any vector space E, a subset C ⊆ E is a convex cone iff C is closed
under positive linear combinations, that is, linear combinations of the form
X
λi vi , with vi ∈ C and λi ≥ 0 for all i ∈ I,
i∈I
where I has finite support (all λi = 0 except for finitely many i ∈ I). Given any set of
vectors S, the positive hull of S, or cone spanned by S, denoted cone(S), is the set of all
positive linear combinations of vectors in S,
( )
X
cone(S) = λi vi | vi ∈ S, λi ≥ 0 .
i∈I
Note that a cone always contains 0. When S consists of a finite number of vector, the con-
vex cone cone(S) is called a polyhedral cone. We have the following version of Carathéodory’s
theorem for convex cones:
Theorem 3.4. Given any vector space E of dimension m, for any (nonvoid) family S =
(vi )i∈L of vectors in E, the cone cone(S) spanned by S is equal to the set of positive combi-
nations of families of m vectors in S.
The proof of Theorem 3.4 can be easily adapted from the proof of Theorem 3.2 and is
left as an exercise.
There is an interesting generalization of Carathéodory’s theorem known as the Colorful
Carathéodory theorem. This theorem due to Bárány and proved in 1982 can be used to give
a fairly short proof of a generalization of Helly’s theorem known as Tverberg’s theorem (see
Section 3.5).
Theorem 3.5. (Colorful Carathéodory theorem) Let E be any affine space of dimension m.
For any point b ∈ E, for any sequence of m + 1 nonempty subsets (S1 , . . . , Sm+1 ) of E, if
b ∈ conv(Si ) for i = 1, . . . , m + 1, then there exists a sequence of m + 1 points (a1 , . . . , am+1 ),
with ai ∈ Si , so that b ∈ conv(a1 , . . . , am+1 ), that is, b is a convex combination of the ai ’s.
Although Theorem 3.5 is not hard to prove, we will not prove it here. Instead, we refer the
reader to Matousek [40], Chapter 8, Section 8.2. There is also a stronger version of Theorem
3.5, in which it is enough to assume that b ∈ conv(Si ∪ Sj ) for all i, j with 1 ≤ i < j ≤ m + 1.
Now that we have given an answer to the first question posed at the end of Section 3.2
we give an answer to the second question.
3.4 Vertices, Extremal Points and Krein and Milman’s

Theorem
First, we define the notions of separation and of separating hyperplanes. For this, recall the
definition of the closed (or open) half–spaces determined by a hyperplane.
Given a hyperplane H, if f : E → R is any nonconstant affine form defining H (i.e.,
H = Ker f ), we define the closed half-spaces associated with f by
H+ (f ) = {a ∈ E | f (a) ≥ 0},
H− (f ) = {a ∈ E | f (a) ≤ 0}.
Observe that if λ > 0, then H+ (λf ) = H+ (f ), but if λ < 0, then H+ (λf ) = H− (f ), and
similarly for H− (λf ).
Thus, the set {H+ (f ), H− (f )} depends only on the hyperplane H, and the choice of a
specific f defining H amounts to the choice of one of the two half-spaces.
We also define the open half–spaces associated with f as the two sets
◦
H + (f ) = {a ∈ E | f (a) > 0},
◦
H − (f ) = {a ∈ E | f (a) < 0}.
◦ ◦ ◦
The set {H + (f ), H − (f )} only depends on the hyperplane H. Clearly, we have H + (f ) =
◦
H+ (f ) − H and H − (f ) = H− (f ) − H.
3.4. VERTICES, EXTREMAL POINTS AND KREIN AND MILMAN’S THEOREM 65
1
A
A
H H
H′
B B
(a) (b)
Figure 3.3: (a) A separating hyperplane H. (b) Strictly separating hyperplanes H and H 0 .
Definition 3.4. Given an affine space E and two nonempty subsets A and B of E, we say
that a hyperplane H separates (resp. strictly separates) A and B if A is in one and B is in
the other of the two half–spaces (resp. open half–spaces) determined by H.
In Figure 3.3 (a), the two closed convex sets A and B are unbounded and B has the
hyperplane H for its boundary, while A is asymptotic to H. The hyperplane H is a separating
hyperplane for A and B but A and B can’t be strictly separated. In Figure 3.3 (b), both A
and B are convex and closed, B is unbounded and asymptotic to the hyperplane, H 0 , but A
is bounded. Both hyperplanes H and H 0 strictly separate A and B.
The special case of separation where A is convex and B = {a}, for some point, a, in A,
is of particular importance.
Definition 3.5. Let E be an affine space and let A be any nonempty subset of E. A
supporting hyperplane of A is any hyperplane H containing some point a of A, and separating
{a} and A. We say that H is a supporting hyperplane of A at a.
Observe that if H is a supporting hyperplane of A at a, then we must have a ∈ ∂A.

Otherwise, there would be some open ball B(a, ) of center a contained in A and so there
would be points of A (in B(a, )) in both half-spaces determined by H, contradicting the
◦
fact that H is a supporting hyperplane of A at a. Furthermore, H ∩ A = ∅.
One should experiment with various pictures and realize that supporting hyperplanes at
a point may not exist (for example, if A is not convex), may not be unique, and may have
several distinct supporting points! (See Figure 3.4).
Next, we need to define various types of boundary points of closed convex sets.
Definition 3.6. Let E be an affine space of dimension d. For any nonempty closed and
convex subset A of dimension d, a point a ∈ ∂A has order k(a) if the intersection of all
the supporting hyperplanes of A at a is an affine subspace of dimension k(a). We say that
Figure 3.4: Examples of supporting hyperplanes
a ∈ ∂A is a vertex if k(a) = 0; we say that a is smooth if k(a) = d − 1, i.e., if the supporting

hyperplane at a is unique. See Figure 3.5
A vertex is a boundary point a such that there are d independent supporting hyperplanes
at a. A d-simplex has boundary points of order 0, 1, . . . , d − 1. This phenomena is illustrated
in Figure 3.5. The following proposition is shown in Berger [8] (Proposition 11.6.2):
Proposition 3.6. The set of vertices of a closed and convex subset is countable.
Another important concept is that of an extremal point.

Definition 3.7. Let E be an affine space. For any nonempty convex subset A, a point
a ∈ ∂A is extremal (or extreme) if A − {a} is still convex.
It is fairly obvious that a point a ∈ ∂A is extremal if it does not belong to the interior of
any closed nontrivial line segment [x, y] ⊆ A (x 6= y, a 6= x and a 6= y).
Observe that a vertex is extremal, but the converse is false. For example, in Figure 3.6,
all the points on the arc of parabola, including v1 and v2 , are extreme points. However, only
v1 and v2 are vertices. Also, if dim E ≥ 3, the set of extremal points of a compact convex
may not be closed. See Berger [8], Chapter 11, Figure 11.6.5.3, which we reproduce in Figure
3.7.
Actually, it is not at all obvious that a nonempty compact convex set possesses extremal
points. In fact, a stronger results holds (Krein and Milman’s theorem). In preparation for
the proof of this important theorem, observe that any compact (nontrivial) interval of A1
has two extremal points, its two endpoints. We need the following lemma:
Lemma 3.7. Let E be an affine space of dimension n, and let A be a nonempty compact
and convex set. Then, A = conv(∂A), i.e., A is equal to the convex hull of its boundary.
Proof. Pick any a in A, and consider any line D through a. Then, D∩A is closed and convex.
However, since A is compact, it follows that D ∩ A is a closed interval [u, v] containing a,
and u, v ∈ ∂A. Therefore, a ∈ conv(∂A), as desired.
v
v
b
a b
a
(i.) k(a) = 2
v
b
a
(ii.) k(b) = 1
b
a
(iii.) k(v) = 0
Figure 3.5: The various types of boundary points for a solid tetrahedron. If the point is in
the interior of a triangular face, it has order 2. If the point is in the interior of an edge, it
has order 1. If the point is a vertex, it has order 0.
v1 v2
Figure 3.6: Examples of vertices and extreme points

(0,0,1)
(0,0,0)
(0,0,-1)
Figure 3.7: Let A the convex set formed by taking a planar unit circle through the origin
and forming the double cone with apex (0, 0, 1) and (0, 0, −1). The extremal points of A are
the points on pink circular boundary minus the origin.
The following important theorem shows that only extremal points matter as far as de-
termining a compact and convex subset from its boundary. The proof of Theorem 3.8 makes
use of a proposition due to Minkowski (Proposition 4.19) which will be proved in Section
4.2.
Theorem 3.8. (Krein and Milman, 1940) Let E be an affine space of dimension n. Every
compact and convex nonempty subset A is equal to the convex hull of its set of extremal
points.
Proof. Denote the set of extremal points of A by Extrem(A). We proceed by induction on
d = dimE. When d = 1, the convex and compact subset A must be a closed interval [u, v], or
a single point. In either cases, the theorem holds trivially. Now, assume d ≥ 2, and assume
that the theorem holds for d − 1. It is easily verified that
Extrem(A ∩ H) = (Extrem(A)) ∩ H,
for every supporting hyperplane H of A (such hyperplanes exist, by Minkowski’s proposition

(Proposition 4.19)). Observe that Lemma 3.7 implies that if we can prove that
∂A ⊆ conv(Extrem(A)),
then, since A = conv(∂A), we will have established that
A = conv(Extrem(A)).
Let a ∈ ∂A, and let H be a supporting hyperplane of A at a (which exists, by Minkowski’s

proposition). Now, A and H are convex so A ∩ H is convex; H is closed and A is compact,
so H ∩ A is a closed subset of a compact subset, A, and thus, A ∩ H is also compact. Since

A ∩ H is a compact and convex subset of H and H has dimension d − 1, by the induction
hypothesis, we have
A ∩ H = conv(Extrem(A ∩ H)).
However,
conv(Extrem(A ∩ H)) = conv((Extrem(A)) ∩ H)

= conv(Extrem(A)) ∩ H ⊆ conv(Extrem(A)),
and so, a ∈ A ∩ H ⊆ conv(Extrem(A)). Therefore, we proved that
∂A ⊆ conv(Extrem(A)),
from which we deduce that A = conv(Extrem(A)), as explained earlier.
Remark: Observe that Krein and Milman’s theorem implies that any nonempty compact
and convex set has a nonempty subset of extremal points. This is intuitively obvious, but
hard to prove! Krein and Milman’s theorem also applies to infinite dimensional affine spaces,
provided that they are locally convex, see Valentine [63], Chapter 11, Bourbaki [13], Chapter
II, Barvinok [4], Chapter 3, or Lax [39], Chapter 13.
An important consequence of Krein and Millman’s theorem is that every convex function
on a convex and compact set achieves its maximum at some extremal point.
Definition 3.8. Let A be a nonempty convex subset of An . A function f : A → R is convex
if
f ((1 − λ)a + λb) ≤ (1 − λ)f (a) + λf (b)
for all a, b ∈ A and for all λ ∈ [0, 1]. The function f : A → R is strictly convex if
f ((1 − λ)a + λb) < (1 − λ)f (a) + λf (b)
for all a, b ∈ A with a 6= b and for all λ with 0 < λ < 1. A function f : A → R is concave
(resp. strictly concave) iff −f is convex (resp. −f is strictly convex). See Figure 3.8.
If f is convex, a simple induction shows that

X X
f λi ai ≤ λi f (ai )
i∈I i∈I
for every finite convex combination

P in A, i.e., for any finite family (ai )i∈I of points in A and
any family (λi )i∈I with i∈I λi = 1 and λi ≥ 0 for all i ∈ I.
Proposition 3.9. Let A be a nonempty convex and compact subset of An and let f : A → R
be any function. If f is convex and continuous, then f achieves its maximum at some extreme
point of A.
l = (1-λ)f(u) + λf(v)
u v
(a)
l = (1-λ)f(u) + λf(v)
u v
(b)
Figure 3.8: Figures (a) and (b) are the graphs of real valued functions. Figure (a) is the
graph of convex function since the blue line lies above the graph of f . Figure (b) shows the
graph of a function which is not convex.
Proof. Since A is compact and f is continuous, f (A) is a closed interval, [m, M ], in R and so
f achieves its minimum m and its maximum M . Say f (c) = M , for some c ∈ A. By Krein
and Millman’s theorem, c is some convex combination of exteme points of A,
k
X
c= λi ai ,
i=1
Pk
with i=1 λi = 1, λi ≥ 0 and each ai an extreme point in A. But then, as f is convex,
k
! k
X X
M = f (c) = f λi ai ≤ λi f (ai )
i=1 i=1
and if we let
f (ai0 ) = max {f (ai )}
1≤i≤k
for some i0 such that 1 ≤ i0 ≤ k, then we get
k k
!
X X
M = f (c) ≤ λi f (ai ) ≤ λi f (ai0 ) = f (ai0 ),
i=1 i=1
3.5. RADON’S, TVERBERG’S, HELLY’S, THEOREMS AND CENTERPOINTS 71
as ki=1 λi = 1. Since M is the maximum value of the function f over A, we have f (ai0 ) ≤ M
P
and so,
M = f (ai0 )
and f achieves its maximum at the extreme point ai0 , as claimed.
Proposition 3.9 plays an important role in convex optimization: It guarantees that the
maximum value of a convex objective function on a compact and convex set is achieved at
some extreme point. Thus, it is enough to look for a maximum at some extreme point of
the domain.
Proposition 3.9 fails for minimal values of a convex function. For example, the function,
x 7→ f (x) = x2 , defined on the compact interval [−1, 1] achieves it minimum at x = 0, which
is not an extreme point of [−1, 1]. However, if f is concave, then f achieves its minimum
value at some extreme point of A. In particular, if f is affine, it achieves its minimum and
its maximum at some extreme points of A.
We conclude this chapter with three other classics of convex geometry.
3.5 Radon’s, Tverberg’s, Helly’s, Theorems and Cen-

terpoints
We begin with Radon’s theorem.
Theorem 3.10. (Radon, 1921) Given any affine space E of dimension m, for every subset X
of E, if X has at least m+2 points, then there is a partition of X into two nonempty disjoint
subsets X1 and X2 such that the convex hulls of X1 and X2 have a nonempty intersection.
Proof. Pick some origin O in E. Write X = (xi )i∈L for some index set L (we can let L = X).
Since by assumption |X| ≥ m + 2 where m = dim(E), X is affinely dependent, and by
Lemma 2.6.5 from Gallier [30], there is a family (µk )k∈L (of finite support) of scalars, not all
null, such that X X
µk = 0 and µk Oxk = 0.
k∈L k∈L
P
Since k∈L µk = 0, the µk are not all null, and (µk )k∈L has finite support, the sets
I = {i ∈ L | µi > 0} and J = {j ∈ L | µj < 0}
are nonempty, finite, and obviously disjoint. Let
X1 = {xi ∈ X | µi > 0} and X2 = {xi ∈ X | µi ≤ 0}.
P
Again, since the µk are not all null and k∈L µk = 0, the sets X1 and X2 are nonempty, and
obviously
X1 ∩ X2 = ∅ and X1 ∪ X2 = X.
Figure 3.9: Examples of Radon Partitions
Furthermore, the definition of I and J implies that (xi )i∈I ⊆ X1 and (xj )j∈J ⊆ X2 . It
remains to prove that conv(X1 ) ∩ conv(X2 ) 6= ∅. The definition of I and J implies that
X
µk Oxk = 0
k∈L
can be written as X X
µi Oxi + µj Oxj = 0,
i∈I j∈J
that is, as X X
µi Oxi = −µj Oxj ,
i∈I j∈J
where X X
µi = −µj = µ,
i∈I j∈J
with µ > 0. Thus, we have

X µi X µj
Oxi = − Oxj ,
i∈I
µ j∈J
µ
with X µi X µj
= − = 1,
i∈I
µ j∈J
µ
P P
proving that i∈I (µi /µ)xi ∈ conv(X1 ) and j∈J −(µj /µ)xj ∈ conv(X2 ) are identical, and
thus that conv(X1 ) ∩ conv(X2 ) 6= ∅.
A partition, (X1 , X2 ), of X satisfying the conditions of Theorem 3.10 is sometimes called

a Radon partition of X and any point in conv(X1 ) ∩ conv(X2 ) is called a Radon point of X.
Figure 3.9 shows two Radon partitions of five points in the plane.
Figure 3.10: The Radon Partitions of four points (in A2 )
It can be shown that a finite set, X ⊆ E, has a unique Radon partition iff it has m + 2
elements and any m + 1 points of X are affinely independent. For example, there are exactly
two possible cases in the plane as shown in Figure 3.10.
There is also a version of Radon’s theorem for the class of cones with an apex. Say that
a convex cone, C ⊆ E, has an apex (or is a pointed cone) iff there is some hyperplane, H,
such that C ⊆ H+ and H ∩ C = {0}. For example, the cone obtained as the intersection of
two half spaces in R3 is not pointed since it is a wedge with a line as part of its boundary.
Here is the version of Radon’s theorem for convex cones:
Theorem 3.11. Given any vector space E of dimension m, for every subset X of E, if
cone(X) is a pointed cone such that X has at least m + 1 nonzero vectors, then there is a
partition of X into two nonempty disjoint subsets, X1 and X2 , such that the cones, cone(X1 )
and cone(X2 ), have a nonempty intersection not reduced to {0}.
The proof of Theorem 3.11 is left as an exercise.

There is a beautiful generalization of Radon’s theorem known as Tverberg’s Theorem.
Theorem 3.12. (Tverberg’s Theorem, 1966) Let E be any affine space of dimension m. For
any natural number, r ≥ 2, for every subset, X, of E, if X has at least (m + 1)(r − 1) + 1
T there is a partition, (X1 , . . . , Xr ), of X into r nonempty pairwise disjoint subsets
points, then
so that ri=1 conv(Xi ) 6= ∅.
A partition as in Theorem 3.12 is called a Tverberg partition and a point in ri=1 conv(Xi )
T
is called a Tverberg point. Theorem 3.12 was conjectured by Birch and proved by Tverberg
in 1966. Tverberg’s original proof was technically quite complicated. Tverberg then gave a
simpler proof in 1981 and other simpler proofs were later given, notably by Sarkaria (1992)
and Onn (1997), using the Colorful Carathéodory theorem. A proof along those lines can be
found in Matousek [40], Chapter 8, Section 8.3. A colored Tverberg theorem and more can
also be found in Matousek [40] (Section 8.3).
Next, we prove a version of Helly’s theorem.
Theorem 3.13. (Helly, 1913) Given any affine space E of dimension m, for T every family
{K1 , . . . , Kn } of n convex subsets of E, if n ≥ m + 2 and the intersection i∈I Ki of any
m + 1 of the Ki is nonempty (where I ⊆ {1, . . . , n}, |I| = m + 1), then ni=1 Ki is nonempty.
T
Proof. The proof is by induction on n ≥ m + 1 and uses Radon’s theorem in the induction
step. For n = m + 1, the assumption of the theorem is that the intersection of any family of
T Next, let L = {1, 2, . . . , n+1},
m+1 of the Ki ’s is nonempty, and the theorem holds trivially.
where n + 1 ≥ m + 2. By the induction hypothesis, Ci = j∈(L−{i}) Kj is nonempty for every
i ∈ L.
We claim that Ci ∩ Cj 6= ∅ for some i 6= j. If so, as Ci ∩ Cj = n+1
T
k=1 Kk , we are done. So,
let us assume that the Ci ’s are pairwise disjoint. Then, we can pick a set X = {a1 , . . . , an+1 }
such that ai ∈ Ci , for every i ∈ L. By Radon’s Theorem, there are two nonempty disjoint
sets X1 , X2 ⊆ X such that X = X1 ∪ X2 and conv(X1 ) ∩ conv(X2 ) 6= ∅. However, X1 ⊆ Kj
for every j with aj ∈/ X1 . This is because aj ∈/ Kj for every j, and so, we get
\
X1 ⊆ Kj .
aj ∈X
/ 1
Symetrically, we also have \

X2 ⊆ Kj .
aj ∈X
/ 2
Since the Kj ’s are convex and

   
\ \ n+1
\
 Kj  ∩  Kj  = Ki ,
aj ∈X
/ 1 aj ∈X
/ 2 i=1
Tn+1 Tn+1
it follows that conv(X1 ) ∩ conv(X2 ) ⊆ i=1 Ki , so that i=1 Ki is nonempty, contradicting
the fact that Ci ∩ Cj = ∅ for all i 6= j.
A more general version of Helly’s theorem is proved in Berger [8].

An amusing corollary of Helly’s theorem is the following result: Consider n ≥ 4 line
segments in the affine plane A2 lying on disjoint parallel lines. If every three of these line
segments meet a line, then all of these line segments meet a common line.
To prove this fact, pick a coordinate system in which the y-axis is parallel to the common
direction of the parallel lines, and for every line segment S, let
CS = {(α, β) ∈ R2 | the line y = αx + β meets S}.
It is not hard to see that CS is convex. Then, by hypothesis the fact that any three line
segments Si , Sj , Sk meet a line means that CSi ∩CSj ∩CSk 6= ∅, any Helly’s Theorem implies
that the family of all the convex sets CSi has a nonempty intersection, which means that
there is a line meeting all the line segments Si . This situation for four lines is illustrated in
Figure 3.11.
We conclude this chapter with a nice application of Helly’s Theorem to the existence
of centerpoints. Centerpoints generalize the notion of median to higher dimensions. Recall
y=
αx
+ S β‘
β
α ‘x +
y=
C(S)
(0,0)
Figure 3.11: The four pink line segments in the affine plane all intersect the horizontal red
line.
that if we have a set of n data points, S = {a1 , . . . , an }, on the real line, a median for S is
a point, x, such that both intervals [x, ∞) and (−∞, x] contain at least n/2 of the points in
S (by n/2, we mean the largest integer greater than or equal to n/2).
Given any hyperplane, H, recall that the closed half-spaces determined by H are denoted
◦ ◦
H+ and H− and that H ⊆ H+ and H ⊆ H− . We let H+ = H+ − H and H− = H− − H be
the open half-spaces determined by H.
Definition 3.9. Let S = {a1 , . . . , an } be a set of n points in Ad . A point, c ∈ Ad , is a
centerpoint of S iff for every hyperplane, H, whenever the closed half-space H+ (resp. H− )
n n
contains c, then H+ (resp. H− ) contains at least d+1 points from S (by d+1 , we mean the
n n n
largest integer greater than or equal to d+1 , namely the ceiling d d+1 e of d+1 ).
So, for d = 2, for each line, D, if the closed half-plane D+ (resp. D− ) contains c, then
D+ (resp. D− ) contains at least a third of the points from S. For d = 3, for each plane, H,
if the closed half-space H+ (resp. H− ) contains c, then H+ (resp. H− ) contains at least a
fourth of the points from S, etc. Example 3.12 shows nine points in the plane and one of
their centerpoints (in red). This example shows that the bound 13 is tight.
Observe that a point, c ∈ Ad , is a centerpoint of S iff c belongs to every open half-space,
◦ ◦
dn dn
H+ (resp. H− ) containing at least d+1
+ 1 points from S (again, we mean d d+1 e + 1).
◦ ◦
Indeed, if c is a centerpoint of S and H is any hyperplane such that H+ (resp. H− )
◦ ◦
dn
contains at least d+1 + 1 points from S, then H+ (resp. H− ) must contain c as otherwise,
dn n
the closed half-space, H− (resp. H+ ) would contain c and at most n − d+1 − 1 = d+1 −1
points from S, a contradiction. Conversely, assume that c belongs to every open half-space,
◦ ◦
dn
H+ (resp. H− ) containing at least d+1 + 1 points from S. Then, for any hyperplane, H,
n
if c ∈ H+ (resp. c ∈ H− ) but H+ contains at most d+1 − 1 points from S, then the open
Figure 3.12: Example of a centerpoint
◦ ◦
n dn
half-space, H− (resp. H+ ) would contain at least n − d+1
+1 = d+1
+ 1 points from S but
not c, a contradiction.
We are now ready to prove the existence of centerpoints.
Theorem 3.14. (Existence of Centerpoints) Every finite set, S = {a1 , . . . , an }, of n points
in Ad has some centerpoint.
Proof. We will use the second characterization of centerpoints involving open half-spaces
dn
containing at least d+1 + 1 points.
Consider the family of sets,

◦ ◦ dn
C = conv(S ∩ H+ ) | (∃H) |S ∩ H+ | >
d+1

◦ ◦ dn
∪ conv(S ∩ H− ) | (∃H) |S ∩ H− | > ,
d+1
where H is a hyperplane.
AsTS is finite, C consists of a finite number of convex sets, say {C1 , . . . , Cm }. If we prove
that m
Tm
C
i=1 i 6
= ∅ we are done, because i=1 i is the set of centerpoints of S.
C
First, we prove by induction on k (with 1 ≤ k ≤ d + 1), that any intersection of k of the
Ci ’s has at least (d+1−k)n
d+1
+ k elements from S. For k = 1, this holds by definition of the Ci ’s.
Next, consider the intersection of k + 1 ≤ d + 1 of the Ci ’s, say Ci1 ∩ · · · ∩ Cik ∩ Cik+1 . Let
A = S ∩ (Ci1 ∩ · · · ∩ Cik ∩ Cik+1 )

B = S ∩ (Ci1 ∩ · · · ∩ Cik )
C = S ∩ Cik+1 .
(d+1−k)n
Note that A = B ∩ C. By the induction hypothesis, B contains at least d+1
+ k elements
dn
from S. As C contains at least d+1 + 1 points from S, and as
|B ∪ C| = |B| + |C| − |B ∩ C| = |B| + |C| − |A|
and |B ∪ C| ≤ n, we get n ≥ |B| + |C| − |A|, that is,
|A| ≥ |B| + |C| − n.
It follows that
(d + 1 − k)n dn
|A| ≥ +k+ +1−n
d+1 d+1
that is,
(d + 1 − k)n + dn − (d + 1)n (d + 1 − (k + 1))n

|A| ≥ +k+1= + k + 1,
d+1 d+1
establishing the induction hypothesis.
Now, if m ≤ d + 1, the above claim for k = m shows that m
T
i=1 Ci 6= ∅ and we are done.
If m ≥ d + 2, the above claim for k = d + 1 shows that any intersection of d + 1 of the Ci ’s
is nonempty. Consequently, the conditions for applying Helly’s Theorem are satisfied and
therefore,
\m
Ci 6= ∅.
i=1
Tm
However, i=1 Ci is the set of centerpoints of S and we are done.
Remark: The above proof actually shows that the set of centerpoints of S is a convex set.
In fact, it is a finite intersection of convex hulls of finitely many points, so it is the convex hull
of finitely many points, in other words, a polytope. It should also be noted that Theorem
3.14 can be proved easily using Tverberg’s theorem (Theorem 3.12). Indeed, for a judicious
choice of r, any Tverberg point is a centerpoint!
Jadhav and Mukhopadhyay have given a linear-time algorithm for computing a center-
point of a finite set of points in the plane. For d ≥ 3, it appears that the best that can
be done (using linear programming) is O(nd ). However, there are good approximation algo-
rithms (Clarkson, Eppstein, Miller, Sturtivant and Teng) and in E3 there is a near quadratic
algorithm (Agarwal, Sharir and Welzl). Recently, Miller and Sheehy (2009) have given an
algorithm for finding an approximate centerpoint in sub-exponential time together with a
polynomial-checkable proof of the approximation guarantee.
Chapter 4
Two Main Tools: Separation and

Polar Duality
4.1 Separation Theorems and Farkas Lemma

It seems intuitively rather obvious that if A and B are two nonempty disjoint convex sets in
A2 , then there is a line, H, separating them, in the sense that A and B belong to the two
(disjoint) open half–planes determined by H. However, this is not always true! For example,
this fails if both A and B are closed and unbounded (find an example). Nevertheless, the
result is true if both A and B are open, or if the notion of separation is weakened a little
bit. The key result, from which most separation results follow, is a geometric version of the
Hahn-Banach theorem. In the sequel, we restrict our attention to real affine spaces of finite
dimension. Then, if X is an affine space of dimension d, there is an affine bijection f between
X and Ad .
In order to prove the Hahn-Banach theorem, we will need two lemmas. Given any two
distinct points x, y ∈ X, we let
]x, y[ = {(1 − λ)x + λy ∈ X | 0 < λ < 1}.
Our first lemma (Lemma 4.1) is intuitively quite obvious so the reader might be puzzled by
the length of its proof. However, after proposing several wrong proofs, we realized that its
proof is more subtle than it might appear. The proof below is due to Valentine [63]. See if
you can find a shorter (and correct) proof!
◦
Lemma 4.1. Let S be a nonempty convex set and let x ∈ S and y ∈ S. Then, we have
◦
]x, y[ ⊆ S.
◦
Proof. Let z ∈ ]x, y[ , that is, z = (1 − λ)x + λy, with 0 < λ < 1. Since x ∈ S, we can find
some open subset, U , contained in S so that x ∈ U . It is easy to check that the central
magnification of center z, Hz, λ−1 , maps x to y. Then, V = Hz, λ−1 (U ) is an open subset
λ λ
79
80 CHAPTER 4. TWO MAIN TOOLS: SEPARATION AND POLAR DUALITY
V
U W v
x z y
Figure 4.1: Illustration for the proof of Lemma 4.1
containing y and as y ∈ S, we have V ∩ S 6= ∅. Let v ∈ V ∩ S be a point of S in this

intersection. Now, there is a unique point, u ∈ U ⊆ S, such that Hz, λ−1 (u) = v and, as S is
λ
convex, we deduce that z = (1 − λ)u + λv ∈ S. Since U is open, the set
W = (1 − λ)U + λv = {(1 − λ)w + λv | w ∈ U } ⊆ S

◦
is also open and z ∈ W , which shows that z ∈ S.
◦ ◦ ◦
Corollary 4.2. If S is convex, then S is also convex, and we have S = S. Furthermore, if
◦ ◦
S 6= ∅, then S = S.
Beware that if S is a closed set, then the convex hull conv(S) of S is not necessarily
closed!
For example, consider the subset S of A2 consisting of the points belonging to the right
branch of the hyperbola of equation x2 − y 2 = 1, that is,
S = {(x, y) ∈ R2 | x2 − y 2 ≥ 1, x ≥ 0}.
Then S is convex, but the convex hull of the set S ∪ {(0, 0} is not closed.
However, if S is compact, then conv(S) is also compact, and thus closed (see Proposition
3.3).
There is a simple criterion to test whether a convex set has an empty interior, based on
the notion of dimension of a convex set (recall that the dimension of a nonempty convex
subset is the dimension of its affine hull).
Proposition 4.3. A nonempty convex set S has a nonempty interior iff dim S = dim X.
4.1. SEPARATION THEOREMS AND FARKAS LEMMA 81
◦
Proof. Let d = dim X. First, assume that S 6= ∅. Then, S contains some open ball of center
a0 , and in it, we can find a frame (a0 , a1 , . . . , ad ) for X. Thus, dim S = dim X. Conversely,
let (a0 , a1 , . . . , ad ) be a frame of X, with ai ∈ S, for i = 0, . . . , d. Then, we have
a0 + · · · + ad ◦
∈ S,
d+1
◦
and S is nonempty.
Proposition 4.3 is false in infinite dimension.
We leave the following property as an exercise:

Proposition 4.4. If S is convex, then S is also convex.
One can also easily prove that convexity is preserved under direct image and inverse
image by an affine map.
The next lemma, which seems intuitively obvious, is the core of the proof of the Hahn-
Banach theorem. This is the case where the affine space has dimension two. First, we need
to define what is a convex cone with vertex x.
Definition 4.1. A convex set, C, is a convex cone with vertex x if C is invariant under all
central magnifications, Hx,λ , of center x and ratio λ, with λ > 0 (i.e., Hx,λ (C) = C). See
Figure 4.2.
Figure 4.2: For the dark pink disk C, Hx,λ (C) is the triangular section, excluding O, between
the two pink lines.
Given a convex set, S, and a point, x ∈

/ S, we can define
[
conex (S) = Hx,λ (S).
λ>0
It is easy to check that this is a convex cone with vertex x.

L
O x
Figure 4.3: Hahn-Banach Theorem in the plane (Lemma 4.5)
Lemma 4.5. Let B be a nonempty open and convex subset of A2 , and let O be a point of
A2 so that O ∈
/ B. Then, there is some line, L, through O, so that L ∩ B = ∅.
Proof. Define the convex cone C = coneO (B). As B is open, it is easy to check that each
HO,λ (B) is open and since C is the union of the HO,λ (B) (for λ > 0), which are open, C
itself is open. Also, O ∈
/ C. We claim that at least one point, x, of the boundary, ∂C, of C,
is distinct from O. Otherwise, ∂C = {O} and we claim that C = A2 − {O}, which is not
convex, a contradiction. Indeed, as C is convex it is connected, A2 − {O} itself is connected
and C ⊆ A2 − {O}. If C 6= A2 − {O}, pick some point a 6= O in A2 − C and some point
c ∈ C. Now, a basic property of connectivity asserts that every continuous path from a (in
the exterior of C) to c (in the interior of C) must intersect the boundary of C, namely, {O}.
However, there are plenty of paths from a to c that avoid O, a contradiction. Therefore,
C = A2 − {O}.
Since C is open and x ∈ ∂C, we have x ∈
/ C. Furthermore, we claim that y = 2O − x (the
◦
symmetric of x w.r.t. O) does not belong to C either. Otherwise, we would have y ∈ C = C
and x ∈ C, and by Lemma 4.1, we would get O ∈ C, a contradiction. Therefore, the line
through O and x misses C entirely (since C is a cone), and thus, B ⊆ C.
Finally, we come to the Hahn-Banach theorem.
Theorem 4.6. (Hahn-Banach Theorem, geometric form) Let X be a (finite-dimensional)

affine space, A be a nonempty open and convex subset of X and L be an affine subspace of
X so that A ∩ L = ∅. Then, there is some hyperplane, H, containing L, that is disjoint from
A.
A H
Figure 4.4: Hahn-Banach Theorem, geometric form (Theorem 4.6)
Proof. The case where dim X = 1 is trivial. Thus, we may assume that dim X ≥ 2. We
reduce the proof to the case where dim X = 2. Let V be an affine subspace of X of maximal
dimension containing L and so that V ∩ A = ∅. Pick an origin O ∈ L in X, and consider the
vector space XO . We would like to prove that V is a hyperplane, i.e., dimV = dimX −1. We
proceed by contradiction. Thus, assume that dim V ≤ dim X − 2. In this case, the quotient
space X/V has dimension at least 2. We also know that X/V is isomorphic to the orthogonal
complement, V ⊥ , of V so we may identify X/V and V ⊥ . The (orthogonal) projection map,
π : X → V ⊥ , is linear, continuous, and we can show that π maps the open subset A to an
open subset π(A), which is also convex (one way to prove that π(A) is open is to observe that
for any point, a ∈ A, a small open ball of center a contained in A is projected by π to an open
ball contained in π(A) and as π is surjective, π(A) is open). Furthermore, O ∈ / π(A). Since
⊥
V has dimension at least 2, there is some plane P (a subspace of dimension 2) intersecting
π(A), and thus, we obtain a nonempty open and convex subset B = π(A) ∩ P in the plane
P ∼ = A2 . So, we can apply Lemma 4.5 to B and the point O = 0 in P ∼ = A2 to find a line,
l, (in P ) through O with l ∩ B = ∅. But then, l ∩ π(A) = ∅ and W = π −1 (l) is an affine
subspace such that W ∩ A = ∅ and W properly contains V , contradicting the maximality of
V . See Figure 4.5.
Remark: The geometric form of the Hahn-Banach theorem also holds when the dimension
of X is infinite but a slightly more sophisticated proof is required. Actually, all that is needed
is to prove that a maximal affine subspace containing L and disjoint from A exists. This can
be done using Zorn’s lemma. For other proofs, see Bourbaki [13], Chapter 2, Valentine [63],
Chapter 2, Barvinok [4], Chapter 2, or Lax [39], Chapter 3.
Theorem 4.6 is false if we omit the assumption that A is open.
For a counter-example, let A ⊆ A2 be the union of the half space y < 0 with the closed
V
O
L=V
π (A ) V
O
l
L=V
A
l V
O
L=V
Figure 4.5: An illustration of the proof of 4.6. Let X = A3 , A be the open spherical ball,
and L the vertical purple line. The blue hyperplane, which strictly separates A from L is
construction using V ⊥ and l.
segment [0, 1] on the x-axis and let L be the point (2, 0) on the boundary of A. It is also
false if A is closed as shown by the following counter-example.
In E3 , consider the closed convex set (cone) A defined by the inequalities
x ≥ 0, y ≥ 0, z ≥ 0, z 2 ≤ xy,
and let D be the line given by x = 0, z = 1. Then D ∩ A = ∅, both A and D are convex and
closed, yet every plane containing D meets A.
Theorem 4.6 has many important corollaries. For example, we will eventually prove that
for any two nonempty disjoint convex sets, A and B, there is a hyperplane separating A and
B, but this will take some work (recall the definition of a separating hyperplane given in
Definition 3.4). We begin with the following version of the Hahn-Banach theorem:
Theorem 4.7. (Hahn-Banach, second version) Let X be a (finite-dimensional) affine space,
A be a nonempty convex subset of X with nonempty interior and L be an affine subspace of
X so that A ∩ L = ∅. Then, there is some hyperplane, H, containing L and separating L
and A.
A H
Figure 4.6: Hahn-Banach Theorem, second version (Theorem 4.7)
◦ ◦
Proof. Since A is convex, by Corollary 4.2, A is also convex. By hypothesis, A is nonempty.
◦
So, we can apply Theorem 4.6 to the nonempty open and convex A and to the affine subspace
◦ ◦ ◦
L. We get a hyperplane H containing L such that A ∩ H = ∅. However, A ⊆ A = A and A
◦
is contained in the closed half space (H+ or H− ) containing A, so H separates A and L.
Corollary 4.8. Given an affine space, X, let A and B be two nonempty disjoint convex
◦
subsets and assume that A has nonempty interior (A 6= ∅). Then, there is a hyperplane
separating A and B.
Proof. Pick some origin O and consider the vector space XO . Define C = A − B (a special
case of the Minkowski sum) as follows:
[
A − B = {a − b | a ∈ A, b ∈ B} = (A − b).
b∈B
It is easily verified that C = A−B is convex and has nonempty interior (as a union of subsets
having a nonempty interior). Furthermore O ∈ / C, since A∩B = ∅.1 (Note that the definition
◦
depends on the choice of O, but this has no effect on the proof.) Since C is nonempty, we
can apply Theorem 4.7 to C and to the affine subspace {O} and we get a hyperplane, H,
1
Readers who prefer a purely affine argument may define C = A − B as the affine subset
A − B = {O + a − b | a ∈ A, b ∈ B}.
Again, O ∈
/ C and C is convex. We can pick the affine form, f , defining a separating hyperplane, H, of C
and {O}, so that f (O + a − b) ≤ f (O), for all a ∈ A and all b ∈ B, i.e., f (a) ≤ f (b).
86 CHAPTER 4. TWO MAIN TOOLS: SEPARATION AND POLAR DUALITY 1
A H
Figure 4.7: Separation Theorem, version 1 (Corollary 4.8)
separating C and {O}. Let f be any linear form defining the hyperplane H. We may assume
that f (a − b) ≤ 0, for all a ∈ A and all b ∈ B, i.e., f (a) ≤ f (b). Consequently, if we let
α = sup{f (a) | a ∈ A} (which makes sense, since the set {f (a) | a ∈ A} is bounded), we have
f (a) ≤ α for all a ∈ A and f (b) ≥ α for all b ∈ B, which shows that the affine hyperplane
defined by f − α separates A and B.
Remark: Theorem 4.7 and Corollary 4.8 also hold in the infinite dimensional case, see Lax
[39], Chapter 3, or Barvinok, Chapter 3.
Since a hyperplane, H, separating A and B as in Corollary 4.8 is the boundary of each
of the two half–spaces that it determines, we also obtain the following corollary:
Corollary 4.9. Given an affine space, X, let A and B be two nonempty disjoint open and
convex subsets. Then, there is a hyperplane strictly separating A and B.
Beware that Corollary 4.9 fails for closed convex sets.
However, Corollary 4.9 holds if we also assume that A (or B) is compact, as shown in
Corollary 4.10.
We need to review the notion of distance from a point to a subset. Let X be a metric
space with distance function, d. Given any point, a ∈ X, and any nonempty subset, B, of
X, we let
d(a, B) = inf d(a, b)
b∈B
(where inf is the notation for least upper bound).

Now, if X is an affine space of dimension d, it can be given a metric structure by giving
the corresponding vector space a metric structure, for instance, the metric induced by a
Euclidean structure. We have the following important property: For any nonempty closed
subset, S ⊆ X (not necessarily convex), and any point, a ∈ X, there is some point s ∈ S
“achieving the distance from a to S,” i.e., so that
d(a, S) = d(a, s).
The proof uses the fact that the distance function is continuous and that a continuous
function attains its minimum on a compact set, and is left as an exercise.
Corollary 4.10. Given an affine space, X, let A and B be two nonempty disjoint closed
and convex subsets, with A compact. Then, there is a hyperplane strictly separating A and
B.
Proof sketch. First, we pick an origin O and we give XO ∼

= An a Euclidean structure. Let d
denote the associated distance. Given any subsets A of X, let
A + B(O, ) = {x ∈ X | d(x, A) < },
where B(a, ) denotes the open ball, B(a, ) = {x ∈ X | d(a, x) < }, of center a and radius
> 0. Note that
[
A + B(O, ) = B(a, ),
a∈A
which shows that A + B(O, ) is open; furthermore it is easy to see that if A is convex, then
A + B(O, ) is also convex. Now, the function a 7→ d(a, B) (where a ∈ A) is continuous and
since A is compact, it achieves its minimum, d(A, B) = mina∈A d(a, B), at some point, a, of A.
Say, d(A, B) = δ. Since B is closed, there is some b ∈ B so that d(A, B) = d(a, B) = d(a, b)
and since A ∩ B = ∅, we must have δ > 0. Thus, if we pick < δ/2, we see that
(A + B(O, )) ∩ (B + B(O, )) = ∅.
Now, A+B(O, ) and B +B(O, ) are open, convex and disjoint and we conclude by applying
Corollary 4.9.
Finally, we have the separation theorem announced earlier for arbitrary nonempty convex
subsets.
Theorem 4.11. (Separation of disjoint convex sets) Given an affine space, X, let A and B
be two nonempty disjoint convex subsets. Then, there is a hyperplane separating A and B.
88 CHAPTER 4. TWO MAIN TOOLS: SEPARATION AND POLAR DUALITY 1
x A+x
O A
H
D B
−x
A−x
Figure 4.8: Separation Theorem, final version (Theorem 4.11)
Proof. The proof is by descending induction on dim A. If dim A = dim X, we know from
Proposition 4.3 that A has nonempty interior and we conclude using Corollary 4.8. Next,
asssume that the induction hypothesis holds if dim A ≥ n and assume dim A = n − 1. Pick
an origin O ∈ A and let H be a hyperplane containing A. Pick x ∈ X outside H and define
C = conv(A ∪ {A + x}) where A + x = {a + x | a ∈ A} and D = conv(A ∪ {A − x}) where
A − x = {a − x | a ∈ A}. Note that C ∪ D is convex. If B ∩ C 6= ∅ and B ∩ D 6= ∅, then
the convexity of B and C ∪ D implies that A ∩ B 6= ∅, a contradiction. Without loss of
generality, assume that B ∩ C = ∅. Since x is outside H, we have dim C = n and by the
induction hypothesis, there is a hyperplane, H1 separating C and B. As A ⊆ C, we see that
H1 also separates A and B.
Remarks:
(1) The reader should compare this proof (from Valentine [63], Chapter II) with Berger’s
proof using compactness of the projective space Pd , see Berger [8] (Corollary 11.4.7).
(2) Rather than using the Hahn-Banach theorem to deduce separation results, one may
proceed differently and use the following intuitively obvious lemma, as in Valentine
[63] (Theorem 2.4):
Lemma 4.12. If A and B are two nonempty convex sets such that A ∪ B = X and
A ∩ B = ∅, then V = A ∩ B is a hyperplane.
One can then deduce Corollaries 4.8 and Theorem 4.11. Yet another approach is
followed in Barvinok [4].
(3) How can some of the above results be generalized to infinite dimensional affine spaces,
especially Theorem 4.6 and Corollary 4.8? One approach is to simultaneously relax
the notion of interior and tighten a little the notion of closure, in a more “linear and
less topological” fashion, as in Valentine [63].
Given any subset A ⊆ X (where X may be infinite dimensional, but is a Hausdorff

topological vector space), say that a point x ∈ X is linearly accessible from A iff there
is some a ∈ A with a 6= x and ]a, x[ ⊆ A. We let lina A be the set of all points linearly
accessible from A and lin A = A ∪ lina A.
A point a ∈ A is a core point of A iff for every y ∈ X, with y 6= a, there is some

z ∈]a, y[ , such that [a, z] ⊆ A. The set of all core points is denoted core A.
◦
It is not difficult to prove that lin A ⊆ A and A ⊆ core A. If A has nonempty interior,
◦
then lin A = A and A = core A. Also, if A is convex, then core A and lin A are convex.
Then, Lemma 4.12 still holds (where X is not necessarily finite dimensional) if we
redefine V as V = lin A ∩ lin B and allow the possibility that V could be X itself.
Corollary 4.8 also holds in the general case if we assume that core A is nonempty. For
details, see Valentine [63], Chapter I and II.
(4) Yet another approach is to define the notion of an algebraically open convex set, as
in Barvinok [4]. A convex set, A, is algebraically open iff the intersection of A with
every line, L, is an open interval, possibly empty or infinite at either end (or all of
L). An open convex set is algebraically open. Then, the Hahn-Banach theorem holds
provided that A is an algebraically open convex set and similarly, Corollary 4.8 also
holds provided A is algebraically open. For details, see Barvinok [4], Chapter 2 and 3.
We do not know how the notion “algebraically open” relates to the concept of core.
(5) Theorems 4.6, 4.7 and Corollary 4.8 are proved in Lax [39] using the notion of gauge
function in the more general case where A has some core point (but beware that Lax
uses the terminology interior point instead of core point!).
An important special case of separation is the case where A is convex and B = {a}, for
some point, a, in A.
A “cute” application of Corollary 4.10 is one of the many versions of “Farkas Lemma”
(1893-1894, 1902), a basic result in the theory of linear programming. For any vector,
x = (x1 , . . . , xn ) ∈ Rn , and any real, α ∈ R, write x ≥ α iff xi ≥ α, for i = 1, . . . , n.
The proof of Farkas Lemma Version I (Proposition 4.14) relies on the fact that a poly-
hedral cone cone(a1 , . . . , am ) is closed. Although it seems obvious that a polyhedral cone
should be closed, a rigorous proof is not entirely trivial.
Indeed, the fact that a polyhedral cone is closed relies crucially on the fact that C is
spanned by a finite number of vectors, because the cone generated by an infinite set may
not be closed. For example, consider the closed disk D ⊆ R2 of center (0, 1) and radius 1,
which is tangent to the x-axis at the origin. Then the cone(D) consists of the open upper
half-plane plus the origin (0, 0), but this set is not closed.
Proposition 4.13. Every polyhedral cone C is closed.
Proof. This is proved by showing that
1. Every primitive cone is closed.
2. A polyhedral cone C is the union of finitely many primitive cones, where a primitive
cone is a polyhedral cone spanned by linearly independent vectors.
Assume that (a1 , . . . , am ) are linearly independent vectors in Rn , and consider any se-
quence (x(k) )k≥0
m
(k)
X
(k)
x = λ i ai
i=1
(k)
of vectors in the primitive cone cone({a1 , . . . , am }), which means that λj ≥ 0 for i =
1, . . . , m and all k ≥ 0. The vectors x(k) belong to the subspace U spanned by (a1 , . . . , am ),
and U is closed. Assume that the sequence (x(k) )k≥0 converges to a limit x ∈ Rn . Since U
is closed and x(k) ∈ U for all k ≥ 0, we have x ∈ U . If we write x = x1 a1 + · · · + xm am , we
would like to prove that xi ≥ 0 for i = 1, . . . , m. The sequence the (x(k) )k≥0 converges to x
iff
lim x(k) − x = 0,

k7→∞
iff
m
X 1/2
(k)
lim |λi − xi |2 =0
k7→∞
i=1
iff
(k)
lim λi = xi , i = 1, . . . , m.
k7→∞
(k)
Since λi ≥ 0 for i = 1, . . . , m and all k ≥ 0, we have xi ≥ 0 for i = 1, . . . , m, so
x ∈ cone({a1 , . . . , am }).
Next, assume that x belongs to the polyhedral cone C. Consider a positive combination
x = λ1 a1 + · · · + λk ak , (∗1 )
for some nonzero a1 , . . . , ak ∈ C, with λi ≥ 0 and with k minimal . Since k is minimal, we

must have λi > 0 for i = 1, . . . , k. We claim that (a1 , . . . , ak ) are linearly independent.
If not, there is some nontrivial linear combination

µ1 a1 + · · · + µk ak = 0, (∗2 )
and since the ai are nonzero, µj 6= 0 for some at least some j. We may assume that µj < 0
for some j (otherwise, we consider the family (−µi )1≤i≤k ), so let
J = {j ∈ {1, . . . , k} | µj < 0}.
For any t ∈ R, since x = λ1 a1 + · · · + λk ak , using (∗2 ) we get
x = (λ1 + tµ1 )a1 + · · · + (λk + tµk )ak , (∗3 )
and if we pick
λj
t = min − ≥ 0,
j∈J µj
we have (λi + tµi ) ≥ 0 for i = 1, . . . , k, but λj + tµj = 0 for some j ∈ J, so (∗3 ) is an
expression of x with less that k nonzero coefficients, contadicting the minimality of k in (∗1 ).
Therefore, (a1 , . . . , ak ) are linearly independent.
Since a polyhedral cone C is spanned by finitely many vectors, there are finitely many
primitive cones (corresponding to linearly independent subfamilies), and since every x ∈ C,
belongs to some primitive cone, C is the union of a finite number of primitive cones. Since
every primitive cone is closed, as a union of finitely many closed sets, C itself is closed.
Lemma 4.14. (Farkas Lemma, Version I) Given any d × n real matrix A, and any vector
z ∈ Rd , exactly one of the following alternatives occurs:
(a) The linear system Ax = z has a solution x = (x1 , . . . , xn ), such that x ≥ 0 and
x1 + · · · + xn = 1, or
(b) There is some c ∈ Rd and some α ∈ R such that c> z < α and c> A ≥ α.
Proof. Let A1 , . . . , An ∈ Rd be the n points corresponding to the columns of A. Then, either
z ∈ conv({A1 , . . . , An }) or z ∈/ conv({A1 , . . . , An }). In the first case, we have a convex
combination
z = x1 A1 + · · · + xn An
where xi ≥ 0 and x1 + · · · + xn = 1, so x = (x1 , . . . , xn ) is a solution satisfying (a).
In the second case, by Corollary 4.10, there is a hyperplane, H, strictly separating {z}
and conv({A1 , . . . , An }), which is closed by Proposition 4.13. In fact, observe that z ∈
/
◦ ◦
conv({A1 , . . . , An }) iff there is a hyperplane, H, such that z ∈H − and Ai ∈ H+ , or z ∈H +
and Ai ∈ H− , for i = 1, . . . , n. As the affine hyperplane, H, is the zero locus of an equation
of the form
c1 y1 + · · · + cd yd = α,
either c> z < α and c> Ai ≥ α for i = 1, . . . , n, that is, c> A ≥ α, or c> z > α and c> A ≤ α.
In the second case, (−c)> z < −α and (−c)> A ≥ −α, so (b) is satisfied by either c and α or
by −c and −α.
Remark: If we relax the requirements on solutions of Ax = z and only require x ≥ 0

(x1 + · · · + xn = 1 is no longer required) then, in condition (b), we can take α = 0. This
is another version of Farkas Lemma. In this case, instead of considering the convex hull of
{A1 , . . . , An } we are considering the convex cone,
cone(A1 , . . . , An ) = {λA1 + · · · + λn An | λi ≥ 0, 1 ≤ i ≤ n},
that is, we are dropping the condition λ1 + · · · + λn = 1. For this version of Farkas Lemma
we need the following separation lemma:
Proposition 4.15. Let C ⊆ Ed be any closed convex cone with vertex O. Then, for every
point a not in C, there is a hyperplane H passing through O separating a and C with a ∈
/ H.
Proof. Since C is closed and convex and {a} is compact and convex, by Corollary 4.10, there
is a hyperplane, H 0 , strictly separating a and C. Let H be the hyperplane through O parallel
to H 0 . Since C and a lie in the two disjoint open half-spaces determined by H 0 , the point a
cannot belong to H. Suppose that some point, b ∈ C, lies in the open half-space determined
by H and a. Then, the line, L, through O and b intersects H 0 in some point, c, and as C
is a cone, the half line determined by O and b is contained in C. So, c ∈ C would belong
to H 0 , a contradiction. Therefore, C is contained in the closed half-space determined by H
that does not contain a, as claimed.
H! H
a
O C
Figure 4.9: Illustration for the proof of Proposition 4.15
Lemma 4.16. (Farkas Lemma, Version II) Given any d × n real matrix A and any vector
(a) The linear system Ax = z has a solution x such that x ≥ 0, or
(b) There is some c ∈ Rd such that c> z < 0 and c> A ≥ 0.
Proof. The proof is analogous to the proof of Lemma 4.14 except that it uses Proposition
4.15 instead of Corollary 4.10 and either z ∈ cone(A1 , . . . , An ) or z ∈
/ cone(A1 , . . . , An ).
One can show that Farkas II implies Farkas I. Here is another version of Farkas Lemma
having to do with a system of inequalities, Ax ≤ z. Although, this version may seem weaker
that Farkas II, it is actually equivalent to it!
Lemma 4.17. (Farkas Lemma, Version III) Given any d × n real matrix A and any vector
(a) The system of inequalities Ax ≤ z has a solution x, or
(b) There is some c ∈ Rd such that c ≥ 0, c> z < 0 and c> A = 0.

Proof. We use two tricks from linear programming:
1. We convert the system of inequalities, Ax ≤ z, into a system of equations by intro-
ducing a vector of “slack variables”, γ = (γ1 , . . . , γd ), where the system of equations
is
x
(A, I) = z,
γ
with γ ≥ 0.
2. We replace each “unconstrained variable”, xi , by xi = Xi − Yi , with Xi , Yi ≥ 0.

Then, the original system Ax ≤ z has a solution, x (unconstrained), iff the system of
equations  
X
(A, −A, I)  Y  = z
γ
has a solution with X, Y, γ ≥ 0. By Farkas II, this system has no solution iff there exists
some c ∈ Rd with c> z < 0 and
c> (A, −A, I) ≥ 0,
that is, c> A ≥ 0, −c> A ≥ 0, and c ≥ 0. However, these four conditions reduce to c> z < 0,
c> A = 0 and c ≥ 0.
These versions of Farkas lemma are statements of the form (P ∨ Q) ∧ ¬(P ∧ Q), which
is easily seen to be equivalent to ¬P ≡ Q, namely, the logical equivalence of ¬P and
Q. Therefore, Farkas-type lemmas can be interpreted as criteria for the unsolvablity of
various kinds of systems of linear equations or systems of linear inequalities, in the form of
a separation property.
For example, Farkas II (Lemma 4.16) says that a system of linear equations, Ax = z,
does not have any solution, x ≥ 0, iff there is some c ∈ Rd such that c> z < 0 and c> A ≥ 0.
This means that there is a hyperplane, H, of equation c> y = 0, such that the columns
vectors, Aj , forming the matrix A all lie in the positive closed half space, H+ , but z lies in
the interior of the other half space, H− , determined by H. Therefore, z can’t be in the cone
spanned by the Aj ’s.
Farkas III says that a system of linear inequalities, Ax ≤ z, does not have any solution
(at all) iff there is some c ∈ Rd such that c ≥ 0, c> z < 0 and c> A = 0. This time, there
is also a hyperplane of equation c> y = 0, with c ≥ 0, such that the columns vectors, Aj ,
forming the matrix A all lie in H but z lies in the interior of the half space, H− , determined
by H. In the “easy” direction, if there is such a vector c and some x satisfying Ax ≤ z, since
c ≥ 0, we get c> Ax ≤ c> z, but c> Ax = 0 and c> z < 0, a contradiction.
What is the crirerion for the insolvability of a system of inequalities Ax ≤ z with x ≥ 0?
This problem is equivalent to the insolvability of the set of inequalities

A z
x≤
−I 0
and by Farkas III, this system has no solution iff there is some vector, (c1 , c2 ), with (c1 , c2 ) ≥
0,
> > A > > z
(c1 , c2 ) = 0 and (c1 , c2 ) < 0.
−I 0
The above conditions are equivalent to c1 ≥ 0, c2 ≥ 0, c> > >
1 A − c2 = 0 and c1 z < 0, which
reduce to c1 ≥ 0, c> >
1 A ≥ 0 and c1 z < 0.
We can put all these versions together to prove the following version of Farkas lemma:
Lemma 4.18. (Farkas Lemma, Version IIIb) For any d × n real matrix A and any vector
z ∈ Rd , the following statements are equivalent:
(1) The system Ax = z has no solution x ≥ 0 iff there is some c ∈ Rd such that c> A ≥ 0
and c> z < 0.
(2) The system Ax ≤ z has no solution iff there is some c ∈ Rd such that c ≥ 0, c> A = 0
and c> z < 0.
(3) The system Ax ≤ z has no solution x ≥ 0 iff there is some c ∈ Rd such that c ≥ 0,
c> A ≥ 0 and c> z < 0.
Proof. We already proved that (1) implies (2) and that (2) implies (3). The proof that (3)
implies (1) is left as an easy exercise.
The reader might wonder what is the criterion for the unsolvability of a system Ax = z,
without any condition on x. However, since the unsolvability of the system Ax = b is
equivalent to the unsolvability of the system

A z
x≤ ,
−A −z
using (2), the above system is unsolvable iff there is some (c1 , c2 ) ≥ (0, 0) such that

> > A > > z
(c1 , c2 ) = 0 and (c1 , c2 ) < 0,
−A −z
4.2. SUPPORTING HYPERPLANES AND MINKOWSKI’S PROPOSITION 95
and these are equivalent to c> > > > > >
1 A − c2 A = 0 and c1 z − c2 z < 0, namely, c A = 0 and c z < 0
where c = c1 − c2 ∈ Rd . However, this simply says that c is orthogonal to the columns
A1 , . . . , An of A and that z is not orthogonal to c, so z cannot belong to the column space
of A, a criterion that we already knew from linear algebra.
As in Matousek and Gartner [41], we can summarize these various criteria in the following
table:
The system The system
Ax ≤ z Ax = z
has no solution ∃c ∈ Rd , such that c ≥ 0, ∃c ∈ Rd , such that
x ≥ 0 iff c> A ≥ 0 and c> z < 0 c> A ≥ 0 and c> z < 0
has no solution ∃c ∈ Rd , such that, c ≥ 0, ∃c ∈ Rd , such that
x ∈ Rn iff c> A = 0 and c> z < 0 c> A = 0 and c> z < 0
Remark: The strong duality theorem in linear programming can be proved using Lemma
4.18(c).
4.2 Supporting Hyperplanes and Minkowski’s Propo-

sition
Recall the definition of a supporting hyperplane given in Definition 3.5. We have the following
important proposition first proved by Minkowski (1896):
Proposition 4.19. (Minkowski) Let A be a nonempty, closed, and convex subset. Then, for
every point a ∈ ∂A, there is a supporting hyperplane to A through a.
Proof. Let d = dim A. If d < dim X (i.e., A has empty interior), then A is contained in some
affine subspace V of dimension d < dim X, and any hyperplane containing V is a supporting
◦
hyperplane for every a ∈ A. Now, assume d = dim X, so that A 6= ∅. If a ∈ ∂A, then
◦ ◦
{a} ∩ A = ∅. By Theorem 4.6, there is a hyperplane H separating A and L = {a}. However,
◦
by Corollary 4.2, since A 6= ∅ and A is closed, we have
◦
A = A = A.
◦ ◦
Now, the half–space containing A is closed, and thus, it contains A = A. Therefore, H
separates A and {a}.
Remark: The assumption that A is closed is convenient but unnecessary. Indeed, the proof
of Proposition 4.19 shows that the proposition holds for every boundary point, a ∈ ∂A
(assuming ∂A 6= ∅).
◦
Beware that Proposition 4.19 is false when the dimension of X is infinite and when A= ∅.
The proposition below gives a sufficient condition for a closed subset to be convex.
Proposition 4.20. Let A be a closed subset with nonempty interior. If there is a supporting
hyperplane for every point a ∈ ∂A, then A is convex.
Proof. We leave it as an exercise (see Berger [8], Proposition 11.5.4).

The condition that A has nonempty interior is crucial!
The proposition below characterizes closed convex sets in terms of (closed) half–spaces.
It is another intuitive fact whose rigorous proof is nontrivial.
Proposition 4.21. Let A be a nonempty closed and convex subset. Then, A is the intersec-
tion of all the closed half–spaces containing it.
Proof. Let A0 be the intersection of all the closed half–spaces containing A. It is immediately
checked that A0 is closed and convex and that A ⊆ A0 . Assume that A0 6= A, and pick
a ∈ A0 − A. Then, we can apply Corollary 4.10 to {a} and A and we find a hyperplane,
H, strictly separating A and {a}; this shows that A belongs to one of the two half-spaces
determined by H, yet a does not belong to the same half-space, contradicting the definition
of A0 .
4.3 Polarity and Duality

Let E = En be the Euclidean affine space of dimension n. We will denote the origin (0, . . . , 0)
in En by O. We know that the inner product on E = En induces a duality between E and
its dual E ∗ (for example, see Chapter 6, Section 2 of Gallier [30]), namely, u 7→ ϕu , where
ϕu is the linear form defined by ϕu (v) = u · v, for all v ∈ E. For geometric purposes, it is
more convenient to recast this duality as a correspondence between points and hyperplanes,
using the notion of polarity with respect to the unit sphere, S n−1 = {a ∈ En | kOak = 1}.
First, we need the following simple fact: For every hyperplane H not passing through O,
there is a unique point h, so that
H = {a ∈ En | Oh · Oa = 1}.
Indeed, any hyperplane H in En is the null set of some equation of the form
α1 x1 + · · · + αn xn = β,
and if O ∈
/ H, then β 6= 0. Thus, any hyperplane H not passing through O is defined by an
equation of the form
h1 x1 + · · · + hn xn = 1,
4.3. POLARITY AND DUALITY 97
if we set hi = αi /β. So, if we let h = (h1 , . . . , hn ), we see that
H = {a ∈ En | Oh · Oa = 1},
as claimed. Now, assume that
H = {a ∈ En | Oh1 · Oa = 1} = {a ∈ En | Oh2 · Oa = 1}.
The functions a 7→ Oh1 · Oa − 1 and a 7→ Oh2 · Oa − 1 are two affine forms defining the
same hyperplane, so there is a nonzero scalar λ so that
Oh1 · Oa − 1 = λ(Oh2 · Oa − 1) for all a ∈ En
(see Gallier [30], Chapter 2, Section 2.10). In particular, for a = O, we find that λ = 1, and
so,
Oh1 · Oa = Oh2 · Oa for all a,
which implies h1 = h2 . This proves the uniqueness of h.
Using the above, we make the following definition:
Definition 4.2. Given any point a 6= O, the polar hyperplane of a (w.r.t. S n−1 ) or dual of
a is the hyperplane a† given by
a† = {b ∈ En | Oa · Ob = 1}.
Given a hyperplane H not containing O, the pole of H (w.r.t S n−1 ) or dual of H is the
(unique) point H † so that
H = {a ∈ En | OH† · Oa = 1}.
We often abbreviate polar hyperplane to polar. We immediately check that a†† = a

and H †† = H, so we obtain a bijective correspondence between En − {O} and the set of
hyperplanes not passing through O.
When a is outside the sphere S n−1 , there is a nice geometric interpetation for the polar
hyperplane H = a† . Indeed, in this case, since
H = a† = {b ∈ En | Oa · Ob = 1}
and kOak > 1, the hyperplane H intersects S n−1 (along an (n − 2)-dimensional sphere)
and if b is any point on H ∩ S n−1 , we claim that Ob and ba are orthogonal. This means
that H ∩ S n−1 is the set of points on S n−1 where the lines through a and tangent to S n−1
touch S n−1 (they form a cone tangent to S n−1 with apex a). Indeed, as Oa = Ob + ba and
b ∈ H ∩ S n−1 i.e., Oa · Ob = 1 and kObk2 = 1, we get
1 = Oa · Ob = (Ob + ba) · Ob = kObk2 + ba · Ob = 1 + ba · Ob,

O a
a†
Figure 4.10: The polar, a† , of a point, a, outside the sphere S n−1
which implies ba · Ob = 0. When a ∈ S n−1 , the hyperplane a† is tangent to S n−1 at a.

Also, observe that for any point a 6= O, and any hyperplane H not passing through O, if
a ∈ H, then, H † ∈ a† , i.e, the pole H † of H belongs to the polar a† of a. Indeed, H † is the
unique point so that
H = {b ∈ En | OH† · Ob = 1}
and
a† = {b ∈ En | Oa · Ob = 1};
since a ∈ H, we have OH† · Oa = 1, which shows that H † ∈ a† .
If a = (a1 , . . . , an ), the equation of the polar hyperplane a† is
a1 X1 + · · · + an Xn = 1.
Remark: As we noted, polarity in a Euclidean space suffers from the minor defect that the
polar of the origin is undefined and, similarly, the pole of a hyperplane through the origin
does not make sense. If we embed En into the projective space, Pn , by adding a “hyperplane
at infinity” (a copy of Pn−1 ), thereby viewing Pn as the disjoint union Pn = En ∪ Pn−1 , then
the polarity correspondence can be defined everywhere. Indeed, the polar of the origin is the
hyperplane at infinity (Pn−1 ) and since Pn−1 can be viewed as the set of hyperplanes through
the origin in En , the pole of a hyperplane through the origin is the corresponding “point at
infinity” in Pn−1 .
Now, we would like to extend this correspondence to subsets of En , in particular, to
convex sets. Given a hyperplane, H, not containing O, we denote by H− the closed half-
space containing O.
Definition 4.3. Given any subset A of En , the set

\
A∗ = {b ∈ En | Oa · Ob ≤ 1, for all a ∈ A} = (a† )− ,
a∈A
a6=O
is called the polar dual or reciprocal of A.
For simplicity of notation, we write a†− for (a† )− . Observe that {O}∗ = En , so it is
†
convenient to set O− = En , even though O† is undefined. By definition, A∗ is convex even if
A is not. Furthermore, note that
(1) A ⊆ A∗∗ .
(2) If A ⊆ B, then B ∗ ⊆ A∗ .
(3) If A is convex and closed, then A∗ = (∂A)∗ .
It follows immediately from (1) and (2) that A∗∗∗ = A∗ . Also, if B n (r) is the (closed)
ball of radius r > 0 and center O, it is obvious by definition that B n (r)∗ = B n (1/r).
In Figure 4.11, the polar dual of the polygon (v1 , v2 , v3 , v4 , v5 ) is the polygon shown in
green. This polygon is cut out by the half-planes determined by the polars of the vertices
(v1 , v2 , v3 , v4 , v5 ) and containing the center of the circle. These polar lines are all easy to
determine by drawing for each vertex, vi , the tangent lines to the circle and joining the
contact points. The construction of the polar of v3 is shown in detail.
Remark: We chose a different notation for polar hyperplanes and polars (a† and H † ) and
polar duals (A∗ ), to avoid the potential confusion between H † and H ∗ , where H is a hy-
perplane (or a† and {a}∗ , where a is a point). Indeed, they are completely different! For
example, the polar dual of a hyperplane is either a line orthogonal to H through O, if O ∈ H,
or a semi-infinite line through O and orthogonal to H whose endpoint is the pole, H † , of H,
whereas, H † is a single point! Ziegler ([67], Chapter 2) use the notation A4 instead of A∗
for the polar dual of A.
We would like to investigate the duality induced by the operation A 7→ A∗ . Unfortunately,
it is not always the case that A∗∗ = A, but this is true when A is closed and convex, as
shown in the following proposition:
Proposition 4.22. Let A be any subset of En (with origin O).

◦ ◦
(i) If A is bounded, then O ∈ A∗ ; if O ∈ A, then A∗ is bounded.
(ii) If A is a closed and convex subset containing O, then A∗∗ = A.

v3
v2 v4
v1 v5
Figure 4.11: The polar dual of a polygon
Proof. (i) If A is bounded, then A ⊆ B n (r) for some r > 0 large enough. Then,
◦ ◦
B n (r)∗ = B n (1/r) ⊆ A∗ , so that O ∈ A∗ . If O ∈ A, then B n (r) ⊆ A for some r small enough,
so A∗ ⊆ B n (r)∗ = B r (1/r) and A∗ is bounded.
(ii) We always have A ⊆ A∗∗ . We prove that if b ∈ / A∗∗ ; this shows that
/ A, then b ∈
∗∗ ∗∗
A ⊆ A and thus, A = A . Since A is closed and convex and {b} is compact (and convex!),
by Corollary 4.10, there is a hyperplane, H, strictly separating A and b and, in particular,
O∈/ H, as O ∈ A. If h = H † is the pole of H, we have
Oh · Ob > 1 and Oh · Oa < 1, for all a ∈ A
/ A∗∗ , since
since H− = {a ∈ En | Oh · Oa ≤ 1}. This shows that b ∈
A∗∗ = {c ∈ En | Od · Oc ≤ 1 for all d ∈ A∗ }

= {c ∈ En | (∀d ∈ En )(if Od · Oa ≤ 1 for all a ∈ A, then Od · Oc ≤ 1)},
just let c = b and d = h.
Remark: For an arbitrary subset A ⊆ En , it can be shown that A∗∗ = conv(A ∪ {O}), the
topological closure of the convex hull of A ∪ {O}.
Proposition 4.22 will play a key role in studying polytopes, but before doing this, we
need one more proposition.
◦
Proposition 4.23. Let A be any closed convex subset of En such that O ∈ A. The polar
hyperplanes of the points of the boundary of A constitute the set of supporting hyperplanes of
A∗ . Furthermore, for any a ∈ ∂A, the points of A∗ where H = a† is a supporting hyperplane
of A∗ are the poles of supporting hyperplanes of A at a.
◦
Proof. Since O ∈ A, we have O ∈ / ∂A, and so, for every a ∈ ∂A, the polar hyperplane a†
is well-defined. Pick any a ∈ ∂A and let H = a† be its polar hyperplane. By definition,
A∗ ⊆ H− , the closed half-space determined by H and containing O. If T is any supporting
hyperplane to A at a, as a ∈ T , we have t = T † ∈ a† = H. Furthermore, it is a simple
exercise to prove that t ∈ (T− )∗ (in fact, (T− )∗ is the interval with endpoints O and t). Since
A ⊆ T− (because T is a supporting hyperplane to A at a), we deduce that t ∈ A∗ , and thus,
H is a supporting hyperplane to A∗ at t. By Proposition 4.22, as A is closed and convex,
A∗∗ = A; it follows that all supporting hyperplanes to A∗ are indeed obtained this way.
Chapter 5
Polyhedra and Polytopes
5.1 Polyhedra, H-Polytopes and V-Polytopes

There are two natural ways to define a convex polyhedron, A:
(1) As the convex hull of a finite set of points.
(2) As a subset of En cut out by a finite number of hyperplanes, more precisely, as the
intersection of a finite number of (closed) half-spaces.
As stated, these two definitions are not equivalent because (1) implies that a polyhedron
is bounded, whereas (2) allows unbounded subsets. Now, if we require in (2) that the convex
set A is bounded, it is quite clear for n = 2 that the two definitions (1) and (2) are equivalent;
for n = 3, it is intuitively clear that definitions (1) and (2) are still equivalent, but proving
this equivalence rigorously does not appear to be that easy. What about the equivalence
when n ≥ 4?
It turns out that definitions (1) and (2) are equivalent for all n, but this is a nontrivial
theorem and a rigorous proof does not come by so cheaply. Fortunately, since we have
Krein and Milman’s theorem at our disposal and polar duality, we can give a rather short
proof. The hard direction of the equivalence consists in proving that Definition (1) implies
Definition (2). This is where the duality induced by polarity becomes handy, especially, the
fact that A∗∗ = A! (under the right hypotheses). First, we give precise definitions (following
Ziegler [67]).
1
Definition 5.1. Let E be any affine Euclidean spaceTp of finite dimension, n. An H-polyhedron
in E, for short, a polyhedron, is any subset, P = i=1 Ci , of E defined as the intersection of a
finite number, p ≥ 1, of closed half-spaces, Ci ; an H-polytope in E is a bounded polyhedron
and a V-polytope is the convex hull, P = conv(S), of a finite set of points, S ⊆ E.
1 →
−
This means that the vector space, E , associated with E is a Euclidean space.
103
1
104 CHAPTER 5. POLYHEDRA AND POLYTOPES
(a) (b)
Figure 5.1: (a) An H-polyhedron. (b) A V-polytope
Obviously, H-polyhedra are convex and closed as intersections of convex and closed half-
spaces (in E). By Proposition 3.3, V-polytopes are also convex and closed (in E). Since the
notions of H-polytope and V-polytope are equivalent (see Theorem 5.7), we often use the
simpler locution polytope. Examples of an H-polyhedron and of a V-polytope are shown in
Figure 5.1.
Note that Definition 5.1 allows H-polytopes and V-polytopes to have an empty interior,
which is somewhat of an inconvenience. This is not a problem, since we may always restrict
ourselves to the affine hull of P (some affine space, E, of dimension d ≤ n, where d = dim(P ),
as in Definition 3.2) as we now show.
Proposition 5.1. Let A ⊆ E be a V-polytope or an H-polyhedron, let E = aff(A) be the

affine hull of A in E (with the Euclidean structure on E induced by the Euclidean structure
on E) and write d = dim(E). Then, the following assertions hold:
(1) The set, A, is a V-polytope in E (i.e., viewed as a subset of E) iff A is a V-polytope

in E.
(2) The set, A, is an H-polyhedron in E (i.e., viewed as a subset of E) iff A is an H-

polyhedron in E.
Proof. (1) This follows immediately because E is an affine subspace of E and every affine sub-
space of E is closed under affine combinations and so, a fortiori , under convex combinations.
We leave the details as an easy exercise.
(2) Assume A is an H-polyhedron in E and that d < n. By definition, A = pi=1 Ci , where
T
the Ci are closed half-spaces determined by some hyperplanes, H1 , . . . , Hp , in E. (Observe
that the hyperplanes, Hi ’s, associated with the closed half-spaces, Ci , may not be distinct.
5.1. POLYHEDRA, H-POLYTOPES AND V-POLYTOPES 105
For example, we may have Ci = (Hi )+ and Cj = (Hi )− , for the two closed half-spaces
determined by Hi .) As A ⊆ E, we have
p
\
A=A∩E = (Ci ∩ E),
i=1
where Ci ∩ E is one of the closed half-spaces determined by the hyperplane, Hi0 = Hi ∩ E,

in E. Thus, A is also an H-polyhedron in E.
Conversely, assume that A is an H-polyhedron in E and that d < n. As any hyperplane,
H, in E can be written as the intersection, H = H− ∩ H+ , of the two closed half-spaces that
it bounds, E itself can be written as the intersection,
p p
\ \
E= Ei = (Ei )+ ∩ (Ei )− ,
i=1 i=1
of finitely many half-spaces in E. Now, as A is an H-polyhedron in E, we have

q
\
A= Cj ,
j=1
where the Cj are closed half-spaces in E determined by some hyperplanes, Hj , in E. However,

each Hj can be extended to a hyperplane, Hj0 , in E, and so, each Cj can be extended to a
closed half-space, Cj0 , in E and we still have
q
\
A= Cj0 .
j=1
Consequently, we get
p q
\ \
A=A∩E = ((Ei )+ ∩ (Ei )− ) ∩ Cj0 ,
i=1 j=1
which proves that A is also an H-polyhedron in E.
The following simple proposition shows that we may assume that E = En :
Proposition 5.2. Given any two affine Euclidean spaces, E and F , if h : E → F is any
affine map then:
(1) If A is any V-polytope in E, then h(E) is a V-polytope in F .
(2) If h is bijective and A is any H-polyhedron in E, then h(E) is an H-polyhedron in F .

