EURO Advanced Tutorials on Operational Research

Series Editors: M. Grazia Speranza · José Fernando Oliveira

Cláudio Alves
François Clautiaux
José Valério de Carvalho
Jürgen Rietz

Dual-Feasible
Functions for Integer
Programming
and Combinatorial
Optimization
Basics, Extensions and Applications
EURO Advanced Tutorials on Operational
Research

Series editors
M. Grazia Speranza, Brescia, Italy
José Fernando Oliveira, Porto, Portugal
More information about this series at http://www.springer.com/series/13840
Cláudio Alves • François Clautiaux •
José Valério de Carvalho • JRurgen Rietz

Dual-Feasible Functions
for Integer Programming
and Combinatorial
Optimization
Basics, Extensions and Applications

123
Cláudio Alves François Clautiaux
Department of Production and Systems Institut de Mathématiques de Bordeaux
University of Minho University of Bordeaux
Braga, Portugal Talence, France

José Valério de Carvalho JRurgen Rietz

Department of Production and Systems Centro Algoritmi
University of Minho University of Minho
Braga, Portugal Braga, Portugal

ISSN 2364-687X ISSN 2364-6888 (electronic)

EURO Advanced Tutorials on Operational Research
ISBN 978-3-319-27602-1 ISBN 978-3-319-27604-5 (eBook)
DOI 10.1007/978-3-319-27604-5

Library of Congress Control Number: 2015960796

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)
Preface

The concept of dual-feasible function (DFF) has been used to improve the resolution
of several combinatorial optimization problems involving knapsack inequalities like
cutting and packing, scheduling, and vehicle routing problems. DFF were used for
the first time by Lueker (1983) to obtain lower bounds for the bin-packing problem.
Since then, the main application of DFF was in the computation of lower bounds,
even though other applications do exist, as, for instance, the generation of valid
inequalities for integer programs (Chvátal 1973).
During many years, DFF were seen only as mere rounding functions that lead to
lower bounds for standard packing problems by changing the value of the input data.
In this tutorial, we bring a broader perspective to the subject by discussing it within
the general framework of duality. A revision of the standard concepts, properties,
and instances is provided with illustrative examples. We show that many lower
bounds derived for packing problems can be expressed as DFF. Also we explore
relevant extensions of standard DFF and their application to different combinatorial
optimization problems.
The link between DFF, column generation models, and the underlying Dantzig-
Wolfe decomposition is strong. The classical DFF rely on the dual perspective of the
well-known column generation model of Gilmore and Gomory for the cutting stock
problem. Many functions were proposed within this specific context. We explore
the general properties that identify the best DFF. Additionally, we describe the
general approaches that can be followed to derive new non-dominated functions. In
particular, it is shown how to derive high-quality DFF from superadditive functions
using symmetry. We show how to use them to derive concrete examples, and we
further illustrate these ideas through the analysis of some of the current DFF that
lead to the best results reported in the literature.
A first generalization of the classical DFF can be done by considering the general
formulation of a set covering model somehow disconnected from the cutting stock
problem. Here, we analyze this generalization and explore how it can influence (or
act upon) the general properties of classical DFF. Extending DFF to nonclassical
formulations is usually not straightforward. In this tutorial, we show how this
extension can be done for different cases including the existence of two-dimensional

v
vi Preface

rectangular items, conflicts, and general domains. Examples are provided to guide
the readers through the rationale behind the development of these extensions and to
provide the basis for future contributions.
Extensions and applications within the scope of integer and linear programming
will be also discussed. Recent developments extending DFF to the domain of
negative values are described. These generalized DFF can be used to derive valid
inequalities for integer programs from a single constraint or a set of generating
constraints.
This monograph was primarily written for graduate students and also advanced
undergraduate students in operations research/management science, computer sci-
ence, and mathematics. It requires a knowledge of linear and integer programming
and duality theory. Chapters 1 and 2 present the basic concepts and definitions
of DFF. Chapter 3 contains extensions of DFF to wider domains. Chapters 4
and 5 address applications in cutting and packing problems and in deriving valid
inequalities, respectively. Exercises are also proposed at the end of each chapter
covering the different parts of the material. Solutions or hints to a selected set of
exercises are provided at the end of the book.
The authors have been using DFF in research for many years, and these proved
to be very useful for the efficient computation of lower bounds for many different
integer programs and combinatorial optimization problems, with nice computational
results. Additionally, they also used them to derive valid inequalities. DFF are
problem dependent. We hope that the insight provided by this monograph may foster
research and a more widespread use of DFF in other problems. The future still holds
much to be discovered in this area.
We would like to thank many colleagues and researchers, who discussed many
issues and stimulated our interest and research in this area. The authors also would
like to thank Gleb Belov, Julia Bennell, Rita Macedo, and Daniel Porumbel, who
did a careful review of an earlier draft of this manuscript and helped to improve its
quality.
This work was supported by FEDER funding through the Programa Operacional
Factores de Competitividade (COMPETE) and by national funding through the
Portuguese Science and Technology Foundation (FCT) in the scope of the project
PTDC/EGE-GES/116676/2010 (Reference from COMPETE: FCOMP-01-0124-
FEDER-020430), and by FCT within the project scope UID/CEC/00319/2013.

