Multiobjective Linear Programming An

Introduction 1st Edition Dinh The Luc Auth

download

https://ebookbell.com/product/multiobjective-linear-programming-
an-introduction-1st-edition-dinh-the-luc-auth-5235464

Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.

Multiobjective Linear And Integer Programming 1st Edition Carlos

Henggeler Antunes

https://ebookbell.com/product/multiobjective-linear-and-integer-
programming-1st-edition-carlos-henggeler-antunes-5483384

Linear And Multiobjective Programming With Fuzzy Stochastic Extensions

1st Edition Masatoshi Sakawa

https://ebookbell.com/product/linear-and-multiobjective-programming-
with-fuzzy-stochastic-extensions-1st-edition-masatoshi-sakawa-4594000

Multiobjective Optimization Algorithms For Bioinformatics Anirban

Mukhopadhyay

https://ebookbell.com/product/multiobjective-optimization-algorithms-
for-bioinformatics-anirban-mukhopadhyay-57514188

Multiobjective Programming And Goal Programming Theoretical Results

And Practical Applications 1st Edition Gabriele Eichfelder Auth

https://ebookbell.com/product/multiobjective-programming-and-goal-
programming-theoretical-results-and-practical-applications-1st-
edition-gabriele-eichfelder-auth-2040824
Multiobjective Problem Solving From Nature From Concepts To
Applications Knowles

https://ebookbell.com/product/multiobjective-problem-solving-from-
nature-from-concepts-to-applications-knowles-22105846

Multiobjective Genetic Algorithms For Clustering Applications In Data

Mining And Bioinformatics 1st Edition Ujjwal Maulik

https://ebookbell.com/product/multiobjective-genetic-algorithms-for-
clustering-applications-in-data-mining-and-bioinformatics-1st-edition-
ujjwal-maulik-2472458

Multiobjective Shape Design In Electricity And Magnetism 1st Edition

Paolo Di Barba Auth

https://ebookbell.com/product/multiobjective-shape-design-in-
electricity-and-magnetism-1st-edition-paolo-di-barba-auth-4194252

Multiobjective Optimization Principles And Case Studies 1st Edition

Phd Yann Collette

https://ebookbell.com/product/multiobjective-optimization-principles-
and-case-studies-1st-edition-phd-yann-collette-4229806

Multiobjective Evolutionary Algorithms And Applications 1st Edition Kc

Tan Phd

https://ebookbell.com/product/multiobjective-evolutionary-algorithms-
and-applications-1st-edition-kc-tan-phd-4239044
Dinh The Luc

Multiobjective
Linear
Programming
An Introduction
Multiobjective Linear Programming
Dinh The Luc

Multiobjective Linear
Programming
An Introduction

123
Dinh The Luc
Avignon University
Avignon
France

ISBN 978-3-319-21090-2 ISBN 978-3-319-21091-9 (eBook)

DOI 10.1007/978-3-319-21091-9

Library of Congress Control Number: 2015943841

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
To Dieu Huyen,
Liuli and The Duc
Preface

Multiobjective optimization problems arise in decision-making processes in many

areas of human activity including economics, engineering, transportation, water
resources, and the social sciences. Although most real-life problems involve non-
linear objective functions and constraints, solution methods are principally
straightforward in problems with a linear structure. Apart from Zeleny’s classic
1974 work entitled “Linear Multiobjective Programming” and Steuer’s 1986 book
“Multiple Criteria Optimization: Theory, Computation and Application,” nearly all
textbooks and monographs on multiobjective optimization are devoted to non-
convex problems in a general setting, sometimes with set-valued data, which are not
always accessible to practitioners. The main purpose of this book is to introduce
readers to the field of multiobjective optimization using problems with fairly simple
structures, namely those in which the objective and constraint functions are linear.
By working with linear problems, readers will easily come to grasp the fundamental
concepts of vector problems, recognize parallelisms in more complicated problems
with scalar linear programming, analyze difficulties related to multi-dimensionality
in the outcome space, and develop effective methods for treating multiobjective
problems.
Because of the introductory nature of the book, we have sought to present the
material in as elementary a fashion as possible, so as to require only a minimum of
mathematical background knowledge. The first part of the book consists of two
chapters providing the necessary concepts and results on convex polyhedral sets
and linear programming to prepare readers for the new area of optimization with
several objective functions. The second part of the book begins with an examination
of the concept of Pareto optimality, distinguishing it from the classical concept of
optimality used in traditional optimization. Two of the most interesting topics in
this part of the book involve duality and stability in multiple objective linear
programming, both of which are discussed in detail. The third part of the book is
devoted to numerical algorithms for solving multiple objective linear programs.
This includes the well-known multiple objective simplex method, the outcome
space method, and a recent method using normal cone directions.

vii
viii Preface

Although some new research results are incorporated into the book, it is well
suited for use in the first part of a course on multiobjective optimization for
undergraduates or first-year graduate students in applied mathematics, engineering,
computer science, operations research, and economics. Neither integer problems
nor fuzzy linear problems are addressed. Further, applications to other domains are
not tackled, though students will certainly have no real difficulty in studying them,
once the basic results of this book assimilated.
During the preparation of this manuscript I have benefited from the assistance of
many people. I am grateful to my Post-Ph.D. and Ph.D. students Anulekha Dhara,
Truong Thi Thanh Phuong, Tran Ngoc Thang, and Moslem Zamani for their careful
reading of the manuscript. I would also like to thank Moslem Zamani for the
illustrative figures he made for this book. I want to take this opportunity to give
special thanks to Juan-Enrique Martinez-Legaz (Autonomous University of
Barcelona), Boris Mordukhovich (Wayne State University), Nguyen Thi Bach Kim
(Hanoi Polytechnical University), Panos Pardalos (University of Florida), Michel
Thera (University of Limoges), Majid Soleimani-Damaneh (University of Tehran),
Ralph E. Steuer (University of Georgia), Michel Volle (University of Avignon), and
Mohammad Yaghoobi (University of Kerman) for their valued support in this
endeavor.

Avignon Dinh The Luc

December 2014
Contents

1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Part I Background

2 Convex Polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 The Space Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 System of Linear Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Convex Polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Basis and Vertices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3 Linear Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Optimal Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Dual Problems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3 The Simplex Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

Part II Theory

4 Pareto Optimality. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1 Pareto Maximal Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 Multiobjective Linear Problems . . . . . . . . . . . . . . . . . . . . . . . . 102
4.3 Scalarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

5 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.1 Dual Sets and Dual Problems . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.2 Ideal Dual Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.3 Strong Dual Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.4 Weak Dual Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.5 Lagrangian Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

ix
x Contents

5.6 Parametric Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

5.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176

6 Sensitivity and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

6.1 Parametric Convex Polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . 183
6.2 Sensitivity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
6.3 Error Bounds and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
6.4 Post-optimal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234

Part III Methods

7 Multiobjective Simplex Method. . . . . . . . . . . . . . . . . . . . . . . . . . . 241

7.1 Description of the Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
7.2 The Multiobjective Simplex Tableau . . . . . . . . . . . . . . . . . . . . 247
7.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258

8 Normal Cone Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

8.1 Normal Index Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
8.2 Positive Index Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
8.3 The Normal Cone Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
8.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284

9 Outcome Space Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289

9.1 Analysis of the Efficient Set in the Outcome Space . . . . . . . . . . 289
9.2 Free Disposal Hull . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
9.3 Outer Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
9.4 The Outcome Space Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 299
9.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305

Bibliographical Notes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309

References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
Notations

N Natural numbers
R Real numbers
Rn Euclidean n-dimensional space
LðRn ; Rm Þ Space of m  n matrices
Bn Closed unit ball in Rn
Sn Unit sphere in Rn
Bmn Closed unit ball in LðRn ; Rm Þ
e Vector of ones
ei i-th coordinate unit vector
Δ Standard simplex
k xk Euclidean norm
k x k1 Max-norm
hx; yi Canonical scalar product
5 Less than or equal to
 Less than but not equal to
\ Strictly less than
affðAÞ Affine hull

clðAÞ, A Closure
intðAÞ Interior
riðAÞ Relative interior
coðAÞ Convex hull
coðAÞ Closed convex hull
coneðAÞ Conic hull
posðAÞ Positive hull
MaxðAÞ Set of maximal elements
WMaxðAÞ Set of weakly maximal elements
MinðAÞ Set of minimal elements
WMinðAÞ Set of weakly minimal elements
S(MOLP) Efficient solution set
WS(MOLP) Weakly efficient solution set

xi
xii Notations

supðAÞ Supremum
IðxÞ Active index set at x
A? Orthogonal
A Negative polar cone
A1 Recession/asymptotic cone
NA ðxÞ Normal cone
dðx; CÞ Distance function
hðA; BÞ Hausdorff distance
grðGÞ Graph
suppðxÞ Support
Chapter 1
Introduction

Mathematical optimization studies the problem of finding the best element from a set
of feasible alternatives with regard to a criterion or objective function. It is written
in the form

optimize f (x)
subject to x ∈ X,

where X is a nonempty set, called a feasible set or a set of feasible alternatives, and
f is a real function on X , called a criterion or objective function. Here “optimize”
stands for either “minimize” or “maximize” which amounts to finding x̄ ∈ X such
that either f (x̄)  f (x) for all x ∈ X , or f (x̄)  f (x) for all x ∈ X .
This model offers a general framework for studying a variety of real-world and
theoretical problems in the sciences and human activities. However, in many practical
situations, we tend to encounter problems that involve not just one criterion, but a
number of criteria, which are often in conflict with each other. It then becomes
impossible to model such problems in the above-mentioned optimization framework.
Here are some instances of such situations.
Automotive design The objective of automotive design is to determine the technical
parameters of a vehicle to minimize (1) production costs, (2) fuel consumption, and
(3) emissions, while maximizing (4) performance and (5) crash safety. These criteria
are not always compatible; for instance a high-performance engine often involves
very high production costs, which means that no design can optimally fulfill all
criteria.
House purchase Buying property is one of life’s weightiest decisions and often
requires the help of real estate agencies. An agency suggests a number of houses
or apartments which roughly meet the potential buyer’s budget and requirements. In
order to make a decision, the buyer assesses the available offers on the basis of his
or her criteria. The final choice should satisfy the following: minimal cost, minimal
maintenance charges, maximal quality and comfort, best environment etc. It is quite

© Springer International Publishing Switzerland 2016 1

D.T. Luc, Multiobjective Linear Programming,
DOI 10.1007/978-3-319-21091-9_1
2 1 Introduction

natural that the higher the quality of the house, the more expensive it is; as such, it
is impossible to make the best choice without compromising.
Distributing electrical power In a system of thermal generators the chief problem
concerns allocating the output of each generator in the system. The aim is not only
to satisfy the demand for electricity, but also to fulfill two main criteria: minimizing
the costs of power generation and minimizing emissions. Since the costs and the
emissions are measured in different units, we cannot combine the two criteria into one.
Queen Dido’s city Queen Dido’s famous problem consists of finding a territory
bounded by a line which has the maximum area for a given perimeter. According to
elementary calculus, the solution is known to be a circle. However, as it is incon-
ceivable to have a city touching the sea without a seashore, Queen Dido set another
objective, namely for her territory to have as large a seashore as possible. As a result,
a semicircle partly satisfies her two objectives, but fails to maximize either aspect.
As we have seen, even in the simplest situations described above there can be
no alternative found that simultaneously satisfies all criteria, which means that the
known concepts of optimization do not apply and there is a real need to develop new
notions of optimality for problems involving multiple objective functions. Such a
concept was introduced by Pareto (1848–1923), an Italian economist who explained
the Pareto optimum as follows: “The optimum allocation of the resources of a society
is not attained so long as it is possible to make at least one individual better off in his
own estimation while keeping others as well off as before in their own estimation.”
Prior to Pareto, the Irish economist Edgeworth (1845–1926) had defined an optimum
for the multiutility problem of two consumers P and Q as “a point (x, y) such that in
whatever direction we take an infinitely small step, P and Q do not increase together
but that, while one increases, the other decreases.” According to the definition put
forward by Pareto, among the feasible alternatives, those that can simultaneously be
improved with respect to all criteria cannot be optimal. And an alternative is optimal
if any alternative better than it with respect to a certain criterion is worse with respect
to some other criterion, that is, if a tradeoff takes place when trying to find a better
alternative. From the mathematical point of view, if one defines a domination order
in the set of feasible alternatives by a set of criteria—an alternative a dominates
an alternative b if the value of every criterion function at a is bigger than that at
b—then an alternative is optimal in the Pareto sense if it is dominated by no other
alternatives. In other words, an alternative is optimal if it is maximal with respect
to the above order. This explains the mathematical origin of the theory of multiple
objective optimization, which stems from the theory of ordered spaces developed by
Cantor (1845–1918) and Hausdorff (1868–1942).
A typical example of ordered spaces, frequently encountered in practice, is the
finite dimensional Euclidean space Rn with n ≥ 2, in which two vectors a and b
are comparable, let’s say a is bigger than or equal to b if all coordinates of a are
bigger than or equal to the corresponding coordinates of b. A multiple objective
optimization problem is then written as
1 Introduction 3

Maximize F(x) := ( f 1 (x), . . . , f k (x))

subject to x ∈ X,

where f 1 , . . . , f k are real objective functions on X and “Maximize” signifies finding

an element x̄ ∈ X such that no value F(x), x ∈ X is bigger than the value F(x̄). It
is essential to note that the solution x̄ is not worse than any other solution, but in no
ways it is the best one, that is, the value F(x̄) cannot be bigger than or equal to all
values F(x), x ∈ X in general. A direct consequence of this observation is the fact
that the set of “optimal values” is not a singleton, which forces practitioners to find
a number of “optimal solutions” before making a final decision. Therefore, solving
a multiple objective optimization problem is commonly understood as finding the
entire set of “optimal solutions” or “optimal values”, or at least a representative
portion of them. Indeed, this is the point that makes multiple objective optimization
a challenging and fascinating field of theoretical research and application.
 Part I
Background
Chapter 2
Convex Polyhedra

We begin the chapter by introducing basic concepts of convex sets and linear func-
tions in a Euclidean space. We review some of fundamental facts about convex
polyhedral sets determined by systems of linear equations and inequalities, includ-
ing Farkas’ theorem of the alternative which is considered a keystone of the theory
of mathematical programming.