Proof. (1) As any affine map preserves affine combinations it also preserves convex combi-
nation. Thus, h(conv(S)) = conv(h(S)), for any S ⊆ E.
(2) Say A = pi=1 Ci in E. Consider any half-space, C, in E and assume that
T
C = {x ∈ E | ϕ(x) ≤ 0},
for some affine form, ϕ, defining the hyperplane, H = {x ∈ E | ϕ(x) = 0}. Then, as h is
bijective, we get
h(C) = {h(x) ∈ F | ϕ(x) ≤ 0}

= {y ∈ F | ϕ(h−1 (y)) ≤ 0}
= {y ∈ F | (ϕ ◦ h−1 )(y) ≤ 0}.
This shows that h(C) is one of the closed half-spaces in F determined by the hyperplane,
H 0 = {y ∈ F | (ϕ ◦ h−1 )(y) = 0}. Furthermore, as h is bijective, it preserves intersections so
p p
!
\ \
h(A) = h Ci = h(Ci ),
i=1 i=1
a finite intersection of closed half-spaces. Therefore, h(A) is an H-polyhedron in F .
By Proposition 5.2 we may assume that E = Ed and by Proposition 5.1 we may assume
that dim(A) = d. These propositions justify the type of argument beginning with: “We may
assume that A ⊆ Ed has dimension d, that is, that A has nonempty interior.” This kind of
reasonning will occur many times.
Since the boundary of a closed half-space, Ci , is a hyperplane, Hi , and since hyperplanes
are defined by affine forms, a closed half-space is defined by the locus of points satisfying a
“linear” inequality of the form ai · x ≤ bi or ai · x ≥ bi , for some vector ai ∈ Rn and some
bi ∈ R. Since ai · x ≥ bi is equivalent to (−ai ) · x ≤ −bi , we may restrict our attention
to inequalities with a ≤ sign. Thus, if A is the p × n matrix whose ith row is ai , we see
that the H-polyhedron, P , is defined by the system of linear inequalities, Ax ≤ b, where
b = (b1 , . . . , bp ) ∈ Rp . We write
P = P (A, b), with P (A, b) = {x ∈ Rn | Ax ≤ b}.
An equation, ai · x = bi , may be handled as the conjunction of the two inequalities ai · x ≤ bi

and (−ai ) · x ≤ −bi . Also, if 0 ∈ P , observe that we must have bi ≥ 0 for i = 1, . . . , p. In
this case, every inequality for which bi > 0 can be normalized by dividing both sides by bi ,
so we may assume that bi = 1 or bi = 0. This observation will be useful to show that the
polar dual of an H-polyhedron is a V-polyhedron.
Remark: Some authors call “convex” polyhedra and “convex” polytopes what we have
simply called polyhedra and polytopes. Since Definition 5.1 implies that these objects are
Figure 5.2: Example of a polytope (a dodecahedron)
convex and since we are not going to consider non-convex polyhedra in this chapter, we stick
to the simpler terminology.
One should consult Ziegler [67], Berger [8], Grunbaum [35] and especially Cromwell [22],
for pictures of polyhedra and polytopes. Figure 5.2 shows the picture a polytope whose faces
are all pentagons. This polytope is called a dodecahedron. The dodecahedron has 12 faces,
30 edges and 20 vertices.
Even better and a lot more entertaining, take a look at the spectacular web sites of
George Hart,
Virtual Polyedra: http://www.georgehart.com/virtual-polyhedra/vp.html,
George Hart’s web site: http://www.georgehart.com/
and also
Zvi Har’El ’s web site: http://www.math.technion.ac.il/ rl/
The Uniform Polyhedra web site: http://www.mathconsult.ch/showroom/unipoly/
Paper Models of Polyhedra: http://www.korthalsaltes.com/
Bulatov’s Polyhedra Collection: http://www.physics.orst.edu/ bulatov/polyhedra/
Paul Getty’s Polyhedral Solids: http://home.teleport.com/ tpgettys/poly.shtml
Jill Britton’s Polyhedra Pastimes: http://ccins.camosun.bc.ca/ jbritton/jbpolyhedra.htm
and many other web sites dealing with polyhedra in one way or another by searching for
“polyhedra” on Google!
Obviously, an n-simplex is a V-polytope. The standard n-cube is the set
{(x1 , . . . , xn ) ∈ En | |xi | ≤ 1, 1 ≤ i ≤ n}.
The standard cube is a V-polytope. The standard n-cross-polytope (or n-co-cube) is the set
n
X
n
{(x1 , . . . , xn ) ∈ E | |xi | ≤ 1}.
i=1
It is also a V-polytope.
In order to prove that an H-polytope is a V-polytope we need to take a closer look at

polyhedra. Basically, we need to make precise the notion of vertex, edge, and face, which
are intuitively clear in dimension three.
Note
Tt that some of the hyperplanes cutting out a polyhedron may be redundant. If
A = i=1 Ci is a polyhedron (where each closedTt half-space, Ci , is associated with a hyper-
plane, Hi , so that ∂Ci = Hi ), we
T say that i=1 Ci is an irredundant decomposition of A if
A cannot be expressed as A = m C
i=1 i
0
with m < t (for some closed half-spaces, Ci0 ). The
following proposition shows that the Ci in an irredundant decomposition of A are uniquely
determined by A.
Proposition 5.3. Let A be a polyhedron with nonempty interior and assume that
A = ti=1 Ci is an irredundant decomposition of A. Then,
T
(i) Up to order, the Ci ’s are uniquely determined by A.
(ii) If Hi = ∂Ci is the boundary of Ci , then Hi ∩ A is a polyhedron with nonempty interior

in Hi , denoted Faceti A, and called a facet of A.
(iii) We have ∂A = ti=1 Faceti A, where the union is irredundant, i.e., Faceti A is not a
S
subset of Facetj A, for all i 6= j.
T Tt
Proof. (ii) Fix any i and consider Ai = j6=i Cj . As A = i=1 Ci is an irredundant decompo-
◦ ◦
sition, there is some x ∈ Ai − Ci . Pick any a ∈ A. By Lemma 4.1, we get b = [a, x] ∩ Hi ∈ Ai ,
so b belongs to the interior of Hi ∩ Ai in Hi .
◦ ◦
(iii) As ∂A = A − A = A ∩ (A)c (where B c denotes the complement of a subset B of En )
and ∂Ci = Hi , we get

! ◦ !
t
\ t
\
∂A = Ci − Cj
i=1 j=1
t
! t
!
\ \ ◦
= Ci − Cj
i=1 j=1
t
! t
!c
\ \ ◦
= Ci ∩ Cj
i=1 j=1
t
! t
!
\ [ ◦
= Ci ∩ (C j )c
i=1 j=1
t t
!
[ \ ◦
= Ci ∩ (C j )c
j=1 i=1
t
!
[ \
= ∂Cj ∩ Ci
j=1 i6=j
t
[ t
[
= (Hj ∩ A) = Facetj A.
j=1 j=1
If we had Faceti A ⊆ Facetj A, for some i 6= j, then, by (ii), as the affine hull of Faceti A is
Hi and the affine hull of Facetj A is Hj , we would have Hi ⊆ Hj , a contradiction.
(i) As the decomposition is irredundant, the Hi are pairwise distinct. Also, by (ii), each
facet, Faceti A, has dimension d − 1 (where d = dim A). Then, in (iii), we can show that the
decomposition of ∂A as a union of polytopes of dimension d − 1 whose pairwise nonempty
intersections have dimension at most d − 2 (since they are contained in pairwise distinct
hyperplanes) is unique up to permutation. Indeed, assume that
∂A = F1 ∪ · · · ∪ Fm = G1 ∪ · · · ∪ Gn ,
where the Fi ’s and G0j are polyhedra of dimension d−1 and each of the unions is irredundant.
Then, we claim that for each Fi , there is some Gϕ(i) such that Fi ⊆ Gϕ(i) . If not, Fi would
be expressed as a union
Fi = (Fi ∩ Gi1 ) ∪ · · · ∪ (Fi ∩ Gik )
where dim(Fi ∩ Gij ) ≤ d − 2, since the hyperplanes containing Fi and the Gj ’s are pairwise
distinct, which is absurd, since dim(Fi ) = d − 1. By symmetry, for each Gj , there is some
Fψ(j) such that Gj ⊆ Fψ(j) . But then, Fi ⊆ Fψ(ϕ(i)) for all i and Gj ⊆ Gϕ(ψ(j)) for all j which
implies ψ(ϕ(i)) = i for all i and ϕ(ψ(j)) = j for all j since the unions are irredundant. Thus,
ϕ and ψ are mutual inverses and the Bj ’s are just a permutation of the Ai ’s, as claimed.
Therefore, the facets, Faceti A, are uniquely determined by A and so are the hyperplanes,
Hi = aff(Faceti A), and the half-spaces, Ci , that they determine.
As a consequence, if A is a polyhedron, then so are its facets and the same holds for
◦
H-polytopes. If A is an H-polytope and H is a hyperplane with H ∩ A 6= ∅, then H ∩ A is
an H-polytope whose facets are of the form H ∩ F , where F is a facet of A.
We can use induction and define k-faces, for 0 ≤ k ≤ n − 1.
Definition 5.2. Let A ⊆ En be a polyhedron with nonempty interior. We define a k-face
of A to be a facet of a (k + 1)-face of A, for k = 0, . . . , n − 2, where an (n − 1)-face is just
a facet of A. The 1-faces are called edges. Two k-faces are adjacent if their intersection is a
(k − 1)-face.
The polyhedron A itself is also called a faceT(of itself) or n-face and the k-faces of A with
k ≤ n − 1 are called proper faces of A. If A = ti=1 Ci is an irredundant decomposition of A
and Hi is the boundary of Ci , then the hyperplane, Hi , is called the supporting hyperplane
of the facet Hi ∩ A. We suspect that the 0-faces of a polyhedron are vertices in the sense
of Definition 3.6. This is true and, in fact, the vertices of a polyhedron coincide with its
extreme points (see Definition 3.7).
Proposition 5.4. Let A ⊆ En be a polyhedron with nonempty interior.
(1) For any point, a ∈ ∂A, on the boundary of A, the intersection of all the supporting
hyperplanes to A at a coincides with the intersection of all the faces that contain a. In
particular, points of order k of A are those points in the relative interior of the k-faces
of A2 ; thus, 0-faces coincide with the vertices of A.
(2) The vertices of A coincide with the extreme points of A.
Proof. (1) If H is a supporting hyperplane to A at a, then, one of the half-spaces, C,
determined by H, satisfies A = A ∩ C. It follows from Proposition 5.3 that if H 6= Hi (where
the hyperplanes Hi are the supporting hyperplanes of the facets of A), then C is redundant,
from which (1) follows.
(2) If a ∈ ∂A is not extreme, then a ∈ [y, z], where y, z ∈ ∂A. However, this implies that
a has order k ≥ 1, i.e, a is not a vertex.
The proof that every H-polytope A is a V-polytope relies on the fact that the extreme
points of an H-polytope coincide with its vertices, which form a finite nonempty set, and by
Krein and Millman’s Theorem (Theorem 3.8), A is the convex hull of its vertices.
The proof that every V-polytope A is an H-polytope relies on the crucial fact that
the polar dual A∗ of a V-polytope A is an H-polyhedron, and that the equations of the
2
Given a convex set, S, in An , its relative interior is its interior in the affine hull of S (which might be
of dimension strictly less than n).
5.2. POLAR DUALS OF V-POLYTOPES AND H-POLYHEDRA 111
hyperplanes cutting out A∗ are obtained in a very simple manner from the points ai specifying
A as A = conv(a1 , . . . , ap ); see Proposition 5.5.
The proof that every V-polytope A is an H-polytope consists of the following steps:
(1) Construct the polar dual A∗ of A. Then we know that A∗ is is an H-polyhedron.

Furthermore, if the center O of the polar duality belongs to the interior of A, then A∗
is bounded, and so it is an H-polytope.
(2) Since A∗ is H-polytope, A∗ is also a V-polytope, as we claimed earlier.
(3) Since A∗ is a V-polytope, its polar dual A∗∗ is an H-polyhedron.
(4) A V-polytope is closed and convex, and since O belong to A (in fact, to the interior of
A), by Proposition 4.22, we have A = A∗∗ , so A is indeed an H-polyhedron; in fact, A
an H-polytope since it is bounded.
5.2 Polar Duals of V-Polytopes and H-Polyhedra of the

Form P (A, 1)
The following proposition gives a simple description of the polar dual A∗ of a V-polyhedron
A = conv(v1 , . . . , vp ) in terms of the hyperplanes a†i . It is a key ingredient in the proof of
the equivalence of V-polytopes and H-polytopes because it implies that if the center of polar
duality is chosen in the interior of a V-polytope, then its poar dual is an H-polytope.
Proposition 5.5. Let S = {ai }pi=1 be a finite set of points in En and let A = conv(S) be
its convex hull. If S 6= {O}, then, the dual A∗ of A w.r.t. the center O is the H-polyhedron
given by
p
(a†i )− .
\
∗
A =
i=1
◦
Furthermore, if O ∈ A, then A∗ is an H-polytope, i.e., the dual of a V-polytope with nonempty
interior is an H-polytope. If A = S = {O}, then A∗ = Ed .
Proof. By definition, we have

p p
X X
∗ n
A = {b ∈ E | Ob · ( λj Oaj ) ≤ 1, λj ≥ 0, λj = 1},
j=1 j=1
(a†i )− , which
Tp Tp
and the right hand side is clearly equal to i=1 {b ∈ En | Ob · Oai ≤ 1} = i=1
◦
is a polyhedron. (Recall that (a†i )− = En if ai = O.) If O ∈ A, then A∗ is bounded (by
Proposition 4.22) and so, A∗ is an H-polytope.
Thus, the dual of the convex hull of a finite set of points {a1 , . . . , ap } is the intersection
of the half-spaces containing O determined by the polar hyperplanes of the points ai .
It is convenient to restate Proposition 5.5 using matrices. First, observe that the proof
of Proposition 5.5 shows that
conv({a1 , . . . , ap })∗ = conv({a1 , . . . , ap } ∪ {O})∗ .
Therefore, we may assume that not all ai = O (1 ≤ i ≤ p). If we pick O as an origin, then
every point aj can be identified with a vector in En and O corresponds to the zero vector,
0. Observe that any set of p points aj ∈ En corresponds to the n × p matrix A whose j th
column is aj . Then, the equation of the the polar hyperplane a†j of any aj (6= 0) is aj · x = 1,
that is
a>
j x = 1.
Consequently, the system of inequalities defining conv({a1 , . . . , ap })∗ can be written in matrix
form as
conv({a1 , . . . , ap })∗ = {x ∈ Rn | A> x ≤ 1},
where 1 denotes the vector of Rp with all coordinates equal to 1. We write
P (A> , 1) = {x ∈ Rn | A> x ≤ 1}. There is a useful converse of this property as proved in
the next proposition.
Proposition 5.6. Given any set of p points {a1 , . . . , ap } in Rn with {a1 , . . . , ap } =

6 {0}, if
th
A is the n × p matrix whose j column is aj , then
conv({a1 , . . . , ap })∗ = P (A> , 1),
with P (A> , 1) = {x ∈ Rn | A> x ≤ 1}.

Conversely, given any p × n matrix A not equal to the zero matrix, we have
P (A, 1)∗ = conv({a1 , . . . , ap } ∪ {0}),
where ai ∈ Rn is the ith row of A or, equivalently,
P (A, 1)∗ = {x ∈ Rn | x = A> t, t ∈ Rp , t ≥ 0, It = 1},
where I is the row vector of length p whose coordinates are all equal to 1.
Proof. Only the second part needs a proof. Let B = conv({a1 , . . . , ap } ∪ {0}), where ai ∈ Rn
is the ith row of A. Then, by the first part,
B ∗ = P (A, 1).
As 0 ∈ B and B is convex and closed (by Proposition 3.3), by Proposition 4.22, B = B ∗∗ =

P (A, 1)∗ , as claimed.
5.3. THE EQUIVALENCE OF H-POLYTOPES AND V-POLYTOPES 113
Remark: Proposition 5.6 still holds if A is the zero matrix because then, the inequalities
A> x ≤ 1 (or Ax ≤ 1) are trivially satisfied. In the first case, P (A> , 1) = Ed , and in the
second case, P (A, 1) = Ed .
Using the above, the reader should check that the dual of a simplex is a simplex and that
the dual of an n-cube is an n-cross polytope.
It is not clear that every H-polyhedron is of the form P (A, 1). This is indeed the case
if we pick O in the interior of A, but this is nontrivial to prove. What we will need is to
find the corresponding “V-definition” of an H-polyhedron. For this we will need to add
positive combinations of vectors to convex combinations of points. Intuitively, these vectors
correspond to “points at infinity.”
5.3 The Equivalence of H-Polytopes and V-Polytopes

We are now ready for the theorem showing the equivalence of V-polytopes and H-polytopes.
This is a nontrivial theorem usually attributed to Weyl and Minkowski (for example, see
Barvinok [4]).
Theorem 5.7. (Weyl-Minkowski) If A is an H-polytope, then A has a finite number of
extreme points (equal to its vertices) and A is the convex hull of its set of vertices; thus, an
H-polytope is a V-polytope. Moreover, A has a finite number of k-faces (for k = 0, . . . , d − 2,
where d = dim(A)). Conversely, the convex hull of a finite set of points is an H-polytope.
As a consequence, a V-polytope is an H-polytope.
Proof. By restricting ourselves to the affine hull of A (some Ed , with d ≤ n) we may assume
that A has nonempty interior. Since an H-polytope has finitely many facets, we deduce
by induction that an H-polytope has a finite number of k-faces, for k = 0, . . . , d − 2. In
particular, an H-polytope has finitely many vertices. By proposition 5.4, these vertices are
the extreme points of A and since an H-polytope is compact and convex, by the theorem of
Krein and Milman (Theorem 3.8), A is the convex hull of its set of vertices.
Conversely, again, we may assume that A has nonempty interior by restricting ourselves
to the affine hull of A. Then, pick an origin O in the interior of A and consider the dual
A∗ of A. By Proposition 5.5, the convex set A∗ is an H-polytope. By the first part of the
proof of Theorem 5.7, the H-polytope A∗ is the convex hull of its vertices. Finally, since A
is convex and closed (by Proposition 3.3), and since O is in the interior of A, Proposition
4.22 and Proposition 5.5 hold, and we deduce that A = A∗∗ is an H-polytope.
In view of Theorem 5.7, we are justified in dropping the V or H in front of polytope, and
will do so from now on. Theorem 5.7 has some interesting corollaries regarding the dual of
a polytope.
Corollary 5.8. If A is any polytope in En such that the interior of A contains the origin
O, then the dual A∗ of A is also a polytope whose interior contains O, and A∗∗ = A.
Corollary 5.9. If A is any polytope in En whose interior contains the origin O, then the
k-faces of A are in bijection with the (n − k − 1)-faces of the dual polytope A∗ . This corre-
spondence is as follows: If Y = aff(F ) is the k-dimensional subspace determining the k-face
F of A then the subspace Y ∗ = aff(F ∗ ) determining the corresponding face F ∗ of A∗ is the
intersection of the polar hyperplanes of points in Y .
Proof. Immediate from Proposition 5.4 and Proposition 4.23.
We also have the following proposition whose proof would not be that simple if we only
had the notion of an H-polytope (as a matter of fact, there is a way of proving Theorem 5.7
using Proposition 5.10).
Proposition 5.10. If A ⊆ En is a polytope and f : En → Em is an affine map, then f (A)

is a polytope in Em .
Proof. Immediate, since an H-polytope is a V-polytope and since affine maps send convex
sets to convex sets.
The reader should check that the Minkowski sum of polytopes is a polytope.
We were able to give a short proof of Theorem 5.7 because we relied on a powerful
theorem, namely, Krein and Milman. A drawback of this approach is that it bypasses the
interesting and important problem of designing algorithms for finding the vertices of an
H-polyhedron from the sets of inequalities defining it. A method for doing this is Fourier–
Motzkin elimination, see Proposition 5.21, and also Ziegler [67] (Chapter 1) and Section 5.4.
This is also a special case of linear programming.
It is also possible to generalize the notion of V-polytope to polyhedra using the notion
of cone and to generalize the equivalence theorem to H-polyhedra and V-polyhedra.
5.4 The Equivalence of H-Polyhedra and V-Polyhedra

The equivalence of H-polytopes and V-polytopes can be generalized to polyhedral sets, i.e.
finite intersections of closed half-spaces that are not necessarily bounded. This equivalence
was first proved by Motzkin in the early 1930’s. It can be proved in several ways, some
involving cones.
Definition 5.3. Let E be any affine Euclidean space of finite dimension n (with associated
→
− →
−
vector space E ). A subset C ⊆ E is a cone if C is closed under linear combinations involving
→
−
only nonnegative scalars called positive combinations. Given a subset, V ⊆ E , the conical
hull or positive hull of V is the set
nX o
cone(V ) = λi vi | {vi }i∈I ⊆ V, λi ≥ 0 for all i ∈ I .
I
5.4. THE EQUIVALENCE OF H-POLYHEDRA AND V-POLYHEDRA 115
A V-polyhedron or polyhedral set is a subset A ⊆ E such that
A = conv(Y ) + cone(V ) = {a + v | a ∈ conv(Y ), v ∈ cone(V )},

→
−
where V ⊆ E is a finite set of vectors and Y ⊆ E is a finite set of points.
→
−
A set C ⊆ E is a V-cone or polyhedral cone if C is the positive hull of a finite set of
vectors, that is,
C = cone({u1 , . . . , up }),
→
− →
−
for some vectors u1 , . . . , up ∈ E . An H-cone is any subset of E given by a finite intersection
of closed half-spaces cut out by hyperplanes through 0.
The positive hull cone(V ) of V is also denoted pos(V ). Observe that a V-cone can be
viewed as a polyhedral set for which Y = {O}, a single point. However, if we take the point
O as the origin, we may view a V-polyhedron A for which Y = {O} as a V-cone. We will
switch back and forth between these two views of cones as we find it convenient, this should
not cause any confusion. In this section, we favor the view that V-cones are special kinds
of V-polyhedra. As a consequence, a (V or H)-cone always contains 0, sometimes called an
apex of the cone.
→
−
A set of the form {a + tu | t ≥ 0}, where a ∈ E is a point and u ∈ E is a nonzero vector,
is called a half-line or ray. Then, we see that a V-polyhedron, A = conv(Y ) + cone(V ), is
the convex hull of the union of a finite set of points with a finite set of rays. In the case of
a V-cone, all these rays meet in a common point, an apex of the cone.
Since an H-polyhedron is an intersection of half-spaces determined by hyperplanes, and
since half-spaces are closed, an H-polyhedron is closed. We know from Proposition 3.3 that
a V-polytope is closed and by Proposition 4.13 that a V-cone is closed. To apply Proposition
4.22 to an arbitrary V-polyhedron we need to know that a V-polyhedron is closed.
Given a V-polyhedron P = conv(Y ) + cone(V ) of dimension d, an easy way to prove that
P is closed is to “lift” P to the hyperplane Hd+1 of equation xd+1 = 1 in Ad+1 , obtaining a
polyhedron Pb contained in Hd+1 homeomorphic to P , and to consider a polyhedral cone (a
V-cone) C(P ) associated with P which has the property that
Pb = C(P ) ∩ Hd+1 .
The details of this construction are given in Section 5.5; see Proposition 5.20(2). Since by
Proposition 4.13 a V-cone is closed and since a hyperplane is closed, Pb = C(P ) ∩ Hd+1 is
closed, and thus P is closed. As a summary, the following propostion holds.
Proposition 5.11. Every V-polyhedron P = conv(Y ) + cone(V ) is closed.
Propositions 5.1 and 5.2 generalize easily to V-polyhedra and cones.

Proposition 5.12. Let A ⊆ E be a V-polyhedron or an H-polyhedron, let E = aff(A) be the

affine hull of A in E (with the Euclidean structure on E induced by the Euclidean structure
on E) and write d = dim(E). Then, the following assertions hold:
(1) The set A is a V-polyhedron in E (i.e., viewed as a subset of E) iff A is a V-polyhedron

in E.
(2) The set A is an H-polyhedron in E (i.e., viewed as a subset of E) iff A is an H-

polyhedron in E.
→
−
Proof. We already proved (2) in Proposition 5.1. For (1), observe that the direction E of
→
− →
−
E is a linear subspace of E . Consequently, E is closed under affine combinations and E is
closed under linear combinations and the result follows immediately.
Proposition 5.13. Given any two affine Euclidean spaces E and F , if h : E → F is any
affine map then:
(1) If A is any V-polyhedron in E, then h(E) is a V-polyhedron in F .

→
− →
− →
−
(2) If g : E → F is any linear map and if C is any V-cone in E , then g(C) is a V-cone
→
−
in F .
(3) If h is bijective and A is any H-polyhedron in E, then h(E) is an H-polyhedron in F .
Proof. We already proved (3) in Proposition 5.2. For (1), using the fact that h(a + u) =
→
− →
−
h(a) + h (u) for any affine map, h, where h is the linear map associated with h, we get
→
−
h(conv(Y ) + cone(V )) = conv(h(Y )) + cone( h (V )).
For (2), as g is linear, we get
g(cone(V )) = cone(g(V )),
establishing the proposition.
Propositions 5.12 and 5.13 allow us to assume that E = Ed and that our (V or H)-
polyhedra, A ⊆ Ed , have nonempty interior (i.e. dim(A) = d).
The generalization of Theorem 5.7 is that every V-polyhedron A is an H-polyhedron and
conversely.
At first glance, it may seem that there is a small problem when A = Ed . Indeed, Definition
5.3 allows the possibility that cone(V ) = Ed for some finite subset, V ⊆ Rd . This is because
it is possible to generate a basis of Rd using finitely many positive combinations. On the
other hand the definition of an H-polyhedron, A, (Definition 5.1) assumes that A ⊆ En is
cut out by at least one hyperplane. So, A is always contained in some half-space of En and
strictly speaking, En is not an H-polyhedron! The simplest way to circumvent this difficulty
is to decree that En itself is a polyhedron, which we do.
Yet another solution is to assume that, unless stated otherwise, every finite set of vectors
V that we consider when defining a polyhedron has the property that there is some hyper-
plane H through the origin so that all the vectors in V lie in only one of the two closed
half-spaces determined by H. But then, the polar dual of a polyhedron can’t be a single
point! Therefore, we stick to our decision that En itself is a polyhedron.
To prove the equivalence of H-polyhedra and V-polyhedra, Ziegler proceeds as follows:
First, he shows that the equivalence of V-polyhedra and H-polyhedra reduces to the equiva-
lence of V-cones and H-cones using an “old trick” of projective geometry, namely, “homog-
enizing” [67] (Chapter 1). Then, he uses two dual versions of Fourier–Motzkin elimination
to pass from V-cones to H-cones and conversely. Since the homogenization method is an
important technique we will describe it in some detail later.
However, it turns out that the double dualization technique used in the proof of Theorem
5.7 can be easily adapted to prove that every V-polyhedron is an H-polyhedron. This is
because if O belongs to the interior of the V-polyhedron A, then its polar dual A∗ is an
H-polytope; see Proposition 5.14. Then, just as in the proof of Theorem 5.7, we can use the
theorem of Krein and Millman to show that A∗ is a V-polytope. By taking the polar dual
of A∗ , we obtain the fact that A∗∗ = A is an H-polyhedron.
Moreover, the dual of an H-polyhedron is a V-polyhedron; see Proposition 5.15. This fact
can be used to prove that every H-polyhedron is a V-polyhedron by using the fact already
shown that every V-polyhedron is an H-polyhedron!
Consequently we will not describe the version of Fourier–Motzkin elimination used to go
from V-cones to H-cones. However, we will present the Fourier–Motzkin elimination method
used to go from H-cones to V-cones; see Proposition 5.21.
The generalization of Proposition 5.5 to polyhedral sets is shown below. As before, the
center of our polar duality is denoted by O. It is taken as the origin of Ed . The new ingredient
is that because a V-polyhderon is defined by points and vectors, its polar dual is still cut
out by hyperplanes, but the hyperplanes corresponding to vectors pass through the origin.
To show this we need to define the “polar hyperplane” u† of a vector u.
Definition 5.4. Given any nonzero vector u ∈ Rd , let u†− be the closed half-space
u†− = {x ∈ Rd | x · u ≤ 0}.
In other words, u†− is the closed half-space bounded by the hyperplane u† through O normal
to u and on the “opposite side” of u.
The following proposition generalizing Proposition 5.5 is inspired by Exercise 7 in Chapter

3 of Grunbaum [35].
Proposition 5.14. Let A = conv(Y ) + cone(V ) ⊆ Ed be a V-polyhedron with Y = {y1 , . . .,

yp } and V = {v1 , . . . , vq }. Then, for any point O, if A 6= {O}, then the polar dual A∗ of A
w.r.t. O is the H-polyhedron given by
p q
(yi† )− (vj† )− .
\ \
∗
A = ∩
i=1 j=1
Furthermore, if A has nonempty interior and O belongs to the interior of A, then A∗ is

bounded, that is, A∗ is an H-polytope. If A = {O}, then A∗ is the special polyhedron
A∗ = Ed .
Proof. By definition of A∗ w.r.t. O, we have
( p q
! p
)
X X X
A∗ = x ∈ Ed Ox · O λi yi + µj vj ≤ 1, λi ≥ 0, λi = 1, µj ≥ 0

i=1 j=1 i=1
( p q p
)
X X X
= x ∈ Ed λi Ox · Oyi + µj Ox · vj ≤ 1, λi ≥ 0, λi = 1, µj ≥ 0 .

i=1 j=1 i=1
When µj = 0 for j = 1, . . . , q, we get

p p
X X
λi Ox · Oyi ≤ 1, λi ≥ 0, λi = 1
i=1 i=1
and we check that

( p p
) p
X X \
d
x∈E λi Ox · Oyi ≤ 1, λi ≥ 0, λi = 1 = {x ∈ Ed | Ox · Oyi ≤ 1}

i=1 i=1 i=1
p
(yi† )− .
\
=
i=1
The points in A∗ must also satisfy the conditions

q
X
µj Ox · vj ≤ 1 − α, µj ≥ 0, µj > 0 for some j, 1 ≤ j ≤ q,
j=1
with α ≤ 1 (here α = pi=1 λi Ox · Oyi ). In particular, for every j ∈ {1, . . . , q}, if we set
P
µk = 0 for k ∈ {1, . . . , q} − {j}, we should have
µj Ox · vj ≤ 1 − α for all µj > 0,
that is,
1−α
Ox · vj ≤ for all µj > 0,
µj
which is equivalent to
Ox · vj ≤ 0.
Consequently, if x ∈ A∗ , we must also have
q q
(vj† )− .
\ \
d
x∈ {x ∈ E | Ox · vj ≤ 0} =
j=1 j=1
Therefore,
p q
(yi† )− (vj† )− .
\ \
∗
A ⊆ ∩
i=1 j=1
However, the reverse inclusion is obvious and thus, we have equality. If O belongs to the
interior of A, we know from Proposition 4.22 that A∗ is bounded. Therefore, A∗ is indeed
an H-polytope of the above form.
It is fruitful to restate Proposition 5.14 in terms of matrices (as we did for Proposition
5.5). First, observe that
(conv(Y ) + cone(V ))∗ = (conv(Y ∪ {O}) + cone(V ))∗ .
If we pick O as an origin then we can represent the points in Y as vectors, and O is now
denoted 0. The zero vector is denoted 0.
If A = conv(Y ) + cone(V ) 6= {0}, let Y be the d × p matrix whose ith column is yi and
let V is the d × q matrix whose j th column is vj . Then Proposition 5.14 says that
(conv(Y ) + cone(V ))∗ = {x ∈ Rd | Y > x ≤ 1, V > x ≤ 0}.
We write P (Y > , 1; V > , 0) = {x ∈ Rd | Y > x ≤ 1, V > x ≤ 0}.

If A = conv(Y ) + cone(V ) = {0}, then both Y and V must be zero matrices but then,
the inequalities Y > x ≤ 1 and V > x ≤ 0 are trivially satisfied by all x ∈ Ed , so even in this
case we have
Ed = (conv(Y ) + cone(V ))∗ = P (Y > , 1; V > , 0).
The converse of Proposition 5.14 also holds as shown below.
Proposition 5.15. Let {y1 , . . . , yp } be any set of points in Ed and let {v1 , . . . , vq } be any
set of nonzero vectors in Rd . If Y is the d × p matrix whose ith column is yi and V is the
d × q matrix whose j th column is vj , then
(conv({y1 , . . . , yp }) + cone({v1 , . . . , vq }))∗ = P (Y > , 1; V > , 0),
with P (Y > , 1; V > , 0) = {x ∈ Rd | Y > x ≤ 1, V > x ≤ 0}.

Conversely, given any p × d matrix Y and any q × d matrix V , we have
P (Y, 1; V, 0)∗ = conv({y1 , . . . , yp } ∪ {0}) + cone({v1 , . . . , vq }),
where yi ∈ Rn is the ith row of Y and vj ∈ Rn is the j th row of V , or equivalently,
P (Y, 1; V, 0)∗ = {x ∈ Rd | x = Y > u + V > t, u ∈ Rp , t ∈ Rq , u, t ≥ 0, Iu = 1},
where I is the row vector of length p whose coordinates are all equal to 1.
Proof. Only the second part needs a proof. Let
B = conv({y1 , . . . , yp } ∪ {0}) + cone({v1 , . . . , vq }),
where yi ∈ Rp is the ith row of Y and vj ∈ Rq is the j th row of V . Then, by the first part,
B ∗ = P (Y, 1; V, 0).
As 0 ∈ B and B is closed (by Proposition 5.11) and convex, by Proposition 4.22, B = B ∗∗ =

P (Y, 1; V, 0)∗ , as claimed.
Proposition 5.15 has the following important corollary:

Proposition 5.16. The following assertions hold:
(1) The polar dual A∗ of every H-polyhedron is a V-polyhedron.
(2) The polar dual A∗ of every V-polyhedron is an H-polyhedron.

Proof. (1) We may assume that 0 ∈ A, in which case A is of the form A = P (Y, 1; V, 0). By
the second part of Proposition 5.15, A∗ is a V-polyhedron.
(2) This is the first part of Proposition 5.15.
We can now use Proposition 5.14, Proposition 4.22, and Krein and Milman’s Theorem
to prove that every V-polyhedron is an H-polyhedron.
Proposition 5.17. If A 6= Ed is a V-polyhedron and if we choose the center of the polarity
◦
O in the interior A of A, then A is of the form A = P (Y, 1). Therefore, every V-polyhedron
is an H-polyhedron.
Proof. Let A 6= Ed be a V-polyhedron of dimension d. Thus A ⊆ Ed has nonempty interior
so we can pick some point O in the interior of A. If d = 0, then A = {0} = E0 and we are
done. Otherwise, by Proposition 5.14, the polar dual A∗ of A w.r.t. O is an H-polytope.
Then, as in the proof of Theorem 5.7, using Krein and Milman’s Theorem we deduce that A∗
is a V-polytope. Now, as O belongs to A and A is closed (by Proposition 5.11) and convex,
by Proposition 4.22 (even if A is not bounded) we have A = A∗∗ , and by Proposition 5.6
(first part), we conclude that A = A∗∗ is an H-polyhedron of the form A = P (Y, 1).
Interestingly, we can now prove easily that every H-polyhedron is a V-polyhedron.
Proposition 5.18. Every H-polyhedron is a V-polyhedron.
Proof. Let A be an H-polyhedron of dimension d. The case where A is bounded is covered

by Theorem 5.7 (Weyl-Minkowski), If A is unbounded, there are two cases to consider.
◦
(1) The pohyhedron A is cut out by hyperplanes not containing 0, that is, 0 ∈ A and A
is of the form A = P (Y, 1). In this case, by Proposition 5.6, A∗ is a V-polytope, but
since A is unbounded 0 lies on the boundary of A∗ . We can translate 0 using some
translation Ω so that the new origin Ω is now in the interior of A∗ , so by Theorem
5.7 (with respect to the origin Ω), A∗ is an H-polytope. We now translate Ω back
to 0 using the translation −Ω, but then 0 lies on the boundary of A∗ , so some of
the hyperplanes cutting out A∗ contain 0, which implies that A∗ is an H-polyhedron
of the form A∗ = P (Y, 1; V, 0). By Proposition 5.15, we deduce that A∗∗ = A is a
V-polyhedron (A = A∗∗ because 0 ∈ A and A is closed and convex).
(2) The pohyhedron A is cut out by hyperplanes, some of which contain 0, which means
A is of the form A = P (Y, 1; V, 0). By Proposition 5.15, the polar dual A∗ of A is a
V-polyhedron. As in the previous case, 0 lies on the boundary of A∗ . We translate 0
using some translation Ω so that the new origin Ω is now in the interior of A∗ , and
by Proposition 5.17, A∗ is an H-polyhedron. As in Case (1) we translate Ω back to 0
using the translation −Ω, so A∗ is of the form A∗ = P (Y, 1; V, 0). By Proposition 5.15,
we deduce that A∗∗ = A is a V-polyhedron (A = A∗∗ because 0 ∈ A and A is closed
and convex).
This concludes the proof that every H-polyhedron is a V-polyhedron.
Putting together Propositions 5.17 and 5.18 we obtain our main theorem:
Theorem 5.19. (Equivalence of H-polyhedra and V-polyhedra) Every H-polyhedron is a

V-polyhedron and conversely.
Both in Proposition 5.17 and in Proposition 5.18, the step that is not automatic is to
find the vertices of an H-polytope from the inequalities defining this H-polytope. A method
to do this algorithmically is Fourier–Motzkin elimination; see Proposition 5.21. This process
can be expansive, in the sense that the number of vertices can be exponential in the number
of inequalities. For example, the standard d-cube is cut out by 2d hyperplanes, but is has
2d vertices.
Here are some examples illustrating Proposition 5.17.
Example 5.1. Let A be the V-polyhedron (a triangle) in A2 defined by set Y = {(−1, −1/2),
(1, −1/2), (0, 1/2)}. By Proposition 5.14, the polar dual A∗ is the H-polytope, a triangle,
cut out by the inequalities:
−x − (1/2)y ≤ 1
x − (1/2)y ≤ 1
(1/2)y ≤ 1.
This is also the V-polytope whose vertices are (−2, 2), (2, 2), (0, −2). By Proposition 5.6,
A = A∗∗ is H-polyhedron cut out by the inequalities
−2x + 2y ≤ 1
2x + 2y ≤ 1
−2y ≤ 1,
which are equivalent to
y ≤ x + 1/2
y ≤ −x + 1/2
−y ≤ 1/2;
see Figure 5.3.
(-2,2) 1/2y = 1 (2,2)

-x -
(0,1/2)
1/2
A*
A
y=
=1
1
y
1/2
(0,1/2)
x-
(-1,-1/2) (1,-1/2)
(-1,-1/2) (1,-1/2)
(-2,2) (2,2)
(0,-2)
(0,1/2)
1 2x A**
= +
2y 2y
+ =
x 1
-2
-2y = 1
(-1,-1/2) (1,-1/2)
(0,-2)
Figure 5.3: The triangle of Example 5.1 written as both a V-polyhedron and an H-
polyhedron.
Example 5.2. Let A be the V-polyhedron (a square) in A2 defined by set Y = {(−1/2, 0),
(0, −1/2), (1/2, 0), (0, 1/2)}. By Proposition 5.14, the polar dual A∗ is the H-polytope,
another square, cut out by the inequalities:
(1/2)y ≤ 1
−(1/2)x ≤ 1
−(1/2)y ≤ 1
(1/2)x ≤ 1.
This is also the V-polytope whose vertices are (−2, 2), (−2, −2), (2, −2), (2, 2). By Proposi-
tion 5.6, A = A∗∗ is H-polyhedron cut out by the inequalities
−2x + 2y ≤1
−2x − 2y ≤1
2x − 2y ≤1
2x + 2y ≤ 1,
y ≤ x + 1/2
y ≥ −x − 1/2
y ≥ x − 1/2
y ≤ −x + 1/2;
see Figure 5.4.

Example 5.3. Let A be the V-polyhedron in A2 defined by the set Y = {(0, −1)} consisting
of a single point and the set of vectors V = {(−1, 1), (1, 1)}. This polyhedron is a cone with
apex (0, −1). By Proposition 5.14, the polar dual A∗ is the H-polytope, a triangle, cut out
by the inequalities:
−x + y ≤ 0
x+y ≤0
−y ≤ 1.
This is also the V-polytope whose vertices are (0, 0), (−1, −1), and (1, −1). By Proposition
5.6, A = A∗∗ is H-polyhedron (a cone) cut out by the inequalities
−x − y ≤ 1
x − y ≤ 1;
see Figure 5.5. Note that there is no line associated with (0, 0) since this point belongs to
the boundary of the triangle.
(-2,2) 1/2y = 1 (2,2)
(0,1/2) A*
A 1/2x = 1
(1/2,0)
(-1/2,0)
(0,1/2)
(0,-1/2)
-1/2x = 1
(1/2,0)
(-1/2,0)
(0,-1/2)
(-2,-2) -1/2y = 1
(-2,2) (2,-2)
(2,2)
(0,1/2)
A** =
1 2x
2y
+
x+
2y
-2 =
1
1 (1/2,0)
-2
(-1/2,0) =
x-
- 2y
2y
2x
=
1
(0,-1/2)
(-2,-2) (2,-2)
Figure 5.4: The diamond of Example 5.2 written as both a V-polyhedron and an H-
polyhedron.
Example 5.4. Let A be the V-polyhedron in A2 defined by set Y = {(−1, −1), (1, −1)}
and the set of vectors V = {(−1, 1), (1, 1)}. By Proposition 5.14, the polar dual A∗ is the
H-polytope, a convex polygon, cut out by the inequalities:
−x + y ≤0
x+y ≤0
−x − y ≤1
x−y ≤ 1.
This is also the V-polytope whose vertices are (0, 0), (−1/2, −1/2), (0, −1), (1/2, −1/2). By
Proposition 5.6, A = A∗∗ is H-polyhedron cut out by the inequalities
−(1/2)x − (1/2)y ≤ 1
−y ≤ 1
(1/2)x − (1/2)y ≤ 1,
)
,1
x+
(-1
=0
(1 y
, 1)
+y
0
=
-x
(0,-1) (0,-1) (1,-1)
(-1,-1) A*
1
y=
x-
-x-
A**
y=
1
(1,-1)
(-1,-1)
Figure 5.5: The triangular cone of Example 5.3 written as both a V-polyhedron and an
H-polyhedron.
y ≥ −x − 2
y ≥ −1
y ≥ x − 2;
see Figure 5.6.
Example 5.5. Let A be the V-polyhedron in A2 defined by sets Y = {(0, 1), (−1, 0), (1, 0)}
and V = {(0, −1)}. By Proposition 5.14, the polar dual A∗ is the H-polytope, a square, cut
out by the inequalities:
y≤1
−x ≤ 1
x≤1
−y ≤ 0.
This is also the V-polytope (square) whose vertices are (−1, 1), (−1, 0), (1, 0), (1, 1). By
A
1)
(-1
,
(1
, 1)
(-1,-1) (1,-1)
-x 1
-y =
= A* y
1 x-
=0
x+
y
+y
0
=
-x
(1,-1)
(-1,-1) (0,-1)
-1 1
/2 y=
x /2
-1
/2 -1
y 2x
= 1/
1 A
A**
(-1/2,-1/2) (1/2,-1/2)
y = -1
(1,-1)
(-1,-1)
Figure 5.6: The trough of Example 5.4 written as both a V-polyhedron and an H-polyhedron.
Proposition 5.6, A = A∗∗ is H-polyhedron cut out by the inequalities
−x + y ≤ 1
−x ≤ 1
x≤1
x + y ≤ 1;
see Figure 5.7.
In all the previous examples, the step that is not automatic is to find the vertices of an
H-polytope from the inequalities defining this H-polytope. For small examples in dimension
2 this is easy, but in general this is an expansive process. A method to do this algorithmically
is Fourier–Motzkin elimination; see Proposition 5.21.
Here are now some example illustrating Proposition 5.18.
Example 5.6. Let A be the H-polyhedron in A2 defined by the inequalities
−x − y ≤ 1
x − y ≤ 1.
(0,1) y=1
(-1,1) (1,1)
x = -1 x=1
(-1,0) (1,0) A*
(0,-1) (0,-1)
(0,1)
(-1,1) 1 x+ (1,1)
y= y=
+
-x 1
(-1,0) (1,0)
(0,-1)
x = -1 x=1
A**
Figure 5.7: The triangular peaked of Example 5.5 written as both a V-polyhedron and an
H-polyhedron.
This is the cone arising in Example 5.3. By Proposition 5.6, the polar dual A∗ a V-polytope,
a triangle, the convex hull of the points (0, 0), (−1, −1), and (1, −1). In Example 5.1, we
computed the equations of the triangle (0, 1/2), (−1, −1/2), and (1, −1/2) obtained by trans-
lating the above triangle by (0, 1/2), namely
y ≤ x + 1/2
y ≤ −x + 1/2
−y ≤ 1/2,
so the triangle (0, 0), (−1, −1), and (1, −1) is also the H-polyhedron (triangle) defined by
the inequalities
−x + y ≤ 0
x+y ≤0
−y ≤ 1,
and by Proposition 5.15, A = A∗∗ is the V-polyhedron given by the set Y = {(0, −1)}
consisting of a single point and the set of vectors V = {(−1, 1), (1, 1)}; see Figure 5.8.
1
y =
x-
-x-
y=
1
=0
x+
y
A
+y
0
=
-x
(-1,-1) y = -1 (1,-1)
A*
A**
)
,1
(-1
(1
, 1)
(0,-1)
Figure 5.8: The triangular cone of Example 5.6 written as first an H-polyhedron and then
V-polyhedron.
−(1/2)x − (1/2)y ≤ 1
−y ≤ 1
(1/2)x − (1/2)y ≤ 1.
This is the H-polyhedron of Example 5.4. By Proposition 5.6, the polar dual A∗ a V-
polytope, the square (0, 0), (−1/2, −1/2), (0, −1), (1/2, −1/2). In Example 5.2, we computed
the equations of the square (−1/2, 0), (0, −1/2), (1/2, 0), (0, 1/2) obtained by translating the
above square by (0, 1/2), namely
y ≤ x + 1/2
y ≥ −x − 1/2
y ≥ x − 1/2
y ≤ −x + 1/2,
so the square (0, 0), (−1/2, −1/2), (0, −1), (1/2, −1/2) is also the H-polyhedron (square) de-
fined by the inequalities
−x + y ≤0
−x − y ≤1
x−y ≤1
x+y ≤ 0.
By Proposition 5.15, A = A∗∗ is the V-polyhedron given by the set Y = {(−1, −1), (1, −1)}
and the set of vectors V = {(−1, 1), (1, 1)}; see Figure 5.9.
-1
/2
x- 1
1/
2y
2 y=
=
1 1/ A
x-
1 /2
y = -1
0
y=
+
-x
-x
-y A*
=
1
(-1/2,1/2)
(-1/2,/-1/2)
1 (0,-1)
x+
y=
y=
x-
0
, 1) A**
(-1
(1
, 1)
(-1,-1) (1,-1)
Figure 5.9: The trough of Example 5.4 written as first an H-polyhedron and then V-
polyhedron.
−x − y ≤ 1
−y ≤ 0
x − y ≤ 1.
By Proposition 5.6, the polar dual A∗ is the V-polyhedron given by the set of points
(0, 0), (−1, −1), and (1, −1) and the vector (0, −1). In example 5.5, we computed the inequal-
ities of the V-polyedron given by Y = {(0, 1), (−1, 0), (1, 0)} and V = {(0, −1)}, obtained
by translating the above V-polyhedron by (0, 1), namely
−x + y ≤ 1
−x ≤ 1
x≤1
x + y ≤ 1,
so the V-polyhedron given by the set of points (0, 0), (−1, −1), and (1, −1) and the vector
(0, −1) is defined by the inequalities
−x + y ≤ 0
−x ≤ 1
x≤1
x + y ≤ 0.
By Proposition 5.15, A = A∗∗ is the V-polyhedron given by the sets Y = {(−1, 0), (1, 0)}
and V = {(−1, 1), (1, 1)}; see Figure 5.10.
Even though we proved the main result of this section, it is instructive to consider a more
computational proof making use of cones and an elimination method known as Fourier–
Motzkin elimination.
5.5 Fourier–Motzkin Elimination and the Polyhedron-

Cone Correspondence
The problem with the converse of Proposition 5.17 when A is unbounded (i.e., not compact)
is that Krein and Milman’s Theorem does not apply. We need to take into account “points
at infinity” corresponding to certain vectors. The trick we used in Proposition 5.17 is that
the polar dual of a V-polyhedron with nonempty interior is an H-polytope. This reduction
to polytopes allowed us to use Krein and Milman to convert an H-polytope to a V-polytope
and then again we took the polar dual.
Another trick is to switch to cones by “homogenizing.” Given any subset, S ⊆ Ed , we
can form the cone C(S) ⊆ Ed+1 by “placing” a copy of S in the hyperplane Hd+1 ⊆ Ed+1 of
equation xd+1 = 1, and drawing all the half-lines from the origin through any point of S. If
S is given by m polynomial inequalities of the form
Pi (x1 , . . . , xd ) ≤ bi ,
5.5. FOURIER–MOTZKIN ELIMINATION AND CONES 131
-x
-y
=
1
1
=
y A
x-
-x
-y
=
1
1
=
y
x-
(-1,-1) (1,-1)
(0,-1)
A*
, 1)
(- 1
(1 A**
, 1)
(-1,0) (1,0)
Figure 5.10: The trough of Example 5.8 written as first an H-polyhedron and then V-
polyhedron.
where Pi (x1 , . . . , xd ) is a polynomial of total degree ni , this amounts to forming the new
homogeneous inequalities

ni x1 xd
xd+1 Pi ,..., − bi xnd+1
i
≤0
xd+1 xd+1
together with xd+1 ≥ 0. In particular, if the Pi ’s are linear forms (which means that ni = 1),
then our inequalities are of the form
ai · x ≤ b i ,
where ai ∈ Rd is some vector, and the homogenized inequalities are
ai · x − bi xd+1 ≤ 0.
It will be convenient to formalize the process of putting a copy of a subset S ⊆ Ed in the

hyperplane Hd+1 ⊆ Ed+1 of equation xd+1 = 1 as follows: For every point a ∈ Ed , let

a
a= ∈ Ed+1 ,
1
b
a | a ∈ S}. Obviously, the map S 7→ Sb is a bijection between the subsets of Ed

and let Sb = {b
and the subsets of Hd+1 . We will use this bijection to identify S and Sb and use the simpler
notation S, unless confusion arises. Let’s apply this to polyhedra.
Let P ⊆ Ed be an H-polyhedron. Then, P is cut out by m hyperplanes Hi , and for each
Hi , there is a nonzero vector ai and some bi ∈ R so that
Hi = {x ∈ Ed | ai · x = bi },
and P is given by
m
\
P = {x ∈ Ed | ai · x ≤ bi }.
i=1
If A denotes the m × d matrix whose i-th row is ai and b is the vector b = (b1 , . . . , bm ), then
we can write
P = P (A, b) = {x ∈ Ed | Ax ≤ b}.
We “homogenize” P (A, b) as follows: Let C(P ) be the subset of Ed+1 defined by

x d+1
C(P ) = ∈R | Ax ≤ xd+1 b, xd+1 ≥ 0
xd+1

x d+1
= ∈R | Ax − xd+1 b ≤ 0, −xd+1 ≤ 0 .
xd+1
Thus, we see that C(P ) is the H-cone given by the system of inequalities

A −b x 0
≤ ,
0 −1 xd+1 0
and that
Pb = C(P ) ∩ Hd+1 .
Conversely, if Q is any H-cone in Ed+1 (in fact, any H-polyhedron), it is clear that
P = Q ∩ Hd+1 is an H-polyhedron in Hd+1 ≈ Ed .
Let us now assume that P ⊆ Ed is a V-polyhedron, P = conv(Y ) + cone(V ), where
Y = {y1 , . . . , yp } and V = {v1 , . . . , vq }. Define Yb = {b y1 , . . . , ybp } ⊆ Ed+1 , and
v1 , . . . , vbq } ⊆ Ed+1 , by
Vb = {b

yi vj
ybi = , vbj = .
1 0
We check immediately that
C(P ) = cone({Yb ∪ Vb })
is a V-cone in Ed+1 such that

Pb = C(P ) ∩ Hd+1 ,
where Hd+1 is the hyperplane of equation xd+1 = 1.
Conversely, if C = cone(W ) is a V-cone in Ed+1 , with wi d+1 ≥ 0 for every wi ∈ W , we
prove next that P = C ∩ Hd+1 is a V-polyhedron.
Proposition 5.20. (Polyhedron–Cone Correspondence) We have the following correspon-
dence between polyhedra in Ed and cones in Ed+1 :
(1) For any H-polyhedron P ⊆ Ed , if P = P (A, b) = {x ∈ Ed | Ax ≤ b}, where A is an
m × d-matrix and b ∈ Rm , then C(P ) given by

A −b x 0
≤
0 −1 xd+1 0
is an H-cone in Ed+1 , and Pb = C(P ) ∩ Hd+1 , where Hd+1 is the hyperplane of equation
xd+1 = 1. Conversely, if Q is any H-cone in Ed+1 (in fact, any H-polyhedron), then
P = Q ∩ Hd+1 is an H-polyhedron in Hd+1 ≈ Ed .
(2) Let P ⊆ Ed be any V-polyhedron, where P = conv(Y ) + cone(V ) with Y = {y1 , . . . , yp }

y1 , . . . , ybp } ⊆ Ed+1 , and Vb = {b
and V = {v1 , . . . , vq }. Define Yb = {b v1 , . . . , vbq } ⊆ Ed+1 ,
by
yi vj
ybi = , vbj = .
1 0
Then,
C(P ) = cone({Yb ∪ Vb })
is a V-cone in Ed+1 such that
Pb = C(P ) ∩ Hd+1 ,
Conversely, if C = cone(W ) is a V-cone in Ed+1 , with wi d+1 ≥ 0 for every wi ∈ W ,

then P = C ∩ Hd+1 is a V-polyhedron in Hd+1 ≈ Ed .
In both (1) and (2), Pb = {b
p | p ∈ P }, with

p
pb = ∈ Ed+1 .
1
Proof. We already proved everything except the last part of the proposition. Let

wi
Y =
b wi ∈ W, wi d+1 > 0
wi d+1
and
Vb = {wj ∈ W | wj d+1 = 0}.
We claim that
P = C ∩ Hd+1 = conv(Yb ) + cone(Vb ),
and thus, P is a V-polyhedron.
Recall that any element z ∈ C can be written as
s
X
z= µk w k , µk ≥ 0.
k=1
Thus, we have
s
X
z= µk wk µk ≥ 0
k=1
X X
= µi w i + µj wj µi , µj ≥ 0
wi d+1 >0 wj d+1 =0
X wi X
= wi d+1 µi + µj wj , µi , µj ≥ 0
wi d+1 w
wi d+1 >0 j d+1 =0
X wi X
= λi + µj wj , λi , µj ≥ 0.
w >0
w i d+1 w =0
i d+1 j d+1
Now, z ∈ C ∩ Hd+1 iff zd+1 = 1 iff pi=1 λi = 1 (where p is the number of elements in Yb ),
P
since the (d + 1)th coordinate of wiwd+1
i
is equal to 1, and the above shows that
P = C ∩ Hd+1 = conv(Yb ) + cone(Vb ),
as claimed.
By Proposition 5.20, if P is an H-polyhedron, then C(P ) is an H-cone. If we can prove

that every H-cone is a V-cone, then again, Proposition 5.20 shows that Pb = C(P ) ∩ Hd+1
is a V-polyhedron and so P is a V-polyhedron. Therefore, in order to prove that every
H-polyhedron is a V-polyhedron it suffices to show that every H-cone is a V-cone.
By a similar argument, Proposition 5.20 shows that in order to prove that every V-
polyhedron is an H-polyhedron it suffices to show that every V-cone is an H-cone. We will
not prove this direction again since we already have it by Proposition 5.17.
It remains to prove that every H-cone is a V-cone. Let C ⊆ Ed be an H-cone. Then, C
is cut out by m hyperplanes Hi through 0. For each Hi , there is a nonzero vector ui so that
Hi = {x ∈ Ed | ui · x = 0},
and C is given by
m
\
C= {x ∈ Ed | ui · x ≤ 0}.
i=1
If A denotes the m × d matrix whose i-th row is ui , then we can write
C = P (A, 0) = {x ∈ Ed | Ax ≤ 0}.
Observe that C = C0 (A) ∩ Hw , where

x d+m
C0 (A) = ∈R | Ax ≤ w
w
is an H-cone in Ed+m and

x d+m
Hw = ∈R |w =0
w
is an affine subspace in Ed+m .

x

We claim that C0 (A) is a V-cone. This follows by observing that for every w
satisfying
Ax ≤ w, we can write
X d X m
x ei 0
= |xi |(sign(xi )) + (wj − (Ax)j ) ,
w i=1
Aei j=1
ej
and then

ei 0
C0 (A) = cone ± |1≤i≤d ∪ |1≤j ≤m .
Aei ej
Since C = C0 (A) ∩ Hw is now the intersection of a V-cone with an affine subspace, to

prove that C is a V-cone it is enough to prove that the intersection of a V-cone with a
hyperplane is also a V-cone. For this, we use Fourier–Motzkin elimination. It suffices to
prove the result for a hyperplane Hk in Ed+m of equation yk = 0 (1 ≤ k ≤ d + m).
Proposition 5.21. (Fourier–Motzkin Elimination) Say C = cone(Y ) ⊆ Ed is a V-cone.

Then, the intersection C ∩ Hk (where Hk is the hyperplane of equation yk = 0) is a V-cone
C ∩ Hk = cone(Y /k ), with
Y /k = {yi | yik = 0} ∪ {yik yj − yjk yi | yik > 0, yjk < 0},
the set of vectors obtained from Y by “eliminating the k-th coordinate.” Here, each yi is a
vector in Rd .
/k
Pdonly nontrivial direction is to prove that C ∩ Hk ⊆ cone(Y ). For this, consider
Proof. The
any v = i=1 ti yi ∈ C ∩ Hk , with ti ≥ 0 and vk = 0. Such a v can be written
X X X
v= ti yi + ti yi + tj yj
i|yik =0 i|yik >0 j|yjk <0
and as vk = 0, we have X X
ti yik + tj yjk = 0.
i|yik >0 j|yjk <0
If ti yik = 0 for i = 1, . . . , d, we are done. Otherwise, we can write

X X
Λ= ti yik = −tj yjk > 0.
i|yik >0 j|yjk <0
Then,
   
X 1 X X 1 X X
v= ti yi +  −tj yjk  ti yi +  ti yik  tj yj
Λ Λ
i|yik =0 i|yik >0 j|yjk <0 j|yjk <0 i|yik >0
X X ti tj
= ti yi + (yik yj − yjk yi ) .
Λ
i|yik =0 i|yik >0
j|yjk <0
Since the k th coordinate of yik yj − yjk yi is 0, the above shows that any v ∈ C ∩ Hk can be
written as a positive combination of vectors in Y /k .
As discussed above, Proposition 5.21 implies (again!)

Corollary 5.22. Every H-polyhedron is a V-polyhedron.
Another way of proving that every V-polyhedron is an H-polyhedron is to use cones.

This method is interesting in its own right so we discuss it briefly.
Let P = conv(Y ) + cone(V ) ⊆ Ed be a V-polyhedron. We can view Y as a d × p matrix
whose ith column is the ith vector in Y and V as d × q matrix whose jth column is the jth
vector in V . Then, we can write
P = {x ∈ Rd | (∃u ∈ Rp )(∃t ∈ Rd )(x = Y u + V t, u ≥ 0, Iu = 1, t ≥ 0)},
where I is the row vector

I = (1, . . . , 1) .
| {z }
p
Now, observe that P can be interpreted as the projection of the H-polyhedron Pe ⊆ Ed+p+q
given by
Pe = {(x, u, t) ∈ Rd+p+q | x = Y u + V t, u ≥ 0, Iu = 1, t ≥ 0}
onto Rd . Consequently, if we can prove that the projection of an H-polyhedron is also
an H-polyhedron, then we will have proved that every V-polyhedron is an H-polyhedron.
Here again, it is possible that P = Ed , but that’s fine since Ed has been declared to be an
H-polyhedron.
In view of Proposition 5.20 and the discussion that followed, it is enough to prove that the
projection of any H-cone is an H-cone. This can be done by using a type of Fourier–Motzkin
elimination dual to the method used in Proposition 5.21. We state the result without proof
and refer the interested reader to Ziegler [67], Section 1.2–1.3, for full details.
Proposition 5.23. If C = P (A, 0) ⊆ Ed is an H-cone, then the projection projk (C) onto
the hyperplane Hk of equation yk = 0 is given by projk (C) = elimk (C) ∩ Hk , with
elimk (C) = {x ∈ Rd | (∃t ∈ R)(x + tek ∈ P )} = {z − tek | z ∈ P, t ∈ R} = P (A/k , 0) and
where the rows of A/k are given by
A/k = {ai | ai k = 0} ∪ {ai k aj − aj k ai | ai k > 0, aj k < 0}.
It should be noted that both Fourier–Motzkin elimination methods generate a quadratic

number of new vectors or inequalities at each step and thus they lead to a combinatorial
explosion. Therefore, these methods become intractable rather quickly. The problem is
that many of the new vectors or inequalities are redundant. Therefore, it is important to
find ways of detecting redundancies and there are various methods for doing so. Again, the
interested reader should consult Ziegler [67], Chapter 1.
There is yet another way of proving that an H-polyhedron is a V-polyhedron without
using Fourier–Motzkin elimination that was inspired to us by the proof of Theorem 1.5
in Chapter II of Ewald [26]. As we already observed, Krein and Milman’s theorem does
not apply if our polyhedron is unbounded. Actually, the full power of Krein and Milman’s
theorem is not needed to show that an H-polytope is a V-polytope.
The crucial point is that if P is an H-polytope with nonempty interior, then every line
` through any point a in the interior of P intersects P in a line segment. This is because P
is compact and ` is closed, so P ∩ ` is a compact subset of a line thus, a closed interval [b, c]
with b < a < c, as a is in the interior of P . Therefore, we can use induction on the dimension
of P to show that every point in P is a convex combination of vertices of the facets of P .
Now, if P is unbounded and cut out by at least two half-spaces (so, P is not a half-space),
then we claim that for every point a in the interior of P , there is some line through a that
intersects two facets of P . This is because if we pick the origin
Tt in the interior of P , we may
assume that P is given by an irredundant intersection, P = i=1 (Hi )− , and for any point a
in the interior of P , there is a line ` through P in general position w.r.t. P , which means
that ` is not parallel to any of the hyperplanes Hi and intersects all of them in distinct
points (see Definition 10.2). Fortunately, lines in general position always exist, as shown in
Proposition 10.3. Using this fact, we can prove the following result:
Proposition 5.24. Let P ⊆ Ed be an H-polyhedron P = ti=1 (Hi )− (an irredundant decom-

T
position), with nonempty interior. If t = 1, that is, P = (H1 )− is a half-space, then
P = a + cone(u1 , . . . , ud−1 , −u1 , . . . , −ud−1 , ud ),

where a is any point in H1 , the vectors u1 , . . . , ud−1 form a basis of the direction of H1 , and
ud is normal to (the direction of ) H1 . (When d = 1, P is the half-line, P = {a+tu1 | t ≥ 0}.)
If t ≥ 2, then every point a ∈ P can be written as a convex combination a = (1 − α)b + αc
(0 ≤ α ≤ 1), where b and c belong to two distinct facets F and G of P , and where
F = conv(YF ) + cone(VF ) and G = conv(YG ) + cone(VG ),
for some finite sets of points YF and YG and some finite sets of vectors VF and VG . (Note:
α = 0 or α = 1 is allowed.) Consequently, every H-polyhedron is a V-polyhedron.
Proof. We proceed by induction on the dimension d of P . If d = 1, then P is either a closed
interval [b, c], or a half-line {a + tu | t ≥ 0}, where u 6= 0. In either case, the proposition is
clear.
For the induction step, assume d > 1. Since every facet F of P has dimension d − 1, the
induction hypothesis holds for F , that is, there exist a finite set of points YF , and a finite
set of vectors VF , so that
F = conv(YF ) + cone(VF ).
Thus, every point on the boundary of P is of the desired form. Next, pick any point a in
the interior of P . Then, from our previous discussion, there is a line ` through a in general
position w.r.t. P . Consequently, the intersection points zi = ` ∩ Hi of the line ` with the
hyperplanes supporting the facets of P exist and are all distinct. If we give ` an orientation,
the zi ’s can be sorted. Since ` contains a which is in the interior of P , any point on ` to the
left of z1 must be outside P , and similarly any point of ` to the right of the rightmost zi ,
say zN , must be outside P , since otherwise there would be a hyperplane cutting ` before z1
or a hyperplane cutting ` after zN . Since P is closed and convex and ` is closed, P ∩ ` is a
closed an convex subset of `. But this subset is bounded since all points outside [z1 , zN ] are
outside P . It follows that P ∩ ` is a closed interval [b, c] with b, c ∈ P , so there is a unique
k such that a lies between b = zk and c = zk+1 . But then, b ∈ Fk = F and c ∈ Fk+1 = G,
where F and G the facets of P supported by Hk and Hk+1 , and a = (1 − α)b + αc, a convex
combination.
Consequently, every point in P is a mixed convex + positive combination of finitely
many points associated with the facets of P and finitely many vectors associated with the
directions of the supporting hyperplanes of the facets P . Conversely, it is easy to see that
any such mixed combination must belong to P and therefore, P is a V-polyhedron.
We conclude this section with a version of Farkas Lemma for polyhedral sets.
Lemma 5.25. (Farkas Lemma, Version IV) Let Y be any d × p matrix and V be any d × q
matrix. For every z ∈ Rd , exactly one of the following alternatives occurs:
(a) There exist u ∈ Rp and t ∈ Rq , with u ≥ 0, t ≥ 0, Iu = 1 and z = Y u + V t.
(b) There is some vector (α, c) ∈ Rd+1 such that c> yi ≥ α for all i with 1 ≤ i ≤ p, c> vj ≥ 0
for all j with 1 ≤ j ≤ q, and c> z < α.
5.6. LINEALITY SPACE AND RECESSION CONE 139
Proof. We use Farkas Lemma, Version II (Lemma 4.16). Observe that (a) is equivalent to
the problem: Find (u, t) ∈ Rp+q , so that

u 0 I O u 1
≥ and = ,
t 0 Y V t z
which is exactly Farkas II(a). Now, the second alternative of Farkas II says that there is no
solution as above if there is some (−α, c) ∈ Rd+1 so that

> 1 > I 0
(−α, c ) < 0 and (−α, c ) ≥ (O, O).
z Y V
These are equivalent to
−α + c> z < 0, −αI + c> Y ≥ O, c> V ≥ O,
namely, c> z < α, c> Y ≥ αI and c> V ≥ O, which are indeed the conditions of Farkas IV(b),
in matrix form.
Observe that Farkas IV can be viewed as a separation criterion for polyhedral sets. This
version subsumes Farkas I and Farkas II.
5.6 Lineality Space and Recession Cone

Given a V-polyhedron P = conv(Y ) + cone(V ), the lines or the rays contained in P play an
important role which suggests the following definition.
Definition 5.5. Given a convex set A ⊆ Ed , the lineality space of A is defined by
lineal(A) = {v ∈ Rd | x + tv ∈ A for all x ∈ A, and all t ∈ R},
and the recession cone of A is defined by
rec(A) = {v ∈ Rd | x + tv ∈ A for all x ∈ A, and all t ≥ 0, where t ∈ R}.
It is immediate from these definitions that lineal(A) is a linear subspace of Rd and that
rec(A) is a convex cone in Rd containing 0.
If we pick a subspace U of Ed such that U and lineal(A) form a direct sum Ed = lineal(A)⊕
U , for example, the othogonal complement of lineal(A), we can decompose A as
A = lineal(A) + (U ∩ A),
with lineal(U ∩ A) = (0). A convex set A such that lineal(A) = (0) is said to be pointed .
If P is an H-polyhedron of the form P = P (A, d), then it is immediate by definition that
lineal(P ) = Ker A = {x ∈ Rd | Ax = 0}.
Regarding recession cones, we have the following proposition.
Proposition 5.26. Let P ⊆ Ed be a convex set.
(1) If P is an H-polyhedron of the form P = P (A, b), then the recession cone rec(P ) is
given by
rec(P ) = P (A, 0).
(2) If P is a V-polyhedron of the form P = conv(Y ) + cone(V ), then the recession cone
rec(P ) is given by
rec(P ) = cone(V ).
Proof. The only part whose proof is nontrivial is the inclusion rec(P ) ⊆ cone(V ). We prove
the contrapositive, if v ∈
/ cone(V ) then v ∈
/ rec(P ), using Farkas Lemma Version IV.
By Farkas Lemma Version IV (Proposition 5.25) with Y = ∅, if v ∈ / cone(V ), then there
is some c ∈ R and some α ∈ R such that c vj ≥ 0 for j = 1, . . . , q, 0 ≥ α, and c> v < α.
d >
This implies that −c> vj ≤ 0 for j = 1, . . . , q, and −c> v > 0. Let d = −c.
For any x ∈ P = conv(Y ) + cone(V ), we have
p q
X X
x= λi yi + µj vj ,
i=1 j=1
Pp
with λi , µj ≥ 0 and i=1 λi = 1. Since d> vj ≤ 0 for j = 1, . . . , q and 0 ≤ λi ≤ 1, we get
p q p
X X X
> > >
d x= λi d yi + µj d v j ≤ λi d> yi ≤ max d> yi = K,
1≤i≤p
i=1 j=1 i=1
where K is a constant that depends only on d and Y . However,
d> (x + tv) = d> x + td> v,
and since d> v > 0, we see that d> (x + tv) tends to +∞ when t tends to +∞, which implies
that x + tv ∈
/ P , and which means that v ∈ / rec(P ).
Proposition 5.26 shows that if P is a V-polyhedron then v ∈ rec(P ) iff there is some
x ∈ P such that x + tv ∈ P for all t ≥ 0, which is much cheaper to check that the condition
of Definition 5.5 which requires checking that x + tv ∈ P for all t ≥ 0 and for all x ∈ P .
Notes. The treatment of polytopes and polyhedra given in this chapter is based on the
following texts (in alphabetic order): Barvinok [4], Berger [8], Ewald [26], Grunbaum [35],
and Ziegler [67]. The terminology V-polyhedron, V-polytope, H-polyhedron, H-polytope, is
borrowed from Ziegler.
The proof of Theorem 5.7 (Weyl-Minkowski) using Krein and Millman’s theorem and po-
lar duality is taken from Berger [8] (Chapter 12, Proposition 12.1.5). Rather different proofs
5.6. LINEALITY SPACE AND RECESSION CONE 141
of the fact that every V-polytope is an H-polytope are given in Grunbaum [35] (Chapter 3,
Theorem 3.1.1), and Ewald [26] (Chapter II, Theorems 1.4 and 1.5).
We believe that the proof of Proposition 5.17 showing that every V-polyhedron is an H-
polyhedron using Proposition 5.14, Krein and Millman’s theorem, and double dualization, is
new. However, this proof is not that original in the sense that it uses the double dualization
trick already found in Berger, and the crucial observation that the polar dual A∗ of a V-
polyhedron A with respect to a center in the interior of A is a bounded H-polyhedron, that
is, an H-polytope. We also believe that the proof of Proposition 5.18 showing that every
H-polyhedron is a V-polyhedron is new (it uses a quadruple polar dualization!).
The equivalence of V-polyhedra is and H-polyhedra is also treated (using different tech-
niques) in some books on convex optimization, among which we recommend Bertsekas [9].
Except for Proposition 5.24, which we believe is new, the results of Section 5.5 on Fourier–
Motzkin elimination and the polyhedron-cone correspondence are taken from Ziegler [67].
Similarly, the material of Section 5.6 is taken from Ziegler [67].
Chapter 6
Linear Programs
6.1 What is Linear Programming?

What is linear programming? At first glance, one might think that this is some style of com-
puter programming. After all, there is imperative programming, functional programming,
object-oriented programming etc. The term linear programming is somewhat misleading,
because it really refers to a method for planning with linear constraints, or more accurately,
an optimization method where both the objective function and the constraints are linear.1
Linear programming was created in the late 1940’s, one of the key players being George
Dantzing, who invented the simplex algorithm. Kantorovitch also did some pioneering work
on linear programming as early as 1939. The term linear programming has a military con-
notation because in the early 1950’s it was used as a synonym for plans or schedules for
training troops, logistical supply, resource allocation, etc. Unfortunately the term linear
programming is well established and we are stuck with it.
Interestingly, even though originally most applications of linear programming were in
the field of economics and industrial engineering, linear programming has become an im-
portant tool in theoretical computer science and in the theory of algorithms. Indeed, linear
programming is often an effective tool for designing approximation algorithms to solve hard
problems (typically NP-hard problems). Linear programming is also the “baby version” of
convex programming, a very effective methodology which has received much attention in
recent years.
Our goal in these notes is to present the mathematical underpinnings of linear pro-
gramming, in particular the existence of an optimal solution if a linear program is feasible
and bounded, and the duality theorem in linear programming, one of the deepest results
in this field. The duality theorem in linear programming also has significant algorithmic
implications but we do not discuss this here. We present the simplex algorithm, the dual
simplex algorithm, and the primal dual algorithm. We also describe the tableau formalism
1
Again, we witness another unfortunate abuse of terminology; the constraints are in fact affine.
143
144 CHAPTER 6. LINEAR PROGRAMS
for running the simplex algorithm and its variants. A particularly nice feature of the tableau
formalism is that the update of a tableau can be performed using elementary row operations
identical to the operations used during the reduction of a matrix to row reduced echelon
form (rref). What differs is the criterion for the choice of the pivot.
However, we do not discuss other methods such as the ellipsoid method or interior points
methods. For these more algorithmic issues, we refer the reader to standard texts on linear
programming. In our opinion, one of the clearest (and among the most concise!) is Matousek
and Gardner [41]; Chvatal [18] and Schrijver [51] are classics. Papadimitriou and Steiglitz
[45] offers a very crisp presentation in the broader context of combinatorial optimization,
and Bertsimas and Tsitsiklis [10] and Vanderbei [64] are very complete.
Linear programming has to do with maximizing a linear cost function c1 x1 + · · · + cn xn
with respect to m “linear” inequalities of the form
ai1 x1 + · · · + ain xn ≤ bi .
These constraints can be put together into an m × n matrix A = (aij ), and written more
concisely as
Ax ≤ b.
For technical reasons that will appear clearer later on, it is often preferable to add the
nonnegativity constaints xi ≥ 0 for i = 1, . . . , n. We write x ≥ 0. It is easy to show that
every linear program is equivalent to another one satisfying the constraints x ≥ 0, at the
expense of adding new variables that are also constrained to be nonnegative. Let P(A, b) be
the set of feasible solutions of our linear program given by
P(A, b) = {x ∈ Rn | Ax ≤ b, x ≥ 0}.
Then, there are two basic questions:
(1) Is P(A, b) nonempty, that is, does our linear program have a chance to have a solution?
(2) Does the objective function c1 x1 + · · · + cn xn have a maximum value on P(A, b)?
The answer to both questions can be no. But if P(A, b) is nonempty and if the objective
function is bounded above (on P(A, b)), then it can be shown that the maximum of c1 x1 +
· · · + cn xn is achieved by some x ∈ P(A, b). Such a solution is called an optimal solution.
Perhaps surprisingly, this result is not so easy to prove (unless one has the simplex method
as its disposal). We will prove this result in full detail (see Proposition 6.1).
The reason why linear constraints are so important is that the domain of potential optimal
solutions P(A, b) is convex . In fact, P(A, b) is a convex polyhedron which is the intersection
of half-spaces cut out by affine hyperplanes. The objective function being linear is convex,
and this is also a crucial fact. Thus, we are led to study convex sets, in particular those that
arise from solutions of inequalities defined by affine forms, but also convex cones.
6.2. NOTATIONAL PRELIMINARIES 145
We give a brief introduction to these topics. As a reward, we provide several criteria for
testing whether a system of inequalities
Ax ≤ b, x ≥ 0
has a solution or not in terms of versions of the Farkas lemma (see Proposition 4.14 and
Proposition 8.4). Then we give a complete proof of the strong duality theorem for linear
programming (see Theorem 8.7). We also discuss the complementary slackness conditions
and show that they can be exploired to design an algorithm for solving a linear program
that uses both the primal problem and its dual. This algorithm known as the primal dual
algorithm, although not used much nowadays, has been the source of inspiration for a whole
class of approximation algorithms also known as primal dual algorithms.
We hope that these notes will be a motivation for learning more about linear program-
ming, convex optimization, but also convex geometry. The “bible” in convex optimization
is Boyd and Vandenberghe [14], and one of the best sources for convex geometry is Ziegler
[67].
6.2 Notational Preliminaries

We view Rn as consisting of column vectors (n × 1 matrices). As usual, row vectors represent
linear forms, that is linear maps ϕ : Rn → R, in the sense that the row vector y (a 1 × n
matrix) represents the linear form ϕ if ϕ(x) = yx for all x ∈ Rn . We denote the space of
linear forms (row vectors) by (Rn )∗ .
Definition 6.1. An affine form ϕ : Rn → R is defined by some linear form c ∈ (Rn )∗ and
some scalar β ∈ R so that
ϕ(x) = cx + β for all x ∈ Rn .
If c 6= 0, the affine form ϕ specified by (c, β) defines the affine hyperplane (for short hyper-
plane) H(ϕ) given by
H(ϕ) = {x ∈ Rn | ϕ(x) = 0} = {x ∈ Rn | cx + β = 0},
and the two (closed) half-spaces
H+ (ϕ) = {x ∈ Rn | ϕ(x) ≥ 0} = {x ∈ Rn | cx + β ≥ 0},

H− (ϕ) = {x ∈ Rn | ϕ(x) ≤ 0} = {x ∈ Rn | cx + β ≤ 0}.
When β = 0, we call H a linear hyperplane.

