Mathematics

MONOGRAPHS AND RESEARCH NOTES IN MATHEMATICS MONOGRAPHS AND RESEARCH NOTES IN MATHEMATICS

Monomial
Monomial Algebras, Second Edition presents algebraic, combina-

Monomial Algebras
Second Edition
torial, and computational methods for studying monomial algebras
and their ideals, including Stanley–Reisner rings, monomial subrings,
Ehrhart rings, and blowup algebras. It emphasizes square-free mono-

Algebras
mials and the corresponding graphs, clutters, or hypergraphs.

New to the Second Edition

• Four new chapters that focus on the algebraic properties of
blowup algebras in combinatorial optimization problems of clut-
ters and hypergraphs
• Two new chapters that explore the algebraic and combinatorial
properties of the edge ideal of clutters and hypergraphs
• Full revisions of existing chapters to provide an up-to-date ac-
Second Edition
count of the subject

Bringing together several areas of pure and applied mathematics, this

book shows how monomial algebras are related to polyhedral geom-
etry, combinatorial optimization, and combinatorics of hypergraphs. It
directly links the algebraic properties of monomial algebras to com-
binatorial structures (such as simplicial complexes, posets, digraphs,
graphs, and clutters) and linear optimization problems.

Features
• Presents computational and combinatorial methods in commuta-
tive algebra
• Shows how to solve a variety of problems of monomial algebras
• Covers various affine and graded rings, including Cohen–Macau-
lay, complete intersection, and normal
• Examines their basic algebraic invariants, such as multiplicity,
Betti numbers, projective dimension, and Hilbert polynomial Villarreal
• Contains more than 550 exercises and over 50 examples, many of
which illustrate the use of computer algebra systems
Rafael H. Villarreal
K23008

w w w. c rc p r e s s . c o m

K23008_cover.indd 1 2/13/15 12:23 PM

MONOGRAPHS AND RESEARCH NOTES IN MATHEMATICS

Monomial
Algebras
Second Edition

Rafael H. Villarreal
Centro de Investigación
y de Estudios Avanzados del IPN
Mexico City, Mexico
 MONOGRAPHS AND RESEARCH NOTES IN MATHEMATICS

Series Editors
John A. Burns
Thomas J. Tucker
Miklos Bona
Michael Ruzhansky
Chi-Kwong Li

Published Titles
Application of Fuzzy Logic to Social Choice Theory, John N. Mordeson, Davender S. Malik
and Terry D. Clark
Blow-up Patterns for Higher-Order: Nonlinear Parabolic, Hyperbolic Dispersion and
Schrödinger Equations, Victor A. Galaktionov, Enzo L. Mitidieri, and Stanislav Pohozaev
Difference Equations: Theory, Applications and Advanced Topics, Third Edition,
Ronald E. Mickens
Dictionary of Inequalities, Second Edition, Peter Bullen
Iterative Optimization in Inverse Problems, Charles L. Byrne
Modeling and Inverse Problems in the Presence of Uncertainty, H. T. Banks, Shuhua Hu,
and W. Clayton Thompson
Monomial Algebras, Second Edition, Rafael H. Villarreal
Set Theoretical Aspects of Real Analysis, Alexander B. Kharazishvili
Signal Processing: A Mathematical Approach, Second Edition, Charles L. Byrne
Sinusoids: Theory and Technological Applications, Prem K. Kythe
Special Integrals of Gradshetyn and Ryzhik: the Proofs – Volume l, Victor H. Moll

Forthcoming Titles
Actions and Invariants of Algebraic Groups, Second Edition, Walter Ferrer Santos
and Alvaro Rittatore
Analytical Methods for Kolmogorov Equations, Second Edition, Luca Lorenzi
Complex Analysis: Conformal Inequalities and the Bierbach Conjecture, Prem K. Kythe
Computational Aspects of Polynomial Identities: Volume l, Kemer’s Theorems, 2nd Edition
Belov Alexey, Yaakov Karasik, Louis Halle Rowen
Cremona Groups and Icosahedron, Ivan Cheltsov and Constantin Shramov
Geometric Modeling and Mesh Generation from Scanned Images, Yongjie Zhang
Groups, Designs, and Linear Algebra, Donald L. Kreher
Handbook of the Tutte Polynomial, Joanna Anthony Ellis-Monaghan and Iain Moffat
Lineability: The Search for Linearity in Mathematics, Juan B. Seoane Sepulveda,
Richard W. Aron, Luis Bernal-Gonzalez, and Daniel M. Pellegrinao
Line Integral Methods and Their Applications, Luigi Brugnano and Felice Iaverno
Microlocal Analysis on Rˆn and on NonCompact Manifolds, Sandro Coriasco
Forthcoming Titles (continued)
Nonlinear Functional Analysis in Banach Spaces and Banach Algebras: Fixed Point Theory
Under Weak Topology for Nonlinear Operators and Block Operators with Applications
Aref Jeribi and Bilel Krichen
Partial Differential Equations with Variable Exponents: Variational Methods and
Quantitative Analysis, Vicentiu Radulescu and Dusan Repovs
Practical Guide to Geometric Regulation for Distributed Parameter Systems,
Eugenio Aulisa and David S. Gilliam
Reconstructions from the Data of Integrals, Victor Palamodov
Special Integrals of Gradshetyn and Ryzhik: the Proofs – Volume ll, Victor H. Moll
Stochastic Cauchy Problems in Infinite Dimensions: Generalized and Regularized
Solutions, Irina V. Melnikova and Alexei Filinkov
Symmetry and Quantum Mechanics, Scott Corry
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2015 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Version Date: 20150202

International Standard Book Number-13: 978-1-4822-3470-1 (eBook - PDF)

This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copy-
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that pro-
vides licenses and registration for a variety of users. For organizations that have been granted a photo-
copy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Contents

Preface ix

Preface to First Edition xv

1 Polyhedral Geometry and Linear Optimization 1

1.1 Polyhedral sets and cones . . . . . . . . . . . . . . . . . . . . 1
1.2 Relative volumes of lattice polytopes . . . . . . . . . . . . . . 20
1.3 Hilbert bases and TDI systems . . . . . . . . . . . . . . . . . 28
1.4 Rees cones and clutters . . . . . . . . . . . . . . . . . . . . . 39
1.5 The integral closure of a semigroup . . . . . . . . . . . . . . . 46
1.6 Unimodularity of matrices and normality . . . . . . . . . . . 48
1.7 Normaliz, a computer program . . . . . . . . . . . . . . . . . 50
1.8 Cut-incidence matrices and integrality . . . . . . . . . . . . . 51
1.9 Elementary vectors and matroids . . . . . . . . . . . . . . . . 55

2 Commutative Algebra 61
2.1 Module theory . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.2 Graded modules and Hilbert polynomials . . . . . . . . . . . 76
2.3 Cohen–Macaulay modules . . . . . . . . . . . . . . . . . . . . 79
2.4 Normal rings . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.5 Valuation rings . . . . . . . . . . . . . . . . . . . . . . . . . . 92
2.6 Krull rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.7 Koszul homology . . . . . . . . . . . . . . . . . . . . . . . . . 104
2.8 A vanishing theorem of Grothendieck . . . . . . . . . . . . . . 107

3 Affine and Graded Algebras 111

3.1 Cohen–Macaulay graded algebras . . . . . . . . . . . . . . . . 111
3.2 Hilbert Nullstellensatz . . . . . . . . . . . . . . . . . . . . . . 123
3.3 Gröbner bases . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.4 Projective closure . . . . . . . . . . . . . . . . . . . . . . . . . 132
3.5 Minimal resolutions . . . . . . . . . . . . . . . . . . . . . . . 134
vi CONTENTS

4 Rees Algebras and Normality 141

4.1 Symmetric algebras . . . . . . . . . . . . . . . . . . . . . . . . 141
4.2 Rees algebras and syzygetic ideals . . . . . . . . . . . . . . . 142
4.3 Complete and normal ideals . . . . . . . . . . . . . . . . . . . 145
4.4 Multiplicities and a criterion of Herzog . . . . . . . . . . . . . 159
4.5 Jacobian criterion . . . . . . . . . . . . . . . . . . . . . . . . . 165

5 Hilbert Series 171

5.1 Hilbert–Serre Theorem . . . . . . . . . . . . . . . . . . . . . . 171
5.2 a-invariants and h-vectors . . . . . . . . . . . . . . . . . . . . 177
5.3 Extremal algebras . . . . . . . . . . . . . . . . . . . . . . . . 182
5.4 Initial degrees of Gorenstein ideals . . . . . . . . . . . . . . . 189
5.5 Koszul homology and Hilbert functions . . . . . . . . . . . . . 196
5.6 Hilbert functions of some graded ideals . . . . . . . . . . . . . 199

6 Stanley–Reisner Rings and Edge Ideals of Clutters 201

6.1 Primary decomposition . . . . . . . . . . . . . . . . . . . . . . 201
6.2 Simplicial complexes and homology . . . . . . . . . . . . . . . 209
6.3 Stanley–Reisner rings . . . . . . . . . . . . . . . . . . . . . . 212
6.4 Regularity and projective dimension . . . . . . . . . . . . . . 226
6.5 Unmixed and shellable clutters . . . . . . . . . . . . . . . . . 234
6.6 Admissible clutters . . . . . . . . . . . . . . . . . . . . . . . . 246
6.7 Hilbert series of face rings . . . . . . . . . . . . . . . . . . . . 250
6.8 Simplicial spheres . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.9 The upper bound conjectures . . . . . . . . . . . . . . . . . . 258

7 Edge Ideals of Graphs 261

7.1 Graph theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
7.2 Edge ideals and B-graphs . . . . . . . . . . . . . . . . . . . . 268
7.3 Cohen–Macaulay and chordal graphs . . . . . . . . . . . . . . 274
7.4 Shellable and sequentially C–M graphs . . . . . . . . . . . . . 282
7.5 Regularity, depth, arithmetic degree . . . . . . . . . . . . . . 293
7.6 Betti numbers of edge ideals . . . . . . . . . . . . . . . . . . . 298
7.7 Associated primes of powers of ideals . . . . . . . . . . . . . . 303

8 Toric Ideals and Affine Varieties 311

8.1 Binomial ideals and their radicals . . . . . . . . . . . . . . . . 311
8.2 Lattice ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
8.3 Monomial subrings and toric ideals . . . . . . . . . . . . . . . 326
8.4 Toric varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
8.5 Affine Hilbert functions . . . . . . . . . . . . . . . . . . . . . 342
8.6 Vanishing ideals over finite fields . . . . . . . . . . . . . . . . 345
8.7 Semigroup rings of numerical semigroups . . . . . . . . . . . . 347
8.8 Toric ideals of monomial curves . . . . . . . . . . . . . . . . . 352
CONTENTS vii

9 Monomial Subrings 365

9.1 Integral closure of monomial subrings . . . . . . . . . . . . . 366
9.2 Homogeneous monomial subrings . . . . . . . . . . . . . . . . 372
9.3 Ehrhart rings . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
9.4 The degree of lattice and toric ideals . . . . . . . . . . . . . . 392
9.5 Laplacian matrices and ideals . . . . . . . . . . . . . . . . . . 396
9.6 Gröbner bases and normal subrings . . . . . . . . . . . . . . . 403
9.7 Toric ideals generated by circuits . . . . . . . . . . . . . . . . 410
9.8 Divisor class groups of semigroup rings . . . . . . . . . . . . . 416

10 Monomial Subrings of Graphs 423

10.1 Edge subrings and ring graphs . . . . . . . . . . . . . . . . . 423
10.2 Incidence matrices and circuits . . . . . . . . . . . . . . . . . 440
10.3 The integral closure of an edge subring . . . . . . . . . . . . . 448
10.4 Ehrhart rings of edge polytopes . . . . . . . . . . . . . . . . . 454
10.5 Integral closure of Rees algebras . . . . . . . . . . . . . . . . 457
10.6 Edge subrings of complete graphs . . . . . . . . . . . . . . . . 461
10.7 Edge cones of graphs . . . . . . . . . . . . . . . . . . . . . . . 467
10.8 Monomial birational extensions . . . . . . . . . . . . . . . . . 477

11 Edge Subrings and Combinatorial Optimization 483

11.1 The canonical module of an edge subring . . . . . . . . . . . 483
11.2 Integrality of the shift polyhedron . . . . . . . . . . . . . . . 484
11.3 Generators for the canonical module . . . . . . . . . . . . . . 487
11.4 Computing the a-invariant . . . . . . . . . . . . . . . . . . . . 489
11.5 Algebraic invariants of edge subrings . . . . . . . . . . . . . . 493

12 Normality of Rees Algebras of Monomial Ideals 499

12.1 Integral closure of monomial ideals . . . . . . . . . . . . . . . 499
12.2 Normality criteria . . . . . . . . . . . . . . . . . . . . . . . . 505
12.3 Rees cones and polymatroidal ideals . . . . . . . . . . . . . . 508
12.4 Veronese subrings and the a-invariant . . . . . . . . . . . . . 513
12.5 Normalizations of Rees algebras . . . . . . . . . . . . . . . . . 519
12.6 Rees algebras of Veronese ideals . . . . . . . . . . . . . . . . . 524
12.7 Divisor class group of a Rees algebra . . . . . . . . . . . . . . 529
12.8 Stochastic matrices and Cremona maps . . . . . . . . . . . . 531

13 Combinatorics of Symbolic Rees Algebras

of Edge Ideals of Clutters 537
13.1 Vertex covers of clutters . . . . . . . . . . . . . . . . . . . . . 537
13.2 Symbolic Rees algebras of edge ideals . . . . . . . . . . . . . 540
13.3 Blowup algebras in perfect graphs . . . . . . . . . . . . . . . 551
13.4 Algebras of vertex covers of graphs . . . . . . . . . . . . . . . 555
13.5 Edge subrings in perfect matchings . . . . . . . . . . . . . . . 558
viii CONTENTS

13.6 Rees cones and perfect graphs . . . . . . . . . . . . . . . . . . 561

13.7 Perfect graphs and algebras of covers . . . . . . . . . . . . . . 564

14 Combinatorial Optimization and Blowup Algebras 567

14.1 Blowup algebras of edge ideals . . . . . . . . . . . . . . . . . 568
14.2 Rees algebras and polyhedral geometry . . . . . . . . . . . . . 570
14.3 Packing problems and blowup algebras . . . . . . . . . . . . . 583
14.4 Uniform ideal clutters . . . . . . . . . . . . . . . . . . . . . . 599
14.5 Clique clutters of comparability graphs . . . . . . . . . . . . . 610
14.6 Duality and integer rounding problems . . . . . . . . . . . . . 615
14.7 Canonical modules and integer rounding . . . . . . . . . . . . 628
14.8 Clique clutters of Meyniel graphs . . . . . . . . . . . . . . . . 633

Appendix Graph Diagrams 635

A.1 Cohen–Macaulay graphs . . . . . . . . . . . . . . . . . . . . . 635
A.2 Unmixed graphs . . . . . . . . . . . . . . . . . . . . . . . . . 638

Bibliography 639

Notation Index 669

Index 673
Preface

The main purpose of this book is to introduce algebraic, combinatorial, and

computational methods to study monomial algebras and their presentation
ideals, including Stanley–Reisner rings, monomial subrings, Ehrhart rings,
blowup algebras, emphasizing square-free monomials and its corresponding
graphs, clutters, or hypergraphs.
Monomial algebras are related to various fields, for instance to numerical
semigroups, semigroup rings, algebraic geometry, commutative algebra and
combinatorics, integer programming and polyhedral geometry, graph theory
and combinatorial optimization. We develop links between the areas shown
as the vertices of the following graph.

 
Polyhedral
Geometry
 






A
A
AA


AA


 AA


AA


Monomial AA


Algebras AA


  AA
, l


, 
l AA 
Combinatorial Combinatorics
Optimization of Hypergraphs
   

This allows us to solve a variety of problems of monomial algebras using

the methods of the other areas and vice versa. An effort has been made to
give a unifying presentation of the techniques and notions of these areas. In
this book we are interested in the algebraic properties of monomial algebras
that can be directly linked to combinatorial structures—such as simplicial
x Preface

complexes, posets, digraphs, graphs, and clutters—and to linear optimiza-

tion problems. We study various types of affine and graded rings (Cohen–
Macaulay, sequentially Cohen–Macaulay, Gorenstein, Artinian, complete
intersection, unmixed, normal, reduced) and examine their basic algebraic
invariants (type, multiplicity, a-invariant, regularity, Betti numbers, Krull
dimension, projective dimension, h-vector, Hilbert polynomial).
In recent years the algebraic properties of blowup algebras have been
linked to combinatorial optimization problems of clutters and hypergraphs.
In this edition, we include four new chapters (Chapters 1, 11, 13, and 14) to
introduce this area of research and to present some of the main advances.
The study of algebraic and combinatorial properties of the edge ideal of
clutters and hypergraphs has attracted a great deal of interest in the last
two decades [163, 189, 224, 326, 408]. In the present edition we include two
chapters with some of the advances in the area (Chapters 6 and 7).
In order to present an up-to-date account of the subject, we have made
a full revision of all chapters in the first edition. The chapters have been
reorganized and arranged in a different order. In particular Chapters 9 and
12 were originally a single chapter, while Chapters 8 and 10 were originally
divided into two chapters for each of them.
This book brings together several areas of pure and applied mathemat-
ics. It contains over 550 exercises and over 50 examples, many of them
illustrating the use of computer algebra systems. It has extensive indices of
terminology and notation.
The contents of this book are as follows. In Chapter 1 we begin by
introducing several notions and results coming from polyhedral geometry,
combinatorial optimization, and linear programming. Relative volumes of
lattice polytope are studied here. We present relations between Hilbert
bases, TDI linear systems of inequalities, max-flow min-cut properties of
clutters, and normality of affine and Rees semigroups. For instance the
max-flow min-cut property is classified in terms of Hilbert bases and the
integrality of certain polyhedra. This is used in Chapter 14 to prove that
the edge ideal of a clutter is normally torsion-free if and only if the clutter
has the max-flow min-cut property. We present a slight generalization of a
theorem of Lucchesi and Younger [298] that is useful to detect TDI systems
arising from incidence matrices of digraphs. This is used in Chapter 11 to
express the a-invariant of the edge subring of a bipartite graph in terms of
directed cuts. Elementary integral vectors and matroids are introduced at
the end of the chapter. The notion of a matroid will appear in several places
in this book in connection with monomial rings and ideals. The computer
program we have used to study linear systems of inequalities and to compute
Hilbert bases of rational cones is Normaliz [68]. The prerequisite for this
chapter is a course on linear algebra and familiarity with point set topology.
Chapter 2 discusses certain topics and results on commutative algebra
Preface xi

(module theory, normal and graded rings). It includes some detailed proofs,
and points the reader to the appropriate references when proofs are omitted.
However we make free use of the standard terminology and notation of
homological algebra (including Tor and Ext) as described in [363] and [428].
This edition has three new sections on valuation rings, Krull rings, and a
vanishing theorem of Grothendieck (Sections 2.5, 2.6, and 2.8).
A number of topics connected to affine and graded algebras are studied in
Chapter 3, e.g., Gröbner bases, Hilbert Nullstellensatz and affine varieties,
projective closure, minimal resolutions, and Betti numbers. We present two
versions of the Noether normalization lemma and some of its applications
to affine algebras and to Cohen–Macaulay graded algebras.
In Chapter 4 a thorough presentation of complete and normal ideals
is given. Here the systematic use of blowup algebras makes clear their
importance for the area. In this chapter we introduce symbolic powers
of ideals and normally torsion-free ideals. Then we present a few cases
where equality of symbolic and ordinary powers can be described in terms
of properties of the associated graded ring. An elegant and useful Cohen–
Macaulay criterion due to Herzog is included here.
Chapter 5 deals with the Hilbert series of graded modules and algebras,
a topic that is quite useful in Stanley’s proof of the upper bound conjecture
for simplicial spheres. The h-vector and a-invariant of a graded algebra are
defined through the Hilbert–Serre theorem. For Cohen–Macaulay graded
algebras we present the main properties of their h-vectors and a-invariants.
Some optimal upper bounds for the number of generators in the least degree
of Gorenstein and Cohen–Macaulay graded ideals are given, which naturally
leads to the notion of an extremal algebra. As an application the Koszul
homology of Cohen–Macaulay ideals with pure resolutions is studied using
Hilbert function techniques. This edition has a new section on Hilbert
functions of a certain type of graded ideals that occur in algebraic coding
theory (Section 5.6).
Chapter 6 is an introduction to monomial ideals, Stanley–Reisner rings,
and edge ideals of clutters. An understated goal here is to highlight some of
the works of T. Hibi, J. Herzog, M. Hochster, G. Reisner, and R. Stanley. We
study algebraic and combinatorial properties of edge ideals and simplicial
complexes. In particular we examine shellable, unmixed, and sequentially
Cohen–Macaulay simplicial complexes and their corresponding edge ideals
and invariants (projective dimension, regularity, depth, Hilbert series). As
for applications, the proofs of the upper bound conjectures for polytopes
are discussed to give a flavor to some of the methods and ideas of the
area. Monomial ideals can also be used to solve certain problems of general
polynomial ideals using the theory of Gröbner bases.
In Chapter 7 we give an introduction to graph theory and study algebraic
properties and invariants of edge ideals using the combinatorial structure of
xii Preface

graphs. The study of edge ideals of graphs has been a very active area of
research in the last decade because of its connections to graph theory. We
present classifications of the following families: unmixed bipartite graphs,
Cohen–Macaulay bipartite graphs, Cohen–Macaulay trees, shellable bipar-
tite graphs, sequentially Cohen–Macaulay bipartite graphs. The invariants
examined here include regularity, depth, Krull dimension, and multiplicity.
Edge ideals of graphs are shown to have the persistence property.
In Chapter 8 we study algebraic and geometric aspects of three special
types of polynomial ideals and their quotient rings:

toric ideals → lattice ideals → binomial ideals.