Braga, Portugal Cláudio Alves

Talence, France François Clautiaux
Braga, Portugal José Valério de Carvalho
Braga, Portugal Jürgen Rietz
October 2015
Contents

1 Linear and Integer Programming .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.2 Dantzig-Wolfe Decomposition . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
1.2.1 Reformulation of the Original Model .. . . . .. . . . . . . . . . . . . . . . . . . . 3
1.2.2 Dantzig-Wolfe Decomposition in Integer Programming.. . . . . 4
1.3 Structure of DW-Decomposition Models . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6
1.3.1 Gilmore and Gomory Model for the Cutting Stock
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7
1.3.2 Block Angular Structure.. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8
1.3.3 Parallel Non-identical Machine Scheduling .. . . . . . . . . . . . . . . . . . 9
1.3.4 Solution of DW-Models with Column Generation .. . . . . . . . . . . 10
1.4 Duality and Bounds from Dual Feasible Solutions .. . . . . . . . . . . . . . . . . . . 11
1.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 12
1.5.1 One-Dimensional Cutting Stock Problem .. . . . . . . . . . . . . . . . . . . . 12
1.5.2 Vector Packing Problem .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 14
1.6 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 16
1.7 Exercises.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 16
2 Classical Dual-Feasible Functions. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
2.2 Properties .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 24
2.2.1 Maximality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 24
2.2.2 Maximality of Convex Functions . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
2.2.3 Extremality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
2.2.4 Extremality of Convex Functions . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 29
2.3 Generating One-Dimensional Dual-Feasible Functions.. . . . . . . . . . . . . . 33
2.3.1 Linear Combination . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 33
2.3.2 Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 34
2.3.3 Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
2.3.4 Using the Limiting Behaviour of a Function . . . . . . . . . . . . . . . . . . 36
2.3.5 Rounding Functions . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 37

vii
viii Contents

2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
2.4.1 Applying Symmetry . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
2.4.2 Using Rounding Functions and Applying Symmetry .. . . . . . . . 39
2.4.3 Improving a Function by Using Its Limiting Behaviour .. . . . . 41
2.4.4 A Special Case: A Staircase Function
with Infinitely Many Stairs . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
2.5 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
2.6 Exercises.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 46
3 General Dual-Feasible Functions . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 51
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 51
3.2 Extension of Dual-Feasible Functions to General Domains . . . . . . . . . . 52
3.2.1 Definition.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 52
3.2.2 Maximality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 54
3.2.3 Extremality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 56
3.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 58
3.4 Properties of Maximal General Dual-Feasible Functions .. . . . . . . . . . . . 60
3.4.1 Structure .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 61
3.4.2 Behaviour at Given Points . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 64
3.4.3 Limits of Possible Convexity .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 66
3.4.4 Composition and Convex Combinations . .. . . . . . . . . . . . . . . . . . . . 66
3.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 67
3.6 Building Maximal General Dual-Feasible Functions . . . . . . . . . . . . . . . . . 73
3.6.1 Method I. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
3.6.2 Method II .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 76
3.6.3 Method III .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 77
3.6.4 Examples .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 80
3.7 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 86
3.8 Exercises.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
4 Applications for Cutting and Packing Problems . . . . .. . . . . . . . . . . . . . . . . . . . 91
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 91
4.2 Set-Covering Dual-Feasible Functions.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 91
4.2.1 Data-Dependent Dual-Feasible Functions .. . . . . . . . . . . . . . . . . . . . 93
4.2.2 Data-Independent Dual-Feasible Functions . . . . . . . . . . . . . . . . . . . 94
4.2.3 General Properties . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 94
4.3 Vector Packing Dual-Feasible Functions . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
4.3.1 Basic Definition .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
4.3.2 General Properties of VP-MDFF . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98
4.3.3 General Classes of VP-MDFF . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 102
4.4 Orthogonal Packing .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 110
4.4.1 DFF for the Oriented Case (m-OPP-O-DFF) .. . . . . . . . . . . . . . . . . 111
4.4.2 DFF for the Case with Rotation (m-OPP-R-DFF) . . . . . . . . . . . . 112
4.5 Bin-Packing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 113
Contents ix

4.6 Bin-Packing Problem with Conflicts . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 115

4.6.1 BPC-DDFF Based on a Knapsack Subproblem .. . . . . . . . . . . . . . 115
4.6.2 A BPC-DDFF Based on Graph Decomposition .. . . . . . . . . . . . . . 117
4.7 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 119
4.8 Exercises.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 120
5 Other Applications in General Integer Programming .. . . . . . . . . . . . . . . . . . 125
5.1 Superadditive Functions in Integer Programming .. . . . . . . . . . . . . . . . . . . . 125
5.2 Valid Inequalities for Integer Programs .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 126
5.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 127
5.4 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 130
5.5 Exercises.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 131

Appendix A Hints and Solutions to Selected Exercises.. . . . . . . . . . . . . . . . . . . . 133

References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 157

Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 159
Acronyms

1D-CSP One-dimensional cutting stock problem

BP Bin-packing problem
BP-DDFF Bin-packing data-dependent dual-feasible function
BPC Bin-packing problem with conflicts
BPC-DDFF Bin-packing problem with conflicts data-dependent dual-feasible
function
CS-DFF Cutting stock dual-feasible function
CS-MDFF Maximal cutting stock dual-feasible function
DFF Dual-feasible function
DW Dantzig-Wolfe
EMDFF Extreme maximal dual-feasible function
IP Integer programming
KP-01 Binary knapsack problem
KPC Binary knapsack problem with conflicts
LP Linear programming
MDFF Maximal dual-feasible function
mD-KP m-dimensional knapsack problem
mD-VPP m-Dimensional vector packing problem
m-OPP m-Dimensional orthogonal bin-packing problem
m-OPP-O m-Dimensional orthogonal bin packing problem with fixed orien-
tation
m-OPP-O-DFF m-Dimensional orthogonal bin-packing with fixed orientation
dual-feasible function
m-OPP-R m-Dimensional orthogonal bin-packing problem with rotation
m-OPP-R-DFF m-Dimensional orthogonal bin packing with rotation dual-
feasible function
SC-DFF Set covering dual-feasible function
SC-DDFF Data-dependent set covering dual-feasible function
VP-DFF Vector packing dual-feasible function
VP-MDFF Maximal vector packing dual-feasible function

xi
Chapter 1
Linear and Integer Programming

1.1 Introduction

Integer Programming (IP) is a modelling tool that has been widely applied in the last
decades to obtain solutions for complex real problems, as those that arise in cutting
and packing, location, routing and many other areas. IP models are of the form:

min zIP WD c> x

s: to Ax  b (1.1)
x  o and integer;

where A 2 Rmn ; b 2 Rm ; c 2 Rn , and x 2 ZnC is a vector of decision variables.