Both H+ (ϕ) and H− (ϕ) are convex and H = H+ (ϕ) ∩ H− (ϕ).
For example, ϕ : R2 → R with ϕ(x, y) = 2x + y + 3 is an affine form defining the line
given by the equation y = −2x − 3. Another example of an affine form is ϕ : R3 → R
y=
-2x
3-
(0,0,1)
H_ H+
H+
H_
(0,1,0)
(1,0,0)
x+y+z=1
i. ii.
Figure 6.1: Figure i. illustrates the hyperplane H(ϕ) for ϕ(x, y) = 2x + y + 3, while Figure
ii. illustrates the hyperplane H(ϕ) for ϕ(x, y, z) = x + y + z − 1.
with ϕ(x, y, z) = x + y + z − 1; this affine form defines the plane given by the equation
x + y + z = 1, which is the plane through the points (0, 0, 1), (0, 1, 0), and (1, 0, 0). Both of
these hyperplanes are illustrated in Figure 6.1.
For any two vector x, y ∈ Rn with x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) we write x ≤ y
iff xi ≤ yi for i = 1, . . . , n, and x ≥ y iff y ≤ x. In particular x ≥ 0 iff xi ≥ 0 for i = 1, . . . , n.
6.3 Linear Programs, Feasible Solutions, Optimal So-

lutions
The purpose of linear programming is to solve the following type of optimization problem.
Definition 6.2. A linear program (P ) is the following kind of optimization problem:
maximize cx
subject to
a1 x ≤ b 1
...
am x ≤ b m
x ≥ 0,
where x ∈ Rn , c, a1 , . . . , am ∈ (Rn )∗ , b1 , . . . , bm ∈ R.
The linear form c defines the objective function x 7→ cx of the program (P ) (from Rn to
R), and the inequalities ai x ≤ bi and xj ≥ 0 are called the constraints of the linear program
(P ).
6.3. LINEAR PROGRAMS, FEASIBLE SOLUTIONS, OPTIMAL SOLUTIONS 147
If we define the m × n matrix  

a1
A =  ... 
 
am
whose rows are the row vectors a1 , . . . , am and b as the column vector
 
b1
 .. 
b =  . ,
bm
the m inequality constraints ai x ≤ bi can be written in matrix form as
Ax ≤ b.
Thus the linear program (P ) can also be stated as the linear program (P ):
maximize cx
subject to Ax ≤ b and x ≥ 0.
Here is an explicit example of a linear program of type (P ):
Example 6.1.
maximize x1 + x2
subject to
x 2 − x1 ≤ 1
x1 + 6x2 ≤ 15
4x1 − x2 ≤ 10
x1 ≥ 0, x2 ≥ 0,
and in matrix form

x1
maximize 1 1
x2
subject to
   
−1 1 1
x
1 6  1 ≤ 15
x2
4 −1 10
x1 ≥ 0, x2 ≥ 0.
10
1
=
1 x2 =
1
3 -x
x2
4x -
(3,2)
2 x1 + 6x
2 = 15
K1 0 1 2 3 4 5
K1
Figure 6.2: The H-polyhedron associated with Example 6.1. The green point (3, 2) is the
unique optimal solution.
It turns out that x1 = 3, x2 = 2 yields the maximum of the objective function x1 + x2 ,

which is 5. This is illustrated in Figure 6.2. Observe that the set of points that satisfy
the above constraints is a convex region cut out by half planes determined by the lines of
equations
x2 − x1 =1
x1 + 6x2 = 15
4x1 − x2 = 10
x1 =0
x2 = 0.
In general, each constraint ai x ≤ bi corresponds to the affine form ϕi given by ϕi (x) =

ai x − bi and defines the half-space H− (ϕi ), and each inequality xj ≥ 0 defines the half-space
H+ (xj ). The intersection of these half-spaces is the set of solutions of all these constraints.
It is a (possibly empty) H-polyhedron denoted P(A, b).
Definition 6.3. If P(A, b) = ∅, we say that the linear program (P ) has no feasible solution,
and otherwise any x ∈ P(A, b) is called a feasible solution of (P ).
The linear program shown in Example 6.2 obtained by reversing the direction of the
inequalities x2 − x1 ≤ 1 and 4x1 − x2 ≤ 10 in the linear program of Example 6.1 has no
feasible solution; see Figure 6.3.
Example 6.2.
maximize x1 + x2
subject to
x1 − x2 ≤ −1
x1 + 6x2 ≤ 15
x2 − 4x1 ≤ −10
x1 ≥ 0, x2 ≥ 0.
4
1
=
-x 1
10
3 x2
1 x2 =
4x -
x1 + 6x
2 = 15
2
K1 0 1 2 3 4 5
K1
Figure 6.3: There is no H-polyhedron associated with Example 6.2 since the blue and purple
regions do not overlap.
Assume P(A, b) 6= ∅, so that the linear program (P ) has a feasible solution. In this case,
consider the image {cx ∈ R | x ∈ P(A, b)} of P(A, b) under the objective function x 7→ cx.
Definition 6.4. If the set {cx ∈ R | x ∈ P(A, b)} is unbounded above, then we say that the
linear program (P ) is unbounded .
The linear program shown in Example 6.3 obtained from the linear program of Example
6.1 by deleting the constraints 4x1 − x2 ≤ 10 and x1 + 6x2 ≤ 15 is unbounded.
Example 6.3.
maximize x1 + x2
subject to
x 2 − x1 ≤ 1
x1 ≥ 0, x2 ≥ 0.
Otherwise, we will prove shortly that if µ is the least upper bound of the set {cx ∈ R |
x ∈ P(A, b)}, then there is some p ∈ P(A, b) such that
cp = µ,
that is, the objective function x 7→ cx has a maximum value µ on P(A, b) which is achieved
by some p ∈ P(A, b).
Definition 6.5. If the set {cx ∈ R | x ∈ P(A, b)} is nonempty and bounded above, any
point p ∈ P(A, b) such that cp = max{cx ∈ R | x ∈ P(A, b)} is called an optimal solution
(or optimum) of (P ). Optimal solutions are often denoted by an upper ∗; for example, p∗ .
The linear program of Example 6.1 has a unique optimal solution (3, 2), but observe
that the linear program of Example 6.4 in which the objective function is (1/6)x1 + x2 has
infinitely many optimal solutions; the maximum of the objective function is 15/6 which
occurs along the points of orange boundary line in Figure 6.2.
Example 6.4.
1
maximize x1 + x2
6
subject to
x 2 − x1 ≤ 1
x1 + 6x2 ≤ 15
4x1 − x2 ≤ 10
x1 ≥ 0, x2 ≥ 0.
The proof that if the set {cx ∈ R | x ∈ P(A, b)} is nonempty and bounded above, then
there is an optimal solution p ∈ P(A, b), is not as trivial as it might seem. It relies on the
fact that a polyhedral cone is closed, a fact that was shown in Section 4.1.
We also use a trick that makes the proof simpler, which is that a linear program (P ) with
inequality constraints Ax ≤ b
maximize cx
subject to Ax ≤ b and x ≥ 0,
is equivalent to the linear program (P2 ) with equality constraints
maximize b cx
b
subject to Ab b ≥ 0,
bx = b and x
c is a linear form in (Rn+m )∗ , and x

b is an m × (n + m) matrix, b
where A b ∈ Rn+m , given by

>
x
A = A Im , b
b c = c 0m , and x b= ,
z
with x ∈ Rn and z ∈ Rm .
Indeed, Ab b ≥ 0 iff
bx = b and x
Ax + z = b, x ≥ 0, z ≥ 0,
iff
Ax ≤ b, x ≥ 0,
and b
cxb = cx.
The variables z are called slack variables, and a linear program of the form (P2 ) is called
a linear program in standard form.
The result of converting the linear program of Example 6.4 to standard form is the
program shown in Example 6.5.
Example 6.5.
1
maximize x 1 + x2
6
subject to
x2 − x1 + z1 = 1
x1 + 6x2 + z2 = 15
4x1 − x2 + z3 = 10
x1 ≥ 0, x2 ≥ 0, z1 ≥ 0, z2 ≥ 0, z3 ≥ 0.
We can now prove that if a linear program has a feasible solution and is bounded, then
it has an optimal solution.
Proposition 6.1. Let (P2 ) be a linear program in standard form, with equality constraint
Ax = b. If P(A, b) is nonempty and bounded above, and if µ is the least upper bound of the
set {cx ∈ R | x ∈ P(A, b)}, then there is some p ∈ P(A, b) such that
cp = µ,
that is, the objective function x 7→ cx has a maximum value µ on P(A, b) which is achieved
by some optimum solution p ∈ P(A, b).
Proof. Since µ = sup{cx ∈ R | x ∈ P(A, b)}, there is a sequence (x(k) )k≥0 of vectors
(k) (k)
x(k) ∈ P(A, b) such that limk7→∞ cx(k) = µ. In particular, if we write x(k) = (x1 , . . . , xn )
(k)
we have xj ≥ 0 for j = 1, . . . , n and for all k ≥ 0. Let A
e be the (m + 1) × n matrix

c
A
e= ,
A
e (k) )k≥0 of vectors Ax

and consider the sequence (Ax e (k) ∈ Rm+1 . We have
(k) (k)
(k) c (k) cx cx
Ax =
e x = (k) = ,
A Ax b
since by hypothesis x(k) ∈ P(A, b), and the constraints are Ax = b and x ≥ 0. Since by
e (k) )k≥0 converges to the vector µ . Now,
hypothesis limk7→∞ cx(k) = µ, the sequence (Ax
b
e (k) can be written as the convex combination
observe that each vector Ax
n
(k) ej
X
e (k) =
Ax xj A ,
j=1
(k) ej ∈ Rm+1 is the jth column of A. e (k) belongs to the

with xj ≥ 0 and where A e Therefore, Ax
polyheral cone
C = cone(Ae1 , . . . , A
en ) = {Ax
e | x ∈ Rn , x ≥ 0},
e (k) ∈ C, which means that there
and since by Proposition 4.13 this cone is closed, limk≥∞ Ax
is some u ∈ Rn with u ≥ 0 such that

µ (k) e = cu ,
= lim Ax
e = Au
b k≥∞ Au
that is, cu = µ and Au = b. Hence, u is an optimal solution of (P2 ).
The next question is, how do we find such an optimal solution? It turns out that for
linear programs in standard form where the constraints are of the form Ax = b and x ≥ 0,
there are always optimal solutions of a special type called basic feasible solutions.
6.4 Basic Feasible Solutions and Vertices

If the system Ax = b has a solution and if some row of A is a linear combination of other
rows, then the corresponding equation is redundant, so we may assume that the rows of A
are linearly independent; that is, we may assume that A has rank m, so m ≤ n.
If A is an m×n matrix, for any nonempty subset K of {1, . . . , n}, let AK be the submatrix
of A consisting of the columns of A whose indices belong to K. We denote the jth column
of the matrix A by Aj .
Definition 6.6. Given a linear program (P2 )
maximize cx
subject to Ax = b and x ≥ 0,
where A has rank m, a vector x ∈ Rn is a basic feasible solution of (P ) if x ∈ P(A, b) 6= ∅,
and if there is some subset K of {1, . . . , n} of size m such that
6.4. BASIC FEASIBLE SOLUTIONS AND VERTICES 153
(1) The matrix AK is invertible (that is, the columns of AK are linearly independent).
(2) xj = 0 for all j ∈

/ K.
The subset K is called a basis of x. Every index k ∈ K is called basic, and every index
/ K is called nonbasic. Similarly, the columns Ak corresponding to indices k ∈ K are
j ∈
called basic, and the columns Aj corresponding to indices j ∈/ K are called nonbasic. The
variables corresponding to basic indices k ∈ K ar called basic variables, and the variables
corresponding to indices j ∈
/ K are called nonbasic.
For example, the linear program
maximize x1 + x2
subject to x1 + x2 + x3 = 1 and x1 ≥ 0, x2 ≥ 0, x3 ≥ 0, (∗)
has three basic feasible solutions; the basic feasible solution K = {1} corresponds to the
point (1, 0, 0); the basic feasible solution K = {2} corresponds to the point (0, 1, 0); the
basic feasible solution K = {3} corresponds to the point (0, 0, 1). Each of these points
corresponds to the vertices of the slanted purple triangle illustrated in Figure 6.4. The
vertices (1, 0, 0) and (0, 1, 0) optimize the objective function with a value of 1.
.7
x+y=0
1
+z=
x+y
Figure 6.4: The H-polytope associated with Linear Program (∗). The objective function
(with x1 → x and x2 → y) is represented by vertical planes parallel to the purple plane
x + y = 0.7, and reaches it maximal value when x + y = 1.
We now show that if the standard linear program (P2 ) as in Definition 6.6 has some
feasible solution and is bounded above, then some basic feasible solution is an optimal
solution. We follow Matousek and Gardner [41] (Chapter 4, Section 2, Theorem 4.2.3).
First we obtain a more convenient characterization of a basic feasible solution.
Proposition 6.2. Given any standard linear program (P2 ) where Ax = b and A is an m × n
matrix of rank m, for any feasible solution x, if J> = {j ∈ {1, . . . , n} | xj > 0}, then x is a
basic feasible solution iff the columns of the matrix AJ> are linearly independent.
Proof. If x is a basic feasible solution then there is some subset K ⊆ {1, . . . , n} of size m such
that the columns of AK are linearly independent and xj = 0 for all j ∈ / K, so by definition
J> ⊆ K, which implies that the columns of the matrix AJ> are linearly independent.
Conversely, assume that x is a feasible solution such that the columns of the matrix AJ>
are linearly independent. If |J> | = m, we are done since we can pick K = J> and then x
is a basic feasible solution. If |J> | < m, we can extend J> to an m-element subset K by
adding m − |J> | column indices so that the columns of AK are linearly independent, which
is possible since A has rank m.
Next we prove that if a linear program in standard form has any feasible solution x0 and
is bounded above, then is has some basic feasible solution x
e which is as good as x0 , in the
x ≥ cx0 .
sense that ce
Proposition 6.3. Let (P2 ) be any standard linear program with objective function cx, where
Ax = b and A is an m × n matrix of rank m. If (P2 ) is bounded above and if x0 is some
feasible solution of (P2 ), then there is some basic feasible solution x x ≥ cx0 .
e such that ce
Proof. Among the feasible solutions x such that cx ≥ cx0 (x0 is one of them) pick one with
the maximum number of coordinates xj equal to 0, say x e. Let
K = J> = {j ∈ {1, . . . , n} | x
ej > 0}
and let s = |K|. We claim that x x ≥ cx0 .

e is a basic feasible solution, and by construction ce
If the columns of AK are linearly independent, then by Proposition 6.2 we know that x
e
is a basic feasible solution and we are done.
Otherwise, the columns of AK are linearly dependent, so there is some nonzero vector
v = (v1 , . . . , vs ) such that AK v = 0. Let w ∈ Rn be the vector obtained by extending v by
setting wj = 0 for all j ∈ / K. By construction,
Aw = AK v = 0.
We will derive a contradiction by exhibiting a feasible solution x(t0 ) such that cx(t0 ) ≥ cx0
with more zero coordinates than xe.
For this we claim that we may assume that w satisfies the following two conditions:
(1) cw ≥ 0.
(2) There is some j ∈ K such that wj < 0.

If cw = 0 and if Condition (2) fails, since w 6= 0, we have wj > 0 for some j ∈ K, in

which case we can use −w, for which wj < 0.
If cw < 0 then c(−w) > 0, so we may assume that cw > 0. If wj > 0 for all j ∈ K, since
x e ≥ 0, and so x(t) = x
e is feasible x e + tw ≥ 0 for all t ≥ 0. Furthermore, since Aw = 0 and
x
e is feasible, we have
Ax(t) = Ae
x + tAw = b,
and thus x(t) is feasible for all t ≥ 0. We also have
cx(t) = ce
x + tcw.
Since cw > 0, as t > 0 goes to infinity the objective function cx(t) also tends to infinity,
contradicting the fact that is is bounded above. Therefore, some w satisfying Conditions (1)
and (2) above must exist.
We show that there is some t0 > 0 such that cx(t0 ) ≥ cx0 and x(t0 ) = x
e + t0 w is feasible,
yet x(t0 ) has more zero coordinates than x
e, a contradiction.
Since x(t) = x
e + tw, we have
x(t)i = x
ei + twi ,
so if we let I = {i ∈ {1, . . . , n} | wi < 0} ⊆ K, which is nonempty since w satisfies Condition
(2) above, if we pick
−e

xi
t0 = min ,
i∈I wi
then t0 > 0, because wi < 0 for all i ∈ I, and by definition of K we have x
ei > 0 for all i ∈ K.
By the definition of t0 > 0 and since xe ≥ 0, we have
ej + t0 wj ≥ 0 for all j ∈ K,
x(t0 )j = x
so x(t0 ) ≥ 0, and x(t0 )i = 0 for some i ∈ I. Since Ax(t0 ) = b (for any t), x(t0 ) is a feasible
solution,
x + t0 cw ≥ cx0 + t0 cw ≥ cx0 ,
cx(t0 ) = ce
and x(t0 )i = 0 for some i ∈ I, we see that x(t0 ) has more zero coordinates than x
e, a
contradiction.
Proposition 6.3 implies the following important result.
Theorem 6.4. Let (P2 ) be any standard linear program with objective function cx, where
Ax = b and A is an m × n matrix of rank m. If (P2 ) has some feasible solution and if it is
bounded above, then some basic feasible solution x
e is an optimal solution of (P2 ).
Proof. By Proposition 6.3, for any feasible solution x there is some basic feasible solution x
e
such that cx ≤ cex. But there are only finitely many basic feasible solutions, so one of them
has to yield the maximum of the objective function.
Geometrically, basic solutions are exactly the vertices of the polyhedron P(A, b), a notion
that we now define.
Definition 6.7. Given an H-polyhedron P ⊆ Rn , a vertex of P is a point v ∈ P with

property that there is some nonzero linear form c ∈ (Rn )∗ and some µ ∈ R, such that v
is the unique point of P for which the map x 7→ cx has the maximum value µ ; that is,
cy < cv = µ for all y ∈ P − {v}. Geometrically this means that the hyperplane of equation
cy = µ touches P exactly at v. More generally, a convex subset F of P is a k-dimensional
face of P if F has dimension k and if there is some affine form ϕ(x) = cx − µ such that
cy = µ for all y ∈ F , and cy < µ for all y ∈ P − F . A 1-dimensional face is called an edge.
The concept of a vertex is illustrated in Figure 6.5, while the concept of an edge is
illustrated in Figure 6.6.
x+y+z=3
(1,1,1)
Figure 6.5: The cube centered at the origin with diagonal through (−1, −1, −1) and (1, 1, 1)
has eight vertices. The vertex (1, 1, 1) is associated with the linear form x + y + z = 3.
Since a k-dimensional face F of P is equal to the intersection of the hyperplane H(ϕ)

of equation cx = µ with P, it is indeed convex and the notion of dimension makes sense.
Observe that a 0-dimensional face of P is a vertex. If P has dimension d, then the (d − 1)-
dimensional faces of P are called its facets.
If (P ) is a linear program in standard form, then its basic feasible solutions are exactly
the vertices of the polyhedron P(A, b). To prove this fact we need the following simple
proposition
Proposition 6.5. Let Ax = b be a linear system where A is an m × n matrix of rank m.

For any subset K ⊆ {1, . . . , n} of size m, if AK is invertible, then there is at most one basic
feasible solution x ∈ Rn with xj = 0 for all j ∈ / K (of course, x ≥ 0)
(1,1,-1)
(1,1,1)
x+y=2
Figure 6.6: The cube centered at the origin with diagonal through (−1, −1, −1) and (1, 1, 1)
has twelve edges. The vertex edge from (1, 1, −1) to (1, 1, 1) is associated with the linear
form x + y = 2.
Proof. In order for x to be feasible we must have Ax = b. Write N = {1, . . . , n} − K, xK

for the vector consisting of the coordinates of x with indices in K, and xN for the vector
consisting of the coordinates of x with indices in N . Then
Ax = AK xK + AN xN = b.
In order for x to be a basic feasible solution we must have xN = 0, so
AK xK = b.
Since by hypothesis AK is invertible, xK = A−1 K b is uniquely determined. If xK ≥ 0 then x

is a basic feasible solution, otherwise it is not. This proves that there is at most one basic
feasible solution x ∈ Rn with xj = 0 for all j ∈/ K.
Theorem 6.6. Let (P ) be a linear program in standard form, where Ax = b and A is an

m × n matrix of rank m. For every v ∈ P(A, b), the following conditions are equivalent:
(1) v is a vertex of the polyhedron P(A, b).
(2) v is a basic feasible solution of the linear program (P ).
Proof. First, assume that v is a vertex of P(A, b), and let ϕ(x) = cx − µ be a linear form
such that cy < µ for all y ∈ P(A, b) and cv = µ. This means that v is the unique point of
P(A, b) for which the objective function x 7→ cx has the maximum value µ on P(A, b), so by
Theorem 6.4, since this maximum is achieved by some basic feasible solution, by uniqueness
v must be a basic feasible solution.
Conversely, suppose v is a basic feasible solution of (P ) corresponding to a subset K ⊆

c ∈ (Rn )∗ be the linear form defined by
{1, . . . , n} of size m. Let b
(
0 if j ∈ K
cj =
b
−1 if j ∈
/ K.
By construction b c x ≤ 0 for any x ≥ 0, hence the function x 7→ b

c v = 0 and b c x on P(A, B)
has a maximum at v. Furthermore, b c x < 0 for any x ≥ 0 such that xj > 0 for some j ∈ / K.
However, by Proposition 6.5, the vector v is the only basic feasible solution such that vj = 0
for all j ∈
/ K, and therefore v is the only point of P(A, b) maximizing the function x 7→ b c x,
so it is a vertex.
n

In theory, to find an optimal solution we try all m possible m-elements subsets K of
{1, . . . , n} and solve for the corresponding unique solution xK of AK x = b. Then we check
whether such a solution satisfies xK ≥ 0, compute cxK , and return some feasible xK for
which the objective function is maximum. This is a totally impractible algorithm.
A practical algorithm is the simplex algorithm. Basically, the simplex algorithm tries to
“climb” in the polyhderon P(A, b) from vertex to vertex along edges (using basic feasible
solutions), trying to maximize the objective function. We present the simplex algorithm in
the next chapter. The reader may also consult texts on linear programming. In particular,
we recommend Matousek and Gardner [41], Chvatal [18], Papadimitriou and Steiglitz [45],
Bertsimas and Tsitsiklis [10], Ciarlet [19], Schrijver [51], and Vanderbei [64].
Observe that Theorem 6.4 asserts that if a linear program (P ) in standard form (where
Ax = b and A is an m×n matrix of rank m) has some feasible solution and is bounded above,
then some basic feasible solution is an optimal solution. By Theorem 6.6, the polyhedron
P(A, b) must have some vertex.
But suppose we only know that P(A, b) is nonempty; that is, we don’t know that the
objective function cx is bounded above. Does P(A, b) have some vertex?
The answer to the above question is yes, and this is important because the simplex
algorithm needs an initial basic feasible solution to get started. Here we prove that if P(A, b)
is nonempty, then it must contain a vertex. This proof still doesn’t constructively yield a
vertex, but we will see in the next chapter that the simplex algorithm always finds a vertex
if there is one (provided that we use a pivot rule that prevents cycling).
Theorem 6.7. Let (P ) be a linear program in standard form, where Ax = b and A is an

m × n matrix of rank m. If P(A, b) is nonempty (there is a feasible solution), then P(A, b)
has some vertex; equivalently, (P ) has some basic feasible solution.
Proof. The proof relies on a trick, which is to add slack variables xn+1 , . . . , xn+m and use the
new objective function −(xn+1 + · · · + xn+m ).
b be the m × (m + n)-matrix, and x, x, and x

If we let A b be the vectors given by
   
x1 xn+1
. .
b = A Im , x =  ..  ∈ R , x =  ..  ∈ R , x
n m x
∈ Rn+m ,

A b=
   
x
xn xn+m
then consider the linear program (Pb) in standard form

maximize − (xn+1 + · · · + xn+m )
subject to A
bxb = b and x b ≥ 0.
Since xi ≥ 0 for all i, the objective function −(xn+1 + · · · + xn+m ) is bounded above by
0. The system Abxb = b is equivalent to the system
Ax + x = b,
so for every feasible solution u ∈ P(A, b), since Au = b, the vector (u, 0m ) is also a feasible
solution of (Pb), in fact an optimal solution since the value of the objective function −(xn+1 +
· · ·+xn+m ) for x = 0 is 0. By Proposition 6.3, the linear program (Pb) has some basic feasible
solution (u∗ , w∗ ) for which the value of the objective function is greater than or equal to the
value of the objective function for (u, 0m ), and since (u, 0m ) is an optimal solution, (u∗ , w∗ )
is also an optimal solution of (Pb). This implies that w∗ = 0, since otherwise the objective
function −(xn+1 + · · · + xn+m ) would have a strictly negative value.
Therefore, (u∗ , 0m ) is a basic feasible solution of (Pb), and thus the columns corresponding
to nonzero components of u∗ are linearly independent. Some of the coordinates of u∗ could
be equal to 0, but since A has rank m we can add columns of A to obtain a basis K associated
with u∗ , and u∗ is indeed a basic feasible solution of (P ).
The definition of a basic feasible solution can be adapted to linear programs where the
constraints are of the form Ax ≤ b, x ≥ 0; see Matousek and Gardner [41] (Chapter 4,
Section 4, Definition 4.4.2).
The most general type of linear program allows constraints of the form ai x ≥ bi or
ai x = bi besides constraints of the form ai x ≤ bi . The variables xi may also take negative
values. It is always possible to convert such programs to the type considered in Definition
6.2. We proceed as follows.
Every constraint ai x ≥ bi is replaced by the constraint −ai x ≤ −bi . Every equality
constraint ai x = bi is replaced by the two constraints ai x ≤ bi and −ai x ≤ −bi .
If there are n variables xi , we create n new variables yi and n new variables zi and
replace every variable xi by yi − zi . We also add the 2n constraints yi ≥ 0 and zi ≥ 0. If the
constraints are given by the inequalities Ax ≤ b, we now have constraints given by

y
A −A ≤ b, y ≥ 0, z ≥ 0.
z
We replace the objective function cx by cy − cz.
Remark: We also showed that we can replace the ineqality constraints Ax ≤ b by equality
constraints Ax = b, by adding slack variables constrained to be nonnegative.
Chapter 7
The Simplex Algorithm
7.1 The Idea Behind the Simplex Algorithm

The simplex algorithm, due to Dantzig, applies to a linear program (P ) in standard form,
where the constraints are given by Ax = b and x ≥ 0, with A a m × n matrix of rank m,
and with an objective function c 7→ cx. This algorithm either reports that (P ) has no
feasible solution, or that (P ) is unbounded, or yields an optimal solution. Geometrically,
the algorithm climbs from vertex to vertex in the polyhedron P(A, b), trying to improve
the value of the objective function. Since vertices correspond to basic feasible solutions, the
simplex algorithm actually works with basic feasible solutions.
Recall that a basic feasible solution x is a feasible solution for which there is a subset
K ⊆ {1, . . . , n} of size m such that the matrix AK consisting of the columns of A whose
indices belong to K are linearly independent, and that xj = 0 for all j ∈ / K. We also let
J> (x) be the set of indices
J> (x) = {j ∈ {1, . . . , n} | xj > 0},
so for a basic feasible solution x associated with K, we have J> (x) ⊆ K. In fact, by
Proposition 6.2, a feasible solution x is a basic feasible solution iff the columns of AJ> (x) are
linearly independent.
If J> (x) had cardinality m for all basic feasible solutions x, then the simplex algorithm
would make progress at every step, in the sense that it would strictly increase the value of the
objective function. Unfortunately, it is possible that |J> (x)| < m for certain basic feasible
solutions, and in this case a step of the simplex algorithm may not increase the value of the
objective function. Worse, in rare cases, it is possible that the algorithm enters an infinite
loop. This phenomenon called cycling can be detected, but in this case the algorithm fails
to give a conclusive answer.
Fortunately, there are ways of preventing the simplex algorithm from cycling (for exam-
ple, Bland’s rule discussed later), although proving that these rules work correctly is quite
involved.
161
162 CHAPTER 7. THE SIMPLEX ALGORITHM
The potential “bad” behavior of a basic feasible solution is recorded in the following
definition.
Definition 7.1. Given a linear program (P ) in standard form where the constraints are
given by Ax = b and x ≥ 0, with A an m × n matrix of rank m, a basic feasible solution x
is degenerate if |J> (x)| < m, otherwise it is nondegenerate.
The origin 0n , if it is a basic feasible solution, is degenerate. For a less trivial example,
x = (0, 0, 0, 2) is a degenerate basic feasible solution of the following linear program in which
m = 2 and n = 4.
Example 7.1.
maximize x2
subject to
− x1 + x2 + x3 = 0
x1 + x4 = 2
x1 ≥ 0, x2 ≥ 0, x3 ≥ 0, x4 ≥ 0.
The matrix A and the vector b are given by

−1 1 1 0 0
A= , b= ,
1 0 0 1 2
and if x = (0, 0, 0, 2), then J> (x) = {4}. There are two ways of forming a set of two linearly
independent columns of A containing the fourth column.
Given a basic feasible solution x associated with a subset K of size m, since the columns
of the matrix AK are linearly independent, by abuse of language we call the columns of AK
a basis of x.
If u is a vertex of (P ), that is, a basic feasible solution of (P ) associated with a basis
K (of size m), in “normal mode,” the simplex algorithm tries to move along an edge from
the vertex u to an adjacent vertex v (with u, v ∈ P(A, b) ⊆ Rn ) corresponding to a basic
feasible solution whose basis is obtained by replacing one of the basic vectors Ak with k ∈ K
by another nonbasic vector Aj for some j ∈ / K, in such a way that the value of the objective
function is increased.
Let us demonstrate this process on an example.
Example 7.2. Let (P ) be the following linear program in standard form.
maximize x1 + x2
subject to
− x1 + x2 + x 3 = 1
x1 + x4 = 3
x2 + x5 = 2
x1 ≥ 0, x2 ≥ 0, x3 ≥ 0, x4 ≥ 0, x5 ≥ 0.
7.1. THE IDEA BEHIND THE SIMPLEX ALGORITHM 163

   
−1 1 1 0 0 1
A=  1 0 0 1 0 , b = 3 .

0 1 0 0 1 2
2 u2
2=1
1+x
-x
1
u0 u1
K1 0 1 2 3 4 5
K1
Figure 7.1: The planar H-polyhedron associated with Example 7.2. The initial basic feasible
solution is the origin. The simplex algorithm first moves along the horizontal orange line to
feasible solution at vertex u1 . It then moves along the vertical red line to obtain the optimal
feasible solution u2 .
The vector u0 = (0, 0, 1, 3, 2) corresponding to the basis K = {3, 4, 5} is a basic feasible

solution, and the corresponding value of the objective function is 0 + 0 = 0. Since the
columns (A3 , A4 , A5 ) corresponding to K = {3, 4, 5} are linearly independent we can express
A1 and A2 as
A1 = −A3 + A4
A2 = A3 + A5 .
Since
1A3 + 3A4 + 2A5 = Au0 = b,
for any θ ∈ R, we have
b = 1A3 + 3A4 + 2A5 − θA1 + θA1

= 1A3 + 3A4 + 2A5 − θ(−A3 + A4 ) + θA1
= θA1 + (1 + θ)A3 + (3 − θ)A4 + 2A5 ,
and
b = 1A3 + 3A4 + 2A5 − θA2 + θA2

= 1A3 + 3A4 + 2A5 − θ(A3 + A5 ) + θA1
= θA2 + (1 − θ)A3 + 3A4 + (2 − θ)A5 .
In the first case, the vector (θ, 0, 1 + θ, 3 − θ, 2) is a feasible solution iff 0 ≤ θ ≤ 3, and
the new value of the objective function is θ.
In the second case, the vector (0, θ, 1 − θ, 3, 2 − θ, 1) is a feasible solution iff 0 ≤ θ ≤ 1,
and the new value of the objective function is also θ.
Consider the first case. It is natural to ask whether we can get another vertex and increase
the objective function by setting to zero one of the coordinates of (θ, 0, 1 + θ, 3 − θ, 2), in this
case the fouth one, by picking θ = 3. This yields the feasible solution (3, 0, 4, 0, 2), which
corresponds to the basis (A1 , A3 , A5 ), and so is indeed a basic feasible solution, with an
improved value of the objective function equal to 3. Note that A4 left the basis (A3 , A4 , A5 )
and A1 entered the new basis (A1 , A3 , A5 ).
We can now express A2 and A4 in terms of the basis (A1 , A3 , A5 ), which is easy to do
since we already have A1 and A2 in term of (A3 , A4 , A5 ), and A1 and A4 are swapped. Such
a step is called a pivoting step. We obtain
A2 = A3 + A5
A4 = A1 + A3 .
Then we repeat the process with u1 = (3, 0, 4, 0, 2) and the basis (A1 , A3 , A5 ). We have
b = 3A1 + 4A3 + 2A5 − θA2 + θA2

= 3A1 + 4A3 + 2A5 − θ(A3 + A5 ) + θA2
= 3A1 + θA2 + (4 − θ)A3 + (2 − θ)A5 ,
and
b = 3A1 + 4A3 + 2A5 − θA4 + θA4

= 3A1 + 4A3 + 2A5 − θ(A1 + A3 ) + θA4
= (3 − θ)A1 + (4 − θ)A3 + θA4 + 2A5 .
In the first case, the point (3, θ, 4 − θ, 0, 2 − θ) is a feasible solution iff 0 ≤ θ ≤ 2, and the
new value of the objective function is 3 + θ. In the second case, the point (3 − θ, 0, 4 − θ, θ, 2)
is a feasible solution iff 0 ≤ θ ≤ 3, and the new value of the objective function is 3 − θ. To
increase the objective function we must choose the first case and we pick θ = 2. Then, we
get the feasible solution u2 = (3, 2, 2, 0, 0), which corresponds to the basis (A1 , A2 , A3 ), and
thus is a basic feasible solution. The new value of the objective function is 5.
Next we express A4 and A5 in terms of the basis (A1 , A2 , A3 ). Again this is easy to do
since we just swapped A5 and A2 (a pivoting step), and we get
A5 = A2 − A3
A4 = A1 + A3 .
We repeat the process with u2 = (3, 2, 2, 0, 0) and the basis (A1 , A2 , A3 ). We have
b = 3A1 + 2A2 + 2A3 − θA4 + θA4
= 3A1 + 2A2 + 2A3 − θ(A1 + A3 ) + θA4
= (3 − θ)A1 + 2A2 + (2 − θ)A3 + θA4 ,
and
b = 3A1 + 2A2 + 2A3 − θA5 + θA5
= 3A1 + 2A2 + 2A3 − θ(A2 − A3 ) + θA5
= 3A1 + (2 − θ)A2 + (2 + θ)A3 + θA5 .
In the first case, the point (3 − θ, 2, 2 − θ, θ, 0) is a feasible solution iff 0 ≤ θ ≤ 2, and the
value of the objective function is 5 − θ. In the second case, the point (3, 2 − θ, 2 + θ, 0, θ) is
a feasible solution iff 0 ≤ θ ≤ 2, and the value of the objective function is also 5 − θ. Since
we must have θ ≥ 0 to have a feasible solution, there is no way to increase the objective
function. In this situation, it turns out that we have reached an optimal solution, in our
case u2 = (3, 2, 2, 0, 0), with the maximum of the objective function equal to 5.
We could also have applied the simplex algorithm to the vertex u0 = (0, 0, 1, 3, 2) and to
the vector (0, θ, 1 − θ, 3, 2 − θ, 1), which is a feasible solution iff 0 ≤ θ ≤ 1, with new value
of the objective function θ. By picking θ = 1, we obtain the feasible solution (0, 1, 0, 3, 1),
corresponding to the basis (A2 , A4 , A5 ), which is indeed a vertex. The new value of the
objective function is 1. Then we express A1 and A3 in terms the basis (A2 , A4 , A5 ) obtaining
A1 = A4 − A3
A3 = A2 − A5 ,
and repeat the process with (0, 1, 0, 3, 1) and the basis (A2 , A4 , A5 ). After three more steps
we will reach the optimal solution u2 = (3, 2, 2, 0, 0).
Let us go back to the linear program of Example 7.1 with objective function x2 and where
the matrix A and the vector b are given by

−1 1 1 0 0
A= , b= .
1 0 0 1 2
Recall that u0 = (0, 0, 0, 2) is a degenerate basic feasible solution, and the objective function
has the value 0. See Figure 7.2 for a planar picture of the H-polyhedron associated with
Example 7.1.
2 u2
x2
u1
K1 0 1 2 3
x1
K1
>
solution is the origin. The simplex algorithm moves along the slanted orange line to the apex
of the triangle.
Pick the basis (A3 , A4 ). Then we have
A1 = −A3 + A4
A2 = A3 ,
and we get
b = 2A4 − θA1 + θA1

= 2A4 − θ(−A3 + A4 ) + θA1
= θA1 + θA3 + (2 − θ)A4 ,
and
b = 2A4 − θA2 + θA2

= 2A4 − θA3 + θA2
= θA2 − θA3 + 2A4 .
In the first case, the point (θ, 0, θ, 2 − θ) is a feasible solution iff 0 ≤ θ ≤ 2, and the value of
the objective function is 0, and in the second case the point (0, θ, −θ, 2) is a feasible solution
iff θ = 0, and the value of the objective function is θ. However, since we must have θ = 0 in
the second case, there is no way to increase the objective function either.
It turns out that in order to make the cases considered by the simplex algorithm as
mutually exclusive as possible, since in the second case the coefficient of θ in the value of
the objective function is nonzero, namely 1, we should choose the second case. We must
pick θ = 0, but we can swap the vectors A3 and A2 (because A2 is coming in and A3 has
the coefficient −θ, which is the reason why θ must be zero), and we obtain the basic feasible
solution u1 = (0, 0, 0, 2) with the new basis (A2 , A4 ). Note that this basic feasible solution
corresponds to the same vertex (0, 0, 0, 2) as before, but the basis has changed. The vectors
A1 and A3 can be expressed in terms of the basis (A2 , A4 ) as
A1 = −A2 + A4
A3 = A2 .
We now repeat the procedure with u1 = (0, 0, 0, 2) and the basis (A2 , A4 ), and we get
b = 2A4 − θA1 + θA1

= 2A4 − θ(−A2 + A4 ) + θA1
= θA1 + θA2 + (2 − θ)A4 ,
and
b = 2A4 − θA3 + θA3

= 2A4 − θA2 + θA3
= −θA2 + θA3 + 2A4 .
In the first case, the point (θ, θ, 0, 2−θ) is a feasible solution iff 0 ≤ θ ≤ 2 and the value of the
objective function is θ, and in the second case the point (0, −θ, θ, 2) is a feasible solution iff
θ = 0 and the value of the objective function is θ. In order to increase the objective function
we must choose the first case and pick θ = 2. We obtain the feasible solution u2 = (2, 2, 0, 0)
whose corresponding basis is (A1 , A2 ) and the value of the objective function is 2.
The vectors A3 and A4 are expressed in terms of the basis (A1 , A2 ) as
A3 = A2
A4 = A1 + A3 ,
and we repeat the procedure with u2 = (2, 2, 0, 0) and the basis (A1 , A2 ). We get
b = 2A1 + 2A2 − θA3 + θA3

= 2A1 + 2A2 − θA2 + θA3
= 2A1 + (2 − θ)A2 + θA3 ,
and
b = 2A1 + 2A2 − θA4 + θA4

= 2A1 + 2A2 − θ(A1 + A3 ) + θA4
= (2 − θ)A1 + 2A2 − θA3 + θA4 .
In the first case, the point (2, 2 − θ, 0, θ) is a feasible solution iff 0 ≤ θ ≤ 2 and the value of
the objective function is 2 − θ, and in the second case, the point (2 − θ, 2, −θ, θ) is a feasible
solution iff θ = 0 and the value of the objective function is 2. This time there is no way
to improve the objective function and we have reached an optimal solution u2 = (2, 2, 0, 0)
with the maximum of the objective function equal to 2.
Let us now consider an example of an unbounded linear program.
Example 7.3. Let (P ) be the following linear program in standard form.
maximize x1
subject to
x1 − x2 + x3 = 1
− x1 + x2 + x4 = 2
x1 ≥ 0, x2 ≥ 0, x3 ≥ 0, x4 ≥ 0.

1 −1 1 0 1
A= , b= .
−1 1 0 1 2
:
inequal x y % 1, x C y % 2, x R 0, y R 0 , x = 1 ..5, y = 1 ..5, color = "Nautical 1"
5
3 =
2
x2
+
-x 1
1
=
2
-x
1 x1
θ(1,1)
K1 0 1 2 3 4 5
K1
>
solution is the origin. The simplex algorithm first moves along the horizontal indigo line to
basic feasible solution at vertex (1, 0). Any optimal feasible solution occurs by moving along
the boundary line parameterized by the orange arrow θ(1, 1).
The vector u0 = (0, 0, 1, 2) corresponding to the basis K = {3, 4} is a basic feasible

solution, and the corresponding value of the objective function is 0. The vectors A1 and A2
are expressed in terms of the basis (A3 , A4 ) by
A1 = A3 − A4
A2 = −A3 + A4 .
Starting with u0 = (0, 0, 1, 2), we get
b = A3 + 2A4 − θA1 + θA1

= A3 + 2A4 − θ(A3 − A4 ) + θA1
= θA1 + (1 − θ)A3 + (2 + θ)A4 ,
and
b = A3 + 2A4 − θA2 + θA2

= A3 + 2A4 − θ(−A3 + A4 ) + θA2
= θA2 + (1 + θ)A3 + (2 − θ)A4 .
In the first case, the point (θ, 0, 1 − θ, 2 + θ) is a feasible solution iff 0 ≤ θ ≤ 1 and the value
of the objective function is θ, and in the second case, the point (0, θ, 1 + θ, 2 − θ) is a feasible
solution iff 0 ≤ θ ≤ 2 and the value of the objective function is 0. In order to increase the
objective function we must choose the first case, and we pick θ = 1. We get the feasible
solution u1 = (1, 0, 0, 3) corresponding to the basis (A1 , A4 ), so it is a basic feasible solution,
and the value of the objective function is 1.
The vectors A2 and A3 are given in terms of the basis (A1 , A4 ) by
A2 = −A1
A3 = A1 + A4 .
Repeating the process with u1 = (1, 0, 0, 3), we get
b = A1 + 3A4 − θA2 + θA2

= A1 + 3A4 − θ(−A1 ) + θA2
= (1 + θ)A1 + θA2 + 3A4 ,
and
b = A1 + 3A4 − θA3 + θA3

= A1 + 3A4 − θ(A1 + A4 ) + θA3
= (1 − θ)A1 + θA3 + (3 − θ)A4 .
In the first case, the point (1 + θ, θ, 0, 3) is a feasible solution for all θ ≥ 0 and the value
of the objective function if 1 + θ, and in the second case, the point (1 − θ, 0, θ, 3 − θ) is a
feasible solution iff 0 ≤ θ ≤ 1 and the value of the objective function is 1 − θ. This time, we
are in the situation where the points
(1 + θ, θ, 0, 3) = (1, 0, 0, 3) + θ(1, 1, 0, 0), θ≥0
form an infinite ray in the set of feasible solutions, and the objective function 1 + θ is
unbounded from above on this ray. This indicates that our linear program, although feasible,
is unbounded.
Let us now describe a step of the simplex algorithm in general.
7.2 The Simplex Algorithm in General

We assume that we already have an initial vertex u0 to start from. This vertex corresponds
to a basic feasible solution with basis K0 . We will show later that it is always possible to
find a basic feasible solution of a linear program (P ) is standard form, or to detect that (P )
has no feasible solution.
The idea behind the simplex algorithm is this: Given a pair (u, K) consisting of a basic
feasible solution u and a basis K for u, find another pair (u+ , K + ) consisting of another basic
feasible solution u+ and a basis K + for u+ , such that K + is obtained from K by deleting
some basic index k − ∈ K and adding some nonbasic index j + ∈ / K, in such a way that the
value of the objective function increases (preferably strictly). The step which consists in
− +
swapping the vectors Ak and Aj is called a pivoting step.
Let u be a given vertex corresponds to a basic feasible solution with basis K. Since the
m vectors Ak corresponding to indices k ∈ K are linearly independent, they form a basis, so
for every nonbasic j ∈
/ K, we write
X j
Aj = γk Ak . (∗)
k∈K
j j
We let γK ∈ Rm be the vector given by γK = (γkj )k∈K . Actually, since the vector γK j
depends
j
on K, to be very precise we should denote its components by (γK )k , but to simplify notation
we usually write γkj instead of (γK
j
)k (unless confusion arises). We will explain later how the
j
coefficients γk can be computed efficiently.
Since u is a feasible solution we have u ≥ 0 and Au = b, that is,
X
uk Ak = b. (∗∗)
k∈K
For every nonbasic j ∈

/ K, a candidate for entering the basis K, we try to find a new vertex
u(θ) that improves the objective function, and for this we add −θAj + θAj = 0 to b in
7.2. THE SIMPLEX ALGORITHM IN GENERAL 171
the equation (∗∗) and then replace the occurrence of Aj in −θAj by the right hand side of
equation (∗) to obtain
X
b= uk Ak − θAj + θAj
k∈K
X X
= k
uk A − θ γkj Ak + θAj
k∈K k∈K
X
= uk − θγkj Ak + θAj .
k∈K
Consequently, the vector u(θ) appearing on the right-hand side of the above equation given
by  j
ui − θγi
 if i ∈ K
u(θ)i = θ if i = j

0 if i ∈
/ K ∪ {j}

automatically satisfies the constraints Au(θ) = b, and this vector is a feasible solution iff
θ ≥ 0 and uk ≥ θγkj for all k ∈ K.
Obviously θ = 0 is a solution, and if

j uk j
γ > 0, k ∈ K > 0,
θ = min j k
γk
then we have a range of feasible solutions for 0 ≤ θ ≤ θj . The value of the objective function
for u(θ) is
X j
X j
cu(θ) = ck (uk − θγk ) + θcj = cu + θ cj − γk ck .
k∈K k∈K
Since the potential change in the objective function is

X j
θ cj − γk ck
k∈K
γkj ck ≤ 0 then the objective function can’t be increased.

P
and θ ≥ 0, if cj − k∈K
+ +
However, if cj + − k∈K γkj ck > 0 for some j + ∈ / K, and if θj > 0, then the objective
P
+
function can be strictly increased by choosing any θ > 0 such that θ ≤ θj , so it is natural
+
to zero at least one coefficient of u(θ) by picking θ = θj , which also maximizes the increase
of the objective function. In this case (Case below (B2)), we obtain a new feasible solution
+
u+ = u(θj ).
+ + + +
Now, if θj > 0, then there is some index k ∈ K such uk > 0, γkj > 0, and θj = uk /γkj ,
−
so we can pick such an index k − for the vector Ak leaving the basis K. We claim that
+
K + = (K − {k − }) ∪ {j + } is a basis. This is because the coefficient γkj − associated with the
− +
column Ak is nonzero (in fact, γkj − > 0), so equation (∗), namely
+ + −
X +
Aj = γkj − Ak + γkj Ak ,
k∈K−{k− }
yields the equation

− + +
X + +
Ak = (γkj − )−1 Aj − (γkj − )−1 γkj Ak ,
k∈K−{k− }
and these equations imply that the subspaces spanned by the vectors (Ak )k∈K and the vectors
(Ak )k∈K + are identical. However, K is a basis of dimension m so this subspace has dimension
+
m, and since K + also has m elements, it must be a basis. Therefore, u+ = u(θj ) is a basic
feasible solution.
The above case is the most common one, but other situations may arise. In what follows,
we discuss all eventualities.
Case (A).
We have cj − k∈K γkj ck ≤ 0 for all j ∈
P
/ K. Then it turns out that u is an optimal
solution. Otherwise, we are in Case (B).
Case (B).
γkj ck > 0 for some j ∈
P
We have cj − k∈K / K (not necessarily unique). There are three
subcases.
Case (B1).
/ K as above we also have γkj ≤ 0 for all k ∈ K, since uk ≥ 0 for all k ∈ K,
If for some j ∈
this places no restriction on θ, and the objective function is unbounded above.
Case (B2).
There is some index j + ∈
/ K such that simultaneously
+
(1) cj + − k∈K γkj ck > 0, which means that the objective function can potentially be
P
increased;
+ +
(2) There is some k ∈ K such that γkj > 0, and for every k ∈ K, if γkj > 0 then uk > 0,
+
which implies that θj > 0.
+
If we pick θ = θj where

j+ uk j +
θ = min j + γk > 0, k ∈ K > 0,
γk
then the feasible solution u+ given by

 +
j+ j
ui − θ γi
 if i ∈ K
u+
+
i = θj if i = j +

0 / K ∪ {j + }
if i ∈

+
is a vertex of P(A, b). If we pick any index k − ∈ K such that θj = uk− /γkj+
− , then
− j+
K = (K − {k }) ∪ {j } is a basis for u . The vector A enters the new basis K + , and the
+ + +
−
vector Ak leaves the old basis K. This is a pivoting step. The objective function increases
strictly.
Case (B3).
/ K such that cj − k∈K γkj ck > 0, and for each of the indices
P
There is some index j ∈
j∈
/ K satisfying the above property we have simultaneously
(1) cj − k∈K γkj ck > 0, which means that the objective function can potentially be in-
P
creased;
(2) There is some k ∈ K such that γkj > 0, and uk = 0, which implies that θj = 0.
Consequently, the objective function does not change. In this case, u is a degenerate basic
feasible solution.
We can associate to u+ = u a new basis K + as follows: Pick any index j + ∈
/ K such that
X j+
cj + − γk ck > 0,
k∈K
−
and any index k ∈ K such that
+
γkj − > 0,
+
and let K + = (K − {k − }) ∪ {j + }. As in Case (B2), The vector Aj enters the new basis K + ,
−
and the vector Ak leaves the old basis K. This is a pivoting step. However, the objective
function does not change since θj+ = 0.
It is easy to prove that in Case (A) the basic feasible solution u is an optimal solution,
and that in Case (B1) the linear program is unbounded. We already proved that in Case
(B2) the vector u+ and its basis K + constitutes a basic feasible solution, and the proof in
Case (B3) is similar. For details, see Ciarlet [19] (Chapter 10).
It is convenient to reinterpret the various cases considered by introducing the followings
sets:
n X j o
B1 = j ∈ / K | cj − γk ck > 0, max γkj ≤ 0
k∈K
k∈K
X j nu o
k
B2 = j ∈
/ K | cj − γk ck > 0, max γkj > 0, min j j
k ∈ K, γk > 0 > 0

k∈K
k∈K γk

X j n uk o
B3 = j ∈
/ K | cj − γk ck > 0, max γkj > 0, min j j
k ∈ K, γk > 0 = 0 ,

k∈K
k∈K γk
and n X j o
B = B1 ∪ B2 ∪ B3 = j ∈
/ K | cj − γk ck > 0 .
k∈K
Then it is easy to see that the following equivalences hold:
Case (A) ⇐⇒ B = ∅, Case (B) ⇐⇒ B 6= ∅

Case (B1) ⇐⇒ B1 6= ∅
Case (B2) ⇐⇒ B2 6= ∅
Case (B3) ⇐⇒ B3 6= ∅.
Furthermore, (A) and (B), (B1) and (B3), (B2) and (B3) are mutually exclusive, while (B1)
and (B2) are not.
If Case (B1) and Case (B2) arise simultaneously, we opt for Case (B1) which says that
the linear program (P ) is unbounded and terminate the algorithm.
Here are a few remarks about the method.
In Case (B2), which is the path followed by the algorithm most frequently, various choices
+
have to be made for the index j + ∈/ K for which θj > 0 (the new index in K + ). Similarly,
various choices have to be made for the index k − ∈ K leaving K, but such choices are
typically less important.
Similarly in Case (B3), various choices have to be made for the new index j + ∈
/ K going
+
into K . In Cases (B2) and (B3), criteria for making such choices are called pivot rules.
Case (B3) only arises when u is a degenerate vertex. But even if u is degenerate, Case
(B2) may arise if uk > 0 whenever γkj > 0. It may also happen that u is nondegenerate but
as a result of Case (B2), the new vertex u+ is degenerate because at least two components
+ + + +
uk1 − θj γkj1 and uk2 − θj γkj2 vanish for some distinct k1 , k2 ∈ K.
Cases (A) and (B1) correspond to situations where the algorithm terminates, and Case
(B2) can only arise a finite number of times during execution of the simplex algorithm, since
the objective function is strictly increased from vertex to vertex and there are only finitely
many vertices. Therefore, if the simplex algorithm is started on any initial basic feasible
solution u0 , then one of three mutually exclusive situations may arise:
(1) There is a finite sequence of occurrences of Case (B2) and/or Case (B3) ending with an
occurrence of Case (A). Then the last vertex produced by the algorithm is an optimal
solution.
(2) There is a finite sequence of occurrences of Case (B2) and/or Case (B3) ending with
an occurrence of Case (B1). We conclude that the problem is unbounded, and thus
has no solution.
(3) There is a finite sequence of occurrences of Case (B2) and/or Case (B3), followed by
an infinite sequence of Case (B3). If this occurs, the algorithm visits the some basis
twice. This a phenomenon known as cycling. In this eventually the algorithm fails to
come to a conclusion.
There are examples for which cycling occur, although this is rare in practice. Such an
example is given in Chvatal [18]; see Chapter 3, pages 31-32, for an example with seven
variables and three equations that cycles after six iterations under a certain pivot rule.
The third possibility can be avoided by the choice of a suitable pivot rule. Two of these
rules are Bland’s rule and the lexicographic rule; see Chvatal [18] (Chapter 3, pages 34-38).
Bland’s rule says: choose the smallest of the elligible incoming indices j + ∈
/ K, and
similarly choose the smallest of the elligible outgoing indices k − ∈ K.
It can be proved that cycling cannot occur if Bland’s rule is chosen as the pivot rule. The
proof is very technical; see Chvatal [18] (Chapter 3, pages 37-38), Matousek and Gardner [41]
(Chapter 5, Theorem 5.8.1), and Papadimitriou and Steiglitz [45] (Section 2.7). Therefore,
assuming that some initial basic feasible solution is provided, and using a suitable pivot rule
(such as Bland’s rule), the simplex algorithm always terminates and either yields an optimal
solution or reports that the linear program is unbounded. Unfortunately Bland’s rules is one
of the slowest pivot rules.
The choice of a pivot rule affects greatly the number of pivoting steps that the simplex
algorithms goes through. It is not our intention here to explain the various pivot rules.
We simply mention the following rules, referring the reader to Matousek and Gardner [41]
(Chapter 5, Section 5.7) or to the texts cited in Section 6.1.
1. Largest coefficient.
2. Largest increase.
3. Steepest edge.
4. Bland’s Rule.
5. Random edge.
The steepest edge rule is one of the most popular. The idea is to maximize the ratio
c(u+ − u)
.
ku+ − uk
The random edge rule picks the index j + ∈

/ K of the entering basis vector uniformly at
random among all elligible indices.
Let us now return to the issue of the initialization of the simplex algorithm. We use the
linear program (Pb) introduced during the proof of Theorem 6.7.
Consider a linear program (P 2)
maximize cx
in standard form where A is an m × n matrix of rank m.

First, observe that since the constraints are equations, we can ensure that b ≥ 0, because
evey equation ai x = bi where bi < 0 can be replaced by −ai x = −bi . The next step is to
introduce the linear program (Pb) in standard form
maximize − (xn+1 + · · · + xn+m )

subject to A
bxb = b and x b ≥ 0,
where A
b and x
b are given by
 
x1
b =  ...  .

b = A Im ,
A x
 
xn+m
Since we assumed that b ≥ 0, the vector x b = (0n , b) is a feasible solution of (Pb), in fact a basic
feasible solutions since the matrix associated with the indices n + 1, . . . , n + m is the identity
matrix Im . Furthermore, since xi ≥ 0 for all i, the objective function −(xn+1 + · · · + xn+m )
is bounded above by 0.
If we execute the simplex algorithm with a pivot rule that prevents cycling, starting with
the basic feasible solution (0n , d), since the objective function is bounded by 0, the simplex
algorithm terminates with an optimal solution given by some basic feasible solution, say
(u∗ , w∗ ), with u∗ ∈ Rn and w∗ ∈ Rm .
As in the proof of Theorem 6.7, for every feasible solution u ∈ P(A, b) the vector (u, 0m )
is an optimal solution of (Pb). Therefore, if w∗ 6= 0, then P(A, b) = ∅, since otherwise for
every feasible solution u ∈ P(A, b) the vector (u, 0m ) would yield a value of the objective
function −(xn+1 + · · · + xn+m ) equal to 0, but (u∗ , w∗ ) yields a strictly negative value since
w∗ 6= 0.
Otherwise, w∗ = 0, and u∗ is a feasible solution of (P ). Since (u∗ , 0m ) is a basic feasible
solution of (Pb) the columns corresponding to nonzero components of u∗ are linearly inde-
pendent. Some of the coordinates of u∗ could be equal to 0, but since A has rank m we can
add columns of A to obtain a basis K ∗ associated with u∗ , and u∗ is indeed a basic feasible
solution of (P ).
Running the simplex algorithm on the linear program Pb to obtain an initial feasible
solution (u0 , K0 ) of the linear program (P 2) is called Phase I of the simplex algorithm.
Running the simplex algorithm on the linear program (P 2) with some initial feasible solution
7.3. HOW PERFORM A PIVOTING STEP EFFICIENTLY 177
(u0 , K0 ) is called Phase II of the simplex algorithm. If a feasible solution of the linear
program (P 2) is readily available then Phase I is skipped. Sometimes, at the end of Phase
I, an optimal solution of (P 2) is already obtained.
In summary, we proved the following fact worth recording.
Proposition 7.1. For any linear program (P2 )
maximize cx
in standard form, where A is an m × n matrix of rank m and b ≥ 0, consider the linear

program (Pb) in standard form
maximize − (xn+1 + · · · + xn+m )

subject to A
bxb = b and x b ≥ 0.
The simplex algorithm with a pivot rule that prevents cycling started on the basic feasible
b = (0n , b) of (Pb) terminates with an optimal solution (u∗ , w∗ ).
solution x
(1) If w∗ 6= 0, then P(A, B) = ∅, that is, the linear program (P ) has no feasible solution.
(2) If w∗ = 0, then P(A, B) 6= ∅, and u∗ is a basic feasible solution of (P ) associated with
some basis K.
Proposition 7.1 shows that determining whether the polyhedron P(A, b) defined by a
system of equations Ax = b and inequalities x ≥ 0 is nonempty is decidable. This decision
procedure uses a fail-safe version of the simplex algorithm (that prevents cycling), and the
proof that it always terminates and returns an answer is nontrivial.
7.3 How Perform a Pivoting Step Efficiently

We now discuss briefly how to perform the computation of (u+ , K + ) from a basic feasible
solution (u, K).
In order to avoid applying permutation matrices it is preferable to allow a basis K to be
a sequence of indices, possibly out of order. Thus, for any m × n matrix A (with m ≤ n)
and any sequence K = (k1 , k2 , · · · , km ) of m elements with ki ∈ {1, . . . , n}, the matrix AK
denotes the m × m matrix whose ith column is the ki th column of A, and similarly for
any vector u ∈ Rn (resp. any linear form c ∈ (Rn )∗ ) the vector uK ∈ Rm (the linear form
cK ∈ (Rm )∗ ) is the vector whose ith entry is the ki th entry in u (resp. the linear whose ith
entry is the ki th entry in c).
For each nonbasic j ∈
/ K, we have
j
Aj = γkj1 Ak1 + · · · + γkjm Akm = AK γK ,
j j
so the vector γK is given by γK = A−1 j
K A , that is, by solving the system
j
AK γK = Aj . (∗γ )
j
To be very precise, since the vector γK depends on K its components should be denoted by
(γK )ki , but as we said before, to simplify notation we write γkji instead of (γK
j j
)ki .
In order toPdecide which case applies ((A), (B1), (B2), (B3)), we need to compute the
numbers cj − k∈K γkj ck for all j ∈
/ K. For this, observe that
X j j
cj − γk ck = cj − cK γK = cj − cK A−1 j
K A .
k∈K
If we write βK = cK A−1
K , then
X
cj − γkj ck = cj − βK Aj .
k∈K
and we see that βK > >

∈ Rm is the solution of the system βK = (A−1 > >
K ) ck , which means that
>
βK is the solution of the system
A> > >
K β K = cK . (∗β )
Remark: Observe that since u is a basis feasible solution of (P ), we have uj = 0 for all
j∈ / K, so u is the solution of the equation AK uK = b. As a consequence, the value of the
objective function for u is cu = cK uK = cK A−1
K b. This fact will play a crucial role in Section
8.2 to show that when the simplex algorithm terminates with an optimal solution of the
linear program (P ), then it also produces an optimal solution of the dual linear program
(D).
Assume that we have a basic feasible solution u, a basis K for u, and that we also have
the matrix AK as well its inverse A−1
K (perhaps implicitly) and also the inverse (AK )
> −1
of
>
AK (perhaps implicitly). Here is a description of an iteration step of the simplex algorithm,
following almost exactly Chvatal (Chvatal [18], Chapter 7, Box 7.1).
An Iteration Step of the (Revised) Simplex Method
Step 1. Compute the numbers cj − k∈K γkj ck = cj − βK Aj for all j ∈
P
/ K, and for this,
>
compute βK as the solution of the system
A> > >
K β K = cK .
If cj − βK Aj ≤ 0 for all j ∈
/ K, stop and return the optimal solution u (Case (A)).
Step 2. If Case (B) arises, use a pivot rule to determine which index j + ∈
/ K should enter
+
the new basis K + (the condition cj + − βK Aj > 0 should hold).
+
Step 3. Compute maxk∈K γkj . For this, solve the linear system
j + +
AK γK = Aj .
7.3. HOW PERFORM A PIVOTING STEP EFFICIENTLY 179
+
Step 4. If maxk∈K γkj ≤ 0, then stop and report that the linear program (P ) is unbounded
(Case (B1)).
+ + +
Step 5. If maxk∈K γkj > 0, use the ratios uk /γkj for all k ∈ K such that γkj > 0 to
+ +
compute θj , and use a pivot rule to determine which index k − ∈ K such that θj = uk− /γkj+
−
should leave K (Case (B2)).

+ +
If maxk∈K γkj = 0, then use a pivot rule to determine which index k − for which γkj − > 0
should leave the basis K (Case (B3)).
Step 6. Update u, K, and AK , to u+ and K + , and AK + . During this step, given the
basis K specified by the sequence K = (k1 , . . . , k` , . . . , km ), with k − = k` , then K + is the
sequence obtained by replacing k` by the incoming index j + , so K + = (k1 , . . . , j + , . . . , km )
with j + in the `th slot.
The vector u is easily updated. To compute AK + from AK we take advantage that AK
+
and AK + only differ by a single column, namely the `th column Aj , which is given by the
linear combination
+ j+
Aj = AK γK .
j +
To simplify notation, denote γK by γ, and recall that k − = k` . If K = (k1 , . . . , km ), then
k− −
AK = [A · · · A · · · A ], and since AK + is the result of replacing the `th column Ak of
k1 im
+
AK by the column Aj , we have
+
AK + = [Ak1 · · · Aj · · · Aim ] = [Ak1 · · · AK γ · · · Aim ] = AK E(γ),
where E(γ) is the following invertible matrix obtained from the identity matrix Im by re-
placing its `th column by γ:
1 γ1
 
 ... ..
.

 
1 γ`−1
 
 
E(γ) =  γ` .
 
γ`+1 1
 
 

 .
.. . .. 

γm 1
+
Since γ` = γkj − > 0, the matrix E(γ) is invertible, and it is easy to check that its inverse is
given by
1 −γ`−1 γ1
 
 ... ..
.

 
1 −γ`−1 γ`−1
 
 
−1 −1
E(γ) =  γ` ,
 
−γ`−1 γ`+1 1
 
 

 .
.. . .. 

−γ`−1 γm 1
which is very cheap to compute. We also have
A−1 −1 −1
K + = E(γ) AK .
Consequently, if AK and A−1 −1

K are available, then AK + and AK + can be computed cheaply
−1
in terms of AK and AK and matrices of the form E(γ). Then the systems (∗γ ) to find the
j
vectors γK can be solved cheaply.
Since
A> > >
K + = E(γ) AK
and
(A>
K+ )
−1
= (A> −1 > −1
K ) (E(γ) ) ,
the matrices A> >

K + and (AK + )
−1
can also be computed cheaply from A> > −1
K , (AK ) , and matrices
of the form E(γ)> . Thus the systems (∗β ) to find the linear forms βK can also be solved
cheaply.
A matrix of the form E(γ) is called an eta matrix ; see Chvatal [18] (Chapter 7). We
showed that the matrix AK s obtained after s steps of the simplex algorithm can be written
as
AK s = AK s−1 Es
for some eta matrix Es , so Aks can be written as the product
AK s = E1 E2 · · · Es
of s beta matrices. Such a factorization is called an eta factorization. The eta factorization
+
can be used to either invert AK s or to solve a system of the form AKs γ = Aj iteratively.
Which method is more efficient depends on the sparsity of the Ei .
In summary, there are cheap methods for finding the next basic feasible solution (u+ , K + )
from (u, K). We simply wanted to give the reader a flavor of these techniques. We refer the
reader to texts on linear programming for detailed presentations of methods for implementing
efficiently the simplex method. In particular, the revised simplex method is presented in
Chvatal [18], Papadimitriou and Steiglitz [45], Bertsimas and Tsitsiklis [10], and Vanderbei
[64].
7.4 The Simplex Algorithm Using Tableaux

We now describe a formalism for presenting the simplex algorithm, namely (full) tableaux .
This is the traditional formalism used in all books, modulo minor variations. A particularly
nice feature of the tableau formalism is that the update of a tableau can be performed using
elementary row operations identical to the operations used during the reduction of a matrix
to row reduced echelon form (rref). What differs is the criterion for the choice of the pivot.
7.4. THE SIMPLEX ALGORITHM USING TABLEAUX 181
j
Since the quantities cj − cK γK play a crucial role in determining which column Aj should
j
come into the basis, the notation cj is used to denote cj − cK γK , which is called the reduced
cost of the variable xj . The reduced costs actually depend on K so to be very precise we
should denote them by (cK )j , but to simplify notation we write cj instead of (cK )j . We will
see shortly how (cK + )i is computed in terms of in terms of (cK )i .
Observe that the data needed to execute the next step of the simplex algorithm are
(1) The current basic solution uK and its basis K = (k1 , . . . , km ).
j
(2) The reduced costs cj = cj − cK A−1 j
K A = cj − cK γK , for all j ∈
/ K.
j
(3) The vectors γK = (γkji )m
i=1 for all j ∈
j
/ K, that allow us to express each Aj as AK γK .
All this information can be packed into a (m + 1) × (n + 1) matrix called a (full) tableau
organized as follows:
cK uK c1 ··· cj ··· cn
uk1 γ11 ··· γ1j ··· γ1n
.. .. .. ..
. . . .
1 j n
ukm γm ··· γm ··· γm
It is convenient to think as the first row as row 0, and of the first column as column 0.
Row 0 contains the current value of the objective function and the reduced costs, column
0 except for its top entry contains the components of the current basic solution uK , and
j
the remaining columns except for their top entry contain the vectors γK . Observe that
j
the γK corresponding to indices j in K constitute a permutation of the identity matrix
+
Im . The entry γkj − is called the pivot element. A tableau together with the new basis
K + = (K − {k − }) ∪ {j + } contains all the data needed to compute the new uK + , the new
j
γK + , and the new reduced costs (cK + )j .
1 n j
If we define the m × n matrix Γ as the matrix Γ = [γK · · · γK ] whose jth column is γK ,
and c as the row vector c = (c1 · · · cn ), then the above tableau is denoted concisely by
cK u K c
uK Γ
We now show that the update of a tableau can be performed using elementary row
operations identical to the operations used during the reduction of a matrix to row reduced
echelon form (rref).
If K = (k1 , . . . , km ), j + is the index of the incoming basis vector, k − = k` is the index
of the column leaving the basis, and if K + = (k1 , . . . , k`−1 , j + , k`+1 , . . . , km ), since AK + =
j+ j j
AK E(γK ), the new columns γK + are computed in terms of the old columns γK using the
equations
j −1 j j + −1 −1 j j + −1 j
γK + = AK + A = E(γK ) AK A = E(γK ) γK .
Consequently the matrix Γ+ is given in terms of Γ by

j −1 +
Γ+ = E(γK ) Γ.
j −1+
But the matrix E(γK ) is of the form
+ +
−(γkj − )−1 γkj1
 
1
 ... .. 
 . 
 + + 

 1 −(γkj − )−1 γkj`−1 

+
E(γ)−1 = (γkj − )−1 ,
 
+ +
−(γkj − )−1 γkj`+1 1
 
 
..
 
..
.
 
 . 
+ +
−(γkj − )−1 γkjm 1
with the column involving the γs in the `th column, and this matrix is the product of the
following elementary row operations:
+
1. Multiply row ` by 1/γkj − (the inverse of the pivot) to make the entry on row ` and
column j + equal to 1.
+
2. subtract γkji × (the normalized) row ` from row i, for i = 1, . . . , ` − 1, ` + 1, . . . , m.
j +
These are exactly the elementary row operations that reduce the `th column γK of Γ
to the `th column of the identity matrix Im . Thus, this step is identical to the sequence of
steps that the procedure to convert a matrix to row reduced echelon from executes on the
`th column of the matrix. The only difference is the criterion for the choice of the pivot.
Since the new basic solution uK + is given by uK + = A−1
K + b, we have
j −1 −1 + j −1 +
uK + = E(γK ) AK b = E(γK ) uK .
This means that u+ is obtained from uK by applying exactly the same elementary row
operations that were applied to Γ. Consequently, just as in the procedure for reducing a
matrix to rref, we can apply elementary row operations to the matrix [uk Γ], which consists
of rows 1, . . . , m of the tableau.
Once the new matrix Γ+ is obtained, the new reduced costs are given by the following
proposition.
Proposition 7.2. Given any linear program (P 2) is standard form
maximize cx
where A is an m × n matrix of rank m, if (u, K) is a basic (not feasible) solution of (P 2)

and if K + = (K − {k − }) ∪ {j + }, with K = (k1 , . . . , km ) and k − = k` , then for i = 1, . . . , n
we have
i i γki − +
j
ci − cK + γK + = ci − cK γK − + (cj + − cK γK ).
γkj −
Using the reduced cost notation, the above equation is
γki −
(cK + )i = (cK )i − + (cK )j + .
γkj −
Proof. Without any loss of generality and to simplify notation assume that K = (1, . . . , m)
j
and write j for j + and ` for km . Since γK
i
= A−1 i i −1 i
K A , γK + = AK + A , and AK + = AK E(γK ),
we have
i −1 i j −1 −1 i j −1 i
ci − cK + γK + = ci − cK + A + A = ci − cK + E(γK )
K AK A = ci − cK + E(γK ) γK ,
where
−(γ`j )−1 γ1j

 
1
 .. ..
.

 . 
 j −1 j 
 1 −(γ` ) γ`−1

j −1 j −1
 
E(γK ) =
 (γ` ) 

j −1 j

 −(γ` ) γ`+1 1 

.. . . . 
 