2.1 The Space Rn

Throughout this book, Rn denotes the n-dimensional Euclidean space of real column
n-vectors. The norm of a vector x with components x1 , · · · , xn is given by
 n 1/2

x = (xi ) 2
.
i=1

The inner product of two vectors x and y in Rn is expressed as


n
x, y = xi yi .
i=1

The closed unit ball, the open unit ball and the unit sphere of Rn are respectively
defined by
 
Bn := x ∈ Rn : x  1 ,
 
int(Bn ) := x ∈ Rn : x < 1 ,
 
Sn := x ∈ Rn : x = 1 .

© Springer International Publishing Switzerland 2016 7

D.T. Luc, Multiobjective Linear Programming,
DOI 10.1007/978-3-319-21091-9_2
8 2 Convex Polyhedra

Given a nonempty set Q ⊆ Rn , we denote the closure of Q by cl(Q) and its interior
by int(Q). The conic hull, the positive hull and the affine hull of Q are respectively
given by
 
cone(Q) := ta : a ∈ Q, t ∈ R, t  0 ,
 k 

pos(Q) := ti a i : a i ∈ Q, ti ∈ R, ti  0, i = 1, · · · , k with k ∈ N ,
i=1
 

k 
k
aff(Q) := ti a : a ∈ Q, ti ∈ R, i = 1, · · · , k and
i i
ti = 1 with k ∈ N ,
i=1 i=1

where N denotes the set of natural numbers (Figs. 2.1, 2.2 and 2.3).

Fig. 2.1 Conic hull (with

Q = Q1 ∪ Q2) cone(Q)

Q1

cone(Q)
Q2

Fig. 2.2 Positive hull (with

Q = Q1 ∪ Q2)
pos(Q)
Q1

Q2
2.1 The Space Rn 9

Fig. 2.3 Affine hull (with

Q = Q1 ∪ Q2)
af f (Q)
Q1

Q2

Among the sets described above cone(Q) and pos(Q) are cones, that is, they are
invariant under multiplication by positive numbers; pos(Q) is also invariant under
addition of its elements; and aff(Q) is an affine subspace of Rn . For two vectors x
and y of Rn , inequalities x > y and x  y mean respectively xi > yi and xi  yi
for all i = 1, · · · , n. When x  y and x = y, we write x ≥ y. So a vector x is
positive, that is x  0, if its components are non-negative; and it is strictly positive
if its components are all strictly positive. The set of all positive vectors of Rn is the
positive orthant Rn+ . Sometimes row vectors are also considered. They are transposes
of column vectors. Operations on row vectors are performed in the same manner as
on column vectors. Thus, for two row n-vectors c and d, their inner product is
expressed by

n
c, d = c T , d T  = ci di ,
i=1

where the upper index T denotes the transpose. On the other hand, if c is a row vector
and x is a column vector, then the product cx is understood as a matrix product which
is equal to the inner product c T , x.

Convex sets
We call a subset Q of Rn convex if the segment joining any two points of Q lies entirely
in Q, which means that for every x, y ∈ Q and for every real number λ ∈ [0, 1], one
has λx + (1 − λ)y ∈ Q (Figs. 2.4, 2.5). It follows directly from the definition that the
intersection of convex sets, the Cartesian product of convex sets, the image and inverse
image of a convex set under a linear transformation, the interior and the closure of a
convex set are convex. In particular, the sum Q 1 + Q 2 := {x + y : x ∈ Q 1 , y ∈ Q 2 }
of two convex sets Q 1 and Q 2 is convex; the conic hull of a convex set is convex.
The positive hull and the affine hull of any set are convex.
The convex hull of Q, denoted co(Q) (Fig. 2.6), consists of all convex combina-
tions of elements of Q, that is,
10 2 Convex Polyhedra

Fig. 2.4 Convex set

y

Fig. 2.5 Nonconvex set

x
y

 

k 
k
co(Q) := λi x : x ∈ Q, λi  0, i = 1, · · · , k and
i i
λi = 1 with k ∈ N .
i=1 i=1

It is the intersection of all convex sets containing Q. The closure of the convex hull
of Q will be denoted by co(Q), which is exactly the intersection of all closed convex
sets containing Q. The positive hull of a set is the conic hull of its convex hull. A
k
convex combination i=1 λi x i is strict if all coefficients λi are strictly positive.
Given a nonempty convex subset Q of Rn , the relative interior of Q, denoted
ri(Q), is its interior relative to its affine hull, that is,
 
ri(Q) := x ∈ Q : (x + εBn ) ∩ aff(Q) ⊆ Q for some ε > 0 .

Equivalently, a point x in Q is a relative interior point if and only if for any point y in
Q there is a positive number δ such that the segment joining the points x − δ(x − y)
and x + δ(x − y) entirely lies in Q. As a consequence, any strict convex combination
of a finite collection {x 1 , · · · , x k } belongs to the relative interior of its convex hull
(see also Lemma 6.4.8). It is important to note also that every nonempty convex set
in Rn has a nonempty relative interior. Moreover, if two convex sets Q 1 and Q 2 have
at least one relative interior point in common, then ri(Q 1 ∩ Q 2 ) = ri(Q 1 ) ∩ ri(Q 2 ).

Fig. 2.6 Convex hull of Q

co(Q)
Q
2.1 The Space Rn 11

Fig. 2.7 The standard

simplex in R3

Example 2.1.1 (Standard simplex) Let ei be the ith coordinate unit vector of Rn ,
that is its components are all zero except for the ith component equal to one. Let Δ
denote the convex hull of e1 , · · · , en . Then a vector x with components x1 , · · · , xn is
n
an element of Δ if and only if xi  0, i = 1, · · · , n and i=1 xi = 1. This set has no
interior point. However, its relative interior consists of x with xi > 0, i = 1, · · · , n
n
and i=1 xi = 1. The set Δ is called the standard simplex of Rn (Fig. 2.7).

Caratheodory’s theorem
It turns out that the convex hull of a set Q in the space Rn can be obtained by convex
combinations of at most n + 1 elements of Q. First we see this for positive hull.

Theorem 2.1.2 Let {a 1 , · · · , a k } be a collection of vectors in Rn . Then for every

nonzero vector x from the positive hull pos{a 1 , · · · , a k } there exists an index set
I ⊆ {1, · · · , k} such that
(i) the vectors a i , i ∈ I are linearly independent;
(ii) x belongs to the positive hull pos{a i , i ∈ I }.

Proof Since the collection {a 1 , · · · , a k } is finite, we may choose an index set I of

minimum cardinality such that x ∈ pos{a i , i ∈ I }. It is evident that there are strictly
positive numbers ti , i ∈ I such that x = i∈I ti a i . We prove that (i) holds for this
I . Indeed, if not, one can find an index j ∈ I and real numbers si such that

aj − si a i = 0.
i∈I \{ j}

Set ti
ε = min t j and − : i ∈ I with si < 0
si

and express
12 2 Convex Polyhedra
 
x= ti a i − ε a j − si a i
i∈I i∈I \{ j}

= (t j − ε)a + j
(ti + εsi )a i .
i∈I \{ j}

It is clear that in the latter sum those coefficients corresponding to the indices that
realize the minimum in the definition of ε are equal to zero. By this, x lies in the
positive hull of less than |I | vectors of the collection. This contradiction completes
the proof. 

A collection of vectors {a 1 , · · · , a k } in Rn is said to be affinely independent if

the dimension of the subspace aff{a 1 , · · · , a k } is equal to k − 1. By convention a set
consisting of a solitary vector is affinely independent. The next result is a version of
Caratheodory’s theorem and well-known in convex analysis.

Corollary 2.1.3 Let {a 1 , · · · , a k } be a collection of vectors in Rn . Then for every

x ∈ co{a 1 , · · · , a k } there exists an index set I ⊆ {1, · · · , k} such that
(i) the vectors a i , i ∈ I are affinely independent
(ii) x belongs to the convex hull of a i , i ∈ I.

Proof We consider the collection of vectors v i = (a i , 1), i = 1, · · · , k in the space

Rn × R. It is easy to verify that x belongs to the convex hull co{a 1 , · · · , a k } if
and only if the vector (x, 1) belongs to the positive hull pos{v 1 , · · · , v k }. Applying
Theorem 2.1.2 to the latter positive hull we deduce the existence of an index set
I ⊆ {1, · · · , k} such that the vector (x, 1) belongs to the positive hull pos{v i , i ∈ I }
and the collection {v i , i ∈ I } is linearly independent. Then x belongs to the convex
hull co{a i , i ∈ I } and the collection {a i , i ∈ I } is affinely independent. 

Linear operators and matrices

A mapping φ : Rn → Rk is called a linear operator between Rn and Rk if
(i) φ(x + y) = φ(x) + φ(y),
(ii) φ(t x) = tφ(x)
for every x, y ∈ Rn and t ∈ R. The kernel and the image of φ are the sets
 
Kerφ = x ∈ Rn : φ(x) = 0 ,
 
Imφ = y ∈ Rk : y = φ(x) for some x ∈ Rn .

These sets are linear subspaces of Rn and Rk respectively.

We denote the k × n-matrix whose columns are c1 , · · · , cn by C, where ci is the
vector image of the ith coordinate unit vector ei by φ. Then for every vector x of Rn
one has
φ(x) = C x.
2.1 The Space Rn 13

The mapping x → C x is clearly a linear operator from Rn to Rk . This explains

why one can identify a linear operator with a matrix. The space of k × n matrices
is denoted by L(Rn , Rk ). The transpose of a matrix C is denoted by C T . The norm
and the inner product in the space of matrices are given by
  1/2
C = |ci j |2 ,
i=1,··· ,n j=1,··· ,n
 
C, B = ci j bi j .
i=1,··· ,n j=1,··· ,n

The norm C is called also the Frobenius norm.

The inner product C, B is nothing but the trace of the matrix C B T . Sometimes
the space L(Rn , Rk ) is identified with the n × k-dimensional Euclidean space Rn×k .

Linear functionals
A particular case of linear operators is when the value space is one-dimensional. This
is the space of linear functionals on Rn and often identified with the space Rn itself.
Thus, each linear functional φ is given by a vector dφ by the formula

φ(x) = dφ , x.

When dφ = 0, the kernel of φ is called a hyperplane; the vector dφ is a normal vector

to this hyperplane. Geometrically, dφ is orthogonal to the hyperplane Kerφ. The sets
 
x ∈ Rn : dφ , x  0 ,
 
x ∈ Rn : dφ , x  0

are closed halfspaces and the sets

 
x ∈ Rn : dφ , x > 0 ,
 
x ∈ Rn : dφ , x < 0

are open halfspaces bounded by the hyperplane Kerφ. Given a real number α and a
nonzero vector d of Rn , one also understands a hyperplane of type
 
H (d, α) = x ∈ Rn : d, x = α .

The sets
 
H+ (d, α) = x ∈ Rn : d, x  α ,
 
H− (d, α) = x ∈ Rn : d, x  α

are positive and negative halfspaces and the sets

14 2 Convex Polyhedra
   
int H+ (d, α) = x ∈ Rn : d, x > α ,
   
int H− (d, α) = x ∈ Rn : d, x < α

are positive and negative open halfspaces.

Theorem 2.1.4 Let Q be a nonempty convex set in Rn and let d, . be a positive
functional on Q, that is d, x  0 for every x ∈ Q. If d, x = 0 for some relative
interior point x of Q, then d, . is zero on Q.

Proof Let y be any point in Q. Since x is a relative interior point, there exists a
positive number δ such that x + t (y − x) ∈ Q for |t|  δ. Applying d, . to this
point we obtain
d, x + t (y − x) = td, y  0

for all t ∈ [−δ, δ]. This implies that d, y = 0 as requested. 

2.2 System of Linear Inequalities

We shall mainly deal with two kinds of systems of linear equations and inequalities.
The first system consists of k inequalities

a i , x  bi , i = 1, · · · , k, (2.1)

where a 1 , · · · , a k are n-dimensional column vectors and b1 , · · · , bk are real num-

bers; and the second system consists of k equations which involves positive vectors
only

a i , x = bi , i = 1, · · · , k (2.2)
x  0.

Denoting by A the k × n-matrix whose rows are the transposes of a 1 , · · · , a k and

by b the column k-vector of components b1 , · · · , bk , we can write the systems (2.1)
and (2.2) in matrix form
Ax  b (2.3)

and

Ax = b (2.4)
x  0.

Notice that any system of linear equations and inequalities can be converted to the
two matrix forms described above. To this end it suffices to perform three operations:
2.2 System of Linear Inequalities 15

(a) Express each variable xi as difference of two non-negative variables xi = xi+ −

xi− where

xi+ = max{xi ; 0},

xi− = max{−xi ; 0}.

(b) Introduce a non-negative slack variable yi in order to obtain equivalence between

inequality a i , x  bi and equality a i , x + yi = bi . Similarly, with a non-
negative surplus variable z i one may express inequality a i , x  bi as equality
a i , x − z i = bi .
(c) Express equality a i , x = bi by two inequalities a i , x  bi and a i , x  bi .

Example 2.2.1 Consider the following system

x1 + 2x2 = 1,
−x1 − x2  0.

It is written in form (2.3) as

⎛ ⎞ ⎛ ⎞
1 2   1
⎝ −1 −2 ⎠ x1  ⎝ −1 ⎠
x2
1 1 0

and in form (2.4) with a surplus variable y as

   
1 −1 2 −2 0  + − + − T 1
x1 , x1 , x2 , x2 , y = ,
−1 1 −1 1 −1 0
(x1+ , x1− , x2+ , x2− , y)T  0.