These ideals are interesting from a computational point of view and are
related to diverse fields, such as combinatorics, algebraic geometry, integer
programming, semigroup rings, coding theory, and algebraic statistics.
Chapter 9 deals with point configurations and their lattice polytopes.
We consider monomial subrings and binomial ideals associated with them,
e.g., Ehrhart rings, Rees algebras, homogeneous subrings, lattice ideals, and
toric ideals. Using Hilbert functions, polyhedral geometry, and Gröbner
bases, we study normalizations of monomial subrings, initial ideals of toric
ideals, normal monomial subrings, primary decompositions and multiplici-
ties of lattice ideals. Algebraic graph theory is used to study matrix ideals
of Laplacian matrices. The reciprocity law of Ehrhart for integral poly-
topes and the Danilov–Stanley formula for canonical modules of monomial
subrings will be introduced here. Applications of these results will be given.
In Chapter 10 we study monomial subrings associated with graphs and
their toric ideals. We relate the even closed walks and circuits of the vector
matroid of a graph with Gröbner bases theory. A description of the integral
closure of the edge subring of a multigraph will be presented along with a
description of the circuits of its toric ideal. We study the family of graphs
whose number of primitive cycles equals its cycle rank. It is shown that this
family is precisely the family of ring graphs. These graphs are characterized
in algebraic and combinatorial terms. We classify edge subrings of bipartite
graphs which are complete intersections. Then we present sharp upper
bounds for the multiplicity of edge subrings. Several connections between
monomial subrings, graph theory, and polyhedral geometry will occur in
this chapter. We study in detail the irreducible representation of an edge
cone of a graph and show some applications to graph theory (for instance
we show the marriage theorem).
Chapter 11 focuses on edge subrings of connected bipartite graphs and
their algebraic invariants. We show how to compute the canonical module
and the a-invariant of an edge subring using linear programming techniques
and introduce an integral polyhedron whose vertices correspond to minimal
generators of the canonical module. The a-invariant of an edge subring will
Preface xiii

be interpreted in combinatorial optimization terms as the maximum number

of edge disjoint directed cuts. We study the Gorenstein property and the
type of an edge subring.
Chapter 12 is about Rees algebras of monomial ideals. We study the
integral closure of a monomial ideal, and the normality and invariants of
its Rees algebra. The normalization of a Rees algebra is examined using
the Danilov–Stanley formula, Carathéodory’s theorem, and Hilbert bases of
Rees cones. Interesting classes of normal ideals, such as ideals of Veronese
type and polymatroidal ideals are introduced in the chapter. The divisor
class group of a normal Rees algebra is computed using polyhedral geometry.
Chapter 13 shows the interaction between graph theory, combinatorial
optimization, commutative algebra, and the theory of blowup algebras.
In this chapter, we give a description—using notions from combinatorial
optimization—of the minimal generators of the symbolic Rees algebra of
the edge ideal of a clutter and show a graph theoretical description of the
minimal generators of the symbolic Rees algebra of the ideal of covers of a
graph. The minimal sets of generators of symbolic Rees algebras of edge
ideals are studied using polyhedral geometry. Indecomposable graphs are
related to the strong perfect graph theorem. We give a description—in
terms of cliques—of the symbolic Rees algebra and the Simis cone of the
edge ideal of a perfect graph.
In Chapter 14 we relate combinatorial optimization with commutative
algebra, and present applications to both areas. We establish some links
between the algebraic properties of blowup algebras of edge ideals and the
combinatorial optimization properties of clutters and polyhedra. A long-
standing conjecture of Conforti and Cornuéjols about packing problems is
examined from an algebraic point of view. We study max-flow min-cut
problems of clutters, packing problems, and integer rounding properties
of various systems of linear inequalities to gain insight about the algebraic
properties of blowup algebras. Systems with integer rounding properties and
clutters with the max-flow min-cut property come from linear optimization
problems. The equality between the Rees algebra and the symbolic Rees
algebra of an edge ideal is characterized in combinatorial and algebraic
terms. A number of properties of clutters and edge ideals are shown to be
closed (under taking) minors, Alexander duals, and parallelizations. The
max-flow min-cut property of a clutter is characterized in algebraic and
combinatorial terms. The structure of ideal uniform clutters is presented
here. If a clutter satisfies the max-flow min-cut property, we prove that all
invariant factors of its incidence matrix are equal to 1 and that the columns
of this matrix form a Hilbert basis. It is shown that the clique clutter of a
comparability graph satisfies the max-flow min-cut property. The normality
of an ideal is described in terms of the integer rounding property of a linear
system and we establish a duality theorem for monomial subrings. We show
xiv Preface

that the Rees algebra of the ideal of covers of a perfect graph is normal and
that clique clutters of Meynel graphs are Ehrhart clutters.
This book stresses the use of computational and combinatorial methods
in commutative algebra because they have been a major factor in discovering
new and interesting results. The main computer algebra programs that we
use in this book are Normaliz [68] and Macaulay2 [199]. These programs
provide an invaluable tool to study monomial algebras and their algebraic
invariants. As a handy reference we include a section summarizing the type
of computations that can be done using Normaliz (see Section 1.7). We
also occasionally use other computer algebra systems, such as CoCoA [88],
TM
Maple [80], and Mathematica  R
[431]. Singular [200] is another system
that can be used for computations in commutative algebra with Gröbner
bases. Porta [84] and polymake [177] are two systems that can be used for
computations in convex polyhedra and finite simplicial complexes.
Combinatorial commutative algebra is an extensive area of mathematics.
The main references for the area are the books of Bruns and Herzog [65], Hibi
[240], Miller and Sturmfels [317], and Stanley [395]. This book emphasizes
the use of discrete mathematics and combinatorial optimization methods in
combinatorial commutative algebra. There are a number of excellent recent
books that offer a complementary view of the subject, namely, Beck and
Robins [21], Bruns and Gubelazde [61], De Loera, Hemmecke and Köppe
[105], Ene and Herzog [142], Herzog and Hibi [224], Huneke and Swanson
[259], and Vasconcelos [414].
Outstanding references for computational and combinatorial aspects that
complement and—in some cases—extend some of the material included here
are the books of Berge [25, 27], Brøndsted [57], Chvátal [87], Cornuéjols [93],
De Loera, Rambau and Santos [106], Diestel [111], Cox, Little and O’Shea
[99], Eisenbud [128], Ewald [151], Gitler and Villarreal [189], Godsil and
Royle [190], Golumbic [191], Greuel and Pfister [200], Harary [208], Kreuzer
and Robbiano [282], Schenck [369], Schrijver [372, 373], Stanley [394, 396],
Sturmfels [400], Vasconcelos [413], and Ziegler [438].
We would like to thank Winfried Bruns, Enrique Reyes, and Aron Simis
for their helpful comments and corrections. Thanks to executive editor
Robert B. Stern for suggesting we prepare an up-to-date second edition
of Monomial Algebras for the new series in mathematics Monographs and
Research Notes in Mathematics, Taylor & Francis (Chapman and Hall/CRC
Press Group). The support of Marsha Pronin and Samantha White at
Taylor & Francis is much appreciated. Finally, we are grateful to Karen
Simon for her careful editorial work on the manuscript.

Rafael H. Villarreal
Cinvestav-IPN
Mexico City, D.F.
Preface to First Edition

Let R = K[x] = K[x1 , . . . , xn ] be a polynomial ring in the indeterminates

x1 , . . . , xn , over the field K. Let

fi = xvi = xv1i1 · · · xvnin , i = 1, . . . , q,

be a finite set of monomials of R. We are interested in studying several

algebras and ideals associated with these monomials. Some of these are:

• the monomial subring: K[f1 , . . . , fq ] ⊂ K[x],

• the Rees algebra: K[x, f1 t, . . . , fq t] ⊂ K[x, t], which is also a mono-
mial subring,
• the face ring or Stanley–Reisner ring: K[x]/(f1 , . . . , fq ), if the mono-
mials are square-free, and
• the toric ideal: the ideal of relations of a monomial subring.

In the following diagram we stress the most relevant relations between

the properties of those algebras that will take place in this text.

Rees algebra
 @ I
@@
@@
R
@
Face ring Monomial subring


Toric ideal

If such monomials are square-free they are indexed by a hypergraph

built on the set of indeterminates, which provides a second combinatorial
structure in addition to the associated Stanley–Reisner simplicial complex.
xvi Preface to First Edition

This book was written with the aim of providing an introduction to the
methods that can be used to study monomial algebras and their presentation
ideals, with emphasis on square-free monomials. We have striven to provide
methods that are effective for computations.
A substantial part of this volume is dedicated to the case of monomial
algebras associated to graphs, that is, those defined by square-free quadratic
monomials defining a simple graph. We will systematically use graph theory
to study those algebras. Such a systematic treatment is a gap in the litera-
ture that we intend to fill. Two outstanding references for graph theory are
[48] and [208].
In the text, special attention is paid to providing means to determine
whether a given monomial algebra or ideal is Cohen–Macaulay or normal.
Those means include diverse characterizations and qualities of those two
properties.
Throughout this work base rings are assumed to be Noetherian and
modules finitely generated.
An effort has been made to make the book self-contained by including
a chapter on commutative algebra (Chapter 2) that includes some detailed
proofs and often points the reader to the appropriate references when proofs
are omitted. However, we make free use of the standard terminology and
notation of homological algebra (including Tor and Ext) as described in
[363] and [428].
The first goal is to present basic properties of monomial algebras. For
this purpose in Chapter 3 we study affine and graded algebras. The topics
include Noether normalizations and their applications, diverse attributes of
Cohen–Macaulay graded algebras, Hilbert Nullstellensatz and affine vari-
eties, some Gröbner bases theory, and minimal resolutions.
In Chapter 4 a thorough presentation of complete and normal ideals is
given. Here the systematic use of Rees algebras and associated graded rings
makes clear their importance for the area.
Chapter 5 deals with the Hilbert series of graded modules and algebras,
a topic that is quite useful in Stanley’s proof of the upper bound conjecture
for simplicial spheres. Here we introduce the h-vector and a-invariant of
graded algebras and give several interpretations of the a-invariant when the
algebra is Cohen–Macaulay. Some optimal upper bounds for the number of
generators in the least degree of Gorenstein and Cohen–Macaulay ideals are
presented, which naturally leads to the notion of an extremal algebra. As
an application the Koszul homology of Cohen–Macaulay ideals with pure
resolutions is studied using Hilbert function techniques.
General monomial ideals and Stanley–Reisner rings are examined in
Chapter 6. The first version of this chapter was some notes originally pre-
pared to teach a short course during the XXVII Congreso Nacional de la
Sociedad Matemática Mexicana in October of 1994. In this course we pre-
Preface to First Edition xvii

sented some applications of commutative algebra to combinatorics. We have

expanded these notes to include a more complete treatment of shellable and
Cohen–Macaulay complexes. The presentation of the last three sections of
this chapter, discussing the Hilbert series of face rings and the upper bound
conjectures, was inspired by [41, 65] and [393].
Since monomial algebras defined by square-free monomials of degree two
have an underlying graph theoretical structure it is natural that some inter-
action will occur between monomial algebras, graph theory, and polyhedral
geometry. We have included three chapters that focus on monomial algebras
associated to graphs. One of them is Chapter 7, where we present connec-
tions between graphs and ideals and study the Cohen–Macaulay property of
the face ring. Another is Chapter 10, where we present a combinatorial des-
cription of the integral closure of the corresponding monomial subring and
give some applications to graph theory. In Chapter 10 we consider mono-
mial subrings and toric ideals of complete graphs with the aim of computing
their Hilbert series, Noether normalizations, and Gröbner bases.
The central topic of Chapters 9 and 12 is the normality of monomial
subrings and ideals; some features of toric ideals are presented here.
Chapter 8 is devoted to the study of monomial curves and their toric
ideals, where the focus of our attention will be on monomial space curves
and monomial curves in four variables. Affine toric varieties and their toric
ideals are studied in Chapter 8.
Most of the material in this textbook has been written keeping in mind
a typical graduate student with a basic knowledge of abstract algebra and
a non-expert who wishes to learn the subject. We hope that this book can
be read by people from diverse subjects and fields, such as combinatorics,
graph theory, and computer algebra. Various units are accessible to upper
undergraduates.
In the last fifteen years a dramatic increase in the number of research
articles and books in commutative algebra that stress its connections with
computational issues in algebraic geometry and combinatorics has taken
place. Excellent references for computational and combinatorial aspects
that complement some of the material included here are [99, 128, 413],
[65, 395, 400] and [57, 438].
A constant concern during the writing of this text was to give appro-
priate credits for the proofs and results that were adapted from printed
material or communicated to us. We apologize for any involuntary omis-
sion and would appreciate any comments and suggestions in this regard.
During the fall of 1999 a course on monomial algebras associated to
graphs was given at the University of Messina covering Chapter 7 to Chapter
10 with the support of the Istituto Nazionale Di Alta Matematica Francesco
Severi. It is a pleasure to thank Vittoria Bonanzinga, Marilena Crupi,
Gaetana Restuccia, Rossana Utano, Maurizio Imbesi, Giancarlo Rinaldo,
xviii Preface to First Edition

Fabio Ciolli, and Giovanni Molica for the opportunity to improve those
chapters and for their hospitality.
We thank Wolmer V. Vasconcelos for his comments and encouragement
to write this book. A number of colleagues and students provided helpful
annotations to some early drafts. We are especially grateful to Adrián
Alcántar, Joe Brennan, Alberto Corso, José Martı́nez-Bernal, Susan Morey,
Carlos Renterı́a, Enrique Reyes, and Aron Simis. We are also grateful to
Laura Valencia for her competent secretarial assistance.
The Consejo Nacional de Ciencia y Tecnologı́a (CONACyT) and the
Sistema Nacional de Investigadores (SNI) deserve special acknowledgment
for their generous support. It should be mentioned that the development of
this book was included in the project Estudios sobre Algebras Monomiales,
which was supported by the CONACyT grant 27931E.
In the homepage “http://www.math.cinvestav.mx/∼vila” we will main-
tain an updated list of corrections.

Rafael H. Villarreal
Cinvestav-IPN
Mexico City, D.F.
Chapter 1

Polyhedral Geometry and

Linear Optimization

In this chapter we introduce several notions and results from polyhedral

geometry, combinatorial optimization and linear programming. Excellent
general references for these areas are [35, 87, 281, 372, 373, 438]. Then
we study relative volumes of lattice polytopes. We present various relations
between Hilbert bases, TDI linear systems of inequalities, the max-flow min-
cut property of clutters, integral linear systems, and the normality of affine
semigroups and Rees semigroups. Elementary integral vectors and matroids
are introduced at the end of the chapter.

1.1 Polyhedral sets and cones

An affine space or linear variety in Rn is by definition a translation of a
linear subspace of Rn . Let A be a subset of Rn . The affine space generated
by A, denoted by aff(A), is the set of all affine combinations of points in A:
aff(A) = {a1 v1 + · · · + ar vr | vi ∈ A, ai ∈ R, a1 + · · · + ar = 1}.
There is a unique linear subspace V of Rn such that aff(A) = x0 + V , for
some x0 ∈ Rn . The dimension of A is defined as dim A = dim R (V ).
Definition 1.1.1 An affine map is a function between two affine spaces
that preserves affine combinations.
A point x ∈ Rn is called a convex combination of v1 , . . . , vr ∈ 
Rn if there
are a1 , . . . , ar in R such that ai ≥ 0 for all i, x = i ai vi and i ai = 1.
Let A be a subset of Rn . The convex hull of A, denoted by conv(A), is the
set of all convex combinations of points in A. If A = conv(A) we say that
A is a convex set.
2 Chapter 1

Definition 1.1.2 A convex polytope P ⊂ Rn is the convex hull of a finite

set of points v1 , . . . , vr in Rn , that is, P = conv(v1 , . . . , vr ).

The inner product of two vector x = (x1 , . . . , xn ) and y = (y1 , . . . , yn )

in Rn is defined by

x, y = x · y = x1 y1 + · · · + xn yn .

Definition 1.1.3 Let x = (x1 , . . . , xn ) be a point in Rn . The Euclidean

norm of x is defined as 
x = x, x.

We shall always assume that a subset of Rn has the topology induced

by the usual topology of Rn .

Definition 1.1.4 A point a of a set A in Rn is said to be a relative interior

point of A if there exists r > 0 such that

Br (a) ∩ aff(A) ⊂ A,

where Br (a) = {x| x − a < r}.

Let A ⊂ Rn . The set of relative interior points of A, denoted by ri(A),

is called the relative interior of A. We denote the closure of A by A. The
set A \ ri(A) is called the relative boundary of A and is denoted by rb(A),
points in rb(A) are called the relative boundary points of A. If dim(A) = n,
the relative interior of A is sometimes denoted by Ao .
An affine space of Rn of dimension n − 1 is called an affine hyperplane.
Given a ∈ Rn \ {0} and c ∈ R, define the affine hyperplane

H(a, c) = {x ∈ Rn | x, a = c}.

Notice that any affine hyperplane of Rn has this form. The two closed
halfspaces bounded by H(a, c) are

H + (a, c) = {x ∈ Rn | x, a ≥ c} and H − (a, c) = H + (−a, −c).

If a is a rational vector and c is a rational number, the affine hyperplane

H(a, c) (resp. the halfspace H + (a, c)) is called a rational hyperplane (resp.
rational halfspace). If c = 0, for simplicity the set Ha will denote the
hyperplane of Rn through the origin with normal vector a, that is,

Ha := H(a, 0) = {x ∈ Rn | x, a = 0}.

The two closed halfspaces bounded by Ha are denoted by

Ha+ = {x ∈ Rn | x, a ≥ 0} and Ha− = {x ∈ Rn | x, a ≤ 0}.