 .
−(γ`j )−1 γm j
1
where the `th column contains the γs. Since cK + = (c1 , . . . , c`−1 , cj , c`+1 , . . . , cm ), we have
m
γkj

j −1 cj X
cK + E(γK ) = c1 , . . . , c`−1 , j − ck , c`+1 , . . . , cm ,
γ` k=1,k6=` γ`j
and
γ1i
 
 .. 
 . 
m ! γ i 
j −1 i
cj X γkj  `−1 
cK + E(γK ) γK = c1 . . . c`−1 − c k c`+1 . . . cm  γì 

γ`j k=1,k6=` γ`j  i 
γ`+1 
 . 
 .. 
i
γm
m m
γì
X X
j
= ck γki+ j cj − ck γk
k=1,k6=`
γ` k=1,k6=`
m m
γì
X X
i j j
= ck γk + j cj + c` γ` − ck γk
k=1,k6=`
γ` k=1
m m
γì
X X
i j
= ck γk + j cj − ck γk
k=1
γ` k=1
i γì j
= cK γK + j (cj − cK γK ),
γ`
and thus
i j −1 i i γì j
ci − cK + γK + = ci − cK + E(γK ) γK = ci − cK γK − j (cj − cK γK ),
γ`
as claimed.
Since (γk1− , . . . , γkn− ) is the `th row of Γ, we see that Proposition 7.2 shows that
(cK )j +
cK + = cK − + Γ` , (†)
γkj −
+
where Γ` denotes the `-th row of Γ and γkj − is the pivot. This means that cK + is obtained
by the elementary row operations which consist first normalizing the `th row by dividing it
+
by the pivot γkj − , and then subtracting (cK )j + × the normalized row ` from cK . These are
exactly the row operations that make the reduced cost (cK )j + zero.
Remark: It easy easy to show that we also have
cK + = c − cK + Γ+ .
We saw in section 7.2 that the change in the objective function after a pivoting step
during which column j + comes in and column k − leaves is given by
X j+
j+ +
θ cj + − γk ck = θj (cK )j + ,
k∈K
where
+ uk −
θj = + .
γkj −
If we denote the value of the objective function cK uK by zK , then we see that
(cK )j +
zK + = zK + + uk − .
γkj −
This means that the new value zK + of the objective function is obtained by first normalizing
+
the `th row by dividing it by the pivot γkj − , and then adding (cK )j + × the zeroth entry of
the normalized `th line by (cK )j + to the zeroth entry of line 0.
In updating the reduced costs, we subtract rather than add (cK )j + × the normalized row `
from row 0. This suggests storing −zK as the zeroth entry on line 0 rather than zK , because
then all the entries row 0 are updated by the same elementary row operations. Therefore,
from now on, we use tableau of the form
−cK uK c1 ··· cj ··· cn

uk1 γ11 ··· γ1j ··· γ1n
.. .. .. ..
. . . .
1 j n
The simplex algorithm first chooses the incoming column j + by picking some column for
+
which cj > 0, and then chooses the outgoing column k − by considering the ratios uk /γkj for
+
which γkj > 0 (along column j + ), and picking k − to achieve the minimum of these ratios.
Here is an illustration of the simplex algorithm using elementary row operations on an
example from Papadimitriou and Steiglitz [45] (Section 2.9).
Example 7.4. Consider the linear program
maximize − 2x2 − x4 − 5x7
subject to
x1 + x2 + x3 + x4 = 4
x1 + x 5 = 2
x3 + x 6 = 3
3x2 + x3 + x7 = 6
x1 , x2 , x3 , x4 , x5 , x6 , x7 ≥ 0.
We have the basic feasible solution u = (0, 0, 0, 4, 2, 3, 6), with K = (4, 5, 6, 7). Since cK =
(−1, 0, 0, −5) and c = (0, −2, 0, −1, 0, 0 − 5) the first tableau is
34 1 14 6 0 0 0 0
u4 = 4 1 1 1 1 0 0 0
u5 = 2 1 0 0 0 1 0 0
u6 = 3 0 0 1 0 0 1 0
u7 = 6 0 3 1 0 0 0 1
Row 0 is obtained by subtracting −1× (row 1) and −5× (row 4) from c = (0, −2, 0, −1, 0,
0, −5). Let us pick column j + = 1 as the incoming column. We have the ratios (for positive
entries on column 1)
4/1, 2/1,
and since the minimum is 2, we pick the outgoing column to be column k − = 5. The pivot
1 is indicated in red. The new basis is K = (4, 1, 6, 7). Next we apply row operations to
reduce column 1 to the second vector of the identity matrix I4 . For this, we subtract row 2
from row 1. We get the tableau
34 1 14 6 0 0 0 0
u4 = 2 0 1 1 1 −1 0 0
u1 = 2 1 0 0 0 1 0 0
u6 = 3 0 0 1 0 0 1 0
u7 = 6 0 3 1 0 0 0 1
To compute the new reduced costs, we want to set c1 to 0 so we subtract row 2 from row
0, and we get the tableau
32 0 14 6 0 −1 0 0
u4 = 2 0 1 1 1 −1 0 0
u1 = 2 1 0 0 0 1 0 0
u6 = 3 0 0 1 0 0 1 0
u7 = 6 0 3 1 0 0 0 1
Next, pick column j + = 3 as the incoming column. We have the ratios (for positive
entries on column 3)
2/1, 3/1, 6/1,
1 is indicated in red and the new basis is K = (3, 1, 6, 7). Next we apply row operations to
reduce column 3 to the first vector of the identity matrix I4 . For this, we subtract row 1
from row 3 and from row 4, to obtain the tableau:
32 0 14 6 0 −1 0 0
u3 = 2 0 1 1 1 −1 0 0
u1 = 2 1 0 0 0 1 0 0
u6 = 1 0 −1 0 −1 1 1 0
u7 = 4 0 2 0 −1 1 0 1
To compute the new reduced costs, we want to set c3 to 0 so we subtract 6× row 1 from
row 0, and we get the tableau
20 0 8 0 −6 5 0 0
u3 = 2 0 1 1 1 −1 0 0
u1 = 2 1 0 0 0 1 0 0
u6 = 1 0 −1 0 −1 1 1 0
u7 = 4 0 2 0 −1 1 0 1
Next we pick j + = 2 as the incoming column. We have the ratios (for positive entries on
column 2)
2/1, 4/2,
1 is indicated in red and the new basis is K = (2, 1, 6, 7). Next we apply row operations to
reduce column 2 to the first vector of the identity matrix I4 . For this, we add row 1 to row
3 and subtract 2× row 1 from row 4 to obtain the tableau:
20 0 8 0 −6 5 0 0
u2 = 2 0 1 1 1 −1 0 0
u1 = 2 1 0 0 0 1 0 0
u6 = 3 0 0 1 0 0 1 0
u7 = 0 0 0 −2 −3 3 0 1
row 0 and we get the tableau
4 0 0 −8 −14 13 0 0
u2 =2 0 1 1 1 −1 0 0
u1 =2 1 0 0 0 1 0 0
u6 =3 0 0 1 0 0 1 0
u7 =0 0 0 −2 −3 3 0 1
The only possible incoming column corresponds to j + = 5. We have the ratios (for
positive entries on column 5)
2/1, 0/3,
3 is indicated in red and the new basis is K = (2, 1, 6, 5). Since the minimum is 0, the basis
K = (2, 1, 6, 5) is degenerate (indeed, the component corresponding to the index 5 is 0).
Next we apply row operations to reduce column 5 to the fourth vector of the identity matrix
I4 . For this, we multiply row 4 by 1/3, and then add the normalized row 4 to row 1 and
subtract the normalized row 4 from row 2, and to obtain the tableau:
4 0 0 −8 −14 13 0 0
u2 = 2 0 1 1/3 0 0 0 1/3
u1 = 2 1 0 2/3 1 0 0 −1/3
u6 = 3 0 0 1 0 0 1 0
u5 = 0 0 0 −2/3 −1 1 0 1/3
row 0 and we get the tableau
4 0 0 2/3 −1 0 0 −13/3
u2 = 2 0 1 1/3 0 0 0 1/3
u1 = 2 1 0 2/3 1 0 0 −1/3
u6 = 3 0 0 1 0 0 1 0
u5 = 0 0 0 −2/3 −1 1 0 1/3
The only possible incoming column corresponds to j + = 3. We have the ratios (for
positive entries on column 3)
2/(1/3) = 6, 2/(2/3) = 3, 3/1 = 3,
2/3 is indicated in red and the new basis is K = (2, 3, 6, 5). Next we apply row operations
to reduce column 3 to the second vector of the identity matrix I4 . For this, we multiply row
2 by 2/3, subtract (1/3)× (normalized row 2) from row 1, and subtract normalized row 2
from row 3, add add row (2/3)× (normalized row 2) to row 4, to obtain the tableau:
4 0 0 2/3 −1 0 0 −13/3
u2 = 1 −1/2 1 0 −1/2 0 0 1/2
u3 = 3 3/2 0 1 3/2 0 0 −1/2
u6 = 0 −3/2 0 0 −3/2 0 1 1/2
u5 = 2 1 0 0 0 1 0 0
To compute the new reduced costs, we want to set c3 to 0 so we subtract (2/3)× row 2
from row 0 and we get the tableau
2 −1 0 0 −2 0 0 −4
u2 = 1 −1/2 1 0 −1/2 0 0 1/2
u3 = 3 3/2 0 1 3/2 0 0 −1/2
u6 = 0 −3/2 0 0 −3/2 0 1 1/2
u5 = 2 1 0 0 0 1 0 0
Since all the reduced cost are ≤ 0, we have reached an optimal solution, namely
(0, 1, 3, 0, 2, 0, 0, 0), with optimal value −2.
The progression of the simplex algorithm from one basic feasible solution to another
corresponds to the visit of vertices of the polyhedron P associated with the constraints of
the linear program illustrated in Figure 7.4.
x3
(0,0,3)
6
(1,0,3)
(0,1,3)
x
1 +x
2 +x
3 =4
3
3x 2
(0,2,2)
1
+x3
(2,0,0)
=6
x2
x1 = 2
(0,2,0)
(2,2,0)
4 = 5 x1
Figure 7.4: The polytope P associated with the linear program optimized by the tableau
method. The red arrowed path traces the progression of the simplex method from the origin
to the vertex (0, 1, 3).
As a final comment, if it is necessary to run Phase I of the simplex algorithm, in the event
that the simplex algorithm terminates with an optimal solution (u∗ , 0m ) and a basis K ∗ such
that some ui = 0, then the basis K ∗ contains indices of basic columns Aj corresponding to
slack variables that need to be driven out of the basis. This is easy to achieve by performing a
pivoting step involving some other column j + corresponding to one of the original variables
+
(not a slack variable) for which (γK ∗ )ji 6= 0. In such a step, it doesn’t matter whether
+
(γK ∗ )ji < 0 or (cK ∗ )j + ≤ 0. If the original matrix A has no redundant equations, such a step
is always possible. Otherwise, (γK ∗ )ji = 0 for all non-slack variables, so we detected that the
ith equation is redundant and we can delete it.
Other presentations of the tableau method can be found in Bertsimas and Tsitsiklis [10]
and Papadimitriou and Steiglitz [45].
7.5 Computational Efficiency of the Simplex Method

Let us conclude with a few comments about the efficiency of the simplex algorithm. In
practice, it was observed by Dantzig that for linear programs with m < 50 and m + n < 200,
the simplex algorithms typically requires less than 3m/2 iterations, but at most 3m iterations.
This fact agrees with more recent empirical experiments with much larger programs that
show that the number iterations is bounded by 3m. Thus, it was somewhat of a shock in
1972 when Klee and Minty found a linear program with n variables and n equations for
which the simplex algorithm with Dantzig’s pivot rule requires requires 2n − 1 iterations.
This program (taken from Chvatal [18], page 47) is reproduced below:
n
X
maximize 10n−j xj
j=1
subject to
X i−1
2 i−j
10 xj + xi ≤ 100i−1
j=1
xj ≥ 0,
for i = 1, . . . , n and j = 1, . . . , n.
If p = max(m, n), then , in terms of worse case behavior, for all currently known pivot
rules, the simplex algorithm has exponential complexity in p. However, as we said earlier, in
practice, nasty examples such as the Klee–Minty example seem to be rare, and the number
of iterations appears to be linear in m.
Whether or not a pivot rule (a clairvoyant rule) for which the simplex algorithms runs
in polynomial time in terms of m is still an open problem.
The Hirsch conjecture claims that there is some pivot rule such that the simplex algorithm
finds an optimal solution in O(p) steps. The best bound known so far due to Kalai and
Kleitman is m1+ln n = (2n)ln m . For more on this topic, see Matousek and Gardner [41]
(Section 5.9) and Bertsimas and Tsitsiklis [10] (Section 3.7).
Researchers have investigated the problem of finding upper bounds on the expected
number of pivoting steps if a randomized pivot rule is used. Bounds better than 2m (but of
course, not polynomial) have been found.
7.5. COMPUTATIONAL EFFICIENCY OF THE SIMPLEX METHOD 191
Understanding the complexity of linear programing, in particular of the simplex algo-

rithm, is still ongoing. The interested reader is referred to Matousek and Gardner [41]
(Chapter 5, Section 5.9) for some pointers.
In the next section we consider important theoretical criteria for determining whether a
set of constraints Ax ≤ b and x ≥ 0 has a solution or not.
Chapter 8
Linear Programming and Duality
8.1 Variants of the Farkas Lemma

This section overlaps Section 4.1, but we believe that most readers will not mind if we review
the versions of Farkas lemma that will be used to prove the strong duality theorem of linear
programming. To avoid a clash with the versions of Farkas I and Farkas II from Section 4.1,
we label the versions of Farkas lemma as Farkas Ib and Farkas IIb, even though Farkas IIb
is the same as Farkas IIIb(3) (see Proposition 4.18).
If A is an m × n matrix and if b ∈ Rm is a vector, it is known from linear algebra that
the linear system Ax = b has no solution iff there is some linear form y ∈ (Rm )∗ such that
yA = 0 and yb 6= 0. This means that the linear from y vanishes on the columns A1 , . . . , An
of A but does not vanish on b. Since the linear form y defines the linear hyperplane H
of equation yz = 0 (with z ∈ Rm ), geometrically the equation Ax = b has no solution iff
there is a linear hyperplane H containing A1 , . . . , An and not containing b. This is a kind of
separation theorem that says that the vectors A1 , . . . , An and b can be separated by some
linear hyperplane H.
What we would like to do is to generalize this kind of criterion, first to a system Ax = b
subject to the constraints x ≥ 0, and next to sets of inequality constraints Ax ≤ b and x ≥ 0.
There are indeed such criteria going under the name of Farkas lemma.
The key is a separation result involving polyhedral cones known as the Farkas–Minkowski
proposition. We have the following fundamental separation lemma, which is just a restate-
ment of Proposition 4.15.
Proposition 8.1. Let C ⊆ Rn be a closed nonempty cone. For any point a ∈ Rn , if a ∈
/ C,
then there is a linear hyperplane H (through 0) such that
1. C lies in one of the two half-spaces determined by H.
2. a ∈
/H
3. a lies in the other half-space determined by H.
193
194 CHAPTER 8. LINEAR PROGRAMMING AND DUALITY
We say that H strictly separates C and a.
The Farkas–Minkowski proposition is Proposition 8.1 applied to a polyhedral cone
C = {λ1 a1 + · · · + λn an | λi ≥ 0, i = 1, . . . , n}
where {a1 , . . . , an } is a finite number of vectors ai ∈ Rn . By Proposition 4.13, any polyhedral

cone is closed, so Proposition 8.1 applies and we obtain the following separation lemma.
Proposition 8.2. (Farkas–Minkowski) Let C ⊆ Rn be a nonempty polyhedral cone C =

cone({a1 , . . . , an }). For any point b ∈ Rn , if b ∈
/ C, then there is a linear hyperplane H
(through 0) such that
1. C lies in one of the two half-spaces determined by H.
2. a ∈
/H
3. a lies in the other half-space determined by H.
Equivalently, there is a nonzero linear form y ∈ (Rn )∗ such that
1. yai ≥ 0 for i = 1, . . . , n.
2. yb < 0.
A direct proof of the Farkas–Minkowski proposition not involving Proposition 8.1 is given
at the end of this section.
Remark: There is a generalization of the Farkas–Minkowski proposition applying to infinite

dimensional real Hilbert spaces; see Ciarlet [19], Chapter 9.
Proposition 8.2 implies our first version of Farkas’ lemma.
Proposition 8.3. (Farkas Lemma, Version Ib) Let A be an m × n matrix and let b ∈ Rm be
any vector. The linear system Ax = b has no solution x ≥ 0 iff there is some nonzero linear
form y ∈ (Rm )∗ such that yA ≥ 0>
n and yb < 0.
Proof. First, assume that there is some nonzero linear form y ∈ (Rm )∗ such that yA ≥ 0
and yb < 0. If x ≥ 0 is a solution of Ax = b, then we get
yAx = yb,
but if yA ≥ 0 and x ≥ 0, then yAx ≥ 0, and yet by hypothesis yb < 0, a contradiction.

Next assume that Ax = b has no solution x ≥ 0. This means that b does not belong to
the polyhedral cone C = cone({A1 , . . . , An }) spanned by the columns of A. By Proposition
8.2, there is a nonzero linear form y ∈ (Rm )∗ such that
8.1. VARIANTS OF THE FARKAS LEMMA 195
1. yAj ≥ 0 for j = 1, . . . , n.
2. yb < 0,
which says that yA ≥ 0>

n and yb < 0.
Next consider the solvability of a system of inequalities of the form Ax ≤ b and x ≥ 0.
Proposition 8.4. (Farkas Lemma, Version IIb) Let A be an m × n matrix and let b ∈ Rm
be any vector. The system of inequalities Ax ≤ b has no solution x ≥ 0 iff there is some
nonzero linear form y ∈ (Rm )∗ such that y ≥ 0> >
m , yA ≥ 0n , and yb < 0.
Proof. We use the trick of linear programming which consists of adding “slack variables”
zi to convert inequalities ai x ≤ bi into equations ai x + zi = bi with zi ≥ 0. If we let
z = (z1 , . . . , zm ), it is obvious that the system Ax ≤ b has a solution x ≥ 0 iff the equation

x
A Im =b
z

x
has a solution with x ≥ 0 and z ≥ 0. Now by Farkas Ib, the above system has no
z
solution with with x ≥ 0 and z ≥ 0 iff there is some nonzero linear form y ∈ (Rm )∗ such that
y A Im ≥ 0>

n+m
and yb < 0, that is, yA ≥ 0> >

n , y ≥ 0m , and yb < 0.
In the next section we use Farkas IIb to prove the duality theorem in linear programming.
Observe that by taking the negation of the equivalence in Farkas IIb we obtain a criterion
of solvability, namely:
The system of inequalities Ax ≤ b has a solution x ≥ 0 iff for every nonzero linear form
y ∈ (Rm )∗ such that y ≥ 0> >
m , if yA ≥ 0n , then yb ≥ 0.
We now prove the Farkas–Minkowski proposition without using Proposition 3.3. This
approach uses a basic property of the distance function from a point to a closed set.
Let X ⊆ Rn be any nonempty set and let a ∈ Rn be any point. The distance d(a, X)
from a to X is defined as
d(a, X) = inf ka − xk .
x∈X
Here, k k denotes the Euclidean norm.
Proposition 8.5. Let X ⊆ Rn be any nonempty set and let a ∈ Rn be any point. If X is
closed, then there is some z ∈ X such that ka − zk = d(a, X).
Proof. Since X is nonempty, pick any x0 ∈ X, and let r = ka − x0 k. If Br (a) is the closed
ball Br (a) = {x ∈ Rn | kx − ak ≤ r}, then clearly
d(a, X) = inf ka − xk = inf ka − xk .

x∈X x∈X∩Br (a)
Since Br (a) is compact and X is closed, K = X ∩ Br (a) is also compact. But the function
x 7→ ka − xk defined on the compact set K is continuous, and the image of a compact set
by a continuous function is compact, so by Heine–Borel it has a minimum that is achieved
by some z ∈ K ⊆ X.
Remark: If U is a nonempty, closed and convex subset of a Hilbert space V , a standard

result of Hilbert space theory (the projection theorem) asserts that for any v ∈ V there is a
unique p ∈ U such that
kv − pk = inf kv − uk = d(v, U ),
u∈U
and
hp − v, u − pi ≥ 0 for all u ∈ U .
p
Here kwk = hw, wi, where h−, −i is the inner product of the Hilbert space V .
We can now give a proof of the Farkas–Minkowski proposition (Proposition 8.2).
Proof of the Farkas–Minkowski proposition. Let C = cone({a1 , . . . , am }) be a polyhedral
cone (nonempty) and assume that b ∈ / C. By Proposition 4.13, the polyhedral cone is
closed, and by Proposition 8.5 there is some z ∈ C such that d(b, C) = kb − zk; that is, z is
a point of C closest to b. Since b ∈
/ C and z ∈ C we have u = z − b 6= 0, and we claim that
the linear hyperplane H orthogonal to u does the job, as illustrated in Figure 8.1.
First let us show that
hu, zi = hz − b, zi = 0. (∗1 )
This is trivial if z = 0, so assume z 6= 0. If hu, zi =
6 0, then either hu, zi > 0 or hu, zi < 0. In
either case we show that we can find some point z 0 ∈ C closer to b than z is, a contradiction.
Case 1 : hu, zi > 0.
Let z 0 = (1 − α)z for any α such that 0 < α < 1. Then z 0 ∈ C and since u = z − b
z 0 − b = (1 − α)z − (z − u) = u − αz,
so
2
kz 0 − bk = ku − αzk2 = kuk2 − 2αhu, zi + α2 kzk2 .
If we pick α > 0 such that α < 2hu, zi/ kzk2 , then −2αhu, zi + α2 kzk2 < 0, so kz 0 − bk2 <
kuk2 = kz − bk2 , contradicting the fact that z is a point of C closest to b.
Case 2 : hu, zi < 0.

8.1. VARIANTS OF THE FARKAS LEMMA 197
z
a3 H
C
a2
a1 b
Figure 8.1: The hyperplane H, perpendicular to z − b, separates the point b from C =

cone({a1 , a2 , a3 }).
Let z 0 = (1 + α)z for any α such that α ≥ −1. Then z 0 ∈ C and since u = z − b we have
z 0 − b = (1 + α)z − (z − u) = u + αz so
2
kz 0 − bk = ku + αzk2 = kuk2 + 2αhu, zi + α2 kzk2 ,
and if
0 < α < −2hu, zi/ kzk2 ,
then 2αhu, zi + α2 kzk2 < 0, so kz 0 − bk2 < kuk2 = kz − bk2 , a contradiction as above.
Therefore hu, zi = 0. We have
hu, ui = hu, z − bi = hu, zi − hu, bi = −hu, bi,
and since u 6= 0, we have hu, ui > 0, so hu, ui = −hu, bi implies that
hu, bi < 0. (∗2 )
It remains to prove that hu, ai i ≥ 0 for i = 1, . . . , m. Pick any x ∈ C such that x 6= z.

We claim that
hb − z, x − zi ≤ 0. (∗3 )
Otherwise hb − z, x − zi > 0, that is, hz − b, x − zi < 0, and we show that we can find some
point z 0 ∈ C on the line segment [z, x] closer to b than z is.
For any α such that 0 ≤ α ≤ 1, we have z 0 = (1 − α)z + αx = z + α(x − z) ∈ C, and

since z 0 − b = z − b + α(x − z) we have
2
kz 0 − bk = kz − b + α(x − z)k2 = kz − bk2 + 2αhz − b, x − zi + α2 kx − zk2 ,
so for any α > 0 such that
α < −2hz − b, x − zi/ kx − zk2 ,
we have 2αhz − b, x − zi + α2 kx − zk2 < 0, which implies that kz 0 − bk2 < kz − bk2 , contra-
dicting that z is a point of C closest to b.
Since hb − z, x − zi ≤ 0, u = z − b, and by (∗1 ) hu, zi = 0, we have
0 ≥ hb − z, x − zi = h−u, x − zi = −hu, xi + hu, zi = −hu, xi,
which means that

hu, xi ≥ 0 for all x ∈ C, (∗3 )
as claimed. In particular,
hu, ai i ≥ 0 for i = 1, . . . , m. (∗4 )
Then, by (∗2 ) and (∗4 ), the linear form defined by y = u> satisfies the properties yb < 0 and
yai ≥ 0 for i = 1, . . . , m, which proves the Farkas–Minkowski proposition.
There are other ways of proving the Farkas–Minkowski proposition, for instance using
minimally infeasible systems or Fourier–Motzkin elimination; see Matousek and Gardner [41]
(Chapter 6, Sections 6.6 and 6.7).
8.2 The Duality Theorem in Linear Programming

Let (P ) be the linear program
maximize cx
with A a m × n matrix, and assume that (P ) has a feasible solution and is bounded above.
Since by hypothesis the objective function x 7→ cx is bounded on P(A, b), it might be useful
to deduce an upper bound for cx from the inequalities Ax ≤ b, for any x ∈ P(A, b). We can
do this as follows: for every inequality
ai x ≤ b i 1 ≤ i ≤ m,
pick a nonnegative scalar yi , multiply both sides of the above inequality by yi obtaining
y i ai x ≤ y i b i 1 ≤ i ≤ m,
8.2. THE DUALITY THEOREM IN LINEAR PROGRAMMING 199
(the direction of the inequality is preserved since yi ≥ 0), and then add up these m equations,
which yields
(y1 a1 + · · · + ym am )x ≤ y1 b1 + · · · + ym bm .
If we can pick the yi ≥ 0 such that
c ≤ y1 a1 + · · · + ym am ,
then since xj ≥ 0 we have
cx ≤ (y1 a1 + · · · + ym am )x ≤ y1 b1 + · · · + ym bm ,
namely we found an upper bound of the value cx of the objective function of (P ) for any
feasible solution x ∈ P(A, b). If we let y be the linear form y = (y1 , . . . , ym ), then since
 
a1
A =  ... 
 
am
y1 a1 + · · · + ym am = yA, and y1 b1 + · · · + ym bm = yb, what we did was to look for some

y ∈ (Rm )∗ such that
c ≤ yA, y ≥ 0,
so that we have
cx ≤ yb for all x ∈ P(A, b). (∗)
Then it is natural to look for a “best” value of yb, namely a minimum value, which leads to
the definition of the dual of the linear program (P ), a notion due to John von Neumann.
Definition 8.1. Given any linear program (P )
maximize cx
with A a m × n matrix, the dual (D) of (P ) is the following optimization problem:
minimize yb
subject to yA ≥ c and y ≥ 0,
where y ∈ (Rm )∗ . The original linear program (P ) is called the primal linear program.
Here is an explicit example of a linear program and its dual.

Example 8.1. Consider the linear program illustrated by Figure 8.3
maximize 2x1 + 3x2

subject to
4x1 + 8x2 ≤ 12
2x1 + x2 ≤ 3
3x1 + 2x2 ≤ 4
x1 ≥ 0, x2 ≥ 0.
Its dual linear program is illustrated in Figure8.2
minimize 12y1 + 3y2 + 4y3

subject to
4y1 + 2y2 + 3y3 ≥ 2
8y1 + y2 + 2y3 ≥ 3
y1 ≥ 0, y2 ≥ 0, y3 ≥ 0.
It can be checked that (x1 , x2 ) = (1/2, 5/4) is an optimal solution of the primal linear
program, with the maximum value of the objective function 2x1 + 3x2 equal to 19/4, and
that (y1 , y2 , y3 ) = (5/16, 0, 1/4) is an optimal solution of the dual linear program, with the
minimum value of the objective function 12y1 + 3y2 + 4y3 also equal to 19/4.
3
2x
+y
=3
2
3x
+
2y
=
y 4
1 4x +
8 y=
12
0
0 0.5 1 1.5 2
x
Figure 8.2: The H-polytope for the linear program of Example 8.1. Note x1 → x and x2 → y.
Observe that in the primal linear program (P ), we are looking for a vector x ∈ Rn
maximizing the form cx, and that the constraints are determined by the action of the rows
of the matrix A on x. On the other hand, in the dual linear program (D), we are looking
for a linear form y ∈ (R∗ )m minimizing the form yb, and the constraints are determined by
8x + y + 2z = 3
y
4x + 2y + 3z = 2
Figure 8.3: The H-polyhedron for the dual linear program of Example 8.1 is the spacial
region “above” the pink plane and in “front” of the blue plane. Note y1 → x, y2 → y, and
y3 → z.
the action of y on the columns of A. This is the sense in which (D) is the dual (P ). In most
presentations, the fact that (P ) and (D) perform a search for a solution in spaces that are
dual to each other is obscured by excessive use of transposition.
To convert the dual program (D) to a standard maximization problem we change the
objective function yb to −b> y > and the inequality yA ≥ c to −A> y > ≤ −c> . The dual
linear program (D) is now stated as (D0 )
maximize − b> y >

subject to − A> y > ≤ −c> and y > ≥ 0,
where y ∈ (Rm )∗ . Observe that the dual in maximization form (D00 ) of the dual program
(D0 ) gives back the primal program (P ).
The above discussion established the following inequality known as weak duality.
Proposition 8.6. (Weak Duality) Given any linear program (P )
maximize cx
with A a m × n matrix, for any feasible solution x ∈ Rn of the primal problem (P ) and every
feasible solution y ∈ (Rm )∗ of the dual problem (D), we have
cx ≤ yb.
We say that the dual linear program (D) is bounded below if {yb | y > ∈ P(−A> , −c> )}
is bounded below.
What happens if x∗ is an optimal solution of (P ) and if y ∗ is an optimal solution of (D)?

We have cx∗ ≤ y ∗ b, but is there a “duality gap,” that is, is it possible that cx∗ < y ∗ b?
The answer is no, this is the strong duality theorem. Actually, the strong duality theorem
asserts more than this.
Theorem 8.7. (Strong Duality for Linear Programming) Let (P ) be any linear program
maximize cx
with A a m × n matrix. The primal problem (P ) has a feasible solution and is bounded above
iff the dual problem (D) has a feasible solution and is bounded below. Furthermore, if (P )
has a feasible solution and is bounded above, then for every optimal solution x∗ of (P ) and
every optimal solution y ∗ of (D), we have
cx∗ = y ∗ b.
Proof. If (P ) has a feasible solution and is bounded above then we know from Proposition
6.1 that (P ) has some optimal solution. Let x∗ be any optimal solution of (P ). First we will
show that (D) has a feasible solution v.
Let µ = cx∗ be the maximum of the objective function x 7→ cx. Then for any > 0, the
system of inequalties
Ax ≤ b, x ≥ 0, cx ≥ µ +
has no solution, since otherwise µ would not be the maximum value of the objective function
cx. We would like to apply Farkas IIb, so first we transform the above system of inequalities
into the system
A b
x≤ .
−c −(µ + )
By Proposition 8.3 (Farkas IIb), there is some linear form (λ, z) ∈ (Rm+1 )∗ such that λ ≥ 0,
z ≥ 0,
A
λ z ≥ 0>m,
−c
and
b
λ z < 0,
−(µ + )
which means that
λA − zc ≥ 0>
m, λb − z(µ + ) < 0,
that is,
λA ≥ zc
λb < z(µ + )
λ ≥ 0, z ≥ 0.
On the other hand, since x∗ ≥ 0 is an optimal solution of the system Ax ≤ b, by Farkas IIb
again (by taking the negation of the equivalence), since λA ≥ 0 (for the same λ as before),
we must have
λb ≥ 0. (∗1 )
We claim that z > 0. Otherwize, since z ≥ 0, we must have z = 0, but then
λb < z(µ + )
implies
λb < 0, (∗2 )
and since λb ≥ 0 by (∗1 ), we have a contradiction. Consequently, we can divide by z > 0
without changing the direction of inequalities, and we obtain
λ
A≥c
z
λ
b<µ+
z
λ
≥ 0,
z
which shows that v = λ/z is a feasible solution of the dual problem (D). However, weak
duality (Proposition 8.6) implies that cx∗ = µ ≤ yb for any feasible solution y ≥ 0 of the
dual program (D), so (D) is bounded below and by Proposition 6.1 applied to the version of
(D) written as a maximization problem, we conclude that (D) has some optimal solution.
For any optimal solution y ∗ of (D), since v is a feasible solution of (D) such that vb < µ + ,
we must have
µ ≤ y ∗ b < µ + ,
and since our reasoning is valid for any > 0, we conclude that cx∗ = µ = y ∗ b.
If we assume that the dual program (D) has a feasible solution and is bounded below,
since the dual of (D) is (P ), we conclude that (P ) is also feasible and bounded above.
The strong duality theorem can also be proved by the simplex method, because when
it terminates with an optimal solution of (P ), the final tableau also produces an optimal
solution y of (D) that can be read off the reduced costs of columns n + 1, . . . , n + m by
flipping their signs. We follow the proof in Ciarlet [19] (Chapter 10).
Theorem 8.8. Consider the linear program (P),
maximize cx
its equivalent version (P2 ) in standard form,
maximize b cx
b
subject to Ab b ≥ 0,
bx = b and x
c is a linear form in (Rn+m )∗ , and x

b is an m × (n + m) matrix, b
where A b ∈ Rn+m , given by
   
x1 xn+1
>
 ..   ..  x
A = A Im , b
b c = c 0m , x =  .  , x =  .  , x b= ,
x
xn xn+m
and the dual (D) of (P ) given by
minimize yb
where y ∈ (Rm )∗ . If the simplex algorithm applied to the linear program (P 2) terminates
with an optimal solution (b u∗ , K ∗ ), where u
b∗ is a basic feasible solution and K ∗ is a basis for
b∗ , then y ∗ = b
u b−1∗ is an optimal solution for (D) such that b
cK ∗ A b∗ = y ∗ b. Furthermore, y ∗
cu
K
is given in terms of the reduced costs by y ∗ = −((cK ∗ )n+1 . . . (cK ∗ )n+m ).
Proof. We know that K ∗ is a subset of {1, . . . , n + m} consisting of m indices such that the
b are linearly independent. Let N ∗ = {1, . . . , n + m} − K ∗ . The
corresponding columns of A
simplex methods terminates with an optimal solution in Case (A), namely when
X j
cj −
b ck ≤ 0 for all j ∈ N ∗ ,
γk b
k∈k
bj = P j bk
where A k∈K ∗ γk A , or using the notations of Section 7.3,
cj − b b−1∗ A
cK ∗ A bj ≤ 0 for all j ∈ N ∗ .
b K
The above inequalities can be written as
cN ∗ − b b−1∗ A
cK ∗ A bN ∗ ≤ 0> ,
b K n
or equivalently as
b−1∗ A
cK ∗ A bN ∗ ≥ b
cN ∗ . (∗1 )
b K
The value of the objective function for the optimal solution u b∗ is b

cub∗ = b b∗K ∗ , and since
cK ∗ u
b∗K ∗ satisfies the equation A
u b∗K ∗ = b, the value of the objective function is
bK ∗ u
b∗K ∗ = b
cK ∗ u b−1∗ b.
cK ∗ A (∗2 )
b K
Then if we let y ∗ = b b−1∗ , obviously we have y ∗ b = b

cK ∗ A bK ∗ , so if we can prove that y ∗ is a
cK ∗ u
K
feasible solution of the dual linear program (D), by weak duality, y ∗ is an optimal solution
of (D). We have
y∗AbK ∗ = b b−1∗ A
cK ∗ A bK ∗ = b
cK ∗ , (∗3 )
K
and by (∗1 ) we get
y∗A
bN ∗ = b b−1∗ A
cK ∗ A bN ∗ ≥ b
cN ∗ . (∗4 )
K
Let P be the (n + m) × (n + m) permutation matrix defined so that

A P = A I m P = AK ∗ AN ∗ .
b b b
Then we also have

c P = c 0>

b m P = cK ∗ cN ∗ .
Using the equations (∗3 ) and (∗4 ) we obtain

y∗ A

bK ∗ AbN ∗ ≥ cK ∗ cN ∗ ,
that is,
y ∗ A Im P ≥ c 0>

m P,
which is equivalent to
y ∗ A Im ≥ c 0>

m ,
that is
y ∗ A ≥ c, y ≥ 0,
and these are exactly the conditions that say that y ∗ is a feasible solution of the dual program
(D).
The reduced costs are given by (b ci − b
cK ∗ )i = b b−1∗ A
cK ∗ A bi , for i = 1, . . . , n + m. But for
K
bn+j is the jth vector of the identity matrix Im , so
i = n + 1, . . . , n + m each column A
cK ∗ )n+j = −(b
(b b−1∗ )j = −yj∗
cK ∗ A j = 1, . . . , m,
K
as claimed.
The fact that the above proof is fairly short is deceptive, because this proof relies on the
fact that there are versions of the simplex algorithm using pivot rules that prevent cycling,
but the proof that such pivot rules work correctly is quite lengthy. Other proofs are given
in Matousek and Gardner [41] (Chapter 6, Sections 6.3), Chvatal [18] (Chapter 5), and
Papadimitriou and Steiglitz [45] (Section 2.7).
Observe that since the last m rows of the final tableau are actually obtained by multipling
[u A]
b by Ab−1∗ , the m × m matrix consisting of the last m columns and last m rows of the final
K
b−1∗ (basically, the simplex algorithm has performed the steps of a Gauss–Jordan
tableau is A K
reduction). This fact allows saving some steps in the primal dual method.
By combining weak duality and strong duality, we obtain the following theorem which
shows that exactly four cases arise.
Theorem 8.9. (Duality Theorem of Linear Programming) Let (P ) be any linear program
maximize cx
and let (D) be its dual program
minimize yb
with A a m × n matrix. Then exactly one of the following possibilities occur:
(1) Neither (P ) nor (D) has a feasible solution.
(2) (P ) is unbounded and (D) has no feasible solution.
(3) (P ) has no feasible solution and (D) is unbounded.
(4) Both (P ) and (D) have a feasible solution. Then both have an optimal solution, and
for every optimal solution x∗ of (P ) and every optimal solution y ∗ of (D), we have
cx∗ = y ∗ b.
An interesting corollary of Theorem 8.9 is that there is a test to determine whether a

linear program (P ) has an optimal solution. Indeed, (P ) has an optimal solution iff the
following set of contraints is satisfiable:
Ax ≤ b
yA ≥ c
cx ≥ yb
x ≥ 0, y ≥ 0>
m.
In fact, for any feasible solution (x∗ , y ∗ ) of the above system, x∗ is an optimal solution of
(P ) and y ∗ is an optimal solution of (D)
8.3 Complementary Slackness Conditions

Another useful corollary of the strong duality theorem is the following result known as the
equilibrium theorem.
Theorem 8.10. (Equilibrium Theorem) For any linear program (P ) and its dual linear
program (D) (with set of inequalities Ax ≤ b where A is an m × n matrix, and objective
8.3. COMPLEMENTARY SLACKNESS CONDITIONS 207
function x 7→ cx), for any feasible solution x of (P ) and any feasible solution y of (D), x
and y are optimal solutions iff
for all i for which nj=1 aij xj < bi
P
yi = 0 (∗D )
and Pm
xj = 0 for all j for which i=1 yi aij > cj . (∗P )
Proof.
Pn First, assume that (∗D ) and (∗P ) hold. The equations in (∗D ) say that yi = 0 unless
j=1 aij xj = bi , hence
m
X m
X n
X m X
X n
yb = y i bi = yi aij xj = yi aij xj .
i=1 i=1 j=1 i=1 j=1
Pm
Similarly, the equations in (∗P ) say that xj = 0 unless i=1 yi aij = cj , hence
n
X n X
X m
cx = cj x j = yi aij xj .
j=1 j=1 i=1
Consequently, we obtain
cx = yb.
By weak duality (Proposition 8.6), we have
cx ≤ yb = cx
for all feasible solutions x of (P ), so x is an optimal solution of (P ). Similarly,
yb = cx ≤ yb
for all feasible solutions y of (D), so y is an optimal solution of (D).
Let us now assume that x is an optimal solution of (P ) and that y is an optimal solution
of (D). Then, as in the proof of Proposition 8.6,
n
X m X
X n m
X
cj x j ≤ yi aij xj ≤ y i bi .
j=1 i=1 j=1 i=1
By strong duality, since x and y are optimal solutions the above inequalities are actually
equalities, so in particular we have
Xn m
X
cj − yi aij xj = 0.
j=1 i=1
Since x and y∗ are feasible, xi ≥ 0 and yj ≥ 0, so if m

P
i=1 yi aij > cj , we must have xj = 0.
Similarly, we have
m
X m
X
yi aij xj − bi = 0,
i=1 j=1
Pm
so if j=1 aij xj < bi , then yi = 0.
The equations in (∗D ) and (∗P ) are often called complementary slackness conditions.
These conditions can be exploited to solve for an optimal solution of the primal problem
with the help of the dual problem, and conversely. Indeed, if we guess a solution to one
problem, then we may solve for a solution of the dual using the complementary slackness
conditions, and then check that our guess was correct. This is the essence of the primal-dual
methods. To present this method, first we need to take a closer look at the dual of a linear
program already in standard form.
8.4 Duality for Linear Programs in Standard Form

Let (P ) be a linear program in standard form, where Ax = b for some m × n matrix of rank
m and some objective function x 7→ cx (of course, x ≥ 0). To obtain the dual of (P ) we
convert the equations Ax = b to the following system of inequalities involving a (2m) × n
matrix.
A b
x≤ .
−A −b
Then, if we denote the 2m dual variables by (y 0 , y 00 ), with y 0 , y 00 ∈ (Rm )∗ , the dual of the
above program is
minimize y 0 b − y 00 b

A
subject to 0
y y 00
≥ c and y 0 , y 00 ≥ 0,
−A
where y 0 , y 00 ∈ (Rm )∗ , which is equivalent to
minimize (y 0 − y 00 )b
subject to (y 0 − y 00 )A ≥ c and y 0 , y 00 ≥ 0,
where y 0 , y 00 ∈ (Rm )∗ . If we write y = y 0 − y 00 , we find that the above linear program is

equivalent to the following linear program (D):
minimize yb
subject to yA ≥ c,
where y ∈ (Rm )∗ . Observe that y is not required to be nonnegative; it is arbitrary.

Next, we would like to know what is the version of Theorem 8.8 for a linear program
already in standard form. This is very simple.
Theorem 8.11. Consider the linear program (P2) in standard form
maximize cx
8.4. DUALITY FOR LINEAR PROGRAMS IN STANDARD FORM 209
and its dual (D) given by
minimize yb
where y ∈ (Rm )∗ . If the simplex algorithm applied to the linear program (P 2) terminates
with an optimal solution (u∗ , K ∗ ), where u∗ is a basic feasible solution and K ∗ is a basis for
u∗ , then y ∗ = cK ∗ A−1 ∗ ∗
K ∗ is an optimal solution for (D) such that cu = y b. Furthermore, if
we assume that the simplex algorithm is started with a basic feasible solution (u0 , K0 ) where
K0 = (n − m + 1, . . . , n) (the indices of the last m columns of A) and A(n−m+1,...,n) = Im (the
last m columns of A constitute the identity matrix Im ), then the optimal solution y ∗ = cK ∗ A−1
K∗
for (D) is given in terms of the reduced costs by
y ∗ = c(n−m+1,...,n) − (cK ∗ )(n−m+1,...,n) ,
and the m×m matrix consisting of last m columns and the last m rows of the final tableau
is A−1
K∗ .
Proof. The proof of Theorem 8.8 applies with A instead of A

b and we can show that
cK ∗ A−1
K ∗ A N ∗ ≥ cN ∗ ,
and that y ∗ = cK ∗ A−1 ∗ ∗

K ∗ satisfies, cu = y b, and
y ∗ AK ∗ = cK ∗ A−1
K ∗ A K ∗ = cK ∗ ,
y AN ∗ = cK ∗ A−1
∗
K ∗ AN ∗ ≥ cN ∗ .
Let P be the n × n permutation matrix defined so that

AP = AK ∗ AN ∗ .
Then we also have

cP = cK ∗ cN ∗ ,
and using the above equations and inequalities we obtain
y ∗ AK ∗ AN ∗ ≥ cK ∗ cN ∗ ,

that is, y ∗ AP ≥ cP , which is equivalent to
y ∗ A ≥ c,
which shows that y ∗ is a feasible solution of (D) (remember, y ∗ is arbitrary so there is no

need for the constraint y ∗ ≥ 0).
The reduced costs are given by
(cK ∗ )i = ci − cK ∗ A−1 i
K∗ A ,
and since for j = n − m + 1, . . . , n the column Aj is the (j + m − n)th column of the identity
matrix Im , we have
(cK ∗ )j = cj − (cK ∗ AK ∗ )j+m−n j = n − m + 1, . . . , n,
that is,
y ∗ = c(n−m+1,...,n) − (cK ∗ )(n−m+1,...,n) ,
as claimed. Since the last m rows of the final tableau is obtained by multiplying [u0 A] by
A−1
K ∗ , and the last m columns of A constitute Im , the last m rows and the last m columns of
the final tableau constitute A−1
K∗ .
Let us now take a look at the complementary slackness conditions of Theorem 8.10. If
we go back to the version of (P ) given by
maximize cx

A b
subject to x≤ and x ≥ 0,
−A −b
and to the version of (D) given by
minimize y 0 b − y 00 b

A
subject to 0
y y 00
≥ c and y 0 , y 00 ≥ 0,
−A
where y 0 , y 00 ∈ (Rm )∗ , since the inequalities Ax ≤ b and −Ax ≤ −b together imply that
Ax = b, we have equality for all these inequality constraints, and so the Conditions (∗D )
place no constraints at all on y 0 and y 00 , while the Conditions (∗P ) assert that
for all j for which m 0 00
P
xj = 0 i=1 (yi − yi )aij > cj .
If we write y = y 0 − y 00 , the above conditions are equivalent to

for all j for which m
P
xj = 0 i=1 yi aij > cj .
Thus we have the following version of Theorem 8.10.

Theorem 8.12. (Equilibrium Theorem, Version 2) For any linear program (P 2) in standard
form (with set of equalities Ax ≤ b where A is an m × n matrix, and objective function
x 7→ cx) and its dual linear program (D), for any feasible solution x of (P ) and any feasible
solution y of (D), x and y are optimal solutions iff
for all j for which m
P
xj = 0 i=1 yi aij > cj . (∗P )
Therefore, the slackness conditions applied to a linear program (P 2) in standard form and
to its dual (D) only impose slackness conditions on the variables xj of the primal problem.
The above fact plays a crucial role in the primal-dual method.
8.5. THE DUAL SIMPLEX ALGORITHM 211
8.5 The Dual Simplex Algorithm

Given a linear program (P 2) in standard form
maximize cx
where A is an m × n matrix of rank m, if no obvious feasible solution is available but if c ≤ 0,

then rather than using the method for finding a feasible solution described in Section 7.2 we
may use a method known as the dual simplex algorithm. This method uses basic solutions
(u, K) where Au = b and uj = 0 for all uj ∈ / K, but does not require u ≥ 0, so u may not
−1
be feasible. However, y = cK AK is required to be feasible for the dual program
minimize yb
where y ∈ (R∗ )m . Since c ≤ 0, observe that y = 0>

m is a feasible solution of the dual.
If a basic solution u of (P 2) is found such that u ≥ 0, then cu = yb for y = cK A−1

K ,
and we have found an optimal solution u for (P 2) and y for (D). The dual simplex method
makes progress by attempting to make negative components of u zero and by decreasing the
objective function of the dual program.
The dual simplex method starts with a basic solution (u, K) of Ax = b which is not
feasible but for which y = cK A−1 K is dual feasible. In many cases, the original linear program
is specified by a set of inequalities Ax ≤ b with some bi < 0, so by adding slack variables it is
easy to find such basic solution u, and if in addition c ≤ 0, then because the cost associated
with slack variables is 0, we see that y = 0 is a feasible solution of the dual.
Given a basic solution (u, K) of Ax = b (feasible or not), y = cK A−1 K is dual feasible
iff cK A−1
K A ≥ c, and since c A−1
K K AK = c K , the inequality c A
K K
−1
A ≥ c is equivalent to
−1
cK AK AN ≥ cN , that is,
cN − cK A−1K AN ≤ 0, (∗1 )
where N = {1, . . . , n} − K. Equation (∗1 ) is equivalent to
j
cj − cK γK ≤ 0 for all j ∈ N , (∗2 )
j j
where γK = A−1 j
K A . Recall that the notation cj is used to denote cj − cK γK , which is called
the reduced cost of the variable xj .
As in the simplex algorithm we need to decide which column Ak leaves the basis K and
which column Aj enters the new basis K + , in such a way that y + = cK + A−1 K + is a feasible
−1
solution of (D), that is, cN + − cK + AK + AN + ≤ 0, where N = {1, . . . , n} − K + . We use
+
Proposition 7.2 to decide wich column k − should leave the basis.

Suppose (u, K) is a solution of Ax = b for which y = cK A−1
K is dual feasible.
Case (A). If u ≥ 0, then u is an optimal solution of (P 2).

Case (B). There is some k ∈ K such that uk < 0. In this case, pick some k − ∈ K such
that uk− < 0 (according to some pivot rule).
Case (B1). Suppose that γkj − ≥ 0 for all j ∈/ K (in fact, for all j, since γkj − ∈ {0, 1} for
all j ∈ K). If so, we we claim that (P 2) is not feasible.
Indeed, let v be some basic feasible solution. We have v ≥ 0 and Av = b, that is,
n
X
vj Aj = b,
j=1
j
so by multiplying both sides by A−1 −1 j
K and using the fact that by definition γK = AK A , we
obtain n
X j
vj γK = A−1
K b = uK .
j=1
But recall that by hypothesis uk− < 0, yet vj ≥ 0 and γkj − ≥ 0 for all j, so the component of
index k − is zero or positive on the left, and negative on the right, a contradiction. Therefore,
(P 2) is indeed not feasible.
Case (B2). We have γkj − < 0 for some j.
We pick the column Aj entering the basis among those for which γkj − < 0. Since we
j
assumed that cj − cK γK ≤ 0 for all j ∈ N by (∗2 ), consider
j
cj − cK γK

+ j cj j
µ = max − γk− < 0, j ∈ N = max − γ j γk− < 0, j ∈ N ≤ 0,

γkj − k−
and the set

+ − cj = µ + .

N (µ ) = j ∈ N γkj −
We pick some index j + ∈ N (µ+ ) as the index of the column entering the basis (using
some pivot rule).
i i
Recall that by hypothesis ci − cK γK ≤ 0 for all j ∈ / K and ci − cK γK = 0 for all i ∈ K.
j + j +
i i
Since γk− < 0, for any index i such that γk− ≥ 0, we have −γk− /γk− ≥ 0, and since by
Proposition 7.2
i i γki − j+
ci − cK + γK + = c i − c K Kγ − + (c j + − cK γ
K ),
γkj −
i
we have ci − cK + γK i
+ ≤ 0. For any index i such that γk − < 0, by the choice of j
+
∈ K ∗,
j +
i
ci − cK γK cj + − cK γK
− ≤ − + ,
γki − γkj −
so
i γki − j +
ci − cK γK − + (cj + − cK γK ) ≤ 0,
γkj −
i
and again, ci −cK + γK + ≤ 0. Therefore, if we let K
+
= (K −{k − })∪{j + }, then y + = cK + A−1
K+
is dual feasible. As in the simplex algorithm, θ+ is given by
+
θ+ = uk− /γkj − ≥ 0,
and u+ is also computed as in the simplex algorithm by
 +
j+ j
ui − θ γi
 if i ∈ K
+ +
ui = θj if i = j + .
 +
0 if i ∈
/ K ∪ {j }

The change in the objective function of the prime and dual program (which is the same,
since uK = A−1 −1
K b and y = cK AK is chosen such that cu = cK uK = yb) is the same as in the
simplex algorithm, namely +
+ j j+
θ c − cK γK .
+ j + j + +
We have θ+ > 0 and cj − cK γK ≤ 0, so if cj − cK γK < 0, then the objective function of
the dual program decreases strictly.
Case (B3). µ+ = 0.
+ j +
The possibity that µ+ = 0, that is, cj − cK γK = 0, may arise. In this case, the objective
function doesn’t change. This is a case of degeneracy similar to the degeneracy that arises
in the simplex algorithm. We still pick j + ∈ N (µ+ ), but we need a pivot rule that prevents
cycling. Such rules exist; see Bertsimas and Tsitsiklis [10] (Section 4.5) and Papadimitriou
and Steiglitz [45] (Section 3.6).
The reader surely noticed that the dual simplex algorithm is very similar to the simplex
algorithm, except that the simplex algorithm preserves the property that (u, K) is (primal)
feasible, whereas the dual simplex algorithm preserves the property that y = cK A−1K is dual
feasible. One might then wonder whether the dual simplex algorithm is equivalent to the
simplex algorithm applied to the dual problem. This is indeed the case, there is a one-to-
one correspondence between the dual simplex algorithm and the simplex algorithm applied
to the dual problem. This correspondence is described in Papadimitriou and Steiglitz [45]
(Section 3.7).
The comparison between the simplex algorithm and the dual simplex algorithm is best
illustrated if we use a description of these methods in terms of (full) tableaux .
Recall that a (full) tableau is an (m + 1) × (n + 1) matrix organized as follows:
−cK uK c1 ··· cj ··· cn
uk1 γ11 ··· γ1j ··· γ1n
.. .. .. ..
. . . .
1 j n
The top row contains the current value of the objective function and the reduced costs,
the first column except for its top entry contain the components of the current basic solution
j
uK , and the remaining columns except for their top entry contain the vectors γK . Observe
j
that the γK corresponding to indices j in K constitute a permutation of the identity matrix
Im . A tableau together with the new basis K + = (K − {k − }) ∪ {j + } contains all the data
j i j+
needed to compute the new uK + , the new γK + , and the new reduced costs ci − (γk − /γk − )cj + .
When executing the simplex algorithm, we have uk ≥ 0 for all k ∈ K (and uj = 0 for
/ K), and the incoming column j + is determined by picking one of the column indices
all j ∈
such that cj > 0. Then, the index k − of the leaving column is determined by looking at the
+ +
minimum of the ratios uk /γkj for which γkj > 0 (along column j + ).
On the other hand, when executing the dual simplex algorithm, we have cj ≤ 0 for all
j∈/ K (and ck = 0 for all k ∈ K), and the outgoing column k − is determined by picking one
of the row indices such that uk < 0. The index j + of the incoming column is determined by
looking at the maximum of the ratios −cj /γkj − for which γkj − < 0 (along row k − ).
More details about the comparison between the simplex algorithm and the dual simplex
algorithm can be found in Bertsimas and Tsitsiklis [10] and Papadimitriou and Steiglitz [45].
Here is an example of the the dual simplex method.
Example 8.2. Consider the following linear program in standard form:
Maximize − 4x1 − 2x2 − x3
 
x1
   x2   
−1 −1 2 1 0 0  
 x3  −3
subject to −4 −2 1 0 1 0   = −4 and (x1 , x2 , x3 , x4 , x5 , x6 ) ≥ 0.
 x4 
1 1 −4 0 0 1  
 x5  2
x6
 
0
0
 
0
We initialize the dual simplex procedure with (u, K) where u =  −3 and K = (4, 5, 6).

 
−4
1
The initial tableau, before explicitly calculating the reduced cost, is
0 c1 c2 c3 c4 c5 c6
u4 = −3 −1 −1 2 1 0 0
.
u5 = −4 −4 −2 1 0 1 0
u6 = 2 1 1 −4 0 0 1
Since u has negative coordinates, Case (B) applies, and we will set k − = 4. We must now
determine whether Case (B1) or Case (B2) applies. This determination is accomplished by
scanning the first three columns in the tableau, and observing each column has a negative
entry. Thus Case (B2) is applicable, and we need to determine the reduced costs. Observe
that c = (−4, −2, −1, 0, 0, 0), which in turn implies c(4,5,6) = (0, 0, 0). Equation (∗2 ) implies
that the nonzero reduced costs are
 
−1
c1 = c1 − c(4,5,6) −4 = −4

1
 
−1
c2 = c2 − c(4,5,6) −2 = −2

1
 
−2
c3 = c3 − c(4,5,6) 1  = −1,

4
and our tableau becomes
0 −4 −2 −1 0 0 0
u4 = −3 −1 −1 2 1 0 0
.
u5 = −4 −4 −2 1 0 1 0
u6 = 2 1 1 −4 0 0 1
Since k − = 4, our pivot row is the first row of the tableau. To determine candidates for j + ,
we scan this row, locate negative entries and compute

−2 −4

+ cj j
µ = max − j γ4 < 0, j ∈ {1, 2, 3} = max , = −2.
γ4 1 1
Since µ+ occurs when j = 2, we set j + = 2. Our new basis is K + = (2, 5, 6). We must
normalize the first row of the tableau, namely multiply by −1, then add twice this normalized
row to the second row, and subtract the normalized row from the third row to obtain the
updated tableau.
0 −4 −2 −1 0 0 0
u2 = 3 1 1 −2 −1 0 0
u5 = 2 −2 0 −3 −2 1 0
u6 = −1 0 0 −2 1 0 1
It remains to update the reduced costs and the value of the objective function by adding
twice the normalized row to the top row.
6 −2 0 −5 −2 0 0
u2 = 3 1 1 −2 −1 0 0
u5 = 2 −2 0 −3 −2 1 0
u6 = −1 0 0 −2 1 0 1
We now repeat the procedure of Case (B2) and set k − = 6 (since this is the only negative
entry of u+ ). Our pivot row is now the third row of the updated tableaux, and the new µ+
becomes
−5

+ cj j 5
µ = max − j γ6 < 0, j ∈ {1, 3, 4} = max =− ,
γ6 2 2
which implies that j + = 3. Hence the new basis is K + = (2, 5, 3), and we update the tableau
by taking − 12 of Row 3, adding twice the normalized Row 3 to Row 1, and adding three
times the normalized Row 3 to Row 2.
6 −2 0 −5 −2 0 0
u2 = 4 1 1 0 −2 0 −1
u5 = 7/2 −2 0 0 −7/2 1 −3/2
u3 = 1/2 0 0 1 −1/2 0 −1/2
It remains to update the objective function and the reduced costs by adding five times the
normalized row to the top row.
17/2 −2 0 0 −9/2 0 −5/2
u2 = 4 1 1 0 −2 0 −1
u5 = 7/2 −2 0 0 − 27 1 −3/2
u3 = 1/2 0 0 1 −1/2 0 −1/2
Since u+ has no negative entries, the dual simplex method terminates and objective function
4x1 − 2x2 − x3 is maximized with − 17 2
at (0, 4, 12 ).
8.6 The Primal-Dual Algorithm

Let (P 2) be a linear program in standard form
maximize cx
where A is an m × n matrix of rank m, and (D) be its dual given by
minimize yb
where y ∈ (Rm )∗ .
Pn
First, we may assume that b ≥ 0 by changing every equation j=1 aij xj = bi with bi < 0
to nj=1 −aij xj = −bi . If we happen to have some feasible solution y of the dual program
P
(D), we know from Theorem 8.12 that a feasible solution x of (P 2) is an optimal solution iff
the equations in (∗P ) hold. If we denote by J the subset of {1, . . . , n} for which the equalities
yAj = cj
8.6. THE PRIMAL-DUAL ALGORITHM 217
hold, then by Theorem 8.12 a feasible solution x of (P 2) is an optimal solution iff
xj = 0 for all j ∈
/ J.
Let |J| = p and N = {1, . . . , n} − J. The above suggests looking for x ∈ Rn such that
X
xj Aj = b
j∈J
xj ≥ 0 for all j ∈ J
xj = 0 for all j ∈
/ J,
or equivalently
AJ xJ = b, xJ ≥ 0, (∗1 )
and
xN = 0n−p .
To search for such an x, and just need to look for a feasible xJ , and for this we can use
the restricted primal linear program (RP ) defined as follows:
maximize − (ξ1 + · · · + ξm )

xJ
subject to AJ Im = b and x, ξ ≥ 0.
ξ
Since by hypothesis b ≥ 0 and the objective function is bounded above by 0, this linear
program has an optimal solution (x∗J , ξ ∗ ).
If ξ ∗ = 0, then the vector u∗ ∈ Rn given by u∗J = x∗J and u∗N = 0n−p is an optimal solution
of (P ).
Otherwise, ξ ∗ > 0 and we have failed to solve (∗1 ). However we may try to use ξ ∗ to
improve y. For this, consider the dual (DRP ) of (RP ):
minimize zb
subject to zAJ ≥ 0
z ≥ −1>
m.
Observe that the program (DRP ) has the same objective function as the original dual
program (D). We know by Theorem 8.11 that the optimal solution (x∗J , ξ ∗ ) of (RP ) yields
an optimal solution z ∗ of (DRP ) such that
z ∗ b = −(ξ1∗ + · · · + ξm
∗
) < 0.
In fact, if K ∗ is the basis associated with (x∗J , ξ ∗ ) and if we write

Ab = AJ Im
c = [0>
and b >
p − 1 ], then by Theorem 8.11 we have
z∗ = b b−1∗ = −1>
cK ∗ A m − (cK ∗ )(p+1,...,p+m) ,
K
where (cK ∗ )(p+1,...,p+m) denotes the row vector of reduced costs in the final tableau corre-
sponding to the last m columns.
If we write
y(θ) = y + θz ∗ ,
then the new value of the objective function of (D) is
y(θ)b = yb + θz ∗ b, (∗2 )
and since z ∗ b < 0, we have a chance of improving the objective function of (D), that is,
decreasing its value for θ > 0 small enough if y(θ) is feasible for (D). This will be the case
iff y(θ)A ≥ c iff
yA + θz ∗ A ≥ c. (∗3 )
Now since y is a feasible solution of (D) we have yA ≥ c, so if z ∗ A ≥ 0 then (∗3 ) is satisfied
and y(θ) is a solution of (D) for all θ > 0, which means that (D) is unbounded. But this
implies that (P ) is not feasible.
Let us take a closer look at the inequalities z ∗ A ≥ 0. For j ∈ J, Since z ∗ is an optimal
solution of (DRP ), we know that z ∗ AJ ≥ 0, so if z ∗ Aj ≥ 0 for all j ∈ N , then (P ) is not
feasible.
Otherwise, there is some j ∈ N = {1, . . . , n} − J such that
z ∗ Aj < 0,
and then since by the definition of J we have yAj > cj for all j ∈ N , if we pick θ > 0 such
that
yAj − cj
θ≤ j ∈ N, z ∗ Aj < 0,
−z ∗ Aj
then we decrease the objective function y(θ)b = yb + θz ∗ b of (D) (since z ∗ b < 0). Therefore
we pick the best θ, namely
j
yA − cj

+ ∗ j
θ = min j∈
/ J, z A < 0 > 0. (∗4 )
−z ∗ Aj
Next, we update y to y + = y(θ+ ) = y + θ+ z ∗ , we create the new restricted primal with

the new subset
J + = {j ∈ {1, . . . , n} | y + Aj = cj },
and repeat the process. Here are the steps of the primal-dual algorithm.
Step 1. Find some feasible solution y of the dual program (D). We will show later
that this is always possible.
Step 2. Compute
J + = {j ∈ {1, . . . , n} | yAj = cj }.
Step 3. Set J = J + and solve the problem (RP ) using the simplex algorithm, starting
from the optimal solution determined during the previous round, obtaining the
optimal solution (x∗J , ξ ∗ ) with the basis K ∗ .
Step 4.
If ξ ∗ = 0, then stop with an optimal solution u∗ for (P ) such that u∗J = x∗J and the
other components of u∗ are zero.
Else let
z ∗ = −1>
m − (cK ∗ )(p+1,...,p+m) ,
be the optimal solution of (DRP ) corresponding to (x∗J , ξ ∗ ) and the basis K ∗ .

If z ∗ Aj ≥ 0 for all j ∈
/ J, then stop; the program (P ) has no feasible solution.
Else compute
yAj − cj

∗
+
y + = y + θ+ z ∗ ,
j
θ = min − ∗ j j ∈
/ J, z A < 0 ,
z A
and
J + = {j ∈ {1, . . . , n} | y + Aj = cj }.
Go back to Step 3.
The following proposition shows that at each iteration we can start the program (RP )
with the optimal solution obtained at the previous iteration.
Proposition 8.13. Every j ∈ J such that Aj is in the basis of the optimal solution ξ ∗
belongs to the next index set J + .
Proof. Such an index j ∈ J correspond to a variable ξj such that ξj > 0, so by complementary
slackness, the constraint z ∗ Aj ≥ 0 of the dual program (DRP ) must be an equality, that is,
z ∗ Aj = 0. But then, we have
y + Aj = yAj + θ+ z ∗ Aj = cj ,
which shows that j ∈ J + .
If (u∗ , ξ ∗ ) with the basis K ∗ is the optimal solution of the program (RP ), Proposition
8.13 together with the last property of Theorem 8.11 allows us to restart the (RP ) in Step 3
with (u∗ , ξ ∗ )K ∗ as initial solution (with basis K ∗ ). For every j ∈ J − J + , column j is deleted,
and for every j ∈ J + − J, the new column Aj is computed by multiplying A b−1∗ and Aj , but
K
Ab−1∗ is the matrix Γ∗ [1:m; p + 1:p + m] consisting of the last m columns of Γ∗ in the final
K
tableau, and the new reduced cj is given by cj − z ∗ Aj . Reusing the optimal solution of the
previous (RP ) may improve efficiency significantly.
Another crucial observation is that for any index j0 ∈ N such that
θ+ = (yAj0 − cj0 )/(−z ∗ Aj0 ), we have
y + Aj0 = yAj0 + θ+ z ∗ Aj0 = cj0 ,
and so j0 ∈ J + . This fact that be used to ensure that the primal-dual algorithm terminates
in a finite number of steps (using a pivot rule that prevents cycling); see Papadimitriou and
Steiglitz [45] (Theorem 5.4).
It remains to discuss how to pick some initial feasible solution y of the dual program (D).
If cj ≤ 0 for j = 1, . . . , n, then we can pick y = 0.
We should note that in many applications, the natural primal optimization problem
is actually the minimization some objective function cx = c1 x1 + · · · + cn xn , rather its
maximization. For example, many of the optimization problems considered in Papadimitriou
and Steiglitz [45] are minimization problems.
Of course, minimizing cx is equivalent to maximizing −cx, so our presentation covers
minimization too. But if we are dealing with a minimization problem, the weight cj are
often nonnegative, so from the point of view of maximization we will have −cj ≤ 0 for all j,
and we will be able to use y = 0 as a starting point.
Going back to our primal problem in maximization form and its dual in minimization
form, we still need to deal with the situation where cj > 0 for some j, in which case there
may not be any obvious y feasible for (D). Preferably we would like to find such a y very
cheaply.
There is a trick to deal with this situation. We pick some very large positive number M
and add to the set of equations Ax = b the new equation
x1 + · · · + xn + xn+1 = M,
with the new variable xn+1 constrained to be nonnegative. If the program (P ) has a fea-
sible solution, such an M exists. In fact, it can shown that for any basic feasible solution
u = (u1 , . . . , un ), each |ui | is bounded by some expression depending only on A and b; see
Papadimitriou and Steiglitz [45] (Lemma 2.1). The proof is not difficult and relies on the fact
that the inverse of a matrix can be expressed in terms of certain determinants (the adjugates).
Unfortunately, this bound contains m! as a factor, which makes it quite impractical.
Having added the new equation above, we obtain the new set of equations

A 0n x b
> = ,
1n 1 xn+1 M
with x ≥ 0, xn+1 ≥ 0, and the new objective function given by

x
c 0 = cx.
xn+1
The dual of the above linear program is
minimize yb + ym+1 M
subject to yAj + ym+1 ≥ cj j = 1, . . . , n
ym+1 ≥ 0.
If cj > 0 for some j, observe that the linear form ye given by
(
0 if 1 ≤ i ≤ m
yei =
max1≤j≤n {cj } > 0
is a feasible solution of the new dual program. In practice, we can choose M to be a number
close to the largest integer representable on the computer being used.
Here is an example of the primal-dual algorithm given in the Math 588 class notes of T.
Molla.
Example 8.3. Consider the following linear program in standard form:
Maximize − x1 − 3x2 − 3x3 − x4

 
  x1  
3 4 −3 1   2
subject to 3 −2 6 −1 x2  = 1 and x1 , x2 , x3 , x4 ≥ 0.
 x3 
6 4 0 1 4
x4
The associated dual program (D) is
Minimize 2y1 + y2 + 4y3
 
  −1
3 4 −3 1 −3
subject to y1 y2 y3 3 −2 6 −1 ≥ 
−3 .

6 4 0 1
−1
We initialize the primal-dual algorithm with the dual feasible point y = (−1/3 0 0).
Observe that only the first inequality of (D) is actually an equality, and hence J = {1}. We
form the restricted primal program (RP 1)
Maximize − (ξ1 + ξ2 + ξ3 )
 
  x1  
3 1 0 0   2
subject to 3 0 1 0  ξ1  = 1 and x1 , ξ1 , ξ2 , ξ3 ≥ 0.
 ξ2 
6 0 0 1 4
ξ3
We now solve (RP 1) via the simplex algorithm. The initial tableau with K = (2, 3, 4) and
J = {1} is
x1 ξ1 ξ2 ξ3
7 12 0 0 0
ξ1 = 2 3 1 0 0 .
ξ2 = 1 3 0 1 0
ξ3 = 4 6 0 0 1
For (RP 1), c = (0, −1, −1, −1), (x1 , ξ1 , ξ2 , ξ3 ) = (0, 2, 1, 4), and the nonzero reduced cost is
given by  
3
0 − (−1 − 1 − 1) 3 = 12.
6
Since there is only one nonzero reduced cost, we must set j + = 1. Since
min{ξ1 /3, ξ2 /3, ξ3 /6} = 1/3, we see that k − = 3 and K = (2, 1, 4). Hence we pivot through
the red circled 3 (namely we divide row 2 by 3, and then subtract 3× (row 2) from row 1,
6× (row 2) from row 3, and 12× (row 2) from row 0), to obtain the tableau
x1 ξ1 ξ2 ξ3
3 0 0 −4 0
ξ1 = 1 0 1 −1 0 .
x1 = 1/3 1 0 1/3 0
ξ3 = 2 0 0 −2 1
At this stage the simplex algorithm for (RP 1) terminates since there are no positive reduced
costs. Since the upper left corner of the final tableau is not zero, we proceed with Step 4 of
the primal dual algorithm and compute
z ∗ = (−1 − 1 − 1) − (0 − 4 0) = (−1 3 − 1),

   
4 4
5
(−1/3 0 0) −2 + 3 = ,
  −(−1 3 − 1) −2 = 14,

3
4 4
   
1 1
2
(−1/3 0 0) −1 + 1 = , −(−1 3 − 1) −1 = 5,
3
1 1
so
+ 5 2 5
θ = min , = ,
42 15 42
and we conclude that the new feasible solution for (D) is
5
y + = (−1/3 0 0) + (−1 3 − 1) = (−19/42 5/14 − 5/42).
42
When we substitute y + into (D), we discover that the first two constraints are equalities,
and that the new J is J = {1, 2}. The new reduced primal (RP 2) is
 
  x1  
3 4 1 0 0  x2 
 2
subject to 3 −2 0 1 0  ξ1  = 1 and x1 , x2 , ξ1 , ξ2 , ξ3 ≥ 0.
 