Redundant equation
Given a system (2.4) we say it is redundant if at least one of the equations (called
redundant equation) can be expressed as a linear combination of the others. In other
words, it is redundant if there is a nonzero k-dimensional vector λ such that

A T λ = 0,
b, λ = 0.

Moreover, redundant equations can be dropped from the system without changing
its solution set. Similarly, an inequation of (2.1) is called redundant if its removal
from the system does not change the solution set.
16 2 Convex Polyhedra

Proposition 2.2.2 Assume that k  n and that the system (2.4) is consistent. Then
it is not redundant if and only if the matrix A has full rank.
Proof If one of equations, say a 1 , x = b1 , is redundant, then a 1 is a linear combina-
tion of a 2 , · · · , a k . Hence the rank of A is not maximal, it is less than k. Conversely,
when the rank of A is maximal (equal to k), no row of A is a linear combination of
the others. Hence no equation of the system can be expressed as a linear combination
of the others. 

Farkas’ theorem
One of the theorems of the alternative that are pillars of the theory of linear and
nonlinear programming is Farkas’ theorem or Farkas’ lemma. There are a variety of
ways to prove it, the one we present here is elementary.
Theorem 2.2.3 (Farkas’ theorem) Exactly one of the following systems has a
solution:
(i) Ax = b and x  0;
(ii) A T y  0 and b, y < 0.
Proof If the first system has a solution x, then for every y with A T y  0 one has

b, y = Ax, y = x, A T y  0,

which shows that the second system has no solution.

Now suppose the first system has no solution. Then either the system

Ax = b

has no solution, or it does have a solution, but every solution of it is not positive. In the
first case, choose m linearly independent columns of A, say a1 , · · · , am , where m is
the rank of A. Then the vectors a1 , · · · , am , b are linearly independent too (because
b does not lie in the space spanned by a1 , · · · , am ). Consequently, the system

ai , y = 0, i = 1, · · · , m,
b, y = −1

admits a solution. This implies that the system (ii) has solutions too. It remains to
prove the solvability of (ii) when Ax = b has solutions and they are all non-positive.
We do it by induction on the dimension of x. Assume n = 1. If the system ai1 x1 =
bi , i = 1, · · · , k has a negative solution x1 , then y = −(b1 , · · · , bk )T is a solution
of (ii) because A T y = −(a11 2 + · · · + a 2 )x > 0 and b, y = −(b2 + · · · + b2 ) < 0.
k1 1 1 k
Now assume n > 1 and that the result is true for the case of dimension n −1. Given an
n-vector x, denote by x the (n − 1)-vector consisting of the first (n − 1) components
of x. Let Ā be the matrix composed of the first (n − 1) columns of A. It is clear that
the system
2.2 System of Linear Inequalities 17

Āx = b and x  0

has no solution. By induction there is some y such that

T
A y  0,
b, y < 0.

If an , y  0, we are done. If an , y < 0, define new vectors

âi = ai , yan − an , yai , i = 1, · · · , n − 1,

b̂ = b, yan − an , yb

and consider a new system

â1 ξ1 + · · · + ân−1 ξn−1 = b̂. (2.5)

We claim that this system of k equations has no positive solution. Indeed, if not, say
ξ1 , · · · , ξn−1 were non-negative solutions, then the vector x with

xi = ξi , i = 1, · · · , n − 1,
1
xn = − a1 ξ1 + · · · + an−1 ξn−1 , y − b, y
an , y

T
should be a positive solution of (i) because −b, y > 0 and Aξ, y = ξ, A y  0
for ξ = (ξ1 , · · · , ξn−1 )T  0, implying xn  0. Applying the induction hypothesis
to (2.5) we deduce the existence of a k-vector ŷ with

âi , ŷ  0, i = 1, · · · , n − 1,
b̂, ŷ < 0.

Then the vector y = an , ŷy − an , y ŷ satisfies the system (ii). The proof is
complete. 

A number of consequences can be derived from Farkas’ theorem which are useful
in the study of linear systems and linear programming problems.

Corollary 2.2.4 Exactly one of the following systems has a solution:

(i) Ax = 0, c, x = 1 and x  0;
(ii) A T y  c.
18 2 Convex Polyhedra

Proof If (ii) has a solution y, then for a positive vector x with Ax = 0 one has

0 = y, Ax = A T y, x  c, x.

So (i) is not solvable. Conversely, if (i) has no solution, then applying Farkas’ theorem
to the inconsistent system
   
A 0
x= and x  0
cT 1

yields the existence of a vector y and of a real number t such that

     
y 0 y
AT c  0 and , < 0.
t 1 t

Hence t < 0 and −y/t is a solution of (ii). 

Corollary 2.2.5 Exactly one of the following systems has a solution:

(i) Ax ≥ 0 and x  0;
(ii) A T y  0 and y > 0.

Proof By introducing a surplus variable z ∈ Rk we convert (i) to an equivalent

system

Ax − I z = 0,
 
x
 0,
z
  
x
c, = 1,
z

where c is an (n + k)-vector whose n first components are all zero and the remaining
components are one. According to Corollary 2.2.4 it has no solution if and only if
the following system has a solution
 
AT
y  c.
−I

It is clear that the latter system is equivalent to (ii). 

The next corollary is known as Motzkin’s theorem of the alternative.

Corollary 2.2.6 (Motzkin’s theorem) Let A and B be two matrices having the same
number of columns. Exactly one of the following systems has a solution:
(i) Ax > 0 and Bx  0;
(ii) A T y + B T z = 0, y ≥ 0 and z  0.
2.2 System of Linear Inequalities 19

Proof The system (ii) is evidently equivalent to the following one

    
AT B T y 0
=
eT 0 z 1
 
y
 0.
z

By Farkas’ theorem it is compatible (has a solution) if and only if the following

system is incompatible:
    
Ae x 0

B0 t 0
   
x 0
, < 0.
t 1

The latter system is evidently equivalent to the system of (i). 

Some classical theorems of alternatives are immediate from Corollary 2.2.6.

• Gordan’s theorem (B is the zero matrix):
Exactly one of the following systems has a solution
(1) Ax > 0;
(2) A T y = 0 and y ≥ 0.
• Ville’s theorem (B is the identity matrix):
Exactly one of the following systems has a solution
(3) Ax > 0 and x  0;
(4) A T y  0 and y ≥ 0.
 
B
• Stiemke’s theorem (A is the identity matrix and B is replaced by ):
−B
Exactly one of the following systems has a solution
(5) Bx = 0 and x > 0;
(6) B T y ≥ 0.

2.3 Convex Polyhedra

A set that can be expressed as the intersection of a finite number of closed half-spaces
is called a convex polyhedron. A convex bounded polyhedron is called a polytope.
According to the definition of closed half-spaces, a convex polyhedron is the solution
set to a finite system of inequalities
20 2 Convex Polyhedra

a i , x  bi , i = 1, · · · , k (2.6)

where a 1 , · · · , a k are n-dimensional column vectors and b1 , · · · , bk are real num-

bers. When bi = 0, i = 1, · · · , k, the solution set to (2.6) is a cone and called a
convex polyhedral cone. We assume throughout this section that the system is not
redundant and solvable.

Supporting hyperplanes and faces

Let P be a convex polyhedron and let

H = {x ∈ Rn : v, x = α}

be a hyperplane with v nonzero. We say H is a supporting hyperplane of P at a point

x ∈ P if the intersection of H with P contains x and P is contained in one of the
closed half-spaces bounded by H (Fig. 2.8). In this case, the nonempty set H ∩ P is
called a face of P. Thus, a nonempty subset F of P is a face if there is a nonzero
vector v ∈ Rn such that

v, y  v, x for all x ∈ F, y ∈ P.

When a face is zero-dimensional, it is called a vertex. A nonempty polyhedron may

have no vertex. By convention P is a face of itself; other faces are called proper
faces. One-dimensional faces are called edges. Two vertices are said to be adjacent
if they are end-points of an edge.

Example 2.3.1 Consider a system of three inequalities in R2 :

x1 + x2  1 (2.7)
−x1 − x2  0 (2.8)
−x1  0. (2.9)

Fig. 2.8 Supporting

hyperplane H

x
2.3 Convex Polyhedra 21

The polyhedron defined by (2.7) and (2.8) has no vertex. It has two one-dimensional
faces determined respectively by x1 + x2 = 1 and x1 + x2 = 0, and one two-
dimensional face, the polyhedron itself. The polyhedron defined by (2.7)–(2.9) has
two vertices (zero-dimensional faces) determined respectively by
 
x1 = 0 x1 = 0
and ,
x2 = 0 x2 = 1

three one-dimensional faces given by

⎧
⎨ x1 + x2  1  
x1 + x2 = 1 −x1 − x2 = 0
−x1 − x2  0 , and ,
⎩ −x1  0 −x1  0
x1 =0

and one two-dimensional face, the polyhedron itself.

Proposition 2.3.2 Let P be a convex polyhedron. The following properties hold.

(i) The intersection of any two faces is a face if it is nonempty.
(ii) Two different faces have no relative interior point in common.

Proof We prove (i) first. Assume F1 and F2 are two faces with nonempty intersection.
If they coincide, there is nothing to prove. If not, let H1 and H2 be two supporting
hyperplanes that generate these faces, say
 
H1 = x ∈ Rn : v 1 , x = α1 ,
 
H2 = x ∈ Rn : v 2 , x = α2 .

Since these hyperplanes contain the intersection of distinct faces F1 and F2 , the
vector v = v 1 + v 2 is not zero. Consider the hyperplane
 
H = x ∈ Rn : v, x = α1 + α2 .

It is a supporting hyperplane of P because it evidently contains the intersection of

the faces F1 and F2 , and for every point x in P, one has

v, x = v 1 , x + v 2 , x  α1 + α2 . (2.10)

It remains to show that the intersection of H and P coincides with the intersection
F1 ∩ F2 . The inclusion
F1 ∩ F2 ⊆ H ∩ P

being clear, we show the converse. Let x be in H ∩ P. Then (2.10) becomes equality
for this x. But v 1 , x  α1 and v 2 , x  α2 , so that equality of (2.10) is possible
only when the two latter inequalities are equalities. This proves that x belongs to
both F1 and F2 .
22 2 Convex Polyhedra

For the second assertion notice that if F1 and F2 have a relative interior point in
common, then in view of Theorem 2.1.4, the functional v 1 , . is constant on F2 . It
follows that F2 ⊆ H1 ∩ P ⊆ F1 . Similarly, one has F1 ⊆ F2 , and hence equality
holds. 

Let x be a solution of the system (2.6). Define the active index set at x to be the
set
I (x) = i ∈ {1, · · · , k} : a i , x = bi .

The remaining indices are called inactive indices.

Theorem 2.3.3 Assume that P is a convex polyhedron given by (2.6). A nonempty

proper convex subset F of P is a face if and only if there is a nonempty maximal
index set I ⊆ {1, · · · , k} such that F is the solution set to the system

a i , x = bi , i ∈ I (2.11)
a , x  b j , j ∈ {1, · · · , k}\I,
j
(2.12)

in which case the dimension of F is equal to n−rank{a i : i ∈ I }.

Proof Denote the solution set to the system (2.11, 2.12) by F that we suppose
nonempty. To prove that it is a face, we set
 
v= a i and α = bi .
i∈I i∈I

Notice that v is nonzero because F is not empty and the system (2.6) is not redundant.
It is clear that the negative half-space H− (v, α) contains P. Moreover, if x is a
solution to the system, then, of course, x belongs to P and to H at the same time,
which implies F ⊆ H ∩ P. Conversely, any point x of the latter intersection satisfies

a i , x  bi , i = 1, · · · , k,
 
a i , x = bi .
i∈I i∈I

The latter equality is possible only when those inequalities with indices from I are
equalities. In other words, x belongs to F .
Now, let F be a proper face of P. Pick a relative interior point x of F and consider
the system (2.11, 2.12) with I = I (x) the active index set of x. Being a proper face of
P, F has no interior point, and so the set I is nonempty. As before, F is the solution
set to that system. By the first part, it is a face. We wish to show that it coincides with
F. For this, in view of Proposition 2.3.2 it suffices to show that x is also a relative
interior point of F . Let x be another point in F . We have to prove that there is a
positive number δ such that the segment [x, x + δ(x − x)] lies in F . Indeed, note
2.3 Convex Polyhedra 23

that for indices j outside the set I , inequalities a j , x  b j are strict. Therefore,
there is δ > 0 such that
a j , x + δa j , x − x  b j

for all j ∈
/ I . Moreover, being a linear combination of x and x, the endpoint x +
δ(x − x) satisfies the equalities (2.11) too. Consequently, this point belongs to F ,
and hence so does the whole segment. Since F and F are two faces with a relative
interior point in common, they must be the same. 
In general, for a given face F of P, there may exist several index sets I for which
F is the solution set to the system (2.11, 2.12). We shall, however, understand that
no inequality can be equality without changing the solution set when saying that the
system (2.11, 2.12) determines the face F. So, if two inequalities combined yields
equality, their indices will be counted in I .
Corollary 2.3.4 If an m-dimensional convex polyhedron has a vertex, then it has
faces of any dimension less than m.
Proof The corollary is evident for a zero-dimensional polyhedron. Suppose P is a
polyhedron of dimension m > 0. By Theorem 2.3.3 without loss of generality we
may assume that P is given by the system (2.11, 2.12) with |I | = n − m and that
the family {a i , i ∈ I } is linearly independent. Since P has a vertex, there is some
i 0 ∈ {1, · · · , k}\I such that the vectors a i , i ∈ I ∪ {i 0 } are linearly independent.
Then the system

a i , x = bi , i ∈ I ∪ {i 0 },
a j , x  b j , j ∈ {1, · · · , k}\(I ∪ {i 0 })

generates an (m − 1)-dimensional face of P. Notice that this system has a solution

because P is generated by the non-redundant system (2.11, 2.12). Continuing the
above process we are able to construct a face of any dimension less than m. 
Corollary 2.3.5 Let F be a face of the polyhedron P determined by the system (2.11,
2.12). Then for every x ∈ F one has

I (x) ⊇ I.