Polyhedral Geometry and Linear Optimization 3

Definition 1.1.5 A polyhedral set or convex polyhedron is a subset of Rn

which is the intersection of a finite number of closed halfspaces of Rn . The
set Rn is considered a polyhedron.
The transpose of a matrix A (resp. vector x) will be denoted by At or
A (resp. xt or x ). Often a vector x will denote a column vector or a row


vector, from the context the meaning should be clear. Thus, a polyhedral
set Q can be represented as

Q = {x ∈ Rn | Ax ≤ b}

for some matrix A and for some vector b. As usual, if a = (ai ) and c = (ci )
are vectors in Rq , then a ≤ c means ai ≤ ci for all i. If A and b have rational
entries, Q is called a rational polyhedron.
Since the intersection of an arbitrary family of convex sets in Rn is
convex, one derives that any polyhedral set is convex and closed.

Definition 1.1.6 Let Q be a closed convex set in Rn . A hyperplane H of

Rn is called a supporting hyperplane of Q if Q is contained in one of the two
closed halfspaces bounded by H and Q ∩ H = ∅.

Definition 1.1.7 A proper face of a polyhedral set Q is a set F ⊂ Q such

that there is a supporting hyperplane H(a, c) satisfying the conditions:
(a) F = Q ∩ H(a, c) = ∅,
(b) Q ⊂ H(a, c), and Q ⊂ H + (a, c) or Q ⊂ H − (a, c).
The improper faces of a polyhedral set Q are Q itself and ∅.

Definition 1.1.8 A proper face F of a polyhedral set Q ⊂ Rn is called a

facet of Q if dim(F ) = dim(Q) − 1.

Proposition 1.1.9 If Q is a polyhedral set in Rn and F1 , F2 are faces of

Q, then their intersection F = F1 ∩ F2 is a face of Q.

Proof. Let Hi = H(ai , ci ) be a supporting hyperplane of Q, where ci ∈ R

and 0 = ai ∈ Rn , such that Fi = Q ∩ Hi and Q ⊂ Hi+ for i = 1, 2. Next we
prove the equality
F = Q ∩ H(a1 + a2 , c1 + c2 ).
The left-hand side is clearly contained in the right-hand side. On the other
hand if x ∈ Q ∩ H(a1 + a2 , c1 + c2 ), then using x, ai  ≥ ci one has:

c1 + c2 = x, a1 + a2  = x, a1  + x, a2  ≥ c1 + c2 ,

hence x, ai  = ci for i = 1, 2 and x ∈ F . As Q ⊂ H + (a1 + a2 , c1 + c2 ), the

set F is a face of Q. 2
4 Chapter 1

Definition 1.1.10 A partially ordered set or poset is a pair P = (V, ),

where V is a finite set of vertices and  is a binary relation on V satisfying:
(a) u  u, ∀u ∈ V (reflexivity).
(b) u  v and v  u, imply u = v (antisymmetry).
(c) u  v and v  w, imply u  w (transitivity).

A poset P = (V, ) with vertex set V = {x1 , . . . , xn } can be displayed

by an inclusion diagram , where xi is joined to xj by raising a line if xi ≺ xj
and there is no other vertex in between.

Example 1.1.11 The inclusion diagram of the divisors of 12 is:

r12
6r @
H @r 4
HHr
3r 2
@ @r 1

A poset in which any two elements x, y have a greatest lower bound

inf{x, y} and a lowest upper bound sup{x, y} is called a lattice. A lattice
is called complete if inf S and sup S exist for any subset S. A mapping ϕ
from one lattice L1 = (V1 , ) onto another lattice L2 = (V2 , ) is called
an isomorphism when it is one-to-one and we have x  y if and only if
ϕ(x)  ϕ(y) for all x, y ∈ V1 .

Corollary 1.1.12 Let Q be a polyhedral set in Rn and let F be the set of

all faces of Q. The partially ordered set (F , ⊂) is a complete lattice with
the lattice operations

inf G = ∩{F | F ∈ G} and sup G = ∩{G ∈ F | ∀F ∈ G; F ⊂ G}.

Proof. It follows from Proposition 1.1.9 and from the fact that a convex
polyhedron has only finitely many faces [427, Theorem 3.2.1(v)]. 2

Definition 1.1.13 The lattice (F , ⊂) is called the face-lattice of Q.

Definition 1.1.14 A polytopal complex C is a finite collection of polytopes

in Rn such that (i) ∅ ∈ C, (ii) if P ∈ C, then all faces of P are in C, and (iii)
the intersection P ∩ Q of two polytopes P, Q ∈ C is a face of both P and Q.

Similarly, one can define the notion of a polyhedral complex replacing

polytope by polyhedron. The dimension of a polytopal complex C, denoted
by dim(C), is the largest dimension of a polytope in C.

Definition 1.1.15 Let P be a convex polytope. All proper faces of P form

the boundary complex C(∂P), whose facets are the facets of P.
Polyhedral Geometry and Linear Optimization 5

The boundary complex C(∂P) of a polytope P of dimension d + 1 is a

pure polytopal complex of dimension d [57, Chapter 2].

Definition 1.1.16 Let P be a convex polytope of dimension d + 1. The

boundary complex C(∂P) is shellable if there is a linear ordering F1 , . . . , Fs
of the facets of P such that for every 1 ≤ i < j ≤ s, there is 1 ≤  < j
satisfying Fi ∩ Fj ⊂ F ∩ Fj and F ∩ Fj is a facet of Fj .

Theorem 1.1.17 [58] The boundary complex of a polytope is shellable.

The set of nonnegative real numbers and the set of nonnegative integers
are denoted by:

R+ = {x ∈ R| x ≥ 0} and N = {0, 1, 2, . . . }

respectively, R+ is also denoted by R≥0 . If A is a set of points in Rn , the

cone generated by A, denoted by R+ A or cone(A), is defined as
 q  
 

R+ A = ai βi  ai ∈ R+ , βi ∈ A for all i .

i=1

The vector space spanned by A is denoted by RA. Given a vector v ∈ Rn ,

we define:
R+ v := {λv| λ ∈ R+ }.

Theorem 1.1.18 (Carathéodory’s theorem) Let v1 , . . . , vq be a sequence

of vectors in Rn not all of them zero. If x ∈ R+ v1 + · · · + R+ vq , then
there is a linearly independent set V ⊂ {v1 , . . . , vq } such that x ∈ R+ V.

Proof. By induction on q. The case q = 1 is clear. If q ≥ 2, we can

write x = a1 v1 + · · · + aq vq , ai ≥ 0 for i = 1, . . . , q. One may assume that
v1 , . . . , vq are linearly dependent and ai > 0 for all i, otherwise the result
follows by induction.  There are real numbers b1 , . . . , bq such that at least
q
one bi is positive and i=1 bi vi = 0. Setting

0 < c = min{ai /bi | bi > 0} = aj /bj ,

q
we get x = x − c(b1 v1 + · · · + bq vq ) = i=1 (ai − cbi )vi , with ai − cbi ≥ 0 for
all i and aj − cbj = 0. Hence, by induction, the result follows. 2

Definition 1.1.19 A set of vectors α1 , . . . , αq ∈ Rn is affinely independent

if forevery sequence λ1 , . . . , λq of real numbers satisfying qi=1 λi αi = 0
q
and i=1 λi = 0, one has λi = 0 for all i.

Corollary 1.1.20 Let v1 , . . . , vq be a set of vectors in Rn and let x be in its

convex hull. Then there exists an affinely independent set V ⊂ {v1 , . . . , vq }
such that x ∈ conv(V).
6 Chapter 1

Proof. Consider B = {(v1 , 1), . . . , (vq , 1)}. Since (x, 1) belongs to R+ B,

the result follows from Carathéodory’s theorem and Exercise 1.1.67. 2

Definition 1.1.21 If C ⊂ Rn is closed under linear combinations with

nonnegative real coefficients, we say that C is a convex cone. A polyhedral
cone is a convex cone which is also a polyhedral set.
Proposition 1.1.22 If C is a closed convex cone and H is a supporting
hyperplane of C, then H is a hyperplane passing through the origin.
Proof. Let H = H(a, c). Since 0 ∈ C ⊂ H + , one has c ≤ 0. If C ∩H = {0},
then c = 0 as required. Assume C ∩ H = {0} and pick 0 = z ∈ C such that
z, a = c. Using that tz ∈ C ⊂ H + (a, c) for all t ≥ 0, one derives
tc = tz, a = tz, a ≥ c ∀ t ≥ 0 ⇒ c = 0. 2
An affine space of dimension 1 is called a line. The following result is
quite useful in determining the facets of a polyhedral cone.
Proposition 1.1.23 Let A be a finite set of points in Zn and let F be a
face of R+ A. The following hold.
(a) If F = {0}, then F = R+ V for some V ⊂ A.
(b) If dim(F ) = 1 and R+ A contains no lines, then F = R+ α with α ∈ A.
(c) If dim(R+ A) = n and F is a facet defined by the supporting hyperplane
Ha , then Ha is generated by a linearly independent subset of A.
Proof. Let F = R+ A ∩ Ha with R+ A ⊂ Ha− . Then F is equal to the cone
generated by the set V = {α ∈ A| α, a = 0}. This proves (a). Parts (b)
and (c) follow from (a). 2
Most of the notions and results considered thus far make sense if we
replace R by an intermediate field Q ⊂ K ⊂ R, i.e., we can work in the
affine space Kn . However with very few exceptions, like for instance Theo-
rem 1.1.24, we will always work in the Euclidean space Rn or Qn .
For convenience we state the fundamental theorem of linear inequalities:
Theorem 1.1.24 [372, Theorem 7.1] Let Q ⊂ K ⊂ R be an intermediate
field and let C ⊂ Kn be a cone generated by A = {α1 , . . . , αq }. If α ∈ Kn \C
and t = rank{α1 , . . . , αq , α}, then there exists a hyperplane Ha containing
t−1 linearly independent vectors from A such that a, α > 0 and a, αi  ≤ 0
for i = 1, . . . , q.
Theorem 1.1.25 (Farkas’s Lemma) Let A be an n × q matrix with entries
in a field K and let α ∈ Kn . Assume Q ⊂ K ⊂ R. Then either there exists
x ∈ Kq with Ax = α and x ≥ 0, or there exists v ∈ Kn with vA ≥ 0 and
v, α < 0, but not both.
Polyhedral Geometry and Linear Optimization 7

Proof. Let A = {α1 , . . . , αq } be the set of column vectors of A. Assume

that there is no x ∈ Kq with Ax = α and x ≥ 0, i.e., α is not in R+ A.
By Theorem 1.1.24 there is a hyperplane Hv such that v, α < 0 and
v, αi  ≥ 0 for all i. Hence vA ≥ 0, as required. If both conditions hold,
then 0 > v, α = v, Ax = vA, x ≥ 0, a contradiction. 2
The reader is referred to [438] for other versions of Farkas’s lemma. The
next result tells us how to separate a point from a cone.
Corollary 1.1.26 Let C ⊂ Kn be a cone generated by A = {a1 , . . . , am }.
If γ ∈ Kn and γ ∈
/ C, then there is a hyperplane H through the origin such
that γ ∈ H − \H and C ⊂ H + .
Proof. Let A be the matrix with column vectors a1 , . . . , am . By Farkas’s
lemma (see Theorem 1.1.25) there exists μ ∈ Kn such that μA ≥ 0 and
γ, μ < 0. Thus μ, ai  ≥ 0 for all i. If H is the hyperplane through the
origin with normal vector μ we get C ⊂ H + , as required. 2
Corollary 1.1.27 Let A be a finite set in Zn , then
ZA ∩ R+ A = ZA ∩ Q+ A and Zn ∩ R+ A = Zn ∩ Q+ A,
where ZA is the subgroup of Zn spanned by A and Q+ = {x ∈ Q| x ≥ 0}.
Proof. It follows at once from Theorem 1.1.25. 2
Definition 1.1.28 We say that C is a finitely generated cone if C = R+ A,
for some finite set A = {v1 , . . . , vq }.
The proof of the next result yields the duality theorem for cones.
Theorem 1.1.29 If C ⊂ Rn , then C is a finitely generated cone (resp.
finitely generated cone by rational vectors) if and only if C is a polyhedral
cone (resp. rational polyhedral cone) in Rn .
Proof. ⇒) Assume that C = (0) is a cone generated by A = {α1 , . . . , αm }.
We set r = dim(R+ A). Notice that aff(C) is the real vector space generated
by A, because 0 ∈ C. If aff(C) = C, then C is a polyhedral cone, because
C is the intersection of n − r hyperplanes of Rn through the origin. Now
assume that C  aff(C). Consider the family
F = {F | F = Ha ∩ C; dim(F ) = r − 1; C ⊂ Ha− }.
By Theorem 1.1.24 the family F is non-empty. Notice that F is a finite
set because each F in F is a cone generated by a subset of A; see Proposi-
tion 1.1.23. Assume that F = {F1 , . . . , Fs }, where Fi = Hai ∩ C. We claim
that the following equality holds
C = Ha−1 ∩ · · · ∩ Ha−s ∩ aff(C). (1.1)
8 Chapter 1

The inclusion “⊂” is clear. To show the inclusion “⊃”, we proceed by

contradiction. Assume that there exists α ∈ C such that α belongs to
the right-hand side of Eq. (1.1). By Theorem 1.1.24 and by reordering
the elements of A if necessary, there is a hyperplane Ha containing linearly
independent vectors α1 , . . . , αr−1 such that (i) a, α > 0, and (ii) a, αi  ≤ 0
for i = 1, . . . , m. Thus F = Ha ∩ C = Hak ∩ C for some 1 ≤ k ≤ s.
We may assume that α1 , . . . , αr form a basis of aff(C). Since α ∈ aff(C)
there are scalars λ1 , . . . , λr with α = λ1 α1 + · · · + λr αr . Using (ii) we get
a, α = λr αr , a > 0. Hence λr < 0. On the other hand ak , αr  < 0
because C ⊂ Ha−k and dim(Fk ) = r − 1. Therefore

ak , α = λr ak , αr  > 0,

 
<0 <0

a contradiction to the fact that α ∈ Ha−k .

⇐) Assume that C = Hb−1 ∩ · · · ∩ Hb−r , where b1 , . . . , br ∈ Rn . Consider
the cone C  generated by b1 , . . . , br . From the first part of the proof we can
write
C  = R+ {b1 , . . . , br } = Hα−1 ∩ · · · ∩ Hα−m , (∗)
for some set of vectors A = {α1 , . . . , αm } in Rn . Next we show the equality
C = R+ A. Notice that bi , αj  ≤ 0 for all i, j, because bi ∈ C  for all i.
Thus R+ A ⊂ C. Assume that there is α ∈ C \ R+ A. By Corollary 1.1.26,
there exists a hyperplane Ha such that R+ A ⊂ Ha− and a, α > 0. Hence
αi , a ≤ 0 for all i, and by Eq. (∗) we conclude that a ∈ R+ {b1 , . . . , br }.
Therefore, we can write a = λ1 b1 + · · · + λr br , λi ≥ 0 for all i. Since α ∈ C,
we have α, a = λ1 α, b1  + · · · + λr α, br  ≤ 0, contradicting a, α > 0.
Thus C = R+ A. The respective statement about the rationality character
of the representations is left as an exercise. 2

Corollary 1.1.30 (Duality theorem for cones) Let B = {β1 , . . . , βr } be a

subset of Rn , and let {α1 , . . . , αm } and {c1 , . . . , cs } be subsets of Rn \ {0}.
− − −
(a) If R+ B = ∩m
i=1 Hαi , then Hβ1 ∩ · · · ∩ Hβr = R+ α1 + · · · + R+ αm .

(b) If R+ B = ∩si=1 Hc+i , then Hβ+1 ∩ · · · ∩ Hβ+r = R+ c1 + · · · + R+ cs .

Proof. Part (a) follows from the second part of the proof of Theorem 1.1.29.
−
Part (b) follows using the equality Hc+i = H−c i
and using part (a). 2
By Theorem 1.1.29 a polyhedral cone C  Rn has two representations:
Minkowski representation C = R+ B with B = {β1 , . . . , βr } a finite set, and
Implicit representation C = Hc+1 ∩ · · · ∩ Hc+s for some c1 , . . . , cs ∈ Rn \ {0}.
Polyhedral Geometry and Linear Optimization 9

From the duality theorem for cones these two representations satisfy:

Hβ+1 ∩ · · · ∩ Hβ+r = R+ c1 + · · · + R+ cs . (1.2)

The dual cone of C is defined as

C ∗ := Hc+ = Ha+ .
c∈C a∈B

∗∗
By the duality theorem one has C = C. An implicit representation
of C is called irredundant or irreducible if none of the closed half-spaces
Hc+1 , . . . , Hc+s can be omitted from the intersection.

Remark 1.1.31 The left-hand side of Eq. (1.2) is an irredundant repre-

sentation of C ∗ if and only if no proper subset of B generates C.

Corollary 1.1.32 Let C ⊂ Rn be a finitely generated cone and let

F = {F | F = Ha ∩ C; dim(F ) = r − 1; C ⊂ Ha− } = {F1 , . . . , Fs } = ∅,

where Fi = C ∩ Hai , dim(C) = r. Then C = aff(C) ∩ Ha−1 ∩ · · · ∩ Ha−s .

Proof. It follows from the first part of the proof of Theorem 1.1.29. 2
One of the fundamental results in polyhedral geometry is the following
remarkable decomposition theorem for polyhedra. See [372, Corollary 7.1b]
and [427, Theorem 4.1.3] for historical comments and for more information
about this result.

Theorem 1.1.33 (Finite basis theorem) If Q is a set in Rn , then Q is a

polyhedron (resp. rational polyhedron) if and only if Q can be expressed as
Q = P + C, where P is a convex polytope (resp. rational polytope) and C
is a finitely generated cone (resp. finitely generated rational cone).

Proof. ⇒) Let Q = {x| Ax ≤ b} be a polyhedron in Rn , where A is a

matrix and b is a vector. Consider the set
  
 x 
C = x ∈ R ; λ ∈ R+ ; Ax − λb ≤ 0 .
n
λ 

Notice that C  can be written as

     
x  A −b x
C = x ∈ R n
; λ ∈ R; ≤ 0 ,
λ  0 −1 λ

where −b is a column vector. Thus C  is a polyhedral cone in Rn+1 . Using

Corollary 1.1.29 we get that C  can be expressed as
   
x1 x
C  = R+ ,..., m (λi ≥ 0; xi ∈ Rn ).
λ1 λm
10 Chapter 1

We may assume that λi ∈ {0, 1} for all i. Set

A = {xi | λi = 1} = {x1 , . . . , xr }, B = {xi | λi = 0} = {xr+1 , . . . , xm },

P = conv(A), and C = R+ B. Notice that x ∈ Q if and only if (x, 1) ∈ C  .

Thus x ∈ Q if and only if (x, 1) can be written as
         
x x1 xr xr+1 x
= μ1 + · · · + μr + μr+1 + · · · + μm m
1 1 1 0 0

with μi ≥ 0 for all i. Consequently x ∈ Q if and only if x ∈ P +C. Therefore

we obtain Q = P + C.
⇐) Assume that Q is equal to P + C with P = conv(x1 , . . . , xr ) and
C = R+ (xr+1 , . . . , xm ). Consider the following finitely generated cone
       
 x1 xr xr+1 x
C = R+ ,..., , ,..., m .
1 1 0 0

By Corollary 1.1.29 the cone C  is a polyhedron. Thus there exists a matrix

A and a vector b such that C  can be written as
  
x 
C = x ∈ R n
; λ ∈ R; Ax + λb ≤ 0 .
λ 

Since x ∈ Q if and only if (x, 1) ∈ C  we conclude that Q = {x| Ax ≤ −b},

that is Q is a polyhedron. This proof is due to Schrijver [372]. 2
Thus, by the finite basis theorem, a polyhedron has two representations.
The computer program PORTA [84] can be used to switch between these two
representations. This program is available for Unix and Windows systems.
For polyhedral cones one can use Normaliz [68].