The zero vector is denoted by o. The set of solutions of the IP model is XIP D fx 2
ZnC W Ax  bg: If the problem is feasible, the set of solutions is a discrete set, either
finite or countably infinite.
A related problem that plays a central role in the resolution of an IP model
is the Linear Programming (LP) model that results from relaxing the integrality
constraints, which impose that the decision variables x can only take integer values.
The LP model that accepts decision variables x 2 RnC is as follows:

min zLP WD c> x

s: to Ax  b (1.2)
x  o:

The set of solutions of the LP relaxation (or continuous relaxation) of the IP

model is a convex set XLP D fx 2 RnC W Ax  bg: The set XLP contains all the
integer solutions, as XIP D XLP \ ZnC :

© Springer International Publishing Switzerland 2016 1

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5_1
2 1 Linear and Integer Programming

Linear Programming Duality theory is a tool to derive bounds to the values of

the optimal solutions of LP models. For instance, given a minimization problem, a
solution that is feasible to the dual model provides a lower bound to the value of
the minimum. Bounds of good quality are often used in optimization, because they
can improve substantially the performance of algorithms. As we will discuss, the
feasible solutions of the dual problems of the LP relaxations of some models may
provide better lower bounds than others.
Typically there are different ways of modelling an IP problem, and any model is
valid if its set of feasible integer solutions is XIP : However, there may be differences
in their LP relaxations. Some models may provide a closer description of the set
of the feasible integer solutions. Let X1  Rn and X2  Rn be the convex sets of
solutions of the LP relaxations of two valid IP models, meaning that XIP D X1 \
ZnC D X2 \ ZnC : A model with a set X1 is said to be stronger than a model with a
set X2 ; if X1  X2 : Alternatively one may state that the second model is weaker.
The convex hull of a set of points is the smallest convex set that contains all the
points in the set. The convex hull of XIP ; denoted as ConvfXIP g; is the strongest
model for an IP problem. Using LP, the optimal solution would never be a fractional
solution, because its extreme points are all integer. The issue is that it may even
be difficult to know all the constraints that are needed to define the set ConvfXIP g:
One has to resort to other alternatives that are weaker than ConvfXIP g, but may be
stronger than the LP relaxation.
One technique to obtain strong models is the Dantzig-Wolfe (DW) decom-
position. One drawback is that the models that result from DW-decomposition
typically have an exponential number of decision variables, and it is not practical to
enumerate them all. One may resort to column generation algorithms, which have
been successfully used in many IP problems, but they are complex, time consuming
to implement and may still represent a heavy computational burden.
We aim at implementing algorithms that use feasible solutions of the dual
polyhedra of the (strong) column generation models, without solving the column
generation model or enumerating any column. We derive the dual feasible solutions
from the structure of the columns, and so they are valid for any instance of a given
problem.
The motivation is the following. Dual Feasible Functions (DFF) provide feasible
dual solutions of strong models whose corresponding lower bounds are often very
close to the lower bounds provided by column generation models. As DFF can
be computed quickly, the computational burden can be small when compared to
column generation algorithms. Furthermore, for a given problem, it is often possible
to derive not only a single DFF but several DFF, or even families of different DFF,
each providing a different dual feasible solution. By using several DFF, we aim at
obtaining at least a feasible dual solution that provides a good lower bound.
This chapter is structured as follows. In the following section, we present the
DW-decomposition, and illustrate how it can lead to a strong model. Then, we focus
on the structure of the models that result from DW-decomposition, which provides
insight on the space of dual feasible solutions. Finally, two small examples are
presented, showing that dual feasible solutions may provide good quality bounds.
1.2 Dantzig-Wolfe Decomposition 3

1.2 Dantzig-Wolfe Decomposition

1.2.1 Reformulation of the Original Model

DW-decomposition may be applied to LP models whose constraints can be divided

into two sets: the first set includes general constraints Ax  b; while the second set
has constraints with a special structure that define a set denoted by X W

min zLP WD c> x

s: to Ax  b
(1.3)
x2X
x  o:

Minkowski’s theorem states that any point x of a non-empty polyhedron X can be

expressed as a convex combination of the extreme points of X plus a non-negative
combination of the extreme rays of X: Therefore, the set X can be defined as follows:
(
X X X
X D x 2 RnC W x D p x p C r yr ; p D 1; p  0; 8p 2 P;
p2P r2R p2P
)
r  0; 8r 2 R ;

where P D fxp g is the set of extreme points of X and R D fyr g is the set of extreme
rays of X: Replacing the value of x in the original model, and rearranging the terms,
we obtain the following reformulation (or DW-model) of the problem:
X X
min zDW WD .c> xp /p C .c> yr /r
p2P r2R
X X
s: to p .Axp / C r .Ayr /  b
p2P r2R
X
p D 1
p2P

p  0; 8p 2 P
r  0; 8r 2 R:

The decision variables of the reformulated problem are the variables p and r :
The elements c> xp and c> yr define the objective function coefficients, while
the columns Axp and Ayr define the constraints of the reformulated problem.
4 1 Linear and Integer Programming

The constraint on the sum of the values of variables p is usually referred to as

the convexity constraint. This DW-model is equivalent to the original model, as the
definition of x provides the correspondence between any solution of the DW-model
to a solution of the original model.