Equality holds if and only if x is a relative interior point of F.

Proof The inclusion I ⊆ I (x) is evident because x ∈ F. For the second part, we
first assume I (x) = I , that is

a i , x = bi , i ∈ I,
a j , x < b j , j ∈ {1, · · · , k}\I.

It is clear that if y ∈ aff(F), then a i , y = bi , i ∈ I, and if y ∈ x + εBk with

ε > 0 sufficiently small, then a j , y < b j , j ∈ {1, · · · , k}\I. We deduce that
24 2 Convex Polyhedra

aff(F) ∩ (x + εBk ) ⊆ F, which shows that x is a relative interior point of F.

Conversely, let x be a relative interior point of F. Using the argument in the proof
of Theorem 2.3.3 we know that F is also a solution set to the system

a i , y = bi , i ∈ I (x),
a j , y  b j , j ∈ {1, · · · , k}\I (x).

Since the system (2.11, 2.12) determines F, we have I (x) ⊆ I , and hence equality
follows. 

Corollary 2.3.6 Let F be a face of the polyhedron P determined by the system (2.11,
2.12). Then a point v ∈ F is a vertex of F if and only if it is a vertex of P.

Proof It is clear that every vertex of P is a vertex of F if it belongs to F. To prove

the converse, let us deduce a system of inequalities from (2.11, 2.12) by expressing
equalities a i , x = bi as two inequalities a i , x  bi and −a i , x  −bi . If v is
a vertex of F, then the active constraints at v consists of the vectors a i , −a i , i ∈ I
and some a j , j ∈ J ⊆ {1, · · · , k}\I , so that the rank of the family {a i , −a i , a j :
i ∈ I, j ∈ J } is equal to n. It follows that the family {a i , a j : i ∈ I, j ∈ J } has rank
equal to n too. In view of Theorem 2.3.3 the point v is a vertex of P. 

Given a face F of a polyhedron, according to the preceding corollary the active

index set I (x) is constant for every relative interior point x of F. Therefore, we call
it active index set of F and denote it by I F .
A collection of subsets of a polyhedron is said to be a partition of it if the elements
of the collection are disjoint and their union contains the entire polyhedron.
Corollary 2.3.7 The collection of all relative interiors of faces of a polyhedron forms
a partition of the polyhedron.

Proof It is clear from Proposition 2.3.2(ii) that relative interiors of different faces
are disjoint. Moreover, given a point x in P, consider the active index set I (x). If
it is empty, then the point belongs to the interior of P and we are done. If it is not
empty, by Corollary 2.3.5, the face determined by system (2.11, 2.12) with I = I (x)
contains x in its relative interior. The proof is complete. 

We now deduce a first result on representation of elements of polyhedra by ver-

tices.

Corollary 2.3.8 A convex polytope is the convex hull of its vertices.

Proof The corollary is evident when the dimension of a polytope is less or equal
to one in which case it is a point or a segment with two end-points. We make the
induction hypothesis that the corollary is true when a polytope has dimension less
than m with 1 < m < n and prove it for the case when P is a polytope determined by
the system (2.6) and has dimension equal to m. Since P is a convex set, the convex
hull of its vertices is included in P itself. Conversely, let x be a point of P. In view of
2.3 Convex Polyhedra 25

Corollary 2.3.7 it is a relative interior point of some face F of P. If F is a proper face

of P, its dimension is less than m, and so we are done. It remains to treat the case
where x is a relative interior point of P. Pick any point y = x in P and consider the
line passing through x and y. Since P is bounded, the intersection of this line with P
is a segment, say with end-points c and d. Let Fc and Fd be faces of P that contain
c and d in their relative interiors. As c and d are not relative interior points of P, the
faces Fc and Fd are proper faces of P, and hence they have dimension strictly less
than m. By induction c and d belong to the convex hulls of the vertices of Fc and Fd
respectively. By Corollary 2.3.6 they belong to the convex hull of the vertices of P,
and hence so does x because x belongs to the convex hull of c and d. 

A similar result is true for polyhedral cones. It explains why one-dimensional

faces of a polyhedral cone are called extreme rays.

Corollary 2.3.9 A nontrivial polyhedral cone with vertex is the convex hull of its
one-dimensional faces.

Proof By definition a polyhedral cone P is defined by a homogeneous system

a i , x  0, i = 1, · · · , k. (2.13)

Choose any nonzero point y in P and consider the hyperplane H given by

a 1 + · · · + a k , x − y = 0. (2.14)

We claim that the vector a 1 + · · · + a k is nonzero. Indeed, if not, the inequalities

(2.13) would become equalities for all x ∈ P, and P would be either a trivial cone, or
a cone without vertex. Moreover, P ∩ H is a bounded polyhedron, because otherwise
one should find a nonzero vector u satisfying a i , u = 0, i = 1, · · · , k and P could
not have vertices. In view of Corollary 2.3.8, P ∩ H is the convex hull of its vertices.
To complete the proof it remains to show that a vertex v of P ∩ H is the intersection of
a one-dimensional face of P with H . Indeed, the polytope P ∩ H being determined
by the system (2.13) and (2.14), there is a set J ⊂ {1, · · · , k} with |J | = n − 1 such
that the vectors a j , j ∈ J and a 1 + · · · + a k are linearly independent and v is given
by system

a j , x = 0, j ∈ J (2.15)
a + · · · + a k , x = a 1 + · · · + a k , y,
1

a i , x  0, i ∈ {1, · · · , k}\J. (2.16)

It is clear that (2.15) and (2.16) determine a one-dimensional face of P whose inter-
section with H is v. 
26 2 Convex Polyhedra

Separation of convex polyhedra

Given two convex polyhedra P and Q in Rn , we say that a nonzero vector v separates
them if
v, x  v, y for all vectors x ∈ P, y ∈ Q

and strict inequality is true for some of them (Fig. 2.9). The following result can be
considered as a version of Farkas’ theorem or Gordan’s theorem.

Theorem 2.3.10 If P and Q are convex polyhedra without relative interior points
in common, then there is a nonzero vector separating them.

Proof We provide a proof for the case where both P and Q have interior points only.
Without loss of generality we may assume that P is determined by the system (2.6)
and Q is determined by the system

d j , x  c j , j = 1, · · · , m.

Thus, the following system:

a i , x < bi , i = 1, · · · , k
d j , x < c j , j = 1, · · · , m

has no solution because the first k inequalities determine the interior of P and the last
m inequalities determine the interior of Q. This system is equivalent to the following
one: ⎛ ⎞ ⎛ ⎞
−A b   0
⎝ −D c ⎠ x > ⎝ 0 ⎠ ,
t
01 0

Fig. 2.9 Separation

H

P
Q
2.3 Convex Polyhedra 27

where A is the k × n-matrix whose rows are transposes of a 1 , · · · , a k , D is the

m × n-matrix whose rows are transposes of d 1 , · · · , d k , b is the k-vector with
the components b1 , · · · , bk and c is the m-vector with the components c1 , · · · , cm .
According to Gordan’s theorem, there exist positive vectors λ ∈ Rk and μ ∈ Rm and
a real number s  0, not all zero, such that

A T λ + D T μ = 0,
b, λ + c, μ + s = 0.

It follows from the latter equality that (λ, μ) is nonzero. We may assume without
loss of generality that λ = 0. We claim that A T λ = 0. Indeed, if not, choose x an
interior point of P and y an interior point of Q. Then D T μ = 0 and hence

b, λ > Ax, λ = x, A T λ = 0

and
c, μ  Dy, μ = y, D T μ = 0,

which is in contradiction with the aforesaid equality. Defining v to be the nonzero

vector −A T λ, we deduce for every x ∈ P and y ∈ Q that

v, x = −A T λ, x = λ, −Ax  λ, −b  μ, c  μ, Dy = v, y.

Of course inequality is strict when x and y are interior points. By this v separates P
and Q as requested. 

Asymptotic cones
Given a nonempty convex and closed subset C of Rn , we say that a vector v is an
asymptotic or a recession direction of C if

x + t x ∈ C for all x ∈ C, t  0.

The set of all asymptotic directions of C is denoted by C∞ (Fig. 2.10). It is a convex

cone. It can be seen that a closed convex set is bounded if and only if its asymptotic
cone is trivial. The set C∞ ∩ (−C∞ ) is a linear subspace and called the lineality
space of C.
An equivalent definition of asymptotic directions is given next.

Theorem 2.3.11 A vector v is an asymptotic direction of a convex and closed set C

if and only if there exist a sequence of elements x s ∈ C and a sequence of positive
numbers ts converging to zero such that v = lims→∞ ts x s .

Proof If v ∈ C∞ and x ∈ C, then x s = x + sv ∈ C for all s ∈ N\{0}. Setting

ts = 1/s we obtain v = lims→∞ ts x s with lims→∞ ts = 0. Conversely, assume that
28 2 Convex Polyhedra

C∞

Fig. 2.10 Asymptotic cone

v = lims→∞ ts x s for x s ∈ C and ts > 0 converging to zero as s tends to ∞. Let

x ∈ C and t > 0 be given. Then tts converges to zero as s → ∞ and 0  tts  1
for s sufficiently large. Hence,

x + tv = lim (x + tts x s )
s→∞

= lim (1 − tts )x + tts x s + tts x

s→∞

= lim (1 − tts )x + tts x s .

s→∞

The set C being closed and convex, the points under the latter limit belong to the set
C, and therefore their limit x + tv belongs to C too. Since x and t > 0 were chosen
arbitrarily we conclude that v ∈ C∞ . 

Below is a formula to compute the asymptotic cone of a polyhedron.

Theorem 2.3.12 The asymptotic cone of the polyhedron P determined by the system
(2.6) is the solution set to system

a i , v  0, i = 1, · · · , k. (2.17)

Proof Let v be an asymptotic direction of P. Then for every positive number t one
has
a i , x + tv  bi , i = 1, · · · , k,

where x is any point in P. By dividing both sides of the above inequalities by t > 0
and letting this t tend to ∞ we derive (2.17). For the converse, if v is a solution of
(2.17), then for every point x in P one has

a i , x + tv = a i , x + ta i , v  bi , i = 1, · · · , k
2.3 Convex Polyhedra 29

for all t  0. Thus, the points x + tv with t  0, belong to P and v is an asymptotic

direction. 

Example 2.3.13 Consider a (nonempty) polyhedron in R3 defined by the system:

−x1 − x2 − x3  −1,
x3  1,
x1 , x2 , x3  0.

The asymptotic cone is given by the system

−x1 − x2 − x3  0,
x3  0,
x1 , x2 , x3  0,

in which the first inequality is redundant, and hence it is simply given by x1  0,

x2  0 and x3 = 0.

Using asymptotic directions we are also able to tell whether a convex polyhedron
has a vertex or not. A cone is called pointed if it contains no straight line. When a
cone C is not pointed, it contains a nontrivial linear subspace C ∩ (−C), called also
the lineality space of C.

Corollary 2.3.14 A convex polyhedron has vertices if and only if its asymptotic cone
is pointed. Consequently, if a convex polyhedron has a vertex, then so does any of its
faces.

Proof It is easy to see that when a polyhedron has a vertex, it contains no straight
line, and hence its asymptotic cone is pointed. We prove the converse by induction
on the dimension of the polyhedron. The case where a polyhedron is of dimension
less or equal to one is evident because a polyhedron with a pointed asymptotic cone
is either a point or a segment or a ray, hence it has a vertex. Assume the induction
hypothesis that the conclusion is true for all polyhedra of dimension less than m with
1 < m < n. Let P be m-dimensional with a pointed asymptotic cone. If P has no
proper face, then the inequalities (2.6) are strict, which implies that P is closed and
open at the same time. This is possible only when P coincides with the space Rn
which contradicts the hypothesis that P∞ is pointed. Now, let F be a proper face of
P. Its asymptotic cone, being a subset of the asymptotic cone of P is pointed too.
By induction, it has a vertex, which in view of Corollary 2.3.6 is also a vertex of P.
To prove the second part of the corollary it suffices to notice that if a face of P
has no vertex, by the first part of the corollary, its asymptotic cone contains a straight
line, hence so does the set P itself. 

A second representation result for elements of a convex polyhedron is now for-

mulated in a more general situation.
30 2 Convex Polyhedra

Corollary 2.3.15 A convex polyhedron with vertex is the convex hull of its vertices
and extreme directions.

Proof We conduct the proof by induction on the dimension of the polyhedron. The
corollary is evident when a polyhedron is zero or one-dimensional. We assume that
it is true for all convex polyhedra of dimension less than m with 1 < m < n and
prove it for an m-dimensional polyhedron P determined by system

a i , x = bi , i ∈ I
a j , x  b j , j ∈ {1, · · · , k}\I,

in which |I | = n − m and the vectors a i , i ∈ I are linearly independent. Let y be

an arbitrary element of P. If it belongs to a proper face of P, then by induction we
can express it as a convex combination of vertices and extreme directions. If it is a
relative interior point of P, then setting a = a 1 + · · · + a k that is nonzero because
the asymptotic cone of P is pointed, P itself having a vertex, and considering the
intersection of P with the hyperplane H determined by equality a, x = a, y we
obtain a bounded polyhedron P ∩ H . By Corollary 2.3.8, the point y belongs to the
convex hull of vertices of P ∩ H . The vertices of P ∩ H belong to proper faces of P,
by induction, they also belong to the convex hull of vertices and extreme directions
of P, hence so does y. 

When a polyhedron has a non-pointed asymptotic cone, it has no vertex. However,

it is possible to express it as a sum of its asymptotic cone and a bounded polyhedron
as well.

Corollary 2.3.16 Every convex polyhedron is the sum of a bounded polyhedron and
its asymptotic cone.