Corollary 1.1.34 A set Q ⊂ Rn is a convex polytope if and only if Q is a

bounded polyhedral set.

Proof. If Q = conv(α1 , . . . , αm ) is a polytope, then by the triangle inequal-

ity for all x ∈ Q we have x ≤ α1 + · · · + αm . Thus Q is bounded.
Conversely if Q is a bounded polyhedron, then by Theorem 1.1.33 we can
decompose Q as Q = P + C, with P a polytope and C a finitely generated
cone. Notice that C = {0}, otherwise fixing p0 ∈ P and c0 ∈ C \ {0} we
get p0 + λc0 ∈ Q for all λ > 0, a contradiction because Q is bounded. Thus
P = Q, as required. 2

Definition 1.1.35 Let Q be a polyhedral set and x0 ∈ Q. The point x0 is

called a vertex or an extreme point of Q if {x0 } is a proper face of Q.

Proposition 1.1.36 [57, Theorem 7.2] If P = conv(α1 , . . . , αr ) ⊂ Rn and

V is the set of vertices of P, then V ⊂ {α1 , . . . , αr } and P = conv(V ).
Polyhedral Geometry and Linear Optimization 11

Lemma 1.1.37 [57, Theorem 5.6] If Q is a polyhedral set in Rn and v ∈ Q

is not a vertex of Q, then there is a face F of Q such that v ∈ ri(F ).
Lemma 1.1.38 Let Q be a convex polyhedron in Rn and x0 ∈ Q. Then x0
is a vertex of Q if and only if Q \ {x0 } is a convex set.
Proof. ⇒) Let H(a, c) be a proper supporting hyperplane of Q such that
{x0 } = Q ∩ H(a, c) and Q ⊂ H + (a, c). Take x, y ∈ Q \ {x0 } and consider
z = ty + (1 − t)x with 0 < t < 1. Note that x, y are not in H(a, c), thus
z, a = ty, a + (1 − t)x, a > tc + (1 − t)c = c.
Hence z, a > c, that is, z = x0 and z ∈ Q, as required.
⇐) Let Q = ∩ri=1 H + (ai , ci ) be a decomposition of Q as an intersection
of closed halfspaces, where 0 = ai ∈ Rn and ci ∈ R for all i. First observe
that x0 is in H(ai , ci ) for some i, otherwise if x0 , ai  > ci for all i, there
is an open ball Bδ (x0 ) in Rn , of radius δ centered at x0 , whose closure lies
in Q. Hence taking two antipodal points x1 , x2 in the boundary of Bδ (x0 )
one obtains x0 ∈ Q \ {x0 }, a contradiction. One may now assume there is
k ≥ 1 such that x0 ∈ H(ai , ci ) for i ≤ k and x0 , ai  > ci for i > k. Set
A = H(a1 , c1 ) ∩ · · · ∩ H(ak , ck ).
We claim that A = {x0 }. If there is x1 ∈ A \ {x0 }, pick Bδ (x0 ) whose
closure is contained in H + (ai , ci ) for all i > k. Since z = tx0 + (1 − t)x1
is in A for all t ∈ R, making t = 1 + δ/ x1 − x0 one derives z ∈ A and
z − x0 = δ, thus z ∈ Q. Note that if k = r, then x1 is already in Q
and in this case we set z = x1 . Making t = −1 in z1 = tz + (1 − t)x0 one
concludes z1 ∈ A and z1 − x0 = δ. Altogether z, z1 are in Q \ {x0 } and
x0 = (z + z1 )/2, a contradiction because Q \ {x0 } is a convex set.
From the equality {x0 } = A = A∩Q one has that {x0 } is an intersection
of faces. Using Proposition 1.1.9 yields that {x0 } is a face, as required. 2
Proposition 1.1.39 If Q = C + P ⊂ Rn with C a polyhedral cone and P
a polytope, then every vertex of Q is a vertex of P.
Proof. Let x be a vertex of Q and write x = c + p for some c ∈ C and
p ∈ P. We claim that p is a vertex of P. If p is not a vertex of P, then by
Lemma 1.1.37 there is a face F of P such that p ∈ ri(F ). Pick p = v ∈ F .
By [57, Theorem 3.5], there is y ∈ F with p ∈]y, v[. Hence p = λy + (1 − λ)v
with 0 < λ < 1. Note that x = y + c ∈ Q, x = v + c ∈ Q, and
x = λ(y + c) + (1 − λ)(v + c) (0 < λ < 1),
thus Q \ {x} is not a convex set, a contradiction to Lemma 1.1.38. This
proves that p is a vertex of P. Assume c = 0, then
p + λc = x = p + c = p + 2c (0 < λ < 1).
12 Chapter 1

Notice that x = p+c = λ0 (p+λc)+(1−λ0 )(p+2c) is a convex combination,

where 0 < λ0 = 1/(2 − λ) < 1. Since p + λc and p + 2c are in Q \ {x}, this
shows that Q \ {x} is not a convex set, a contradiction. Thus c = 0 and x
is a vertex of P, as required. 2

Theorem 1.1.40 [372, Theorem 8.4] Let Q ⊂ Rn be a polyhedral set. Then

Q has at least one vertex if and only if Q does not contain any lines in Rn .

Proposition 1.1.41 Let Q be a polyhedron containing no lines and let

f (x) = x, a. If max{f (x)| x ∈ Q} < ∞, then the maximum is attained at
a vertex of Q.

Proof. Let x0 ∈ Q be an optimal solution and let λ = f (x0 ) be the optimal

value. Note that F = Q ∩ H(a, λ) is a face of Q because Q ⊂ H − (a, λ) and
x0 ∈ F . Since F contains no lines, by Theorem 1.1.40, the face F contains
at least one vertex z0 , which is also a vertex of Q by transitivity. Thus the
optimal value λ is attained at the vertex z0 , as required. 2

Theorem 1.1.42 If Q ⊂ Rn is a polyhedral set containing no lines and

α1 , . . . , αr are the vertices of Q, then there are β1 , . . . , βs ∈ Rn such that

Q = conv(α1 , . . . , αr ) + (R+ β1 + · · · + R+ βs ).

Proof. It follows from the proof of [427, Theorem 4.1.3]. 2

Definition 1.1.43 If a polyhedron Q in Rn is represented as

Q = aff(Q) ∩ H + (a1 , c1 ) ∩ · · · ∩ H + (ar , cr ) (1.3)

with ai ∈ Rn \ {0}, ci ∈ R for all i, and none of the closed halfspaces

H + (a1 , c1 ), . . . , H + (ar , cr ) can be omitted from the intersection, we say
that Eq. (1.3) is an irreducible representation of Q.

Theorem 1.1.44 [427, Theorem 3.2.1] Let Q be a polyhedral set in Rn

which is not an affine space and let

Q = aff(Q) ∩ H + (a1 , c1 ) ∩ · · · ∩ H + (ar , cr )

be an irreducible representation. If Fi = Q ∩ H(ai , ci ), i = 1, . . . , r, then

(a) ri(Q) = {x ∈ Q| x, a1  > c1 , . . . , x, ar  > cr };
(b) rb(Q) = F1 ∪ · · · ∪ Fr ;
(c) the facets of Q are precisely the sets F1 , . . . , Fr ;
(d) each face F  Q is the intersection of all Fi such that F ⊂ Fi .
Polyhedral Geometry and Linear Optimization 13

Corollary 1.1.45 Let Q be a polyhedron and let β be a vector in ri(Q).

If F is a proper face of Q, then β ∈ F .
r
Proposition 1.1.46 Let Q = i=1 H + (ai , ci ) be a polyhedral set in Rn
which is not an affine space, where 0 = ai ∈ Rn and ci ∈ R for all i. If
x0 ∈ Rn , then x0 is a vertex of Q if and only if
(a) {x0 } = ∩i∈I H(ai , ci ) for some I ⊂ {1, . . . , r}, and
(b) x0 , ai  ≥ ci for all i = 1, . . . , r.

Proof. ⇒) This direction follows at once from the proof of Lemma 1.1.38.
⇐) Note x0 ∈ Q. Since the intersection of faces of Q is a face (see
Proposition 1.1.9), one has that {x0 } is a face. 2

Corollary 1.1.47 Let A = (aij ) be an r×n matrix, let c = (ci ) be a column

vector and let a1 , . . . , ar be the rows of A. If Q = {x ∈ Rn | Ax ≥ c} and
x0 ∈ Q, then x0 is a vertex of Q if and only if there is J ⊂ {1, . . . , r} with
|J| = n such that the set {ai | i ∈ J} is linearly independent and

{x0 } = {(xi ) ∈ Rn | ai1 x1 + · · · + ain xn = ci for all i ∈ J}.

Proof. Assume x0 is a vertex of Q. Hence, by Proposition 1.1.46, x0 is the

unique solution of a system of linear equations:

ai1 x1 + · · · + ain xn = ci (i ∈ I)

for some I ⊂ {1, . . . , r}. Let [A |c ] be the augmented matrix of this system,
by Gaussian elimination one obtains that this matrix reduces to
 
In | c
.
0 | 0

Hence the rank of A is equal to n. Thus there are n linearly independent

rows of A and x0 is the unique solution of the system

ai1 x1 + · · · + ain xn = ci (i ∈ J ⊂ I)

for some J ⊂ I with |J| = n. The converse is clear. 2

Let Q be a polyhedron in Rn represented by a system of linear con-
straints
ai , x ≥ bi (i = 1, . . . , r),
where ai ∈ Rn and bi ∈ R for all i. With a slight abuse of language, we will
say that the constraints are linearly independent if the corresponding ai are
linearly independent.
14 Chapter 1

Definition 1.1.48 A vector x0 in Rn is called a basic feasible solution of a

system of linear constraints ai , x ≥ bi , i = 1, . . . , r, if
(a) x0 satisfies all linear constraints, and
(b) out of the constraints that satisfy ai , x0  = bi , there are n of them
that are linearly independent.

The next result is a restatement of the corollary above.

Corollary 1.1.49 Let Q be a polyhedron in Rn . A vector x0 in Rn is a

vertex of Q if and only if x0 is a basic feasible solution of any system of
linear constraints that represents Q.

This result yields a method to find the vertices of a polyhedron which

is by no means the best in practice. See [10, 11] for a thorough discussion
on how to find all the vertices of a polyhedron.

Corollary 1.1.50 If C = Rn is a polyhedral cone of dimension n, then

there is a unique irreducible representation

C = Ha+1 ∩ · · · ∩ Ha+r , where ai ∈ Rn \ {0}.

Proof. According to Proposition 1.1.22 a set F is a proper face of C if

there is a supporting hyperplane Ha of C such that F = C ∩ Ha = ∅ and
C ⊂ Ha . In particular the facets of C are defined by hyperplanes through
the origin. Therefore by Theorem 1.1.44 the irreducible representation of C
has the required form and is unique. 2

Proposition 1.1.51 If Q = Rn is a rational polyhedral cone of dimension

n in Rn , then there are unique (up to sign) a1 , . . . , ar in Zn with relatively
prime entries such that the irreducible representation of Q is

Q = Ha+1 ∩ Ha+2 ∩ · · · ∩ Ha+r

Proof. By the finite basis theorem, there are α1 , . . . , αq in Qn such that

Q = R+ α1 + · · · + R+ αq . First note that if Hb is a supporting hyperplane of
Q generated by a set of n − 1 linearly independent vectors in {α1 , . . . , αq },
then by the Gram–Schmidt process Hb has an orthogonal basis of vectors
in Qn , and consequently there is a normal vector a to Hb such that a ∈ Qn
and Ha = Hb . Hence multiplying a by a suitable integer and then dividing
by the greatest common divisor of the entries, one may assume Ha = Hb ,
where a is in Zn and has relative prime entries. Observe that a, b are linearly
dependent because the orthogonal complement of Ha is one dimensional. It
is readily seen that a is uniquely determined up to sign.
To complete the proof use Proposition 1.1.23 (c) to see that any facet
of Q is defined by a supporting hyperplane Hb as above. 2
Polyhedral Geometry and Linear Optimization 15

Definition 1.1.52 A polyhedron containing no lines is called pointed.

If F is a face of dimension 1 of a pointed polyhedral cone, then F = R+ v

for some v (see Proposition 1.1.23). The face F is called an extremal ray.

Lemma 1.1.53 [372, Section 8.8] If C is a pointed polyhedral cone in Rn ,

then C is generated by non-zero representatives of its extremal rays.

Theorem 1.1.54 Let A be a finite set of non-zero points in Zn and let V

be the set of all α ∈ A such that R+ α is a face of R+ A of dimension 1. If
R+ A is a pointed cone, then R+ A = R+ V.

Proof. We set A = {α1 , . . . , αq }. Let F be a face of R+ A of dimension 1.

Then, by Proposition 1.1.23, F = R+ αi for some i. Thus, the result follows
from Lemma 1.1.53. 2

Definition 1.1.55 Let A be an n × q real matrix and let b, c be two real

vectors of sizes q and n, respectively. The primal problem is defined as
n
Maximize f (x) = i=1 ci xi (∗)
Subject to xA ≤ b and x ≥ 0.

Its dual problem is defined as


Minimize g(y) = qi=1 bi yi (∗∗)
Subject to Ay ≥ c and y ≥ 0.

A solution x0 ∈ Rn of the primal (resp. dual) problem that maximizes the

objective function f is called and optimal solution, the corresponding value
f (x0 ) of the objective function is called an optimal value.

The most important theorem in Linear Programming (LP) theory is:

Theorem 1.1.56 (LP duality theorem, [372, Corollary 7.1.g]) If the primal
problem (∗) has an optimal solution (x1 , . . . , xn
), then the dual
q problem (∗∗)
n
has an optimal solution (y1 , . . . , yq ) such that i=1 ci xi = i=1 bi yi .

Theorem 1.1.57 [372, Corollary 7.1g, p. 92] Let A be a matrix and b, c

vectors. Then

max{c, x | xA ≤ b} = min{b, y | y ≥ 0, Ay = c} (1.4)

provided that at least one of the sets in (1.4) is non-empty.

An immediate consequence of the duality theorem is:

16 Chapter 1

Corollary 1.1.58 Consider the LP primal-dual pair

max{x, c| xA ≤ b} = min{y, b| y ≥ 0; Ay = c}.

Let x and y be feasible solutions, i.e., xA ≤ b, y ≥ 0 and Ay = c. Then the

following conditions are equivalent :
(a) x and y are both optimum solutions.
(b) x, c = y, b.
(c) (b − xA)y = 0.

Condition (c) is called complementary slackness. See Exercise 1.1.81 for

another form of complementary slackness

Proposition 1.1.59 [281, Proposition 3.3] Let Q = {x ∈ Rn | xA ≤ b} = ∅

be a polyhedral set and let F = ∅ be a proper subset of Q. Then the following
conditions are equivalent :
(a) F = {x ∈ Q | A x = b } for some subsystem xA ≤ b of xA ≤ b.
(b) F is a proper face of Q.
(c) F = {x ∈ Q | x, c = δ} for some vector 0 = c ∈ Rn such that the
maximum value δ = max{x, c | x ∈ Q} is finite.

Definition 1.1.60 A minimal face of a polyhedron is a face not containing

any other face.

Lemma 1.1.61 [372, p. 104] Let Q = {x ∈ Rn | xA ≤ b} = ∅ be a convex

polyhedron. A set F is a minimal face of Q if and only if ∅ = F  Q and F
is equal to F = {x ∈ Rn | A x = b } for some subsystem xA ≤ b of xA ≤ b.

Definition 1.1.62 A rational polyhedron Q ⊂ Rn is called integral if Q

is equal to conv(Zn ∩ Q).

If Q is a pointed rational polyhedron, then Q is integral if and only if Q

has only integral vertices. This is shown below.

Theorem 1.1.63 [281, Theorem 5.12] If Q is a rational polyhedron, then

Q is integral if and only if any of the following statements hold:
(a) Each face of Q contains integral vectors.
(b) Each minimal face of Q contains integral vectors.
(c) max{c, x| x ∈ Q} is attained by an integral vector for each c for which
the maximum is finite.
(d) max{c, x| x ∈ Q} is an integer for each integral vector c for which
the maximum is finite.
Polyhedral Geometry and Linear Optimization 17

Corollary 1.1.64 If Q is a pointed polyhedron, then Q is integral if and

only if all vertices of Q are integral.

Proof. Let x0 be a vertex of Q. There is a supporting hyperplane H =

H(a, c) such that {x0 } = H ∩ Q and Q ⊂ H + . Since x0 is a convex
combination of points in Q ∩ Zn , it is seen that x0 is integral. The converse
follows from Theorem 1.1.63 and Proposition 1.1.41. 2

Proposition 1.1.65 Let Q ⊂ Rn be a rational polyhedron and let QI =

conv(Q ∩ Zn ) be its integral hull. Then
(a) QI is a polyhedron.
(b) If Q is a pointed polyhedron, then QI is integral.

Proof. We may assume that QI is non-empty. By the finite basis theorem

we can write Q = P + C, where C is a cone generated by integral vectors
α1 , . . . , αs and P is a polytope. Consider the linear map T : Rs → Rn
induced by T (ei ) = αi . Notice that B := T ([0, 1]s ) is a polytope whose
elements have the form λ1 α1 + · · · + λs αs , 0 ≤ λi ≤ 1 for all i. It is not hard
to see that QI = conv((P +B)∩Zn )+C. Since P +B is bounded, we get that
QI is a polyhedron by the finite basis theorem. This proves (a). To prove
part (b) observe that the vertices of QI are contained in (P + B) ∩ Zn , this
follows from Proposition 1.1.39. Hence by Corollary 1.1.64, QI is integral.
This proof was adapted from [372]. 2

Exercises
1.1.66 If f : Rq → Rn is an affine map, then f (x) = Ax + b, for some n × q
matrix A and some vector b.

1.1.67 Let A = {α0 , . . . , αr } be a sequence of distinct vectors in Rn . Prove:

(a) A is affinely independent ⇔ α1 − α0 , . . . , αr − α0 are linearly indepen-
dent ⇔ (α0 , 1), . . . , (αr , 1) are linearly independent.
(b) If A is contained in an affine hyperplane not containing the origin, then
A is affinely independent if and only if it is linearly independent.

1.1.68 If V ⊂ Rn , prove that V is an affine space in Rn if and only if

λ1 V + λ2 V ⊂ V for all λ1 , λ2 in R such that λ1 + λ2 = 1.

1.1.69 If C ⊂ Rn , prove that C is a convex cone if and only if λx + μy ∈ C

for all x, y ∈ C and λ, μ ≥ 0.

1.1.70 Let π1 : R2 → R be the projection π1 (x1 , x2 ) = x1 . Prove that π1 is

not a closed linear map.
Hint Consider the closed set {(x, 1/x)| x > 0}.
18 Chapter 1

1.1.71 Let Q be a polyhedral set in Rn and let f : Rn → Rm be a linear

function. Prove that f (Q) is a polyhedral set.

1.1.72 Let Q be a polyhedron in Rn and let f : Q → R be a linear function.

If f is bounded from above and Q contains no lines, prove that f attains
its maximum at a vertex of Q. Notice that f (Q) is a closed convex set.

1.1.73 Determine the facets of the convex polytope

P = conv(±e1 , . . . , ±en ) ⊂ Rn .

1.1.74 Let C be a polyhedral cone in Rn and let C ∗ be its dual cone. Then
C is a pointed cone if and only if dim(C ∗ ) = n.

1.1.75 If C = Rn and C ∗ is its dual cone, then C ∗ = (0).

1.1.76 If C = He+2 ⊂ R2 , then C = R+ e1 + R+ (−e1 ) + R+ e2 . Use the

duality theorem to show that C ∗ = R+ e2 .

1.1.77 If C = R+ (1, 1), prove that C = H(1,−1)

+
∩ H(−1,1)
+
∩ He+1 and that
∗ +
the dual cone is given by C = H(1,1) .