6 4 0 0 1  ξ2  4
ξ3
Once again, we solve (RP 2) via the simplex algorithm, where c = (0, 0, −1, −1, −1), (x1 , x2 ,
ξ1 , ξ2 , ξ3 ) = (1/3, 0, 1, 0, 2) and K = (3, 1, 5). The initial tableau is obtained from the final
tableau of the previous (RP 1) by adding a column corresponding the the variable x2 , namely
    
1 −1 0 4 6
Ab−1 A2 = 0 1/3 0 −2 = −2/3 ,
K
0 −2 1 4 8
with  
4
c2 = c2 − z ∗ A2 = 0 − −1 3 −1 −2 = 14,

4
and we get
x1 x2 ξ1 ξ2 ξ3
3 0 14 0 −4 0
ξ1 = 1 0 6 1 −1 0 .
x1 = 1/3 1 −2/3 0 1/3 0
ξ3 = 2 0 8 0 −2 1
Note that j + = 2 since the only positive reduced cost occurs in column 2. Also observe
that since min{ξ1 /6, ξ3 /8} = ξ1 /6 = 1/6, we set k − = 3, K = (2, 1, 5) and pivot along the
red 6 to obtain the tableau
x1 x2 ξ1 ξ2 ξ3
2/3 0 0 −7/3 −5/3 0
x2 = 1/6 0 1 1/6 −1/6 0 .
x1 = 4/9 1 0 1/9 2/9 0
ξ3 = 2/3 0 0 −4/3 −2/3 1
Since the reduced costs are either zero or negative the simplex algorithm terminates, and
we compute
z ∗ = (−1 − 1 − 1) − (−7/3 − 5/3 0) = (4/3 2/3 − 1),
   
1 1
(−19/42 5/14 − 5/42) −1 + 1 = 1/14, −(4/3 2/3 − 1) −1 = 1/3,
1 1
so
3
θ+ = ,
14
5
y + = (−19/42 5/14 − 5/42) + (4/3 2/3 − 1) = (−1/6 1/2 − 1/3).
14
When we plug y + into (D), we discover that the first, second, and fourth constraints are
equalities, which implies J = {1, 2, 4}. Hence the new restricted primal (RP 3) is
 
x1
   x2   
3 4 1 1 0 0  
 x4  2
subject to 3 −2 −1 0 1 0   = 1 and x1 , x2 , x4 , ξ1 , ξ2 , ξ3 ≥ 0.
 ξ1 
6 4 1 0 0 1  
 ξ2  4
ξ3
The initial tableau for (RP 3), with c = (0, 0, 0, −1, −1, −1), (x1 , x2 , x4 , ξ1 , ξ2 , ξ3 ) = (4/9, 1/6,
0, 0, 0, 2/3) and K = (2, 1, 6), is obtained from the final tableau of the previous (RP 2) by
adding a column corresponding the the variable x4 , namely
    
1/6 −1/6 0 1 1/3
b−1 A4 =  1/9
A 2/9 0 −1 = −1/9 ,
K
−4/3 −2/3 1 1 1/3
with  
1
∗ 4

c4 = c4 − z A = 0 − 4/3 2/3 −1 −1 = 1/3,

1
and we get
x1 x2 x4 ξ1 ξ2 ξ3
2/3 0 0 1/3 −7/3 −5/3 0
x2 = 1/6 0 1 1/3 1/6 −1/6 0 .
x1 = 4/9 1 0 −1/9 1/9 2/9 0
ξ3 = 2/3 0 0 1/3 −4/3 −2/3 1
Since the only positive reduced cost occurs in column 3, we set j + = 3. Furthermore
since min{x2 /(1/3), ξ3 /(1/3)} = x2 /(1/3) = 1/2, we let k − = 2, K = (3, 1, 6), and pivot
around the red circled 1/3 to obtain
x1 x2 x4 ξ1 ξ2 ξ3
1/2 0 −1 0 −5/2 −3/2 0
x4 = 1/2 0 3 1 1/2 −1/2 0 .
x1 = 1/2 1 1/3 0 1/6 1/6 0
ξ3 = 1/2 0 −1 0 −3/2 −1/2 1
At this stage, there are no positive reduced costs, and we must compute
z ∗ = (−1 − 1 − 1) − (−5/2 − 3/2 0) = (3/2 1/2 − 1),

   
−3 −3
(−1/6 1/2 − 1/3)  6  + 3 = 13/2, −(3/2 1/2 − 1)  6  = 3/2,
0 0
so
13
θ+ = ,
3
13
y + = (−1/6 1/2 − 1/3) + (3/2 1/2 − 1) = (19/3 8/3 − 14/3).
3
We plug y + into (D) and discover that the first, third, and fourth constraints are equalities.
Thus, J = {1, 3, 4} and the restricted primal (RP 4) is
 
x1
  x3   
3 −3 1 1 0 0 x4 
 2
subject to 3 6 −1 0 1 0   = 1 and x1 , x3 , x4 , ξ1 , ξ2 , ξ3 ≥ 0.
 ξ1 
6 0 1 0 0 1  
 ξ2  4
ξ3
The initial tableau for (RP 4), with c = (0, 0, 0, −1, −1, −1), (x1 , x3 , x4 , ξ1 , ξ2 , ξ3 ) = (1/2,
0, 1/2, 0, 0, 1/2) and K = (3, 1, 6) is obtained from the final tableau of the previous (RP 3)
by replacing the column corresponding to the variable x2 by a column corresponding to the
variable x3 , namely
    
1/2 −1/2 0 −3 −9/2
b−1 A3 =  1/6
A 1/6 0  6  =  1/2  ,
K
−3/2 −1/2 1 0 3/2
with  
−3
c3 = c3 − z ∗ A3 = 0 − 3/2 1/2 −1  6  = 3/2,

0
and we get
x1 x3 x4 ξ1 ξ2 ξ3
1/2 0 3/2 0 −5/2 −3/2 0
x4 = 1/2 0 −9/2 1 1/2 −1/2 0
.
x1 = 1/2 1 1/2 0 1/6 1/6 0
ξ3 = 1/2 0 3/2 0 −3/2 −1/2 1
By analyzing the top row of reduced cost, we see that j + = 2. Furthermore, since
min{x1 /(1/2), ξ3 /(3/2)} = ξ3 /(3/2) = 1/3, we let k − = 6, K = (3, 1, 2), and pivot along the
red circled 3/2 to obtain
x1 x3 x4 ξ1 ξ2 ξ3
0 0 0 0 −1 −1 −1
x4 = 2 0 0 1 −4 −2 3 .
x1 = 1/3 1 0 0 2/3 1/3 −1/3
x3 = 1/3 0 1 0 −1 −1/3 2/3
Since the upper left corner of the final tableau is zero and the reduced costs are all ≤ 0,
we are finally finished. Then y = (19/3 8/3 − 14/3) is an optimal solution of (D), but more
importantly (x1 , x2 , x3 , x4 ) = (1/3, 0, 1/3, 2) is an optimal solution for our original linear
program and provides an optimal value of −10/3.
The primal-dual algorithm for linear programming doesn’t seem to be the favorite method
to solve linear programs nowadays. But it is important because its basic principle, to use
a restricted (simpler) primal problem involving an objective function with fixed weights,
namely 1, and the dual problem to provide feedback to the primal by improving the ob-
jective function of the dual, has led to a whole class of combinatorial algorithms (often
approximation algorithms) based on the primal-dual paradigm. The reader will get a taste
of this kind of algorithm by consulting Papadimitriou and Steiglitz [45], where it is explained
how classical algorithms such as Dijkstra’s algorithm for the shortest path problem, and Ford
and Fulkerson’s algorithm for max flow can be derived from the primal-dual paradigm.
Chapter 9
Basics of Combinatorial Topology
In order to study and manipulate complex shapes it is convenient to discretize these shapes
and to view them as the union of simple building blocks glued together in a “clean fashion.”
The building blocks should be simple geometric objects, for example, points, lines segments,
triangles, tetrahedra and more generally simplices, or even convex polytopes. We will begin
by using simplices as building blocks.
The material presented in this chapter consists of the most basic notions of combinatorial
topology, going back roughly to the 1900-1930 period and it is covered in nearly every alge-
braic topology book (certainly the “classics”). A classic text (slightly old fashion especially
for the notation and terminology) is Alexandrov [1], Volume 1 and another more “modern”
source is Munkres [43]. An excellent treatment from the point of view of computational
geometry can be found is Boissonnat and Yvinec [12], especially Chapters 7 and 10. An-
other fascinating book covering a lot of the basics but devoted mostly to three-dimensional
topology and geometry is Thurston [61].
One of the main goals of this chapter is to define a discrete (combinatorial) analog of
the notion of a topological manifold (with or without boundary). The key for doing this is
to define a combinatorial notion of nonsingularity of a face, and technically this is achieved
by defining the notions of star and link of a face. There are actually two variants of the
notion of star: closed stars and open stars. It turns out that the notion of nonsingularity is
captured well by defining a face to be nonsingular if its link is homeomorphic to a sphere or
to a closed ball. It is intuitively clear that if every face is nonsingular then the open star of
every face is a “nice” open set, either an open ball or the intersection of an open ball with
a half space.
However, proving this fact rigorously takes a surprising amount of work and requires the
introduction of new concepts such as the suspension of a complex and the join of complexes.
Once again, our geometric intuition in dimension greater than three is very unreliable, and
we have to resort to algebraic arguments involving induction to be on solid grounds.
227
228 CHAPTER 9. BASICS OF COMBINATORIAL TOPOLOGY
9.1 Simplicial Complexes

Recall that a simplex is just the convex hull of a finite number of affinely independent points.
We also need to define faces, the boundary, and the interior of a simplex.
Definition 9.1. Let E be any normed affine space, say E = Em with its usual Euclidean
norm. Given any n+1 affinely independent points a0 , . . . , an in E, the n-simplex (or simplex)
σ defined by a0 , . . . , an is the convex hull of the points a0 , . . . , an , that is, the set of all convex
combinations λ0 a0 + · · · + λn an , where λ0 + · · · + λn = 1 and λi ≥ 0 for all i, 0 ≤ i ≤ n;
the simplex σ is often denoted by (a0 , . . . , an ). We call n the dimension of the n-simplex σ,
and the points a0 , . . . , an are the vertices of σ; we denote the set of vertices {a0 , . . . , an } by
vert(σ). Given any subset {ai0 , . . . , aik } of {a0 , . . . , an } (where 0 ≤ k ≤ n), the k-simplex
generated by ai0 , . . . , aik is called a k-face or simply a face of σ. A face s of σ is a proper
face if s 6= σ (we agree that the empty set is a face of any simplex). For any vertex ai , the
face generated by a0 , . . . , ai−1 , ai+1 , . . . , an (i.e., omitting ai ) is called the face opposite ai .
Every face that is an (n − 1)-simplex is called a boundary face or facet. The union of the
boundary faces is the boundary of σ, denoted by ∂σ, and the complement of ∂σ in σ is the
interior Int σ = σ − ∂σ of σ. The interior Int σ of σ is sometimes called an open simplex .
See Figure 9.1.
a0
0 - simplex
a0 a1 a0 a1
interior
1 - simplex boundary
a2 a2
a0 a1 a0 a1
interior
boundary
2 - simplex
a2
a2
a3
a3
a0
a0 a1
a1 interior boundary
3 -simplex
Figure 9.1: Illustrations of low-dimensional simplicies in E3 , along with their corresponding

interior and boundaries. The boundary of the 3-simplex (solid tetrahedron) consist of four
triangles.
9.1. SIMPLICIAL COMPLEXES 229
It should be noted that for a 0-simplex consisting of a single point {a0 }, ∂{a0 } = ∅, and
Int {a0 } = {a0 }. Of course, a 0-simplex is a single point, a 1-simplex is the line segment
(a0 , a1 ), a 2-simplex is a triangle (a0 , a1 , a2 ) (with its interior), and a 3-simplex is a tetrahe-
dron (a0 , a1 , a2 , a3 ) (with its interior). The inclusion relation between any two faces σ and τ
of some simplex, s, is written σ τ .
We now state a number of properties of simplices, whose proofs are left as an exercise.
Clearly, a point x belongs to the boundary ∂σ of σ iff at least one of its barycentric co-
ordinates (λ0 , . . . , λn ) is zero, and a point x belongs to the interior Int σ of σ iff all of its
barycentric coordinates (λ0 , . . . , λn ) are positive, i.e., λi > 0 for all i, 0 ≤ i ≤ n. Then, for
every x ∈ σ, there is a unique face s such that x ∈ Int s, the face generated by those points
ai for which λi > 0, where (λ0 , . . . , λn ) are the barycentric coordinates of x.
A simplex σ is convex, arcwise connected, compact, and closed. The interior Int σ of a
simplex is convex, arcwise connected, open, and σ is the closure of Int σ.
We now put simplices together to form more complex shapes, following Munkres [43].
The intuition behind the next definition is that the building blocks should be “glued cleanly.”
Definition 9.2. A simplicial complex in Em (for short, a complex in Em ) is a set K consisting
of a (finite or infinite) set of simplices in Em satisfying the following conditions:
(1) Every face of a simplex in K also belongs to K.
(2) For any two simplices σ1 and σ2 in K, if σ1 ∩ σ2 6= ∅, then σ1 ∩ σ2 is a common face of

both σ1 and σ2 .
Every k-simplex, σ ∈ K, is called a k-face (or face) of K. A 0-face {v} is called a vertex and
a 1-face is called an edge. The dimension of the simplicial complex K is the maximum of
the dimensions of all simplices in K. If dim K = d, then every face of dimension d is called
a cell and every face of dimension d − 1 is called a facet.
Condition (2) guarantees that the various simplices forming a complex intersect nicely.
It is easily shown that the following condition is equivalent to condition (2):
(20 ) For any two distinct simplices σ1 , σ2 , Int σ1 ∩ Int σ2 = ∅.
Remarks:
1. A simplicial complex, K, is a combinatorial object, namely, a set of simplices satisfying
certain conditions but not a subset of Em . However, every complex, K, yields a subset
of Em called the geometric realization of K and denoted |K|. This object will be
defined shortly and should not be confused with the complex. Figure 9.2 illustrates
this aspect of the definition of a complex. For clarity, the two triangles (2-simplices)
are drawn as disjoint objects even though they share the common edge, (v2 , v3 ) (a
1-simplex) and similarly for the edges that meet at some common vertex.
230 CHAPTER 9. BASICS OF COMBINATORIAL TOPOLOGY 1
v3 v3
v1 v4
v2 v2
Figure 9.2: A set of simplices forming a complex
2. Unlike the situation for polyhedra, where all faces are external in the sense that they
belong to the boundary of the polyhedron, the situation for simplicial complexes is
more subtle; a face of a simplicial complex can be internal or external. For example,
the 1-simplex (v2 , v3 ) for the simplicial complex shown in Figure 9.2 is internal, but the
the 1-simplex (v1 , v2 ) is external. If we consider the simplicial complex consisting of the
faces of a tetrahedron, then every edge (1-simplex) is internal. However, if we consider
the simplicial complex consisting of a (solid) tetrahedron, then its facets (2-simplices)
and edges (1-simplices) are external. These matters will be clarified in Definition 9.7.
3. Some authors define a facet of a complex, K, of dimension d to be a d-simplex in K,

as opposed to a (d − 1)-simplex, as we did. This practice is not consistent with the
notion of facet of a polyhedron and this is why we prefer the terminology cell for the
d-simplices in K.
4. It is important to note that in order for a complex, K, of dimension d to be realized in

Em , the dimension of the “ambient space,” m, must be big enough. For example, there
are 2-complexes that can’t be realized in E3 or even in E4 . There has to be enough
room in order for condition (2) to be satisfied. It is not hard to prove that m = 2d + 1
is always sufficient. Sometimes, 2d works, for example in the case of surfaces (where
d = 2).
Some collections of simplices violating some of the conditions of Definition 9.2 are shown
in Figure 9.3. On the left, the intersection of the two 2-simplices is neither an edge nor a
vertex of either triangle. In the middle case, two simplices meet along an edge which is not
an edge of either triangle. On the right, there is a missing edge and a missing vertex.
Some “legal” simplicial complexes are shown in Figure 9.5.
The union |K| of all the simplices in K is a subset of Em . We can define a topology
on |K| by defining a subset F of |K| to be closed iff F ∩ σ is closed in σ for every face
σ ∈ K. It is immediately verified that the axioms of a topological space are indeed satisfied.
9.1. SIMPLICIAL COMPLEXES 1
231
Figure 9.3: Collections of simplices not forming a complex
v3
v1 v4
v2
Figure 9.4: The geometric realization of the complex of Figure 9.2
The resulting topological space |K| is called the geometric realization of K. The geometric
realization of the complex from Figure 9.2 is shown in Figure 9.4.
Obviously, |σ| = σ for every simplex, σ. Also, note that distinct complexes may have the
same geometric realization. In fact, all the complexes obtained by subdividing the simplices
of a given complex yield the same geometric realization.
A polytope is the geometric realization of some simplicial complex. A polytope of di-
mension 1 is usually called a polygon, and a polytope of dimension 2 is usually called a
polyhedron. Unfortunately the term “polytope” is overloaded since the polytopes induced
by simplicial complexes are generally not convex. Consequently, if we use the term polytope
for the objects defined in Chapter 5, we should really say “convex polytope” to avoid am-
biguity. When K consists of infinitely many simplices we usually require that K be locally
finite, which means that every vertex belongs to finitely many faces. If K is locally finite,
then its geometric realization, |K|, is locally compact.
In the sequel, we will consider only finite simplicial complexes, that is, complexes K
consisting of a finite number of simplices. In this case, the topology of |K| defined above
232 CHAPTER 9. BASICS OF COMBINATORIAL TOPOLOGY 1
Figure 9.5: Examples of simplicial complexes
is identical to the topology induced from Em . Also, for any simplex σ in K, Int σ coincides
◦
with the interior σ of σ in the topological sense, and ∂σ coincides with the boundary of σ in
the topological sense.
Definition 9.3. Given any complex, K2 , a subset K1 ⊆ K2 of K2 is a subcomplex of K2
iff it is also a complex. For any complex, K, of dimension d, for any i with 0 ≤ i ≤ d, the
subset
K (i) = {σ ∈ K | dim σ ≤ i}
is called the i-skeleton of K. Clearly, K (i) is a subcomplex of K. See Figure 9.6. We also let
K i = {σ ∈ K | dim σ = i}.
Observe that K 0 is the set of vertices of K and K i is not a complex. A simplicial complex,
K1 is a subdivision of a complex K2 iff |K1 | = |K2 | and if every face of K1 is a subset of some
face of K2 . A complex K of dimension d is pure (or homogeneous) iff every face of K is a
face of some d-simplex of K (i.e., some cell of K). See Figure 9.7. A complex is connected
iff |K| is connected.
It is easy to see that a complex is connected iff its 1-skeleton is connected. The intuition
behind the notion of a pure complex, K, of dimension d is that a pure complex is the result
of gluing pieces all having the same dimension, namely, d-simplices. For example, in Figure
9.8, the complex on the left is not pure but the complex on the right is pure of dimension 2.
9.2 Nonsingular Faces; Stars and Links

Most of the shapes that we will be interested in are well approximated by pure complexes,
in particular, surfaces or solids. However, pure complexes may still have undesirable “singu-
larities” such as the vertex v in Figure 9.8(b). The notions star and link of a face provide a
technical way to deal with singularities.
9.2. NONSINGULAR FACES; STARS AND LINKS 233
|K|
K(1)
Figure 9.6: The one skeleton of a two-dimensional complex.
Definition 9.4. Let K be any complex and let σ be any face of K. The star St(σ) (or if
we need to be very precise St(σ, K)) of σ is the subcomplex of K consisting of all faces τ
containing σ and of all faces of τ , that is,
St(σ) = {s ∈ K | (∃τ ∈ K)(σ τ and s τ )}.
The link Lk(σ) (or Lk(σ, K)) of σ is the subcomplex of K consisting of all faces in St(σ)
that do not intersect σ, that is,
Lk(σ) = {τ ∈ K | τ ∈ St(σ) and σ ∩ τ = ∅}.
To simplify notation, if σ = {v} is a vertex we write St(v) for St({v}) and Lk(v) for
Lk({v}). Figure 9.9 shows
(a) A complex (on the left).
(b) The star of the vertex v, indicated in mint green and the link of v, shown as thicker
red lines.
If K is pure and of dimension d, then St(σ) is also pure of dimension d and if dim σ = k,
then Lk(σ) is pure of dimension d − k − 1.
For technical reasons, following Munkres [43], besides defining the complex St(σ), it is
useful to introduce the open star of σ.
Definition 9.5. Given a complex K, for any simplex σ in K, the open star of σ, denoted
st(σ), is defined as the subspace of |K| consisting of the union of the interiors Int(τ ) = τ −∂ τ
of all the faces τ containing σ.
|K| K1
Figure 9.7: The complex K1 is a subdivision of the two-dimensional complex K. 1
(a) (b)
Figure 9.8: (a) A complex that is not pure. (b) A pure complex
According to this definition, the open star of σ is not a complex but instead a subset of
|K|. Note that
st(σ) = |St(σ)|,
that is, the closure of st(σ) is the geometric realization of the complex St(σ). Then lk(σ) =
|Lk(σ)| is the union of the simplices in St(σ) that are disjoint from σ. If σ is a vertex v, we
have
lk(v) = st(v) − st(v).
However, beware that if σ is not a vertex, then lk(σ) is properly contained in st(σ) − st(σ)!
See Figures 9.10 and 9.11.
One of the nice properties of the open star st(σ) of σ is that it is open. To see this,
observe that for any point a ∈ |K|, there is a unique smallest simplex σ = (v0 , . . . , vk ) such
(a) (b)
v v
Figure 9.9: (a) A complex. (b) Star and Link of v.
that a ∈ Int(σ), that is, such that
a = λ0 v0 + · · · + λk vk
with λi > 0 for all i, with 0 ≤ i ≤ k (and of course, λ0 + · · · + λk = 1). (When k = 0, we

have v0 = a and λ0 = 1.) For every arbitrary vertex v of K, we define tv (a) by

λi if v = vi , with 0 ≤ i ≤ k,
tv (a) =
0 if v ∈/ {v0 , . . . , vk }.
Using the above notation, observe that
st(v) = {a ∈ |K| | tv (a) > 0},
and thus, |K| − st(v) is the union of all the faces of K that do not contain v as a vertex,
obviously a closed set; see Figure 9.12. Thus, st(v) is open in |K|. It is also quite clear that
st(v) is path connected. Moreover, for any k-face σ of K, if σ = (v0 , . . . , vk ), then
st(σ) = {a ∈ |K| | tvi (a) > 0, 0 ≤ i ≤ k},
that is,
st(σ) = st(v0 ) ∩ · · · ∩ st(vk ).
Consequently, st(σ) is open and path connected, as illustrated in Figure 9.13.
Unfortunately, the “nice” equation
St(σ) = St(v0 ) ∩ · · · ∩ St(vk )
is false! (and anagolously for Lk(σ).) For a counter-example, (which is illustrated in Figure
9.14), consider the boundary of a tetrahedron and the star of a facet (a 2-simplex).
v v
|K| |St(v)|
v v
st(v)
lk(v) = |St(v)| - st(v)
Figure 9.10: For the pure 2-dimensional complex |K|, an illustration of |St(v)|, st(v), and
lk(v).
Recall that in Ed , the (open) unit ball B d is defined by
B d = {x ∈ Ed | kxk < 1},

d
the closed unit ball B is defined by
d
B = {x ∈ Ed | kxk ≤ 1},
and the (d − 1)-sphere S d−1 , by
S d−1 = {x ∈ Ed | kxk = 1}.

d
Obviously, S d−1 is the boundary of B (and B d ). The notion of link allows us to define
precisely what we mean by a nonsingular face.
Definition 9.6. Let K be a pure complex of dimension d and let σ be any k-face of K,
with 0 ≤ k ≤ d − 1. We say that σ is nonsingular iff the geometric realization lk(σ) of the
d−k−1
link of σ is homeomorphic to either S d−k−1 or to B ; this is written as lk(σ) ∼
= S d−k−1
d−k−1
or lk(σ) ∼
=B , where ∼
= means homeomorphic.
1
In Figure 9.9, note that the link of v is not homeomorphic to S 1 or B , so v is singular.
Given a pure complex, it is necessary to distinguish between two kinds of faces.
σ σ
|K| |St(σ )|
σ σ
|St(σ )| - st( σ)
st(σ )
lk( σ )
Figure 9.11: Given the edge σ in the pure 2-dimensional complex |K|, observe that lk(σ) ⊂
st(σ) − st(σ).
Definition 9.7. Let K be any pure complex of dimension d. A k-face σ of K is a boundary

or external face iff it belongs to a single cell (i.e., a d-simplex) of K, and otherwise it is called
an internal face (0 ≤ k ≤ d − 1). The boundary of K, denoted bd(K), is the subcomplex of
K consisting of all boundary facets of K together with their faces.
It is clear by definition that bd(K) is a pure complex of dimension d − 1. Even if K

is connected, bd(K) is not connected, in general. For example, if K is a 2-complex in the
plane, the boundary of K usually consists of several simple closed polygons (i.e, 1 dimensional
complexes homeomorphic to the circle, S 1 ).
Proposition 9.1. Let K be any pure complex of dimension d. For any k-face σ of K the
boundary complex bd(Lk(σ)) is nonempty iff σ is a boundary face of K (0 ≤ k ≤ d − 2).
Furthermore, Lkbd(K) (σ) = bd(Lk(σ)) for every face σ of bd(K), where Lkbd(K) (σ) denotes
the link of σ in bd(K).
Proof. Let F be any facet of K containing σ. We may assume that F = (v0 , . . . , vd−1 ) and
v
1
v1
σ
v
2
|K|
st( v1)
v
1
v
2
|K| - st( v1)
Figure 9.12: The construction of the closed set |K| − st(v1 ).
σ = (v0 , . . . , vk ), in which case, F 0 = (vk+1 , . . . , vd−1 ) is a (d − k − 2)-face of K and by

definition of Lk(σ), we have F 0 ∈ Lk(σ). Now, every cell (i.e., d-simplex) s containing F is
of the form s = conv(F ∪ {v}) for some vertex v, and s0 = conv(F 0 ∪ {v}) is a (d − k − 1)-face
in Lk(σ) containing F 0 . Consequently, F 0 is an external face of Lk(σ) iff F is an external
facet of K, establishing the proposition. The second statement follows immediately from the
proof of the first.
Proposition 9.1 shows that if every face of K is nonsingular, then the link of every internal
face is a sphere whereas the link of every external face is a ball.
The main goal of the rest of this section is to show that if K is a pure complex of
dimension d and if all its k-faces are nonsingular (0 ≤ k ≤ d − 1), then lk(σ) is either
d d−1
homeomorphic to B d or to B − B . As a consequence, the geometric realization |K| of
K is a manifold.
Although the above facts are easy to check for d = 1, 2, and in some simple cases for
d = 3, a rigorous proof requires a fair amount of work and the introduction of several new
concepts. The key point is that we need to express St(σ) in terms of Lk(σ), and for this, we
need the notion of join of complexes, two special cases of which are the notion of cone and
d+1
of suspension. These two notions allow building the sphere S d+1 and the closed ball B
d
from the sphere S d and the closed ball B , and allow an inductive argument on d. There are
many technical details which we will omit to simplify the exposition. Complete details and
proofs can be found in Munkres [43] (Chapter 8, Section 62).
v1
v
2
|K|
v
1
v2 v2
st( v1)
v
1
v2
st(σ ) = st( v1) h st( v2)
Figure 9.13: Given the edge σ in the pure 2-dimensional complex |K|, st(σ) = st(v1 ) ∩ st(v2 ).
We begin with yet another notion of cone.
Definition 9.8. Given any complex K in En , if dim K = d < n, for any point v ∈ En such
that v does not belong to the affine hull of |K|, the cone on K with vertex v, denoted, v ∗ K,
is the complex consisting of all simplices of the form (v, a0 , . . . , ak ) and their faces, where
(a0 , . . . , ak ) is any k-face of K. If K = ∅, we set v ∗ K = v. See Figure 9.15
It is not hard to check that v ∗ K is indeed a complex of dimension d + 1 containing K

as a subcomplex.
Remark: Unfortunately, the word “cone” is overloaded. It might have been better to use
the locution pyramid instead of cone as some authors do (for example, Ziegler). However,
since we have been following Munkres [43], a standard reference in algebraic topology, we
decided to stick with the terminology used in that book, namely, “cone.”
If σ is a simplex in a complex K, we will need to express St(σ) in terms of σ and its link
Lk(σ), and |St(σ)| − st(σ) in terms of ∂σ and Lk(σ). For this, we will need a generalization
v
v 0
0
v1
v1
v
v 3
3
v
v 2
2
|St( v0 v1 v 2)|
|K|
v
v v0 0
0
v1
v1 v1
v
v v 3
3 3
v
v v 2
2 2
|St( v2)|
|St( v0 )|
|St( v1)|
v
0
v1
v
3
v
2
|St( v0 )| h |St( v1)| h |St( v2)|
Figure 9.14: Let |K| be the boundary of the solid tetrahedron. The star of a triangular face
is itself and contains only three edges. It is not the intersection of the star of its vertices,
since the star of a vertex contains all six edges of the tetrahedron.
of the above notion of cone to two simplicial complexes K and L, called the join of two
complexes.
Definition 9.9. Given any two disjoint nonempty complexes K and L in En such that
dim(K) + dim(L) ≤ n − 1, if for any simplex σ = (v0 , . . . , vh ) in K and any simplex
τ = (w0 , . . . , wk ) in L, the points (v0 , . . . , vh , w0 , . . . , wk ) are affinely independent, then we
define σ ∗ τ as the simplex
σ ∗ τ = (v0 , . . . , vh , w0 , . . . , wk );
more rigorously, σ ∗τ is the (h+k +1)-simplex spanned by the points (v0 , . . . , vh , w0 , . . . , wk ).
If the collection of all the simplicies σ ∗ τ and their faces is a simplicial complex, then this
complex is denoted by K ∗ L and is called the join of K and L.
Note that if K ∗ L is a complex, then its dimension is dim(K) + dim(L) + 1, which implies
that dim(K) + dim(L) ≤ n − 1.
|K|
v * |K|
Figure 9.15: On the left is the two-dimensional planar complex |K|. On the right is the
geometric realization of |v∗K|. It consists of a solid blue tetrahedron and a peach tetrahedral
shell.
Observe that a cone v ∗ L corresponds to the special case where K is a complex consisting
of the single vertex v. If K = {v0 , v1 } is the complex consisting of two distinct vertices (with
no edge between them), and if {v0 , v1 } ∗ L is a complex, then it is called a suspension of L
and it is denoted by S(L) or susp(L). The suspension of L is the complex consisting of the
union of the two cones v0 ∗ L and v1 ∗ L.
Two problems immediately come to mind:
(1) Characterize the geometric realization |K ∗ L| of the join K ∗ L of two complexes K
and L in terms of the geometric realizations |K| and |L| of K and L, if K ∗ L is indeed
a complex.
(2) Find a sufficient condition of |K| and |L| that implies that K ∗ L is a complex.
The following proposition gives answers to these problems and gives a necessary and
sufficient condition for K ∗ L to be a complex. The proof is quite technical and not very
illuminating so we refer the reader to Munkres [43] (Chapter 8, Lemma 62.1).
Proposition 9.2. Let K and L be any two disjoint nonempty complexes in En such that
dim(K) + dim(L) ≤ n − 1.
(a) If K ∗L is a complex, then its geometric realization |K ∗L| is the union of all the closed
line segments [x, y] joining some point x in |K| to some point y in |L|. Two such line
segments intersect in a most a common endpoint.
(b) Conversely, if every pair of line segments joining points of |K| and points of |L| inter-
sect in at most a common endpoint, then K ∗ L is a complex.
Proposition 9.2 shows that |K ∗ L| can be expressed in terms of the realizations of cones
of the form v ∗ L, where v ∈ K, as
[
|K ∗ L| = |v ∗ L|.
v∈|K|
A few more technical propositions, all proved in Munkres [43] (see Chapter 8, Section
62), will be needed.
A surjective function f : X → Y between two topological spaces X and Y is called
a quotient map if a subset V of Y is open iff f −1 (V ) is open in X. A quotient map is
automatically continuous.
Proposition 9.3. Suppose K ∗ L is a well-defined complex where K and L are finite com-
plexes (it suffices to assume that K is locally finite). Then the map
π : |K| × |L| × [0, 1] → |K ∗ L|
given by
π(x, y, t) = (1 − t)x + ty
is a quotient map. For every x ∈ |K| and every y ∈ |L|, the map π collapes {x} × |L| × {0}
to the point x and |K| × {y} × {1} to the point y. Otherwise, π is injective.
Using Proposition 9.3 we can prove the following “obvious” proposition which turns out
to be very handy.
Proposition 9.4. Suppose K ∗ L and M ∗ N are well-defined finite complexes (it suffices to
assume that K is locally finite). If |K| ∼
= |M | and |L| ∼
= |N |, then |K ∗ L| ∼
= |M ∗ N |.
The next proposition follows immediately from the definitions.

Proposition 9.5. Let J, K, L be complexes in En . If J ∗ K and (J ∗ K) ∗ L are well-defined,
then K ∗ J and J ∗ (K ∗ L) are also well-defined and J ∗ K = K ∗ J, (J ∗ K) ∗ L = J ∗ (K ∗ L).
The following proposition shown in Munkres [43] (Chapter 8, Lemma 62.6) is crucial.
The proof actually follows pretty much from the definitions.
Proposition 9.6. For any complex K of dimension d and any k-simplex σ ∈ K (0 ≤ k ≤
d − 1), we have
St(σ) = σ ∗ Lk(σ),
and
st(σ) ∼
= |St(σ)| − |∂σ ∗ Lk(σ)|.
By convention, σ ∗ ∅ = σ if Lk(σ) = ∅, and ∅ ∗ Lk(σ) = Lk(σ) if ∂σ = ∅. See Figures 9.16
and 9.17.
v
v3
v1
v
v2 3
v1
|K|
lk(v)
v2
v v v
v
v3 v
3
v1
v1
v2 v2 v3
v1
v3
v1 v3
v1 v2
v2 v2 |St(v)| = |v * Lk(v)|
v
3
v1
v2
v * Lk(v)
Figure 9.16: An illustration of the formula St(v) ∼

= v ∗Lk(v) when |K| is the two-dimensional
tetrahedral shell.
Figure 9.18 shows a 3-dimensional complex. The link of the edge (v6 , v7 ) is the pentagon
2
P = (v1 , v2 , v3 , v4 , v5 ) ∼
= S 1 . The link of the vertex v7 is the cone v6 ∗ P ∼
= B . The link
1
of (v1 , v2 ) is (v6 , v7 ) ∼= B and the link of v1 is the union of the triangles (v2 , v6 , v7 ) and
2
(v5 , v6 , v7 ), which is homeomorphic to B .
The following technical propositions are needed to show that if K is any pure complex
of dimension d, nonsingularity of all the faces implies that every open star is an open subset
homeomorphic either to B d or to B d ∩ Hd , where
Hd = {(x1 , . . . , xd ) ∈ Rd | xd ≥ 0}.
The standard simplex ∆d is the convex subset of Rd+1 given by
∆d = {(t0 , . . . , td ) ∈ Rd+1 | t0 + · · · + td = 1, ti ≥ 0}.
It is easy to show that

d
|∆d | ∼
=B , |∂∆d+1 | ∼
= S d, for all d ≥ 0;
v4
v4
v3
v1 v3
v1
v2
st( v1 v2) v2
|K|
v4 v4
v3 v3
v1 v1
v v
v2 v
| v * Lk(v1 v2) |
v4
v4
minus
v3
v1 v3
v1
v2 | S 0 * Lk(v1 v2) |
B 1 x | v * Lk(v1 v2) | v2
Figure 9.17: An illustration of the formula st(σ) ∼

= |σ ∗ Lk(σ)| − |S 0 ∗ Lk(σ)| when |K| is the
two-dimensional tetrahedral shell and σ = (v1 , v2 ).
see Munkres [43] (Chapter 1, Lemma 1.1). The following formulae are also easy to show but
they are essential to carry out induction on the dimension of spheres or (closed) balls.
Proposition 9.7. The following homeomorphisms hold for all d ≥ 0:

d+1
|v ∗ ∂∆d+1 | ∼
= |∂∆d+1 ∗ v| ∼
=B
d+1
|v ∗ ∆d | ∼ = |∆d ∗ v| ∼
=B
|∂∆d+1 ∗ {v0 , v1 }| ∼
= S d+1
d+1
|∆d ∗ {v0 , v1 }| ∼
=B ,
for any points v, v0 6= v1 not in ∆d .
Informally, the first formula says that a cone over the sphere S d is homeomorphic to the
d+1 d
closed ball B , the second formula says that a cone over the closed ball B , is homeomorphic
d+1
to the closed ball B , the third formula says that the suspension of the sphere S d is
v6
v5 v4
v1 v3
v2
v7
Figure 9.18: More examples of links and stars
homeomorphic to the sphere S d+1 , and the fourth formula says that the suspension of the
d d+1
closed ball B is homeomorphic to the closed ball B .
Proposition 9.8. For every d ≥ 1 and every k-simplex σ (0 ≤ k ≤ d − 1), we have
d
|σ ∗ ∂∆d−k | ∼
=B
|∂σ ∗ ∂∆d−k | ∼= S d−1 .
Proof. We proceed by induction on d ≥ 1. For the base case d = 1, we must have k = 0 so
σ = v and ∂σ = ∅ for some vertex v, and then by Proposition 9.7
1
|v ∗ ∂∆1 | ∼
=B ,
and
|∅ ∗ ∂∆1 | = |∂∆1 | ∼
= S 0.
For the induction step for the first formula, we use Proposition 9.7 which says that
|∂∆d−k ∗ {v0 , v1 }| ∼
= S d−k+1 ∼
= |∂∆d−k+1 |.
For any k-simplex σ with 0 ≤ k ≤ d − 1, by Proposition 9.4 we have
|σ ∗ ∂∆d−k+1 | ∼
= |σ ∗ (∂∆d−k ∗ ({v0 , v1 })|
∼
= |(σ ∗ ∂∆d−k ) ∗ {v0 , v1 }|.
By the induction hypothesis, we have
d
|σ ∗ ∂∆d−k | ∼
= B = |∆d |,
so by Proposition 9.4 and Proposition 9.7, we have

|σ ∗ ∂∆d−k+1 | ∼
= |(σ ∗ ∂∆d−k ) ∗ {v0 , v1 }|
∼
= |∆d ∗ {v0 , v1 }|
d+1
∼
=B .
For a d-simplex σ, since |σ| ∼
= |∆d | and |∂∆1 | = |{v0 , v1 }|, by Proposition 9.4 and Proposition
9.7, we have
|σ ∗ ∂∆1 | ∼
= |∆d ∗ {v0 , v1 }|
d+1
∼
=B .
This concludes the induction step for the first formula and proves that
d
|σ ∗ ∂∆d−k | ∼
=B .
For the second formula, if k = 0 then σ = v is a vertex and ∂σ = ∅ so
|∅ ∗ ∂∆d | = |∂∆d | ∼
= S d−1 .
|∂σ ∗ ∂∆d−k+1 | ∼
= |∂σ ∗ (∂∆d−k ∗ ({v0 , v1 })|
∼
= |(∂σ ∗ ∂∆d−k ) ∗ {v0 , v1 }|.
|∂σ ∗ ∂∆d−k | ∼
= S d−1 = |∂∆d |,
|∂σ ∗ ∂∆d−k+1 | ∼
= |(∂σ ∗ ∂∆d−k ) ∗ {v0 , v1 }|
∼
= |∂∆d ∗ {v0 , v1 }|
∼
= S d.
For a d-simplex σ, since |∂σ| ∼
= |∂∆d | and |∂∆1 | ∼
= |{v0 , v1 }|, by Proposition 9.4 and Propo-
sition 9.7, we have
|∂σ ∗ ∂∆1 | ∼
= |∂σ ∗ {v0 , v1 }|
∼
= |∂∆d ∗ {v0 , v1 }|
∼
= S d.
This concludes the induction step for the second formula and proves that
|∂σ ∗ ∂∆d−k | ∼
= S d−1 ,
as claimed.
Since |∂∆d−k | ∼
= S d−k−1 , with a slight abuse of notation the formulae of Proposition 9.8
can be written as
d
|σ ∗ S d−k−1 | ∼
=B
|∂σ ∗ S d−k−1 | ∼= S d−1 .
The following proposition is the counterpart of Proposition 9.8 for balls instead of spheres.
Proposition 9.9. For every d ≥ 1 and every k-simplex σ (0 ≤ k ≤ d − 1), we have
d
|σ ∗ ∆d−k−1 | ∼
=B
d−1
|∂σ ∗ ∆d−k−1 | ∼
=B .
Proof. We proceed by induction on d ≥ 1. For the base case d = 1, we must have k = 0 so
σ = v for some vertex v, and then by Proposition 9.7
1
|v ∗ ∆0 | ∼
=B ,
and by definition
0
|∅ ∗ ∆0 | = |∆0 | ∼
=B .
For the induction step for the first formula, we use Proposition 9.7 which says that
d−k
|∆d−k−1 ∗ v| ∼
=B = |∆d−k |.
|σ ∗ ∆d−k | ∼
= |σ ∗ (∆d−k−1 ∗ v)|
∼
= |(σ ∗ ∆d−k−1 ) ∗ v|.
d
|σ ∗ ∆d−k−1 | ∼
= B = |∆d |,
|σ ∗ ∆d−k | ∼
= |(σ ∗ ∆d−k−1 ) ∗ v|
∼
= |∆d ∗ v|
d+1
∼
=B .
For a d-simplex σ, since |σ| ∼
= |∆d | and |∆0 | = |v|, by Proposition 9.4 and Proposition 9.7,
we have
|σ ∗ ∆0 | ∼
= |∆d ∗ v|
d+1
∼
=B .
This concludes the induction step for the first formula and proves that
d
|σ ∗ ∆d−k−1 | ∼
=B .
For the second formula, if k = 0 then σ = v is a vertex and ∂σ = ∅ so
d−1
|∅ ∗ ∆d−1 | = |∆d−1 | ∼
=B .
|∂σ ∗ ∆d−k | ∼
= |∂σ ∗ (∆d−k−1 ∗ v)|
∼
= |(∂σ ∗ ∆d−k−1 ) ∗ v|.
d−1
|∂σ ∗ ∆d−k−1 | ∼
=B = |∆d−1 |,
|∂σ ∗ ∆d−k | ∼
= |(∂σ ∗ ∆d−k−1 ) ∗ v|
∼
= |∆d−1 ∗ v|
d
∼
=B .
For a d-simplex σ, since |∂σ| ∼
= |∂∆d | and |∆0 | ∼
= |v|, by Proposition 9.4 and Proposition
9.7, we have
|∂σ ∗ ∆0 | ∼
= |∂σ ∗ v|
∼
= |∂∆d ∗ v|
d
∼
=B .
This concludes the induction step for the second formula and proves that
d−1
|∂σ ∗ ∆d−k−1 | ∼
=B ,
as claimed.
d−k−1
Since |∆d−k−1 | ∼
=B , with a slight abuse of notation the formulae of Proposition 9.9
can be written as
d−k−1 d
|σ ∗ B |∼
=B
d−k−1 d−1
|∂σ ∗ B |∼
=B .
Finally, we can prove that for any pure complex K of dimension d, nonsingularity of all
the faces implies that the open star of any internal face is homeomorphic to B d , and that
d d−1
the open star of any boundary face is homeomorphic to B − B . This result for pure
complexes K without boundaries is stated in Thurston [61] (Chapter 3, Proposition 3.2.5).
Theorem 9.10. Let K be any pure complex of dimension d. If every face of K is nonsin-
d d−1
gular, then st(σ) ∼
= B d for every every internal k-face σ of K, and st(σ) ∼
= B −B for
every every boundary k-face σ of K (0 ≤ k ≤ d − 1).
Proof. By Proposition 9.6, for any complex K of dimension d and any k-simplex σ ∈ K
(0 ≤ k ≤ d − 1), we have
St(σ) = σ ∗ Lk(σ),
and
st(σ) ∼
= |St(σ)| − |∂σ ∗ Lk(σ)|.
If σ is an internal face then
= S d−k−1 ∼
|Lk(σ)| ∼ = |∂∆d−k |,
so by Proposition 9.8 and Proposition 9.4
st(σ) ∼
= |St(σ)| − |∂σ ∗ Lk(σ)|
∼
= |σ ∗ ∂∆d−k | − |∂σ ∗ ∂∆d−k |
d
∼
= B − S d−1
∼
= Bd.
If σ is a boundary face then

d−k−1
|Lk(σ)| ∼
=B ∼
= |∆d−k−1 |,
so by Proposition 9.9 and Proposition 9.4
st(σ) ∼
= |St(σ)| − |∂σ ∗ Lk(σ)|
∼
= |σ ∗ ∆d−k−1 | − |∂σ ∗ ∆d−k−1 |
d d−1
∼
=B −B ,
as claimed.
Theorem 9.10 has the following corollary which shows that any pure complex for which
every face is nonsingular is a manifold.
Proposition 9.11. Let K be any pure complex of dimension d. If every face of K is

nonsingular, then for every point a ∈ |K|, there is an open subset U ⊆ |K| containing a
such that if a does not belong to the boundary of |K| then U ∼
= B d , and if a belongs to the
boundary of |K| then U ∼ d
=B ∩H . d
Proof. Any point a ∈ |K| belongs to some simplex σ, so we proceed by induction on the
dimension of σ. If σ is a vertex v, then by Proposition 9.10 we have st(v) ∼
= B d or st(σ) ∼
=
d d−1
B −B ∼ d d
= B ∩ H . If σ is a simplex of dimension k + 1, then any point a ∈ ∂σ
on the boundary of σ belongs to a simplex of dimension at most k, and the induction
hypothesis implies that there is an open subset U ⊆ |K| containing a such that U ∼ = B d or
U ∼ = B d ∩ Hd . Otherwise, a belongs to the interior of σ, and we conclude by Proposition
d d−1
9.10 since st(σ) ∼
= B d or st(σ) ∼
=B −B ∼
= B d ∩ Hd .
Remark: Thurston states that Proposition 9.11 holds for pure complexes without bound-
aries under the weaker assumption that lk(v) ∼ = S d−1 for every vertex; see Thurston [61],
Chapter 3, Proposition 3.2.5. A proof of the more general fact that if lk(v) ∼ = S d−1 or
d−1
lk(v) ∼
=B for every vertex then every face is nonsingular can be found in Stallings [54];
see Section 4.4, Proposition 4.4.12. The proof requires several technical lemmas and is quite
involved.
Here are more useful propositions about pure complexes without singularities.
Proposition 9.12. Let K be any pure complex of dimension d. If every facet of K is
nonsingular, then every facet of K is contained in at most two cells (d-simplices).
Proof. If |K| ⊆ Ed , then this is an immediate consequence of the definition of a complex.
Otherwise, consider lk(σ). By hypothesis, either lk(σ) ∼
= B 0 or lk(σ) ∼
= S 0 . As B 0 = {0},
0
S = {−1, 1} and dim Lk(σ) = 0, we deduce that Lk(σ) has either one or two points, which
proves that σ belongs to at most two d-simplices.
Proposition 9.13. Let K be any pure and connected complex of dimension d. If every face
of K is nonsingular, then for every pair of cells (d-simplices), σ and σ 0 , there is a sequence
of cells, σ0 , . . . , σp , with σ0 = σ and σp = σ 0 , and such that σi and σi+1 have a common facet,
for i = 0, . . . , p − 1.
Proof. We proceed by induction on d, using the fact that the links are connected for d ≥
2.
Proposition 9.14. Let K be any pure complex of dimension d. If every facet of K is
nonsingular, then the boundary, bd(K), of K is a pure complex of dimension d − 1 with an
empty boundary. Furthermore, if every face of K is nonsingular, then every face of bd(K)
is also nonsingular.
Proof. Left as an exercise.
9.3 Polyhedral Complexes

The building blocks of simplicial complexes, namely, simplicies, are in some sense mathemat-
ically ideal. However, in practice, it may be desirable to use a more flexible set of building
blocks. We can indeed do this and use convex polytopes as our building blocks.
9.3. POLYHEDRAL COMPLEXES 251
Definition 9.10. A polyhedral complex in Em (for short, a complex in Em ) is a set K

consisting of a (finite or infinite) set of convex polytopes in Em satisfying the following
conditions:
(1) Every face of a polytope in K also belongs to K.
(2) For any two polytopes σ1 and σ2 in K, if σ1 ∩ σ2 6= ∅, then σ1 ∩ σ2 is a common face

of both σ1 and σ2 .
Every polytope σ ∈ K of dimension k is called a k-face (or face) of K. A 0-face {v} is

called a vertex and a 1-face is called an edge. The dimension of the polyhedral complex K
is the maximum of the dimensions of all polytopes in K. If dim K = d, then every face of
dimension d is called a cell , and every face of dimension d − 1 is called a facet.
Remark: Since the building blocks of a polyhedral complex are convex polytopes it might
be more appropriate to use the term “polytopal complex” rather than “polyhedral complex”
and some authors do that. On the other hand, most of the traditional litterature uses the
terminology polyhedral complex so we will stick to it. There is a notion of complex where
the building blocks are cones but these are called fans.
Every convex polytope, P , yields two natural polyhedral complexes:
(i) The polyhedral complex K(P ) consisting of P together with all of its faces. This
complex has a single cell, namely P itself.
(ii) The boundary complex K(∂P ) consisting of all faces of P other than P itself. The cells
of K(∂P ) are the facets of P .
The notions of k-skeleton and pureness are defined just as in the simplicial case. The
notions of star and link are defined for polyhedral complexes just as they are defined for
simplicial complexes except that the word “face” now means face of a polytope. Now, by
Theorem 5.7, every polytope σ is the convex hull of its vertices.
Let vert(σ) denote the set of vertices of σ. Then, we have the following crucial observation:
Given any polyhedral complex K, for every point x ∈ |K|, there is a unique polytope σx ∈ K
such that x ∈ Int(σx ) = σx − ∂ σx . We define a function t : V → R+ that tests whether x
belongs to the interior of any face (polytope) of K having v as a vertex as follows: For every
vertex v of K,
1 if v ∈ vert(σx )
tv (x) =
0 if v ∈
/ vert(σx ),
where σx is the unique face of K such that x ∈ Int(σx ).
Now, just as in the simplicial case, the open star st(v) of a vertex v ∈ K is given by
st(v) = {x ∈ |K| | tv (x) = 1},

and it is an open subset of |K| (the set |K| − st(v) is the union of the polytopes of K that
do not contain v as a vertex, a closed subset of |K|). Also, for any face σ, of K, the open
star st(σ) of σ is given by
\
st(σ) = {x ∈ |K| | tv (x) = 1, for all v ∈ vert(σ)} = st(v).
v∈vert(σ)
Therefore, st(σ) is also open in |K|.

The next proposition is another result that seems quite obvious, yet a rigorous proof is
more involved that we might think. In fact, the only place that I am aware of where a proof
is mentioned is the survey article by Carl Lee, Subdivisions and Triangulations of Polytopes
(Chapter 17), in Goodman and O’Rourke [32]. Actually, the “proof” that Lee is referring
to is a proof sketch whose details are “left to the reader.” It turns out that a proof can be
given using an inductive construction described in Grünbaum [35] (Chapter 5).
The proposition below states that a convex polytope can always be cut up into simplices,
that is, it can be subdivided into a simplicial complex. In other words, every convex polytope
can be triangulated. This implies that simplicial complexes are as general as polyhedral
complexes.
One should be warned that even though, in the plane, every bounded region (not nec-
essarily convex) whose boundary consists of a finite number of closed polygons (polygons
homeomorphic to the circle S 1 ) can be triangulated, this is no longer true in three dimensions!
For example, the 3D polyhedron known as the Schönhart polyhedron cannot be triangulated;
see Boissonnat and Yvinec [12] (Chapters 13, Section 13.2).
Proposition 9.15. Every convex d-polytope P can be subdivided into a simplicial complex
without adding any new vertices, i.e., every convex polytope can be triangulated.
Proof sketch. It would be tempting to proceed by induction on the dimension, d, of P but
we do not know any correct proof of this kind. Instead, we proceed by induction on the
number, p, of vertices of P . Since dim(P ) = d, we must have p ≥ d + 1. The case p = d + 1
corresponds to a simplex, so the base case holds.
For p > d + 1, we can pick some vertex, v ∈ P , such that the convex hull, Q, of the
remaining p − 1 vertices still has dimension d. Then, by the induction hypothesis, Q, has
a simplicial subdivision. Now, we say that a facet, F , of Q is visible from v iff v and the
interior of Q are strictly separated by the supporting hyperplane of F . Then, we add the
d-simplices, conv(F ∪ {v}) = v ∗ F , for every facet, F , of Q visible from v to those in the
triangulation of Q. We claim that the resulting collection of simplices (with their faces)
constitutes a simplicial complex subdividing P .
This is the part of the proof that requires some details. We say that v is beneath a facet
F of Q iff v belongs to the open half–space determined by the supporting hyperplane of F
which contains the interior of Q. We make use of a theorem of Grünbaum [35] (Theorem 1,
Chapter 5, Section 5.2) which states the following:
9.4. COMBINATORIAL AND TOPOLOGICAL MANIFOLDS 253
Theorem (Grünbaum). If P and Q are two polytopes as above with P = conv(Q ∪ {v}),
then the following properties hold:
(i) A face F of Q is a face of P iff there is a facet F 0 of Q such that F ⊆ F 0 and v is

beneath F 0 .
(ii) If F is a face of Q, then F ∗ = conv(F ∪ {v}) is a face of P iff either
(a) v ∈ aff(F ); or
(b) among the facets of Q containing F there is at least one such that v is beneath it,
and at least one which is visible from v.
Moreover, each face of P is of one and only one of those types.
The above theorem implies that the new simplices that need to be added to form a
triangulation of P are the convex hulls conv(F ∪ {v}) associated with facets F of Q visible
from v. The reader should check that everything really works out!
With all this preparation, it is now quite natural to define combinatorial manifolds.
9.4 Combinatorial and Topological Manifolds

The notion of pure complex without singular faces turns out to be a very good “discrete”
approximation of the notion of (topological) manifold because of its highly computational
nature. This motivates the following definition:
Definition 9.11. A combinatorial d-manifold is any space X homeomorphic to the geomet-

ric realization |K| ⊆ En of some pure (simplicial or polyhedral) complex K of dimension
d whose faces are all nonsingular. If the link of every k-face of K is homeomorphic to the
sphere S d−k−1 , we say that X is a combinatorial manifold without boundary, else it is a
combinatorial manifold with boundary.
Other authors use the term triangulation of PL-manifold for what we call a combinatorial
manifold.
It is easy to see that the connected components of a combinatorial 1-manifold are either
simple closed polygons or simple chains (“simple” means that the interiors of distinct edges
are disjoint). A combinatorial 2-manifold which is connected is also called a combinatorial
surface (with or without boundary). Proposition 9.14 immediately yields the following result:
Proposition 9.16. If X is a combinatorial d-manifold with boundary, then bd(X) is a

combinatorial (d − 1)-manifold without boundary.
Now, because we are assuming that X sits in some Euclidean space, En , the space X
is Hausdorff and second-countable. (Recall that a topological space is second-countable iff
there is a countable family {Ui }i≥0 of open sets of X such that every open subset of X is
the union of open sets from this family.) Since it is desirable to have a good match between
manifolds and combinatorial manifolds, we are led to the definition below.
Recall that
Hd = {(x1 , . . . , xd ) ∈ Rd | xd ≥ 0}.
Definition 9.12. For any d ≥ 1, a (topological) d-manifold with boundary is a second-
countable, topological Hausdorff space M , together with an open cover (Ui )i∈I of open sets
in M and a family (ϕi )i∈I of homeomorphisms ϕi : Ui → Ωi , where each Ωi is some open
subset of Hd in the subset topology. Each pair (U, ϕ) is called a coordinate system or
chart of M , each homeomorphism ϕi : Ui → Ωi is called a coordinate map, and its inverse
ϕ−1
i : Ωi → Ui is called a parameterization of Ui . The family (Ui , ϕi )i∈I is often called an
atlas for M . A (topological) bordered surface is a connected 2-manifold with boundary. If
for every homeomorphism ϕi : Ui → Ωi , the open set Ωi ⊆ Hd is actually an open set in Rd
(which means that xd > 0 for every (x1 , . . . , xd ) ∈ Ωi ), then we say that M is a d-manifold .
Note that a d-manifold is also a d-manifold with boundary.

If ϕi : Ui → Ωi is some homeomorphism onto some open set Ωi of Hd in the subset
topology, some p ∈ Ui may be mapped into Rd−1 × R+ , or into the “boundary” Rd−1 × {0}
of Hd . Letting ∂Hd = Rd−1 × {0}, it can be shown using homology that if some coordinate
map ϕ, defined on p maps p into ∂Hd , then every coordinate map ψ, defined on p maps p
into ∂Hd .
Thus, M is the disjoint union of two sets ∂M and Int M , where ∂M is the subset consisting
of all points p ∈ M that are mapped by some (in fact, all) coordinate map ϕ defined on
p into ∂Hd , and where Int M = M − ∂M . The set ∂M is called the boundary of M , and
the set Int M is called the interior of M , even though this terminology clashes with some
prior topological definitions. A good example of a bordered surface is the Möbius strip. The
boundary of the Möbius strip is a circle.
The boundary ∂M of M may be empty, but Int M is nonempty. Also, it can be shown
using homology that the integer d is unique. It is clear that Int M is open and a d-manifold,
and that ∂M is closed. If p ∈ ∂M , and ϕ is some coordinate map defined on p, since Ω = ϕ(U )
is an open subset of ∂Hd , there is some open half ball Bo+
d
centered at ϕ(p) and contained in
Ω which intersects ∂H along an open ball Bo , and if we consider W = ϕ−1 (Bo+
d d−1 d
), we have
d
an open subset of M containing p which is mapped homeomorphically onto Bo+ in such that
way that every point in W ∩ ∂M is mapped onto the open ball Bod−1 . Thus, it is easy to see
that ∂M is a (d − 1)-manifold.
Proposition 9.17. Every combinatorial d-manifold is a d-manifold with boundary.
Proof. This is an immediate consequence of Proposition 9.11.
9.4. COMBINATORIAL AND TOPOLOGICAL MANIFOLDS 255
Is the converse of Proposition 9.17 true?

It turns out that answer is yes for d = 1, 2, 3 but no for d ≥ 4. This is not hard to
prove for d = 1. For d = 2 and d = 3, this is quite hard to prove; among other things, it is
necessary to prove that triangulations exist and this is very technical. For d ≥ 4, not every
manifold can be triangulated (in fact, this is undecidable!).
What if we assume that M is a triangulated manifold, which means that M is a manifold
and that M ∼= |K| for some pure d-dimensional complex K?
Surprisingly, for d ≥ 5, there are triangulated manifolds whose links are not spherical
d−k−1
(i.e., not homeomorphic to B or S d−k−1 ). Such an example is the double suspension of
Poincaré space; see Thurston [61], Example 3.2.11.
Fortunately, we will only have to deal with d = 2, 3! Another issue that must be addressed
is orientability.
Assume that we fix a total ordering of the vertices of a complex, K. Let σ = (v0 , . . . , vk )
be any simplex. Recall that every permutation (of {0, . . . , k}) is a product of transpositions,
where a transposition swaps two distinct elements, say i and j, and leaves every other element
fixed. Furthermore, for any permutation, π, the parity of the number of transpositions
needed to obtain π only depends on π and it called the signature of π. We say that two
permutations are equivalent iff they have the same signature. Consequently, there are two
equivalence classes of permutations: Those of even signature and those of odd signature.
Then, an orientation of σ is the choice of one of the two equivalence classes of permutations
of its vertices. If σ has been given an orientation, then we denote by −σ the result of
assigning the other orientation to it (we call it the opposite orientation).
For example, (0, 1, 2) has the two orientation classes:
{(0, 1, 2), (1, 2, 0), (2, 0, 1)} and {(2, 1, 0), (1, 0, 2), (0, 2, 1)}.
Definition 9.13. Let X ∼ = |K| be a combinatorial d-manifold. We say that X is orientable

if it is possible to assign an orientation to all of its cells (d-simplices) so that whenever two
cells σ1 and σ2 have a common facet, σ, the two orientations induced by σ1 and σ2 on σ are
opposite. A combinatorial d-manifold together with a specific orientation of its cells is called
an oriented manifold . If X is not orientable we say that it is non-orientable.
Remark: It is possible to define the notion of orientation of a manifold but this is quite
technical and we prefer to avoid digressing into this matter. This shows another advantage
of combinatorial manifolds: The definition of orientability is simple and quite natural.
There are non-orientable (combinatorial) surfaces, for example, the Möbius strip which
can be realized in E3 . The Möbius strip is a surface with boundary, its boundary being a
circle. There are also non-orientable (combinatorial) surfaces such as the Klein bottle or
the projective plane but they can only be realized in E4 (in E3 , they must have singularities
such as self-intersection). We will only be dealing with orientable manifolds and, most of
the time, surfaces.
One of the most important invariants of combinatorial (and topological) manifolds is
their Euler(-Poincaré) characteristic. In the next chapter, we prove a famous formula due
to Poincaré giving the Euler characteristic of a convex polytope. For this, we will introduce
a technique of independent interest called shelling.
Chapter 10
Shellings, the Euler–Poincaré Formula

for Polytopes, the Dehn-Sommerville
Equations and the Upper Bound
Theorem
10.1 Shellings
The notion of shellability is motivated by the desire to give an inductive proof of the Euler–
Poincaré formula in any dimension. Historically, this formula was discovered by Euler for
three dimensional polytopes in 1752 (but it was already known to Descartes around 1640).
If f0 , f1 and f2 denote the number of vertices, edges and triangles of the three dimensional
polytope, P , (i.e., the number of i-faces of P for i = 0, 1, 2), then the Euler formula states
that
f0 − f1 + f2 = 2.
The proof of Euler’s formula is not very difficult but one still has to exercise caution. Euler’s
formula was generalized to arbitrary d-dimensional polytopes by Schläfli (1852) but the
first correct proof was given by Poincaré. For this, Poincaré had to lay the foundations of
algebraic topology and after a first “proof” given in 1893 (containing some flaws) he finally
gave the first correct proof in 1899. If fi denotes the number of i-faces of the d-dimensional
polytope, P , (with f−1 = 1 and fd = 1), the Euler–Poincaré formula states that:
d−1
X
(−1)i fi = 1 − (−1)d ,
i=0
which can also be written as

d
X
(−1)i fi = 1,
i=0
257
258 CHAPTER 10. SHELLINGS AND THE EULER–POINCARÉ FORMULA
by incorporating fd = 1 in the first formula or as

d
X
(−1)i fi = 0,
i=−1
by incorporating both f−1 = 1 and fd = 1 in the first formula.

Earlier inductive “proofs” of the above formula were proposed, notably a proof by Schläfli
in 1852, but it was later observed that all these proofs assume that the boundary of every
polytope can be built up inductively in a nice way, what is called shellability. Actually,
counter-examples of shellability for various simplicial complexes suggested that polytopes
were perhaps not shellable. However, the fact that polytopes are shellable was finally proved
in 1970 by Bruggesser and Mani [17] and soon after that (also in 1970) a striking application
of shellability was made by McMullen [42] who gave the first proof of the so-called “upper
bound theorem”.
As shellability of polytopes is an important tool and as it yields one of the cleanest
inductive proof of the Euler–Poincaré formula, we will sketch its proof in some details. This
Chapter is heavily inspired by Ziegler’s excellent treatment [67], Chapter 8. We begin with
the definition of shellability. It’s a bit technical, so please be patient!
Definition 10.1. Let K be a pure polyhedral complex of dimension d. A shelling of K is a

list F1 , . . . , Fs of all the cells (i.e., d-faces) of K such that either d = 0 (and thus, all Fi are
points), or the following conditions hold:
(i) The boundary complex K(∂F1 ) of the first cell F1 of K has a shelling.
(ii) For any j, 1 < j ≤ s, the intersection of the cell Fj with the previous cells is nonempty
and is an initial segment of a shelling of the (d − 1)-dimensional boundary complex of
Fj , that is,
j−1
!
[
Fj ∩ Fi = G1 ∪ G2 ∪ · · · ∪ Gr ,
i=1
for some shelling G1 , G2 , . . . , Gr , . . . , Gt of K(∂Fj ), with 1 ≤ r ≤ t. As the intersection

should be the initial segment of a shelling for the (d − 1)-dimensional complex ∂Fj , it
has to be pure (d − 1)-dimensional and connected for d > 1.
A polyhedral complex is shellable if it is pure and has a shelling.
Note that shellabiliy is only defined for pure complexes. Here are some examples of
shellable complexes:
(1) Every 0-dimensional complex, that is, every set of points is shellable, by definition.
10.1. SHELLINGS 259 1
1 2 3 4 3 5
2
8 4 2
1
7 6 5 1
Figure 10.1: Non shellable and Shellable 2-complexes
(2) A 1-dimensional complex is a graph without loops and parallel edges. A 1-dimensional
complex is shellable iff it is connected, which implies that it has no isolated vertices.
Any ordering of the edges e1 , . . . , es such that {e1 , . . . , ei } induces a connected subgraph
for every i will do. Such an ordering can be defined inductively, due to the connectivity
of the graph.
(3) Every simplex is shellable. In fact, any ordering of its facets yields a shelling. This is
easily shown by induction on the dimension, since the intersection of any two facets Fi
and Fj is a facet of both Fi and Fj .
(4) The d-cubes are shellable. By induction on the dimension, it can be shown that
every ordering of the 2d facets F1 , . . . , F2d such that F1 and F2d are opposite (that is,
F2d = −F1 ) yields a shelling.
However, already for 2-complexes, problems arise. For example, in Figure 10.1, the left
and the middle 2-complexes are not shellable but the right complex is shellable.
The problem with the left complex is that cells 1 and 2 intersect at a vertex, which is not
1-dimensional. In the middle complex shown in Figure 10.1, the intersection of cell 8 with its
predecessors is not connected, so the particular order chosen is not a shelling. However, there
are other orders that constitute a shelling. In contrast, the ordering of the right complex is
a shelling. However, observe that the reverse ordering is not a shelling because cell 4 has an
empty intersection with cell 5!
Remarks:
1. Condition (i) in Definition 10.1 is redundant because, as we shall prove shortly, every
polytope is shellable. However, if we want to use this definition for more general
complexes, then Condition (i) is necessary.
2. When K is a simplicial complex, Condition (i) is of course redundant, as every simplex

is shellable but Condition (ii) can also be simplified to:
(ii’) For any j, with 1 < j ≤ s, the intersection of Fj with the previous cells is
nonempty and pure (d − 1)-dimensional. This means that for every i < j there is
some l < j such that Fi ∩ Fj ⊆ Fl ∩ Fj and Fl ∩ Fj is a facet of Fj .
The following proposition yields an important piece of information about the local struc-
ture of shellable simplicial complexes; see Ziegler [67], Chapter 8.
Proposition 10.1. Let K be a shellable simplicial complex and say F1 , . . . , Fs is a shelling
for K. Then, for every vertex v, the restriction of the above sequence to the link Lk(v), and
to the star St(v), are shellings.
Since the complex K(P ) associated with a polytope P has a single cell, namely P itself,
note that by condition (i) in the definition of a shelling, K(P ) is shellable iff the complex
K(∂P ) is shellable. We will say simply say that “P is shellable” instead of “K(∂P ) is
shellable.”
We have the following useful property of shellings of polytopes whose proof is left as an
exercise (use induction on the dimension):
Proposition 10.2. Given any polytope, P , if F1 , . . . , Fs is a shelling of P , then the reverse
sequence Fs , . . . , F1 is also a shelling of P .
Proposition 10.2 generally fails for complexes that are not polytopes, see the right 2-
complex in Figure 10.1.
We will now present the proof that every polytope is shellable, using a technique invented
by Bruggesser and Mani (1970) known as line shelling [17]. This is quite a simple and
natural idea if one is willing to ignore the technical details involved in actually checking that
it works. We begin by explaining this idea in the 2-dimensional case, a convex polygon, since
it is particularly simple.
Consider the 2-polytope P shown in Figure 10.2 (a polygon) whose faces are labeled
F1 , F2 , F3 , F4 , F5 . Pick any line ` intersecting the interior of P and intersecting the supporting
lines of the facets of P (i.e., the edges of P ) in distinct points labeled z1 , z2 , z3 , z4 , z5 (such
a line can always be found, as will be shown shortly). Orient the line ` (say, upward), and
travel on ` starting from the point of P where ` leaves P , namely z1 . For a while, only face
F1 is visible, but when we reach the intersection z2 of ` with the supporting line of F2 , the
face F2 becomes visible, and F1 becomes invisible as it is now hidden by the supporting line
of F2 . So far, we have seen the faces F1 and F2 , in that order . As we continue traveling along
`, no new face becomes visible but for a more complicated polygon, other faces Fi would
become visible one at a time as we reach the intersection zi of ` with the supporting line of
Fi , and the order in which these faces become visible corresponds to the ordering of the zi ’s
along the line `. Then, we imagine that we travel very fast and when we reach “+∞” in the
upward direction on `, we instantly come back on ` from below at “−∞”. At this point, we
only see the face of P corresponding to the lowest supporting line of faces of P , i.e., the line
10.1. SHELLINGS 261
corresponding to the smallest zi , in our case z3 . At this stage, the only visible face is F3 .
We continue traveling upward on ` and we reach z3 , the intersection of the supporting line
of F3 with `. At this moment, F4 becomes visible, and F3 disappears as it is now hidden
by the supporting line of F4 . Note that F5 is not visible at this stage. Finally, we reach z4 ,
the intersection of the supporting line of F4 with `, and at this moment the last facet F5
becomes visible (and F4 becomes invisible, F3 being also invisible). Our trip stops when we
reach z5 , the intersection of F5 and `. During the second phase of our trip, we saw F3 , F4
and F5 , and the entire trip yields the sequence F1 , F2 , F3 , F4 , F5 , which is easily seen to be a
shelling of P . 1
z2
F1 z1
F2
F4
F3
z5
F5
z4
z3
Figure 10.2: Shelling a polygon by travelling along a line
This is the crux of the Bruggesser-Mani method for shelling a polytope: We travel along
a suitably chosen line and record the order in which the faces become visible during this
trip. This is why such shellings are called line shellings.
In order to prove that polytopes are shellable we need the notion of points and lines
in “general position.” Recall from the equivalence of V-polytopes and H-polytopes that a
polytope P in Ed with nonempty interior is cut out by t irredundant hyperplanes Hi , and
by picking the origin in the interior of P the equations of the Hi may be assumed to be of
the form
ai · z = 1
where ai and aj are not proportional for all i 6= j, so that
P = {z ∈ Ed | ai · z ≤ 1, 1 ≤ i ≤ t}.
Definition 10.2. Let P be any polytope in Ed with nonempty interior and assume that P
is cut out by the irredudant hyperplanes Hi of equations ai · z = 1, for i = 1, . . . , t. A point
c ∈ Ed is said to be in general position w.r.t. P is c does not belong to any of the Hi , that
is, if ai · c 6= 1 for i = 1, . . . , t. A line ` is said to be in general position w.r.t. P if ` is not
parallel to any of the Hi , and if ` intersects the Hi in distinct points.
The following proposition showing the existence of lines in general position w.r.t. a
polytope illustrates a very useful technique, the “perturbation method.” The “trick” behind
this particular perturbation method is that polynomials (in one variable) have a finite number
of zeros.
Proposition 10.3. Let P be any polytope in Ed with nonempty interior. For any two points
x and y in Ed , with x outside of P ; y in the interior of P ; and x in general position w.r.t.
P ; for λ ∈ R small enough, the line, `λ , through x and yλ with
yλ = y + (λ, λ2 , . . . , λd ),
intersects P in its interior and is in general position w.r.t. P .
Proof. Assume that P is defined by t irredundant hyperplanes Hi , where Hi is given by the

equation ai · z = 1, and write Λ = (λ, λ2 , . . . , λd ) and u = y − x. Then the line `λ is given by
`λ = {x + s(yλ − x) | s ∈ R} = {x + s(u + Λ) | s ∈ R}.
The line `λ is not parallel to the hyperplane Hi iff
ai · (u + Λ) 6= 0, i = 1, . . . , t,
and it intersects the Hi in distinct points iff there is no s ∈ R such that
ai · (x + s(u + Λ)) = 1 and aj · (x + s(u + Λ)) = 1 for some i 6= j.

10.1. SHELLINGS 263
Observe that ai · (u + Λ) = pi (λ) is a nonzero polynomial in λ of degree at most d. Since

a polynomial of degree d has at most d zeros, if we let Z(pi ) be the (finite) St set of zeros of
pSi we can ensure that `λ is not parallel to any of the Hi by picking λ ∈ / i=1 Z(pi ) (where
t
i=1 Z(p i ) is a finite set). Now, as x is in general position w.r.t. P , we have ai · x 6= 1, for
i = 1 . . . , t. The condition stating that `λ intersects the Hi in distinct points can be written
ai · x + sai · (u + Λ) = 1 and aj · x + saj · (u + Λ) = 1 for some i 6= j,
or
spi (λ) = αi and spj (λ) = αj for some i 6= j,
where αi = 1−ai ·x and αj = 1−aj ·x. As x is in general position w.r.t. P , we have αi , αj 6= 0
and as the Hi are irredundant, the polynomials pi (λ) = ai · (u + Λ) and pj (λ) = aj · (u + Λ)
are not proportional. Now, if λ ∈
/ Z(pi ) ∪ Z(pj ), in order for the system
spi (λ) = αi
spj (λ) = αj
to have a solution in s we must have
qij (λ) = αi pj (λ) − αj pi (λ) = 0,
where qij (λ) is not the zero polynomial since pi (λ) and pj (λ) are not proportional and
αi , αj 6= 0. If we pick λ ∈
/ Z(qij ), then qij (λ) 6= 0. Therefore, if we pick
t
[ t
[
λ∈
/ Z(pi ) ∪ Z(qij ),
i=1 i6=j
the line `λ is in general position w.r.t. P . Finally, we can pick λ small enough so that
yλ = y + Λ is close enough to y so that it is in the interior of P .
It should be noted that the perturbation method involving Λ = (λ, λ2 , . . . , λd ) is quite

flexible. For example, by adapting the proof of Proposition 10.3 we can prove that for any
two distinct facets Fi and Fj of P , there is a line in general position w.r.t. P intersecting
Fi and Fj . Start with x outside P and very close to Fi and y in the interior of P and very
close to Fj .
Finally, before proving the existence of line shellings for polytopes, we need more termi-
nology.
Definition 10.3. Given any point x strictly outside a polytope P , we say that a facet F
of P is visible from x iff for every y ∈ F the line through x and y intersects P only in y
(equivalently, x and the interior of P are strictly separared by the supporting hyperplane of
F ).
z3
z2
z1 F1
F4 F3
F2
Figure 10.3: Shelling a polytope by travelling along a line, `
We now prove the following fundamental theorem due to Bruggesser and Mani [17] (1970):
Theorem 10.4. (Existence of Line Shellings for Polytopes) Let P be any polytope in Ed of
dimension d. For every point x outside P and in general position w.r.t. P , there is a shelling
of P in which the facets of P that are visible from x come first.
Proof. By Proposition 10.3, we can find a line ` through x such that ` is in general position
w.r.t. P and ` intersects the interior of P . Pick one of the two faces in which ` intersects
P , say F1 , let z1 = ` ∩ F1 , and orient ` from the inside of P to z1 . As ` intersects the
supporting hyperplanes of the facets of P in distinct points, we get a linearly ordered list of
these intersection points along `,
z1 , z2 , · · · , zm , zm+1 , · · · , zs ,
where zm+1 is the smallest element, zm is the largest element, and where z1 and zs belong to
the faces of P where ` intersects P . Then, as in the example illustrated by Figure 10.2, by
travelling “upward” along the line ` starting from z1 we get a total ordering of the facets of
P,
F1 , F2 , . . . , Fm , Fm+1 , . . . , Fs
where Fi is the facet whose supporting hyperplane cuts ` in zi .
10.1. SHELLINGS 265
We claim that the above sequence is a shelling of P . This is proved by induction on d.

For d = 1, P consists a line segment and the theorem clearly holds.
Consider the intersection ∂Fj ∩ (F1 ∪ · · · ∪ Fj−1 ). We need to show that this is an initial
segment of a shelling of ∂Fj . If j ≤ m, i.e., if Fj become visible before we reach ∞, then
the above intersection is exactly the set of facets of Fj that are visible from zj = ` ∩ aff(Fj ).
Therefore, by induction on the dimension, these facets are shellable and they form an initial
segment of a shelling of the whole boundary ∂Fj .
If j ≥ m+1, that is, after “passing through ∞” and reentering from −∞, the intersection
∂Fj ∩ (F1 ∪ · · · ∪ Fj−1 ) is the set of non-visible facets. By reversing the orientation of the line,
`, we see that the facets of this intersection are shellable and we get the reversed ordering
of the facets.
Finally, when we reach the point x starting from z1 , the facets visible from x form an
initial segment of the shelling, as claimed.
Remark: The trip along the line ` is often described as a rocket flight starting from the
surface of P viewed as a little planet (for instance, this is the description given by Ziegler
[67] (Chapter 8)). Observe that if we reverse the direction of `, we obtain the reversal of the
original line shelling. Thus, the reversal of a line shelling is not only a shelling but a line
shelling as well.
We can easily prove the following corollary:
Corollary 10.5. Given any polytope P , the following facts hold:
(1) For any two facets F and F 0 , there is a shelling of P in which F comes first and F 0
comes last.
(2) For any vertex v of P , there is a shelling of P in which the facets containing v form
an initial segment of the shelling.
Proof. For (1), we use a line in general position and intersecting F and F 0 in their interior.
For (2), we pick a point x beyond v and pick a line in general position through x intersecting
the interior of P . Pick the origin O in the interior of P . A point x is beyond v iff x and O
lies on different sides of every hyperplane Hi supporting a facet of P containing x but on
the same side of Hi for every hyperplane Hi supporting a facet of P not containing x. Such
a point can be found on a line through O and v, as the reader should check.
Remark: A plane triangulation K is a pure two-dimensional complex in the plane such that
|K| is homeomorphic to a closed disk. Edelsbrunner proves that every plane triangulation
has a shelling, and from this, that χ(K) = 1, where χ(K) = f0 −f1 +f2 is the Euler–Poincaré
characteristic of K, where f0 is the number of vertices, f1 is the number of edges and f2 is
the number of triangles in K (see Edelsbrunner [25], Chapter 3). This result is an immediate
consequence of Corollary 10.5 if one knows about the stereographic projection map, which
will be discussed in the next Chapter.
We now have all the tools needed to prove the famous Euler–Poincaré Formula for Poly-
topes.
10.2 The Euler–Poincaré Formula for Polytopes

We begin by defining a very important topological concept, the Euler–Poincaré characteristic
of a complex.
Definition 10.4. Let K be a d-dimensional polyhedral complex. For every i, with 0 ≤ i ≤ d,

we let fi denote the number of i-faces of K and we let
f (K) = (f0 , · · · , fd ) ∈ Nd+1
be the f -vector associated with K (if necessary we write fi (K) instead of fi ). The Euler–
Poincaré characteristic χ(K) of K is defined by
d
X
χ(K) = f0 − f1 + f2 + · · · + (−1)d fd = (−1)i fi .
i=0
Given any d-dimensional polytope P , the f -vector associated with P is the f -vector associ-
ated with K(P ), that is,
f (P ) = (f0 , · · · , fd ) ∈ Nd+1 ,
where fi , is the number of i-faces of P (= the number of i-faces of K(P ) and thus, fd = 1),
and the Euler–Poincaré characteristic χ(P ) of P is defined by
d
X
χ(P ) = f0 − f1 + f2 + · · · + (−1)d fd = (−1)i fi .
i=0
Moreover, the f -vector associated with the boundary ∂P of P is the f -vector associated
with K(∂P ), that is,
f (∂P ) = (f0 , · · · , fd−1 ) ∈ Nd ,
where fi , is the number of i-faces of ∂P (with 0 ≤ i ≤ d − 1), and the Euler–Poincaré
characteristic χ(∂P ) of ∂P is defined by
d−1
X
d−1
χ(∂P ) = f0 − f1 + f2 + · · · + (−1) fd−1 = (−1)i fi .
i=0
Observe that χ(P ) = χ(∂P ) + (−1)d , since fd = 1.

10.2. THE EULER–POINCARÉ FORMULA FOR POLYTOPES 267
Remark: It is convenient to set f−1 = 1. Then, some authors, including Ziegler [67] (Chap-
ter 8), define the reduced Euler–Poincaré characteristic χ0 (K) of a polyhedral complex (or
a polytope) K as
d
X
0 d
χ (K) = −f−1 + f0 − f1 + f2 + · · · + (−1) fd = (−1)i fi = −1 + χ(K),
i=−1
i.e., they incorporate f−1 = 1 into the formula.

A crucial observation for proving the Euler–Poincaré formula is that the Euler–Poincaré
characteristic is additive, which means that if K1 and K2 are any two complexes such that
K1 ∪ K2 is also a complex, which implies that K1 ∩ K2 is also a complex (because we must
have F1 ∩ F2 ∈ K1 ∩ K2 for every face F1 of K1 and every face F2 of K2 ), then
χ(K1 ∪ K2 ) = χ(K1 ) + χ(K2 ) − χ(K1 ∩ K2 ). (∗)
This follows immediately because for any two sets A and B
|A ∪ B| = |A| + |B| − |A ∩ B|.
To prove our next theorem we will use complete induction on N × N ordered by the
lexicographic ordering. Recall that the lexicographic ordering on N × N is defined as follows:
 m = m0 and n < n0

(m, n) < (m0 , n0 ) iff or

m < m0 .

Theorem 10.6. (Euler–Poincaré Formula) For every polytope P of dimension d, we have

d
X
χ(P ) = (−1)i fi = 1 (d ≥ 0),
i=0
and so
d−1
X
χ(∂P ) = (−1)i fi = 1 − (−1)d (d ≥ 1).
i=0
Proof. We prove the following statement: For every d-dimensional polytope P , if d = 0 then
χ(P ) = 1,
else if d ≥ 1 then for every shelling F1 , . . . , Ffd−1 of P , for every j, with 1 ≤ j ≤ fd−1 , we
have
1 if 1 ≤ j < fd−1
χ(F1 ∪ · · · ∪ Fj ) = d
1 − (−1) if j = fd−1 .
We proceed by complete induction on (d, j) ≥ (0, 1). For d = 0 and j = 1, the polytope P
consists of a single point and so, χ(P ) = f0 = 1, as claimed.
For the induction step, assume that d ≥ 1. For 1 = j < fd−1 , since F1 is a polytope of
dimension d − 1, by the induction hypothesis, χ(F1 ) = 1, as desired.
For 1 < j < fd−1 , by (∗) we have
j−1 j−1
! ! !
[ [
χ(F1 ∪ · · · Fj−1 ∪ Fj ) = χ Fi + χ(Fj ) − χ Fi ∩ Fj . (∗∗)
i=1 i=1
Since (d, j − 1) < (d, j), by the induction hypothesis,

j−1
!
[
χ Fi =1
i=1
and since dim(Fj ) = d − 1, again by the induction hypothesis,
χ(Fj ) = 1.
Now, as F1 , . . . , Ffd−1 is a shelling and j < fd−1 , we have
j−1
!
[
Fi ∩ Fj = G1 ∪ · · · ∪ Gr ,
i=1
for some shelling G1 , . . . , Gr , . . . , Gt of K(∂Fj ), with r < t = fd−2 (∂Fj ). The fact that
r < fd−2 (∂Fj ), i.e., that G1 ∪ · · · ∪ Gr is not the whole boundary of Fj is a property of line
shellings and also follows from Proposition 10.2. As dim(∂Fj ) = d − 2, and r < fd−2 (∂Fj ),
by the induction hypothesis, we have
j−1
! !
[
χ Fi ∩ Fj = χ(G1 ∪ · · · ∪ Gr ) = 1.
i=1
Consequently, by (∗∗) we obtain
χ(F1 ∪ · · · Fj−1 ∪ Fj ) = 1 + 1 − 1 = 1,
as claimed (when j < fd−1 ).