Proof Denote the lineality space of the asymptotic cone of a polyhedron P by M. If

M is trivial, we are done in view of Corollary 2.3.15 (the convex hull of all vertices
of P serves as a bounded polyhedron). If M is not trivial, we decompose the space
Rn into the direct sum of M and its orthogonal M ⊥ . Denote by P⊥ the projection of
P on M ⊥ . Then P = P⊥ + M. Indeed, let x be an element in P and let x = x 1 + x 2
with x 1 ∈ M and x 2 ∈ M ⊥ . Then x 2 ∈ P⊥ and one has x ∈ P⊥ + M. Conversely,
let x 1 ∈ M and x 2 ∈ P⊥ . By definition there is some y ∈ P, say y = y 1 + y 2 with
y 1 ∈ M and y 2 ∈ M ⊥ such that y 2 = x 2 . Since M is a part of the asymptotic cone
of P, one deduces that

x = x 1 + x 2 = y 1 + (x 1 − y 1 ) + y 2 = y + (x 1 − y 1 ) ∈ y + M ⊆ P,

showing that x belongs to P. Further, we claim that the asymptotic cone of P⊥

is pointed. In fact, if not, say it contains a straight line d. Then the convexity of
P implies that the space M + d belongs to the asymptotic cone of P. This is a
contradiction because d lies in M ⊥ and M is already the biggest linear subspace
contained in P. Let Q denote the convex hull of the set of all vertices of P⊥ which
2.3 Convex Polyhedra 31

is nonempty by Corollary 2.3.14. It follows from Corollary 2.3.18 below that the
asymptotic cone of P is the sum of the asymptotic cone of P⊥ and M. We deduce
P = Q + (P⊥ )∞ + M = Q + P∞ as requested. 

The following calculus rule for asymptotic directions under linear transformations
is useful.

Corollary 2.3.17 Let P be the polyhedron determined by the system (2.6) and let L
be a linear operator from Rn to Rm . Then

L(P∞ ) = [L(P)]∞ .

Proof The inclusion L(P∞ ) ⊆ [L(P)]∞ is true for any closed convex set. Indeed,
if u is an asymptotic direction of P, then for every x in P and for every positive
number t one has x + tu ∈ P. Consequently, L(x) + t L(u) belongs to L(P) for all
t  0. This means that L(u) is an asymptotic direction of L(P). For the converse
inclusion, let v be a nonzero asymptotic direction of L(P). By definition, for a fixed
x of P, vectors L(x)+tv belong to L(P) for any t  0. Thus, there are x 1 , x 2 , · · · in
ν
P such that L(x ν ) = L(x) + νv, or equivalently v = L( x ν−x ) for all ν = 1, 2, · · ·
ν
Without loss of generality we may assume that the vectors x ν−x converge to some
nonzero vector u as ν tends to ∞. Then

xν x
a i , u = lim a i ,  − a i ,   0
ν→∞ ν ν
for all i = 1, · · · , k. In view of Theorem 2.3.12 the vector u is an asymptotic direction
of P and v = L(u) ∈ L(P∞ ) as requested. 

Corollary 2.3.18 Let P, P1 and P2 be polyhedra in Rn with P ⊆ P1 . Then

(P)∞ ⊆ (P1 )∞
(P1 × P2 )∞ = (P1 )∞ × (P2 )∞
(P1 + P2 )∞ = (P1 )∞ + (P2 )∞ .

Proof The first two expressions are direct from the definition of asymptotic direc-
tions. For the third expression consider the linear transformation L from Rn × Rn to
Rn defined by L(x, y) = x + y, and apply Corollary 2.3.17 and the second expression
to conclude. 

Polar cones
Given a cone C in Rn , the (negative) polar cone of C (Fig. 2.11) is the set

C ◦ := v ∈ Rn : v, x  0 for all x ∈ C .

32 2 Convex Polyhedra

The polar cone of C ◦ is called the bipolar cone of C. Here is a formula to compute
the polar cone of a polyhedral cone.
Theorem 2.3.19 The polar cone of the polyhedral cone determined by the system

a i , x  0, i = 1, · · · , k,

is the positive hull of the vectors a 1 , · · · , a k .

Proof It is clear that any positive combination of vectors a 1 , · · · , a k belongs to the
polar cone of the polyhedral cone. Let v be a nonzero vector in the polar cone. Then
the following system has no solution

a i , x  0, i = 1, · · · , k,
v, x > 0.

According to Farkas’ theorem the system

y1 a 1 + · · · + yk a k = v,
y1 , · · · , yk  0

has a solution, which completes the proof. 

Example 2.3.20 Let C be a polyhedral cone in R3 defined by the system:

x1 − x2  0,
x3 = 0.

By expressing the latter equality as two inequalities x3  0 and −x3  0,

we deduce that the polar cone of C is the positive hull of the three vectors
(1, −1, 0)T , (0, 0, −1)T and (0, 0, 1)T . In other words, the polar cone C ◦ consists
of vectors (t, −t, s)T with t ∈ R+ and s ∈ R.
Corollary 2.3.21 Let C1 and C2 be polyhedral cones in Rn . Then the following
calculus rules hold
Fig. 2.11 Polar cone

C◦
2.3 Convex Polyhedra 33

(C1 + C2 )◦ = C1◦ ∩ C2◦

(C1 ∩ C2 )◦ = C1◦ + C2◦ .

Proof Let v ∈ (C1 + C2 )◦ . We have

v, x + y  0 for all x ∈ C1 , y ∈ C2 .

By setting y = 0 in this inequality we deduce v ∈ C1◦ . Similarly, by setting x = 0

we obtain v ∈ C2◦ , and hence v ∈ C1◦ ∩ C2◦ . Conversely, if v belongs to both C1◦
and C2◦ , then v, · is negative on C1 and C2 . Consequently, it is negative on the sum
C1 + C2 by linearity, which shows that v ∈ (C1 + C2 )◦ .
For the second equality we observe that the inclusion C1◦ + C2◦ ⊆ (C1 ∩ C2 )◦
follows from the definition. To prove the opposite inclusion we assume that C1 is
determined by the system described in Theorem 2.3.19 with i = 1, · · · , k1 and C2 is
determined by that system with i = k1 + 1, · · · , k1 + k2 . Then the polyhedral
cone C1 ∩ C2 is determined by that system with i = 1, · · · , k1 + k2 . In view
of Theorem 2.3.19, the polar cone of C1 ∩ C2 is the positive hull of the vectors
a 1 , · · · , a k1 +k2 , which is evidently the sum of the positive hulls pos{a 1 , · · · , a k1 }
and pos{a k1 +1 , · · · , a k1 +k2 }, that is the sum of the polar cones C1◦ and C2◦ . 

Corollary 2.3.22 The bipolar cone of a polyhedral cone C coincides with the cone
C itself.

Proof According to Theorem 2.3.19 a vector v belongs to the bipolar cone C ◦◦ if

and only if
  k 
v, λi a i  0 for all λi  0, i = 1, · · · , k.
i=1

The latter system is equivalent to

a i , v  0, i = 1, · · · , k,

which is exactly the system determining the cone C. 

Corollary 2.3.23 A vector v belongs to the polar cone of the asymptotic cone of a
convex polyhedron if and only if the linear functional v, . attains its maximum on
the polyhedron.

Proof It suffices to consider the case where v is nonzero. Assume v belongs to the
polar cone of the asymptotic cone P∞ . In virtue of Theorems 2.3.12 and 2.3.19, it is
a positive combination of the vectors a 1 , · · · , a k . Then the linear functional v, . is
majorized by the same combination of real numbers b1 , · · · , bk on P. Let α be its
supremum on P. Our aim is to show that this value is realizable, or equivalently, the
system
34 2 Convex Polyhedra

a i , x  bi , i = 1, · · · , k
v, x  α

is solvable. Suppose to the contrary that the system has no solution. In view of
Corollary 2.2.4, there are a positive vector y and a real number t  0 such that

tv = A T y,
tα = b, y + 1.

We claim that t is strictly positive. Indeed, if t = 0, then A T y = 0 and b, y = −1

and for a vector x in P we would deduce

0 = A T y, x = y, Ax  y, b = −1,

a contradiction. We obtain expressions for v and α as follows

1 T 1 
v= A y and α = b, y + 1 .
t t

Let {x r }r 1 be a maximizing sequence of the functional v, . on P, which means

limr →∞ v, x r  = α. Then, for every r one has

1 T
v, x r  = A y, x r 
t
1
 y, b
t
1
 α− ,
t
which is a contradiction when r is sufficiently large.
For the converse part, let x be a point in P where the functional v, . achieves its
maximum. Then
v, x − x  0 for all x ∈ P.

In particular,
v, u  0 for all u ∈ P∞ ,

and hence v belongs to the polar cone of P∞ . 

Normal cones
Given a convex polyhedron P determined by the system (2.6) and a point x in P, we
say that a vector v is a normal vector to P at x if

v, y − x  0 for all y ∈ P.

2.3 Convex Polyhedra 35

The set of all normal vectors to P at x forms a convex cone called the normal cone to
P at x and denoted N P (x) (Fig. 2.12). When x is an interior point of P, the normal
cone at that point is zero. When x is a boundary point, the normal cone is computed
by the next result.
Theorem 2.3.24 The normal cone to the polyhedron P at a boundary point x of P
is the positive hull of the vectors a i with i being active indices at the point x.
Proof Let x be a boundary point in P. Then the active index set I (x) is nonempty.
Let v be an element of the positive hull of the vectors a i , i ∈ I (x), say

v= λi a i with λi  0, i ∈ I (x).
i∈I (x)

Then for every point x in P and every active index i ∈ I (x), one has

a i , x − x = a i , x − bi  0,

which yields 
v, x − x = λi a i , x − x  0.
i∈I (x)

Hence v is normal to P at x. For the converse, assume that v is a nonzero vector

satisfying
v, x − x  0 for all x ∈ P. (2.18)

We wish to establish that v is a normal vector at 0 to the polyhedron, denoted Q, that

is determined by the system

a i , y  0, i ∈ I (x).

This will certainly complete the proof because the normal cone to that polyhedron
is exactly its polar cone, the formula of which was already given in Theorem 2.3.19.
Observe that normality condition (2.18) can be written as

v, y  0 for all y ∈ cone(P − x).

Therefore, v will be a normal vector to Q at zero if Q coincides with cone(P − x).

Indeed, let y be a vector of cone(P − x), say y = t (x − x) for some x in P and some
positive number t. Then

a i , y = ta i , x − x  0,

which yields y ∈ Q. Thus, cone(P − x) is a subset of Q. For the reverse inclusion

we notice that inequalities with inactive indices are strict at x. Therefore, given a
vector y in Q, one can find a small positive number t such that
36 2 Convex Polyhedra

Fig. 2.12 Normal cone

P
y
x

NP (y) + y
NP (x) + x

a j , x + ta j , y  b j

for all j inactive. Of course, when i is active, it is true that

a i , x + t y = a i , x + ta i , y  bi .

Hence, x + t y belongs to P, or equivalently y belongs to cone(P − x). This achieves

the proof. 

Example 2.3.25 Consider the polyhedron in R3 defined by the system:

x1 + x2 + x3  1,
−2x1 − 3x2  −1,
x1 , x2 , x3  0.

This is a convex polytope with six vertices

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 0 0
v1 = ⎝ 0 ⎠ , v2 = ⎝ 1 ⎠ , v3 = ⎝ 1/3 ⎠ ,
0 0 0
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1/2 1/2 0
v4 = ⎝ 0 ⎠ , v5 = ⎝ 0 ⎠ , v6 = ⎝ 1/3 ⎠
0 1/2 2/3

and five two-dimensional faces

co{v1 , v2 , v5 , v6 }, co{v1 , v2 , v3 , v4 }, co{v1 , v4 , v5 },

co{v3 , v4 , v5 , v6 }, co{v2 , v3 , v6 }.
2.3 Convex Polyhedra 37

At the vertex v1 there are three active constraints:

x1 + x2 + x3 = 1,
x2 = 0,
x3 = 0,

and two non-active constraints

−2x1 − 3x2  −1,

−x1  0.

Hence the normal cone at the vertex v1 is the positive hull of the vectors u 1 =
(1, 1, 1)T , u 2 = (0, −1, 0)T and u 3 = (0, 0, −1)T . Notice that u 1 generates
the normal cone at the point (1/3, 1/3, 1/3)T on the two-dimensional face F1 =
co{v1 , v2 , v6 , v5 }, u 2 generates the normal cone at the point (2/3, 0, 1/4)T on the
two-dimensional face F2 = co{v1 , v4 , v5 }, and the positive hull of u 1 and u 2 is the
normal cone at the point (3/4, 0, 1/4)T on the one-dimensional face [v1 , v5 ] that is
the intersection of the two-dimensional faces F1 and F2 .

As a direct consequence of Theorem 2.3.24, we observe that the normal cone is

the same at any relative interior point of a face. We refer to this cone as the normal
cone to a face. In view of Corollary 2.3.7 we obtain a collection of all normal cones
of faces, whose union is called the normal cone of P and denoted by N P . Thus, if
F := {F1 , · · · , Fq } is the collection of all faces of P, then

q
NP = N (Fi ).
i=1

It is to point out a distinction between this cone and the cone N (P), the normal cone
to P when P is considered as a face of itself. We shall see now that the collection N
of all normal cones N (Fi ), i = 1, · · · , q, is a nice dual object of the collection F.

Theorem 2.3.26 Assume that P is a convex polyhedron given by the system

a i , x  bi , i = 1, · · · , k.

Then the following assertions hold.

(i) The normal cone of P is composed of all normal cones to P at its points, that is

NP = N P (x)
x∈P

and coincides with the polar cone of the asymptotic cone of P. In particular,
it is a polyhedral cone, and it is the whole space if and only if P is a polytope
(bounded polyhedron).
38 2 Convex Polyhedra

(ii) In the collection F, if Fi is a face of F j , then N (F j ) is a face of N (Fi ). Moreover,

if i = j, then the normal cones N (Fi ) and N (F j ) have no relative interior point
in common.
(iii) In the collection N , if N is a face of N (Fi ), then there is a face F containing
the face Fi such that N = N (F ).