1.2.2 Dantzig-Wolfe Decomposition in Integer Programming

A motivation for using DW-decomposition in IP is to obtain a model stronger than

the LP relaxation, which is possible when the set X has special characteristics.
DW-decomposition provides a strong model if the polyhedron X does not have the
integrality property, meaning that its extreme points and rays are not all integer. To
obtain a strong model, we use x 2 Convfx 2 X and integerg  X; instead of using
x 2 X: This corresponds to imposing the integrality constraints on the decision
variables x only in the second set of constraints of the original model, and using
integer extreme points and rays.
The set of feasible solutions of the reformulated model, expressed in terms of the
variables x of the original model, is XDWI D fx 2 Convfx 2 X and integerg W Ax 
b; x  og: The corresponding DWI-model is as follows:

min zDWI WD c> x

s: to Ax  b
(1.4)
x 2 Convfx 2 X and integerg
x  o:

Given that ConvfXIP g  XDWI  XLP ; it follows that zIP  zDWI  zLP for
minimization problems. Note that, if the polyhedron X has the integrality property,
then Convfx 2 X and integerg D X; and XDWI D fx 2 X W Ax  b; x  og: In this
case, the DWI-model is as strong as the LP relaxation, and zIP  zDWI D zLP :
Example 1.1 The comparison of the sets XLP and XDWI in Fig. 1.1 illustrates how
DW-decomposition may provide a strong model. The domain of the IP problem is
the finite set of full dots that belong to the space of the LP relaxation, XLP ; shown in
Fig. 1.1a, delimited by the double line.
The set of constraints is decomposed into the first set of general constraints,
which is the single constraint A1 x  b1 ; and the second set, which is the single
constraint A2 x  b2 and the non-negativity constraints x  o; which defines the
set X: The set X does not have the integrality property, because it has a fractional
extreme point in the x1 axis. The set Convfx 2 X and integerg is delimited by the
double line, in Fig. 1.1b, and its extreme points are all integer.
The solution space of the reformulated model, XDWI ; is shown in Fig. 1.1c,
delimited by the double line. It is the set of points x 2 ConvfA2 x  b2 ; x 
o and integerg that also obey the constraint in the first set. t
u
1.2 Dantzig-Wolfe Decomposition 5

(a) x2
A2 x ≤ b2

XLP = {x : A1 x ≤ b1 ,
A2 x ≤ b2 ,
A1 x ≤ b1 x ≥ o}

x1

(b) x2
A2 x ≤ b2

Conv{A2 x ≤ b2 , x ≥ o and integer}

x1

(c) x2
A2 x ≤ b2

XDW I = {x : A1 x ≤ b1 ,
A1 x ≤ b1 x ∈ Conv{A2 x ≤ b2 , x ≥ o and integer}}

x1

Fig. 1.1 Getting a strong model with Dantzig-Wolfe decomposition. (a) The set XLP D
fx W A1 x  b1 ; A2 x  b2 ; x  og. (b) The set ConvA2 x  b2 ; x  o and integerg. (c) The set
XDWI D fx W A1 x  b1 ; x 2 ConvfA2 x  b2 ; x  o and integergg

Example 1.2 The IP model in Fig. 1.1 is maxfx1 C x2 W x1 C 3x2  6; 8x1 C

5x2  20; x1 ; x2  0 and integerg: Note that this is a maximization problem
with  constraints, but the same reformulation principles do apply. Keeping the
single constraint x1 C 3x2  6 in the master problem and letting the set X WD
fx W 8x1 C 5x2  20; x1 ; x2  0g; the matrix A is just the vector .1; 3/
and the vector c D .1; 1/> : The extreme points of Convfx 2 X and integerg
6 1 Linear and Integer Programming

are x1 D .0; 0/>; x2 D .2; 0/> and x3 D .0; 4/> ; and there are no extreme
rays, because X is a bounded set. Therefore, Œc> x1 ; c> x2 ; c> x3  D Œ0; 2; 4 and
ŒAx1 ; Ax2 ; Ax3  D Œ0; 2; 12: The DWI-model resulting from this decomposition
is:

max zDWI WD 01 C 22 C 43 ;

s: to 01 C 22 C 123  6;
1 C 2 C 3 D 1;
1 ; 2 ; 3  0:

The .1 ; 2 ; 3 / coordinates of the extreme points of XDWI are .1; 0; 0/; .0; 1; 0/;
. 12 ; 0; 12 / and .0; 35 ; 25 /; respectively. The last extreme point, which is the optimal
solution, maps to the solution in the original space x D 0 x1 C 35 x2 C 25 x3 D
. 65 ; 85 /> : t
u

1.3 Structure of DW-Decomposition Models

Many integer programming and combinatorial optimization problems were derived

using DW-decomposition, leading to reformulated models with columns that have
a special structure. A column may represent a cutting pattern in a cutting stock
problem, a plan for a machine in a machine scheduling problem or a trip of a
vehicle in a vehicle routing problem, and the feasible solutions of the models are
built combining a subset of columns. We focus on the structure of the columns,
because it determines the structure of the constraints of their corresponding dual
models.
There are many examples of models that were developed without any reference
to DW-decomposition. They were naturally derived, because their feasible solutions
can be expressed as a combination of columns with some special structure.
Nevertheless, we should always keep in mind the DW-decomposition, because the
integrality property helps in recognizing when those “column models” are weak or
strong.
In the following sections, we present the Gilmore and Gomory model for the
Cutting Stock Problem, which is the model underlying the classical dual-feasible
functions addressed in Chap. 2, and the application of DW-decomposition to models
with a block angular structure, which provides insight on how to derive dual-feasible
functions for more general domains. Finally, we briefly describe column generation
algorithms.
1.3 Structure of DW-Decomposition Models 7