If j = fd−1 , then we have a complete shelling of ∂Ffd−1 , that is,
fd−1 −1
!
[
Fi ∩ Ffd−1 = G1 ∪ · · · ∪ Gfd−2 (Ffd−1 ) = ∂Ffd−1 .
i=1
10.3. DEHN–SOMMERVILLE EQUATIONS FOR SIMPLICIAL POLYTOPES 269
As dim(∂Fj ) = d − 2, by the induction hypothesis,
χ(∂Ffd−1 ) = χ(G1 ∪ · · · ∪ Gfd−2 (Ffd−1 ) ) = 1 − (−1)d−1 ,
and by (∗∗) it follows that
χ(F1 ∪ · · · ∪ Ffd−1 ) = 1 + 1 − (1 − (−1)d−1 ) = 1 + (−1)d−1 = 1 − (−1)d ,
establishing the induction hypothesis in this last case. But then,
χ(∂P ) = χ(F1 ∪ · · · ∪ Ffd−1 ) = 1 − (−1)d
and
χ(P ) = χ(∂P ) + (−1)d = 1,
proving our theorem.
Remark: Other combinatorial proofs of the Euler–Poincaré formula are given in Grünbaum
[35] (Chapter 8), Boissonnat and Yvinec [12] (Chapter 7) and Ewald [26] (Chapter III).
Coxeter gives a proof very close to Poincaré’s own proof using notions of homology theory
[20] (Chapter IX). We feel that the proof based on shellings is the most direct and one of
the most elegant. Incidently, the above proof of the Euler–Poincaré formula is very close to
Schläfli proof from 1852 but Schläfli did not have shellings at his disposal so his “proof” had
a gap. The Bruggesser-Mani proof that polytopes are shellable fills this gap!
10.3 Dehn–Sommerville Equations for Simplicial

Polytopes and h-Vectors
If a d-polytope P has the property that its faces are all simplices, then it is called a simplicial
polytope.
It is easily shown that a polytope is simplicial iff its facets are simplices, in which case,
every facet has d vertices. The polar dual of a simplicial polytope is called a simple polytope.
We see immediately that every vertex of a simple polytope belongs to d facets.
For simplicial (and simple) polytopes it turns out that other remarkable equations be-
sides the Euler–Poincaré formula hold among the number of i-faces. These equations were
discovered by Dehn for d = 4, 5 (1905) and by Sommerville in the general case (1927). Al-
though it is possible (and not difficult) to prove the Dehn–Sommerville equations by “double
counting,” as in Grünbaum [35] (Chapter 9) or Boissonnat and Yvinec (Chapter 7, but be-
ware, these are the dual formulae for simple polytopes), it turns out that instead of using
the f -vector associated with a polytope it is preferable to use what’s known as the h-vector,
because for simplicial polytopes the h-numbers have a natural interpretation in terms of
shellings. Furthermore, the statement of the Dehn–Sommerville equations in terms of h-

vectors is transparent:
hi = hd−i ,
and the proof is very simple in terms of shellings.
If K is a simplicial polytope and V is the set of vertices of K, then every i-face of K can
be identified with an (i + 1)-subset of V (that is, a subset of V of cardinality i + 1).
In the rest of this section, we restrict our attention to pure simplicial complexes. In order
to motivate h-vectors, we begin by examining more closely the structure of the new faces that
are created during a shelling when the cell Fj is added to the partial shelling F1 , . . . , Fj−1 .
Definition 10.5. For any shelling F1 , . . . , Fs of a pure simplicial complex K of dimension

d − 1, for every j, with 1 ≤ j ≤ s, the restriction Rj of the facet Fj is the set of “obligatory”
vertices
Rj = {v ∈ Fj | Fj − {v} ⊆ Fi , for some i with 1 ≤ i < j}.
Observe that R1 = ∅. The crucial property of the Rj is that the new faces G added at
step j (when Fj is added to the shelling) are precisely the faces in the set
Ij = {G ⊆ V | Rj ⊆ G ⊆ Fj }.
The proof of the above fact is left as an exercise to the reader, or see Ziegler [67] (Chapter
8, Section 8.3).
But then, we obtain a partition {I1 , . . . , Is } of the set of faces of the simplicial complex
(other that K itself). Note that the empty face is allowed. Now, if we define
hi = |{j | |Rj | = i, 1 ≤ j ≤ s}|,
for i = 0, . . . , d, then it turns out that we can recover the fk in terms of the hi as follows:
s k
d − |Rj | d−i
X X
fk−1 = = hi ,
j=1
k − |Rj | i=0
k−i
with 1 ≤ k ≤ d.
But more is true: The above equations are invertible and the hk can be expressed in
terms of the fi as follows:
k
k−i d − i
X
hk = (−1) fi−1 ,
i=0
d−k
with 0 ≤ k ≤ d (remember, f−1 = 1).
Let us explain all this in more detail. Consider the example of a connected graph (a
simplicial 1-dimensional complex) from Ziegler [67] (Section 8.3) shown in Figure 10.4.
1 4 5
2 3 6
Figure 10.4: A connected 1-dimensional complex, C.
A shelling order of its 7 edges is given by the sequence

12, 13, 34, 35, 45, 36, 56.
The partial order of the faces of C together with the blocks of the partition {I1 , . . . , I7 }
associated with the seven edges of C are shown in Figure 10.5, with the blocks Ij shown in
boldface.
12 13 34 35 45 36 56
1 2 3 4 5 6
Figure 10.5: The partition associated with a shelling of C.
In the above example, we have R1 = {∅}, R2 = {3}, R3 = {4}, R4 = {5}, R5 = {4, 5},
R6 = {6} and R7 = {5, 6}, I1 = {∅, 1, 2, 12}, I2 = {3, 13}, I3 = {4, 34}, I4 = {5, 35},
I5 = {45}, I6 = {6, 36}, I7 = {56}, and the “minimal” new faces (corresponding to the Rj ’s)
added at every stage of the shelling are
∅, 3, 4, 5, 45, 6, 56.
Definition 10.6. For any shellable pure simplicial complex K of dimension d − 1, if hi is
the number of blocks Ij such that the corresponding restriction set Rj has size i, that is,
hi = |{j | |Rj | = i, 1 ≤ j ≤ s}| for i = 0, . . . , d,
then we define the h-vector associated with K as
h(K) = (h0 , . . . , hd ).
In other words, hi is the number of minimal faces in the partitition that have i vertices,
with h0 = 1.
In our example, as R1 = {∅}, R2 = {3}, R3 = {4}, R4 = {5}, R5 = {4, 5}, R6 = {6} and
R7 = {5, 6}, we have h0 = 1, h1 = 4, and h2 = 2, that is,
h(C) = (1, 4, 2).
Looking at Figure 10.5, we see that for every horizontal layer i (starting from 0) of the
lattice, hi is the numbers of nodes (in bold) that are minimal in some block of the partition.
Now, let us show that if K is a shellable simplicial complex, then the f -vector can be
recovered from the h-vector. Indeed, K is a pure simplicial complex so every face is contained
in a face of dimension d − 1 which has d vertices, and if |Rj | = i, then each (k − 1)-face in
the block Ij must use all i nodes in Rj , so that there are only d − i nodes available and,
among those, k − i must be chosen. Therefore,
s
d − |Rj |
X
fk−1 = ,
j=1
k − |Rj |
and by definition of hi , we get

k
d−i d−k+1 d−1

X d
fk−1 = hi = hk + hk−1 + · · · + h1 + h0 , (∗)
i=0
k−i 1 k−1 k
where 1 ≤ k ≤ d. Moreover, the formulae are invertible, that is, the hi can be expressed in
terms of the fk . For this, form the two polynomials
d
X
f (x) = fi−1 xd−i = fd−1 + fd−2 x + · · · + f0 xd−1 + f−1 xd
i=0
with f−1 = 1 and

d
X
h(x) = hi xd−i = hd + hd−1 x + · · · + h1 xd−1 + h0 xd .
i=0
Then, it is easy to see that

d
X
f (x) = hi (x + 1)d−i = h(x + 1).
i=0
Consequently, h(x) = f (x − 1) and by comparing the coefficients of xd−k on both sides of

the above equation, we get
k
k−i d − i
X
hk = (−1) fi−1 .
i=0
d − k
In particular, h0 = 1, h1 = f0 − d, and
hd = fd−1 − fd−2 + fd−3 + · · · + (−1)d−1 f0 + (−1)d .
It is also easy to check that
h0 + h1 + · · · + hd = fd−1 .
Now, we just showed that if K is shellable, then its f -vector and its h-vector are related
as above. But even if K is not shellable, the above suggests defining the h-vector from the
f -vector as above. Thus, we make the definition:
Definition 10.7. For any (d − 1)-dimensional pure simplicial complex K, the h-vector
associated with K is the vector
h(K) = (h0 , . . . , hd ) ∈ Zd+1 ,
given by
k
d−i
X
k−i
hk = (−1) fi−1 .
i=0
d−k
Note that if K is shellable, then the interpretation of hi as the number of cells, Fj , such
that the corresponding restriction set, Rj , has size i shows that hi ≥ 0. However, for an
arbitrary simplicial complex, some of the hi can be strictly negative. Such an example is
given in Ziegler [67] (Section 8.3).
We summarize below most of what we just showed:
Proposition 10.7. Let K be a (d−1)-dimensional pure simplicial complex. If K is shellable,
then its h-vector is nonnegative and hi counts the number of cells in a shelling whose restric-
tion set has size i. Moreover, the hi do not depend on the particular shelling of K.
There is a way of computing the h-vector of a pure simplicial complex from its f -vector
reminiscent of the Pascal triangle (except that negative entries can turn up). This method is
known as Stanley’s trick ; see Stanley [55]. For this we write the numbers fi to thelast entries
of the rows of Pascal’s triangle (to the place where ordinarily we would put i+1 i+1
= 1), and
then we compute the other entries using the rule
upper right neighbor − upper left neighbor.
For example, for the graph C of Figure 10.4, for which the f -vector is f = (1, 6, 7), we obtain
the following table
1
1 6
1 5 7
1 4 2
and we get h(C) = (1, 4, 7).

If we now consider the boundary of an octahedron we have f = (1, 6, 12, 8), we have the
following table
1
1 6
1 5 12
1 4 7 8
1 3 3 1
so h = (1, 3, 3, 1).
For a simplicial complex that is not shellable, it is possible to obtain an h-vector with
negative entries; see Ziegler [67] (Section 8.3, Example (iii)).
We are now ready to prove the Dehn–Sommerville equations. For d = 3, these are easily
obtained by double counting. Indeed, for a simplicial polytope, every edge belongs to two
facets and every facet has three edges. It follows that
2f1 = 3f2 .
Together with Euler’s formula

f0 − f1 + f2 = 2,
we see that
f1 = 3f0 − 6 and f2 = 2f0 − 4,
namely, that the number of vertices of a simplicial 3-polytope determines its number of edges
and faces, these being linear functions of the number of vertices. For arbitrary dimension d,
we have
Theorem 10.8. (Dehn–Sommerville Equations) If K is any simplicial d-polytope, then the
components of the h-vector satisfy
hk = hd−k k = 0, 1, . . . , d.
Equivalently,
d
X
d−i i
fk−1 = (−1) fi−1 k = 0, . . . , d.
i=k
k
Furthermore, the equation h0 = hd is equivalent to the Euler–Poincaré formula.
Proof. We present a short and elegant proof due to McMullen. Recall from Proposition 10.2
that the reversal Fs , . . . , F1 of a shelling F1 , . . . , Fs of a polytope is also a shelling. From this,
we see that for every Fj , the restriction set of Fj in the reversed shelling is equal to Rj − Fj ,
the complement of the restriction set of Fj in the original shelling. Therefore, if |Rj | = k,
then Fj contributes “1” to hk in the original shelling iff it contributes “1” to hd−k in the
reversed shelling (where |Rj − Fj | = d − k). It follows that the value of hk computed in the
original shelling is the same as the value of hd−k computed in the reversed shelling. However,
by Proposition 10.7, the h-vector is independent of the shelling and hence, hk = hd−k .
To prove the second equation, following Ewald [26] (Chapter III, Theorem 3.7), define
the polynomials F (x) and H(x) by
d
X
i d x
F (x) = fi−1 x ; H(x) = (1 − x) F .
i=0
1−x
Pd
Note that H(x) = i=0 fi−1 xi (1 − x)d−i , and an easy computation shows that the coefficient
of xk is equal to
k
d−i
X
k−i
(−1) fi−1 = hk .
i=0
d−k
Now, the equations hk = hd−k are equivalent to
H(x) = xd H(x−1 ). (†)
Substituting the definition of H in terms of F in equation (†) yields

d x d d 1/x
(1 − x) F = x (1 − 1/x) F
1−x 1 − 1/x
d
d (x − 1)

d x 1
(1 − x) F =x F
1−x xd x−1

d x d 1
(1 − x) F = (−(1 − x)) F
1−x x−1

x 1
F = (−1)d F ,
1−x x−1
where the last equation holds for all x 6= 1. If we let

1 1
y=− = ,
x−1 1−x
then
1 x
y−1= −1= ,
1−x 1−x
so the equation
x d 1
F = (−1) F
1−x x−1
which holds for all x 6= 1 yields
F (y − 1) = (−1)d F (−y)
for all y 6= 0 (since y = 1/(1 − x)). But F (x − 1) and (−1)d F (−x) are polynomials that
have the same value for infinitely many real values, so in fact the polynomials F (x − 1) and
(−1)d F (−x) are identical. As
d d i
X
i
X X i
F (x − 1) = fi−1 (x − 1) = fi−1 xi−j (−1)j ,
i=0 i=0 j=0
i − j
we see that the coefficient of xk in F (x − 1) (obtained when i − j = k, that is, j = i − k) is

d d
i−k i i−k i
X X
(−1) fi−1 = (−1) fi−1 .
i=0
k i=k
k
On the other hand, the coefficient of xk in (−1)d F (−x) is (−1)d+k fk−1 . By equating the
coefficients of xk , we get
d
d+k
X
i−k i
(−1) fk−1 = (−1) fi−1 ,
i=k
k
which, by multiplying both sides by (−1)d+k , is equivalent to

d d
X
d+i i X
d−i i
fk−1 = (−1) fi−1 = (−1) fi−1 ,
i=k
k i=k
k
as claimed. Finally, as we already know that
hd = fd−1 − fd−2 + fd−3 + · · · + (−1)d−1 f0 + (−1)d
and h0 = 1, by multiplying both sides of the equation hd = h0 = 1 by (−1)d−1 and moving

(−1)d (−1)d−1 = −1 to the right hand side, we get the Euler–Poincaré formula.
Clearly, the Dehn–Sommerville equations, hk = hd−k , are linearly independent for

0 ≤ k < b d+1
2
c. For example, for d = 3, we have the two independent equations
h0 = h3 , h1 = h2 ,
and for d = 4, we also have two independent equations
h0 = h4 , h1 = h3 ,
since h2 = h2 is trivial. When d = 3, we know that h1 = h2 is equivalent to 2f1 = 3f2 and

when d = 4, if one unravels h1 = h3 in terms of the fi ’ one finds
2f2 = 4f3 ,
10.4. THE UPPER BOUND THEOREM 277
that is f2 = 2f3 . More generally, it is easy to check that
2fd−2 = dfd−1
for all d. For d = 5, we find three independent equations
h0 = h5 , h1 = h4 , h2 = h3 ,
and so on.
It can be shown that for general d-polytopes, the Euler–Poincaré formula is the only equa-
tion satisfied by all h-vectors and for simplicial d-polytopes, the b d+1
2
c Dehn–Sommerville
equations, hk = hd−k , are the only equations satisfied by all h-vectors (see Grünbaum [35],
Chapter 9).
Remark: Readers familiar with homology and cohomology may suspect that the Dehn–
Sommerville equations are a consequence of a type of Poincaré duality. Stanley proved that
this is indeed the case. It turns out that the hi are the dimensions of cohomology groups of
a certain toric variety associated with the polytope. For more on this topic, see Stanley [56]
(Chapters II and III) and Fulton [28] (Section 5.6).
As we saw for 3-dimensional simplicial polytopes, the number of vertices, n = f0 , de-
termines the number of edges and the number of faces, and these are linear in f0 . For
d ≥ 4, this is no longer true and the number of facets is no longer linear in n but in fact
quadratic. It is then natural to ask which d-polytopes with a prescribed number of vertices
have the maximum number of k-faces. This question which remained an open problem for
some twenty years was eventually settled by McMullen in 1970 [42]. We will present this
result (without proof) in the next section.
10.4 The Upper Bound Theorem and Cyclic Polytopes

Given a d-polytope with n vertices, what is an upper bound on the number of its i-faces? This
question is not only important from a theoretical point of view but also from a computational
point of view because of its implications for algorithms in combinatorial optimization and in
computational geometry.
The answer to the above problem is that there is a class of polytopes called cyclic polytopes
such that the cyclic d-polytope, Cd (n), has the maximum number of i-faces among all d-
polytopes with n vertices.
This result stated by Motzkin in 1957 became known as the upper bound conjecture until
it was proved by McMullen in 1970, using shellings [42] (just after Bruggesser and Mani’s
proof that polytopes are shellable). It is now known as the upper bound theorem. Another
proof of the upper bound theorem was given later by Alon and Kalai [2] (1985). A version
of this proof can also be found in Ewald [26] (Chapter 3).
McMullen’s proof is not really very difficult but it is still quite involved so we will only
state some propositions needed for its proof. We urge the reader to read Ziegler’s account
of this beautiful proof [67] (Chapter 8). We begin with cyclic polytopes.
First, consider the cases d = 2 and d = 3. When d = 2, our polytope is a polygon in
which case n = f0 = f1 . Thus, this case is trivial.
For d = 3, we claim that 2f1 ≥ 3f2 . Indeed, every edge belongs to exactly two faces so if
we add up the number of sides for all faces, we get 2f1 . Since every face has at least three
sides, we get 2f1 ≥ 3f2 . Then, using Euler’s relation, it is easy to show that
f1 ≤ 6n − 3 f2 ≤ 2n − 4
and we know that equality is achieved for simplicial polytopes.

Let us now consider the general case. The rational curve, c : R → Rd , given parametrically
by
c(t) = (t, t2 , . . . , td )
is at the heart of the story. This curve if often called the moment curve or rational normal
curve of degree d. For d = 3, it is known as the twisted cubic. Here is the definition of the
cyclic polytope, Cd (n).
Definition 10.8. For any sequence, t1 < . . . < tn , of distinct real number ti ∈ R, with
n > d, the convex hull,
Cd (n) = conv(c(t1 ), . . . , c(tn ))
of the n points c(t1 ), . . . , c(tn ) on the moment curve of degree d is called a cyclic polytope.
The first interesting fact about the cyclic polytope is that it is simplicial.
Proposition 10.9. Every d + 1 of the points c(t1 ), . . . , c(tn ) are affinely independent. Con-
sequently, Cd (n) is a simplicial polytope and the c(ti ) are vertices.
Proof. We may assume that n = d + 1. Say c(t1 ), . . . , c(tn ) belong to a hyperplane, H, given
by
α1 x1 + · · · + αd xd = β.
(Of course, not all the αi are zero.) Then, we have the polynomial, H(t), given by
H(t) = −β + α1 t + α2 t2 + · · · + αd td ,
of degree at most d and as each c(ti ) belong to H, we see that each c(ti ) is a zero of H(t).
However, there are d + 1 distinct c(ti ), so H(t) would have d + 1 distinct roots. As H(t) has
degree at most d, it must be the zero polynomial, a contradiction. Returing to the original
n > d + 1, we just proved every d + 1 of the points c(t1 ), . . . , c(tn ) are affinely independent.
Then, every proper face of Cd (n) has at most d independent vertices, which means that it is
a simplex.
n

The following proposition already shows that the cyclic polytope, Cd (n), has k
(k − 1)-
faces if 1 ≤ k ≤ b d2 c.
Proposition 10.10. For any k with 2 ≤ 2k ≤ d, every subset of k vertices of Cd (n) is a

(k − 1)-face of Cd (n). Hence

n d
fk (Cd (n)) = if 0 ≤ k < .
k+1 2
Proof. Consider any sequence ti1 < ti2 < · · · < tik . We will prove that there is a hyperplane
separating F = conv({c(ti1 ), . . . , c(tik )}) and Cd (n). Consider the polynomial
k
Y
p(t) = (t − tij )2
j=1
and write
p(t) = a0 + a1 t + · · · + a2k t2k .
Consider the vector
a = (a1 , a2 , . . . , a2k , 0, . . . , 0) ∈ Rd
and the hyperplane, H, given by
H = {x ∈ Rd | x · a = −a0 }.
Then, for each j with 1 ≤ j ≤ k, we have
c(tij ) · a = a1 tij + · · · + a2k t2k

ij = p(tij ) − a0 = −a0 ,
and so, c(tij ) ∈ H. On the other hand, for any other point, c(ti ), distinct from any of the
c(tij ), we have
k
Y
c(ti ) · a = −a0 + p(ti ) = −a0 + (ti − tij )2 > −a0 ,
j=1
proving that c(ti ) ∈ H+ . But then, H is a supporting hyperplane of F for Cd (n) and F is a
(k − 1)-face.
Observe that Proposition 10.10 shows that any subset of b d2 c vertices of Cd (n) forms
a face of Cd (n). When a d-polytope has this property it is called a neighborly polytope.
Therefore, cyclic polytopes are neighborly. Proposition 10.10 also shows a phenomenon that
only manifests itself in dimension at least 4: For d ≥ 4, the polytope Cd (n) has n pairwise
adjacent vertices. For n >> d, this is counter-intuitive.
Finally, the combinatorial structure of cyclic polytopes is completely determined as fol-
lows:
Proposition 10.11. (Gale evenness condition, Gale (1963)). Let n and d be integers with
2 ≤ d < n. For any sequence t1 < t2 < · · · < tn , consider the cyclic polytope
Cd (n) = conv(c(t1 ), . . . , c(tn )).
A subset S ⊆ {t1 , . . . , tn } with |S| = d determines a facet of Cd (n) iff for all i < j not in S,
then the number of k ∈ S between i and j is even:
|{k ∈ S | i < k < j}| ≡ 0 (mod 2) for i, j ∈

/S
Proof. Write S = {s1 , . . . , sd } ⊆ {t1 , . . . , tn }. Consider the polyomial

d
Y d
X
q(t) = (t − si ) = bj tj ,
i=1 j=0
let b = (b1 , . . . , bd ), and let H be the hyperplane given by
H = {x ∈ Rd | x · b = −b0 }.
Then, for each i, with 1 ≤ i ≤ d, we have

d
X
c(si ) · b = bj sji = q(si ) − b0 = −b0 ,
j=1
so that c(si ) ∈ H. For all other t 6= si ,
q(t) = c(t) · b + b0 6= 0,
that is, c(t) ∈

/ H. Therefore, F = {c(s1 ), . . . , c(sd )} is a facet of Cd (n) iff {c(t1 ), . . . , c(tn )}−F
lies in one of the two open half-spaces determined by H. This is equivalent to q(t) changing
its sign an even number of times while, increasing t, we pass through the vertices in F .
Therefore, the proposition is proved.
In particular, Proposition 10.11 shows that the combinatorial structure of Cd (n) does
not depend on the specific choice of the sequence t1 < · · · < tn . This justifies our notation
Cd (n).
Here is the celebrated upper bound theorem first proved by McMullen [42].
Theorem 10.12. (Upper Bound Theorem, McMullen (1970)) Let P be any d-polytope with n
vertices. Then, for every k, with 1 ≤ k ≤ d, the polytope P has at most as many (k −1)-faces
as the cyclic polytope Cd (n), that is
fk−1 (P ) ≤ fk−1 (Cd (n)).
Moreover, equality for some k with b d2 c ≤ k ≤ d implies that P is neighborly.

The first step in the proof of Theorem 10.12 is to prove that among all d-polytopes
with a given number n of vertices, the maximum number of i-faces is achieved by simplicial
d-polytopes.
Proposition 10.13. Given any d-polytope P with n-vertices, it is possible to form a sim-
plicial polytope P 0 by perturbing the vertices of P such that P 0 also has n vertices and
fk−1 (P ) ≤ fk−1 (P 0 ) for 1 ≤ k ≤ d.
Furthermore, equality for k > b d2 c can occur only if P is simplicial.
Sketch of proof. First, we apply Proposition 9.15 to triangulate the facets of P without
adding any vertices. Then, we can perturb the vertices to obtain a simplicial polytope P 0
with at least as many facets (and thus, faces) as P .
Proposition 10.13 allows us to restict our attention to simplicial polytopes. Now, it is

obvious that
n
fk−1 ≤
k
for any polytope P (simplicial or not) and we also know that equality holds if k ≤ b d2 c for
neighborly polytopes such as the cyclic polytopes. For k > b d2 c, it turns out that equality
can only be achieved for simplices.
However, for a simplicial polytope, the Dehn–Sommerville equations hk = hd−k together
with the equations (∗) giving fk in terms of the hi ’s show that f0 , f1 , . . . , fb d c already deter-
2
mine the whole f -vector. Thus, it is possible to express the fk−1 in terms of h0 , h1 , . . . , hb d c
2
for k ≥ b d2 c. It turns out that we get
2
bdc
d−i

X ∗ i
fk−1 = + hi ,
i=0
k−i k−d+i
where the meaning of the superscript ∗ is that when d is even we only take half of the last
term for i = d2 and when d is odd we take the whole last term for i = d−1
2
(for details, see
Ziegler [67], Chapter 8). As a consequence if we can show that the neighborly polytopes
maximize not only fk−1 but also hk−1 when k ≤ b d2 c, then the upper bound theorem will be
proved. Indeed, McMullen proved the following theorem which is “more than enough” to
yield the desired result ([42]):
Theorem 10.14. (McMullen (1970)) For every simplicial d-polytope with f0 = n vertices,
we have
n−d−1+k

hk (P ) ≤ for 0 ≤ k ≤ d.
k
Furthermore, equality holds for all l and all k with 0 ≤ k ≤ l iff l ≤ b d2 c, and P is l-
neighborly. (a polytope is l-neighborly iff any subset of l or less vertices determine a face of
P .)
The proof of Theorem 10.14 is too involved to be given here, which is unfortunate since it
is really beautiful. It makes a clever use of shellings and a careful analysis of the h-numbers
of links of vertices. Again, the reader is referred to Ziegler [67], Chapter 8.
Since cyclic d-polytopes are neighborly (which means that they are b d2 c-neighborly), The-
orem 10.12 follows from Proposition 10.13, and Theorem 10.14.
Corollary 10.15. For every simplicial neighborly d-polytope with n vertices, we have
bdc
2
d−i n−d−1+i

X ∗ i
fk−1 = + for 1 ≤ k ≤ d.
i=0
k−i k−d+i i
This gives the maximum number of (k − 1)-faces for any d-polytope with n-vertices, for all
k with 1 ≤ k ≤ d. In particular, the number of facets of the cyclic polytope Cd (n), is
b d2 c
n−d−1+i

X ∗
fd−1 = 2 ,
i=0
i
and more explicitly,

n − b d+1 c n − b d+2 c

fd−1 = 2 + 2 .
n−d n−d
d
Corollary 10.15 implies that the number of facets of any d-polytope is O(nb 2 c ). An
unfortunate consequence of this upper bound is that the complexity of any convex hull
d
algorithms for n points in Ed is O(nb 2 c ).
d
The O(nb 2 c ) upper bound can be obtained more directly using a pretty argument using
shellings due to R. Seidel [52].
Consider any shelling of any simplicial d-polytope, P . For every facet, Fj , of a shelling
either the restriction set Rj or its complement Fj − Rj has at most b d2 c elements. So, either
in the shelling or in the reversed shelling, the restriction set of Fj has at most b d2 c elements.
Moreover, the restriction sets are all distinct, by construction. Thus, the number of facets
is at most twice the number of k-faces of P with k ≤ b d2 c. It follows that
b d2 c
X n
fd−1 ≤ 2
i=0
i
d
and this rough estimate yields a O(nb 2 c ) bound.
Remark: There is also a lower bound theorem due to Barnette (1971, 1973) which gives a
lower bound on the f -vectors all d-polytopes with n vertices. In this case, there is an analog of
the cyclic polytopes called stacked polytopes. These polytopes Pd (n) are simplicial polytopes
obtained from a simplex by building shallow pyramids over the facets of the simplex. Then,
it turns out that if d ≥ 2, then
d d+1

n − k if 0 ≤ k ≤ d − 2
fk ≥ k k+1
(d − 1)n − (d + 1)(d − 2) if k = d − 1.
There has been a lot of progress on the combinatorics of f -vectors and h-vectors since
1971, especially by R. Stanley, G. Kalai and L. Billera, and K. Lee, among others. We
recommend two excellent surveys:
1. Bayer and Lee [5] summarizes progress in this area up to 1993.
2. Billera and Björner [11] is a more advanced survey which reports on results up to 1997.
In fact, many of the chapters in Goodman and O’Rourke [32] should be of interest to the
reader.
Generalizations of the Upper Bound Theorem using sophisticated techniques (face rings)
due to Stanley can be found in Stanley [56] (Chapters II) and connections with toric varieties
can be found in Stanley [56] (Chapters III) and Fulton [28].
Chapter 11
Projective Spaces, Projective

Polyhedra, Polar Duality w.r.t. a
Nondegenerate Quadric
The fact that not just points but also vectors are needed to deal with unbounded polyhedra
is a hint that perhaps the notions of polytope and polyhedra can be unified by “going
projective”. Indeed, the goal of this chapter is to define a notion of projective polyhedron
which is a natural extension of the notion of polyhedron in affine space, and retains many
of the properties of polyhedra.
However, we have to be careful because projective geometry does not accommodate well
the notion of convexity. This is because convexity has to do with convex combinations,
but the essence of projective geometry is that everything is defined up to non-zero scalars,
without any requirement that these scalars be positive.
It is possible to develop a theory of oriented projective geometry (due to J. Stolfi [57]) in
which convexity is nicely accommodated. However, in this approach, every point comes as a
pair, (positive point, negative point), and although it is a very elegant theory, we find it a bit
unwieldy. However, since all we really need is to “embed” Ed into its projective completion,
Pd , so that we can deal with “points at infinity” and “normal points” in a uniform manner
in particular, with respect to projective transformations, we will content ourselves with
a definition of a notion of projective polyhedron using the notion of polyhedral cone. This
notion is just what is needed in Chapter 12 to deal with the correspondence between Voronoi
diagrams and Delaunay triangulations in terms of the lifting to a paraboloid or the lifting
to a sphere. We will not attempt to define a general notion of convexity.
11.1 Projective Spaces

We begin with a “crash course” on (real) projective spaces. There are many texts on pro-
jective geometry. We suggest starting with Gallier [30] and then move on to far more
285
286 CHAPTER 11. PROJECTIVE SPACES AND POLYHEDRA, POLAR DUALITY
comprehensive treatments such as Berger (Geometry II) [8] or Samuel [50].
Definition 11.1. The (real) projective space RPn is the set of all lines through the origin
in Rn+1 , i.e., the set of one-dimensional subspaces of Rn+1 (where n ≥ 0). Since a one-
dimensional subspace L ⊆ Rn+1 is spanned by any nonzero vector u ∈ L, we can view RPn
as the set of equivalence classes of nonzero vectors in Rn+1 − {0} modulo the equivalence
relation,
u ∼ v iff v = λu, for some λ ∈ R, λ 6= 0.
We have the projection p : (Rn+1 − {0}) → RPn given by p(u) = [u]∼ , the equivalence class
of u modulo ∼. Write [u] (or hui) for the line
[u] = {λu | λ ∈ R}
defined by the nonzero vector u. Note that [u]∼ = [u] − {0} for every u 6= 0, so the map
[u]∼ 7→ [u] is a bijection which allows us to identify [u]∼ and [u]. Thus, we will use both
notations interchangeably as convenient.
The projective space RPn is sometimes denoted P(Rn+1 ). Since every line L in Rn+1
intersects the sphere S n in two antipodal points, we can view RPn as the quotient of the
sphere S n by identification of antipodal points. We call this the spherical model of RPn ,
which we illustrate in Figure 11.1.
y x
x
x x
y
(i.) (ii.)
Figure 11.1: The geometric construction for RP1 and RP2 via the identification of antipodal
points of S 1 and S 2 respectively.
A more subtle construction consists in considering the (upper) half-sphere instead of the
sphere, where the upper half-sphere S+n is set of points on the sphere S n such that xn+1 ≥ 0.
This time, every line through the center intersects the (upper) half-sphere in a single point,
except on the boundary of the half-sphere, where it intersects in two antipodal points a+
11.1. PROJECTIVE SPACES 287
and a− . Thus, the projective space RPn is the quotient space obtained from the (upper)
half-sphere S+n by identifying antipodal points a+ and a− on the boundary of the half-sphere.
We call this model of RPn the half-spherical model , which we illustrate in Figure 11.2.
x
x x
(i.)
x
x
(ii.)
Figure 11.2: The geometric construction for RP1 ∼ S 1 and RP2 in terms of the antipodal
boundary points of S+1 and S+2 respectively.
When n = 2, we get a circle. When n = 3, the upper half-sphere is homeomorphic

to a closed disk (say, by orthogonal projection onto the xy-plane), and RP2 is in bijection
with a closed disk in which antipodal points on its boundary (a unit circle) have been
identified. This is hard to visualize! In this model of the real projective space, projective
lines are great semicircles on the upper half-sphere, with antipodal points on the boundary
identified. Boundary points correspond to points at infinity. By orthogonal projection,
these great semicircles correspond to semiellipses, with antipodal points on the boundary
identified. Traveling along such a projective “line,” when we reach a boundary point, we
“wrap around”! In general, the upper half-sphere S+n is homeomorphic to the closed unit
ball in Rn , whose boundary is the (n − 1)-sphere S n−1 . For example, the projective space
RP3 is in bijection with the closed unit ball in R3 , with antipodal points on its boundary
(the sphere S 2 ) identified!
Another useful way of “visualizing” RPn is to use the hyperplane Hn+1 ⊆ Rn+1 of equation
xn+1 = 1. Observe that for every line [u] through the origin in Rn+1 , if u does not belong
to the hyperplan Hn+1 (0) ∼= Rn of equation xn+1 = 0, then [u] intersects Hn+1 is a unique
point, namely

u1 un
,..., ,1 ,
un+1 un+1
where u = (u1 , . . . , un+1 ). The lines [u] for which un+1 = 0 are “points at infinity”. See
Figure 11.3.
[v] ~ [u] ~ [v]~ [u]~
y=1
v u z=1
L∞
(i.)
(ii.)
Figure 11.3: The plane model construction for RP1 and RP2 , where points at infinity corre-
spond to the x-axis and the xy-plane respectively.
Observe that the set of lines in Hn+1 (0) ∼

= Rn is the set of points of the projective space
n−1 n
RP , and so RP can be written as the disjoint union
RPn = Rn q RPn−1 .
We can repeat the above analysis on RPn−1 and so we can think of RPn as the disjoint
union
RPn = Rn q Rn−1 q · · · q R1 q R0 ,
where R0 = {0} consist of a single point. The above shows that there is an embedding
Rn ,→ RPn given by (u1 , . . . , un ) 7→ (u1 , . . . , un , 1).
It will also be very useful to use homogeneous coordinates.
Definition 11.2. Given any point, a = [u]∼ ∈ RPn , the set
{(λu1 , . . . , λun+1 ) | λ 6= 0}
is called the set of homogeneous coordinates of a. Since u 6= 0, observe that for all homoge-
neous coordinates, (u1 , . . . , un+1 ), for a, some ui must be non-zero. The traditional notation
for the homogeneous coordinates of a point a = [u]∼ is
(u1 : · · · : un : un+1 ).
There is a useful bijection between certain kinds of subsets of Rd+1 and subsets of RPd .
For any subset S of Rd+1 , let
−S = {−u | u ∈ S}.
Geometrically, −S is the reflexion of S about 0. Note that for any nonempty subset, S ⊆
Rd+1 , with S 6= {0}, the sets S, −S, and S ∪ −S all induce the same set of points in
projective space RPd , since
p(S − {0}) = {[u]∼ | u ∈ S − {0}}

= {[−u]∼ | u ∈ S − {0}}
= {[u]∼ | u ∈ −S − {0}} = p((−S) − {0})
= x{[u]∼ | u ∈ S − {0}} ∪ {[u]∼ | u ∈ (−S) − {0}} = p((S ∪ −S) − {0}),
because [u]∼ = [−u]∼ . Using these facts we obtain a bijection between subsets of RPd and
certain subsets of Rd+1 .
Definition 11.3. We say that a set S ⊆ Rd+1 is symmetric iff S = −S. Obviously, S ∪ −S
is symmetric for any set S. Say that a subset C ⊆ Rd+1 is a double cone iff for every
u ∈ C − {0}, the entire line [u] spanned by u is contained in C. See Figure 11.4.
u S=C
[u]
-u
-S
Figure 11.4: The double cone for the green curve S.
We exclude the trivial double cone, C = {0}, since the trivial vector space does not yield
a projective space. Thus, every double cone can be viewed as a set of lines through 0. Note
that a double cone is symmetric. Given any nonempty subset, S ⊆ RPd , let v(S) ⊆ Rd+1 be
the set of vectors, [
v(S) = [u]∼ ∪ {0}.
[u]∼ ∈S
Note that v(S) is a double cone.
Proposition 11.1. The map, v : S 7→ v(S), from the set of nonempty subsets of RPd to the
set of nonempty, nontrivial double cones in Rd+1 is a bijection.
Proof. We already noted that v(S) is nontrivial double cone. Consider the map,
ps : S 7→ p(S) = {[u]∼ ∈ RPd | u ∈ S − {0}}.
We leave it as an easy exercise to check that ps ◦ v = id and v ◦ ps = id, which shows that v
and ps are mutual inverses.
Definition 11.4. Given any subspace X ⊆ Rn+1 with dim X = k + 1 ≥ 1 and 0 ≤ k ≤ n,

a k-dimensional projective subspace of RPn is the image Y = p(X − {0}) of X − {0} under
the projection p. We often write Y = P(X). When k = n − 1, we say that Y is a projective
hyperplane or simply a hyperplane. When k = 1, we say that Y is a projective line or simply
a line. See Figure 11.5.
P2
x*
x*
Figure 11.5: In the half-spherical model, a projective line is the maroon semi-circle obtained
by intersecting the hemisphere with a plane through the origin.
It is easy to see that every projective hyperplane, H, is the kernel (zero set) of some
linear equation of the form
a1 x1 + · · · + an+1 xn+1 = 0,
where one of the ai is nonzero, in the sense that

H = {(x1 : · · · : xn+1 ) ∈ RPn | a1 x1 + · · · + an+1 xn+1 = 0}.
Conversely, the kernel of any such linear equation defines a projective hyperplane. Further-
more, given a projective hyperplane, H ⊆ RPn , the linear equation defining H is unique up
to a nonzero scalar.
Definition 11.5. For any i, with 1 ≤ i ≤ n + 1, the set
Ui = {(x1 : · · · : xn+1 ) ∈ RPn | xi 6= 0}
is a subset of RPn called an affine patch of RPn .
We have a bijection, ϕi : Ui → Rn , between Ui and Rn given by

x1 xi−1 xi+1 xn+1
ϕi : (x1 : · · · : xn+1 ) 7→ ,..., , ,..., .
xi xi xi xi
This map is well defined because if (y1 , . . . , yn+1 ) ∼ (x1 , . . . , xn+1 ), that is,
(y1 , . . . , yn+1 ) = λ(x1 , . . . , xn+1 ), with λ 6= 0, then
yj λxj xj
= = (1 ≤ j ≤ n + 1),
yi λxi xi
since λ 6= 0 and xi , yi 6= 0. The inverse, ψi : Rn → Ui ⊆ RPn , of ϕi is given by
ψi : (x1 , · · · , xn ) 7→ (x1 : · · · xi−1 : 1 : xi : · · · : xn ).
Observe that the bijection, ϕi , between Ui and Rn can also be viewed as the bijection

x1 xi−1 xi+1 xn+1
(x1 : · · · : xn+1 ) 7→ ,..., , 1, ,..., ,
xi xi xi xi
between Ui and the hyperplane, Hi ⊆ Rn+1 , of equation xi = 1. We will make heavy use of
these bijections. For example, for any subset, S ⊆ RPn , the “view of S from the patch Ui ”,
S Ui , is in bijection with v(S) ∩ Hi , where v(S) is the double cone associated with S (see
Proposition 11.1).
The affine patches, U1 , . . . , Un+1 , cover the projective space RPn , in the sense that every
(x1 : · · · : xn+1 ) ∈ RPn belongs to one of the Ui ’s, as not all xi = 0. See Figures 11.6 and
11.7. The Ui ’s turn out to be open subsets of RPn and they have nonempty overlaps. When
we restrict ourselves to one of the Ui , we have an “affine view of RPn from Ui .” In particular,
on the affine patch Un+1 , we have the “standard view” of Rn embedded into RPn as Hn+1 ,
the hyperplane of equation xn+1 = 1. The complement Hi (0) of Ui in RPn is the (projective)
hyperplane of equation xi = 0 (a copy of RPn−1 ). With respect to the affine patch Ui , the
hyperplane Hi (0) plays the role of hyperplane (of points) at infinity.
From now on, for simplicity of notation, we will write Pn for RPn . We need to define
projective maps. Such maps are induced by linear maps.
y=1
x=1
(-1,0) (1,0)
Figure 11.6: The space RP1 , visualized by the spherical model, is covered by the two affine
patches y = 1 or U2 , and x = 1 or U1 .
Definition 11.6. Any injective linear map, h : Rm+1 → Rn+1 , induces a map, P(h) : Pm →
Pn , defined by
P(h)([u]∼ ) = [h(u)]∼
and called a projective map. When m = n and h is bijective, the map P(h) is also bijective
and it is called a projectivity.
We have to check that this definition makes sense, that is, it is compatible with the
equivalence relation, ∼. For this, assume that u ∼ v, that is
v = λu,
with λ 6= 0 (of course, u, v 6= 0). As h is linear, we get
h(v) = h(λu) = λh(u),
that is, h(u) ∼ h(v), which shows that [h(u)]∼ does not depend on the representative chosen
in the equivalence class of [u]∼ . It is also easy to check that whenever two linear maps, h1
and h2 , induce the same projective map, i.e., if P(h1 ) = P(h2 ), then there is a nonzero scalar,
λ, so that h2 = λh1 .
Why did we require h to be injective? Because if h has a nontrivial kernel, then, any
nonzero vector u ∈ Ker (h) is mapped to 0, but as 0 does not correspond to any point of
Pn , the map P(h) is undefined on P(Ker (h)).
In some case, we allow projective maps induced by non-injective linear maps h. In this
case, P(h) is a map whose domain is Pn − P(Ker (h)). An example is the map, σN : P3 → P2 ,
given by
(x1 : x2 : x3 : x4 ) 7→ (x1 : x2 : x4 − x3 ),
which is undefined at the point (0 : 0 : 1 : 1). This map is the “homogenization” of the central
projection (from the north pole, N = (0, 0, 1)) from E3 onto E2 .
z=1
x
x
x
x
x x
y=1
x x
x x
x=1
x x
Figure 11.7: The space RP2 , visualized by the spherical model, is covered by the three affine
patches z = 1 or U3 , y = 1 or U2 , and x = 1 or U1 . The plane z = 1 covers everything except
the pink circle x2 + y 2 = 1. The plane y = 1 will cover this circle, excluding the x-intercepts.
These x-intercepts are then covered by x = 1.
Although a projective map, f : Pm → Pn , is induced by some linear map, h, the map f is

not linear! This is because linear combinations of points in Pm do not make any sense!
Another way of defining functions (possibly partial) between projective spaces involves
using homogeneous polynomials. If p1 (x1 , . . . , xm+1 ), . . . , pn+1 (x1 , . . . , xm+1 ) are n + 1 homo-
geneous polynomials all of the same degree d, and if these n + 1 polynomials do not vanish
simultaneously, then we claim that the function f given by
f (x1 : · · · : xm+1 ) = (p1 (x1 , . . . , xm+1 ) : · · · : pn+1 (x1 , . . . , xm+1 ))
is indeed a well-defined function from Pm to Pn . Indeed, if (y1 , . . . , ym+1 ) ∼ (x1 , . . . , xm+1 ),

that is, (y1 , . . . , ym+1 ) = λ(x1 , . . . , xm+1 ), with λ 6= 0, as the pi are homogeneous of degree d,
pi (y1 , . . . , ym+1 ) = pi (λx1 , . . . , λxm+1 ) = λd pi (x1 , . . . , xm+1 ),

and so,
f (y1 : · · · : ym+1 ) = (p1 (y1 , . . . , ym+1 ) : · · · : pn+1 (y1 , . . . , ym+1 ))

= (λd p1 (x1 , . . . , xm+1 ) : · · · : λd pn+1 (x1 , . . . , xm+1 ))
= λd (p1 (x1 , . . . , xm+1 ) : · · · : pn+1 (x1 , . . . , xm+1 ))
= λd f (x1 : · · · : xm+1 ),
which shows that f (y1 : · · · : ym+1 ) ∼ f (x1 : · · · : xm+1 ), as required.

For example, the map, τN : P2 → P3 , given by
(x1 : x2 , : x3 ) 7→ (2x1 x3 : 2x2 x3 : x21 + x22 − x23 : x21 + x22 + x23 ),
is well-defined. It turns out to be the “homogenization” of the inverse stereographic map

from E2 to S 2 (see Section 12.5). Observe that
τN (x1 : x2 : 0) = (0 : 0 : x21 + x22 : x21 + x22 ) = (0 : 0 : 1 : 1),
that is, τN maps all the points at infinity (in H3 (0)) to the “north pole,” (0 : 0 : 1 : 1).
However, when x3 6= 0, we can prove that τN is injective (in fact, its inverse is σN , defined
earlier).
Most interesting subsets of projective space arise as the collection of zeros of a (finite)
set of homogeneous polynomials. Let us begin with a single homogeneous polynomial,
p(x1 , . . . , xn+1 ), of degree d and set
V (p) = {(x1 : · · · : xn+1 ) ∈ Pn | p(x1 , . . . , xn+1 ) = 0}.
As usual, we need to check that this definition does not depend on the specific representative
chosen in the equivalence class of [(x1 , . . . , xn+1 )]∼ . If (y1 , . . . , yn+1 ) ∼ (x1 , . . . , xn+1 ), that
is, (y1 , . . . , yn+1 ) = λ(x1 , . . . , xn+1 ), with λ 6= 0, as p is homogeneous of degree d,
p(y1 , . . . , yn+1 ) = p(λx1 , . . . , λxn+1 ) = λd p(x1 , . . . , xn+1 ),
and as λ 6= 0,
p(y1 , . . . , yn+1 ) = 0 iff p(x1 , . . . , xn+1 ) = 0,
which shows that V (p) is well defined.
Definition 11.7. For a set of homogeneous polynomials (not necessarily of the same degree)
E = {p1 (x1 , . . . , xn+1 ), . . . , ps (x1 , . . . , xn+1 )}, we set
s
\
V (E) = V (pi ) = {(x1 : · · · : xn+1 ) ∈ Pn | pi (x1 , . . . , xn+1 ) = 0, i = 1 . . . , s}.
i=1
The set, V (E), is usually called the projective variety defined by E (or cut out by E). When
E consists of a single polynomial p, the set V (p) is called a (projective) hypersurface.
For example, if
p(x1 , x2 , x3 , x4 ) = x21 + x22 + x23 − x24 ,
then V (p) is the projective sphere in P3 , denoted S f2 . Indeed, if we “look” at V (p) on the
affine patch U4 , where x4 6= 0, we know that this amounts to setting x4 = 1, and we do
get the set of points (x1 , x2 , x3 , 1) ∈ U4 satisfying x21 + x22 + x23 − 1 = 0, our usual 2-sphere!
However, if we look at V (p) on the patch U1 , where x1 6= 0, we see the quadric of equation
1 + x22 + x23 = x24 , which is not a sphere but a hyperboloid of two sheets! Nevertheless, if we
pick x4 = 0 as the plane at infinity, note that the projective sphere does not have points at
infinity since the only real solution of x21 + x22 + x23 = 0 is (0, 0, 0), but (0, 0, 0, 0) does not
correspond to any point of P3 .
Another example is given by
q = (x1 , x2 , x3 , x4 ) = x21 + x22 − x3 x4 ,
for which V (q) corresponds to a paraboloid in the patch U4 . Indeed, if we set x4 = 1, we get
the set of points in U4 satisfying x3 = x21 + x22 . For this reason, we denote V (q) by P
e and
call it a (projective) paraboloid .
Definition 11.8. Given any homogeneous polynomial F (x1 , . . . , xd+1 ), we will also make
use of the hypersurface cone C(F ) ⊆ Rd+1 , defined by
C(F ) = {(x1 , . . . , xd+1 ) ∈ Rd+1 | F (x1 , . . . , xd+1 ) = 0}.
Observe that V (F ) = P(C(F )).
Remark: Every variety V (E), defined by a set of polynomials, E = {p1 (x1 , . . . , xn+1 ), . . .,
ps (x1 , . . . , xn+1 )}, is also the hypersurface defined by the single polynomial equation
p21 + · · · + p2s = 0.
This fact, peculiar to the real field R is a mixed blessing. On the one-hand, the study of
varieties is reduced to the study of hypersurfaces. On the other-hand, this is a hint that we
should expect that such a study will be hard.
Perhaps to the surprise of the novice, there is a bijective projective map (a projectivity)
f2 to P.
sending S e This map, θ, is given by
θ(x1 : x2 : x3 : x4 ) = (x1 : x2 : x3 + x4 : x4 − x3 ),
whose inverse is given by
x3 − x4 x3 + x4

−1
θ (x1 : x2 : x3 : x4 ) = x1 : x2 : : .
2 2
Indeed, if (x1 : x2 : x3 : x4 ) satisfies

x21 + x22 + x23 − x24 = 0,
and if (z1 : z2 : z3 : z4 ) = θ(x1 : x2 : x3 : x4 ), then from above,
z3 − z4 z3 + z4

(x1 : x2 : x3 : x4 ) = z1 : z2 : : ,
2 2
and by plugging the right-hand sides in the equation of the sphere, we get
2 2
z3 − z4

2 2 z3 + z4 1
z1 + z2 + − = z12 + z22 + (z32 + z42 − 2z3 z4 − (z32 + z42 + 2z3 z4 ))
2 2 4
2 2
= z1 + z2 − z3 z4 = 0,
which is the equation of the paraboloid P.
e
11.2 Projective Polyhedra

Following the “projective doctrine” which consists in replacing points by lines through the
origin, that is, to “conify” everything, we will define a projective polyhedron as any set of
points in Pd induced by a polyhedral cone in Rd+1 . To do so, it is preferable to consider
cones as sets of positive combinations of vectors (see Definition 5.3). Just to refresh our
memory, a set, C ⊆ Rd , is a V-cone or polyhedral cone if C is the positive hull of a finite set
of vectors, that is,
C = cone({u1 , . . . , up }),
for some vectors, u1 , . . . , up ∈ Rd . An H-cone is any subset of Rd given by a finite intersection
of closed half-spaces cut out by hyperplanes through 0.
A good place to learn about cones (and much more) is Fulton [28]. See also Ewald [26].
By Theorem 5.19, V-cones and H-cones form the same collection of convex sets (for every
d ≥ 0). Naturally, we can think of these cones as sets of rays (half-lines) of the form
hui+ = {λu | λ ∈ R, λ ≥ 0},
where u ∈ Rd is any nonzero vector. We exclude the trivial cone, {0}, since 0 does not define
any point in projective space. When we “go projective,” each ray corresponds to the full
line, hui, spanned by u which can be expressed as
hui = hui+ ∪ −hui+ ,
where −hui+ = hui− = {λu | λ ∈ R, λ ≤ 0}. Now, if C ⊆ Rd is a polyhedral cone, obviously
−C is also a polyhedral cone and the set C ∪ −C consists of the union of the two polyhedral
cones C and −C. Note that C ∪ −C can be viewed as the set of all lines determined by the
nonzero vectors in C (and −C). It is a double cone. Unless C is a closed half-space, C ∪ −C
is not convex. See Figure 11.8. It seems perfectly natural to define a projective polyhedron
as any set of lines induced by a set of the form C ∪ −C, where C is a polyhedral cone.
11.2. PROJECTIVE POLYHEDRA 297
(0,1,1)
(1,0,1)
(1,1,1)
C
(-1,-1,-1)
(-1,0,-1)
-C
(0,-1,-1)
(0,1,1)
(1,0,1) (1,1,1)
(-1,0,-1)
(-1,-1,-1)
-C
(0,-1,-1)
Figure 11.8: The double cone C ∪ −C, where C is the V-cone C = cone{(1, 0, 1),
(0, 1, 1), (1, 1, 1)}.
Definition 11.9. A projective polyhedron is any subset, P ⊆ Pd , of the form
P = p((C ∪ −C) − {0}) = p(C − {0}),
where C is any polyhedral cone (V or H cone) in Rd+1 (with C 6= {0}). We write

P = P(C ∪ −C) or P = P(C). See Figure 11.9.
It is important to observe that because C∪−C is a double cone there is a bijection between
nontrivial double polyhedral cones and projective polyhedra. So, projective polyhedra are
equivalent to double polyhedral cones. However, the projective interpretation of the lines
induced by C ∪ −C as points in Pd makes the study of projective polyhedra geometrically
more interesting.
Projective polyhedra inherit many of the properties of cones but we have to be careful
because we are really dealing with double cones, C ∪ −C, and not cones. As a consequence,
there are a few unpleasant surprises, for example, the fact that the collection of projective
polyhedra is not closed under intersection!
Before dealing with these issues, let us show that every “standard” polyhedron P ⊆ Ed
has a natural projective completion, Pe ⊆ Pd , such that on the affine patch Ud+1 (where xd+1 6=
0), Pe Ud+1 = P . For this, we use our theorem on the Polyhedron–Cone Correspondence
(Theorem 5.20, part (2)).
Let A = X + U , where X is a set of points in Ed and U is a cone in Rd . For every point,
x ∈ X, and every vector, u ∈ U , let

x u
x
b= , u
b= ,
1 0
b = {b
and let X x | x ∈ X}, U
b = {b
u | u ∈ U } and Ab = {b
a | a ∈ A}, with

a
a= .
1
b
Then,
b ∪U
C(A) = cone({X b })
is a cone in Rd+1 such that
b = C(A) ∩ Hd+1 ,
A
where Hd+1 is the hyperplane of equation xd+1 = 1. If we set A e = P(C(A)), then we get
a subset of Pd and in the patch Ud+1 , the set A
e Ud+1 is in bijection with the intersection
(C(A) ∪ −C(A)) ∩ Hd+1 = A, b and thus, in bijection with A.
e the projective completion of A. We have an injection, A −→ A,
We call A e given by
(a1 , . . . , ad ) 7→ (a1 : · · · : ad : 1),
which is just the map, ψd+1 : Rd → Ud+1 .

What the projective completion does is to add to A the “points at infinity” corresponding
to the vectors in U , that is, the points of Pd corresponding to the lines in the cone, U .
Definition 11.10. If X = conv(Y ) and U = cone(V ) for some finite sets Y = {y1 , . . . , yp }
and V = {v1 , . . . , vq }, then P = conv(Y ) + cone(V ) is a V-polyhedron and Pe = P(C(P ))
is a projective polyhedron. The projective polyhedron Pe = P(C(P )) is called the projective
completion of P . See Figure 11.9.
Observe that if C is a closed half-space in Rd+1 , then P = P(C ∪ −C) = Pd . Now, if

C ⊆ Rd+1 is a polyhedral cone and C is contained in a closed half-space, it is still possible that
C contains some nontrivial linear subspace and we would like to understand this situation.
The first thing to observe is that U = C ∩ (−C) is the largest linear subspace contained
in C.
Definition 11.11. If C ∩ (−C) = {0}, we say that C is a pointed or strongly convex cone.
In this case, one immediately realizes that 0 is an extreme point of C and so, there is a
hyperplane, H, through 0 so that C ∩ H = {0}, that is, except for its apex, C lies in one of
the open half-spaces determined by H. As a consequence, by a linear change of coordinates,
we may assume that this hyperplane is Hd+1 (0) and so, for every projective polyhedron,
(-1,1) (1,1)
A
(-1,0) (1,0)
E
(-1,0,1) D
B
(1,0,1)
(-1,1,0)
C(A)
z=1
(1,1,0)
C
B E
P = P 2(C(A) )
C D
Figure 11.9: The bottom figure shows a projective polyhedron, which is the projective
completion (in the halfsphere model of P2 ) of the infinite trough A = X + U , where
X = {(−1, 0), (1, 0)} and U = {(−1, 1), (1, 1)}.
P = P(C), if C is pointed then there is an affine patch (say, Ud+1 ) where P has no points at
infinity, that is, P is a polytope! On the other hand, from another patch, Ui , as P Ui is in
bijection with (C ∪ −C) ∩ Hi , the projective polyhedron P viewed on Ui may consist of two
disjoint polyhedra.
The situation is very similar to the classical theory of projective conics or quadrics (for
example, see Brannan, Esplen and Gray, [15]). The case where C is a pointed cone corre-
sponds to the nondegenerate conics or quadrics. In the case of the conics, depending how
we slice a cone, we see an ellipse, a parabola or a hyperbola.
For projective polyhedra, when we slice a polyhedral double cone, C ∪ −C, we may see
a polytope (elliptic type) a single unbounded polyhedron (parabolic type) or two unbounded
polyhedra (hyperbolic type). See Figure 11.10.
Now, when U = C ∩ (−C) 6= {0}, the polyhedral cone, C, contains the linear subspace,
U , and if C 6= Rd+1 , then for every hyperplane, H, such that C is contained in one of the two
closed half-spaces determined by H, the subspace U ∩ H is nontrivial. An example is the
cone, C ⊆ R3 , determined by the intersection of two planes through 0 (a wedge). In this case,
U is equal to the line of intersection of these two planes. Also observe that C ∩ (−C) = C
(0,1,1)
(1,0,1) (1,1,1)
R
C P Q
(-1,0,-1) (0,1,1)
(-1,-1,-1)
-C (1,0,1)
(0,-1,-1)
(i.)
(1,1,1)
C W
(-1,0,-1)
-C
(0,-1,-1)
(ii.)
(0,1,1)
(1,0,1)
(1,1,1)
C S
T
(-1,0,-1)
(-1,-1,-1)
-C
(0,-1,-1)
(iii.)
Figure 11.10: For the sea green double cone C ∪ −C of Figure 11.8, Figure (i.) illustrates
an elliptic type polytope, Figure (ii.) illustrates a parabolic type polyhedron, while Figure
(iii.) illustrates a hyperbolic type polyhedron.
iff C = −C, that is, iff C is a linear subspace.

The situation where C ∩ (−C) 6= {0} is reminiscent of the case of cylinders in the theory
of quadric surfaces (see Brannan, Esplen and Gray [15] or Berger [8]). Now, every cylinder
can be viewed as the ruled surface defined as the family of lines orthogonal to a plane and
touching some nondegenerate conic.
A similar decomposition holds for polyhedral cones as shown below in a proposition
borrowed from Ewald [26] (Chapter V, Lemma 1.6). We should warn the reader that we
have some doubts about the proof given there, and so we offer a different proof adapted
from the proof of Lemma 16.2 in Barvinok [4]. See Figure 11.11. Given any two subsets,
V, W ⊆ Rd , as usual, we write V +W = {v+w | v ∈ V, w ∈ W } and v+W = {v+w | w ∈ W },
for any v ∈ Rd .
Proposition 11.2. For every polyhedral cone C ⊆ Rd , if U = C ∩ (−C), then there is some
pointed cone C0 so that U and C0 are orthogonal and
C = U + C0 ,
with dim(U ) + dim(C0 ) = dim(C).
-C
C0 C = U + C0
Figure 11.11: In R3 , C is the cone determined by the pink and peach half planes, and
U = C ∩ −C is the red line of intersection. Then C = U + C0 , where C0 it the peach pointed
cone contained in plane perpendicular to U .
Proof. We already know that U = C ∩ (−C) is the largest linear subspace of C. Let U ⊥ be
the orthogonal complement of U in Rd and let π : Rd → U ⊥ be the orthogonal projection
onto U ⊥ . By Proposition 5.13, the projection, C0 = π(C), of C onto U ⊥ is a polyhedral
cone. We claim that C0 is pointed and that
C = U + C0 .
Since π −1 (v) = v + U for every v ∈ C0 , we have U + C0 ⊆ C. On the other hand, by

definition of C0 , we also have C ⊆ U + C0 , so C = U + C0 . If C0 was not pointed, then
it would contain a linear subspace, V , of dimension at least 1 but then, U + V would be
a linear subspace of C of dimension strictly greater than U , which is impossible. Finally,
dim(U ) + dim(C0 ) = dim(C) is obvious by orthogonality.
Definition 11.12. The linear subspace U = C ∩ (−C) is called the cospan of C.
Both U and C0 are uniquely determined by C. To a great extent, Proposition 11.2 reduces
the study of non-pointed cones to the study of pointed cones.
Definition 11.13. We call the projective polyhedra of the form P = P(C), where C is
a cone with a non-trivial cospan (a non-pointed cone) a projective polyhedral cylinder , by
analogy with the quadric surfaces. We also propose to call the projective polyhedra of the
form P = P(C), where C is a pointed cone, a projective polytope (or nondegenerate projective
polyhedron).
The following propositions show that projective polyhedra behave well under projective
maps and intersection with a hyperplane:
Proposition 11.3. Given any projective map, h : Pm → Pn , for any projective polyhedron,
P ⊆ Pm , the image, h(P ), of P is a projective polyhedron in Pn . Even if h : Pm → Pn is a
partial map but h is defined on P , then h(P ) is a projective polyhedron.
Proof. The projective map, h : Pm → Pn , is of the form h = P(b h), for some injective linear
m+1 n+1
map, h : R
b → R . Moreover, the projective polyhedron, P , is of the form P = P(C),
for some polyhedral cone, C ⊆ Rm+1 , with C = cone({u1 , . . . , up }), for some nonzero vector
ui ∈ Rm+1 . By definition,
P(h)(P ) = P(h)(P(C)) = P(b

h(C)).
h is linear,
As b
b h(cone({u1 , . . . , up })) = cone({b
h(C) = b h(u1 ), . . . , b
h(up )}).
If we let C
b = cone({b
h(u1 ), . . . , b
h(up )}), then b
h(C) = C
b is a polyhedral cone and so,
P(h)(P ) = P(b
h(C)) = P(C)
b
is a projective cone. This argument does not depend on the injectivity of b

h, as long as
C ∩ Ker (b
h) = {0}.
Proposition 11.3 together with earlier arguments shows that every projective polytope,
P ⊆ Pd , is equivalent under some suitable projectivity to another projective polytope, P 0 ,
which is a polytope when viewed in the affine patch, Ud+1 . This property is similar to the
fact that every (non-degenerate) projective conic is projectively equivalent to an ellipse.
Since the notion of a face is defined for arbitrary polyhedra it is also defined for cones.
Consequently, we can define the notion of a face for projective polyhedra.
Definition 11.14. Given a projective polyhedron P ⊆ Pd , where P = P(C) for some

polyhedral cone (uniquely determined by P ) C ⊆ Rd+1 , a face of P is any subset of P of the
form P(F ) = p(F − {0}), for any nontrivial face F ⊆ C of C (F 6= {0}). Consequently, we
say that P(F ) is a vertex iff dim(F ) = 1, an edge iff dim(F ) = 2 and a facet iff dim(F ) =
dim(C) − 1. The projective polyhedron P and the empty set are the improper faces of P .
If C is strongly convex, then it is easy to prove that C is generated by its edges (its one-
dimensional faces, these are rays) in the sense that any set of nonzero vectors spanning these
edges generates C (using positive linear combinations). As a consequence, if C is strongly
convex, we may say that P is “spanned” by its vertices, since P is equal to P(all positive
combinations of vectors representing its edges).
Remark: Even though we did not define the notion of convex combination of points in Pd ,
the notion of projective polyhedron gives us a way to mimic certain properties of convex
sets in the framework of projective geometry. That’s because every projective polyhedron
corresponds to a unique polyhedral cone.
If our projective polyhedron is the completion Pe = P(C(P )) ⊆ Pd of some polyhedron
P ⊆ Rd , then each face of the cone C(P ) is of the form C(F ), where F is a face of P and
so, each face of Pe is of the form P(C(F )), for some face F of P . In particular, in the affine
patch Ud+1 the face P(C(F )) is in bijection with the face F of P . We will usually identify
P(C(F )) and F .
We now consider the intersection of projective polyhedra but first, let us make some
general remarks about the intersection of subsets of Pd . Given any two nonempty subsets,
P(S) and P(S 0 ), of Pd where S and S 0 are polyhedral cones (or more generally cones with
vertex 0), what is P(S) ∩ P(S 0 )? It is tempting to say that
P(S) ∩ P(S 0 ) = P(S ∩ S 0 ),
but unfortunately this is generally false! The problem is that P(S) ∩ P(S 0 ) is the set of all
lines determined by vectors both in S and S 0 but there may be some line spanned by some
vector u ∈ (−S) ∩ S 0 or u ∈ S ∩ (−S 0 ) such that u does not belong to S ∩ S 0 or −(S ∩ S 0 ).
Observe that
−(−S) = S
−(S ∩ S 0 ) = (−S) ∩ (−S 0 ).
Then, the correct intersection is given by
(S ∪ −S) ∩ (S 0 ∪ −S 0 ) = (S ∩ S 0 ) ∪ ((−S) ∩ (−S 0 )) ∪ (S ∩ (−S 0 )) ∪ ((−S) ∩ S 0 )

= (S ∩ S 0 ) ∪ −(S ∩ S 0 ) ∪ (S ∩ (−S 0 )) ∪ −(S ∩ (−S 0 )),
which is the union of two double cones (except for 0, which belongs to both). Therefore, if
P(S) ∩ P(S 0 ) 6= ∅, then S ∩ S 0 6= {0} or S ∩ (−S 0 ) 6= {0}, and so
P(S) ∩ P(S 0 ) = P(S ∩ S 0 ) ∪ P(S ∩ (−S 0 )) = P(S ∩ S 0 ) ∪ P((−S) ∩ S 0 ),
since P(S ∩(−S 0 )) = P((−S)∩S 0 ), with the understanding that if S ∩S = {0} or S ∩(−S 0 ) =
{0}, then the corresponding term should be omitted.
Furthermore, if S 0 is symmetric (i.e., S 0 = −S 0 ), then
(S ∪ −S) ∩ (S 0 ∪ −S 0 ) = (S ∪ −S) ∩ S 0
= (S ∩ S 0 ) ∪ ((−S) ∩ S 0 )
= (S ∩ S 0 ) ∪ −(S ∩ (−S 0 ))
= (S ∩ S 0 ) ∪ −(S ∩ S 0 ).
Thus, if either S or S 0 is symmetric and if P(S) ∩ P(S 0 ) 6= ∅ then
P(S) ∩ P(S 0 ) = P(S ∩ S 0 ).
Now, if C is a pointed polyhedral cone then C ∩ (−C) = {0}. Consequently, for any other
polyhedral cone C 0 we have (C ∩ C 0 ) ∩ ((−C) ∩ C 0 ) = {0}. Using these facts and adopting
the convention that P({0}) = ∅, we obtain the following result:
Proposition 11.4. Let P = P(C) and P 0 = P(C 0 ) be any two projective polyhedra in Pd . If
P(C) ∩ P(C 0 ) 6= ∅, then the following properties hold:
(1)
P(C) ∩ P(C 0 ) = P(C ∩ C 0 ) ∪ P(C ∩ (−C 0 )),
the union of two projective polyhedra. If C or C 0 is a pointed cone i.e., P or P 0 is a
projective polytope, then P(C ∩ C 0 ) and P(C ∩ (−C 0 )) are disjoint (if both are defined).
(2) If P 0 = H for some hyperplane H ⊆ Pd , then P ∩ H is a projective polyhedron.
1
a1= (-1, 1) a1= (-1, 1) P (C) a2= (1, 1)
a2= (1, 1)
C C
-C -C
-a 2 -a 1 -a 2 -a 1
-b2 -b2 b1= (1,2)

b1= (1,2)
P 1 (C’)
C‘ C‘
- C‘ - C‘
b = (1, -2)
b = (1, -2) 2
-b 1 2 -b 1
Figure 11.12: Let C = cone{(1, 1), (−1, 1)} and C 0 = cone{(1, 2), (1, −2)}. In the half-
spherical model of P1 , P(C) is the bold red arc, while P(C 0 ) is the bold blue arc.
b1= (1,2)
-b2 b1= (1,2)

P 1 (C h C’ )
a1= (-1, 1) a2= (1, 1)
a1= (-1, 1) a2= (1, 1)

C
C‘
C
C‘
- C‘
-C
-a 2 -a 1
b = (1, -2)
2
b = (1, -2)
2
-b 1 1
P (C) h P 1 (C’)
-b2
P 1 (C h (-C’))
a1= (-1, 1) a2= (1, 1)
- C‘
-b 1
Figure 11.13: For the cones C and C 0 defined in Figure 11.12, P(C) ∩ P(C 0 ) is illustrated
by two disjoint purple arcs; the light purple arc is P(C ∩ C 0 ), while the dark purple arc is
P(C ∩ (−C 0 )) .
Proof. We already proved (1) so only (2) remains to be proved. Of course, we may assume
that P 6= Pd . This time, using the equivalence theorem of T
V-cones and H-cones (Theorem
5.19), we know that P is of the form P = P(C), with C = pi=1 Ci , where the Ci are closed
half-spaces in Rd+1 . Moreover, H = P(H),
b for some hyperplane, Hb ⊆ Rd+1 , through 0. Now,
as H
b is symmetric,
P ∩ H = P(C) ∩ P(H) b = P(C ∩ H).
b
Consequently,
P ∩ H = P(C ∩ H)
b
p
! !
\
= P Ci ∩ H
b .
i=1
However, H b+ ∩ H
b =H b − , where H
b + and H
b − are the two closed half-spaces determined by H
b
and so,
p p
! !
\ \
C
b= Ci ∩ H b= Ci ∩ H b+ ∩ Hb−
i=1 i=1
is a polyhedral cone. Therefore, P ∩ H = P(C)

b is a projective polyhedron.
Proposition 11.4 can be sharpened a little.

Proposition 11.5. Let P = P(C) and P 0 = P(C 0 ) be any two projective polyhedra in Pd . If
P(C) ∩ P(C 0 ) 6= ∅, then
P(C) ∩ P(C 0 ) = P(C ∩ C 0 ) ∪ P(C ∩ (−C 0 )),
the union of two projective polyhedra. If C = −C, i.e., C is a linear subspace (or if C 0 is a
linear subspace), then
P(C) ∩ P(C 0 ) = P(C ∩ C 0 ).
Furthermore, if either C or C 0 is pointed, the above projective polyhedra are disjoint, else if
C and C 0 both have nontrivial cospan and P(C ∩ C 0 ) and P(C ∩ (−C 0 )) intersect then
P(C ∩ C 0 ) ∩ P(C ∩ (−C 0 )) = P(C ∩ (C 0 ∩ (−C 0 ))) ∪ P(C 0 ∩ (C ∩ (−C))).
Finally, if the two projective polyhedra on the right-hand side intersect, then
P(C ∩ (C 0 ∩ (−C 0 ))) ∩ P(C 0 ∩ (C ∩ (−C))) = P((C ∩ (−C)) ∩ (C 0 ∩ (−C 0 ))).
Proof. Left as a simple exercise in boolean algebra.
In preparation for Section 12.7, we also need the notion of tangent space at a point of a
variety.
11.3 Tangent Spaces of Hypersurfaces and Projective

Hypersurfaces
Since we only need to consider the case of hypersurfaces we restrict attention to this case
(but the general case is a straightforward generalization). Let us begin with a hypersurface
of equation p(x1 , . . . , xd ) = 0 in Rd , that is, the set
S = V (p) = {(x1 , . . . , xd ) ∈ Rd | p(x1 , . . . , xd ) = 0},
where p(x1 , . . . , xd ) is a polynomial of total degree m.

Pick any point a = (a1 , . . . , ad ) ∈ Rd . Recall that there is a version of the Taylor
expansion formula for polynomials such that, for any polynomial p(x1 , . . . , xd ) of total degree
m, for every h = (h1 , . . . , hd ) ∈ Rd , we have
X Dα p(a) α
p(a + h) = p(a) + h
α!
1≤|α|≤m
d
X X Dα p(a) α
= p(a) + pxi (a)hi + h ,
i=1
α!
2≤|α|≤m
11.3. TANGENT SPACES OF HYPERSURFACES 307
where we use the multi-index notation, with α = (i1 , . . . , id ) ∈ Nd , |α| = i1 + · · · + id ,

α! = i1 ! · · · id !, hα = hi11 · · · hidd ,
∂ i1 +···+id p
Dα p(a) = (a),
∂xi11 · · · ∂xidd
and
∂p
pxi (a) = (a).
∂xi
Consider any line ` through a, given parametrically by
` = {a + th | t ∈ R},
with h 6= 0 and say a ∈ S is a point on the hypersurface S = V (p), which means that
p(a) = 0. The intuitive idea behind the notion of the tangent space to S at a is that it is
the set of lines that intersect S at a in a point of multiplicity at least two, which means that
the equation giving the intersection, S ∩ `, namely
p(a + th) = p(a1 + th1 , . . . , ad + thd ) = 0,
is of the form
t2 q(a, h)(t) = 0,
where q(a, h)(t) is some polynomial in t. Using Taylor’s formula, as p(a) = 0, we have
d
X
p(a + th) = t pxi (a)hi + t2 q(a, h)(t),
i=1
for some polynomial q(a, h)(t). From this, we see that a is an intersection point of multiplicity
at least 2 iff
Xd
pxi (a)hi = 0. (†)
i=1
Consequently, if ∇p(a) = (px1 (a), . . . , pxd (a)) 6= 0 (that is, if the gradient of p at a is
nonzero), we see that ` intersects S at a in a point of multiplicity at least 2 iff h belongs to
the hyperplane of equation (†).
Definition 11.15. Let S = V (p) be a hypersurface in Rd . For any point a ∈ S, if ∇p(a) 6= 0,
then we say that a is a non-singular point of S. When a is nonsingular, the (affine) tangent
space Ta (S) (or simply Ta S) to S at a is the hyperplane through a of equation
d
X
pxi (a)(xi − ai ) = 0.
i=1
See Figure 11.14.

Vp(a)
x-a x
Ta(S)
a
S = { (x,y,z) | p(x,y,z) = 0 }
Figure 11.14: The tangent plane to the surface S = V (p).
Observe that the hyperplane of the direction of Ta S is the hyperplane through 0 and
parallel to Ta S given by
Xd
pxi (a)xi = 0.
i=1
When ∇p(a) = 0, we either say that Ta S is undefined or we set Ta S = Rd .