Proof For the first property it is evident that N P is contained in the union of the right
hand side. Let x ∈ P. There exists an index i ∈ {1, · · · , k} such that x ∈ ri(Fi ).
Then N P (x) = N (Fi ) and equality of (i) is satisfied. To prove that N P coincides
with (P∞ )◦ , let v be a vector of the normal cone N (Fi ) for some i. Choose a relative
interior point x0 of the face Fi . Then, by definition,

v, x − x0   0 for all x ∈ P.

By Corollary 2.3.23 the vector v belongs to the polar cone of the cone P∞ . Conversely,
let v be in (P∞ )◦ . In view of the same corollary, the linear functional v, . attains
its maximum on P at some point x, which means that

v, x − x  0 for all x ∈ P.

By definition, v is a normal vector to P at x.

For (ii), assume that Fi is a face of F j with i = j, which implies that the active
index set I Fi of Fi contains the active index set I F j of F j . Let x j be a relative interior
point of F j . Then one has

N (F j ) = N P (x j ) ⊂ N (Fi ).

Suppose that N (F j ) is not a face of N (Fi ). There exists a face

N0 = pos{a  :  ∈ I0 } ⊆ N (Fi )

for some I0 ⊆ I Fi , which contains N (F j ) as a proper subset and such that its relative
interior meets N (F j ) at some point, say v0 . Let F0 be the solution set to the system

a  , x = b ,  ∈ I0 ,
a  , x  b ,  ∈ {1, · · · , p}\I0 .

We see that I F j ⊆ I0 ⊆ I Fi , hence Fi ⊆ F0 ⊆ F j . In particular F0 = ∅, hence it is

a face of P. Let x0 be a relative interior point of F0 . We claim that

v, x j − x0  = 0 for all v ∈ N0 .

Indeed, consider the linear functional v → v, x j − x0  on N0 . On the one hand,

v, x j − x0   0 for all v ∈ N0 because x0 ∈ ri(F0 ). On the other hand, for
v0 ∈ ri(N0 ) ∩ N (F j ) above, one has v0 , x0 − x j   0, hence v0 , x j − x0  = 0.
2.3 Convex Polyhedra 39

Consequently, v, x j − x0  = 0 on N0 . Using this fact we derive for every v ∈ N0

that
v, x − x j  = v, x − x0  + v, x0 − x j   0,

for all x ∈ M, which implies v ∈ N (F j ) and arrive at the contradiction N (F j ) = N0 .

To prove the second part of assertion (ii), suppose to the contrary that the normal
cones N (Fi ) and N (F j ) have a relative interior point v in common. Then for each
x ∈ Fi and y ∈ F j one has
v, x − y = 0.

Since u, y − x  0 for all u ∈ N (Fi ) and v is a relative interior point of N (Fi ),
one deduces
u, x − y = 0 for all u ∈ N (Fi ).

Consequently, for u ∈ N (Fi ) it is true that

u, z − y = u, z − x + u, x − y  0 for all z ∈ P,

which shows u ∈ N (F j ). In other words N (Fi ) ⊆ N (F j ). The same argument with

i and j interchanging the roles, leads to equality N (Fi ) = N (F j ). In view of the
first part we arrive at the contradiction Fi = F j .
We proceed to (iii). Let N be a face of N (Fi ) for some i : 1  i  k. The case
N = N (Fi ) being trivial, we may assume N = N (Fi ). Let I ⊆ I Fi be a subset of
indices such that

N = cone{a  :  ∈ I } ⊆ N (Fi ) = cone{a  :  ∈ I Fi } .

Let F be the solution set to the system

a  , x = b ,  ∈ I,
a  , x  b ,  ∈ {1, · · · , p}\I .

Since I ⊆ I Fi , we have Fi ⊆ F. In particular F = ∅ and F is a face of P. Now we

show that N (F) = N and Fi is a proper face of F. Indeed, as N is a proper face of
N (Fi ), I is a proper subset of I Fi and there is a nonzero vector u ∈ R n such that

a  , u = 0 for  ∈ I,
a  , u < 0 for  ∈ I Fi \I .

Take x ∈ ri(Fi ) and consider the point x + tu with t > 0. One obtains

a  , x + tu = b ,  ∈ I,

a  , x + tu = a  , x + ta  , u < b ,  ∈ I Fi \I .

40 2 Convex Polyhedra

Moreover, since a  , x < b for  ∈ {1, · · · , p}\I Fi , when t is sufficiently small,

one also has
a  , x + tu < b ,  ∈ {1, · · · p}\I Fi .

Consequently,
N (F) = N P (x + tu) = pos{a  :  ∈ I } = N

when t is sufficiently small. It is evident that F = Fi . The proof is complete. 

Example 2.3.27 Consider the polyhedron P in R2 defined by the system:

−x1 − x2  −1,
−x1 + x2  1,
−x2  0.

It has two vertices F1 and F2 determined respectively by

⎧ ⎧
⎨ −x1 −x2 = −1 ⎨ −x1 −x2 = −1
−x1 +x2  1 and −x1 +x2 = 1
⎩ ⎩
−x2 = 0 −x2  0

three one-dimensional faces F3 , F4 and F5 determined respectively by

⎧ ⎧ ⎧
⎨ −x1 −x2  −1 ⎨ −x1 −x2 = −1 ⎨ −x1 −x2  −1
−x1 +x2  1 , −x1 +x2  1 and −x1 +x2 = 1
⎩ ⎩ ⎩
−x2 = 0 −x2  0 −x2  0

and P itself is the unique two-dimensional face. Denote by v1 = (−1, −1)T , v2 =

(−1, 1)T and v3 = (0, −1)T . Then the normal cones of the faces F1 ,· · · , F5 are
respectively the positive hulls of the families {v1 , v3 }, {v1 , v2 }, {v3 }, {v1 } and {v2 }.
The normal cone of P is zero. Moreover, the union N P of these normal cones is the
positive hull of the vectors v2 and v3 . It is the polar cone of the asymptotic cone of
P which is defined by the system

−x1 − x2  0,
−x1 + x2  0,
−x2  0,

in which the first inequality is redundant and hence it is reduced to x1  x2  0.

Next we prove that the normal cone of a face is obtained from the normal cones
of its vertices.
2.3 Convex Polyhedra 41

Corollary 2.3.28 Assume that a face F of P is the convex hull of its vertices
v 1 , · · · , v q . Then
q
N (F) = N P (v i ).
i=1

q
Proof The inclusion N (F) ⊆ i=1 N P (v i ) is clear from (ii) of Theorem 2.3.26.
We prove the converse inclusion. Let u be a nonzero vector of the intersection
q
i=1 N P (v ). Let x be a relative interior point of F. Then x is a convex combi-
i

nation of the vertices v 1 , · · · , v q :


q
x= λi v i
i=1

with λi  0, i = 1, · · · , q and λ1 + · · · + λq = 1. We have then

u, x − v i   0 for all x ∈ P, i = 1, · · · , q.

This implies


q 
q
u, x − x = u, λi x − λi v i 
i=1 i=1

q
= λi u, x − v i   0
i=1

By this, u is a normal vector to P at x, and u ∈ N (F). 

Combining this corollary with Corollary 2.3.8 we conclude that the normal cone
of a bounded face is the intersection of the normal cones of all proper faces of that
bounded face. This is not true for unbounded faces, for instance when a face has no
proper face.

2.4 Basis and Vertices

In this section we consider a polyhedron P given by the system

Ax = b (2.19)
x  0.
Random documents with unrelated
content Scribd suggests to you:
 CHAPTER VII
MYSTERY SHIP

“I can’t get over the way that pitcher came to us,”

Goggles Short murmured low to Johnny Thompson.
They were seated in the bleachers. The Saturday game
was about to begin. The new pitcher from the
laboratories, cap drawn low, eyes gleaming, was putting
over a few to the catcher.

“It is strange,” Johnny said. “Prince of India!” he

exclaimed. “I gave him that name and I’m proud of it.”
In his publicity produced by the thought-camera Johnny
had played up the name “Prince of India.” He liked the
sound of it. “He looks the part too! Look at that slim
nose of his,” he went on, “those thin lips, that high
forehead. You’d take him for a Frenchman, or perhaps
an Englishman, if it weren’t for that dark skin of his. If
he’s not a Prince of India, he should be. Watch him
pitch!” The slender man on the mound, moving with the
smooth agility of a cat, seemed to fairly slide the ball
over the plate.

“Listen to the crowd!” Goggles cried. “And is it a crowd! [71]

That publicity stuff of yours was great! We’ll get nearly
half the money we need for that first payment today.
And Wednesday! It’s in the bag.”
“Don’t be too sure,” Johnny warned.

“Listen to that crowd!” Goggles exclaimed once more.

Led by Meggy Strawn, a streak of gold and blue that

danced across the grass, the crowd was chanting:

A Prince! A Prince! A Prince!

No quince! No quince! No quince!
A Peach! A Peach! A Peach!
We win! Yea! Yea!

As for the “Prince,” he seemed totally unconscious of his

surroundings as he slid one more stinger over the plate.

“It is strange,” Johnny said to Goggles, “strange about [72]

that pitcher, I mean. Colonel Chamberlain has had him
working in his laboratories for more than three months.
The pay-roll proves that. But who knew it? The pay-
master and Colonel Chamberlain, that’s all. Queer, isn’t
it? And now, when everything seems lost for old
Hillcrest, he walks right into the picture. He takes the
ball, and whang! How it pops into that old mit! Not a
man will get to first. See! There goes one of ’em. Three
strikes and out. Great, I’d say! Suppose he can keep it
up?”

He did not wait for an answer. Instead, he allowed his

eyes to seek a spot in the sky. Something up there
interested him.

“Nope!” he murmured. “It’s not coming down.”

“What’s not coming down?” Goggles asked quickly.

“That airplane. It’s been circling way up high there for a

long time.”
“I should hope it wouldn’t come down,” Goggles
laughed good-naturedly. “What d’ye think? Think they’d
come right on down and land square in the middle of
the ball field?” He laughed again.

Johnny did not reply. Truth was, he did not know what [73]
he had expected. It was strange about that airplane. He
had been watching it off and on for twenty minutes. All
that time it had been circling above the ball field. At first
it had seemed little more than a speck against the dull
gray of a leaden sky. Moment by moment it had circled
lower.

“Saw an eagle do that once,” he had told himself as a

little thrill ran up his spine. “Old eagle soared and
soared and soared until he was maybe a hundred feet
from the ground. Then he folded his wings and
dropped. And such a drop! Straight down! When he
came up he held a half-grown rabbit in his talons. He’d
had his eye on that rabbit all the time.”

Strangely enough, as he watched the airplane circle

above the ball field where two fine teams were
contending for high honors, fantastic as it might seem,
he had gained the impression that this plane, circling as
the eagle had circled, would in the end make one
straight drop to the ball field.

“What nonsense!” he whispered to himself. “Why should

they do that? Crack up! Everyone in the plane would be
killed. Eagle’s a different sort of bird. He could recover
balance and rise again. That plane—”

All the same, the impression remained a haunting [74]

suggestion until, with the end of the first half, a shut-
out for the opposing team, the Centralia boys went
trotting off the field. Only then did the airplane go
skimming away into the hazy distance.

“It is as if the eagle had been watching the rabbit only

to see the rabbit scurry into his hole,” he told himself.

“But the rabbit will come out again? Another inning?” a

voice seemed to whisper in his ear.

With that, for a time at least, he forgot the strange

airplane and gave his attention to the ball game.

“Hello Meggy,” he said a moment later as she slid into

the place beside him. “We’re going to win, Meg!” he
cried.

Meg’s voice was low. “Yes, we must, Johnny!”

Suddenly Meggy pinched Johnny’s arm. “Look! He—he’s

up to bat! Isn’t he mysterious! The—the ‘Prince of
India’—that’s what they call him.”

Once again Johnny’s eye was on the ball. The opposing

pitcher shot it through to the Prince, but it went high
and wide. The dark-faced one never moved a muscle.

“Believe he can bat,” was Johnny’s mental comment. His [75]

practiced eye swept over the diamond. Arthur Lowe was
on first, Fred Frame on second. There were two men
out. No score on either side.

“Now,” he whispered hoarsely, “just one good swat!

That’s all we need! Get a grand lead! We—”

He did not finish. Came the crack of a bat and the ball
went soaring high and far.
“Yea! Yea! Yea!” The crowd sprang to its feet and
howled madly. “Yea! Yea! Yea! Prince! Prince! Prince!”

When the crowd settled back to its seats the new

pitcher was on third base. Two men had come romping
home.

“Two to nothing!” Meg exulted. “Watch us climb!”

Little Artie Snow was up next. He swung wildly and

fanned. The inning was over.

“Well!” Johnny stretched himself. “Looks as if we’d lick

’em all right.”

All Meggy said was, “Isn’t he mysterious?” She was

thinking of the “Prince.”

Then, as her mood changed, Meggy seized her

megaphone and, grasping Johnny by the arm,
screamed, “Come on! Cart wheels!”

Johnny had done cart wheels with Meggy on many [76]

another occasion, but always in private. But now! Oh
well, Meg was Meg. Her word was law. Cart wheels it
was, an even dozen, then a rousing cheer led by Meg:

Yea! Hillcrest! Yea! Hillcrest!

Beat ’em! Beat ’em! Beat ’em!

Scarcely had Johnny got his breath than he discovered

that the “Prince” was once more on the mound, the
second inning about to begin. Quite automatically his
eyes swept the sky. They came to a focus.

“The airplane!” he whispered excitedly. “Like the eagle,

it is circling back.”
It was strange the excitement this stirred up within his
being. Why was it? It seemed absurd, yet in his soul
there was a feeling that the dark pitcher must hurry,
that the men who came up to bat must go down as they
had before, one, two, three, or else the eagle would
drop. “What nonsense!” he muttered once more.

For all that, the airplane did circle lower and lower.
There was too in the mysterious pitcher’s action a
suggestion of tense nervousness that was hard to
explain.

A bat cracked. A ball popped into the air. The pitcher [77]
had it. One man down.