1.1.78 If C is a rational cone and CI := conv(C ∩ Zn ), then CI = C.

1.1.79 Let Q = {x| Ax ≤ b} = ∅ be a polyhedron. If Q = P + C, where P

is a polytope and C is a polyhedral cone, prove that

{y|Ay ≤ 0} = {y|x + y ∈ Q, ∀ x ∈ Q}.

The cone C = {y|Ay ≤ 0} is called the characteristic cone of Q.

1.1.80 Let Q = {x| Ax ≤ b} be a polyhedron. The lineality space of Q is

the linear space lin.space(Q) = {x| Ax = 0}. Prove that Q is pointed if and
only if lin.space(Q) = (0).

1.1.81 (Complementary slackness) Consider the LP primal-dual pair

min{x, c| x ≥ 0; xA ≥ b} = max{y, b| y ≥ 0; Ay ≤ c}.

Let x and y be feasible solutions, i.e., xA ≥ b, Ay ≤ c and x, y ≥ 0. Use

Theorem 1.1.56 to show that the following conditions are equivalent:
(a) x and y are both optimum solutions.
(b) x, c = y, b.
(c) (b − xA)y = 0 and x(c − Ay) = 0.
Polyhedral Geometry and Linear Optimization 19

1.1.82 (P. Gordan [194]) The system Ax < 0 is unsolvable if and only if the
system yA = 0, y ≥ 0, y = 0 is solvable. The vector inequality (ai ) < (bi )
means that ai < bi for all i.

In the following exercises K will denote an intermediate field Q ⊂ K ⊂ R.

1.1.83 Let H ⊂ Kn be a linear subspace and {v1 , v2 , . . . , vm } a basis of H.

Prove that H has an orthogonal basis using the Gram–Schmidt process:
1. Set u1 := v1 and u2 := v2 − ((v2 · u1 )/(u1 · u1 ))u1 . Verify that {u1 , u2 }
is an orthogonal set and Ku1 + Ku2 = Kv1 + Kv2 .
2. If u1 , u2 , . . . , uk (k < m) have been constructed so that they form an
orthogonal basis for the linear space generated by v1 , . . . , vk , setting
 
vk+1 · u1 vk+1 · ui vk+1 · uk
uk+1 = vk+1 − u1 + · · · + ui + · · · + uk ,
u1 · u1 ui · ui uk · uk

verify that {u1 , u2 , . . . , uk , uk+1 } is an orthogonal set and that this set
generates Kv1 + Kv2 + · · · + Kvk + Kvk+1 .

1.1.84 If A = (aij ) is a matrix with entries in K, then dim(RA ) is equal to

dim(im(A)), where RA is the row space of A and im(A) is the image of the
linear map defined by A.

1.1.85 If H is a linear subspace of Kn , then H ⊕ H ⊥ = Kn , where

H ⊥ = {x ∈ Kn | x · h = 0 for all h ∈ H}.

1.1.86 If A ∈ Mn×q (K), then

(im(A))⊥ = ker(At ), im(A) ⊕ ker(At ) = Kn , and (im(At ))⊥ = ker(A).

1.1.87 Let α, β, γ be three vectors in Kn such that α, β ≤ 0 and β, γ > 0.
Prove that the vector α = α − α,β  
γ,β γ is in K , α , β = 0, and α belongs
n

to the cone generated by α and γ.

1.1.88 Using the previous exercises show that the proof of Farkas’s lemma
given in [304, p. 86] is valid for any intermediate field Q ⊂ K ⊂ R.

1.1.89 Show that the proof of the fundamental theorem of linear inequali-
ties given in [372, Theorem 7.1] works for any intermediate field Q ⊂ K ⊂ R.

1.1.90 If A ∈ Mn×q (K) and γ ∈ Kn , then either there exists a vector

x ∈ Kq with Ax ≤ γ, or there exists a vector α ∈ Kn with α ≥ 0, αA = 0
and α, γ < 0 , but not both.
20 Chapter 1

1.2 Relative volumes of lattice polytopes

In this section we introduce relative volumes and unimodular coverings.
Two main references for relative volumes are [21, 146].
Let A = {v1 , . . . , vq } be a set of distinct vectors in Zn and let

P = conv(v1 , . . . , vq )

be the convex hull of A. A point in Zn is called a lattice point and P is

called a lattice polytope. Here a lattice point (resp. lattice polytope) is used
as a synonym of integral point (resp. integral polytope).
In the sequel to simplify the exposition and the proofs we set m = q − 1
and α0 = v1 , . . . , αm = vq .
If V = {0, α1 − α0 , . . . , αm − α0 } is the image of A under the translation
f
f : Rn −→ Rn , x −→ x − α0 ,

then aff(A) = α0 + RV. In particular, from this expression we get

d = dim(P) = dimR (RV),

where RV is the linear space spanned by V.

Theorem 1.2.1 [261, Theorem 3.7] Let Zn be the free Z-module of rank n.
Then any submodule of Zn is free of rank at most n.
Using this result, and the inclusions ZV ⊂ Zn ∩ RV ⊂ RV, it follows
that Zn ∩ RV is a free abelian group of rank d. A constructive proof of this
fact, which is useful for computing relative volumes, is included below in
the proof of Lemma 1.2.3.

Theorem 1.2.2 [332, Theorem II.9, pp. 26-27] Let A be an integral matrix
of rank d. Then there are invertible integral matrices U and Q such that

U −1 AQ = diag{s1 , . . . , sd , 0, . . . , 0},

si > 0 for 1 ≤ i ≤ d and si divides si+1 for 1 ≤ i ≤ d − 1.

The matrix D = diag{s1 , . . . , sd , 0, . . . , 0} is called the Smith normal

form of A and s1 , . . . , sd are called the invariant factors of A.

Lemma 1.2.3 RV ∩ Zn = Zγ1 ⊕ · · · ⊕ Zγd for some γ1 , . . . , γd in Zn .

Proof. Consider the matrix A of size n × m whose column vectors corre-

spond to the non-zero vectors in V. By Theorem 1.2.2, there are unimodular
integral matrices U and Q, of orders n and m, respectively, such that

U −1 AQ = D = diag{s1 , . . . , sd , 0, . . . , 0} (si = 0; si |si+1 ∀ i).

Polyhedral Geometry and Linear Optimization 21

Let u1 , . . . , un be the columns of U . We claim that the leftmost d columns

of U can serve as γ1 , . . . , γd . We regard some vectors as column vectors.
Take α ∈ RV ∩ Zn , thus α = Aβ for some β ∈ Rm . Set β  = Q−1 β and
denote the ith entry of β  by βi . Notice the equalities

α = U (Dβ  ) and U −1 α = Dβ  = (s1 β1 , . . . , sd βd , 0, . . . , 0)t .

Hence si βi ∈ Z for all i, and consequently α is in Zu1 + · · · + Zud . For the
reverse inclusion it suffices to prove that ui belongs to RV for i = 1, . . . , d.
Using the equality (AQ)ei = (U D)ei (ei =ith unit vector) we derive

Aqi = U (si ei ) = si (U ei ) = si ui (1 ≤ i ≤ d),

where qi = Qei is the ith column of Q. Hence ui ∈ RV for i = 1, . . . , d. 2

Lemma 1.2.4 There is an affine isomorphism φ : aff(A) → Rd such that

φ(aff(A) ∩ Zn ) = Zd , where d = dim(P).

Proof. By Lemma 1.2.3 one has RV ∩ Zn = Zγ1 ⊕ · · · ⊕ Zγd for some

γ1 , . . . , γd in Zn . Therefore there is a linear map
ψ
ψ : RV −→ Rd , γi −→ ei .

Hence φ = ψf : aff(A) −→ Rd satisfies the required condition. 2

Observe that φ(P) is an integral polytope of dimension d with a positive
Lebesgue measure denoted by m(φ(P)).

Definition 1.2.5 The relative volume of the integral polytope P is:

vol(P) := m(φ(P)).

If P has dimension n, then the relative volume of P agrees with its usual
volume. To see that the relative volume is independent of φ we need the
following fact about the standard volume.

Theorem 1.2.6 [427, Theorem 6.2.14] Let T : Rn → Rn be an affine map

given by T (x) = Ax + x0 , where A is a real matrix of order n and x0 ∈ Rn .
If P is a bounded convex set in Rn , then T (P) has volume |det(A)|vol(P).

Lemma 1.2.7 Let T : Zn → Zn be the affine map given by T (x) = Ax + β,

where A is an integral matrix of order n and β ∈ Zn . If T is bijective, then
det(A) = ±1.

Proof. Note that the matrix A determines a bijective linear map, thus A
has an inverse with integral entries. Indeed, for each i there is a column
vector βi in Zn such that Aβi = ei . Therefore A (β1 · · · βn ) = I, and
consequently det(A) = ±1. 2
22 Chapter 1

Proposition 1.2.8 The relative volume of P is independent of φ.

Proof. If φ1 is another affine isomorphism φ1 : aff(A) → Rd such that

φ1 (aff(A) ∩ Zn ) = Zd , then the affine map φ1 φ−1 : Zd −→ Zd is bijective.
Using Theorem 1.2.6 and Lemma 1.2.7 one obtains

vol(φ(P)) = vol((φ1 φ−1 )(φ(P))) = vol(φ1 (P)). 2

In practice one can compute volumes of lattice polytopes using Normaliz

[68] and Polyprob [104].

Example 1.2.9 We give an illustration in R3 of the procedure outlined

above to compute relative volumes. We begin by setting

A = {v1 , v2 , v3 , v4 } = {(1, 3, 1), (1, 5, 3), (4, 9, 4), (2, 8, 5)}

and P = conv(A). Then V = {(0, 0, 0), (0, 2, 2), (3, 6, 3), (1, 5, 4)} and
f
f : R3 −→ R3 , x −→ x − α0 ,

where α0 = (1, 3, 1). Consider the matrix

⎡ ⎤
0 3 1
A = ⎣ 2 6 5 ⎦.
2 3 4

Next using Maple [80] to compute the Smith normal form of A we obtain
invertible integral matrices U and Q such that U AQ = D, where
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0 0 −2 1 0 0 1 0
D = ⎣ 0 1 0 ⎦; U =⎣ 1 0 0 ⎦; U −1 = ⎣ 1 2 0 ⎦.
0 0 0 1 −1 1 1 1 1

By the proof of Lemma 1.2.3, the first two columns of U −1 form a Z-basis
for RV ∩Z3 . Using the affine map ψ : RV → R3 induced by ψ(0, 1, 1) = (1, 0)
and ψ(1, 2, 1) = (0, 1), and the map φ = ψf : aff(A) → R, we get

φ(P) = conv((0, 0), (2, 0), (0, 3), (3, 1)) and vol(P) = m(φ(P)) = 11/2.

Proposition 1.2.10 ([176], [394, Proposition 4.6.30]) The relative volume

of P is given by:
|Zn ∩ iP|
vol(P) = lim .
i→∞ id
Definition 1.2.11 For an abelian group (M, +) its torsion subgroup T (M )
is the set of all x in M such that px = 0 for some 0 = p ∈ N. The group M
is torsion-free if T (M ) = (0).
Polyhedral Geometry and Linear Optimization 23

Lemma 1.2.12 If B = {β1 , . . . , βs } ⊂ Zn , then

(a) (RB ∩ Zn )/ZB = T (Zn /ZB), and (∗)

(b) RB ∩ Zn = ZB if and only if Zn /ZB is torsion-free.

Proof. (a): If z ∈ T (Zn /ZB) ⊂ Zn /ZB, then kz ∈ ZB for some 0 = k ∈ N,

so z ∈ QB ⊂ RB. Since also z ∈ Zn , z is in the left-hand side of Eq. (∗).
This shows the inclusion “⊃”. The other inclusion is left as an exercise.
(b): This follows from (a). 2

Lemma 1.2.13 [397, pp. 32-33] If H ⊂ G are free abelian groups  of the
same rank d with Z-bases δ1 , . . . , δd and γ1 , . . . , γd related by δi = j gij γj ,
where gij ∈ Z for all i, j, then |G/H| = | det(gij )|.

Theorem 1.2.14 [146] If A = {v1 , . . . , vq } ⊂ Zn is a set of vectors lying

on an affine hyperplane not containing the origin and P = conv(A) has
dimension d, then

|ZA ∩ iP|
vol(P) = |T (Zn /(v2 − v1 , . . . , vq − v1 ))| lim .
i→∞ id
Proof. As in the beginning of this section for convenience of notation we set
m = q − 1 and α0 = v1 , . . . , αm = vq . Let x0 ∈ Qn be such that αi , x0  = 1
for all i. As αi − α0 , x0  = 0, there is a decomposition ZA = Zα0 ⊕ ZV
with V = {α1 − α0 , . . . , αm − α0 }. Pick δ1 , . . . , δd ∈ Zn such that

ZV = Zδ1 ⊕ · · · ⊕ Zδd . (1.5)

Therefore one can write

αi = α0 + fi1 δ1 + · · · + fid δd (fij ∈ Z). (1.6)

Consider the lattice polytope

P1 = conv((1, 0, . . . , 0), (1, f11 , . . . , f1d ), . . . , (1, fm1 , . . . , fmd )) ⊂ Rd+1 .

ϕ
There is a bijective map ZA∩iP −→ Zd+1 ∩iP1 , namely ϕ is the restriction
of the linear map from ZA to Zd+1 that maps each vector into its coordinate
vector with respect to the basis {α0 , δ1 , . . . , δd }. Therefore

|ZA ∩ iP| |Zd+1 ∩ iP1 |

lim = lim = vol(P1 ).
i→∞ id i→∞ id
To estimate vol(P1 ) consider the lattice polytope

P2 = conv((0, . . . , 0), (f11 , . . . , f1d ), . . . , (fm1 , . . . , fmd )) ⊂ Rd

24 Chapter 1

and note that applying the translation α → α − e1 to P1 gives:

vol(P1 ) = vol(conv((0, . . . , 0), (0, f11 , . . . , f1d ), . . . , (0, fm1 , . . . , fmd ))).

Thus vol(P1 ) = vol(P2 ). The next step is to relate vol(P2 ) with vol(P) by
obtaining a second expression for vol(P2 ). According to Lemma 1.2.3, there
are γ1 , . . . , γd ∈ Zn such that

RV ∩ Zn = Zγ1 ⊕ · · · ⊕ Zγd . (1.7)

Writing

δi = gi1 γ1 + · · · + gid γd (gij ∈ Z; i, j = 1, . . . , d), (1.8)

from Eqs. (1.6) and (1.8) one derives the matrix equality

(fij )(gij ) = (cij ), (1.9)

where αi − α0 = ci1 γ1 + · · · + cid γd , 1 ≤ i ≤ m. By definition

vol(P) = vol(conv((0, . . . , 0), (c11 , . . . , c1d ), . . . , (cm1 , . . . , cmd ))).

By Eq. (1.9), the linear transformation

σ
σ : Zd −→ Zd , x −→ x(gij ),

satisfies σ(fi1 , . . . , fid ) = (ci1 , . . . , cid ). Hence by Theorem 1.2.6 we get

vol(P) = vol(σ(P2 )) = | det(gij )|vol(P2 ).

To finish the proof it suffices to show that | det(gij )| is the order of the
torsion subgroup of Zn /ZV. Now we have
(a) (b)
T (Zn /ZV) = (RV ∩ Zn )/ZV = (Zγ1 ⊕ · · · ⊕ Zγd )/(Zδ1 ⊕ · · · ⊕ Zδd ),

where in (a) we use Lemma 1.2.12, and in (b) the identities (1.5) and (1.7).
Hence, from the identity (1.8) and Lemma 1.2.13, we get

|T (Zn /ZV)| = | det(gij )|. 2

Proposition 1.2.15 Let α0 , . . . , αn be a set of affinely independent points

in Rn and Δ = conv(α0 , . . . , αn ). Then the volume of the simplex Δ is
 ⎛ ⎞  ⎛ ⎞
 α0 1   α1 − α0 
 
 ⎜ . .. ⎟  ⎜ .. ⎟
det ⎝ .. . ⎠ det ⎝ . ⎠
  
 αn 1   αn − α0 
vol(Δ) = = .
n! n!
Polyhedral Geometry and Linear Optimization 25

Proof. The result follows using linear algebra or applying the change of
variables formula. See Theorem 1.2.6. 2

Definition 1.2.16 A set Δ in Rn is called a lattice d-simplex if Δ is the

convex hull of a set of d + 1 affinely independent points in Zn .

Let Δ be a lattice d-simplex in Rn . The normalized volume of Δ is

defined as d!vol(Δ). If d = n, then from Proposition 1.2.15 one has
1
vol(Δ) ≥ .
n!
Definition 1.2.17 A lattice d-simplex Δ in Rn is called unimodular if
d!vol(Δ) = 1.

Proposition 1.2.18 Let Δ = conv(α0 , . . . , αd ) be a lattice d-simplex. If

R(α1 − α0 , . . . , αd − α0 ) ∩ Zn = Zγ1 ⊕ · · · ⊕ Zγd


for some γ1 , . . . , γd ∈ Zn and αi − α0 = j cij γj , then

| det(cij )|
vol(Δ) = .
d!
Proof. Note vol(Δ) = vol(conv(0, c1 , . . . , cd )), where ci = (ci1 , . . . , cin ).
Hence the formula follows from Proposition 1.2.15. 2

Lemma 1.2.19 If β1 , . . . , βm ∈ Zn , then

R(β1 , . . . βm ) ∩ Zn = Zβ1 + · · · + Zβm + Zδ1 + · · · + Zδs ,

where δ1 , . . . , δs generate the torsion subgroup of Zn /(β1 , . . . , βm ).

Proof. It is straightforward and will be left as an exercise. 2

Lemma 1.2.20 If α0 , . . . , αm ∈ Zn , then there is an isomorphism of groups
 
ϕ : T (Zn /(α1 − α0 , . . . , αm − α0 )) −→ T Zn+1 /((α0 , 1), . . . , (αm , 1)) ,

given by ϕ(α) = (α, 0). Here T (M ) denotes the torsion subgroup of M .

Proof. The map ϕ is clearly well-defined and injective. To prove that ϕ

is onto consider an equation s(α, b) = λ0 (α0 , 1) + · · · + λm (αm , 1), where
λi ∈ Z, 0 = s ∈ N, α ∈ Zn and b ∈ Z. Then

sα = λ0 α0 + · · · + λm αm ,
sb = λ0 + · · · + λm .

Hence s(α − bα0 ) = λ1 (α1 − α0 ) + · · · + λm (αm − α0 ). It follows readily

that ϕ(α − bα0 ) = (α, b), as required. 2
26 Chapter 1

Proposition 1.2.21 Let α = α0 , . . . , αd be a set of affinely independent

vectors in Zn defining a simplex Δ. Then the following are equivalent:

(a) d!vol(Δ) = 1.

(b) R(α1 − α0 , . . . , αd − α0 ) ∩ Zn = Z(α1 − α0 ) + · · · + Z(αd − α0 ).

(c) Zn /Z{α1 − α0 , . . . , αd − α0 } is torsion-free.

(d) Zn+1 /Z{(α0 , 1), . . . , (αd , 1)} is torsion-free.

Proof. (a) ⇔ (b) follows from Proposition 1.2.18 and (b) ⇔ (c) ⇔ (d)
follows from Lemmas 1.2.19 and 1.2.20. 2

Corollary 1.2.22 If Δ = conv(α0 , α1 , . . . , αd ) is a unimodular lattice d-

simplex in Rn , then Δ ∩ Zn = {α0 , α1 , . . . , αd }.