We now extend the notion of tangent space to projective varieties. As we will see, this
amounts to homogenizing and the result turns out to be simpler than the affine case!
Definition 11.16. Let S = V (F ) ⊆ Pd be a projective hypersurface, which means that
S = V (F ) = {(x1 : · · · : xd+1 ) ∈ Pd | F (x1 , . . . , xd+1 ) = 0},
where F (x1 , . . . , xd+1 ) is a homogeneous polynomial of total degree m. We say that a point
a ∈ S is non-singular iff ∇F (a) = (Fx1 (a), . . . , Fxd+1 (a)) 6= 0.
For every i = 1, . . . , d + 1, let

xj
zji = ,
xi
where j = 1, . . . , d + 1 and j 6= i, and let f i be the result of “dehomogenizing” F at i, that
is,
f i (z1i , . . . , zi−1
i i
, zi+1 i
, . . . , zd+1 ) = F (z1i , . . . , zi−1
i i
, 1, zi+1 i
, . . . , zd+1 ).
Definition 11.17. We define the (projective) tangent space Ta S to a at S as the hyperplane

H such that for each affine patch Ui where ai 6= 0, if we let
aj
ai
j = ,
ai
where j = 1, . . . , d + 1 and j 6= i, then the restriction H Ui of H to Ui is the affine
hyperplane tangent to S Ui given by
d+1
X
fzii (ai )(zji − ai
j ) = 0.
j
j=1
j6=i
Thus, on the affine patch Ui , the tangent space Ta S is given by the homogeneous equation
d+1
X
fzii (ai )(xj − ai
j xi ) = 0.
j
j=1
j6=i
This looks awful but we can make it pretty if we remember that F is a homogeneous
polynomial of degree m and that we have the Euler relation:
d+1
X
Fxj (a)aj = mF (a),
j=1
for every a = (a1 , . . . , ad+1 ) ∈ Rd+1 . Using this, we can come up with a clean equation for
our projective tangent hyperplane. It is enough to carry out the computations for i = d + 1.
Our tangent hyperplane has the equation
d
X
Fxj (ad+1
1 , . . . , ad+1
d , 1)(xj − ad+1
j xd+1 ) = 0,
j=1
that is,
d
X d
X
Fxj (ad+1
1 , . . . , ad+1
d , 1)xj + Fxj (ad+1
1 , . . . , ad+1
d , 1)(−ad+1
j xd+1 ) = 0.
j=1 j=1
As F (x1 , . . . , xd+1 ) is homogeneous of degree m, and as ad+1 6= 0 on Ud+1 , we have

d+1
am
d+1 F (a1 , . . . , ad+1
d , 1) = F (a1 , . . . , ad , ad+1 ),
so from the above equation we get

d
X d
X
Fxj (a1 , . . . , ad+1 )xj + Fxj (a1 , . . . , ad+1 )(−ad+1
j xd+1 ) = 0. (∗)
j=1 j=1
Since a ∈ S, we have F (a) = 0, so the Euler relation yields

d
X
Fxj (a1 , . . . , ad+1 )aj + Fxd+1 (a1 , . . . , ad+1 )ad+1 = 0,
j=1
which, by dividing by ad+1 and multiplying by xd+1 , yields

d
X
Fxj (a1 , . . . , ad+1 )(−ad+1
j xd+1 ) = Fxd+1 (a1 , . . . , ad+1 )xd+1 ,
j=1
and by plugging this in (∗), we get

d
X
Fxj (a1 , . . . , ad+1 )xj + Fxd+1 (a1 , . . . , ad+1 )xd+1 = 0.
j=1
Consequently, the tangent hyperplane to S at a is given by the equation

d+1
X
Fxj (a)xj = 0.
j=1
Definition 11.18. Let S = V (F ) be a hypersurface in Pd , where F (x1 , . . . , xd+1 ) is a

homogeneous polynomial. For any point a ∈ S, if ∇F (a) 6= 0, then we say that a is a
non-singular point of S. When a is nonsingular, the (projective) tangent space Ta (S) (or
simply Ta S) to S at a is the hyperplane through a of equation
d+1
X
Fxi (a)xi = 0.
i=1
For example, if we consider the sphere S 2 ⊆ P3 of equation
x2 + y 2 + z 2 − w2 = 0,
the tangent plane to S 2 at a = (a1 , a2 , a3 , a4 ) is given by
a1 x + a2 y + a3 z − a4 w = 0.
Remark: If a ∈ S = V (F ), as F (a) = d+1

P
i=1 Fxi (a)ai = 0 (by Euler), the equation of the
tangent plane Ta S to S at a can also be written as
d+1
X
Fxi (a)(xi − ai ) = 0.
i=1
Now, if a = (a1 : · · · : ad : 1) is a point in the affine patch Ud+1 , then the equation of the
intersection of Ta S with Ud+1 is obtained by setting ad+1 = xd+1 = 1, that is
d
X
Fxi (a1 , . . . , ad , 1)(xi − ai ) = 0,
i=1
which is just the equation of the affine hyperplane to S ∩ Ud+1 at a ∈ Ud+1 .

It will be convenient to adopt the following notational convention: Given any point
x = (x1 , . . . , xd ) ∈ Rd written as a row vector, we let x denote the corresponding column
vector such that x> = x.
Projectivities behave well with respect to hypersurfaces and their tangent spaces. Let
S = V (F ) ⊆ Pd be a projective hypersurface, where F is a homogeneous polynomial of
degree m and let h : Pd → Pd be a projectivity (a bijective projective map). Assume that h
is induced by the invertible (d + 1) × (d + 1) matrix, A = (ai j ), and write A−1 = (a−1
i j ). For
d+1
any hyperplane, H ⊆ R , if ϕ is any linear from defining ϕ, i.e., H = Ker (ϕ), then
h(H) = {h(x) ∈ Rd+1 | ϕ(x) = 0}
= {y ∈ Rd+1 | (∃x ∈ Rd+1 )(y = h(x), ϕ(x) = 0)}
= {y ∈ Rd+1 | (ϕ ◦ h−1 )(y) = 0}.
Consequently, if H is given by
α1 x1 + · · · + αd+1 xd+1 = 0
and if we write α = (α1 , . . . , αd+1 ), then h(H) is the hyperplane given by the equation
αA−1 y = 0.
Similarly,
h(S) = {h(x) ∈ Rd+1 | F (x) = 0}
= {y ∈ Rd+1 | (∃x ∈ Rd+1 )(y = h(x), F (x) = 0)}
= {y ∈ Rd+1 | F ((A−1 y)> ) = 0}
is the hypersurface defined by the polynomial
d+1 d+1
!
X X
G(x1 , . . . , xd+1 ) = F a−1
1 j xj , . . . , a−1
d+1 j xj .
j=1 j=1
Furthermore, using the chain rule, we get

(Gx1 , . . . , Gxd+1 ) = (Fx1 , . . . , Fxd+1 )A−1 ,
which shows that a point, a ∈ S, is non-singular iff its image, h(a) ∈ h(S), is non-singular
on h(S). This also shows that
h(Ta S) = Th(a) h(S),
that is, the projectivity, h, preserves tangent spaces. In summary, we have
Proposition 11.6. Let S = V (F ) ⊆ Pd be a projective hypersurface, where F is a homo-

geneous polynomial of degree m, and let h : Pd → Pd be a projectivity (a bijective projective
map). Then, h(S) is a hypersurface in Pd and a point a ∈ S is nonsingular for S iff h(a) is
nonsingular for h(S). Furthermore,
h(Ta S) = Th(a) h(S),
that is, the projectivity h preserves tangent spaces.
Remark: If h : Pm → Pn is a projective map, say induced by an injective linear map given

by the (n + 1) × (m + 1) matrix A = (ai j ), given any hypersurface S = V (F ) ⊆ Pn , we can
define the pull-back h∗ (S) ⊆ Pm of S, by
h∗ (S) = {x ∈ Pm | F (h(x)) = 0}.
This is indeed a hypersurface because F (x1 , . . . , xn+1 ) is a homogenous polynomial and h∗ (S)
is the zero locus of the homogeneous polynomial
m+1 m+1
!
X X
G(x1 , . . . , xm+1 ) = F a1 j x j , . . . , an+1 j xj .
j=1 j=1
If m = n and h is a projectivity, then we have
h(S) = (h−1 )∗ (S).
11.4 Quadrics (Affine, Projective) and Polar Duality

The case where S = V (Φ) ⊆ Pd is a hypersurface given by a homogeneous polynomial
Φ(x1 , . . . , xd+1 ) of degree 2 will come up a lot and deserves a little more attention. In this
case, if we write x = (x1 , . . . , xd+1 ), then Φ(x) = Φ(x1 , . . . , xd+1 ) is completely determined
by a (d + 1) × (d + 1) symmetric matrix, say F = (fi j ), and we have
d+1
X
>
Φ(x) = x F x = f i j x i xj .
i,j=1
Since F is symmetric, we can write

d+1
X d+1
X d+1
X
Φ(x) = f i j xi xj = fi i x2i +2 f i j xi xj .
i,j=1 i=1 i,j=1
i<j
11.4. QUADRICS (AFFINE, PROJECTIVE) AND POLAR DUALITY 313
Definition 11.19. The polar form ϕ(x, y) of Φ(x), is given by

d+1
X
ϕ(x, y) = x> F y = fi j xi yj ,
i,j=1
where x = (x1 , . . . , xd+1 ) and y = (y1 , . . . , yd+1 ).

Of course,
2ϕ(x, y) = Φ(x + y) − Φ(x) − Φ(y).
We also check immediately that
d+1
>
X ∂Φ(x)
2ϕ(x, y) = 2x F y = yj ,
j=1
∂xj
and so,
∂Φ(x) ∂Φ(x)
,..., = 2x> F.
∂x1 ∂xd+1
Definition 11.20. The hypersurface S = V (Φ) ⊆ Pd is called a (projective) (hyper-)quadric
surface. We say that a quadric surface S = V (Φ) is nondegenerate iff the matrix F defining
Φ is invertible.
For example, the sphere, S d ⊆ Pd+1 , is the nondegenerate quadric given by

> Id+1 0
x x=0
O −1
and the paraboloid, P ⊆ Pd+1 , is the nongenerate quadric given by

 
Id 0 0
x>  O 0 − 21  x = 0.
O − 12 0
If h : Pd → Pd is a projectivity induced by some invertible matrix, A = (ai j ), and if

S = V (Φ) is a quadric defined by the matrix F , we immediately check that h(S) is the
quadric defined by the matrix (A−1 )> F A−1 . Furthermore, as A is invertible, we see that S
is nondegenerate iff h(S) is nondegenerate.
Observe that polar duality w.r.t. the sphere, S d−1 , can be expressed by

∗ d > Id 0 y
X = x ∈ R | (∀y ∈ X) (x , 1) ≤0 ,
O −1 1
where X is any subset of Rd . The above suggests generalizing polar duality with respect to
any nondegenerate quadric.
Let Q = V (Φ(x1 , . . . , xd+1 )) be a nondegenerate quadric given by the homogeneous poly-

nomial Φ with corresponding polar form ϕ and matrix F = (fi j ). Then, we know that
ϕ induces a natural duality between Rd+1 and (Rd+1 )∗ , namely, for every u ∈ Rd+1 , if
ϕu ∈ (Rd+1 )∗ is the linear form given by
ϕu (v) = ϕ(u, v)
for every v ∈ Rd+1 , then the map u 7→ ϕu , from Rd+1 to (Rd+1 )∗ , is a linear isomorphism.
Definition 11.21. Let Q = V (Φ(x1 , . . . , xd+1 )) be a nondegenerate quadric with corre-
sponding polar form ϕ. For any u ∈ Rd+1 , with u 6= 0, the set
u† = {v ∈ Rd+1 | ϕ(u, v) = 0} = {v ∈ Rd+1 | ϕu (v) = 0} = Ker ϕu
is a hyperplane called the polar of u (w.r.t. Q). In terms of the matrix representation of Q,
the polar of u is given by the equation
u> F x = 0,
or
d+1
X ∂Φ(u)
xj = 0.
j=1
∂xj
Going over to Pd , we say that P(u† ) is the polar (hyperplane) of the point a = [u] ∈ Pd and
we write a† for P(u† ).
Note that the equation of the polar hyperplane a† of a point a ∈ Pd is identical to the
equation of the tangent plane to Q at a, except that a is not necessarily on Q. However, if
a ∈ Q, then the polar of a is indeed the tangent hyperplane Ta Q to Q at a.
Proposition 11.7. Let Q = V (Φ(x1 , . . . , xd+1 )) ⊆ Pd be a nondegenerate quadric with
corresponding polar form, ϕ, and matrix, F . Then, every point, a ∈ Q, is nonsingular.
Proof. Since
∂Φ(a) ∂Φ(a)
,..., = 2a> F,
∂x1 ∂xd+1
if a ∈ Q is singular, then a> F = 0 with a 6= 0, contradicting the fact that F is invertible.
The reader should prove the following simple proposition:

Proposition 11.8. Let Q = V (Φ(x1 , . . . , xd+1 )) be a nondegenerate quadric with correspond-
ing polar form, ϕ. Then, the following properties hold: For any two points, a, b ∈ Pd ,
(1) a ∈ b† iff b ∈ a† ;
(2) a ∈ a† iff a ∈ Q;
(3) Q does not contain any hyperplane.
Remark: As in the case of the sphere, if Q is a nondegenerate quadric and a ∈ Pd is

any point such that the polar hyperplane a† intersects Q, then there is a nice geometric
interpretation for a† . Observe that for every b ∈ Q ∩ a† , the polar hyperplane b† is the
tangent hyperplane Tb Q to Q at b and that a ∈ Tb Q. Also, if a ∈ Tb Q for any b ∈ Q, as
b† = Tb Q, then b ∈ a† . Therefore, Q ∩ a† is the set of contact points of all the tangent
hyperplanes to Q passing through a.
Proposition 11.9. Every hyperplane H ⊆ Pd is the polar of a single point a ∈ Pd .
Proof. Indeed, if H is defined by a nonzero linear form f ∈ (Rd+1 )∗ , as Φ is nondegenerate,

there is a unique u ∈ Rd+1 , with u 6= 0, so that f = ϕu , and as ϕu vanishes on H, we see
that H is the polar of the point a = [u]. If H is also the polar of another point b = [v], then
ϕv vanishes on H, which means that
ϕv = λϕu = ϕλu ,
with λ 6= 0 and this implies v = λu, that is, a = [u] = [v] = b, and the pole of H is indeed
unique.

sponding polar form ϕ. The polar dual (w.r.t. Q) X ∗ of a subset X ⊆ Rd+1 is given by
X ∗ = {v ∈ Rd+1 | (∀u ∈ X)(ϕ(u, v) ≤ 0)}.
For every subset X ⊆ Pd , we let

X ∗ = P((v(X))∗ ),
where v(X) is the unique double cone associated with X as in Proposition 11.1.
Observe that X ∗ is always a cone, even if X ⊆ Rd+1 is not. By analogy with the Euclidean
case, for any nonzero vector u ∈ Rd+1 , let
(u† )− = {v ∈ Rd+1 | ϕ(u, v) ≤ 0}.
Now, we have the following version of Proposition 5.5:
Proposition 11.10. Let Q = V (Φ(x1 , . . . , xd+1 )) be a nondegenerate quadric with corre-

sponding polar form ϕ, and matrix F = (fi j ). For any nontrivial polyhedral cone C =
cone(u1 , . . . , up ), where ui ∈ Rd+1 , ui 6= 0, we have
p
(u†i )− .
\
∗
C =
i=1
If U is the (d + 1) × p matrix whose ith column is ui , then we can also write
C ∗ = P (U > F, 0),
where
P (U > F, 0) = {v ∈ Rd+1 | U > F v ≤ 0}.
Consequently, the polar dual of a polyhedral cone w.r.t. a nondegenerate quadric is a poly-
hedral cone.
Proof. The proof is essentially the same as the proof of Proposition 5.5. As
C = cone(u1 , . . . , up ) = {λ1 u1 + · · · + λp up | λi ≥ 0, 1 ≤ i ≤ p},
we have
C ∗ = {v ∈ Rd+1 | (∀u ∈ C)(ϕ(u, v) ≤ 0)}

= {v ∈ Rd+1 | ϕ(λ1 u1 + · · · + λp up , v) ≤ 0, λi ≥ 0, 1 ≤ i ≤ p}
= {v ∈ Rd+1 | λ1 ϕ(u1 , v) + · · · + λp ϕ(up , v) ≤ 0, λi ≥ 0, 1 ≤ i ≤ p}
p
\
= {v ∈ Rd+1 | ϕ(ui , v) ≤ 0}
i=1
p
(u†i )− .
\
=
i=1
By the equivalence theorem for H-polyhedra and V-polyhedra, we conclude that C ∗ is a

polyhedral cone.
Proposition 11.10 allows us to make the following definition:

sponding polar form ϕ. Given any projective polyhedron P = P(C), where C is a polyhedral
cone, the polar dual (w.r.t. Q) P ∗ of P is the projective polyhedron
P ∗ = P(C ∗ ).
We also show that projectivities behave well with respect to polar duality.
Proposition 11.11. Let Q = V (Φ(x1 , . . . , xd+1 )) be a nondegenerate quadric with corre-

sponding polar form ϕ, and matrix F = (fi j ). For every projectivity h : Pd → Pd , if h
h given by the invertible matrix A = (ai j ), for every subset
is induced by the linear map b
d+1
X ⊆ R , we have
h(X ∗ ) = (b
b h(X))∗ ,
where on the left-hand side X ∗ is the polar dual of X w.r.t. Q, and on the right-hand side
h(X))∗ is the polar dual of b
(b h(X) w.r.t. the nondegenerate quadric h(Q) given by the matrix
−1 > −1
(A ) F A . Consequently, if X 6= {0}, then
h((P(X))∗ ) = (h(P(X)))∗
and for every projective polyhedron P , we have
h(P ∗ ) = (h(P ))∗ .
Proof. As
X ∗ = {v ∈ Rd+1 | (∀u ∈ X)(u> F v ≤ 0)},
we have
h(X ∗ ) = {b
b h(v) ∈ Rd+1 | (∀u ∈ X)(u> F v ≤ 0)}
= {y ∈ Rd+1 | (∀u ∈ X)(u> F A−1 y ≤ 0)}
h(X))(x> (A−1 )> F A−1 y ≤ 0)}
= {y ∈ Rd+1 | (∀x ∈ b
h(X))∗ ,
= (b
where (b h(X))∗ is the polar dual of b

h(X) w.r.t. the quadric whose matrix is (A−1 )> F A−1 ,
that is, the polar dual w.r.t. h(Q).
The second part of the proposition follows immediately by setting X = C, where C is
the polyhedral cone defining the projective polyhedron, P = P(C).
We will also need the notion of an affine quadric and polar duality with respect to an
affine quadric. Fortunately, the properties we need in the affine case are easily derived from
the projective case using the “trick” that the affine space Ed can be viewed as the hyperplane
Hd+1 ⊆ Rd+1 of equation, xd+1 = 1, and that its associated vector space Rd can be viewed as
the hyperplane Hd+1 (0) ⊆ Rd+1 of equation xd+1 = 0. A point, a ∈ Ad , corresponds to the
a u
d+1 d
d+1
vector ba = 1 ∈ R , and a vector u ∈ R corresponds to the vector, u b = 0 ∈ R . This
way, the projective space P = P(R ) is the natural projective completion of Ed , which
d d+1
is isomorphic to the affine patch Ud+1 where xd+1 6= 0. The hyperplane xd+1 = 0 is the
“hyperplane at infinity” in Pd .
If we write x = (x1 , . . . , xd ), a polynomial, Φ(x) = Φ(x1 , . . . , xd ), of degree 2 can be
written as
Xd d
X
Φ(x) = ai j x i x j + 2 bi xi + c,
i,j=1 i=1
where A = (ai j ) is a symmetric matrix. If we write b> = (b1 , . . . , bd ), then we have

> A b x > A b
Φ(x) = (x , 1) > =x b.
x
b c 1 b> c
b
Therefore, as in the projective case, Φ is completely determined by a (d + 1) × (d + 1)

symmetric matrix, say F = (fi j ), and we have

> x
Φ(x) = (x , 1)F =xb> F x
b.
1
Definition 11.24. We say that Q ⊆ Rd is a nondegenerate affine quadric iff

d > x
Q = V (Φ) = x ∈ R | (x , 1)F =0 ,
1
where F is symmetric and invertible. Given any point a ∈ Rd , the polar hyperplane a† of a
w.r.t. Q is defined by
† d > x
a = x ∈ R | (a , 1)F =0 .
1
From a previous discussion, the equation of the polar hyperplane a† is
d
X ∂Φ(a)
(xi − ai ) = 0.
i=1
∂xi
Definition 11.25. Given any subset X ⊆ Rd , the polar dual X ∗ of X is defined by

∗ d > y
X = y ∈ R | (∀x ∈ X) (x , 1)F ≤0 .
1
As noted before, polar duality with respect to the affine sphere S d ⊆ Rd+1 corresponds
to the case where
Id 0
F = ,
O −1
and polar duality with respect to the affine paraboloid P ⊆ Rd+1 corresponds to the case
where  
Id−1 0 0
F = O 0 − 21  .
O − 12 0
We will need the following version of Proposition 5.15:
Proposition 11.12. Let Q be a nondegenerate affine quadric given by the (d + 1) × (d + 1)
symmetric matrix F , let {y1 , . . . , yp } be any set of points in Ed , and let {v1 , . . . , vq } be any
set of nonzero vectors in Rd . If Yb is the (d + 1) × p matrix whose ith column is ybi and Vb is
the (d + 1) × q matrix whose j th column is vbj , then
(conv({y1 , . . . , yp }) ∪ cone({v1 , . . . , vq }))∗ = P (Yb > F, 0; Vb > F, 0),

with
x x
P (Yb > F, 0; Vb > F, 0) = >
d
x∈R |Y F
b >
≤ 0, V F
b ≤0 .
1 0
Proof. The proof is immediately adapted from that of Proposition 5.15.
Using Proposition 11.12, we can prove the following Proposition showing that projective
completion and polar duality commute:
Proposition 11.13. Let Q be a nondegenerate affine quadric given by the (d + 1) × (d + 1)
symmetric, invertible matrix F . For every polyhedron P ⊆ Rd , we have
f∗ = (Pe)∗ ,
P
where on the right-hand side, we use polar duality w.r.t. the nondegenerate projective quadric
Q
e defined by F .
Proof. By definition, we have Pe = P(C(P )), (Pe)∗ = P((C(P ))∗ ) and P

f∗ = P(C(P ∗ )).
Therefore, it suffices to prove that
(C(P ))∗ = C(P ∗ ).
Now, P = conv(Y ) + cone(V ), for some finite set of points Y and some finite set of vectors
V , and we know that
C(P ) = cone(Yb ∪ Vb ).
From Proposition 11.10,
(C(P ))∗ = {v ∈ Rd+1 | Yb > F v ≤ 0, Vb > F v ≤ 0}
and by Proposition 11.12,

∗ d> x > x
P = x∈R |Y F
b ≤ 0, V F
b ≤0 .
1 0
But, by definition of C(P ∗ ) (see Section 5.5, especially Proposition 5.20), the hyperplanes
cutting out C(P ∗ ) are obtained by homogenizing the equations of the hyperplanes cutting
out P ∗ and so,

∗ x d+1 b > x > x
C(P ) = ∈R |Y F ≤ 0, V F
b ≤ 0 = (C(P ))∗ ,
xd+1 xd+1 xd+1
as claimed.
Remark: If Q = V (Φ(x1 , . . . , xd+1 )) is a projective or an affine quadric, it is obvious that
V (Φ(x1 , . . . , xd+1 )) = V (λΦ(x1 , . . . , xd+1 ))
for every λ 6= 0. This raises the following question: If
Q = V (Φ1 (x1 , . . . , xd+1 )) = V (Φ2 (x1 , . . . , xd+1 )),

what is the relationship between Φ1 and Φ2 ?

The answer depends crucially on the field over which projective space or affine space is
defined (i.e., whether Q ⊆ RPd or Q ⊆ CPd in the projective case or whether Q ⊆ Rd+1 or
Q ⊆ Cd+1 in the affine case).
For example, over R, the polynomials Φ1 (x1 , x2 , x3 ) = x21 + x22 and Φ2 (x1 , x2 , x3 ) =
2x21+ 3x22 both define the point (0 : 0 : 1) ∈ P2 , since the only real solution of Φ1 and Φ2 are
of the form (0, 0, z). However, if Q has some nonsingular point, the following can be proved
(see Samuel [50], Theorem 46 (Chapter 3)):
Theorem 11.14. Let Q = V (Φ(x1 , . . . , xd+1 ) be a projective or an affine quadric over RPd or
Rd+1 . If Q has a nonsingular point, then for every polynomial Φ0 such that Q = V (Φ0 (x1 , . . .,
xd+1 ), there is some λ 6= 0 (λ ∈ R) so that Φ0 = λΦ.
In particular, Theorem 11.14 shows that the equation of a nondegenerate quadric is

unique up to a scalar.
Actually, more is true. It turns out that if we allow complex solutions, that is, if Q ⊆ CPd
in the projective case or Q ⊆ Cd+1 in the affine case, then Q = V (Φ1 ) = V (Φ2 ) always implies
Φ2 = λΦ1 for some λ ∈ C, with λ 6= 0. In the real case, the above holds (for some λ ∈ R,
with λ 6= 0) unless Q is an affine subspace (resp. a projective subspace) of dimension at
most d − 1 (resp. of dimension at most d − 2). Even in this case, there is a bijective affine
map f (resp. a bijective projective map h) such that Φ2 = Φ1 ◦ f −1 (resp. Φ2 = Φ1 ◦ h−1 ).
A proof of these facts (and more) can be found in Tisseron [62] (Chapter 3).
We now have everything we need for a rigorous presentation of the material of Section
12.7. For a comprehensive treatment of the affine and projective quadrics and related mate-
rial, the reader should consult Berger (Geometry II) [8] or Samuel [50].
Chapter 12
Dirichlet–Voronoi Diagrams and

In this chapter we present the concepts of a Voronoi diagram and of a Delaunay triangu-
lation. These are important tools in computational geometry and Delaunay triangulations
are important in problems where it is necessary to fit 3D data using surface splines. It is
usually useful to compute a good mesh for the projection of this set of data points onto the
xy-plane, and a Delaunay triangulation is a good candidate.
Our presentation of Voronoi diagrams and Delaunay triangulations is far from thor-
ough. We are primarily interested in defining these concepts and stating their most impor-
tant properties. For a comprehensive exposition of Voronoi diagrams, Delaunay triangula-
tions, and more topics in computational geometry, our readers may consult O’Rourke [44],
Preparata and Shamos [47], Boissonnat and Yvinec [12], de Berg, Van Kreveld, Overmars,
and Schwarzkopf [6], or Risler [48]. The survey by Graham and Yao [33] contains a very
gentle and lucid introduction to computational geometry.
In Section 12.7 (which relies on Sections 12.5 and 12.6), we show that the Delaunay
triangulation of a set of points P is the stereographic projection of the convex hull of the set
of points obtained by mapping the points in P onto the sphere using inverse stereographic
projection. We also prove in Section 12.8 that the Voronoi diagram of P is obtained by
taking the polar dual of the above convex hull and projecting it from the north pole (back
onto the hyperplane containing P ). A rigorous proof of this second fact is not trivial because
the central projection from the north pole is only a partial map. To give a rigorous proof,
we have to use projective completions. This requires defining convex polyhedra in projective
space, and we use the results of Chapter 11 (especially, Section 11.2).
12.1 Dirichlet–Voronoi Diagrams

Let E be a Euclidean space of finite dimension, that is, an affine space E whose underlying
→
−
vector space E is equipped with an inner product (and has finite dimension). For concrete-
321
322 CHAPTER 12. DIRICHLET–VORONOI DIAGRAMS
ness, one may safely assume that E = Em , although what follows applies to any Euclidean
space of finite dimension. Given a set P = {p1 , . . . , pn } of n points in E, it is often useful to
find a partition of the space E into regions each containing a single point of P and having
some nice properties. It is also often useful to find triangulations of the convex hull of P
having some nice properties. We shall see that this can be done and that the two problems
are closely related. In order to solve the first problem, we need to introduce bisector lines
and bisector planes.
For simplicity, let us first assume that E is a plane i.e., has dimension 2. Given any two
distinct points a, b ∈ E, the line orthogonal to the line segment (a, b) and passing through
the midpoint of this segment is the locus of all points having equal distance to a and b. It
is called the bisector line of a and b. The bisector line of two points is illustrated in Figure
12.1. 1
L b
Figure 12.1: The bisector line L of a and b
If h = 21 a + 21 b is the midpoint of the line segment (a, b), letting m be an arbitrary point
on the bisector line, the equation of this line can be found by writing that hm is orthogonal
to ab. In any orthogonal frame, letting m = (x, y), a = (a1 , a2 ), b = (b1 , b2 ), the equation of
this line is
(b1 − a1 )(x − (a1 + b1 )/2) + (b2 − a2 )(y − (a2 + b2 )/2) = 0,
(b1 − a1 )x + (b2 − a2 )y = (b21 + b22 )/2 − (a21 + a22 )/2.
The closed half-plane H(a, b) containing a and with boundary the bisector line is the locus
of all points such that
(b1 − a1 )x + (b2 − a2 )y ≤ (b21 + b22 )/2 − (a21 + a22 )/2,
and the closed half-plane H(b, a) containing b and with boundary the bisector line is the
locus of all points such that
(b1 − a1 )x + (b2 − a2 )y ≥ (b21 + b22 )/2 − (a21 + a22 )/2.
12.1. DIRICHLET–VORONOI DIAGRAMS 323
The closed half-plane H(a, b) is the set of all points whose distance to a is less that or equal
to the distance to b, and vice versa for H(b, a). Thus, points in the closed half-plane H(a, b)
are closer to a than they are to b.
We now consider a problem called the post office problem by Graham and Yao [33]. Given
any set P = {p1 , . . . , pn } of n points in the plane (considered as post offices or sites), for
any arbitrary point x, find out which post office is closest to x. Since x can be arbitrary,
it seems desirable to precompute the sets V (pi ) consisting of all points that are closer to pi
than to any other point pj 6= pi . Indeed, if the sets V (pi ) are known, the answer is any post
office pi such that x ∈ V (pi ). Thus, it remains to compute the sets V (pi ). For this, if x is
closer to pi than to any other point pj 6= pi , then x is on the same side as pi with respect to
the bisector line of pi and pj for every j 6= i, and thus
\
V (pi ) = H(pi , pj ).
j6=i
If E has dimension 3, the locus of all points having equal distance to a and b is a plane.
It is called the bisector plane of a and b. The equation of this plane is also found by writing
that hm is orthogonal to ab. The equation of this plane is
(b1 − a1 )(x − (a1 + b1 )/2) + (b2 − a2 )(y − (a2 + b2 )/2) + (b3 − a3 )(z − (a3 + b3 )/2) = 0,
(b1 − a1 )x + (b2 − a2 )y + (b3 − a3 )z = (b21 + b22 + b23 )/2 − (a21 + a22 + a23 )/2.
The closed half-space H(a, b) containing a and with boundary the bisector plane is the locus
of all points such that
(b1 − a1 )x + (b2 − a2 )y + (b3 − a3 )z ≤ (b21 + b22 + b23 )/2 − (a21 + a22 + a23 )/2,
and the closed half-space H(b, a) containing b and with boundary the bisector plane is the
(b1 − a1 )x + (b2 − a2 )y + (b3 − a3 )z ≥ (b21 + b22 + b23 )/2 − (a21 + a22 + a23 )/2.
The closed half-space H(a, b) is the set of all points whose distance to a is less that or equal
to the distance to b, and vice versa for H(b, a). Again, points in the closed half-space H(a, b)
are closer to a than they are to b.
Given any set P = {p1 , . . . , pn } of n points in E (of dimension m = 2, 3), it is often useful
to find for every point pi the region consisting of all points that are closer to pi than to any
other point pj 6= pi , that is, the set
V (pi ) = {x ∈ E | d(x, pi ) ≤ d(x, pj ), for all j 6= i},

where d(x, y) = (xy · xy)1/2 , the Euclidean distance associated with the inner product · on
E. From the definition of the bisector line (or plane), it is immediate that
\
V (pi ) = H(pi , pj ).
j6=i
Families of sets of the form V (pi ) were investigated by Dirichlet [23] (1850) and Voronoi
[66] (1908). Voronoi diagrams also arise in crystallography (Gilbert [31]). Other applications,
including facility location and path planning, are discussed in O’Rourke [44]. For simplicity,
we also denote the set V (pi ) by Vi , and we introduce the following definition.
Definition 12.1. Let E be a Euclidean space of dimension m = 2, 3. Given any set P = {p1 ,
. . ., pn } of n points in E, the Dirichlet–VoronoiTdiagram Vor (P ) of P = {p1 , . . . , pn } is the
family of subsets of E consisting of the sets Vi = j6=i H(pi , pj ) and of all of their intersections.
Dirichlet–Voronoi diagrams are also called Voronoi diagrams, Voronoi tessellations, or

Thiessen polygons. Following common usage, we will use the terminology Voronoi diagram.
As intersections of convex sets (closed half-planes or closed half-spaces), the Voronoi regions
V (pi ) are convex sets. In dimension two, the boundaries of these regions are convex polygons,
and in dimension three, the boundaries are convex polyhedra.
Whether a region V (pi ) is bounded or not depends on the location of pi . If pi belongs
to the boundary of the convex hull of the set P , then V (pi ) is unbounded, and otherwise
bounded. In dimension two, the convex hull is a convex polygon, and in dimension three,
the convex hull is a convex polyhedron. As we will see later, there is an intimate relationship
between convex hulls and Voronoi diagrams.
Generally, if E is a Euclidean space of dimension m, given any two distinct points a, b ∈ E,
the locus of all points having equal distance to a and b is a hyperplane. It is called the bisector
hyperplane of a and b. The equation of this hyperplane is still found by writing that hm is
orthogonal to ab. The equation of this hyperplane is
(b1 − a1 )(x1 − (a1 + b1 )/2) + · · · + (bm − am )(xm − (am + bm )/2) = 0,
(b1 − a1 )x1 + · · · + (bm − am )xm = (b21 + · · · + b2m )/2 − (a21 + · · · + a2m )/2.
The closed half-space H(a, b) containing a and with boundary the bisector hyperplane is the
(b1 − a1 )x1 + · · · + (bm − am )xm ≤ (b21 + · · · + b2m )/2 − (a21 + · · · + a2m )/2,
and the closed half-space H(b, a) containing b and with boundary the bisector hyperplane is
the locus of all points such that
(b1 − a1 )x1 + · · · + (bm − am )xm ≥ (b21 + · · · + b2m )/2 − (a21 + · · · + a2m )/2.
The closed half-space H(a, b) is the set of all points whose distance to a is less than or equal
to the distance to b, and vice versa for H(b, a). Figure 12.2 shows the Voronoi diagram of a
set of twelve points.
In the general case where E has dimension m, the definition of the Voronoi diagram
Vor (P ) of P is the same as Definition 12.1, except that H(pi , pj ) is the closed half-space
containing pi and having the bisector hyperplane of pi and pj as boundary. Also, observe
that the convex hull of P is a convex polytope.
Figure 12.2: A Voronoi diagram
We will now state a proposition listing the main properties of Voronoi diagrams. It turns
out that certain degenerate situations can be avoided if we assume that the points in the set
P are in general position.
Definition 12.2. If P is a set of points in an affine space of dimension m, then we say
that the points of P are in general position if no m + 2 points from P belong to the same
(m − 1)-sphere.
Thus when m = 2, no 4 points in P are cocyclic, and when m = 3, no 5 points in P are
on the same sphere.
Proposition 12.1. Given a set P = {p1 , . . . , pn } of n points in some Euclidean space E
of dimension m (say Em ), if the points in P are in general position and not in a common
hyperplane then the Voronoi diagram of P satisfies the following conditions:
(1) Each region Vi is convex and contains pi in its interior.
(2) Each vertex of Vi belongs to m + 1 regions Vj and to m + 1 edges.
(3) The region Vi is unbounded iff pi belongs to the boundary of the convex hull of P .
(4) If p is a vertex that belongs to the regions V1 , . . . , Vm+1 , then p is the center of the
(m − 1)-sphere S(p) determined by p1 , . . . , pm+1 . Furthermore, no point in P is inside
the sphere S(p) (i.e., in the open ball associated with the sphere S(p)).
(5) If pj is a nearest neighbor of pi , then one of the faces of Vi is contained in the bisector
hyperplane of (pi , pj ).
(6)
n
[ ◦ ◦
Vi = E, and V i ∩ V j = ∅, for all i, j, with i 6= j,
i=1
◦
where V i denotes the interior of Vi .
Proof. We prove only some of the statements, leaving the others as an exercise (or see Risler
[48]).
T
(1) Since Vi = j6=i H(pi , pj ) and each half-space H(pi , pj ) is convex, as an intersection
of convex sets, Vi is convex. Also, since pi belongs to the interior of each H(pi , pj ), the point
pi belongs to the interior of Vi .
(2) Let Fi,j denote Vi ∩ Vj . Any vertex p of the Vononoi diagram of P must belong to r
faces Fi,j . Let us pick the origin of our affine space to be p. Now, given a vector space E
and any two subspaces M and N of E, recall that we have the Grassmann relation
dim(M ) + dim(N ) = dim(M + N ) + dim (M ∩ N ).
Then since p belongs to the intersection of hyperplanes that support the boundaries of the Vi ,
and since a hyperplane has dimension m − 1, by the Grassmann relation, in order to obtain
{p}, a subspace of dimension 0, as the intersection of hyperplanes, we must intersect at least
m hyperplanes, so we must have r ≥ m. We can rename the r + 1 points pi corresponding
the regions Vi inducing the faces containing p by p1 , . . . , pr+1 , so that the r faces containing
p are denoted F1,2 , F2,3 , . . . , Fr,r+1 . Since Fi,j = Vi ∩ Vj , we have
Fi,j = {p | d(p, pi ) = d(p, pj ) ≤ d(p, pk ), for all k 6= i, j},
and since p ∈ F1,2 ∩ F2,3 ∩ · · · ∩ Fr,r+1 , we have
d(p, p1 ) = · · · = d(p, pr+1 ) < d(p, pk ) for all k ∈

/ {1, . . . , r + 1}.
This means that p is the center of a sphere passing through p1 , . . . , pr+1 and containing no
other point in P . By the assumption that points in P are in general position, since there
are r + 1 points pi on a sphere, we must have r + 1 ≤ m + 1, that is, r ≤ m, and thus
r = m. Thus, p belongs to V1 ∩ · · · ∩ Vm+1 , but to no other Vj with j ∈ / {1, . . . , m + 1}.
Furthermore, every edge of the Voronoi diagram containing p is the intersection of m of the
regions V1 , . . . , Vm+1 , and so there are m + 1 of them.
Figure 12.3: Another Voronoi diagram
For simplicity, let us again consider the case where E is a plane. It should be noted that
certain Voronoi regions, although closed, may extend very far. Figure 12.3 shows such an
example.
It is also possible for certain unbounded regions to have parallel edges.
There are a number of methods for computing Voronoi diagrams. A fairly simple (al-
though not very efficient) method is to compute each Voronoi region V (pi ) by intersecting
the half-planes H(pi , pj ) (with pi fixed). One way to do this is to construct for each pi suc-
cessive convex polygons that converge to the boundary of the region V (pi ). At every step
we intersect the current convex polygon with the bisector line of pi and pj . There are at
most two intersection points. We also need a starting polygon, and for this we can pick a
square containing all the points. A naive implementation will run in O(n3 ). However, the
intersection of half-planes can be done in O(n log n), using the fact that the vertices of a
convex polygon can be sorted. Thus, the above method runs in O(n2 log n). Actually, there
are faster methods (see Preparata and Shamos [47] or O’Rourke [44]), and it is possible to
design algorithms running in O(n log n). The most direct method to obtain fast algorithms
is to use the “lifting method” discussed in Section 12.4, whereby the original set of points is
lifted onto a paraboloid, and to use fast algorithms for finding a convex hull.
A very interesting (undirected) graph can be obtained from the Voronoi diagram as
follows: The vertices of this graph are the points pi (each corresponding to a unique region
of Vor (P )), and there is an edge between pi and pj iff the regions Vi and Vj share an edge.
The resulting graph is called a Delaunay triangulation of the convex hull of P , after Delaunay,
who invented this concept in 1934. Such triangulations have remarkable properties.
Figure 12.4 shows the Delaunay triangulation associated with the earlier Voronoi diagram
of a set of twelve points.
Figure 12.4: Delaunay triangulation associated with a Voronoi diagram
One has to be careful to make sure that all the Voronoi vertices have been computed
before computing a Delaunay triangulation, since otherwise, some edges could be missed. In
Figure 12.5 illustrating such a situation, if the lowest Voronoi vertex had not been computed
(not shown on the diagram!), the lowest edge of the Delaunay triangulation would be missing.
The concept of a triangulation can be generalized to any dimension m ≥ 3.
12.2 Triangulations
The concept of a triangulation relies on the notion of pure simplicial complex defined in
Chapter 9. The reader should review Definition 9.2 and Definition 9.3.
Definition 12.3. Given a subset, S ⊆ Em (where m ≥ 1), a triangulation of S is a pure

(finite) simplicial complex, K, of dimension m such that S = |K|, that is, S is equal to the
geometric realization of K.
Given a finite set P of n points in the plane, and given a triangulation of the convex hull
of P having P as its set of vertices, observe that the boundary of P is a convex polygon.
Similarly, given a finite set P of points in 3-space, and given a triangulation of the convex hull
of P having P as its set of vertices, observe that the boundary of P is a convex polyhedron.
It is interesting to know how many triangulations exist for a set of n points (in the plane
12.2. TRIANGULATIONS 329
Figure 12.5: Another Delaunay triangulation associated with a Voronoi diagram
or in 3-space), and it is also interesting to know the number of edges and faces in terms
of the number of vertices in P . These questions can be settled using the Euler–Poincaré
characteristic. We say that a polygon in the plane is a simple polygon iff it is a connected
closed polygon such that no two edges intersect (except at a common vertex).
Proposition 12.2.
(1) For any triangulation of a region of the plane whose boundary is a simple polygon,
letting v be the number of vertices, e the number of edges, and f the number of triangles,
we have the “Euler formula”
v − e + f = 1.
(2) For any region, S, in E3 homeomorphic to a closed ball and for any triangulation of S,
letting v be the number of vertices, e the number of edges, f the number of triangles,
and t the number of tetrahedra, we have the “Euler formula”
v − e + f − t = 1.
(3) Furthermore, for any triangulation of the combinatorial surface, B(S), that is the
boundary of S, letting v 0 be the number of vertices, e0 the number of edges, and f 0 the
number of triangles, we have the “Euler formula”
v 0 − e0 + f 0 = 2.
Proof. All the statements are immediate consequences of Theorem 10.6. For example, part
(1) is obtained by mapping the triangulation onto a sphere using inverse stereographic pro-
jection, say from the North pole. Then, we get a polytope on the sphere with an extra facet
corresponding to the “outside” of the triangulation. We have to deduct this facet from the
Euler characteristic of the polytope and this is why we get 1 instead of 2.
It is now easy to see that in case (1), the number of edges and faces is a linear function
of the number of vertices and boundary edges, and that in case (3), the number of edges
and faces is a linear function of the number of vertices. Indeed, in the case of a planar
triangulation, each face has 3 edges, and if there are eb edges in the boundary and ei edges
not in the boundary, each nonboundary edge is shared by two faces, and thus 3f = eb + 2ei .
Since v − eb − ei + f = 1, we get
v − eb − ei + eb /3 + 2ei /3 = 1,
2eb /3 + ei /3 = v − 1,
and thus ei = 3v − 3 − 2eb . Since f = eb /3 + 2ei /3, we have f = 2v − 2 − eb .

Similarly, since v 0 − e0 + f 0 = 2 and 3f 0 = 2e0 , we easily get e = 3v − 6 and f = 2v − 4.
Thus, given a set P of n points, the number of triangles (and edges) for any triangulation
of the convex hull of P using the n points in P for its vertices is fixed.
Case (2) is trickier, but it can be shown that
v − 3 ≤ t ≤ (v − 1)(v − 2)/2.
Thus, there can be different numbers of tetrahedra for different triangulations of the convex
hull of P .
Remark: The numbers of the form v − e + f and v − e + f − t are called Euler–Poincaré
characteristics. They are topological invariants, in the sense that they are the same for all
triangulations of a given polytope. This is a fundamental fact of algebraic topology.
We shall now investigate triangulations induced by Voronoi diagrams.
12.3 Delaunay Triangulations

Given a set P = {p1 , . . . , pn } of n points in the plane and the Voronoi diagram Vor (P ) for
P , we explained in Section 12.1 how to define an (undirected) graph: The vertices of this
graph are the points pi (each corresponding to a unique region of Vor (P )), and there is an
edge between pi and pj iff the regions Vi and Vj share an edge. The resulting graph turns out
to be a triangulation of the convex hull of P having P as its set of vertices. Such a complex
can be defined in general. For any set P = {p1 , . . . , pn } of n points in Em , we say that a
triangulation of the convex hull of P is associated with P if its set of vertices is the set P .
12.3. DELAUNAY TRIANGULATIONS 331
Definition 12.4. Let P = {p1 , . . . , pn } be a set of n points in Em , and let Vor (P ) be

the Voronoi diagram of P . We define a complex Del (P ) as follows. The complex Del (P )
contains the k-simplex {p1 , . . . , pk+1 } iff V1 ∩ · · · ∩ Vk+1 6= ∅, where 0 ≤ k ≤ m. The complex
Del (P ) is called the Delaunay triangulation of the convex hull of P .
Thus, {pi , pj } is an edge iff Vi ∩ Vj 6= ∅, {pi , pj , ph } is a triangle iff Vi ∩ Vj ∩ Vh 6= ∅,

{pi , pj , ph , pk } is a tetrahedron iff Vi ∩ Vj ∩ Vh ∩ Vk 6= ∅, etc.
For simplicity, we often write Del instead of Del (P ). A Delaunay triangulation for a set
of twelve points is shown in Figure 12.6.
Figure 12.6: A Delaunay triangulation
Actually, it is not obvious that Del (P ) is a triangulation of the convex hull of P , but
this can be shown, as well as the properties listed in the following proposition.
Proposition 12.3. Let P = {p1 , . . . , pn } be a set of n points in Em , and assume that they
are in general position. Then the Delaunay triangulation of the convex hull of P is indeed a
triangulation associated with P , and it satisfies the following properties:
(1) The boundary of Del (P ) is the convex hull of P .
(2) A triangulation T associated with P is the Delaunay triangulation Del (P ) iff every
(m − 1)-sphere S(σ) circumscribed about an m-simplex σ of T contains no other point
from P (i.e., the open ball associated with S(σ) contains no point from P ).
The proof can be found in Risler [48] and O’Rourke [44]. In the case of a planar set P , it
can also be shown that the Delaunay triangulation has the property that it maximizes the
minimum angle of the triangles involved in any triangulation of P . However, this does not
characterize the Delaunay triangulation. Given a connected graph in the plane, it can also
be shown that any minimal spanning tree is contained in the Delaunay triangulation of the
convex hull of the set of vertices of the graph (O’Rourke [44]).
We will now explore briefly the connection between Delaunay triangulations and convex
hulls.
12.4 Delaunay Triangulations and Convex Hulls

In this section we show that there is an intimate relationship between convex hulls and
Delaunay triangulations. We will see that given a set P of points in the Euclidean space
Em of dimension m, we can “lift” these points onto a paraboloid living in the space Em+1 (a
hypersurface), and that the Delaunay triangulation of P is the projection of the downward-
facing faces of the convex hull of the set of lifted points. This remarkable connection was
first discovered by Edelsbrunner and Seidel [24]. For simplicity, we consider the case of a set
P of points in the plane E2 , and we assume that they are in general position.
Consider the paraboloid of revolution of equation z = x2 + y 2 . A point p = (x, y) in the
plane is lifted to the point l(p) = (X, Y, Z) in E3 , where X = x, Y = y, and Z = x2 + y 2 .
The first crucial observation is that a circle in the plane is lifted into a plane curve (an
ellipse). Indeed, if such a circle C is defined by the equation
x2 + y 2 + ax + by + c = 0,
since X = x, Y = y, and Z = x2 + y 2 , by eliminating x2 + y 2 we get
Z = −ax − by − c,
and thus X, Y, Z satisfy the linear equation
aX + bY + Z + c = 0,
which is the equation of a plane. Thus, the intersection of the cylinder of revolution consisting
of the lines parallel to the z-axis and passing through a point of the circle C with the
paraboloid z = x2 + y 2 is a planar curve (an ellipse) as illustrated in Figure 12.7.
We can compute the convex hull of the set of lifted points. Let us focus on the downward-
facing faces of this convex hull.
A downward-facing face is a face such that the z-coordinate of the unit normal to the
plane supporting this face pointing towards the interior of the convex hull of the set of lifted
points is positive.
Let (l(p1 ), l(p2 ), l(p3 )) be a downward-facing face, where the points p1 , p2 , p3 belong to
the set P . We claim that no other point from P is inside the circle C circumscribed about
p1 , p2 , p3 .
12.4. DELAUNAY TRIANGULATIONS AND CONVEX HULLS 333
>
Figure 12.7: The intersection of the paraboloid x2 +y 2 = z with the cylinder x2 +(y−1)2 = 0.
The intersection is an ellipse in the plane z = 2y − 1.
Indeed, a point p inside the circle C would lift to a point l(p) on the paraboloid. Since
no four points are cocyclic, one of the four points p1 , p2 , p3 , p is further from O than the
others; say this point is p3 . Then, the face (l(p1 ), l(p2 ), l(p)) would be below the face
(l(p1 ), l(p2 ), l(p3 )), contradicting the fact that (l(p1 ), l(p2 ), l(p3 )) is one of the downward-
facing faces of the convex hull of P . See Figure 12.8. But then, by Property (2) of Proposition
12.3, the triangle (p1 , p2 , p3 ) would belong to the Delaunay triangulation of P .
l ( p3 )
l ( p3 )
l (p)
l ( p1 ) l ( p2 ) l (p)
l ( p1 ) l ( p2 )
p2
p2
p1 p
p1 p
p3
p3
Figure 12.8: The lift of four points p1 , p2 , p3 , p. Since p is inside the green circle, the blue
triangle (l(p1 ), l(p), l(p2 )) is beneath the green triangle (l(p1 ), l(p2 ), l(p3 )), which implies that
(l(p1 ), l(p2 ), l(p3 )) is not downward facing.
Therefore, we have shown that the projection of the part of the convex hull of the lifted
set l(P ) consisting of the downward-facing faces is the Delaunay triangulation of P . Figure
12.9 shows the lifting of the Delaunay triangulation shown earlier. Another example of the
lifting of a Delaunay triangulation is shown in Figure 12.10.
10
y
7.5
5
2.5
0
10
z
5
0
2.5
5
x 7.5
10
Figure 12.9: A Delaunay triangulation and its lifting to a paraboloid
10
y
7.5
5
2.5
0
10
z
5
0
2.5
5
x 7.5
10
Figure 12.10: Another Delaunay triangulation and its lifting to a paraboloid
The fact that a Delaunay triangulation can be obtained by projecting a lower convex
hull can be used to find efficient algorithms for computing a Delaunay triangulation. It also
holds for higher dimensions.
12.4. DELAUNAY TRIANGULATIONS AND CONVEX HULLS 335
The Voronoi diagram itself can also be obtained from the lifted set l(P ). However, this
time, we need to consider tangent planes to the paraboloid at the lifted points. It is fairly
obvious that the tangent plane at the lifted point (a, b, a2 + b2 ) is
z = 2ax + 2by − (a2 + b2 ).
Given two distinct lifted points (a1 , b1 , a21 + b21 ) and (a2 , b2 , a22 + b22 ), the intersection of the
tangent planes at these points is a line belonging to the plane of equation
(b1 − a1 )x + (b2 − a2 )y = (b21 + b22 )/2 − (a21 + a22 )/2.
Now, if we project this plane onto the xy-plane, we see that the above is precisely the equation
of the bisector line of the two points (a1 , b1 ) and (a2 , b2 ). See Figure 12.11. Therefore, if we
look at the paraboloid from z = +∞ (with the paraboloid transparent), the projection of the
boundary of the polyhedron V(P ) consisting of the intersection of the half spaces containing
the origin cut out by the tangent planes at the lifted points is the Voronoi diagram!
display f, g, h ;
Figure 12.11: The intersection of the tangent plane at (0, 1, 1) with equation z = 2y − 1, and
the tangent plane at (1, 0, 1) with equation z = 2x − 1, has intersection y − x = 0, namely
the bisecting hyperplane between (0, 1, 0) and (1, 0, 0).
It should be noted that the “duality” between the Delaunay triangulation, which is the
projection of the convex hull of the lifted set l(P ) viewed from z = −∞, and the Voronoi
diagram, which is the projection of the boundary of the polyhedron V(P ) cut out by the
tangent planes at the points of the lifted set l(P ) viewed from z = +∞, is reminiscent of
the polar duality with respect to a quadric. This duality will be thoroughly investigated in
Section 12.7.
The reader interested in algorithms for finding Voronoi diagrams and Delaunay triangu-
lations is referred to O’Rourke [44], Preparata and Shamos [47], Boissonnat and Yvinec [12],
de Berg, Van Kreveld, Overmars, and Schwarzkopf [6], and Risler [48].
12.5 Stereographic Projection and the Space of

Generalized Spheres
We saw in Section 12.4 that lifting a set of points P ⊆ E2 to the paraboloid P via the lifting
function l : E2 → P given by l(x, y) = (x, y, x2 + y 2 ) yields a definition of the Delaunay
triangulation Del (P ) of the set of points P that does not require any knowledge of the Voronoi
diagram of P . Namely, Del (P ) is the orthogonal projection of the part of the convex hull of
the lifted set l(P ) consisting of its downward-facing faces. The Voronoi diagram Vor (P ) is
also obtained from the lifted set l(P ); it is the projection of the boundary of the polyhedron
V(P ) cut out by the tangent planes at the points of the lifted set l(P ).
The generalization to any dimension d ≥ 2 is immediate. Recall that the paraboloid P
in Ed+1 is given by the equation
Xd
xd+1 = x2i ,
i=1
and of course, the sphere S d is given by

d+1
X
x2i = 1.
i=1
Then the lifting map l : Ed → P is given by

d
X
l(x1 , . . . , xd ) = (x1 , . . . , xd , x2i ),
i=1
and the orthogonal projection pd+1 : Ed+1 → Ed is given by
pd+1 (x1 , . . . , xd , xd+1 ) = (x1 , . . . , xd ).
As far as we know, Edelsbrunner and Seidel [24] were the first to find the relationship
between Voronoi diagrams and the polar dual of the convex hull of a lifted set of points onto
a paraboloid. This connection is described in Note 3.1 of Section 3 in [24]. The connection
between the Delaunay triangulation and the convex hull of the lifted set of points is described
in Note 3.2 of the same paper. Polar duality is not mentioned and seems to enter the scene
only with Boissonnat and Yvinec [12].
Brown appears to be the first person who observed that Voronoi diagrams and convex
hulls are related via inversion with respect to a sphere [16]. Brown takes a set of points
P , for simplicity assumed to be in the plane, first lifts these points to the unit sphere S 2
using inverse stereographic projection from the north pole τN : E2 → (S 2 − {N }) (which is
equivalent to an inversion of power 2 centered at the north pole), getting τN (P ), and then
takes the convex hull D(P ) = conv(τN (P )) of the lifted set. Now, in order to obtain the
12.5. STEREOGRAPHIC PROJECTION AND THE SPACE OF SPHERES 337
N = (0, 0, 1)
M = (1/2, 1, 1/2)
π N (M) ~ (1,2,0)
Figure 12.12: The geometric construction for πN : (E3 − H3 ) → E2 .
Voronoi diagram of P , apply our inversion (of power 2 centered at the north pole) to each
of the faces of conv(τN (P )), obtaining spheres passing through the north pole, and then
intersect these spheres with the plane containing P , obtaining circles. The centers of some
of these circles are the Voronoi vertices. Finally, a simple criterion can be used to retain the
“nearest Voronoi points” and to connect up these vertices; see Brown [16], page 225.
Note that Brown’s method is not the method that uses the polar dual of the poly-
hedron D(P ) = conv(τN (P )), as we might have expected from the lifting method using
a paraboloid. However, Brown’s method suggests a method for obtaining the Delaunay
triangulation Del (P ) of P by lifting the set P to the sphere S d by applying the inverse
stereographic projection τN : Ed → (S d − {N }) (see Definition 12.5) instead of the lifting
function l, computing the convex hull D(P ) = conv(τN (P )) of the lifted set τN (P ), and then
applying the central projection πN from the north pole N to the hyperplane xd+1 = 0 instead
of the orthogonal projection pd+1 to the facets of the polyhedron D(P ) that do not contain
the north pole, as we will prove in Section 12.7. The central projection πN is the partial
map πN : (Ed+1 − Hd+1 ) → Ed given by
1
πN (x1 , . . . , xd , xd+1 ) = (x1 , . . . , xd );
1 − xd+1
see Definition 12.5. For any point M = (x1 , . . . , xd , xd+1 ) not in the hyperplane Hd+1 of
equation xd+1 = 1, the point πN (M ) is the intersection of the line hN, M i through M and
N with the hyperplane Hd+1 (0) of equation xd+1 = 0. See Figure 12.12.
Thus, instead of using a paraboloid we can use a sphere, and instead of the lifting function
l we can use the the inverse stereographic projection τN . Then, to get back down to Ed , we
use the central projection πN instead of the orthogonal projection pd+1 . As D(P ) is strictly
below the hyperplane xd+1 = 1, there are no problems.
It turns out that there is a “projective transformation” Θ of Ed+1 that maps the sphere
d
S minus the north pole to the paraboloid P, and this maps satisfies the equation
l = Θ ◦ τN .
The map Θ is given by

xi
zi = , 1≤i≤d
1 − xd+1
xd+1 + 1
zd+1 = .
1 − xd+1
Observe that Θ is actually a partial function which is undefined on the hyperplane Hd+1
tangent to S d at the north pole, and that its first d component are identical to those of the
stereographic projection! Then, we immediately find that
2zi
xi = , 1≤i≤d
1 + zd+1
zd+1 − 1
xd+1 = .
1 + zd+1
Consequently, Θ is a bijection between Ed+1 − Hd+1 and Ed+1 − Hd+1 (−1), where Hd+1 (−1)
is the hyperplane of equation xd+1 = −1. As we said earlier, Θ maps the sphere S d minus
the north pole to the paraboloid P, (see Figure 12.13),and
l = Θ ◦ τN .
What this means is that if we think of the inverse stereographic projection τN as a lifting of
points in Ed to the sphere S d , then lifting points from Ed to S d and then mapping S d − {N }
to P by applying Θ is equivalent to lifting points from Ed to the paraboloid P using l.
It would be tempting to define the Voronoi diagram Vor (P ) as the central projection of
the polar dual D(P )∗ of D(P ). However, we have to be careful because Θ does not map
all convex polyhedra to convex polyhedra. In particular, Θ is not well-defined on any face
of D(P )∗ intersecting the hyperplane Hd+1 (of equation xd+1 = 1). Fortunately, we can
circumvent these difficulties by using the concept of a projective polyhedron introduced in
Chapter 11 and defining a projective version θ of Θ which is a total function. We can also
define projective versions of σN , τN , l, and πN , to prove that the Voronoi diagram of P is
indeed obtained from a suitable projection of the polar dual of D(P ) (actually, a projective
version of D(P )).
In summary, Voronoi diagrams, Delaunay Triangulations, and their properties, can also
be nicely explained using inverse stereographic projection and the central projection from
N , but a rigorous justification of why this “works” is not as simple as it might appear.
z=3
N = (0,0,1)
z=1
Θ (p) = (1,0,1)
z = 1/2
z = 1/3
z=0
Θ (S) = (0,0,0)
p = (1,0,0) Θ
z = -1/2
S = (0,0,-1)
Figure 12.13: The projective transformation Θ which maps S 2 − {N } to z = x2 + y 2 .
The advantage of stereographic projection over the lifting onto a paraboloid is that the
(d-)sphere is compact. Since the stereographic projection and its inverse map (d − 1)-spheres
to (d − 1)-spheres (or hyperplanes), all the crucial properties of Delaunay triangulations
are preserved. The purpose of this section is to establish the properties of stereographic
projection (and its inverse) that will be needed in Section 12.7.
Recall that the d-sphere S d ⊆ Ed+1 is given by
S d = {(x1 , . . . , xd+1 ) ∈ Ed+1 | x21 + · · · + x2d + x2d+1 = 1}.
It will be convenient to write a point (x1 , . . . , xd+1 ) ∈ Ed+1 as z = (x, xd+1 ), with
x = (x1 , . . . , xd ). We denote N = (0, . . . , 0, 1) (with d zeros) as (0, 1) and call it the north
pole, and S = (0, . . . , 0, −1) (with d zeros) as (0, −1) and call it the south pole. We also
1 1 1
write kzk = (x21 + · · · + x2d+1 ) 2 = (kxk2 + x2d+1 ) 2 (with kxk = (x21 + · · · + x2d ) 2 ). With these
notations,
S d = {(x, xd+1 ) ∈ Ed+1 | kxk2 + x2d+1 = 1}.
The stereographic projection from the north pole σN : (S d −{N }) → Ed is the restriction to
S d of the central projection πN : (Ed+1 − Hd+1 ) → Ed from N onto the hyperplane Hd+1 (0) ∼ =
d
E of equation xd+1 = 0; that is, M 7→ πN (M ) where πN (M ) is the intersection of the line
hN, M i through N and M with Hd+1 (0). Since the line through N and M = (x, xd+1 ) is
given parametrically by
hN, M i = {(1 − λ)(0, 1) + λ(x, xd+1 ) | λ ∈ R},
the intersection πN (M ) of this line with the hyperplane xd+1 = 0 corresponds to the value
of λ such that
(1 − λ) + λxd+1 = 0,
N = (0,0,1)
M = (x , y, z)
σ N (M) = (x/(1-z), y/(1-z))
Figure 12.14: The stereographic projection σN : (S 2 − {N }) → E2 .
that is,
1
λ= .
1 − xd+1
Therefore, the coordinates of πN (M ), with M = (x, xd+1 ), are given by
x
πN (x, xd+1 ) = .
1 − xd+1
See Figure 12.14. The central projection πN is undefined on the hyperplane Hd+1 of equation
xd+1 = 1, and the stereographic projection σN from the north pole, which is the restriction
of πN to the sphere S d , is undefined at the north pole.
Let us find the inverse τN = σN −1
(P ) of any P ∈ Hd+1 (0) ∼
= Ed . This time, τN (P ) is the
intersection of the line hN, P i through P ∈ Hd+1 (0) and N with the sphere S d . Since the
line through N and P = (x, 0) is given parametrically by
hN, P i = {(1 − λ)(0, 1) + λ(x, 0) | λ ∈ R},
the intersection τN (P ) of this line with the sphere S d corresponds to the nonzero value of λ
such that
λ2 kxk2 + (1 − λ)2 = 1,
that is
λ(λ(kxk2 + 1) − 2) = 0.
Thus, we get
2
λ= 2 ,
kxk + 1
N = (0,0,1)
τ N (P) = (2x/ ( x 2 + y 2 + 1 ) , 2y/ ( x 2 + y 2 + 1 ), 1 - 2/( x 2 + y 2 + 1 ) )
P = (x,y)
Figure 12.15: The inverse stereographic projection τN : E2 → (S 2 − {N }).
from which we get

2x 2
τN (x) = 2 ,1 − 2
kxk + 1 kxk + 1
!
2
2x kxk − 1
= , .
kxk + 1 kxk2 + 1
2
See Figure 12.15.

We leave it as an exercise to the reader to verify that τN ◦ σN = id and σN ◦ τN = id.
We can also define the stereographic projection from the south pole σS : (S d − {S}) → Ed ,
and its inverse τS . Again, the computations are left as a simple exercise to the reader. The
above computations are summarized in the following definition:
Definition 12.5. The central projection πN : (Ed+1 − Hd+1 ) → Ed from N onto the hyper-
plane Hd+1 (0) ∼
= Ed of equation xd+1 = 0 is given by
x
πN (x, xd+1 ) = , (xd+1 6= 1).
1 − xd+1
The stereographic projection from the north pole σN : (S d − {N }) → Ed is the restriction of

πN to the sphere S d . The inverse of σN , denoted τN : Ed → (S d − {N }) and called inverse
stereographic projection from the north pole, is given by
!
2x kxk2 − 1
τN (x) = , .
kxk2 + 1 kxk2 + 1
Remark: An inversion of center C and power ρ > 0 is a geometric transformation

f : (Ed+1 − {C}) → Ed+1 defined so that for any M 6= C, the points C, M , and f (M ) are
collinear, and
kCMkkCf (M)k = ρ.
Equivalently, f (M ) is given by
ρ
f (M ) = C + CM.
kCMk2
Clearly, f ◦ f = id on Ed+1 − {C}, so f is invertible and the reader will check that if we
pick the center of inversion to be the north pole and if we set ρ = 2, then the coordinates of
f (M ) are given by
2xi
yi = , 1≤i≤d
+ · · · + + x2d+1 − 2xd+1 + 1
x21 x2d
x2 + · · · + x2d + x2d+1 − 1
yd+1 = 2 1 ,
x1 + · · · + x2d + x2d+1 − 2xd+1 + 1
where (x1 , . . . , xd+1 ) are the coordinates of M . In particular, if we restrict our inversion to
the unit sphere S d , as x21 + · · · + x2d + x2d+1 = 1, we get
xi
yi = , 1≤i≤d
1 − xd+1
yd+1 = 0,
which means that our inversion restricted to S d is simply the stereographic projection σN
(and the inverse of our inversion restricted to the hyperplane xd+1 = 0 is the inverse stereo-
graphic projection τN ).
We will now show that the image of any (d − 1)-sphere S on S d not passing through the
north pole, that is, the intersection S = S d ∩ H of S d with any hyperplane H not passing
through N , is a (d − 1)-sphere. Here, we are assuming that S has positive radius, that is, H
is not tangent to S d .
Assume that H is given by
a1 x1 + · · · + ad xd + ad+1 xd+1 + b = 0.
Since N ∈/ H, we must have ad+1 + b 6= 0. For any (x, xd+1 ) ∈ S d , write σN (x, xd+1 ) = X.
Since
x
X= ,
1 − xd+1
we get x = X(1 − xd+1 ), and using the fact that (x, xd+1 ) also belongs to H we will express
xd+1 in terms of X and then find an equation for X which will show that X belongs to a
(d − 1)-sphere. Indeed, (x, xd+1 ) ∈ H implies that
d
X
ai Xi (1 − xd+1 ) + ad+1 xd+1 + b = 0,
i=1
that is,
d
X d
X
ai Xi + (ad+1 − aj Xj )xd+1 + b = 0.
i=1 j=1
Pd
If j=1 aj Xj = ad+1 , then ad+1 + b = 0, which is impossible. Therefore, we get
Pd
−b − ai X i
xd+1 = Pi=1
d
,
ad+1 − i=1 ai Xi
and so
ad+1 + b
1 − xd+1 = .
ad+1 − di=1 ai Xi
P
Plugging x = X(1 − xd+1 ) in the equation kxk2 + xdd+1 = 1 of S d , we get
(1 − xd+1 )2 kXk2 + x2d+1 = 1,
and replacing xd+1 and 1 − xd+1 by their expression in terms of X, we get

d
X d
X
2 2 2
(ad+1 + b) kXk + (−b − ai Xi ) = (ad+1 − ai Xi )2 ,
i=1 i=1
that is,
d
X d
X
2 2 2
(ad+1 + b) kXk = (ad+1 − ai Xi ) − (b + ai Xi )2
i=1 i=1
d
X
= (ad+1 + b)(ad+1 − b − 2 ai Xi ),
i=1
which yields
Xd
2 2
(ad+1 + b) kXk + 2(ad+1 + b)( ai Xi ) = (ad+1 + b)(ad+1 − b),
i=1
that is,
d
2
X ai ad+1 − b
kXk + 2 Xi − = 0,
i=1
ad+1 + b ad+1 + b
which is indeed the equation of a (d − 1)-sphere in Ed . By “completing the square,” the
above equation can be written as
d 2 d
X ai X a2i ad+1 − b
Xi + − 2
− = 0,
i=1
ad+1 + b i=1
(ad+1 + b) ad+1 + b
which yields
d 2 Pd
X ai i=1 a2i + (ad+1 − b)(ad+1 + b)
Xi + = ,
i=1
ad+1 + b (ad+1 + b)2
that is,
d 2 Pd+1
X ai i=1a2i − b2
Xi + = . (∗)
i=1
ad+1 + b (ad+1 + b)2
However, the distance from the origin to the hyperplane H of equation
a1 x1 + · · · + ad xd + ad+1 xd+1 + b = 0
is
|b|
δ= ,
Pd+1 2 1/2
i=1 ai
and since we are assuming that H intersects the unit sphere S d in a sphere of positive radius
we must have δ < 1, so
d+1
X
2
b < a2i ,
i=1
and (∗) is indeed the equation of a real sphere (its radius is positive). Therefore, when
N ∈/ H, the image of S = S d ∩ H by σN is a (d − 1)-sphere in Hd+1 (0) = Ed . See Figure
12.16.
If the hyperplane H contains the north pole, then ad+1 + b = 0, in which case, for every
(x, xd+1 ) ∈ S d ∩ H, we have
d
X
ai xi + ad+1 xd+1 − ad+1 = 0,
i=1
that is,
d
X
ai xi − ad+1 (1 − xd+1 ) = 0,
i=1
and except for the north pole, we have

d
X xi
ai − ad+1 = 0,
i=1
1 − xd+1
which shows that

d
X
ai Xi − ad+1 = 0,
i=1
Figure 12.16: Two views of the plane −x−y+z = 0 intersecting S 2 . The bottom figure shows
the stereographic projection of the intersection, namely the circle (x − 1)2 + (y − 1)2 = 1.
the intersection of the hyperplanes H and Hd+1 (0). Therefore, the image of S d ∩ H by σN
is the hyperplane in Ed which is the intersection of H with Hd+1 (0). See Figure 12.17.
We will also prove that τN maps (d − 1)-spheres in Hd+1 (0) to (d − 1)-spheres on S d
not passing through the north pole. Assume that X ∈ Ed belongs to the (d − 1)-sphere of
equation
Xd d
X
Xi2 + aj Xj + b = 0.
i=1 j=1
For any (X, 0) ∈ Hd+1 (0), we know that (x, xd+1 ) = τN (X) is given by
!
2X kXk2 − 1
(x, xd+1 ) = , .
kXk2 + 1 kXk2 + 1
Using the equation of the (d − 1)-sphere, we get
2X
x=
−b + 1 − dj=1 aj Xj
P
and Pd
−b − 1 − j=1 aj X j
xd+1 = Pd .
−b + 1 − j=1 aj X j
Figure 12.17: The plane x + y + z − 1 = 0 intersecting S 2 , along with the stereographic

projection of the intersection, namely the line x + y = 1.
Then, we get
d Pd
X 2 aj Xj
j=1
ai xi = Pd ,
i=1 −b + 1 − j=1 aj Xj
which yields
d
X Xd Xd d
X
(−b + 1)( ai x i ) − ( ai xi )( aj X j ) = 2 aj X j .
i=1 i=1 j=1 j=1
From the above, we get

d
(−b + 1)( di=1 ai xi )
X P
ai X i = Pd .
i=1 i=1 a i x i + 2
Plugging this expression in the formula for xd+1 above, we get
−b − 1 − di=1 ai xi
P
xd+1 = ,
−b + 1
which yields
d
X
ai xi + (−b + 1)xd+1 + (b + 1) = 0,
i=1
the equation of a hyperplane H not passing through the north pole. Therefore, the image
of a (d − 1)-sphere in Hd+1 (0) is indeed the intersection H ∩ S d of S d with a hyperplane not
passing through N , that is, a (d − 1)-sphere on S d .
Given any hyperplane H 0 in Hd+1 (0) = Ed , say of equation
d
X
ai Xi + b = 0,
i=1
the image of H 0 under τN is a (d−1)-sphere on S d , the intersection of S d with the hyperplane

H passing through N and determined as follows: For any (X, 0) ∈ Hd+1 (0), if τN (X) =
(x, xd+1 ), then
x
X= ,
1 − xd+1
and so (x, xd+1 ) satisfies the equation
d
X
ai xi + b(1 − xd+1 ) = 0,
i=1
that is,
d
X
ai xi − bxd+1 + b = 0,
i=1
which is indeed the equation of a hyperplane H passing through N . We summarize all this
in the following proposition:
Proposition 12.4. The stereographic projection σN : (S d − {N }) → Ed induces a bijection
between the set of (d − 1)-spheres on S d and the union of the set of (d − 1)-spheres in Ed
with the set of hyperplanes in Ed ; every (d − 1)-sphere on S d not passing through the north
pole is mapped to a (d − 1)-sphere in Ed , and every (d − 1)-sphere on S d passing through
the north pole is mapped to a hyperplane in Ed . In fact, σN maps the (d − 1)-sphere on S d
determined by the hyperplane
a1 x1 + · · · + ad xd + ad+1 xd+1 + b = 0
not passing through the north pole (ad+1 + b 6= 0) to the (d − 1)-sphere
d 2 Pd+1 2
X ai a − b2
Xi + = i=1 i 2 ,
i=1
ad+1 + b (ad+1 + b)
and the (d − 1)-sphere on S d determined by the hyperplane

d
X
ai xi + ad+1 xd+1 − ad+1 = 0
i=1
through the north pole to the hyperplane

d
X
ai Xi − ad+1 = 0;
i=1
−1
the map τN = σN maps the (d − 1)-sphere
d
X d
X
Xi2 + aj X j + b = 0
i=1 j=1
to the (d − 1)-sphere on S d determined by the hyperplane

d
X
ai xi + (−b + 1)xd+1 + (b + 1) = 0
i=1
not passing through the north pole, and the hyperplane

d
X
ai X i + b = 0
i=1
to the (d − 1)-sphere on S d determined by the hyperplane

d
X
ai xi − bxd+1 + b = 0
i=1
through the north pole.
Proposition 12.4 raises a natural question: What do the hyperplanes H in Ed+1 that do
not intersect S d correspond to, if they correspond to anything at all?
The first thing to observe is that the geometric definition of the stereographic projection
and its inverse makes it clear that the hyperplanes corresponding to (d − 1)-spheres in Ed
(by τN ) do intersect S d . Now, when we write the equation of a (d − 1)-sphere S, say
d
X d
X
Xi2 + ai Xi + b = 0,
i=1 i=1
we are implicitly assuming a condition on the ai ’s and b that ensures that S is not the empty
sphere, that is, that its radius R is positive (or zero). By “completing the square,” the above
equation can be rewritten as
d d
X ai 2 1 X 2
Xi + = a − b,
i=1
2 4 i=1 i
and so the radius R of our sphere is given by

d
2 1X 2
R = a −b
4 i=1 i
whereas its center is the point c = − 12 (a1 , . . . , ad ). Thus, our sphere is a “real” sphere of
positive radius iff
X d
a2i > 4b,
i=1
Pd
or a single point, c = − 21 (a1 , . . . , ad ), iff i=1 a2i = 4b.
What happens when
d
X
a2i < 4b?
i=1
In this case, if we allow “complex points,” that is, if we consider solutions of our equation
d
X d
X
Xi2 + ai X i + b = 0
i=1 i=1
q
d
4b − di=1 a2i . The
i
P
over C , then we get a “complex” sphere of (pure) imaginary radius 2
funny thing is that our computations carry over unchanged and the image of the complex
sphere S is still the intersection of the complex sphere S d with the hyperplane H given
d
X
ai xi + (−b + 1)xd+1 + (b + 1) = 0.
i=1
However, this time, even though H does not have any “real” intersection points with S d , we
can show that it does intersect the “complex sphere,”
S d = {(z1 , . . . , zd+1 ) ∈ Cd+1 | z12 + · · · + zd+1

2
= 1}
in a nonempty set of points in Cd+1 .