1.3.1 Gilmore and Gomory Model for the Cutting Stock

Problem

Given a set of m 2 N different item lengths `i (i 2 f1; : : : ; mg) to be cut from stock
rolls of length L > 0, the 1-dimensional cutting stock problem (1D-CSP) consists
of finding how to cut these items such that the number of used rolls is minimized.
The Gilmore and Gomory model for the Cutting Stock Problem is as follows:
X
min z WD xj (1.5)
j2J

s: to Ax  b; (1.6)
xj 2 N; 8j 2 J; (1.7)

where b 2 Nm describes the order demands of the items, and J is the index set of
all feasible patterns aj , which form the matrix A. The quantity of the j-th pattern to
be used is xj . A pattern aj 2 Nm is feasible if and only if

l> aj  L; (1.8)

i.e., the total sum of the lengths of all the items to be cut in the corresponding
quantity does not exceed the length of the roll L.
The dual of the continuous relaxation of (1.5)–(1.7) is

max zD WD b> u (1.9)

> j
s: to u a  1; 8j 2 J; (1.10)
u 2 Rm
C: (1.11)

Both the primal and the dual problem are solvable, if no item is longer than the
length of the stock rolls. Due to the strong duality theorem, the optimal objective
function values of the continuous relaxation of (1.5)–(1.7) and its dual (1.9)–(1.11)
are the same, and for all feasible solutions of both problems, one has z  zD . Note
that because of the constraints (1.10) and (1.11), it follows that ui  1 for all i.
The solutions of the Gilmore and Gomory model result from a non-negative
combination of columns. It can be shown that this structure results from a DW-
decomposition of an original arc-flow model, whose solutions are extreme rays
that can be associated to solutions of a knapsack problem (see Sects. 1.6 and 1.7).
Therefore, the convexity constraint is not needed.
8 1 Linear and Integer Programming

Example 1.3 Consider a 1D-CSP instance with rolls of length 8 and items of
lengths 4, 3 and 2, with order demands of 5, 4 and 8, respectively. A model is as
follows:

Cutting patterns
LD8 x1 x2 x3 x4 x5 x6 Demand bi
`i D 4 2 1 1 5
3 1 2 1 4
2 2 1 2 4 8
min 1 1 1 1 1 1

Each column describes the number of items of each length produced in a cutting
pattern. A feasible cutting plan is a combination of cutting patterns that satisfies the
demand. The objective is to minimize the number of rolls used.
The structure of the model is appealing. It is a covering model, in which one has
to select a set of columns that cover the demand. If the values of the demands were
all equal to one, clearly the feasible cutting patterns should only have one item of
each type. In this special case, the model would be a set covering model. t
u

1.3.2 Block Angular Structure

Consider a vector of decision variables x partitioned in x D Œx1 ; : : : ; xi ; : : : ; xK ;

with xi 2 RnCi , ni 2 N, i D 1; : : : ; K. There are K sets of constraints that define
independent sets X i , i D 1; 2; : : : ; K, and a set of general linking constraints. The
model has a block angular structure and is as follows:
> > >
min zLP WD c1 x1 C c2 x2 C : : : C cK x K
s: to A1 x 1 C A2 x 2 C : : : C AK x K b
x1 2 X 1
x2 2 X 2
:::
xK 2 X K
i
x 2 RnCi ; ni 2 N; i D 1; 2; : : : ; K ,

where Ai 2 Rmni , ci 2 Rni , with ni 2 N, i D 1; 2; : : : ; K, and b 2 Rm .

In this case, consider that the sets X i are bounded, and so there are no extreme
rays. The set X i can be defined as follows:
X X
X i D fxi 2 RnCi W xi D ip xip ; p D 1; ip  0; 8p 2 Pi g; i D 1; 2; : : : ; K;
p2Pi p2Pi
1.3 Structure of DW-Decomposition Models 9

where Pi D fxip g is the set of extreme points of X i : The reformulation is as follows:

X >
X >
X >
min zDW WD .c1 x1p /1p C .c2 x2p /2p C : : : C .cK xKp /Kp
1 2
X
p2P X
p2P X
p2PK
s: to 1p .A1 x1p / C 2p .A2 x2p / C ::: C Kp .AK xKp /  b
p2P1 p2P2
X p2PK
1p D1
p2P1
X
2p D1
p2P2
::: X
Kp D1
p2PK
ip  0; 8p 2 Pi ; i D 1; 2; : : : ; K:

In this reformulated model, a column may still represent a cutting pattern

or a plan for a machine. Nevertheless, this more general structure allows us to
model the cases in which the cutting patterns come from rolls of different sizes,
as happens in the multiple roll lengths cutting stock problem, or the machines
have different capabilities or availabilities, as in the parallel non-identical machine
scheduling problem. Even though the columns are generated from different entities,
their contributions are used to build feasible solutions that must obey the linking
constraints.