Proof. Let α ∈ Δ ∩ Zn . Then α = λ0 α0 + · · · + λd αd , where λi ∈ Q+ for

all i and λ0 + · · · + λd = 1. Using the equality

α − α0 = λ1 (α1 − α0 ) + · · · + λd (αd − α0 )

and Proposition 1.2.21(c) we get

α − α0 = η1 (α1 − α0 ) + · · · + ηd (αd − α0 )
= λ1 (α1 − α0 ) + · · · + λd (αd − α0 ),

where ηi ∈ Z for all i. Since α1 − α0 , . . . , αd − α0 are linearly independent

we obtain λi = ηi for all i. Hence exactly one of the λi ’s is equal to 1 and
the others are equal to zero, that is, α ∈ {α0 , α1 , . . . , αd }. 2

Definition 1.2.23 A lattice polytope P = conv(A) of dimension d is said

to have a unimodular covering with support in A if there are simplices
Δ1 , . . . , Δm of dimension d such that the following two conditions hold
(i) P = ∪m
i=1 Δi and the vertices of Δi are contained in A for all i.

(ii) d!vol(Δi ) = 1.
If condition (ii) is replaced by
(ii)’ [ZA : ZAi ] = 1 for all i,

where Ai denotes the vertex set of Δi , then we say that Δ1 , . . . , Δm is a

weakly unimodular covering of P with support in A.
For a discussion on the existence of unimodular covers of rational cones
see [60] and the references therein.
Polyhedral Geometry and Linear Optimization 27

Proposition 1.2.24 If A = {v1 , . . . , vq } ⊂ Zn and the vectors in A lie in

a hyperplane of Rn not containing the origin, then any unimodular covering
of P = conv(A) with support in A is weakly unimodular.

Proof. Let Δi = conv(Ai ) be any unimodular simplex of dimension d with

vertex set Ai ⊂ A and d = dim(P). For convenience of notation assume
that Ai = {v1 , . . . , vd+1 }. To show the equality ZA = ZAi it suffices to
prove ZA ⊂ ZAi . Take an arbitrary vector α in ZA and write:

α = η1 v1 + · · · + ηq vq (ηi ∈ Z),
(α, η) = η1 (v1 , 1) + · · · + ηq (vq , 1) (η = η1 + · · · + ηq ).

Since A lies in an affine hyperplane not containing the origin and Ai is

affinely independent, the set Ai is seen to be linearly independent. Therefore

d + 1 = dim(R{(v1 , 1), . . . , (vd+1 , 1)}) = dim(R{(v1 , 1), . . . , (vq , 1)}).

Thus one can write (α, η) = λ1 (v1 , 1) + · · · + λd+1 (vd+1 , 1), λi ∈ Q. By

Proposition 1.2.21 the group Zn+1 /Z{(v1 , 1), . . . , (vd+1 , 1)} is torsion-free.
Hence it follows from Lemma 1.2.12(b) that λi ∈ Z for all i, and conse-
quently α is in ZAi . 2

Example 1.2.25 Let A = {v1 , . . . , v5 } be the set of vectors in Z6 given by

v1 = (1, 0, 1, 0, 0, 1), v2 = (1, 0, 0, 1, 1, 0), v3 = (0, 1, 1, 0, 1, 0),

v4 = (0, 1, 0, 1, 0, 1), v5 = (1, 0, 0, 1, 0, 1).

Consider A1 = {v2 , v3 , v4 , v5 }, A2 = {v1 , v3 , v4 , v5 }, A3 = {v1 , v2 , v3 , v5 },

Δ1 = conv(A1 ), Δ2 = conv(A2 ), Δ3 = conv(A3 ),
and P = conv(A). The groups Z6 /ZA and Z6 /ZAi are torsion-free for
all i. This readily implies that ZA = ZAi , i = 1, 2, 3. Using the relation
v1 + v2 + v4 = v3 + 2v5 it is not hard to see that P = Δ1 ∪ Δ2 ∪ Δ3 . Thus
Δ1 , Δ2 , Δ3 is a weakly unimodular covering. The relative volume of P can
be computed using the procedure described before. It is left to the reader
to verify that dim(P) = 3 and vol(P) = 1/2.

Exercises
1.2.26 Let Δ be an n-dimensional simplex in Rn with vertices α0 , . . . , αn
in Zn . Prove that Δ is unimodular if and only if any of the following two
equivalent conditions hold
(a) Zn+1 = Z(α0 , 1) + · · · + Z(αn , 1).
(b) Zn = Z(α1 − α0 ) + · · · + Z(αn − α0 ).
28 Chapter 1

1.2.27 Construct a lattice simplex of normalized volume greater than 1.

Then prove that the converse of Proposition 1.2.24 fails.

1.2.28 If P is the integral polytope with vertices (0, 0) and (1, 3):

y
6
3 s


2 

1 


s -
0 1 2 x

Prove that the relative volume of P is equal to 1.

1.2.29 If P = conv((3, 1), (1, 3)), then vol(P) = 2.

1.2.30 Consider the set A = {e1 + e2 , e2 + e3 , e3 + e4 , e1 + e4 } and the

polytope P = conv(A). Prove that dim(P) = 2 and vol(P) = 1.

1.2.31 Let P be the tetrahedron in R3 whose vertices are v1 = (1, 1, 0),

v2 = (2, 1, −1), v3 = (2, −5, 0), v4 = (7, 1, −8). Prove that vol(P) = 2.
Then verify this using Maple [80].

1.2.32 If A = {v1 , . . . , vq } ⊂ Zn , prove that the rank of RA ∩ Zn , as a free

abelian group, is equal to the dimension of RA as a real vector space.

1.3 Hilbert bases and TDI systems

The set of integral points of a rational polyhedral cone form a semigroup that
arises in many branches of mathematics, such as combinatorial commutative
algebra [65], toric varieties [176], and integer programming [372].
A finite set H ⊂ Rn is called a Hilbert basis if

Zn ∩ R+ H = NH,

where NH is the semigroup spanned by H consisting of all linear combina-

tions of H with coefficients in N. Notice that all vectors in a Hilbert basis
are integral. A nice introduction to Hilbert bases can be found in [215].
Let C ⊂ Rn be a rational polyhedral cone. A finite set H is called a
Hilbert basis of C if C = R+ H and H is a Hilbert basis. A Hilbert basis of
C is minimal if it does not strictly contain any other Hilbert basis of C.
Polyhedral Geometry and Linear Optimization 29

Hilbert bases of rational polyhedral cones always exist. For a pointed

cone there is only one minimal Hilbert basis. To prove these two facts, we
need Gordan’s lemma. Below we give two versions of this lemma.

Definition 1.3.1 Given a = (ai ) ∈ Rn , we define the value of a as:

|a| = a1 + · · · + an .

Lemma 1.3.2 (Gordan, version 1) If A = {v1 , . . . , vq } ⊂ Zn and ZA is

the subgroup generated by A, then there exist γ1 , . . . , γm in Zn such that

ZA ∩ R+ A = Nγ1 + · · · + Nγm

and γi ∈ [N, M ]n for all i, where N = −q max |vi− | and M = q max |vi+ |.
1≤i≤q 1≤i≤q

Proof. Recall that C = ZA ∩ R+ A = ZA ∩ Q+ A; see Corollary 1.1.27. Let

β ∈ C. Then one can write
q 
 
xi
β= vi ,
i=1
yi

where xi ∈ N and 0 = yi ∈ N. By the division algorithm there are ri , ni in

N such that xi = ni yi + ri and 0 ≤ ri < yi . Therefore one can write


q 
q
β= ni vi + ai vi ,
i=1 i=1
q
with ai ∈ [0, 1] ∩ Q. As i=1 ai vi ∈ C ∩ [N, M ]n , the set A ∪ (C ∩ [N, M ]n )
is a generating set for C with the required property. Our argument was
based on the proof of Gordan’s lemma given in [65]. 2

Definition 1.3.3 A semigroup (S, +, 0) of Zn is said to be finitely generated

if there exists a finite set Γ = {γ1 , . . . , γr } ⊂ S such that:

S = NΓ := Nγ1 + · · · + Nγr .

A set of generators Γ of S is called minimal if γi = 0 ∀i and none of its

elements is a linear combination with coefficients in N of the others.

Remark 1.3.4 There are examples of subsemigroups of Nn , with n ≥ 2,

which are not finitely generated; see Exercise 1.3.34.

Lemma 1.3.5 (Gordan, version 2) If R+ A is a cone in Rn generated by

a finite set A ⊂ Zn , then the semigroup Zn ∩ R+ A is finitely generated.
30 Chapter 1

Proof. If A = {v1 , . . . , vq }, consider the finite set of integral points:

{a1 v1 + · · · + aq vq | 0 ≤ ai ≤ 1} ∩ Zn = {γ1 , . . . , γr }.

It is left to the reader to prove that γ1 , . . . , γr is the required set of generators

for Zn ∩ R+ A. See [372, p. 233]. 2
Let A be a finite set in Zn and let G = Zn or G = ZA. Then, by
Gordan’s lemma (versions 1 and 2) there exists γ1 , . . . , γr ∈ Zn such that:

G ∩ R+ A = Nγ1 + · · · + Nγr .

Computing the γi ’s is in general difficult [69]. Fortunately, the γi ’s can be

computed using Normaliz [68].
Therefore Hilbert bases of rational polyhedral cones always exist:

Proposition 1.3.6 Let A be a finite set in Zn . Then there exist γ1 , . . . , γr

such that
R+ A ∩ Zn = Nγ1 + · · · + Nγr ,
and H = {γ1 , . . . , γr } is a Hilbert basis of R+ A.

Proof. The existence follows from Lemma 1.3.5. That H is a Hilbert basis
of R+ A follows from the equality R+ A = R+ γ1 + · · · + R+ γr . 2
In Definition 1.1.52 we considered pointed polyhedra. For cones we have
the following equivalent definition (see Exercise 1.3.36).

Definition 1.3.7 A polyhedral cone C = {x| Ax ≤ 0} is called pointed if

the lineality space {x| Ax = 0} is equal to {0}.

Lemma 1.3.8 Let C be a rational polyhedral cone. If C is pointed, then

there exists an integral vector b such that b, x > 0 for all 0 = x ∈ C.

Proof. There is a rational matrix A such that C = {x| Ax ≤ 0}; see

Theorem 1.1.29. We may assume that A is integral. If u1 , . . . , ut are the
rows of A, then b = −u1 − · · · − ut satisfies the required condition. 2

Theorem 1.3.9 [371] Let A be a finite set in Zn and let C = R+ A. If C

is pointed, then there exists a unique minimal Hilbert basis of C given by

H = {x ∈ C ∩ Zn | 0 = x ∈
/ Ny1 + Ny2 ; ∀y1 , y2 ∈ (C \ {0}) ∩ Zn }.

Proof. Let Γ = {γ1 , . . . , γr } be an arbitrary Hilbert basis of C.

We claim
r
that H ⊂ Γ. Let x ∈ H. Since Zn ∩ C = NΓ we can write x = i=1 ai γi ,
ai ∈ N. Thus, by construction of H, we get x = γi for some i, which proves
the claim. To finish the proof it suffices to prove NH = NΓ, because this
Polyhedral Geometry and Linear Optimization 31

equality implies R+ H = R+ Γ and consequently Zn ∩ R+ H = NH. We

proceed by contradiction by assuming that the set:

V = {x| x ∈ NΓ = C ∩ Zn ; x ∈
/ NH}

is not empty. By Lemma 1.3.8 there exists b ∈ Zn such that b, x > 0 for
all 0 = x ∈ C. Let x0 ∈ V such that

x0 , b = min{x, b| x ∈ V}.

Since x0 ∈ / H we can write x0 = x1 + x2 with x1 , x2 in (C \ {0}) ∩ Zn . Thus,

since xi , b < x0 , b for i = 1, 2, we have x1 , x2 ∈ NH and x1 + x2 ∈ NH,
a contradiction. 2

Definition 1.3.10 (Edmonds and Giles [125]) A rational system xA ≤ b

of linear inequalities is called totally dual integral, abbreviated TDI, if the
minimum in the LP-duality equation

max{x, c| xA ≤ b} = min{y, b| y ≥ 0; Ay = c} (1.10)

has an integral optimum solution y for each integral vector c for which the
minimum is finite.

Note that there are rational systems of linear inequalities which define
the same polyhedron and such that one system is totally dual integral while
the other is not (see Exercise 1.3.39).
TDI systems occur in the theory of Gröbner bases of toric ideals [252],
in the theory of perfect graphs [296, 86], and in combinatorial commutative
algebra [147, 185, 188]. See Theorems 9.6.21, 13.6.8, and 14.3.6.

Proposition 1.3.11 [372, Corollary 22.1c] If xA ≤ b is a TDI-system and

b is integral, then the polyhedron {x | xA ≤ b} is integral.

Proof. By hypothesis min{y, b| y ≥ 0; Ay = c} has an integral optimum

solution y for each integral vector c for which the minimum is finite. Then
by the LP duality theorem (see Theorem 1.1.57), we get that

max{x, c| xA ≤ b}

is an integer for each integral vector c for which the maximum is finite.
Thus the polyhedron {x | xA ≤ b} is integral by Theorem 1.1.63. 2

Definition 1.3.12 A set P in Rn is called a parallelotope if it is the image

of [0, 1]n under a non-singular linear transformation, i.e., P has the form

P = {λ1 v1 + · · · + λn vn | 0 ≤ λi ≤ 1}

for some linearly independent vectors v1 , . . . , vn in Rn .

32 Chapter 1

Lemma 1.3.13 Let v1 , . . . , vn be a basis of Rn and let

P = {λ1 v1 + · · · + λn vn | 0 ≤ λi ≤ 1}
be a parallelotope. Then vol(P) = | det[v1 , . . . , vn ]|.
Proof. Since P is the image of [0, 1]n under the linear map T induced by
ei → vi , the formula follows from Theorem 1.2.6. 2
Proposition 1.3.14 Let A be an integral matrix of order n × n and let
A = {v1 , . . . , vn } be the set of columns of A. If det(A) = 0, then A is a
Hilbert basis if and only if | det(A)| = 1.
Proof. ⇒) Let Q = [0, 1]n be the unit cube and let P be the parallelotope
P = {λ1 v1 + · · · + λn vn | 0 ≤ λi ≤ 1}.
By Lemma 1.3.13 one has vol(P) = | det(A)|. As A is a Hilbert basis and
A is linearly independent, we have
(k + 1)n = |kQ ∩ Zn | = |kP ∩ Zn |, ∀ k ∈ N.
Therefore by Proposition 1.2.10 and the fact that [0, 1]n has volume 1, we
conclude that 1 = vol(P) = | det(A)|, as required.
⇐) Let α be a vector in R+ A ∩ Zn . As | det(A)| = 1 and since A is
linearly independent, by Cramer’s rule, it follows that α ∈ NA. Thus A is
a Hilbert basis. 2
Notation Let A = (0) be an integral matrix. The greatest common divisor
of all the non-zero r × r sub determinants of A will be denoted by Δr (A).
Theorem 1.3.15 [261, Theorem 3.9] Let A be an integral matrix of rank r
and let d1 , . . . , dr be the invariant factors of A. Then
d1 = Δ1 (A), d2 = Δ2 (A)Δ1 (A)−1 , . . . , dr = Δr (A)Δr−1 (A)−1 .
The next result is called the fundamental structure theorem for finitely
generated abelian groups.
Theorem 1.3.16 [261, pp. 187-188] Let M be a finitely generated Z-module
with a presentation M  Zn /(a1 , . . . , aq ). If A is the matrix of rank r with
columns a1 , . . . , aq and d1 , . . . , dr are the invariant factors of A, then
M  Z/(d1 ) ⊕ Z/(d2 ) ⊕ · · · ⊕ Z/(dr ) ⊕ Zn−r .
Lemma 1.3.17 If A is an integral matrix of size n × q and rank r, then
Δr (A) = |T (Zn /(a1 , . . . , aq ))|,
where ai is the ith column of A. In particular Δr (A) = 1 if and only if the
quotient group Zn /(a1 , . . . , aq ) is torsion-free.
Polyhedral Geometry and Linear Optimization 33

Proof. Let d1 , . . . , dr be the invariant factors of the matrix A. On one

hand using Theorem 1.3.15, we get the equality d1 d2 · · · dr = Δr (A). On
the other hand, by Theorem 1.3.16, we obtain an isomorphism

T (Zn /(a1 , . . . , aq ))  Z/(d1 ) ⊕ · · · ⊕ Z/(dr ).

Therefore the order of the torsion subgroup is Δr (A), as required. 2

Lemma 1.3.18 Let A = {v1 , . . . , vd } ⊂ Zn be a set of linearly independent

vectors and let A be the matrix with column vectors v1 , . . . , vd . Then

d!vol(conv(0, v1 , . . . , vd )) = Δd (A).

Proof. Let Γ = {γ1 , . . . , γd } be a set of vectors such that

RA ∩ Zn = Zγ1 ⊕ · · · ⊕ Zγd , (1.11)

see Lemma 1.2.3. Then we can write

vi = ci1 γ1 + · · · + cid γd (i = 1, . . . , d) (1.12)

where C = (cij ) is an integral matrix. By Proposition 1.2.18, we have

d!vol(conv(0, v1 , . . . , vd )) = | det(C)|.

Using Eqs. (1.11) and (1.12), we get that the map

T (Zn /ZA) −→ T (Zd /Z{c1 , . . . , cd }), α → [α]Γ ,

is an isomorphism of groups, where ci is the ith row of C and [α]Γ is the

coordinate vector of α in the basis Γ. Hence, by Lemma 1.3.17, we get

Δd (A) = |T (Zn /ZA)| = |T (Zd /Z{c1 , . . . , cd })| = | det(C)|. 2

Proposition 1.3.19 Let A = {v1 , . . . , vd } ⊂ Zn be a linearly independent

set and let A be the matrix with column vectors v1 , . . . , vd . Then A is a
Hilbert basis if and only if Δd (A) = 1.
Proof. ⇒) Let C = (cij ) be as in the proof of Lemma 1.3.18. As A is
a Hilbert basis, it is seen that the rows of C form a Hilbert basis. Then
| det(C)| = 1 by Proposition 1.3.14. Therefore Δd (A) = 1 by Lemma 1.3.18.
⇐) If Δd (A) = 1, then Zn /ZA is torsion-free. Since the vi ’s are linearly
independent, it follows readily that A is a Hilbert basis. 2

Theorem 1.3.20 [179] Let A = {v1 , . . . , vq } ⊂ Zn be a Hilbert basis and

let r = rank (ZA). If R+ A is pointed, then there exist H ⊂ A such that H
is linearly independent, H is a Hilbert basis and |H| = r.
34 Chapter 1

Corollary 1.3.21 If A = {v1 , . . . , vq } ⊂ Zn and R+ A is a pointed cone,

then A is a Hilbert basis if and only if R+ A ∩ ZA = NA and Zn /ZA is a
torsion-free group.

Proof. Assume that A is a Hilbert basis. Let A be the matrix with column
vectors v1 , . . . , vq and let r = rank (A). Then clearly R+ A ∩ ZA = NA. By
Theorem 1.3.20 and Proposition 1.3.19, we get Δr (A) = 1. Thus Zn /ZA
is a torsion-free group. The converse follows readily. Notice that for the
converse the hypothesis that the cone R+ A is pointed is not needed. 2
Let P = {x | xA ≤ b} be a rational polyhedron and let F be a face of P.
A column of A is active in F if the corresponding inequality in xA ≤ b is
satisfied with equality for all vectors x in F . A minimal face of P is a face
not containing any other face. A face F of P is minimal if and only if F is
an affine subspace [372].

Theorem 1.3.22 [372, Theorem 22.5] A rational system xA ≤ b is TDI if

and only if for each minimal face F of the polyhedron P = {x | xA ≤ b}, the
columns of A which are active in F form a Hilbert basis.

Proof. ⇒) Let v1 , . . . , vq be the column vectors of A and let b = (bi ).