It follows from all this that σN and τN establish a bijection between the set of all hyper-
planes in Ed+1 minus the hyperplane Hd+1 (of equation xd+1 = 1) tangent to S d at the north
pole, with the union of four sets:
(1) The set of all (real) (d − 1)-spheres of positive radius; see Figure 12.16.
(2) The set of all (complex) (d − 1)-spheres of imaginary radius;
(3) The set of all hyperplanes in Ed ; see Figure 12.17.
(4) The set of all points of Ed (viewed as spheres of radius 0); see Figure 12.18.
Moreover, Set (1) corresponds to the hyperplanes that intersect the interior of S d and do not
pass through the north pole; Set (2) corresponds to the hyperplanes that do not intersect S d ;
Set (3) corresponds to the hyperplanes that pass through the north pole minus the tangent
hyperplane at the north pole; and Set (4) corresponds to the hyperplanes that are tangent
to S d , minus the tangent hyperplane at the north pole.
It is convenient to add the “point at infinity” ∞ to Ed , because then the above bijection
can be extended to map the tangent hyperplane at the north pole to ∞. The union of these
four sets (with ∞ added) is called the set of generalized spheres, sometimes denoted S(Ed ).
N = (0,0,1)
√ p √
Figure 12.18: The plane −2y − 2z − 2 2 = 0 tangent to the point (0, −1/ √ (2), −1/ 2),
along with its corresponding stereographic projection, the point (0, −1/( 2 + 1), 0).
This is a fairly complicated space. For one thing, topologically S(Ed ) is homeomorphic to
the projective space Pd+1 with one point removed (the point corresponding to the “hyperplane
at infinity”), and this is not a simple space. We can get a slightly more concrete “‘picture”
of S(Ed ) by looking at the polars of the hyperplanes w.r.t. S d . Then, the “real” spheres
correspond to the points strictly outside S d which do not belong to the tangent hyperplane
at the north pole, (i.e. Figure 12.19); the complex spheres correspond to the points in the
interior of S d ; the points of Ed ∪ {∞} correspond to the points on S d , (i.e. Figure 12.18);
the hyperplanes in Ed correspond to the points in the tangent hyperplane at the north
pole expect for the north pole, (i.e. Figure 12.20). Unfortunately, the poles of hyperplanes
through the origin are undefined. This can be fixed by embedding Ed+1 in its projective
completion Pd+1 , but we will not go into this.
There are other ways of dealing rigorously with the set of generalized spheres. One
method described by Boissonnat [12] is to use the embedding where the sphere S of equation
d
X d
X
Xi2 −2 ai X i + b = 0
i=1 i=1
is mapped to the point

ϕ(S) = (a1 , . . . , ad , b) ∈ Ed+1 .
This gives us another way of dealing with the sets of type (1), (2), and (4) described earlier.
Now, by a previous computation we know that
d
X
b= a2i − R2 ,
i=1
where c = (a1 , . . . , ad ) is the center of S and R is its radius.

(2,2,2)
Figure 12.19: The plane 2x + 2y + 2z = 1 with its dual (2, 2, 2). Also shown is the stereo-
graphic projection of the intersection, namely the circle (x + 2)2 + (y + 2)2 = 11.
The quantity di=1 a2i − R2 is known as the power of the origin w.r.t. S. In general,
P
the power of a point X ∈ Ed is defined as ρ(X) = kcXk2 − R2 , which, after a moment of
thought, is just
Xd d
X
2
ρ(X) = Xi − 2 ai Xi + b.
i=1 i=1
Now, since points correspond to spheres of radius 0, we see that the image of the point
X = (X1 , . . . , Xd ) is
d
X
l(X) = (X1 , . . . , Xd , Xi2 ).
i=1
Thus, in this model, points of E are lifted to the paraboloid P ⊆ Ed+1 of equation
d
d
X
xd+1 = x2i .
i=1
Actually, this method does not deal with hyperplanes but it is possible to do so. The
trick is to consider equations of a slightly more general form that capture both spheres and
hyperplanes, namely, equations of the form
d
X d
X
c Xi2 + ai Xi + b = 0.
i=1 i=1
Indeed, when c = 0, we do get a hyperplane! Now, to carry out this method we really
need to consider equations up to a nonzero scalars, that is, we consider the projective space
(1,1,1)
(0,0,1)
Figure 12.20: The plane x + y + z = 1 with its dual (1, 1, 1). Also shown is the stereographic
projection of the intersection, namely the line x + y = 1.
b d )) associated with the vector space S(E

P(S(E b d ) consisting of the above equations. Then, it
turns out that the quantity
d
1 X 2
%(a, b, c) = ( a − 4bc)
4 i=1 i
b d ) whose corresponding bilinear form
(with a = (a1 , . . . , ad )) defines a quadratic form on S(E
d
0 0 0 1 X 0
ρ((a, b, c), (a , b , c )) = ( ai ai − 2bc0 − 2b0 c)
4 i=1
has a natural interpretation (with a = (a1 , . . . , ad ) and a0 = (a01 , . . . , a0d )). Indeed, orthogo-
nality with respect to ρ (that is, when ρ((a, b, c), (a0 , b0 , c0 )) = 0) says that the corresponding
spheres defined by (a, b, c) and (a0 , b0 , c0 ) are orthogonal, that the corresponding hyperplanes
defined by (a, b, 0) and (a0 , b0 , 0) are orthogonal, etc. The reader who wants to read more
about this approach should consult Berger (Volume II) [8].
12.6 Relating Lifting to a Paraboloid and Lifting to a

Sphere
We explained in Section 12.5 that there is a simple relationship between the lifting onto a
paraboloid and the lifting onto S d using the inverse stereographic projection map because
12.6. RELATING LIFTING TO A PARABOLOID AND LIFTING TO A SPHERE 353
the sphere and the paraboloid are projectively equivalent, as we showed for S 2 in Section
11.1.
We defined the map Θ given by
xi
zi = , 1≤i≤d
1 − xd+1
xd+1 + 1
zd+1 = ,
1 − xd+1
and showed that Θ is a bijection between Ed+1 −Hd+1 and Ed+1 −Hd+1 (−1), where Hd+1 (−1)
is the hyperplane of equation xd+1 = −1. We will show a little later that Θ maps the sphere
S d minus the north pole to the paraboloid P, and satisfies the equation
l = Θ ◦ τN .
The fact that Θ is undefined on the hyperplane Hd+1 is not a problem as far as mapping
the sphere to the paraboloid because the north pole is the only point that does not have an
image. However, later on when we consider the Voronoi polyhedron V(P ) of a lifted set of
points P , we will have more serious problems because in general, such a polyhedron intersects
both hyperplanes Hd+1 and Hd+1 (−1). This means that Θ will not be well-defined on the
whole of V(P ) nor will it be surjective on its image. To remedy this difficulty, we work with
projective completions. Basically, this amounts to chasing denominators and homogenizing
equations, but we also have to be careful in dealing with convexity, and this is where the
projective polyhedra (studied in Section 11.2) will come handy.
fd ⊆ Pd+1 given by the
So, let us consider the projective completion of the sphere S
equation
d+1
X
x2i = x2d+2 ,
i=1
e ⊆ Pd+1 given by the equation

and the projective completion of the paraboloid P
d
X
xd+1 xd+2 = x2i .
i=1
Definition 12.6. Let θ : Pd+1 → Pd+1 be the projectivity induced by the linear map
θb: Rd+2 → Rd+2 given by
zi = xi , 1≤i≤d
zd+1 = xd+1 + xd+2
zd+2 = xd+2 − xd+1 ,
whose inverse is given by

xi = zi , 1≤i≤d
zd+1 − zd+2
xd+1 =
2
zd+1 + zd+2
xd+2 = .
2
The map θ is a projective version of Θ, but is better behaved because it is a total function.
If we plug these formulae in the equation of Sfd , we get
d
X
4( zi2 ) + (zd+1 − zd+2 )2 = (zd+1 + zd+2 )2 ,
i=1
which simplifies to
d
X
zd+1 zd+2 = zi2 .
i=1
Therefore, θ(S fd ) = P,
e that is, θ maps the projective completion of the sphere to the
projective completion of the paraboloid. Observe that the projective north pole N e =
(0 : · · · : 0 : 1 : 1) is mapped to the point at infinity (0 : · · · : 0 : 1 : 0).
Recall from Definition 11.5 that for any i, with 1 ≤ i ≤ d + 1, the set
Ui = {(x1 : · · · : xd+1 ) ∈ Pd | xi 6= 0}
is a subset of Pd called an affine patch of Pd . We have a bijection ϕi : Ui → Rd between Ui
and Rd given by

x1 xi−1 xi+1 xd+1
ϕi : (x1 : · · · : xd+1 ) 7→ ,..., , ,..., ,
xi xi xi xi
with inverse ψi : Rd → Ui ⊆ Pd given by
ψi : (x1 , · · · , xd ) 7→ (x1 : · · · xi−1 : 1 : xi : · · · : xd ).
The map Θ is the restriction of θ to the affine patch Ud+1 , and as such, it can be fruitfully
described as the composition of θb with a suitable projection onto Ed+1 . For this, as we have
done before, we identify Ed+1 with the hyperplane Hd+2 ⊆ Ed+2 of equation xd+2 = 1 (using
the injection, id+2 : Ed+1 → Ed+2 , where ij : Ed+1 → Ed+2 is the injection given by
(x1 , . . . , xd+1 ) 7→ (x1 , . . . , xj−1 , 1, xj+1 , . . . , xd+1 )
for any (x1 , . . . , xd+1 ) ∈ Ed+1 ). For each i, with 1 ≤ i ≤ d + 2, let πi : (Ed+2 − Hi (0)) → Ed+1
be the projection of center 0 ∈ Ed+2 onto the hyperplane Hi ⊆ Ed+2 of equation xi = 1
(Hi ∼= Ed+1 and Hi (0) ⊆ Ed+2 is the hyperplane of equation xi = 0) given by

x1 xi−1 xi+1 xd+2
πi (x1 , . . . , xd+2 ) = ,..., , ,..., (xi 6= 0).
xi xi xi xi
Geometrically, for any x ∈

/ Hi (0), the image πi (x) of x is the intersection of the line through
the origin and x with the hyperplane Hi ⊆ Ed+2 of equation xi = 1. This is illustrated in
Figure 12.21. Observe that the map πi : (Ed+2 − Hd+2 (0)) → Ed+1 is an “affine” version of
the bijection ϕi : Ui → Rd+1 of Section 11.1. Then, we have
Θ = πd+2 ◦ θb ◦ id+2 .
If we identify Hd+2 and Ed+1 , we may write with a slight abuse of notation Θ = πd+2 ◦ θ.
b
z=1
π3 (p) ~ (1,1,1)
p = (1/2, 1/2, 1/2 )
Figure 12.21: The geometric realization of image of π3 (p), where π3 : (E3 − H3 (0)) → E2 .
We will need some properties of the projection πd+2 and of Θ, and for this, let
Hd+ = {(x1 , . . . , xd ) ∈ Ed | xd > 0} and Hd− = {(x1 , . . . , xd ) ∈ Ed | xd < 0}.
Proposition 12.5. The maps πd+2 , πN , and Θ have the following properties:
(1) For every hyperplane H through the origin, πd+2 (H) is a hyperplane in Hd+2 . See
Figure 12.22.
(2) Given any set of points {a1 , . . . , an } ⊆ Ed+2 , if {a1 , . . . , an } is contained in the open
half-space above the hyperplane xd+2 = 0 or {a1 , . . . , an } is contained in the open half-
space below the hyperplane xd+2 = 0, then the image by πd+2 of the convex hull of the
ai ’s is the convex hull of the images of these points, that is,
πd+2 (conv({a1 , . . . , an })) = conv({πd+2 (a1 ), . . . , πd+2 (an )}).
See Figure 12.23.

half-space above the hyperplane xd+2 = 1 or {a1 , . . . , an } is contained in the open half-
space below the hyperplane xd+2 = 1, then the image by πN of the convex hull of the
ai ’s is the convex hull of the images of these points, that is,
πN (conv({a1 , . . . , an })) = conv({πN (a1 ), . . . , πN (an )}).
half-space above the hyperplane Hd+1 or {a1 , . . . , an } is contained in the open half-space
below Hd+1 , then
Θ(conv({a1 , . . . , an })) = conv({Θ(a1 ), . . . , Θ(an )}).
(5) For any set S ⊆ Ed+1 , if conv(S) does not intersect Hd+1 , then
Θ(conv(S)) = conv(Θ(S)).
Figure 12.22: The intersection of the teal plane x + y + z = 0 with the magenta plane z = 1
results in the pink line x + y = −1. This line is the also the projection of the teal plane via
π3 as shown by the lime green rays through the origin.
Proof. (1) The image, πd+2 (H), of a hyperplane H through the origin is the intersection of
H with Hd+2 , which is a hyperplane in Hd+2 .
(2) This seems fairly clear geometrically but the result fails for arbitrary sets of points,
so to be on the safe side, we give an algebraic proof. We will prove the following two facts
by induction on n ≥ 1:
π3(a3)
π3 (a2)
a3
a
2
z=1
a π3 (a1)
1
Figure 12.23: The convex hull of {a1 , a2 , a3 }, namely the dusty rose triangle above z = 0
and below z = 1, under π3 , is projected to the lavender triangle in the plane z = 1.
(1) For all λ1 , . . . , λn ∈ R with λ1 + · · · + λn = 1 and λi ≥ 0, for all a1 , . . . , an ∈ Hd+2 +

(resp. ∈ Hd+2
− ), there exist some µ 1 , . . . , µ n ∈ R with µ 1 + · · · + µ n = 1 and µ i ≥ 0, so
that
πd+2 (λ1 a1 + · · · + λn an ) = µ1 πd+2 (a1 ) + · · · + µn πd+2 (an ).
(2) For all µ1 , . . . , µn ∈ R with µ1 + · · · + µn = 1 and µi ≥ 0, for all a1 , . . . , an ∈ Hd+2 +

(resp. ∈ Hd+2
− ), there exist some λ1 , . . . , λ n ∈ R with λ1 + · · · + λ n = 1 and λi ≥ 0, so
that
πd+2 (λ1 a1 + · · · + λn an ) = µ1 πd+2 (a1 ) + · · · + µn πd+2 (an ).
(1) The base case is clear. Let us assume for the moment that we proved (1) for n = 2
and consider the induction step for n ≥ 2. Since λ1 + · · · + λn+1 = 1 and n ≥ 2, there is
some i such that λi 6= 1, and without loss of generality, say λ1 6= 1. Then, we can write

λ2 λn+1
λ1 a1 + · · · + λn+1 an+1 = λ1 a1 + (1 − λ1 ) a2 + · · · + an+1
1 − λ1 1 − λ1
and since λ1 + λ2 + · · · + λn+1 = 1, we have
λ2 λn+1
+ ··· + = 1.
1 − λ1 1 − λ1
By the induction hypothesis, for n = 2, there exist α1 with 0 ≤ α1 ≤ 1, such that

λ2 λn+1
πd+2 (λ1 a1 + · · · + λn+1 an+1 ) = πd+2 λ1 a1 + (1 − λ1 ) a2 + · · · + an+1
1 − λ1 1 − λ1

λ2 λn+1
= (1 − α1 )πd+2 (a1 ) + α1 πd+2 a2 + · · · + an+1 .
1 − λ1 1 − λ1
Again, by the induction hypothesis (for n), there exist β2 , . . . , βn+1 with β2 + · · · + βn+1 = 1
and βi ≥ 0, so that

λ2 λn+1
πd+2 a2 + · · · + an+1 = β2 πd+2 (a2 ) + · · · + βn+1 πd+2 (an+1 ),
1 − λ1 1 − λ1
so we get
πd+2 (λ1 a1 + · · · + λn+1 an+1 ) = (1 − α1 )πd+2 (a1 ) + α1 (β2 πd+2 (a2 ) + · · · + βn+1 πd+2 (an+1 ))
= (1 − α1 )πd+2 (a1 ) + α1 β2 πd+2 (a2 ) + · · · + α1 βn+1 πd+2 (an+1 ),
and clearly, 1 − α1 + α1 β2 + · · · + α1 βn+1 = 1 as β2 + · · · + βn+1 = 1; 1 − α1 ≥ 0; and α1 βi ≥ 0,

as 0 ≤ α1 ≤ 1 and βi ≥ 0. This establishes the induction step, and thus all is left is to prove
is the case n = 2.
(2) The base case n = 1 is also clear. As in (1), let us assume for a moment that (2) is
proved for n = 2 and consider the induction step. The proof is quite similar to that of (1),
but this time, we may assume that µ1 6= 1, and we write
µ1 πd+2 (a1 ) + · · · + µn+1 πd+2 (an+1 )

µ2 µn+1
= µ1 πd+2 (a1 ) + (1 − µ1 ) πd+2 (a2 ) · · · + πd+2 (an+1 ) .
1 − µ1 1 − µ1
By the induction hypothesis, there are some α2 , . . . , αn+1 with α2 +· · ·+αn+1 = 1 and αi ≥ 0
such that
µ2 µn+1
πd+2 (α2 a2 + · · · + αn+1 an+1 ) = πd+2 (a2 ) + · · · + πd+2 (an+1 ).
1 − µ1 1 − µ1
By the induction hypothesis for n = 2, there is some β1 with 0 ≤ β1 ≤ 1, so that
πd+2 ((1−β1 )a1 +β1 (α2 a2 +· · ·+αn+1 an+1 )) = µ1 πd+2 (a1 )+(1−µ1 )πd+2 (α2 a2 +· · ·+αn+1 an+1 ),
which establishes the induction hypothesis. Therefore, all that remains is to prove (1) and
(2) for n = 2.
As πd+2 is given by

x1 xd+1
πd+2 (x1 , . . . , xd+2 ) = ,..., (xd+2 6= 0),
xd+2 xd+2
it is enough to treat the case when d = 0, that is,

e
π2 (e, f ) = .
f
Let a = (a1 , b1 ) and b = (a2 , b2 ). To prove (1), we need to show that for any λ, with
0 ≤ λ ≤ 1,
π2 ((1 − λ)a + λb) = conv({π2 (a), π2 (b)}).
But since
a1 a2
π2 (a) = , π2 (b) =
b1 b2
(1 − λ)a1 + λa2
π2 ((1 − λ)a + λb) = π2 ((1 − λ)a1 + λa2 , (1 − λ)b1 + λb2 ) = ,
(1 − λ)b1 + λb2
it is enough to show that for any λ, with 0 ≤ λ ≤ 1, if b1 b2 > 0 then
a1 (1 − λ)a1 + λa2 a2 a1 a2
≤ ≤ if ≤ ,
b1 (1 − λ)b1 + λb2 b2 b1 b2
and
a2 (1 − λ)a1 + λa2 a1 a2 a1
≤ ≤ if ≤ ,
b2 (1 − λ)b1 + λb2 b1 b2 b1
where, of course, (1−λ)b1 +λb2 6= 0. For this, we compute (leaving some steps as an exercise)
(1 − λ)a1 + λa2 a1 λ(a2 b1 − a1 b2 )

− =
(1 − λ)b1 + λb2 b1 ((1 − λ)b1 + λb2 )b1
and
(1 − λ)a1 + λa2 a2 (1 − λ)(a2 b1 − a1 b2 )
− =− .
(1 − λ)b1 + λb2 b2 ((1 − λ)b1 + λb2 )b2
Now, as b1 b2 > 0, that is, b1 and b2 have the same sign, and as 0 ≤ λ ≤ 1, we have both
((1 − λ)b1 + λb2 )b1 > 0 and ((1 − λ)b1 + λb2 )b2 > 0. Then, if a2 b1 − a1 b2 ≥ 0, that is ab11 ≤ ab22
(since b1 b2 > 0), the first two inequalities hold, and if a2 b1 − a1 b2 ≤ 0, that is ab22 ≤ ab11 (since
b1 b2 > 0), the last two inequalities hold. This proves (1).
In order to prove (2), once again set a = (a1 , b1 ) and b = (a2 , b2 ). Then given any µ, with
0 ≤ µ ≤ 1, if b1 b2 > 0, we show that we can find λ with 0 ≤ λ ≤ 1, so that
a1 a2 (1 − λ)a1 + λa2
(1 − µ) +µ =
b1 b2 (1 − λ)b1 + λb2
If we let
a1 a2
α = (1 − µ) +µ ,
b1 b2
we find that λ is given by the equation
λ(a2 − a1 + α(b1 − b2 )) = αb1 − a1 .
After some (tedious) computations (check for yourself!) we find
((1 − µ)b2 + µb1 )(a2 b1 − a1 b2 )

a2 − a1 + α(b1 − b2 ) =
b1 b2
µb1 (a2 b1 − a1 b2 )
αb1 − a1 = .
b1 b2
a1 a2
If a2 b1 − a1 b2 = 0, then b1
= b2
and λ = 0 works. If a2 b1 − a1 b2 6= 0, then
µb1 µ
λ= = .
(1 − µ)b2 + µb1 (1 − µ) bb21 + µ
b2
Since b1 b2 > 0, we have b1
> 0, and since 0 ≤ µ ≤ 1, we conclude that 0 ≤ λ ≤ 1, which
proves (2).
(3) This proof is completely analogous to the proof of (2).
(4) Since
Θ = πd+2 ◦ θb ◦ id+2 ,
as id+2 and θb are linear, they preserve convex hulls, so by (2), we simply have to show that
either θb ◦ id+2 ({a1 , . . . , an }) is strictly below the hyperplane, xd+2 = 0, or strictly above it.
But
b 1 , . . . , xd+2 )d+2 = xd+2 − xd+1
θ(x
and id+2 (x1 , . . . , xd+1 ) = (x1 , . . . , xd+1 , 1), so
(θb ◦ id+2 )(x1 , . . . , xd+1 )d+2 = 1 − xd+1 ,
and this quantity is positive iff xd+1 < 1, negative iff xd+1 > 1; that is, either all the points
ai are strictly below the hyperplane Hd+1 or all strictly above it.
(5) This follows immediately from (4) as conv(S) consists of all finite convex combinations
of points in S.
If a set {a1 , . . . , an } ⊆ Ed+2 contains points on both sides of the hyperplane xd+2 = 0,
then πd+2 (conv({a1 , . . . , an })) is not necessarily convex; see Figure 12.24.
Besides θ, we need to define a few more maps in order to establish the connection between
the Delaunay complex on S d and the Delaunay complex on P. We use the convention of
denoting the extension to projective spaces of a map f defined between Euclidean spaces by
fe.
(-1, 0, 1)
z=1
(1,0,1)
(1/2, 0, 1/2)
(1/2, 0 , 0)
(1/2, 0, -1/2)
Figure 12.24: Let a1 = (1/2, 0, 1/2) and a2 = (1/2, 0, −1/2). Since π3 ((1/2, 0, 0)) is unde-
fined, the image of π3 (conv({a1 , a2 }) is two disconnected infinite rays.
Definition 12.7. The Euclidean orthogonal projection pi : Rd+1 → Rd is given by

pi (x1 , . . . , xd+1 ) = (x1 , . . . , xi−1 , xi+1 , . . . , xd+1 ),
and pei : Pd+1 → Pd denotes the projection from Pd+1 onto Pd given by
pei (x1 : · · · : xd+2 ) = (x1 : · · · : xi−1 : xi+1 : · · · : xd+2 ),
which is undefined at the point (0 : · · · : 1 : 0 : · · · : 0), where the “1” is in the ith slot. The
map πeN : (Pd+1 − {Ne }) → Pd is the central projection from the projective north pole onto
Pd given by
eN (x1 : · · · : xd+1 : xd+2 ) = (x1 : · · · : xd : xd+2 − xd+1 ) .
π
A geometric interpretation of π eN will be needed later in certain proofs. If we identify Pd
d+1
with the hyperplane Hd+1 (0) ⊆ P of equation xd+1 = 0, then we claim that for any x 6= N e,
the point π
eN (x) is the intersection of the line through N
e and x with the hyperplane Hd+1 (0).
See Figure 12.25. Indeed, parametrically, the line hN , xi through N
e e = (0 : · · · : 0 : 1 : 1) and
x is given by
hN
e , xi = {(µx1 : · · · : µxd : λ + µxd+1 : λ + µxd+2 ) | λ, µ ∈ R, λ 6= 0 or µ 6= 0}.
The line hN
e , xi intersects the hyperplane xd+1 = 0 iff
λ + µxd+1 = 0,
so we can pick λ = −xd+1 and µ = 1, which yields the intersection point,
(x1 : · · · : xd : 0 : xd+2 − xd+1 ),
as claimed.
1
P
~ P2
~ x N = (0:1:1)
π N(x)
Figure 12.25: A schematic representation, via the plane model for P2 , of the geometric image
eN (x). The copy of P1 corresponds to the intersection of the xz- plane with the plane
of π
z = 1.
Definition 12.8. The projective versions of σN and τN , denoted σ

eN : (S e }) → Pd and
fd − {N
τeN : Pd → S
fd ⊆ Pd+1 , are given by
eN (x1 : · · · : xd+2 ) = (x1 : · · · : xd : xd+2 − xd+1 ),

σ
and
d d
!
X X
τeN (x1 : · · · : xd+1 ) = 2x1 xd+1 : · · · : 2xd xd+1 : x2i − x2d+1 : x2i + x2d+1 .
i=1 i=1
fd − {N
It is an easy exercise to check that the image of S e } by σ
eN is Ud+1 , and that σ
eN
and τeN Ud+1 are mutual inverses.
Observe that σ
eN = π fd , the restriction of the projection π
eN S fd .
eN to the sphere S
l : Ed → P
Definition 12.9. The lifting e e ⊆ Pd+1 is given by
d
!
X
l(x1 , . . . , xd ) =
e x1 : · · · : xd : x2i : 1 ,
i=1
and the embedding ψd+1 : Ed → Pd (the map ψd+1 defined in Section 11.1) is given by
ψd+1 (x1 , . . . , xd ) = (x1 : · · · : xd : 1).

Then, we easily check the following facts.
Proposition 12.6. The maps θ, π

eN , τeN , ped+1 , e
l and ψd+1 defined before satisfy the equations
l = θ ◦ τeN ◦ ψd+1
e
eN = ped+1 ◦ θ
π
τeN ◦ ψd+1 = ψd+2 ◦ τN
l = ψd+2 ◦ l
e
l = Θ ◦ τN .
Proof. Recall that θ is given by
θ(x1 : · · · : xd+2 ) = (x1 : · · · : xd : xd+1 + xd+2 : xd+2 − xd+1 ).
Then, as
d d
!
X X
τeN ◦ ψd+1 (x1 , . . . , xd ) = 2x1 : · · · : 2xd : x2i − 1 : x2i + 1 ,
i=1 i=1
we get
d
!
X
θ ◦ τeN ◦ ψd+1 (x1 , . . . , xd ) = 2x1 : · · · : 2xd : 2 x2i : 2
i=1
d
!
X
= x1 : · · · : xd : x2i : 1 =e
l(x1 , . . . , xd ),
i=1
which proves the first equation.

For the second equation, since ped+1 drops the (d + 1)th component of a (d + 2)-tuple, the
eN = ped+1 ◦ θ follows immediately from the definitions.
equation π
We have
τeN (ψd+1 (x)) = τeN (x1 : · · · : xd : 1)

= (2x1 : · · · : 2xd : kxk2 − 1 : kxk2 + 1),
and
!
2x1 2xd kxk2 − 1
ψd+2 (τN (x)) = : · · · : : :1
kxk2 + 1 kxk2 + 1 kxk2 + 1
= (2x1 : · · · : 2xd : kxk2 − 1 : kxk2 + 1),
so the third equation holds.

Since ψd+2 adds a 1 as a (d + 2)th component, the fourth equation follows immediately
from the definitions.
For the fifth equation, since
!
2
2x kxk − 1
τN (x) = ,
kxk + 1 kxk2 + 1
2
and
xi
Θ(x)i = , 1≤i≤d
1 − xd+1
xd+1 + 1
Θ(x)d+1 = ,
1 − xd+1
we get
!
kxk2 − 1

2xi
Θ(τN (x))i = 1−
kxk2 + 1 kxk2 + 1
2xi
= = xi , 1 ≤ i ≤ d,
2
! !
kxk2 − 1 kxk2 − 1
θ(τN (x))d+1 = +1 1−
kxk2 + 1 kxk2 + 1
2 kxk2
= = kxk2 ,
2
and since
d
X
l(x) = (x1 , . . . , xd , x2i ) = (x1 , . . . , xd , kxk2 ),
i=1
we have shown that l = Θ ◦ τN , as claimed.
12.7 Lifted Delaunay Complexes and Delaunay

Complexes via Lifting to a Sphere
In order to define precisely Delaunay complexes as projections of objects obtained by delet-
ing some faces from a projective polyhedron we need to define the notion of “projective
(polyhedral) complex.” However, this is easily done by defining the notion of cell complex
where the cells are polyhedral cones. Such objects are known as fans. The definition below
is basically Definition 9.10 in which the cells are cones as opposed to polytopes.
12.7. LIFTED DELAUNAY COMPLEXES AND DELAUNAY COMPLEXES 365
Definition 12.10. A fan in Em is a set K consisting of a (finite or infinite) set of polyhedral

cones in Em satisfying the following conditions:
(1) Every face of a cone in K also belongs to K.
(2) For any two cones σ1 and σ2 in K, if σ1 ∩ σ2 6= ∅, then σ1 ∩ σ2 is a common face of
both σ1 and σ2 . See Figure 12.26.
Every cone σ ∈ K of dimension k is called a k-face (or face) of K. A 0-face {v} is called a
vertex , and a 1-face is called an edge. The dimension of the fan K is the maximum of the
dimensions of all cones in K. If dim K = d, then every face of dimension d is called a cell ,
and every face of dimension d − 1 is called a facet.
Figure 12.26: Five polyhedral cones in E3 combined into a fan.

d
A projective (polyhedral) complex K ⊆ 2P is a set of projective polyhedra of the form
d+1
{P(C) | C ∈ K}, where K ⊆ 2R is a fan.
Given a projective complex, the notions of face, vertex, edge, cell, facet, are defined in
the obvious way.
d
If K ⊆ 2R is a polyhedral complex, then it is easy to check that the set {C(σ) | σ ∈
d+1
K} ⊆ 2R (where C(σ) is the V-cone associated with σ defined in Section 5.5) is a fan.
d
Definition 12.11. Given a polyhedral complex K ⊆ 2R , the projective complex
e = {P(C(σ)) | σ ∈ K} ⊆ 2Pd
K
is called the projective completion of K. See Figure 12.27.
~
K
P2
Figure 12.27: The projective completion, in the plane model of P2 , of K, the two dimensional
complex in the xy-plane consisting of the teal triangle and the periwinkle quadrilateral.
Also, it is easy to check that if f : P → P 0 is an injective affine map between two

polyhedra P and P 0 , then f extends uniquely to a projectivity fe: Pe → Pf0 between the
projective completions of P and P 0 .
We now have all the facts needed to show that Delaunay triangulations and Voronoi
diagrams can be defined in terms of the lifting τeN ◦ ψd+1 , and the projection π
eN , and to
d
establish their duality via polar duality with respect to S .
Definition 12.12. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , the polytope D(P ) ⊆
Rd+1 , called the Delaunay polytope associated with P , is the convex hull of the union of the
lifting of the points of P onto the sphere S d (via inverse stereographic projection) with the
north pole; that is, D(P ) = conv(τN (P ) ∪ {N }). See Figures 12.28, 12.29, and 12.30. The
e ) ⊆ Pd+1 associated with P is the projective completion of
projective Delaunay polytope D(P
d+1
D(P ). The polyhedral complex DC(P ) ⊆ 2R , called the lifted Delaunay complex of P , is
the complex obtained from the boundary of D(P ) by deleting the facets containing the north
f ) ⊆ 2Pd+1 is the projective
pole (and their faces), as illustrated in Figure 12.31, and DC(P
completion of DC(P ). The polyhedral complex Del (P ) = ϕd+1 ◦ π f )) ⊆ 2Ed is the
eN (DC(P
Delaunay complex of P .
The above is not the “standard” definition of the Delaunay triangulation of P , but it
is equivalent to the definition given in Section 17.3.1 of Boissonnat and Yvinec [12], as we
will prove shortly. It also has certain advantages over lifting onto a paraboloid, as we will
explain. Furthermore, to be perfectly rigorous, we should define Del (P ) by
Del (P ) = ϕd+1 (e f )) ∩ 2Ud+1 ),

πN (DC(P
Figure 12.28: The inverse stereographic projection of P = {(0, 0, 0), (1, 0, 0), (0, 1, 0),
(2, 2, 0)}. Note that τN ((0, 0, 0)) = (0, 0, −1) and τN ((2, 2, 0)) = (4/9, 4/9, 7/9).
f )) ⊆ 2Ud+1 because DC(P ) is strictly below the hyperplane Hd+1 .

eN (DC(P
but π
The Delaunay complex Del (P ) is often called the Delaunay triangulation of P , but this
terminology is slightly misleading because Del (P ) is not simplicial unless P is in general
position; see Proposition 12.11.
It it possible and useful to define Del (P ) more directly in terms of DC(P ). The projection
d+1 e }) → Pd comes from the linear map π
πf
N : (P − {N bN : Rd+2 → Rd+1 given by
bN (x1 , . . . , xd+1 , xd+2 ) = (x1 , . . . , xd , xd+2 − xd+1 ).

π
Consequently, as DC(P
f ) = DC(P
^) = P(C(DC(P ))), we immediately check that
Del (P ) = ϕd+1 ◦ π
eN (DC(P
f )) = ϕd+1 ◦ π
bN (C(DC(P ))) = ϕd+1 ◦ π
bN (cone(DC(P
\))),
where DC(P
\) = {b
u | u ∈ DC(P )} and u
b = (u, 1).
This suggests defining the map πN : (Rd+1 − Hd+1 ) → Rd by
πN = ϕd+1 ◦ π
bN ◦ id+2 ,
which is explicity given by

1
πN (x1 , . . . , xd , xd+1 ) = (x1 , . . . , xd ).
1 − xd+1
Observe that the map πN is just the central projection from the north pole to the hyperplane
xd+1 = 0. Then, as DC(P ) is strictly below the hyperplane Hd+1 , we have
Del (P ) = ϕd+1 ◦ π
eN (DC(P
f )) = πN (DC(P )).

Figure 12.29: Two views of the Delaunay polytope D(P ), where P = {(0, 0, 0), (1, 0, 0),
(0, 1, 0), (2, 2, 0)}.
First, note that Del (P ) = ϕd+1 ◦ πeN (DC(P

f )) = πN (DC(P )) is indeed a polyhedral
complex whose geometric realization is the convex hull conv(P ) of P . Indeed, by Proposition
12.5, the images of the facets of DC(P ) are polytopes, and when any two such polytopes
meet, they meet along a common face. Furthermore, if dim(conv(P )) = m, then Del (P )
is pure m-dimensional. First, Del (P ) contains at least one m-dimensional cell. If Del (P )
was not pure, as the complex is connected there would be some cell σ of dimension s < m
meeting some other cell τ of dimension m along a common face of dimension at most s, and
because σ is not contained in any face of dimension m, no facet of τ containing σ ∩ τ can be
adjacent to any cell of dimension m, and so Del (P ) would not be convex, a contradiction.
Our next goal is to show that the Delaunay complex Del (P ) coincides with the “standard”
Delaunay complex Del 0 (P ), as defined in Section 17.3.1 of Boissonnat and Yvinec [12]. To
define Del 0 (P ), we need the notion of lower-facing facet.
For any polytope P ⊆ Ed , given any point x not in P , recall that a facet F of P is
visible from x iff for every point y ∈ F , the line through x and y intersects P only in y. If
dim(P ) = d, this is equivalent to saying that x and the interior of P are strictly separated
by the supporting hyperplane of F . Note that if dim(P ) < d, it possible that every facet of
P is visible from x. See Figure 12.33.
Definition 12.13. Assume that P ⊆ Ed is a polytope with dim(P ) = d. We say that a

facet F of P is a lower-facing facet of P iff the unit normal to the supporting hyperplane
of F pointing towards the interior of P has non-negative xd+1 -coordinate. A facet F that is
not lower-facing is called an upper-facing facet (note that in this case, the xd+1 coordinate
of the unit normal to the supporting hyperplane of F pointing towards the interior of P is
strictly negative). See Figure 12.34.
Figure 12.30: Two additional views of the Delaunay polytope D(P ) circumscribed by the
unit sphere.
Here is a convenient way to characterize lower-facing facets.

Proposition 12.7. Given any polytope P ⊆ Ed with dim(P ) = d, for any point c on the
Oxd -axis, if c lies strictly above all the intersection points of the Oxd -axis with the supporting
hyperplanes of all the upper-facing facets of F , then the lower-facing facets of P are exactly
the facets not visible from c. See Figure 12.35.
Proof. Note that the intersection points of the Oxd -axis with the supporting hyperplanes
of all the upper-facing facets of P are strictly above the intersection points of the Oxd -axis
with the supporting hyperplanes of all the lower-facing facets. Suppose F is visible from c.
Then, F must not be lower-facing, as otherwise, for any y ∈ F , the line through c and y has
to intersect some upper-facing facet and F is not be visible from c, a contradiction.
Now, as P is the intersection of the closed half-spaces determined by the supporting
hyperplanes of its facets, by the definition of an upper-facing facet, any point c on the
Oxd -axis that lies strictly above the intersection points of the Oxd -axis with the supporting
hyperplanes of all the upper-facing facets of F has the property that c and the interior
of P are strictly separated by all these supporting hyperplanes. Therefore, all the upper-
facing facets of P are visible from c. It follows that the facets visible from c are exactly the
upper-facing facets, as claimed.
We will also need the following fact.

Proposition 12.8. Given any polytope P ⊆ Ed , if dim(P ) = d, then there is a point c
on the Oxd -axis such that for all points x on the Oxd -axis and above c, the set of facets of
Figure 12.31: The lifted Delaunay complex DC(P ), where P = {(0, 0, 0), (1, 0, 0), (0, 1, 0),
(2, 2, 0)}.
conv(P ∪ {x}) not containing x is identical. Moreover, the set of facets of P not visible from
x is the set of facets of conv(P ∪ {x}) that do not contain x.
Proof. If dim(P ) = d then pick any c on the Oxd -axis above the intersection points of the
Oxd -axis with the supporting hyperplanes of all the upper-facing facets of F . Then, c is in
general position w.r.t. P in the sense that c and any d vertices of P do not lie in a common
hyperplane. Now, our result follows by Lemma 8.3.1 of Boissonnat and Yvinec [12].
Corollary 12.9. Given any polytope P ⊆ Ed with dim(P ) = d, there is a point c on the
Oxd -axis so that for all x on the Oxd -axis and above c, the lower-facing facets of P are
exactly the facets of conv(P ∪ {x}) that do not contain x. See Figure 12.36.
As usual, let ed+1 = (0, . . . , 0, 1) ∈ Rd+1 .

The standard Delaunay polyhedron and the standard Delaunay complex are defined as
follows (compare Boissonnat and Yvinec [12], Section 17.3.1).
Definition 12.14. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , let D0 (P ) denote the
polyhedron conv(l(P )) + cone(ed+1 ), and let D e 0 (P ) be the projective completion of D0 (P ).
Also, let DC 0 (P ) be the polyhedral complex consisting of the bounded facets of the polytope
D0 (P ), and let DCf 0 (P ) be the projective completion of DC 0 (P ). See Figure 12.37. The
f 0 (P )) = pd+1 (DC 0 (P )) is the standard Delaunay complex
complex Del 0 (P ) = ϕd+1 ◦ ped+1 (DC
of P , that is, the orthogonal projection of DC 0 (P ) onto Ed . See Figure 12.38.
Intuitively, adding to conv(l(P )) all the verttical rays parallel to ed+1 based on points
in conv(l(P )) washes out the upper-facing faces of conv(l(P )). Then the bounded facets of
conv(l(P ))+cone(ed+1 ) are precisely the lower-facing facets of conv(l(P )) (if dim(conv(P )) =
d).
The first of the two main theorems of this chapter is that the two notions of Delaunay
complexes coincide.
Figure 12.32: The Delaunay complex Del (P ), for P = {(0, 0, 0), (1, 0, 0), (0, 1, 0), (2, 2, 0)}
obtained by applying πN to DC(P ) of Figure 12.31. The bottom triangle projects onto
the planar green triangle with vertices {(0, 0, 0), (1, 0, 0), (0, 1, 0)}, while the top triangle
projection onto the aqua triangle with vertices {(1, 0, 0), (0, 1, 0), (2, 2, 0)}.
Theorem 12.10. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , we have
θ(D(P e 0 (P )
e )) = D and θ(DC(P f 0 (P ).
f )) = DC
Furthermore,
Del (P ) = Del 0 (P ).
Therefore, the two notions of a Delaunay complex agree. If dim(conv(P )) = d, then the
bounded facets of conv(l(P )) + cone(ed+1 ) are precisely the lower-facing facets of conv(l(P )).
Proof. Recall that
D(P ) = conv(τN (P ) ∪ {N }),
and D(P
e ) = P(C(D(P ))) is the projective completion of D(P ). If we write τ\
N (P ) for
{τN (pi ) | pi ∈ P }, then
\
C(D(P )) = cone(τ\N (P ) ∪ {N }).
b
By definition, we have
θ(D)
e = P(θ(C(D))).
b
Now, as θb is linear,
N (P ) ∪ {N })) = cone(θ(τN (P )) ∪ {θ(N )}).

τ\
θ(C(D))
b = θ(cone(
b b b \ b b
w
(0,0,0)
y3
y
2
P y
1
(i.)
(ii.)
Figure 12.33: In E3 , all three edges of the planar triangle in Figure (i.) are visible from x,
while in Figure (ii.), only the top face of the solid rectangular box is visible from x.
We claim that
N (P )) ∪ {θ(N )}) = cone(l(P ) ∪ {(0, . . . , 0, 1, 1)})

b τ\
cone(θ( b b d
= C(D0 (P )),
where
D0 (P ) = conv(l(P )) + cone(ed+1 ).
Indeed,
b 1 , . . . , xd+2 ) = (x1 , . . . , xd , xd+1 + xd+2 , xd+2 − xd+1 ),
θ(x
and for any pi = (x1 , . . . , xd ) ∈ P ,
! Pd
2
2x1 2xd x i − 1
τ\
N (pi ) = Pd 2
, . . . , Pd 2
, Pi=1
d 2
,1
i=1 xi + 1 i=1 xi + 1 i=1 xi + 1
d d
!
1 X X
= Pd 2
2x1 , . . . , 2xd , x2i − 1, x2i + 1 ,
i=1 xi + 1 i=1 i=1
so we get
d
!
2 X 2
b τ\
θ( N (pi )) = Pd
2
x1 , . . . , x d , x2i , 1 = Pd 2
l(p
di ).
i=1 xi +1 i=1 i=1 xi +1
Also, we have
θ(
bN b ) = θ(0,
b . . . , 0, 1, 1) = (0, . . . , 0, 2, 0) = 2ed
d+1 ,
and by definition of cone(−) (scalar factors are irrelevant), we get

0
N (P )) ∪ {θ(N )}) = cone(l(P ) ∪ {(0, . . . , 0, 1, 1)}) = C(D (P )),
b τ\
cone(θ( b b d
Figure 12.34: Let P be the solid mint green triangular bipyramid. The three faces on the
top are upper-facing, while the three faces on the bottom are lower-facing.
p1
p2
Figure 12.35: Let P be the solid mint green triangular bipyramid. The intersections of the
upper-facing facets and the z-axis is given by the two points p1 and p2 . Since c is above p1 ,
none of the three lower-facing facets in the bottom of the bipyramid are visible from c.
with D0 (P ) = conv(l(P )) + cone(ed+1 ), as claimed. This proves that
θ(D(P e 0 (P ).
e )) = D
Now, it is clear that the facets of conv(τN (P ) ∪ {N }) that do not contain N are mapped
to the bounded facets of conv(l(P )) + cone(ed+1 ), since N goes the point at infinity, so
θ(DC(P f 0 (P ).
f )) = DC
eN = ped+1 ◦ θ, by Proposition 12.6, we get

As π
0
Del 0 (P ) = ϕd+1 ◦ ped+1 (DC
f (P )) = ϕd+1 ◦ (e
pd+1 ◦ θ)(DC(P
f )) = ϕd+1 ◦ π
eN (DC(P
f )) = Del (P ),
Figure 12.36: Let P be the solid mint green triangular bipyramid. Then conv(P ∪ {x}) is
the larger solid bipyramid with gray top and mint green bottom. The lower-facing facets of
P are the three mint green faces on the bottom of both P and conv(P ∪ {x}).
Figure 12.37: Two views of the tetrahedron l(P ), where P = {(0, 0, 0), (1, 0, 0), (0, 1, 0),
(2, 2, 0)}.
as claimed. Finally, if dim(conv(P )) = d, then, by Corollary 12.9, we can pick a point c on

the Oxd+1 -axis so that the facets of conv(l(P ) ∪ {c}) that do not contain c are precisely the
lower-facing facets of conv(l(P )). However, it is also clear that the facets of conv(l(P ) ∪ {c})
that contain c tend to the unbounded facets of DC 0 (P ) = conv(l(P )) + cone(ed+1 ) when c
goes to +∞.
We can also characterize when the Delaunay complex Del (P ) is simplicial. Recall that
we say that a set of points P ⊆ Ed is in general position if no d + 2 of the points in P belong
to a common (d − 1)-sphere.
Proposition 12.11. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , if P is in general

position, then the Delaunay complex Del (P ) is a pure simplicial complex. The lifted Delaunay
complex DC(P ) is also a pure simplicial complex.
12.8. LIFTED VORONOI COMPLEXES AND VORONOI COMPLEXES 375
Figure 12.38: The polyhedral complex DC 0 (P ) for P = {(0, 0, 0), (1, 0, 0), (0, 1, 0), (2, 2, 0)},
and its orthogonal projection (in red) onto the xy-plane. Note this orthogonal projection
gives the same Delaunay complex as in Figure 12.32.
Proof. Let dim(conv(P )) = r. Then, τN (P ) is contained in a (r − 1)-sphere of S d , so we may

assume that r = d. Suppose Del (P ) has some facet F which is not a d-simplex. If so, F is
the convex hull of at least d + 2 points p1 , . . . , pk of P , and since F = πN (Fb) for some facet
Fb of DC(P ), we deduce that τN (p1 ), . . . , τN (pk ) belong to the supporting hyperplane H of
Fb. Now, if H passes through the north pole, then we know that p1 , . . . , pk belong to some
hyperplane of Ed , which is impossible since p1 , . . . , pk are the vertices of a facet of dimension
d. Thus, H does not pass through N , and so p1 , . . . , pk belong to some (d − 1)-sphere in Ed .
As k ≥ d + 2, this contradicts the assumption that the points in P are in general position.
The proof that DC(P ) is a pure simplicial complex is similar. A similar proof is also
given in Boissonnat and Yvinec [12], Section 17.3.2.
Remark: Even when the points in P are in general position, the Delaunay complex D(P )
may not be a simplicial polytope. For example, if d + 1 points belong to a hyperplane in
Ed , then the lifted points belong to a hyperplane passing through the north pole, and these
d + 1 lifted points together with N form a non-simplicial facet. For example, consider the
polytope obtained by lifting our original d + 1 points on a hyperplane H plus one more point
not in the hyperplane H; see Figure 12.39.
12.8 Lifted Voronoi Complexes and Voronoi

Complexes via Lifting to a Sphere
Our final goal is to characterize the Voronoi diagram of P in terms of the polar dual D(P )∗
of D(P ) (with respect to the sphere S d ). The polar dual D(P )∗ of D(P ) is the polyhedron
obtained by intersecting the half-spaces containing the origin associated with the tangent
N
H
p
C
B
A
Figure 12.39: Let d = 2. The three points on the line H and the orange point p are in
general position. However, when lifted to S 2 , the convex hull of these three points and N
form a solid pyramid with base ABCN .
hyperplanes to S d at the lifted points τN (pi ) (with pi ∈ P ), and at the north pole N .
See Figures 12.40, and 12.41. It follows that the polyhedron D(P )∗ has exactly one facet
containing the north pole. The Voronoi diagram of P is the result of applying the central
projection πN from N to the polyhedron D(P )∗ . Under this central projection, the facet
containing the north pole goes to infinity, so instead of considering the polar dual D(P )∗ we
should consider the polar dual DC(P )∗ of the lifted Delaunay complex DC(P ) which does
not have the north pole as a vertex. Then the Voronoi diagram of P is the result of applying
the central projection πN from N to the complex DC(P )∗ . See Figures 12.42 through 12.45.
The polyhedron DC(P )∗ still contains faces intersecting the tangent hyperplane to S d
at the north pole, so we can’t simply map it to the corresponding complex obtained from
the polar dual of the lifted points l(pi ) on the paraboloid P. However, using projective
completions, we can indeed define this mapping and recover the Voronoi diagram of P .
Definition 12.15. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , the lifted Voronoi complex
associated with P is the polar dual (w.r.t. S d ⊆ Rd+1 ) V(P ) = (DC(P ))∗ ⊆ Rd+1 of the
e ) ⊆ Pd+1 is the projective completion of V(P ).
lifted Delaunay complex DC(P ), and V(P
See Figure 12.46. The polyhedral complex Vor (P ) = ϕd+1 (e e )) ∩ 2Ud+1 ) ⊆ 2Ed is the
πN (V(P
Voronoi complex of P , or Voronoi diagram of P . See Figure 12.47.
As in Section 12.7 (just after Definition 12.12), it is easy to see that
Vor (P ) = ϕd+1 (e e )) ∩ 2Ud+1 ) = πN (V(P )).

πN (V(P
Figure 12.40: The polar dual D(P )∗ , where P = {(0, 0, 0), (1, 0, 0), (0, 1, 0), (2, 2, 0)}. The
inverse stereographic projection of each point of P , along with N , is the depicted by a red
dot, and the five teal faces of the unbounded wedge are tangent to each red point.
Definition 12.16. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , let V 0 (P ) = (DC 0 (P ))∗
be the polar dual (w.r.t. P ⊆ Rd+1 ) of the “standard” Delaunay complex of Definition 12.14,
and let V f 0 (P ) ⊆ Pd be its projective completion. The standard Voronoi diagram
e 0 (P ) = DC
is given by Vor 0 (P ) = pd+1 (V 0 (P )); see Definition 17.2.7 of Boissonnat and Yvinec [12].
It is not hard to check that
Vor 0 (P ) = pd+1 (V 0 (P )) = ϕd+1 (e e 0 (P )) ∩ Ud+1 ).

pd+1 (V
In order to prove our second main theorem we need to show that θ has a good be-
havior with respect to tangent spaces. Recall from Section 11.2 that for any point a =
(a1 : · · · : ad+2 ) ∈ Pd+1 , the tangent hyperplane Ta S fd to the sphere S
fd at a is given by the
equation
d+1
X
ai xi − ad+2 xd+2 = 0.
i=1
Similarly, the tangent hyperplane Ta P

e to the paraboloid P
e at a is given by the equation
d
X
2 ai xi − ad+2 xd+1 − ad+1 xd+2 = 0.
i=1
If we lift a point a ∈ Ed to S
fd by τeN ◦ ψd+1 and to P
e by e
l, it turns out that the image of the
tangent hyperplane to S at τeN ◦ ψd+1 (a) by θ is the tangent hyperplane to P
f d e at e
l(a).
Figure 12.41: The polar dual D(P )∗ , where P = {(0, 0, 0), (1, 0, 0), (0, 1, 0), (2, 2, 0)}, and
D(P ) is the Delaunay polytope of Figures 12.29 and 12.30.
Proposition 12.12. The map θ has the following properties:
(1) For any point a = (a1 , . . . , ad ) ∈ Ed , we have
fd ) = Te P,
θ(TτeN ◦ψd+1 (a) S e
l(a)
that is, θ preserves tangent hyperplanes.
(2) For every (d − 1)-sphere S ⊆ Ed , we have
τN ◦ ψd+1 (S)) = e
θ(e l(S),
e
that is, θ preserves lifted (d − 1)-spheres.
Proof. (1) By Proposition 12.6, we know that
l = θ ◦ τeN ◦ ψd+1
e
and we proved in Section 11.3 (Proposition 11.6) that projectivities preserve tangent spaces.
Thus,
θ(Tτe ◦ψ (a) S fd ) = Te P,
fd ) = Tθ◦eτ ◦ψ (a) θ(S e
N d+1 N d+1 l(a)
as claimed.
l = θ ◦ τeN ◦ ψd+1 .
(2) This follows immediately from the equation e
Figure 12.42: Five views of the dual to the lifted Delaunay complex of Figure 12.31.
Given any two distinct points a = (a1 , . . . , ad ) and b = (b1 , . . . , bd ) in Ed , recall that the
bisector hyperplane Ha,b of a and b is given by
(b1 − a1 )x1 + · · · + (bd − ad )xd = (b21 + · · · + b2d )/2 − (a21 + · · · + a2d )/2.
We have the following useful proposition:
Proposition 12.13. Given any two distinct points a = (a1 , . . . , ad ) and b = (b1 , . . . , bd ) in
Ed , the image under the projection π fd ∩ Tτe ◦ψ (b) S
eN of the intersection TτeN ◦ψd+1 (a) S N d+1
fd of the
tangent hyperplanes at the lifted points τeN ◦ ψd+1 (a) and τeN ◦ ψd+1 (b) on the sphere Sfd ⊆ Pd+1
d
is the embedding of the bisector hyperplane Ha,b of a and b into P ; that is,
π fd ∩ Tτe ◦ψ (b) S
eN (TτeN ◦ψd+1 (a) S fd ) = ψd+1 (Ha,b ).
N d+1
Proof. In view of the geometric interpretation of π eN given earlier, we need to find the
equation of the hyperplane H passing through the intersection of the tangent hyperplanes
fd and Tτe ◦ψ (b) S
TτeN ◦ψd+1 (a) S fd , and passing through the north pole, and then it is geomet-
N d+1
rically obvious that
π fd ∩ Tτe ◦ψ (b) S
eN (TτeN ◦ψd+1 (a) S fd ) = H ∩ Hd+1 (0),
N d+1
Figure 12.43: Four views of the relationship between DC(P )∗ of Figure 12.42 and the De-
launay complex of Figure 12.32.
where Hd+1 (0) is the hyperplane (in Pd+1 ) of equation xd+1 = 0. Recall that TτeN ◦ψd+1 (a) S
fd
fd are given by
and TτeN ◦ψd+1 (b) S
d
X d
X d
X
E1 = 2 ai x i + ( a2i − 1)xd+1 − ( a2i + 1)xd+2 = 0
i=1 i=1 i=1
and
d
X d
X Xd
E2 = 2 bi x i + ( b2i − 1)xd+1 − ( b2i + 1)xd+2 = 0.
i=1 i=1 i=1
fd ∩ Tτe ◦ψ (b) S
The hyperplanes passing through TτeN ◦ψd+1 (a) S fd are given by an equation of
N d+1
the form
λE1 + µE2 = 0,
with λ, µ ∈ R. Furthermore, in order to contain the north pole, this equation must vanish
for x = (0 : · · · : 0 : 1 : 1). But, observe that setting λ = −1 and µ = 1 gives a solution since
the corresponding equation is
d
X d
X d
X Xd d
X
2 (bi − ai )xi + ( b2i − a2i )xd+1 −( 2
bi − a2i )xd+2 = 0,
i=1 i=1 i=1 i=1 i=1
F = (-1/2, 1, 1)
E = (1, -1/2, 1)
D = (-2, 1, -1)
C = (1, -2, -1) (2,2,0)

B = (1,1,-1)
πN ( C ) πN ( D )
(1/2, -1, 0) (-1, 1/2, 0)
π (B)
N (1/2, 1/2, 0)
π N (A )
(7/6, 7/6, 0)
(2,2,0)
Figure 12.44: The Voronoi diagram for the Delaunay triangulation of the red dots P =
{(0, 0, 0), (1, 0, 0), (0, 1, 0), (2, 2, 0)}. The green and black dots are projected, via πN , onto
the aqua dots. Note that E and F are mapped to infinity.
and it vanishes on (0 : · · · : 0 : 1 : 1). But then, the intersection of H with the hyperplane
Hd+1 (0) of equation xd+1 = 0 is given by
d
X Xd d
X
2 (bi − ai )xi − ( b2i − a2i )xd+2 = 0.
i=1 i=1 i=1
Since we view Pd as the hyperplane Hd+1 (0) ⊆ Pd+1 and since the coordinates of points
in Hd+1 (0) are of the form (x1 : · · · : xd : 0 : xd+2 ), the above equation is equivalent to the
equation of ψd+1 (Ha,b ) in Pd in which xd+1 is replaced by xd+2 .
Here is the second main theorem of this chapter.
Theorem 12.14. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , we have
θ(V(P e 0 (P )
e )) = V
and
Vor (P ) = Vor 0 (P ).
Therefore, the two notions of Voronoi diagrams agree.
(2,2)
(7/6, 7/6)
(0,1)
(1/2, 1/2)
(0,0) (1,0)
Figure 12.45: Another view of the Voronoi diagram (in blue) for the Delaunay triangulation
(in red) of P = {(0, 0, 0), (1, 0, 0), (0, 1, 0), (2, 2, 0)}.
Figure 12.46: The polar dual of DC 0 (P ) from Figure 12.38.
Proof. By definition,
V(P
e ) = V(P ^)∗ ,
]) = DC(P
and by Proposition 11.13,
∗
∗ f ))∗ ,
DC(P ) = DC(P ) = (DC(P
^ ^
so
V(P f ))∗ .
e ) = (DC(P
By Proposition 11.11,
θ(V(P f ))∗ ) = (θ(DC(P
e )) = θ((DC(P f )))∗ ,
and by Theorem 12.10,

0
θ(DC(P
f )) = DC
f (P ),
(7/6, 7/6, 4/3)
Figure 12.47: The orthogonal projection of the dark green lines onto the lighter green
lines (which lie in the xy-plane, provide the Voronoi diagram of P = {(0, 0, 0), (1, 0, 0),
(2, 2, 0), (0, 1, 0)}. This is precisely the same Voronoi diagram as in Figure 12.45.
so we get
θ(V(P f 0 (P ))∗ .
e )) = (DC
But, by Proposition 11.13 again,

∗
f 0 (P ))∗ = DC

0 0 e 0 (P ).
(DC ^ ^
(P ) = DC (P )∗ = V
^ 0 (P ) = V
Therefore,
θ(V(P e 0 (P ),
e )) = V
as claimed.
eN = ped+1 ◦ θ by Proposition 12.6, we get
As π
Vor 0 (P ) = ϕd+1 (e e 0 (P )) ∩ 2Ud+1 )

pd+1 (V
e )) ∩ 2Ud+1 )
pd+1 ◦ θ(V(P
= ϕd+1 (e
e )) ∩ 2Ud+1 )
πN (V(P
= ϕd+1 (e
= Vor (P ),
finishing the proof.
We can also prove the proposition below which shows directly that Vor (P ) is the Voronoi
diagram of P . Recall that that V(P
e ) is the projective completion of V(P ). We observed in
Section 11.2 (see page 302) that in the patch Ud+1 , there is a bijection between the faces of
V(P
e ) and the faces of V(P ). Furthermore, the projective completion H e of every hyperplane
d
H ⊆ R is also a hyperplane, and it is easy to see that if H is tangent to V(P ), then H e is
tangent to V(P
e ).
Proposition 12.15. Given any set of points P = {p1 , . . . , pn } ⊆ Ed , for every p ∈ P , if F

is the facet of V(P ) that contains τN (p), if H is the tangent hyperplane at τN (p) to S d , and
if F is cut out by the hyperplanes H, H1 , . . . , Hkp , in the sense that
F = (H ∩ H1 )− ∩ · · · ∩ (H ∩ Hkp )− ,
where (H ∩ Hi )− denotes the closed half-space in H containing τN (p) determined by H ∩ Hi ,

then
V (p) = ϕd+1 (e e ∩H
πN ( H e 1 )− ∩ · · · ∩ π e ∩H
eN (H e kp )− ∩ Ud+1 )
is the Voronoi region of p (where ϕd+1 (e e ∩H

πN ( H e i )− ∩Ud+1 ) is the closed half-space containing
p). If P is in general position and dim(conv(P )) = d, then V(P ) and Vor (P ) are simple
polyhedra (every vertex belongs to d + 1 facets).
Proof. Recall that by Proposition 12.6,
τeN ◦ ψd+1 = ψd+2 ◦ τN .
Each Hi = TτN (pi ) S d is the tangent hyperplane to S d at τN (pi ), for some pi ∈ P . Now,
by definition of the projective completion, the embedding V(P ) −→ V(P e ) is given by a 7→
ψd+2 (a). Thus, every point p ∈ P is mapped to the point ψd+2 (τN (p)) = τf N (ψd+1 (p)), and
d d
we also have Hi = TτeN ◦ψd+1 (pi ) S and H = TτeN ◦ψd+1 (p) S . By Proposition 12.13,
e e
eN (TτeN ◦ψd+1 (p) S d ∩ TτeN ◦ψd+1 (pi ) S d ) = ψd+1 (Hp,pi )

π
is the embedding of the bisector hyperplane of p and pi in Pd , so the first part holds.
Since dim(conv(P )) = d every vertex of V(P ) must belong to at least d + 1 faces.Now,
assume that some vertex v ∈ V(P ) = DC(P )∗ belongs to k ≥ d + 2 facets of V(P ). By
polar duality, this means that the facet F dual of v has k ≥ d + 2 vertices τN (p1 ), . . . , τN (pk )
of DC(P ). However, this contradicts Proposition 12.11. The fact that Vor (P ) is a simple
polyhedron was aready proved in Proposition 12.1.
Note that if m = dim(conv(P )) < d, then the Voronoi complex V(P ) may not have any
vertices.
We conclude our presentation of Voronoi diagrams and Delaunay triangulations with a
short section on applications.
12.9 Applications of Voronoi Diagrams and Delaunay

Triangulations
The examples below are taken from O’Rourke [44]. Other examples can be found in Preparata
and Shamos [47], Boissonnat and Yvinec [12], and de Berg, Van Kreveld, Overmars, and
Schwarzkopf [6].
12.9. APPLICATIONS 385
The first example is the nearest neighbors problem. There are actually two subproblems:
Nearest neighbor queries and all nearest neighbors.
The nearest neighbor queries problem is as follows. Given a set P of points and a query
point q, find the nearest neighbor(s) of q in P . This problem can be solved by computing the
Voronoi diagram of P and determining in which Voronoi region q falls. This last problem,
called point location, has been heavily studied (see O’Rourke [44]). The all neighbors problem
is as follows: Given a set P of points, find the nearest neighbor(s) to all points in P . This
problem can be solved by building a graph, the nearest neighbor graph, for short nng. The
nodes of this undirected graph are the points in P , and there is an arc from p to q iff p is
a nearest neighbor of q or vice versa. Then it can be shown that this graph is contained in
the Delaunay triangulation of P .
The second example is the largest empty circle. Some practical applications of this
problem are to locate a new store (to avoid competition), or to locate a nuclear plant as
far as possible from a set of towns. More precisely, the problem is as follows. Given a set
P of points, find a largest empty circle whose center is in the (closed) convex hull of P ,
empty in that it contains no points from P inside it, and largest in the sense that there is no
other circle with strictly larger radius. The Voronoi diagram of P can be used to solve this
problem. It can be shown that if the center p of a largest empty circle is strictly inside the
convex hull of P , then p coincides with a Voronoi vertex. However, not every Voronoi vertex
is a good candidate. It can also be shown that if the center p of a largest empty circle lies
on the boundary of the convex hull of P , then p lies on a Voronoi edge.
The third example is the minimum spanning tree. Given a graph G, a minimum spanning
tree of G is a subgraph of G that is a tree, contains every vertex of the graph G, and minimizes
the sum of the lengths of the tree edges. It can be shown that a minimum spanning tree
is a subgraph of the Delaunay triangulation of the vertices of the graph. This can be used
to improve algorithms for finding minimum spanning trees, for example Kruskal’s algorithm
(see O’Rourke [44]).
We conclude by mentioning that Voronoi diagrams have applications to motion planning.
For example, consider the problem of moving a disk on a plane while avoiding a set of
polygonal obstacles. If we “extend” the obstacles by the diameter of the disk, the problem
reduces to finding a collision–free path between two points in the extended obstacle space.
One needs to generalize the notion of a Voronoi diagram. Indeed, we need to define the
distance to an object, and medial curves (consisting of points equidistant to two objects)
may no longer be straight lines. A collision–free path with maximal clearance from the
obstacles can be found by moving along the edges of the generalized Voronoi diagram. This
is an active area of research in robotics. For more on this topic, see O’Rourke [44].
Acknowledgement. I wish to thank Marcelo Siqueira for suggesting many improvements

and for catching many bugs with his “eagle eye.”
Bibliography
[1] P.S. Alexandrov. Combinatorial Topology. Dover, first edition, 1998. Three volumes
bound as one.
[2] Noga Alon and Gil Kalai. A simple proof of the upper-bound theorem. European J.
Comb., 6:211–214, 1985.
[3] Michael Artin. Algebra. Prentice Hall, first edition, 1991.
[4] Alexander Barvinok. A Course in Convexity. GSM, Vol. 54. AMS, first edition, 2002.
[5] Margaret M. Bayer and Carl W. Lee. Combinatorial aspects of convex polytopes. In
P.M. Gruber and J.M. Wills, editors, Handbook of Convex Geometry, pages 485–534.
Elsevier Science, 1993.
[6] M. Berg, M. Van Kreveld, M. Overmars, and O. Schwarzkopf. Computational Geometry.

Algorithms and Applications. Springer, first edition, 1997.
[7] Marcel Berger. Géométrie 1. Nathan, 1990. English edition: Geometry 1, Universitext,
Springer Verlag.
[8] Marcel Berger. Géométrie 2. Nathan, 1990. English edition: Geometry 2, Universitext,
Springer Verlag.
[9] Dimitri P. Bertsekas. Convex Optimization Theory. Athena Scientific, first edition,
2009.
[10] Dimitris Bertsimas and John N. Tsitsiklis. Introduction to Linear Optimization. Athena
Scientific, third edition, 1997.
[11] J. Billera, Louis and Anders Björner. Faces numbers of polytopes and complexes. In
J.E. Goodman and Joe O’Rourke, editors, Handbook of Discrete and Computational
Geometry, pages 291–310. CRC Press, 1997.
[12] J.-D. Boissonnat and M. Yvinec. Géométrie Algorithmique. Ediscience International,

first edition, 1995.
387
388 BIBLIOGRAPHY
[13] Nicolas Bourbaki. Espaces Vectoriels Topologiques. Eléments de Mathématiques. Mas-

son, 1981.
[14] Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University
Press, first edition, 2004.
[15] David A. Brannan, Matthew F. Esplen, and Jeremy J. Gray. Geometry. Cambridge
University Press, first edition, 1999.
[16] K.Q. Brown. Voronoi diagrams from convex hulls. Inform. Process. Lett., 9:223–228,
1979.
[17] Heinz Bruggesser and Peter Mani. Shellable decompositions of cells and spheres. Math.
Scand., 29:197–205, 1971.
[18] Vasek Chvatal. Linear Programming. W.H. Freeman, first edition, 1983.
[19] P.G. Ciarlet. Introduction to Numerical Matrix Analysis and Optimization. Cambridge
University Press, first edition, 1989. French edition: Masson, 1994.
[20] H.S.M. Coxeter. Regular Polytopes. Dover, third edition, 1973.
[21] H.S.M. Coxeter. Introduction to Geometry. Wiley, second edition, 1989.
[22] Peter Cromwell. Polyhedra. Cambridge University Press, first edition, 1994.
[23] G.L. Dirichlet. Über die reduktion der positiven quadratischen formen mid drei unbes-
timmten ganzen zahlen. Journal für die reine und angewandte Mathematik, 40:209–227,
1850.
[24] H. Edelsbrunner and R. Seidel. Voronoi diagrams and arrangements. Discrete Compu-
tational Geometry, 1:25–44, 1986.
[25] Herbert Edelsbrunner. Geometry and Topology for Mesh Generation. Cambridge Uni-
versity Press, first edition, 2001.
[26] Günter Ewald. Combinatorial Convexity and Algebraic Geometry. GTM No. 168.
Springer Verlag, first edition, 1996.
[27] Jean Fresnel. Méthodes Modernes En Géométrie. Hermann, first edition, 1998.
[28] William Fulton. Introduction to Toric Varieties. Annals of Mathematical Studies, No.
131. Princeton University Press, 1997.
[29] Jean H. Gallier. Curves and Surfaces In Geometric Modeling: Theory And Algorithms.
Morgan Kaufmann, 1999.
BIBLIOGRAPHY 389
[30] Jean H. Gallier. Geometric Methods and Applications, For Computer Science and En-
gineering. TAM, Vol. 38. Springer, second edition, 2011.
[31] E.N. Gilbert. Random subdivisions of space into crystals. Annals of Math. Stat.,
33:958–972, 1962.
[32] Jacob E. Goodman and Joseph O’Rourke. Handbook of Discrete and Computational
Geometry. CRC Press, second edition, 2004.
[33] R. Graham and F. Yao. A whirlwind tour of computational geometry. American Math-
ematical Monthly, 97(8):687–701, 1990.
[34] Donald T. Greenwood. Principles of Dynamics. Prentice Hall, second edition, 1988.
[35] Branko Grünbaum. Convex Polytopes. GTM No. 221. Springer Verlag, second edition,
2003.
[36] D. Hilbert and S. Cohn-Vossen. Geometry and the Imagination. Chelsea Publishing
Co., 1952.
[37] Serge Lang. Algebra. Addison Wesley, third edition, 1993.
[38] Serge Lang. Real and Functional Analysis. GTM 142. Springer Verlag, third edition,
1996.
[39] Peter Lax. Functional Analysis. Wiley, first edition, 2002.
[40] Jiri Matousek. Lectures on Discrete Geometry. GTM No. 212. Springer Verlag, first
edition, 2002.
[41] Jiri Matousek and Bernd Gartner. Understanding and Using Linear Programming.
Universitext. Springer Verlag, first edition, 2007.
[42] Peter McMullen. The maximum number of faces of a convex polytope. Mathematika,
17:179–184, 1970.
[43] James R. Munkres. Elements of Algebraic Topology. Addison-Wesley, first edition, 1984.
[44] Joseph O’Rourke. Computational Geometry in C. Cambridge University Press, second

edition, 1998.
[45] Christos H. Papadimitriou and Kenneth Steiglitz. Combinatorial Optimization. Algo-

rithms and Complexity. Dover, first edition, 1998.
[46] Dan Pedoe. Geometry, A comprehensive Course. Dover, first edition, 1988.
[47] F.P. Preparata and M.I. Shamos. Computational Geometry: An Introduction. Springer
Verlag, first edition, 1988.
390 BIBLIOGRAPHY
[48] J.-J. Risler. Mathematical Methods for CAD. Masson, first edition, 1992.
[49] R. Tyrrell Rockafellar. Convex Analysis. Princeton Landmarks in Mathematics. Prince-

ton University Press, 1970.
[50] Pierre Samuel. Projective Geometry. Undergraduate Texts in Mathematics. Springer

Verlag, first edition, 1988.
[51] Alexander Schrijver. Theory of Linear and Integer Programming. Wiley, first edition,
1999.
[52] Raimund Seidel. The upper-bound theorem for polytopes: an easy proof of its asymp-
totic version. Comput. Geometry: Theory and Applications, 5:115–116, 1995.
[53] Ernst Snapper and Troyer Robert J. Metric Affine Geometry. Dover, first edition, 1989.
[54] John Stallings. Lectures on Polyhedral Topology. Tata Institute, first edition, 1967.
[55] Richard P. Stanley. The number of faces of simplicial polytopes and spheres. In J.E
Goodman, E. Lutwak, J. Malkevitch, and P. Pollack, editors, Discrete Geometry and
Convexity, pages 212–223. Annals New York Academy of Sciences, 1985.
[56] Richard P. Stanley. Combinatorics and Commutative Algebra. Progress in Mathematics,

No. 41. Birkhäuser, second edition, 1996.
[57] J. Stolfi. Oriented Projective Geometry. Academic Press, first edition, 1991.
[58] Gilbert Strang. Linear Algebra and its Applications. Saunders HBJ, third edition, 1988.
[59] Bernd Sturmfels. Gröbner Bases and Convex Polytopes. ULS, Vol. 8. AMS, first edition,
1996.
[60] Rekha R. Thomas. Lectures in Geometric Combinatorics. STML, Vol. 33. AMS, first
edition, 2006.
[61] P. Thurston, William. Three-Dimensional Geometry and Topology, Vol. 1. Princeton

University Press, first edition, 1997. Edited by Silvio Levy.
[62] Claude Tisseron. Géométries affines, projectives, et euclidiennes. Hermann, first edition,
1994.
[63] Frederick A. Valentine. Convex Sets. McGraw–Hill, first edition, 1964.
[64] Robert J. Vanderbei. Linear Programming: Foundations and Extensions. Springer,

fourth edition, 2014.
[65] Lucas Vienne. Présentation algébrique de la géométrie classique. Vuibert, first edition,
1996.
BIBLIOGRAPHY 391
[66] M.G. Voronoi. Nouvelles applications des paramètres continus à la théorie des formes
quadratiques. J. Reine u. Agnew. Math., 134:198–287, 1908.
[67] Gunter Ziegler. Lectures on Polytopes. GTM No. 152. Springer Verlag, first edition,
1997.

Comb Topol

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Comb Topol

Uploaded by

Document Information

Original Description:

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

Comb Topol

Uploaded by

Copyright:

Available Formats

Aspects of Convex Geometry

Polyhedra, Linear Programming,

Jean Gallier and Jocelyn Quaintance

Key-words: Convex sets, polytopes, polyhedra, linear programming, simplex algorithm,

2 Basics of Affine Geometry 11

3 Basic Properties of Convex Sets 57

4 Two Main Tools: Separation and Polar Duality 79

5 Polyhedra and Polytopes 103

5.4 The Equivalence of H-Polyhedra and V-Polyhedra . . . . . . . . . . . . . . . 114

6 Linear Programs 143

7 The Simplex Algorithm 161

8 Linear Programming and Duality 193

9 Basics of Combinatorial Topology 227

10 Shellings and the Euler–Poincaré Formula 257

11 Projective Spaces and Polyhedra, Polar Duality 285

12 Dirichlet–Voronoi Diagrams 321

1.1 Motivations and Goals

Basics of Affine Geometry

2.1 Affine Spaces

and a vector b ∈ Rm , the set U = {x ∈ Rn | Ax = b} of solutions of the system Ax = b is an

Use coordinate systems only when needed!

Figure 2.1: Points and free vectors.

Figure 2.2: The two position vectors for the point x.

(λu1 + µv1 , λu2 + µv2 , λu3 + µv3 ).

(λu01 + µv10 , λu02 + µv20 , λu03 + µv30 )

of λu + µv with respect to the basis (e01 , e02 , e03 ) are given by

(a01 , a02 , a03 ) = (a1 − ω1 , a2 − ω2 , a3 − ω3 )

(λa1 + µb1 , λa2 + µb2 , λa3 + µb3 ),

but the coordinates

(λa1 + µb1 − (λ + µ)ω1 , λa2 + µb2 − (λ + µ)ω2 , λa3 + µb3 − (λ + µ)ω3 ),

which are different from

(λa1 + µb1 − ω1 , λa2 + µb2 − ω2 , λa3 + µb3 − ω3 ),

unless λ + µ = 1. See Figure 2.3.

a + b = (3,4,2) = (0, 0, -3)

a + b = (-3, -4, -8) = (0, 0, -3)

Figure 2.4: Intuitive picture of an affine space.

where b ∈ E. The composition of the first mapping with the second is

Figure 2.5: An affine space: the line of equation x + y − 1 = 0

2.2 Examples of Affine Spaces

Figure 2.6: An affine space: the plane x + y + z − 1 = 0.

2.3 Chasles’s Identity

Figure 2.7: The paraboloid of revolution P viewed as a two-dimensional affine space.

by (A2), and thus, by (A3),

Thus, letting a = c in Chasles’s identity, we get

Given any four points a, b, c, d ∈ E, since by Chasles’s identity

we have the parallelogram law

2.4 Affine Combinations, Barycenters

Figure 2.8: Points and corresponding vectors in affine geometry.

(λa1 + µb1 , λa2 + µb2 ).

a = (-1,-1) a = (-2, -2)

Figure 2.9: The example from the beginning of Section 2.4.

Proof. (1) By Chasles’s identity (see Section 2.3), we have

An illustration of this calculation in A2 is provided by Figure 2.10.

Figure 2.10: Part (1) of Proposition 2.1.

(which is independent of a ∈ E, by Proposition 2.1) is called the barycenter (or barycentric

In dealing with barycenters, it is convenient to introduce the notion of a weighted point,

and setting a = x, the point x is the unique point such that

(1) Since the barycenter of a family ((ai , λi ))i∈I of weighted