A second man came up. Ball! Strike! Ball! Crack! Up

went the ball again. Down it came, right into that
pitching wizard’s mit. Two out.

The plane circled lower. In the damp, cloudy air it

seemed nearer than it really was.

Third man to bat. Strike! Strike! Strike! You’re out!

“Just like that!” Johnny exulted. He did not so much as

glance at the plane. He knew that once again it had
gone skimming away.

“It’s strange,” he murmured.

“What’s strange?” Meggy asked.

“Oh—everything,” he evaded, “everything’s strange

today.” How could he tell Meggy of this fantastic
daydream?
Again the opposing team took their places in the field.
Once more Hillcrest came to bat. And how they did bat!
Inspired by rosy dreams of victory, they sent the ball
spinning, right, left and center. By the time Centralia
had them stopped, the score stood 5 to 0 in favor of
Hillcrest, and the crowd had gone mad.

“We’ll win!” Meggy screamed. [78]

“We’ll win!” Goggles roared.

As for Johnny, he merely murmured, “Wait!”

The wait was destined to be longer than he dreamed it

might be. Four wild balls put the lead-off man of
Centralia on first with no one out.

It was then that Johnny once more began noticing that

haunting airplane. It had returned. Once again it was
circling downward.

The mysterious pitcher was slipping, there could be no

doubting this. A hard-hit liner put the second batter on
base.

Then the pitcher seemed to tighten up. He fanned the

third man.

“But that plane!” Johnny was truly startled now. The

plane did actually seem to be in a nose-dive. Down,
down, down it came, straight at that lone figure, the
pitcher, on the mound.

“They—they—” In his excitement Johnny stood up. He [79]

crushed his cap within his tight clenched hands. “No!
No! Thank—” He did not finish. With a burst of speed, a
thunder of motors, the airplane righted itself, then shot
upward. But what was that? Did Johnny’s eyes deceive
him? Did he catch a gleam of fire—or was it only a
brilliant flash of light? Half unconsciously he waited the
report of a shot fired. It did not come.

“It’s the strangest thing!” he murmured as he settled

back in his place. Already the airplane was a long way
off.

So filled was the boy’s mind with wild speculations that

he failed to follow the game. Perhaps this was just as
well. Dame Fortune appeared to have deserted the
mysterious pitcher. He walked another man. The bases
were full.

“But look at him,” Meggy whispered in Johnny’s ear.

“Look at him wind up! You’d think he was doing it in his
sleep!”

Indeed, as Johnny focussed his attention upon this

mysterious stranger, he appeared to waver, as if he
might fall.

“Something awfully queer about that,” Johnny

murmured.

With what appeared to be tremendous effort the pitcher

hurled the ball. It would have cut the plate squarely in
the middle had not a stout bat met it to send it high and
far.

When the commotion was over, the score stood 5 to 6 [80]

in favor of Centralia. There were men on second and
third. What was more, the “Prince” was walking
unsteadily toward the bench.
“Listen!” Meggy exclaimed. “They’re calling for Fred
Frame.”

“Something queer about that!” Johnny repeated as he

turned to watch the “Prince” walk away toward the
showers. “The eagle swooped downward, and now—”
he did not finish.

“He walks as if he were half blind. Poor ‘Prince!’” Meg

sympathized. “What could have happened?”

Johnny would have given much to know the answer. For

some time to come it was to remain a veiled secret.

“The mystery ship,” Johnny thought as he watched that

airplane glide away toward the clouds. Then he
murmured low, “Mystery wings.”

“‘Mystery wings!’ What makes you say that?” Meg

whispered.

“Because that’s the way I think of a plane,” he replied [81]

soberly. “You can’t say the planes of an airplane. Don’t
sound right. Why not wings of a plane? And, for my
part, every plane that passes over my head has wings
of mystery.”

“You’re queer,” was Meg’s only reply.

[82]
 CHAPTER VIII
STRANGE PASSENGERS

Among the Hillcrest fans feeling was running high. That

something strange and rather terrible had happened to
their new and quite marvelous pitcher, they appeared to
realize. “But what did happen?” they were asking.
“Who’s to blame? Who were the men in that plane?”
Two men had been seen. They were not close enough
to be recognized. Had the Centralia crowd hired them to
heckle the new pitcher? This they found it difficult to
believe. The friendliest of relations had always existed
between the two small cities, even though there was a
keen rivalry. “But who? Who? Who?” they were asking
on every side. The mystery of the dark-skinned pitcher
from the laboratories deepened.

As for Doug Danby, on whose shoulders rested

Hillcrest’s hopes of victory, he found no time for solving
mysteries.

“Fred, old boy,” he said to Fred Frame, “you’ll have to [83]

go in there and win the game. And you can!” He gave
him a slap on the back. “If—”

“If my arm holds out,” Fred finished.

Tall, angular, red-headed, silent and droll, Fred was a
universal favorite. He had been a successful pitcher until
his arm had taken to going wrong. “I’ll go in,” he said
simply, “and do my best.”

A loud cheer greeted him as he walked toward the

mound. Despite all this, he felt a chill run up his spine.
The score stood 6 to 5 against him. This wonderful
crowd had turned out to see their team win. They had
banked heavily on the mysterious “Prince.” In this they
had lost. Would they lose the game as well?

“Not if I can help it!” Fred set his teeth hard.

“What if that plane returns?” He shuddered. “What if

they do to me the thing they did to the ‘Prince,’
whatever that was! Oh well!” He set his shoulders
squarely.

But now the shouts of the throng brought him back to

earth. Motioning the batter to one side, he prepared to
“throw a few over.”

As his hand grasped the ball, as his muscles began [84]

playing like iron bands, as the ball went speeding to cut
the plate and land with a loud plop in the catcher’s mit,
all else but the game was forgotten.

“We must win!” He set his lips tight.

And indeed they must. They had lost one game, could
not afford to lose another.

That he was in a hard spot he knew quite well. With the

score standing 6 to 5 against him, with men on second
and third and only one man out, the game might be lost
with a single crack of the bat. It was with a rapidly
beating heart that he motioned the batter up.

Yet, even as his arm went back, two questions flashed

through his mind: “Who is this ‘Prince’? What happened
that after such a brilliant start he was unable to finish?

“Something queer!” he muttered for the third time as he

sent the ball spinning.

“Ball!” the umpire called.

Then, like a bolt from the blue came a thought. He

made a sign to the catcher. They met half way between
the mound and the home plate. After a few whispered
words they parted.

Fred’s second offering went very wide of the plate. He [85]

did not seem to care. Then, just as he wound up for the
third pitch, someone caught on.

“He’s goin’ to walk that batter!” a big voice bellowed

from the bleachers of the opposing team. “Big League
stuff! Walking Billy to get at Vern!”

At once there was a mad roar that ended in hisses and

boos.

Little Fred cared for that. If he wished to walk

Centralia’s toughest batter to get at a weak one, it was
his privilege. “And after that?” an Imp seemed to be
whispering in his ear. All the same the passed batter
went down to first. The bags were loaded.

“If I slip now—” he thought. “Just listen to them howl!”

He gripped the ball hard.
“Wow! He’s got a rubber arm!” a big voice roared as the
umpire called another ball.

There was silence as Fred slipped over a strike.

Again that roar with the second ball.

“Strike!”

“Ball!”

“There you are!” the big voice roared. “Two and three! [86]
Let’s see you get out of that!”

Fred caught his breath. Bases full. Three balls, two

strikes, and—“If only the old soup-bone holds out!” he
murmured.

His hand went out. It came back. He shot the ball

straight from the shoulder. Then, without knowing why,
he followed the ball. Lucky break! The batter connected.
He sent a bouncer straight into Fred’s mitt and he half
way to the plate. With a mad dash he was there to cut
off the run to the plate. Next he sent the ball speeding
to first.

“Double play! Double play!” the crowd roared. And so it

was. The inning was over. For the moment, at least, all
was well.

Inspired by his unusual success in pulling his team out

of the hole, Fred pitched the remaining innings with the
skill of a genius. He allowed only five hits, and left but
three men on base. Hillcrest scored three runs in the
seventh, to cinch the game. In the end Fred was carried
from the field in triumph.
“Another big day Wednesday, and we’ll win!” exulted
Doug Danby.

“Don’t get too much excited,” he warned Johnny and [87]

Meg as they came rushing up to congratulate him. “This
is not the end. It is only the beginning. We must win
again and again. It’s going to take a real campaign to
gain our end.”

“Don’t worry!” Johnny laughed. “The way Fred pitched

those last innings, there’s not a team that can stop us.”

“There’s where you’re wrong.” It was Fred who spoke.

He had just come up to them.

“What do you mean?” Johnny asked in surprise.

“Well—” Fred paused to ponder. “Well, you know there

are times when you do things and you say to yourself, ‘I
can do this as often as I choose.’ Then there are times
when you feel all sort of lifted out of yourself and you
do things well without seeming to try. But when it’s all
over you say, ‘That was great! But I better never try
that again. If I do, I’ll fail.’ This afternoon was just like
that. Johnny, I wouldn’t like to face that situation again,
ever!” Fred’s tone was so serious that for a full moment
no one spoke.

It was Fred himself who at last broke that silence.

“But then, there’ll not be the need.” He smiled. “Our old [88]
friend, the ‘Prince’ will lead us to sure victory next time.”

“The ‘Prince’!” Doug turned to Meggy. “Where did your

uncle find him, Meggy? Who is he? Where’s he been
hiding?” Meggy was Colonel Chamberlain’s favorite
niece.
“I don’t know,” Meggy admitted.

“But your uncle said he’d been working down at his

laboratories for more than three months!” Johnny
protested.

“Ye-es,” Meggy replied slowly, “and I suppose that

should make him my first cousin! But it doesn’t. I never
saw him before, nor heard of him either. Uncle doesn’t
tell me much about the laboratories. There are always
so many secret investigations going on down there, so
many processes being developed—things he can’t talk
about—that—well, I guess he thinks it’s best to say
nothing at all about any of it. And I suppose,” she
added, “this pitcher is just one more secret.”

“But why would he hide out so?” Doug Danby asked.

“He just doesn’t wish to be recognized, that’s why,”

Johnny said in a tone that carried conviction.

“In a town like this?” Doug exclaimed. “It sure does [89]
seem strange!” Had he but known it, those were the
very words that were passing from lip to lip all over this
quiet little city. “A strange pitcher! A mysterious dark
stranger! And in a town like this!” That was what they
were saying. And, almost without exception, the answer
was, “Just think, in a town like this!”

“Well anyway,” Fred said, “he can pitch! And that’s just
what we need. We’ll just have to have him next
Wednesday when we go against Fairfield. They’re the
toughest battling bunch we’ll play for a long time. You
can’t count on me to lick them.”

“The ‘Prince’ only lasted two and a half innings,” Doug

suggested.
“Yes, but some—” Johnny did not finish. What he
started to say was, “Something rather terrible happened
to him.” After all, he had only guessed that; could not
prove it.

“Well,” Johnny said, “I gotta be anklin’ on home.

Goodbye, Meggy. Goodbye, boys.”

A half hour later he was seated on a ridge that lay

above the town. Beneath him was a long, low building.

“The laboratories!” he whispered. “Place of mystery. [90]

Home of the mysterious ‘Prince.’”

His whole being was stirred. It was not that he

suspected any wrong of those who worked behind
heavily glazed windows in the laboratories. Far from
that. Colonel Chamberlain had always been counted
among Hillcrest’s foremost citizens. The laboratories
belonged to him.

“I’ll have to hunt up Goggles,” Johnny told himself.

“Wonder where he went? He always knows a lot. He
may know more than I do about this pitcher.”

Goggles was a thinker. He was the only boy ever

entrusted with Colonel Chamberlain’s secrets. He alone,
of all the town’s boys, had crossed the threshold of the
laboratories. Only he had seen something of that which
went on inside.

“They test all sorts of things in there,” he had confided [91]

to Johnny one day, “soap and silk, dyes, and all sorts of
powerful drugs. They try to find things out, to do things
that have never been done before, like making rubber
out of crude petroleum or paper out of sunflower stalks.
They succeed sometimes, too. See!” He had pulled a
sheet of paper from his pocket. “Made from a sunflower
stalk. Pretty good paper, eh?

“When they make a real discovery,” he went on, “they

sell it to some great manufacturer.

“Colonel Chamberlain—” he had taken a deep breath.

“He showed me a lot of things I can’t talk about. He
says maybe some day I can work with him in the
laboratories. Boy! Won’t that be grand!”

“Yes, I shouldn’t wonder if Goggles knows something

about this ‘Prince,’” Johnny said to himself now.

He broke short off to stare down at the laboratories.

Someone had come walking down the gravel path. He
walked slowly. “Seems to drag his feet,” Johnny
whispered. Just then the newcomer looked up toward
the sun. Johnny got a full view of his slim, dark face. It
was the ‘Prince.’ A moment more and the long, low
place of mysteries had swallowed him up.

That evening Johnny searched in vain for Goggles. [92]

Goggles’ mother did not know where he was, nor did
anyone else. Johnny decided to go on a little detective
cruise all by himself. Mounting his bicycle, he rode east
nine miles to the Shady Valley landing field. In the office
he found two men in aviators’ uniforms playing
checkers.

“Say!” he said in a subdued voice, “Did any of you fly a

plane over the Hillcrest ball field this afternoon?”

“Yes, I did.” The younger of the two men looked up

quickly. “Why?”
“Oh nothing I guess.” Johnny dropped into a seat
prepared to watch the game.

Though for a full quarter hour he said never a word, the

young aviator looked at Johnny in a queer way many
times.

“Well, what about it?” he said, turning to Johnny when

the game was over.

“Nothing I guess,” Johnny repeated.

“That was a queer business,” the aviator chuckled, “that

flying over your field. Had two passengers, sort of hard
lookers, but well-dressed. Said they lived in Hillcrest.
They wanted to go over the ball game. Kept telling me
to circle down, down, down. Then they’d say, ‘No! Not
now! Up again!’ They repeated that little trick three
times.”

“I know,” Johnny breathed.