Assume that v1 , . . . , vr are the columns of A which are active in F . Take an
integral vector c in R+ {v1 , . . . , vr }. Then the maximum in the LP-duality
equation
max{x, c| xA ≤ b} = min{y, b| y ≥ 0; Ay = c} (1.13)
is attained by each vector of F . Indeed if x0 ∈ F and c = λ1 v1 + · · · + λr vr ,
with λi ≥ 0 for all i, then x0 , c = λ1 b1 + · · · + λr br and

min{y, b| y ≥ 0; Ay = c} ≤ λ1 b1 + · · · + λr br ≤ max{x, c| xA ≤ b}.

Thus we have equality everywhere and the maximum is attained at any

vector of F . The minimum has an integral optimum solution y. Then by
complementary slackness (see Corollary 1.1.58), we get (b − xA)y = 0 for
any x ∈ F . Therefore (bi − x, vi )yi = 0 for i = 1, . . . , q. If vi is inactive in
F , then x, vi  < bi for some x ∈ F . Then yi = 0. This proves that yi = 0
for i > r, i.e., c ∈ N{v1 , . . . , vr }. Altogether v1 , . . . , vr is a Hilbert basis.
⇐) Let c be an integral vector for which the optimum values of Eq. (1.13)
are finite. Let F be a face of P such that each vector in F attains the
maximum in Eq. (1.13). We may assume that F is a minimal face of P.
By the minimality of F if a vector x0 in F satisfies x0 , vi  = bi for some
i, then any other vector x in F satisfies x0 , vi  = bi , i.e., vi is active in F .
Thus we may assume that there is 1 ≤ r ≤ q such that x, vi  = bi for i ≤ r
and x ∈ F , and x, vi  < bi for i > r and x ∈ F . Thus by hypothesis the
set A1 = {v1 , . . . , vr } is a Hilbert basis. Let x and y = (yi ) be optimum
solutions of Eq. (1.13) with x ∈ F . Then by complementary slackness (see
Polyhedral Geometry and Linear Optimization 35

Corollary 1.1.58), we get (b − xA)y = 0, Ay = c and y ≥ 0. If yi > 0, then

x, vi  = bi , i.e., i ≤ r. Therefore c is in R+ A1  . Hence c is in NA1 because
r
A1 is a Hilbert basis. Then we can write c = i=1 ηi vi , ηi ∈ N. Consider
the vector y0 = (η1 , . . . , ηr , 0 . . . , 0) in Nq whose last q − r entries are equal
to zero. Thus c = Ay0 . Since bi −x, vi  = 0 for i ≤ r, we get (b−xA)y0 = 0.
Thus the minimum in Eq. (1.13) is attained in y0 , as required. 2
Notice that we have in fact shown that Theorem 1.3.22 is valid if we
replace “each minimal face F ” by “each face F ”.

Definition 1.3.23 Let A be an integral matrix and let w be an integral

vector. The system xA ≤ w is said to have the integer rounding property if

min{y, w| y ≥ 0; Ay = a} = min{y, w| Ay = a; y ≥ 0 ; y integral}

for each integral vector a for which min{y, w| y ≥ 0; Ay = a} is finite.

Theorem 1.3.24 [372, Theorem 22.18] Let A be an integral matrix of size

n × q with column vectors v1 , . . . , vq and w = (wi ) ∈ Zq . Then the system
xA ≤ w has the integer rounding property if and only if the set

H = {(v1 , w1 ), . . . , (vq , wq ), en+1 } ⊂ Zn+1

is a Hilbert basis.

Corollary 1.3.25 Let A be an integral matrix and let w = (wi ) be an

integral vector. The system xA ≤ w is TDI if and only if {x| xA ≤ w} is
an integral polyhedron and the set H ⊂ Zn+1 is a Hilbert basis.

Proof. It follows from Proposition 1.3.11, Theorem 1.3.24, and using the
definition of a TDI system. 2

Theorem 1.3.26 Let A be an integer matrix with column vectors v1 , . . . , vq

and let w = (wi ) be an integral vector. If the polyhedron P = {x| xA ≤ w} is
integral and H = {(v1 , w1 ), . . . , (vq , wq )} is a Hilbert basis, then the system
xA ≤ w is TDI.

Proof. Let F be a minimal face of P. Recall that a column of A is active

in F if the corresponding inequality in xA ≤ w is satisfied with equality for
all vectors in F . We may assume that v1 , . . . , vr are the columns of A which
are active in F . Then x, vi  = wi for x ∈ F and 1 ≤ i ≤ r.
If y, vi  < wi for some y ∈ F , then x, vi  < wi for any other x ∈ F .
Indeed if x, vi  = wi for some x ∈ F , consider the supporting hyperplane
of P given by H = {x|x, vi  = wi }, then x ∈ F ∩ H  F because y ∈ F
and y ∈ / F ∩ H, a contradiction to the minimality of the face F . Thus we
may also assume that x, vi  < wi for x ∈ F and i > r.
36 Chapter 1

Since P is integral, by Theorem 1.1.63, each face of P contains integral

vectors. Pick an integral vector x0 ∈ F . By Theorem 1.3.22, it suffices to
prove that B = {v1 , . . . , vr } is a Hilbert basis. Take a ∈ R+ B ∩ Zn . Then
a = λ1 v1 + · · · + λr vr with λi ≥ 0 for all i. Thus we have
b = a, x0  = λ1 v1 , x0  + · · · + λr vr , x0 
= λ1 w1 + · · · + λr wr .
r
Hence b is an integer andq (a, b) = i=1 λi (vi , wi ). As H is a Hilbert basis,
we can write (a, b) = i=1 ηi (vi , wi ), ηi ∈ N for all i. Therefore on one
hand 0 = (a, b), (x0 , −1). On the other hand (a, b), (x0 , −1) is equal to

r 
q
ηi (vi , wi ), (x0 , −1) + ηi (vi , wi ), (x0 , −1) .
 
i=1 i=r+1
=0 <0

Hence ηi = 0 for i > r and a = η1 v1 + · · · + ηr vr . Thus a ∈ NB. 2

The converse is not true in general. However there are some interesting
linear systems where the converse holds.

Definition 1.3.27 Let A be an integral matrix and w an integral vector.

The system x ≥ 0; xA ≤ w is TDI if the minimum in the LP-duality equation
max{a, x| x ≥ 0; xA ≤ w} = min{y, w| y ≥ 0; Ay ≥ a} (1.14)
has an integral optimum solution y for each integral vector a with finite
minimum.
Proposition 1.3.28 Let A be a nonnegative integral matrix of size n × q
with column vectors v1 , . . . , vq and let w = (wi ) ∈ Nq . Then, the system
x ≥ 0; xA ≤ w is TDI if and only if the polyhedron P = {x| x ≥ 0; xA ≤ w}
is integral and H = {(v1 , w1 ), . . . , (vq , wq ), −e1 , . . . , −en } is a Hilbert basis.
Proof. Assume the system x ≥ 0; xA ≤ w is TDI. By Proposition 1.3.11,
we get that P is integral. Next we prove that H is a Hilbert basis. Take
(a, b) ∈ R+ H ∩Zn+1 , where a ∈ Zn and b ∈ Z. By hypothesis, the minimum
in Eq. (1.14) has an integral optimum solution y = (yi ) such that y, w ≤ b.
Since y ≥ 0 and a ≤ Ay, we can write
a = y1 v1 + · · · + yq vq − δ1 e1 − · · · − δn en (δi ∈ N) =⇒
(a, b) = y1 (v1 , w1 ) + · · · + yq−1 (vq−1 , wq−1 )
+(yq + b − y, w)(vq , wq ) − (b − y, w)vq − δ,

where δ = (δi ). As the entries of A are in N, the vector −vq can be written
as a nonnegative integer combination of −e1 , . . . , −en . Thus (a, b) ∈ NH.
This proves that H is a Hilbert basis. The converse follows readily from
Theorem 1.3.26. 2
Polyhedral Geometry and Linear Optimization 37

Lemma 1.3.29 Let A be an integral matrix of size n×q with column vectors
v1 , . . . , vq and w = (wi ) ∈ Zq . If P = {x| xA ≤ w} is pointed, then α is a
−
vertex of P if and only if H(α,−1) is a closed halfspace defining a facet of
the cone C = R+ {(v1 , w1 ), . . . , (vq , wq ), en+1 }.

Proof. Since P is pointed, the rank of A is equal to n and C is a cone

of full dimension n + 1. Using that α is a vertex of P if and only if α is
a basic feasible solution of the system xA ≤ w (see Corollary 1.1.49), the
proof follows from Theorem 1.1.44. 2

Procedure 1.3.30 If P = {x| xA ≤ w} is pointed, using Corollary 1.3.25

and Lemma 1.3.29 we can check whether a system xA ≤ w is TDI. Using
Normaliz [68] this can be achieved using a single input file of the form:
q+1
n+1
v1,w1
...
vq,wq
0 1
0
The first and third block of the output file contain the Hilbert basis and
the support hyperplanes. To verify if P is integral, by Lemma 1.3.29, it
suffices to verify that all rows of the support hyperplanes having its last
entry positive are integral and its last entry is equal to 1.

Example 1.3.31 To illustrate Procedure 1.3.30 consider the integer vector

w = (3, 0, 0, 1, 0) and the matrix
⎛ ⎞
1 1 1 1 1
A=⎝ 0 1 0 1 2 ⎠
0 0 1 1 2

We now verify that the system xA ≤ w is not TDI. Using the following
input file for Normaliz
6
4
1 0 0 3
1 1 0 0
1 0 1 0
1 1 1 1
1 2 2 0
0 0 0 1
0
38 Chapter 1

We obtain the following output file (we only show the first and third block):
6 Hilbert basis elements: 6 support hyperplanes:
1 0 0 3 -3 3 3 1
1 1 0 0 0 0 0 1
1 0 1 0 0 0 1 0
0 0 0 1 0 1 0 0
1 2 2 0 2 -2 1 0
1 1 1 0 2 1 -2 0
Thus the polyhedron {x| xA ≤ w} is integral and the rows in the input file
do not form a Hilbert basis. For instance the vector (1, 1, 1, 0) is in the cone
generated by the rows of the input file but it is not an N-linear combination
of the rows.

Corollary 1.3.32 Let xA ≤ w be a system satisfying the integer rounding

property and let A = {v1 , . . . , vq } be the set of column vectors of A. Then
(a) A is a Hilbert basis.
(b) Zn /ZA is a torsion-free group provided that R+ A is a pointed cone.
q
Proof. (a): Take a ∈ R+ A ∩ Zn , then we can write a = i=1 λi vi , for
some λ1 , . . . , λq in R+ . Hence

(a,  i λi wi ) = λ1 (v1 , w1 ) + · · · + λq (vq , wq ) + δ(0, 1),

where δ ≥ 0. Therefore, by Theorem 1.3.24, there are λ1 , . . . λq ∈ N and

δ  ∈ N such that

(a,  i wi λi ) = λ1 (v1 , w1 ) + · · · + λq (vq , wq ) + δ  (0, 1),

Thus a ∈ NA. (b): It follows at once from (a) and Corollary 1.3.21. 2

Exercises
1.3.33 Prove that any subsemigroup of N is finitely generated.

1.3.34 Prove that the subsemigroup of N2 given by S = (N+ × N) ∪ {(0, 0)}

is not finitely generated, where N+ = {1, 2, . . .}.

1.3.35 (Gordan’s lemma) Let L be a lattice in Rn , i.e., L is an additive

subgroup of Zn , and let C be a rational polyhedral cone in Rn . Prove that
L ∩ C is a finitely generated semigroup.
Hint Adapt the proof of Lemma 1.3.2.

1.3.36 Let C = {x| Ax ≤ 0} be a rational polyhedral cone. Prove that C

contains no lines if and only if {x| Ax = 0} = (0).
Polyhedral Geometry and Linear Optimization 39

1.3.37 Let A be an integral matrix. If C = {x| x ≥ 0; Ax = 0}, prove that

C is a rational polyhedral cone.

1.3.38 If w = (3, 0, 0, −1, 0) and A is the matrix

⎛ ⎞
1 1 1 1 1
A=⎝ 0 1 0 1 1 ⎠,
0 0 1 1 2

use Normaliz to verify that the system xA ≤ w is TDI.

1.3.39 Consider the two systems

⎛ ⎞
⎛ ⎞ 2 ⎛ ⎞ ⎛ ⎞
1 0 1 1 ⎜ 2 ⎟ 1 0 1 2
(x1 , x2 , x3 ) ⎝ 1 1 0 1 ⎠≤⎜ ⎟ ; (x1 , x2 , x3 ) ⎝ 1
⎝ 2 ⎠ 1 0 ⎠≤⎝ 2 ⎠
0 1 1 1 0 1 1 2
3

Prove that they define the same polyhedron. Then prove that the first
system is TDI but the second is not. Use Normaliz to verify these assertions.

1.3.40 Let A be an n × d integral matrix of rank d whose set of column

vectors A = {v1 , . . . , vd } form a Hilbert basis. Prove that the system xA ≤ b
is TDI for each integral vector b.

1.4 Rees cones and clutters

In this section we characterize the max-flow min-cut property of clutters in
terms of Hilbert bases of Rees cones and the integrality of polyhedra.
Let A be an n × q matrix with entries in N, let A = {v1 , . . . , vq } be the
set of columns of A, and let

A := {e1 , . . . , en , (v1 , 1), . . . , (vq , 1)} ⊂ Rn+1 ,

where ei is the ith unit vector. The cone R+ A , generated by A , is called

the Rees cone of A or the Rees cone of A.
The term Rees cone was coined in [147] because this cone encodes some
information about the Rees algebra of the monomial ideal associated to A;
see Chapters 12–14. The first aim of this section is to study the irreducible
representation, as an intersection of closed halfspaces, of a Rees cone.
We shall always assume that all rows and columns of A are non-zero. The
set covering polyhedron (in Proposition 13.1.2 we clarify this terminology)
is by definition the rational polyhedron:

Q(A) := {x ∈ Rn | x ≥ 0, xA ≥ 1},
40 Chapter 1

where 0 and 1 are vectors whose entries are equal to 0 and 1, respectively.
Often we denote the vectors 0, 1 simply by 0, 1. As is seen below, one can
express the irreducible representation of R+ A in terms of Q(A).
Some of the facets of a Rees cone are easy to identify. Consider the
index set

J = {i | 1 ≤ i ≤ n and ei , vj  = 0 for some j} ∪ {n + 1}.

Notice that R+ A has dimension n + 1. It is not hard to see that the set

F = R+ A ∩ Hei (1 ≤ i ≤ n + 1)

defines a facet of R+ A if and only i ∈ J . Therefore, by Theorem 1.1.44

and Proposition 1.1.51, the Rees cone has the following unique irreducible
representation ! ! r
R+ A = He+i Hα+i (1.15)
i∈J i=1

such that 0 = αi ∈ Qn+1 and αi , en+1  = −1 for all i. In many interesting
cases, from the viewpoint of commutative algebra, one has the equality
J = {1, . . . , n}; see Section 14.2 and Exercise 14.2.34.

Lemma 1.4.1 Let A be a matrix with entries in N and let a = (ai1 , . . . , aiq )
be its ith row. Set k = min{aij | 1 ≤ j ≤ q}. If aij > 0 for all j, then ei /k
is a vertex of Q(A).

Proof. Set x0 = ei /k. Clearly x0 ∈ Q(A), x0 , vj  = 1 for some j, and

x0 , e  = 0 for  = i. It is seen that x0 is a vertex of Q(A). 2
The irreducible representation of a Rees cone can be expressed in terms
of the vertices of the set covering polyhedron.

Theorem 1.4.2 Let V be the vertex set of Q(A). Then the irreducible
representation of the Rees cone is given by
! !
R+ A = He+i +
H(u,−1) .
i∈J u∈V

Proof. We set B = {ei | i ∈ J } ∪ {(u, −1)| u ∈ V } and V = {u1 , . . . , up }.

First we dualize Eq. (1.15) and use the duality theorem for cones to obtain

(R+ A )∗ = {y ∈ Rn+1 | y, x ≥ 0, ∀ x ∈ R+ A }

= He+1 ∩ · · · ∩ He+n ∩ H(v
+
1 ,1)
∩ · · · ∩ H(v
+
q ,1)

= R+ ei + R+ α1 + · · · + R+ αr . (1.16)
i∈J
Polyhedral Geometry and Linear Optimization 41

Next we show the equality (R+ A )∗ = R+ B. The right-hand side is clearly

contained in the left-hand side because a vector α belongs to Q(A) if and
only if (α, −1) is in (R+ A )∗ . To prove the reverse inclusion observe that by
Eq. (1.16) it suffices to show that αk ∈ R+ B for all k. Writing αk = (ck , −1)
and using αk ∈ (R+ A )∗ gives ck ∈ Q(A). The set covering polyhedron can
be written as
Q(A) = R+ e1 + · · · + R+ en + conv(V ),
where conv(V ) denotes the convex hull of V, this follows from the structure
of polyhedra (see Theorem 1.1.33) by noticing that the characteristic cone
of Q(A) is precisely Rn+ . Thus we can write
ck = λ1 e1 + · · · + λn en + μ1 u1 + · · · + μp up ,
where λi ≥ 0, μj ≥ 0 for all i, j and μ1 +· · ·+μp = 1. If 1 ≤ i ≤ n and i ∈
/ J,
then the ith row of A has all its entries positive. Thus by Lemma 1.4.1 we
get that ei /ki is a vertex of Q(A) for some ki > 0. To avoid cumbersome
notation we denote ei and (ei , 0) simply by ei , from the context the meaning
of ei should be clear. Therefore from the equalities
     !
  ei ei 
λi ei = λi ki = λi ki , −1 + λi ki en+1
ki ki
i∈J
/ i∈J
/ i∈J
/ i∈J
/

we conclude that i∈J
/ λi ei is in R+ B. From the identities
αk = (ck , −1) = λ1 e1 + · · · + λn en + μ1 (u1 , −1) + · · · + μp (up , −1)
  p
= λi ei + λi ei + μi (ui , −1)
i∈J
/ i∈J \{n+1} i=1

we obtain αk ∈ R+ B, as required. Taking duals in (R+ A )∗ = R+ B yields

R+ A = Ha+ . (1.17)
a∈B

Thus, by Remark 1.1.31, the proof reduces to showing that β ∈ / R+ (B \ {β})

for all β ∈ B. To show this we will assume that β ∈ R+ (B \ {β}) for some
β ∈ B and derive a contradiction.
Case (I): β = (uj , −1). For simplicity assume β = (up , −1). Then

 
p−1
(up , −1) = λi ei + μj (uj , −1), (λi ≥ 0; μj ≥ 0) ⇒
i∈J j=1

 
p−1
up = λi ei + μj u j (1.18)
i∈J \{n+1} j=1

−1 = λn+1 − (μ1 + · · · + μp−1 ). (1.19)

42 Chapter 1

To derive a contradiction we claim that Q(A) = Rn+ + conv(u1 , . . . , up−1 ),

which is impossible because by Proposition 1.1.39 the vertices of Q(A) would
be contained in {u1 , . . . , up−1 }. To prove the claim note that the right-hand
side is clearly contained in the left-hand side. For the other inclusion take
γ ∈ Q(A) and write


n 
p 
p
γ = bi ei + ci u i (bi , ci ≥ 0; ci = 1)
i=1 i=1 i=1

(1.18) 
p−1
= δ+ (ci + cp μi )ui (δ ∈ Rn+ ).
i=1

Therefore using the inequality

!

p−1 
p−1 
p−1
(1.19)
(ci + cp μi ) = ci + cp μi = (1 − cp ) + cp (1 + λn+1 ) ≥ 1
i=1 i=1 i=1

we get γ ∈ Rn+ + conv(u1 , . . . , up−1 ). This proves the claim.