1.3.3 Parallel Non-identical Machine Scheduling

Consider the problem of building a feasible schedule for a set of parallel non-
identical machines. A plan for a machine is feasible if the jobs assigned to the
machine can be executed within a given time slot. There is a set of feasible plans, one
for each different machine. The plan for each machine is one selected from several
plans. Each plan for a given machine represents an extreme point of the solution set
of the machine. Therefore, a convexity constraint is needed to select just one plan
for each machine.
The model that results from DW-decomposition is a set partitioning model (the
derivation is left as an exercise). The columns represent the feasible plans for the
machines. The model has a set of constraints for the jobs, which indicate that each
job is executed once, and a set of constraints for the machines, which are convexity
constraints.
10 1 Linear and Integer Programming

Example 1.4 Consider an example with four jobs and two machines. Let yik be a
decision variable that represents a feasible plan, indexed by k; that assigns a set of
jobs to machine i: Each feasible plan for a machine has a cost, and the objective is
to minimize the total cost.

y10 y11 y12 y13 y14 y15 y20 y21 y22 y23 y24 y25
Job 1 1 1 1 1 1 1 D1
2 1 1 1 1 1 D1
3 1 1 1 1 D1
4 1 1 1 1 1 D1
Machine 1 1 1 1 1 1 1 D1
2 1 1 1 1 1 1 D1
min 0 27 25 24 22 21 0 12 16 14 10 14

Note that each machine has a schedule that is the null solution, meaning that
the machine is idle in the plan. Clearly, the machine constraints might be replaced
by  constraints, because the idle machine columns are slack variables for those
constraints. t
u

1.3.4 Solution of DW-Models with Column Generation

The DWI-model may be stronger, but it comes at a price. The first issue is that
there is generally an exponential number of extreme points and extreme rays in the
set Convfx 2 X and integerg; and the second is that to find them it is necessary to
solve an integer optimization problem. To overcome the first issue, DWI-models are
solved in practice using column generation algorithms, which only pick attractive
variables.
The solution of DWI-models with column generation algorithms is not a central
topic in this monograph. Nevertheless, it is described succinctly as follows. The
DWI-model has a master problem, defined by the first set of constraints in the DW-
decomposition. The column generation algorithm starts with a restricted master
problem, a model that has a restricted set of variables, which is optimized. The
dual information from the restricted master problem is used in one or several
subproblems to find the most attractive column to be inserted in the restricted master
problem, which is then re-optimized. The iterative algorithm is repeated until no
more attractive columns are found, yielding a solution that is provably optimal.
In each iteration, the optimal solution of the subproblem, which is stated in terms
of the original variables x; is an extreme point or an extreme ray of the set Convfx 2
X and integerg: Therefore, the subproblem is an integer optimization problem. This
second issue may not be too hard to overcome. For instance, in the Cutting Stock
1.4 Duality and Bounds from Dual Feasible Solutions 11

Problem, it amounts to finding integer solutions of a subproblem that is a knapsack

problem.
Example 1.5 For the solution of the Gilmore and Gomory model (1.5)–(1.7), the
restricted master problem is initialized with a set of columns (for instance, each
column is a cutting pattern with multiple copies of the same item in quantities
b `Li c; 8i/: Let J .J  J/ denote the set of cutting patterns in the restricted master
problem, and u D .u1 ; u2 ; : : : ; um / 2 Rm C be the corresponding optimal dual
solution.
The subproblem is then used to find the most attractive column that does not
belong to the restricted master problem (in J n J/: The structure of a feasible column
is well defined: aj 2 Nm and it obeys (1.8). It will be attractive if its reduced cost,
cN j D 1  u> aj ; is negative. The most attractive cutting pattern is the column amin
with the most negative reduced cost cN min D 1  u> amin D minj2JnJ .1  u> aj /:
As all the columns in J have non-negative reduced costs, we may search over the
entire set J: Furthermore, we may use the opposite function, to find the column j
such that maxj2J .u> aj  1/:
Therefore, given the dual optimal solution u of a restricted master problem,
the most attractive feasible column is provided by the solution of the following
maximization knapsack subproblem:

max z.u/ WD u> aj  1 (1.12)

s: to l> aj  L (1.13)
a 2N :
j m
(1.14)

If the optimal solution of the knapsack subproblem is positive (corresponding to a

negative reduced cost cutting pattern), the feasible column is added to the restricted
master problem, which is re-optimized; otherwise, the current solution solves the
linear relaxation of the Gilmore and Gomory model, because there are no more
attractive columns. t
u

1.4 Duality and Bounds from Dual Feasible Solutions

Consider the following pair of primal and dual problems:

min z WD c> x max zD WD b> u

(Primal) s: to Ax  b (Dual) s: to A> u  c
x  o; u  o;
where u 2 Rm is the vector of dual decision variables.
12 1 Linear and Integer Programming

The Weak Duality theorem states that, given a primal feasible solution xO and
a dual feasible solution u;O then b> uO  zD  z  c> xO : It follows that any feasible
solution uO to the dual problem provides a lower bound, b> u;
O to the optimum solution
of the primal problem, z. On the other hand, the Strong Duality theorem states that
if the optima of the two problems are finite, then zD D z:
Let us analyze what happens when we consider two primal models, one weak
and one strong. For instance, recall that the primal minimization DWI-model that is
stronger than the LP relaxation model yields a better optimum solution, i.e., zLP 
zDWI  zIP . According to the Strong Duality theorem, the optimal value of the dual
maximization problem of the DWI-model will also be greater than or equal to the
optimal value of the LP relaxation model. Therefore, one may expect to find feasible
solutions of the dual maximization problem of the DWI-model that have an objective
function value that is greater than or equal to zLP .
Recall that we aim at finding feasible solutions of the dual models of strong
DWI-models without enumerating any of the (exponentially many) columns that
correspond to all the extreme solutions of the sets in the subproblem(s). Instead,
by analyzing the structure of the dual of the DWI-models, we aim at deriving
functions that provide dual feasible solutions that obey all the (exponentially many)
constraints.