“You know?” the young aviator stared. [93]

“Of course I do. Go on.”

“Well—” the aviator cleared his throat. “The third time

we went down closer than I like to. Then we flew away.
Sort of queer, I’d say!” He shot Johnny an enquiring
look.

“Did they carrying anything?” Johnny asked.

“Nothing that I saw.”

“No gun or anything like that?”

“Of course not. What do you think? Think we operate a
bombing plane or something?”

“No, not quite that.” Johnny lapsed into silence.

“Queer business!” The aviator stared at him hard. “What

do you know about it?”

“Nothing much I guess.” Johnny’s tone did not change.

“Only thought I might.”

“But look!” the aviator exclaimed. “If you think that’s

queer, listen to this one. A short while back I took a long
trip, thousand miles or more. Flew it at night. Passenger
told me where to go and where to land.

“Place we landed was all light when we were coming

down. It went dark the minute we landed.

“Two men in uniform came rushing up. One said, ‘Say! [94]
Where do you think you are?’

“‘Don’t know,’ I said.

“‘Well, you’d better,’ one of them yelled. ‘This is a

Federal prison. Move out of here quick!’

“‘Guess we’d better leave right away.’ That’s what my

fare said to me.”

The aviator paused for breath. Johnny was staring.

“Wait! That’s not all!” The aviator waved a hand. “The

lights came on, bright as day, just long enough for me
to taxi across the enclosure and rise; then all went dark.

“And listen!” He paused once more. “When my fare left

the plane, there was a man with him, a slim, dark-faced
man. He came from that prison. I’d swear to it! Can you
beat that?”

“Looks like a jail delivery.” Johnny spoke low. “Should

think you’d be afraid!”

“I would,” the aviator settled back in his chair, “only the

man who went with me that night, my passenger, was
one of the best known and most highly respected
citizens in this part of the country. I was hired by him.”

“Slim, dark-faced man,” Johnny murmured to himself, [95]

recalling the aviator’s words as he rode home a short
time later.

[96]
 CHAPTER IX
“WHO’S AFRAID OF A
CHINAMAN?”

Next morning Johnny wandered over to the Sentinel

office. He wanted to thank the editor for the fine
publicity he had given the game. More than this he
always had enjoyed a half hour in the box-like office of
C. K. Lovell, or “old C. K.” as the people of the city had
come to call him.

C. K. was something of a character. More than six feet

tall, a broad-shouldered, slouching figure of a man, with
masses of gray hair and bushy eyebrows, slumped down
in his office chair, he resembled a shaggy St. Bernard
dog basking in the sun.

“H’llo Johnny!” he greeted. “Fine game yesterday. Sort

of queer, though. Rather unusual about that pitcher!
And did you notice that airplane? What did you make of
that?”

“Haven’t got it made yet.” Johnny dropped into a chair.

“Tough about that pitcher though. It must not happen
again.
“But say!” he enthused. “Wasn’t that a grand crowd! [97]
Boys owe you a lot.”

“Oh, that’s nothing,” the editor laughed good-naturedly.

“Boys deserved it. Fine lot of boys. Be a bigger crowd
than ever next week. What about that electric umpire?
Think it will work?”

“Sure will.”

“Call strikes and balls, and all that?”

“Sure will, C. K.”

“Dodge pop bottles too?” C. K. laughed.

“No. Pop bottles would be bad for his eyes. Got forty
eyes, that umpire has.” Johnny laughed. “Guess the
crowd will go easy on that, though.

“You see,” Johnny went on as the editor showed his

interest by hitching up in his chair, “an electric eye is
like a radio tube. When a beam of light is sent to it from
across a space it stays just so until the light is shut off
by some object, say a baseball. Then it sets up a howl.
If you connect it with a phonograph attachment, you
can make it call out ‘Foul ball!’”

“Interesting if true,” C.K. drawled. “Sure ought to draw

a crowd.

“Say Johnny!” The editor leaned forward to speak in a [98]

tone little more than a whisper. “Heard anything about
Federal agents being around town?”

“Federal agents!” Johnny stared. “No. What for?”

“I’ve heard they’re looking for a Chinaman, a little fellow
—name’s Tao Sing, I believe.”

“Tao Sing!” Johnny started. A mental picture of Tao Sing

in the small room at the back of the Chinese spice shop
flashed into his mind.

“Thought I knew them all,” said C.K. “This must be a

new one.”

“Why should Federal agents want a Chinaman? Who’s

afraid of a Chinaman?” This last slipped from Johnny’s
lips unbidden.

“Who’s afraid of a Chinaman!” C.K. sat up straight quite

suddenly. “Plenty of people afraid of a Chinaman,
Johnny. Plenty!

“A Chinaman looks dull and sleepy enough,” he went on.

“So does a big old tom cat. But let a dog come around
the corner and see what the cat does to him. A
Chinaman’s like that. He’ll go up like a rocket most any
time.

“I worked down near Frisco’s old Chinatown, Johnny, [99]

years ago,” he went on. “Got to be sort of an amateur
guide. Went with the police when they raided Chinese
gambling joints and opium dens. Say! I can hear the
steel door bang yet when the first Chink gave the
warning. Bang! Bang! Bang! And sometimes it wasn’t a
door that banged either.” His voice dropped. “Johnny,
things happened there I wouldn’t dare tell about—not
even now. And that was a long time ago, a long time
ago.” C.K. settled back in his chair.

“Well, I—” Johnny got to his feet a trifle unsteadily.

“Guess I better get going.”
“Don’t hurry, Johnny.”

“Got to go.”

Johnny did hurry. He was afraid he might tell what he

knew about Tao Sing. He was not ready to do that—not
just yet.

“But boy, oh boy!” he whispered. “Would what I know

about that little Chink make front page stuff! First
column in every city!” He could see it now: “CHINAMAN
INVENTS THOUGHT-RECORDING CAMERA. NO MAN’S THOUGHTS
HIS OWN.”

He was sorely tempted to release the story at once. On

sober thought, however, he decided he was not ready to
do that—not yet!

“So they’re looking for Tao Sing, those Federal agents,” [100]
he thought. “Wonder why? Wonder if the think-o-graphs
and the thought-camera have anything to do with that?”
He recalled his visit to the Chinese Chamber of
Commerce, of the pictures he had taken of Wung Lu’s
thoughts and how he had delivered them to Tao Sing.
The thought was disturbing. “Ought not to have been
snooping round gathering up another fellow’s thoughts,
then peddling them to someone else,” he grumbled.
“And yet—” ah yes, and yet—if he had not done that he
could not have had the thought-camera for his own use.

“I’ll use it a lot more,” he assured himself. “Find out all

sorts of queer things for C.K. He’ll run them in his paper
and make a scoop.”

But would he return to the Chinese Chamber of

Commerce? Well, not right away. He recalled what C.K.
had said of things that had happened in old Frisco’s
Chinatown, and a chill ran up his spine.

“Fellow’d think—”

No matter what he’d think. Here was Goggles.

“Look here, Johnny!” Goggles exclaimed. “I heard you [101]

found out about that airplane that was over the ball
field yesterday.”

“Didn’t find out much—just that a pilot from over at the

flying field took two men up.”

“Who were they?”

“Wish I knew. The pilot said they were from Hillcrest.”

“If they were it should be easy to find them. Not many

new people in Hillcrest. Only about half a dozen stop at
the hotel. Rest live in houses. I’ll get ’em. Give me
time.” Goggles’ big eyes gleamed behind his thick
glasses. “I’m an amateur detective, Johnny.”

“Done a little of that myself,” Johnny said with a grin.

“Not in a small city, though. Guess I’ll leave that to you.”

“I’ll find ’em, Johnny.” Goggles was away.

Johnny smiled as he watched him hurry down the

street. Goggles sure was an interesting boy. He dug into
everything just as a gopher digs into the earth.
Chemistry, electricity, detective work, it was all the same
to him.

“Little cities are surely interesting,” was Johnny’s mental [102]

comment. “In big cities everyone tries to be just like
everyone else. People think alike, walk alike, dress the
same, everything. In a little city everyone is different.”

Then he brought himself up with a jerk. There was the

thought-camera. Somehow, since talking to C.K. about
the Chinese, he found himself all but overcome with a
desire to hide the thought-camera in some very dark
and secret spot. In the end, after hurrying home, he
buried it deep among the clothes in his trunk, locked
the trunk, then hid the key.

“So they’re after Tao Sing!” he murmured low. “Wonder

if they’ll find him. And if they do, I wonder—” he did not
finish that last wonder.

[103]
 CHAPTER X
CLUES FROM THE DUST

“In cases like this—” Goggles’ eyes bulged behind his

thick glasses. His beak-like nose appeared to wrinkle
and wriggle as a rabbit’s. “In a case like this,” he
repeated, “one may learn a great deal from dust. Take a
vacuum cleaner now. It’s queer. I’ve helped clean
dozens of furnished houses and apartments after the
tenants were gone. Some of them would scrub the
place till it shone like a new dollar. But the vacuum
cleaner! What do you think?” He paused. “Always half
full of dust!

“And yes!” he exclaimed. “Same here. A good big lot of

dust. I’m prepared. See!” He drew a stout paper sack
from his pocket. Unfastening the cloth dust-bag from
the vacuum cleaner, he proceeded to empty its contents
into the paper sack.

“Dust?” said Johnny, “What can you do with dust?”

“You wait,” said Goggles, “You’ll see.” [104]

“Well, you can have your dust,” Johnny grumbled. “Can’t

see how that can help any.”
Since his visit to the landing field, Johnny had been
more convinced than ever that the presence of that
airplane above Hillcrest baseball grounds on that day
when the mysterious “Prince” had somehow been forced
from the mound, had meant something very strange.

“Up to something, that’s what they were!” he had told

himself. “And I’m going to find out what.”

Recalling Goggles’ suggestion regarding the manner in

which these men might be found, he hunted him up on
the following day.

“Found out anything?” he asked.

“No, but I’m going to,” Goggles replied. “It should not
be hard. They live here. They’re strangers in town.
They’d rent furnished rooms. All we have to do is to
check up on rentals.”

They had checked up and they had, they believed,

found the very place they were looking for. The
description of the two men who had rented a small
furnished bungalow tallied with that of the men they
sought.

There was only one hitch—the men had checked out of [105]
the bungalow.

“That’s too bad!” Johnny had mourned. “I hoped to

catch up with them. It’s not so much what they’ve done
as what they may do. It’s my theory that they have a
grudge of some sort against the ‘Prince.’ He’s got to
pitch some more games for us if we are to win. Those
men will do something more, perhaps something a
great deal worse.”
“What will we do if we find them?” Goggles had asked.
“You can’t prove anything.”

“Proof is what we want.”

“You can find clues in an empty house,” Goggles had

declared. “Plenty of them. It doesn’t matter that they’re
gone. Left all sorts of clues behind. Take dust, for
instance. You get the keys and we’ll go right over
there.”

So here they were in the recently deserted bungalow.

Here was Goggles industriously collecting dust while
Johnny tiptoed softly from room to room, pulling out
drawers without a sound and, after peering within,
softly closing them again.

“Dust!” he mumbled, “What good is a lot of dust? You’d

think—”

He broke off short to stare. In the drawer just before [106]

him his eyes took in two objects. One was a small dry
battery of an unusual shape. On the end of this was a
threaded attachment that apparently just fitted into the
small end of the other object. This second object was a
funnel-shaped tube a foot long. It was an inch across at
one end and three inches at the other. The inside of this
tube shone with an unusual brilliancy.

“Queer business!” Johnny murmured. These objects

were quickly transferred to the inner pocket of his coat.
The drawer was softly closed.

It would seem that he was not a second too soon, for

from below came the sound of an opening door, then a
gruff voice:
“Well son, you’re cleanin’ the place up a bit.”

The voice sent a chill coursing up Johnny’s spine. It was

the voice of a stranger. He was talking to Goggles.

“Yes, I—” Goggles’ answering voice sounded unsteady

and weak. “I do this sort of thing quite—quite a lot. Sort
of—of dust up a bit.”

“Well now that’s fine!” (It did not sound fine to Johnny.) [107]
“But me and my pardners here moved out of this place
a short while back. We came here to get a few things
we forgot, didn’t we Joe?”

“Yep, that’s right,” a second gruff voice replied.

“Them shoes now,” the first voice went on. “We left ’em.
See you got ’em all cleaned up for us.” Goggles had
found a pair of shoes and had scraped the mud from
them in search of clues.

“Yes, I—” Goggles’ voice faded out.

“Well that’s O.K., buddy,” said the first voice again.

“We’ll just get on into the little bedroom and look for a
thing or two.”

“The little bedroom.” That was where Johnny found

himself at that instant. Like a rabbit that has sighted a
dog, he was up and silently away. In truth he went out
of the side door to vanish into the shadows of a broad
old pine tree.

Well enough that he did too, for a moment later he

heard one of the strangers say to Goggles in a tone not
so friendly:
“Boy! We left something in a dresser drawer in that little [108]
bedroom. You cleaned in there yet?”

“No, I—I’ve not been out of this room.” Goggles

stammered a little, but had spoken the simple truth.

After looking him over from head to foot, the speaker

turned on his heel and left the house. He was followed
by his pardner.

“Whew!” Goggles breathed five minutes later, “What do

you think of that?”

“I think,” said Johnny, “think—. Come on! Let’s get out

of here! I got ’em in my pocket.”

“Got what?”

“The things they came back after.”

“Let’s see!” Goggles held out a hand.

“Not now. I say, let’s go!”

“All right,” Goggles agreed reluctantly. “Guess I’ve got all

the dust I need.”

After locking the door, they hurried away to Goggles’

basement where he had rigged up a sort of laboratory
and workshop.

“Now,” Goggles breathed, snapping on the light, “we’ll

have a look at that stuff from the sweeper.” He emptied
the contents of the paper sack into a sheet of wrapping
paper.

“Now.” With a needle set into the end of an old pen- [109]
holder, he began dragging the stuff about, at the same
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookbell.com