Case (II): β = ek for some k ∈ J . We only consider the subcase k ≤ n.
The subcase k = n + 1 can be treated similarly. We can write

 
p
ek = λi ei + μi (ui , −1), (λi ≥ 0; μi ≥ 0).
i∈J \{k} i=1

p p
From this equality we get ek = i=1 μi ui . Hence ek A ≥ ( i=1 μi )1 > 0, a
contradiction because k ∈ J and ek , vj  = 0 for some j. 2
As a consequence we obtain:

Theorem 1.4.3 [188, Theorem 3.2] The mapping ϕ : Qn → Qn+1 given by

ϕ(α) = (α, −1) induces a bijective mapping

ϕ : V −→ {α1 , . . . , αr }

between the set V of vertices of Q(A) and the set {α1 , . . . , αr } of vectors
that occur in the irreducible representation of R+ A given in Eq. (1.15).

Definition 1.4.4 Let A be a matrix with entries in {0, 1}. The system
x ≥ 0; xA ≥ 1 is called totally dual integral (TDI) if the maximum in the
LP-duality equation

min{α, x| x ≥ 0; xA ≥ 1} = max{y, 1| y ≥ 0; Ay ≤ α}

has an integral optimum solution y for each integral vector α with finite
maximum.
Polyhedral Geometry and Linear Optimization 43

Proposition 1.4.5 If the system x ≥ 0; xA ≥ 1 is TDI, then Q(A) has

only integral vertices.

Proof. It follows from Theorem 1.1.63. 2

Definition 1.4.6 A clutter C with vertex set X = {x1 , . . . , xn } is a family

of subsets of X, called edges, none of which is included in another. The set
of vertices and edges of C are denoted by V (C) and E(C), respectively.

Let C be a clutter with vertex set X = {x1 , . . . , xn } and let f1 , . . . , fq

be the edges of C. The incidence matrix of C is the n × q matrix A = (aij )
given by aij = 1 if xi ∈ fj and aij = 0 otherwise.

Definition 1.4.7 A clutter C, with incidence matrix A, has the max-flow

min-cut (MFMC) property if both sides of the LP-duality equation

min{α, x| x ≥ 0; xA ≥ 1} = max{y, 1| y ≥ 0; Ay ≤ α} (1.20)

have integral optimum solutions x, y for each nonnegative integral vector α.

It turns out that the max-flow min-cut property is equivalent to require

that a certain blowup ring has no nilpotent elements; see Theorem 14.3.6.

Proposition 1.4.8 Let C be a clutter and let A be its incidence matrix.

The system x ≥ 0; xA ≥ 1 is TDI if and only if C has the max-flow min-cut
property.

Proof. If follows from Propositions 1.4.5 and 1.1.41. 2

Let C be a clutter with n vertices and let A = {v1 , . . . , vq } be the set of
columns of its incidence matrix. For use below we denote by A the Rees
configuration of C:

A = {e1 , . . . , en , (v1 , 1), . . . , (vq , 1)} ⊂ Nn+1 ,

where ei is the ith unit vector of Rn

Theorem 1.4.9 [188, Theorem 3.4] If C is a clutter and A is its incidence

matrix, then C has the max-flow min-cut property if and only if Q(A) is an
integral polyhedron and A is a Hilbert basis of R+ A .

Proof. ⇒) By Proposition 1.4.5 the polyhedron Q(A) is integral. Next we

show that A is an integral Hilbert basis. Take (α, αn+1 ) ∈ Zn+1 ∩ R+ A .
Then Ay ≤ α and y, 1 = αn+1 for some vector y ≥ 0. Therefore one
concludes that the optimal value of the linear program

max{y, 1| y ≥ 0; Ay ≤ α}
44 Chapter 1

is greater than or equal to αn+1 . Since C has the MFMC property, this
linear program has an optimal integral solution y0 . By Exercise 1.4.12,
there exists an integral vector y0 such that

0 ≤ y0 ≤ y0 and |y0 | = αn+1 .

Therefore
         
α A  A  α A
= y0 + (y0 − y0 ) + − y0
αn+1 1 0 0 0

and (α, αn+1 ) ∈ NA , as required.

⇐) Assume that C does not satisfy the MFMC property. Then there is
an α0 ∈ Nn such that if y0 is an optimal solution of the linear program:

max{y, 1| y ≥ 0; Ay ≤ α0 }, (∗)

then y0 is not integral. We claim that also the optimal value |y0 | = y0 , 1
of this linear program is not integral. If |y0 | is integral, then (α0 , |y0 |) is
in Zn+1 ∩ R+ A . As A is a Hilbert basis we get that (α0 , |y0 |) is in NA ,
but this readily yields that the linear program (∗) has an integral optimal
solution, a contradiction. This completes the proof of the claim.
Now, consider the dual linear program:

min{x, α0 | x ≥ 0, xA ≥ 1}.

By Proposition 1.1.41 the optimal value of this linear program is attained

at a vertex x0 of Q(A). Then by the duality theorem (Theorem 1.1.56) we
get x0 , α0  = |y0 | ∈
/ Z. Hence x0 is not integral, a contradiction to the
integrality of the set covering polyhedron Q(A). 2

Remark 1.4.10 Normaliz [68] computes the irreducible representation and

the minimal integral Hilbert basis of a Rees cone. Thus we can effectively
use Theorems 1.4.2 and 1.4.9 to determine whether a given clutter has the
max-flow min-cut property; see example below.

Example 1.4.11 Let A be the transpose of the matrix:

⎡ ⎤
1 0 0 0 1
⎢ 0 1 0 1 0 ⎥
⎢ ⎥
⎣ 0 0 1 1 1 ⎦
1 1 1 0 0

and let C be the clutter with incidence matrix A. Consider the following
input file for Normaliz that we call “example.in”:
Polyhedral Geometry and Linear Optimization 45

4
5
1 0 0 0 1
0 1 0 1 0
0 0 1 1 1
1 1 1 0 0
3
Applying Normaliz , i.e., typing “normaliz example” in the directory where
one has the file normaliz.exe we get the output file “example.out”:
9 generators of integral closure of Rees algebra:
1 0 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0
0 0 0 1 0 0
0 0 0 0 1 0
1 0 0 0 1 1
0 1 0 1 0 1
0 0 1 1 1 1
1 1 1 0 0 1
10 support hyperplanes:
0 0 1 1 1 -1
1 0 0 0 0 0
0 1 0 0 0 0
0 0 0 0 0 1
0 0 1 0 0 0
1 0 0 1 0 -1
0 0 0 1 0 0
0 0 0 0 1 0
0 1 0 0 1 -1
1 1 1 0 0 -1
semigroup is not homogeneous
The first block shows that A is a Hilbert basis for the Rees cone. The second
block shows the irreducible representation of the Rees cone of A, thus using
Theorem 1.4.2 we obtain that Q(A) is integral. Altogether Theorem 1.4.9
proves that the clutter C has the max-flow min-cut property.

Exercises
1.4.12 Let b, η1 , . . . , ηq be a sequence in N. If η1 + · · · + ηq ≥ b, then there
are 1 , . . . , q ∈ N such that 0 ≤ i ≤ ηi for all i and 1 + · · · + q = b.
1.4.13 Let A be the incidence matrix of a clutter C and let v1 , . . . , vq be its
column vectors. The system x ≥ 0; xA ≥ 1 is TDI if and only if the set
Au = {vi | vi , u = 1} ∪ {ei | ei , u = 0}
is a Hilbert basis for any vertex u of Q(A).
46 Chapter 1

1.5 The integral closure of a semigroup

Let A = {v1 , . . . , vq } be a set of vectors in Nn \ {0}. The integral closure or
normalization of the affine semigroup

NA := Nv1 + · · · + Nvq ⊂ Nn ,

is defined as NA := ZA ∩ R+ A, where ZA is the subgroup of Zn generated

by A. The semigroup NA is called normal or integrally closed if NA = NA.
The Rees semigroup of A is by definition NA ⊂ Nn+1 , where

A = {(v1 , 1), . . . , (vq , 1), e1 , . . . , en }.

Notice that en+1 = (v1 , 1) − (v1 , 1), e1 e1 − · · · − (v1 , 1), en en . Hence
ZA = Zn+1 . As a consequence we get

NA = R+ A ∩ Zn+1 .

From this equality we obtain a characterization of the normality of a Rees

semigroup in terms of Hilbert bases:

Proposition 1.5.1 NA is normal if and only if A is a Hilbert basis.

The algebraic invariants of semigroup rings of semigroups generated by

a Hilbert basis are easier to understand; see [405, Section 6]. This family of
semigroups and their cones are studied throughout this book.

Proposition 1.5.2 If A is a Hilbert basis, then NA is normal.

Proof. Since NA ⊂ R+ A ∩ ZA ⊂ R+ A ∩ Zn = NA we obtain the equality

NA = R+ A ∩ ZA. 2

Example 1.5.3 If A = {(1, 1, 0), (0, 1, 1), (1, 0, 1)}, the affine semigroup
NA is normal because A is linearly independent, but A is not a Hilbert
basis because NA  Z3 ∩ R+ A (see Exercise 1.5.8).

Notation Given β = (βi ) ∈ Rn+1 , we define degn+1 (β) = βn+1 .

Proposition 1.5.4 [147] Let R+ A be the Rees cone of A and let B be its
minimal integral Hilbert basis. If β ∈ B, then degn+1 (β) < n.

Proof. We set β = (α, b), with α ∈ Nn and b = degn+1 (β) ∈ N. Note

that R+ A ∩ Zn+1 = Q+ A ∩ Zn+1 and dim(R+ A ) = n + 1. Thus, since
(α, b) ∈ R+ A , by Carathéodory’s theorem (Theorem 1.1.18) we can write

(α, b) = λ1 (vi1 , 1) + · · ·+ λr (vir , 1) + μ1 ej1 + · · · + μs ejs (λi , μk ∈ Q+ ), (∗)

Polyhedral Geometry and Linear Optimization 47

where {(vi1 , 1), . . . , (vir , 1), ej1 , . . . , ejs } is a linearly independent subset of
A . By the minimality of B it is seen that 0 ≤ λi < 1 and 0 ≤ μk < 1 for all
i, k. From Eq. (∗) we get b = λ1 + · · · + λr < r. Thus b ≤ r − 1. If r ≤ n,
then b ≤ n − 1. Thus from now on we may assume r = n + 1 and Eq. (∗)
takes the simpler form
(α, b) = λ1 (vi1 , 1) + · · · + λn+1 (vin+1 , 1).
Consider the cone C generated by A = {(vi1 , 1), . . . , (vin+1 , 1)}. Since −e1
is not in C, using Exercise 1.5.9, we obtain a point
x0 = (1 − λ0 )(α, b) + λ0 (−e1 ) (0 ≤ λ0 < 1)
in the relative boundary of C. By Theorem 1.1.44 the relative boundary
of C is the union of its facets. Hence using that any facet of C is an
n-dimensional cone generated by a subset of A (see Proposition 1.1.23)
together with the minimality of B it follows that we can write (α, b) as:
(α, b) = ρ0 e1 + ρ1 (vj1 , 1) + · · · + ρn (vjn , 1) (0 ≤ ρi < 1 ∀i),
and consequently b ≤ n − 1, as required. 2
Definition 1.5.5 A matrix A is called totally unimodular if each i × i
subdeterminant of A is 0 or ±1 for all i ≥ 1.
Examples of totally unimodular matrices include incidence matrices of
clutters without odd cycles [27, Chapter 5], incidence matrices of bipartite
graphs and digraphs [372, p. 274], and network matrices [372, Chapter 19].
Theorem 1.5.6 (Hoffman, Kruskal) If A is a totally unimodular matrix,
then Q = {x | x ≥ 0; Ax ≤ b} is integral for each integral vector b.
Proof. Assume that A has size n × q. Notice that Q contains no lines, thus
it suffices to show that Q has only integral vertices. Let x0 be a vertex of
Q. By Corollary 1.1.47, A x0 = b for some subsystem A x ≤ b of
   
A b
≤
−Iq 0
such that A is a square matrix of order q with linearly independent rows.
By hypothesis det(A ) = ±1, hence x0 is integral because the inverse of A
is an integral matrix by Cramer’s rule. 2
Corollary 1.5.7 [49] If A = (aij ) is a totally unimodular {0, 1}-matrix of
size n × q with column vectors v1 , . . . , vq , then R+ A is normal.
Proof. By the Hoffman–Kruskal theorem, the system x ≥ 0; xA ≥ 1 is
TDI. Thus, by Theorem 1.4.9, the set A is a Hilbert basis, i.e., the Rees
semigroup R+ A is normal. 2
48 Chapter 1

Exercises
1.5.8 If A = {(1, 1, 0), (0, 1, 1), (1, 0, 1)}, use Normaliz , with option 0, to
verify that the minimal integral Hilbert basis of R+ A is A ∪ {(1, 1, 1)}.

1.5.9 Let a be a point of a set V in Rn and let x be a point in aff(V) not

in V. If λ0 = sup{λ ∈ [0, 1]| (1 − λ)a + λx ∈ V}, then x0 = (1 − λ0 )a + λ0 x
is a relative boundary point of V lying between a and x.

1.5.10 If A is a totally unimodular matrix and A is obtained from A by

adding columns of unit vectors, then the matrix A is totally unimodular.

1.6 Unimodularity of matrices and normality

In this section, we characterize when a vector belongs to a subgroup of Zn
and prove Heger’s theorem for the existence of solutions of an integer linear
system. Then we show the normality of semigroups arising from unimodular
matrices and unimodular coverings.
A minor of order r (r-minor for short) of a matrix A is defined as the
determinant of a square submatrix of A of order r.

Definition 1.6.1 An integral matrix A = (0) is t-unimodular if all the non-

zero r-minors of A have absolute value equal to t, where r is the rank of A.
If A is 1-unimodular, we say that A is a unimodular matrix .

Theorem 1.6.2 [372, Theorem 19.2] Let A be an integral matrix of full row
rank. Then the polyhedron {x| x ≥ 0; Ax = b} is integral for each integral
vector b if and only if A is unimodular.

Theorem 1.6.3 Let A be an n × q integral matrix whose set of column

vectors is A = {v1 , . . . , vq } and let b ∈ Zn be a column vector such that
rank(A) = rank([A b]). The following conditions are equivalent:
(a) b ∈ ZA.
(b) Zn /ZA and Zn /Z(A ∪ {b}) have the same invariant factors.
(c) The matrices [A 0] and [A b] have the same Smith normal form.
(d) Δr (A) = Δr ([A b]), where r = rank(A) and Δr (A) is the gcd of all
the non-zero r-minors of A.
q
Proof. There is k ∈ N \ {0} such that kb = i=1 λi vi , λi ∈ Z. Therefore,
there is a canonical epimorphism of finite groups
ϕ
ϕ : T (Zn /ZA) −→ T (Zn /ZB) (α + ZA −→ α + ZB),

where B = A ∪ {b}. Notice that ϕ is injective if and only if (a) holds.

Polyhedral Geometry and Linear Optimization 49

The implication (a) ⇒ (b) is straightforward. There are invertible inte-

gral matrices Pi and Qi such that

D1 = Q1 [A 0]P1 = diag(d1 , . . . , dr , 0, . . . , 0),

D2 = Q2 [A b]P2 = diag(e1 , . . . , er , 0, . . . , 0),

are the Smith normal forms of [A 0] and [A b], respectively; that is, di , ei are
positive integers satisfying that di divides di+1 and ei divides ei+1 for all i.
By the fundamental structure theorem for finitely generated abelian groups
(see Theorem 1.3.16) there are isomorphisms:

T (Zn /ZA)  (Z/d1 Z) × · · · × (Z/dr Z),

T (Zn /ZB)  (Z/e1 Z) × · · · × (Z/er Z).

Thus (b) ⇔ (c). Note Δi (A) = d1 · · · di and Δi ([A b]) = e1 · · · ei for all i.
Hence (c) ⇒ (d). To prove (d) ⇒ (i) observe |T (Zn /ZA)| = |T (Zn /ZB)|
and consequently ϕ must be injective. 2
As an immediate consequence of Theorem 1.6.3 we get:

Theorem 1.6.4 (I. Heger, [372, p. 51]) Let A be an n × q integral matrix

and let b ∈ Zn be a column vector. If rank(A) = rank([A b]), then the system
Ax = b has an integral solution if and only if Δr (A) = Δr ([A b]).

Remark 1.6.5 Let A ⊂ Zn be a finite set. Note that the finite basis
theorem together with Theorem 1.6.3 yield a membership test to decide
when a given vector b in Zn belongs to NA = ZA ∩ R+ A.

Proposition 1.6.6 If A is a t-unimodular matrix with columns v1 , . . . , vq

and vi1 , . . . , vir is a Q-basis for the column space of A, then

ZA = Zvi1 ⊕ · · · ⊕ Zvir

Proof. For each j one has Δr (vi1 · · · vir ) = Δr (vi1 · · · vir vj ) = t. Then by
Theorem 1.6.3, we get vj ∈ Zvi1 ⊕ · · · ⊕ Zvir , as required. 2

Theorem 1.6.7 If A is a t-unimodular matrix whose set of columns is A,

then the affine semigroup NA is normal.

Proof. If b ∈ ZA ∩ R+ A, by Theorem 1.1.18 and Proposition 1.6.6 there

are vi1 , . . . , vir linearly independent vectors in A, r = rank(A), such that

b ∈ R+ vi1 + · · · + R+ vir and b ∈ Zvi1 + · · · + Zvir . (1.21)

By comparing the coefficients of b with respect to the two representations

obtained from Eq. (1.21), one derives b ∈ NA. 2
50 Chapter 1

Corollary 1.6.8 If A is a totally unimodular matrix whose set of columns

is A, then NA = Zn ∩ R+ A; that is, A is a Hilbert basis.

Proof. Let A = {v1 , . . . , vq } be the set of column vectors of A. Take

b ∈ Zn ∩ R+ A, then by Carathéodory’s theorem r(see Theorem 1.1.18) and
after permutation of the vi ’s we can write b = i=1 ηi vi with ηi ≥ 0 for all
i, where r is the rank of A and v1 , . . . , vr are linearly independent.
The submatrix A = (v1 · · · vr ) is totally unimodular. Therefore, by
Theorem 1.6.3, the system of equations A x = b has an integral solution.
Thus b is a linear combination of v1 , . . . , vr with coefficients in Z. It follows
that ηi ∈ N for all i, that is, b ∈ NA. The other inclusion is clear. 2

Proposition 1.6.9 If A = {v1 , . . . , vq } ⊂ Zn lie in a hyperplane of Rn not

containing the origin and P = conv(A) has a weakly unimodular covering
with support in A, then NA is a normal semigroup.

Proof. Take 0 = β ∈ R+ A ∩ ZA. There exists λ1 . . . , λq ∈ Q+ such that

β = λ1 v1 +· · ·+λq vq . We set |λ| = λ1 +· · ·+λq . Since β/|λ| is in P and using
that P has a weakly unimodular covering, there is an affinely independent
set A1 ⊂ A defining a lattice simplex Δ = conv(A1 ) such that β/|λ| ∈ Δ
and ZA1 = ZA. Altogether β belongs to ZA1 ∩ R+ A1 . Note that A1 is a
linearly independent set because A1 lie in a hyperplane not containing the
origin. Therefore it follows rapidly that β ∈ NA1 , as required. 2

Exercises
1.6.10 If A = {v1 , . . . , vq } ⊂ Zn and P = conv(A) has a unimodular
covering with support in A, then N(v1 , 1) + · · · + N(vq , 1) is normal.

1.7 Normaliz, a computer program

Throughout this book we will frequently use Normaliz [68], a program that
provides an invaluable effective tool to study monomial subrings and their
algebraic invariants. This program computes the following:

• normalization (or integral closure) of an affine semigroup;

• Hilbert basis of a pointed rational cone;
• lattice points of an integral polytope;
• support hyperplanes of a Rees cone;
• generators of the integral closure of the Rees algebra of a monomial
ideal I = (xv1 , . . . , xvq ) ⊂ R = K[x1 , . . . , xn ];
• generators of the integral closure of I;