1.5 Examples

Two examples are presented below, illustrating that feasible solutions of the dual of
a DWI-model can provide lower bounds that are better than trivial lower bounds.
In Sect. 2.1, for the cutting stock problem, and in Sect. 4.3.3, for the vector packing
problem, we will see how these dual feasible solutions are derived from DFF suited
to each problem.

1.5.1 One-Dimensional Cutting Stock Problem

A trivial lower bound for the 1D-CSP results from calculating the minimum number
of rolls of size L that are needed to place the sum of the sizes of all items:

X
m
LBT D d bi `i =Le:
iD1

The dual polytope of the Gilmore and Gomory model may have dual feasible
solutions that provide better lower bounds. Recall that the dual polytope has all the
(exponentially many) constraints that correspond to all feasible cutting patterns. As
we will see, only maximal patterns (in which there is no room for any other item)
are needed to define the dual polytope.
1.5 Examples 13

Example 1.6 A company has to deliver ten items with weight 0.4 and 40 items
with weight 0.3. Each vehicle can carry a weight of 1. The lower bound LBT D
d.10  0:4 C 40  0:3/=1e D 16, meaning that the sum of all weights is 16, and so,
at least, 16 vehicles are needed. However, the optimal solution requires 17 vehicles.
One optimal solution is having ten vehicles carrying items of sizes 0.4, 0.3 and 0.3,
six vehicles carrying three items of size 0.3, and one vehicle carrying two items of
size 0.3.
The set of all maximal feasible cutting patterns is JM D f.a1 ; a2 / 2 N2 W 0:4 a1 C
0:3 a2  1g D f.2; 0/; .1; 2/; .0; 3/g  J: The pair of primal-dual problems is as
follows:

min z WD 1x1 C 1x2 C 1x3 max zD WD 10u1 C 40u2

s: to 2x1 C 1x2  10 s: to 2u1 1
(Primal) 2x2 C 3x3  40 (Dual) 1u1 C 2u2  1
x1 ; x2 ; x3  0; 3u2  1
u1 ; u2  0:
The space of the dual problem is depicted in Fig. 1.2. Note that cutting patterns
that are not maximal would lead to redundant constraints.
The extreme points of the dual polytope are O.0; 0/, A.0; 13 /, B. 31 ; 13 /, C. 12 ; 14 /,
D. 12 ; 0/, respectively. Besides the extreme points, we consider the dual feasible
solution T.0:4; 0:3/: The corresponding lower bounds are as follows:

LBO D d.10  0 C 40  0/=1e D 0 bins,

LBA D d.10  0 C 40  13 /=1e D 14 bins,
LBB D d.10  13 C 40  13 /=1e D 17 bins,
LBC D d.10  12 C 40  14 /=1e D 15 bins,
LBD D d.10  12 C 40  0/=1e D 5 bins,
LBT D d.10  0:4 C 40  0:3/=1e D 16 bins. t
u

Fig. 1.2 Space of dual of u2

Cutting Stock Problem
Example

1
2
A B
1
3 TC

O D
1 1 u1
2
14 1 Linear and Integer Programming

PuO i D `i =L; i D 1; : : : ; m; is always feasible. It provides

Note that the dual solution
the lower bound LBT D d m O i e: In the example, it corresponds to the point
iD1 bi u
T.0:4; 0:3/, which is not an extreme point of the dual polytope. There may even be
cases, when all dual constraints have slack, where it is an interior point of the dual
space. However, in other cases, the lower bound LBT may be equal to the optimal
solution of the 1D-CSP.

1.5.2 Vector Packing Problem

In the m-dimensional vector packing problem (mD-VPP), with m 2 Nnf0; 1g; items
with m independent dimensions (for instance, volume and weight in 2-dimensional
problems) have to be packed into a minimum number of larger objects, which are
m-dimensional bins. There are m capacity constraints, one for each dimension of
the problem, i.e., the sum of the lengths of all packed items must not exceed the
bin size in any of the m directions. The m-dimensional bins are all equal and have
lengths Ld , d D 1; : : : ; m, and there are n 2 N different items with lengths `id
(i D 1; : : : ; n; d D 1; : : : ; m/:
A trivial lower bound for the m-dimensional Vector Packing Problem, which
amounts to applying the trivial lower bound for the 1D-CSP to all the dimensions
and then taking the best value, is:
(& ')
X
n
LVPP D max `id =Ld : (1.15)
dD1;:::;m
iD1

A DWI-model for the mD-VPP is similar to the Gilmore and Gomory model for
the 1D-CSP (1.5)–(1.7), but the packing aj 2 Nn is feasible if and only if

X
n
j
ai  `id  Ld ; d D 1; : : : ; m; (1.16)
iD1

i.e., for each dimension, the sum of the lengths of the items packed in a bin does not
exceed the length of the bin. Again, the dual polytope has dual feasible solutions
that provide better lower bounds.
Example 1.7 A company has vehicles that can carry a volume of 4 and a weight of
5, and has to deliver four items with the following volumes and weights:

Item 1 2 3 4 Vehicle
Volume 2 3 1 2 4
Weight 3 2 4 1 5
1.5 Examples 15

Fig. 1.3 One optimal weight weight weight

solution of the 2-dimensional
VPP

volume volume volume

The trivial lower bound is LVPP D maxfd.2 C 3 C 1 C 2/=4e; d.3 C 2 C 4 C

1/=5eg D 2I at least two vehicles are needed. However, the optimal solution of the
2-dimensional VPP instance requires three vehicles. One optimal solution is shown
in Fig. 1.3. Vehicle 1 carries items 1 and 4, vehicle 2 carries item 2 and vehicle 3
carries item 3.
Let xj be a decision variable that represents a feasible packing that assigns a set
of items to a vehicle. The maximal packings for the 2D-VPP instance are:

packings
.v; w/ item x1 x2 x3
.2; 3/ 1 1
.3; 2/ 2 1
.1; 4/ 3 1
.2; 1/ 4 1 1
.4; 5/

and the corresponding pair of primal-dual problems is as follows:

min z WD x1 C x2 C x3 max zD WD u1 C u2 C u3 C u4
s: to x1 1 s: to u1 C u4  1
x2 1 u2 1
(Primal) (Dual)
x3  1 u3 C u4  1
x1 C x3  1 u1 ; u2 ; u3 ; u4  0:
x1 ; x2 ; x3  0;
Some feasible solutions of the dual space are uO 1 D . 12 ; 1; 0; 12 /> ; uO 2 D
.1; 1; 0; 0/>; uO 3 D .1; 1; 1; 0/> and uO 4 D .1; 0; 1; 0/>. The dual solution uO 3 is the
one that provides the best lower bound, equal to 3, for this 2D-VPP instance. t
u
16 1 Linear and Integer Programming

1.6 Related Literature

For a revision of integer and linear programming, as well as DW-decomposition,

see the textbooks by Nemhauser and Wolsey (1998) or Bazaraa et al. (2010). DW-
decomposition was first proposed by Dantzig and Wolfe (1960).
Geoffrion (1974) shows how to exploit problem structure in integer program-
ming, and develops the general theory that relates to the bounds from the LP
relaxation and from the models with subproblems without the integrality property.
The theory is developed in the context of Lagrangian relaxation, which is closely
related to DW-decomposition, as shown by Nemhauser and Wolsey (1998).
Gilmore and Gomory (1961) introduce the solution of the LP relaxation of the
cutting stock problem using column generation, by solving an auxiliary integer
knapsack problem. Valério de Carvalho (1999) introduces the arc-flow model for
the cutting stock problem, and shows how the Gilmore and Gomory model can be
derived as a non-negative combination of extreme rays, which are feasible solutions
of the arc-flow model, and a null extreme point. Vance (1998) showed that the
Gilmore and Gomory model can also be derived from a DW-decomposition of a
model with a block angular structure. The derivation is left as an exercise.
Lueker (1983) uses the dual of the Gilmore and Gomory model to derive lower
bounds for the bin-packing problem. In fact, the cutting stock problem and the
bin-packing problem have structures that are very similar. The lower bound LBT
is due to Martello and Toth (1990) and the lower bound for the vector packing
problem, LBVPP , is due to Spieksma (1994).

1.7 Exercises

1. Consider the problem presented in Example 1.2, p. 5.

(a) Apply a similar DW-decomposition, but considering the set X instead of the
set Convfx W x 2 Xand integerg. Call ZDW the value of the optimal solution
of the DW-model.
(b) Find the values of zIP , zDWI , zDW and zLP , and the relationship between them.
2. Consider the following LP problem with a block angular structure:
max z WD 3 x11 + 5 x12 + 1 x21 + 2 x22
s: to 1 x11 + 2 x12 + 2 x21 + 1 x22  6
3 x11 + 2 x12 + 1 x21 + 1 x22  8
1 x11 + 2 x12  4
1 x11  2
3 x21 + 1 x22  3
x11 ; x12 ; x21 ; x22  0.
1.7 Exercises 17

(a) Apply a DW-decomposition considering the following two bounded sets X 1

and X 2 in different subproblems:

X 1 D f.x11 ; x12 / 2 R2 W x11 C 2x12  4; x11  2; x11 ; x12  0g

X 2 D f.x21 ; x22 / 2 R2 W 3x21 C x22  3; x21 ; x22  0g;

(b) Draw the solution space of the two sets X 1 and X 2 :

(c) Using an LP software package, solve the DW-model. Determine the optimal
values of variables of the original LP model.
3. The parallel non-identical machine scheduling example, presented in Sect. 1.3.3,
is an application of the generalized assignment problem (GAP). There is a set of
jobs J; indexed by j; and a set of machines I; indexed by i; and we want to minimize
the total processing cost of assigning jobs to machines, which is the sum of the costs
of assigning each job to a machine, denoted by cij : The processing time of job j in
machine i is denoted as pij ; the time available in machine i is Pi :
The original model has binary decision variables xij ; defined as follows:

1; if job j is processed in machine i
xij D
0; otherwise,

and the IP model is as follows:

XX
min zIP WD cij xij (1.17)
i2I j2J
X
s: to pij xij  Pi ; 8i 2 I (1.18)
j2J
X
xij D 1; 8j 2 J (1.19)
i2I

xij 2 f0; 1g; 8i 2 I; j 2 J: (1.20)

There are two different DW-decompositions for the LP relaxation of this IP

depending on the set of constraints that are placed in the subproblem, either
constraints (1.18) or constraints (1.19).
(a) For each decomposition, describe the resulting master problem and subprob-
lem.
(b) Discuss the quality of the bounds obtained from both DW-models, and
compare them with the bound from the LP relaxation of the IP model above.
(c) Build both DW-models considering the data given in the following tables
for a problem instance with two machines and four jobs. The times available
in the machines are 10 and 12, respectively. The values of the processing
18 1 Linear and Integer Programming

costs C D Œcij  and the processing times P D Œ pij  are as follows:

1 2 3 4 1 2 3 4
C D 1 15 12 10 9 PD 1 3 4 5 6
2 8 4 8 6 2 4 5 8 4
(d) Let u D .u1 ; u2 ; u3 ; u4 / and v D .v1 ; v2 / be the vectors of dual variables
associated with the job and machine constraints, respectively. Check that
.u; v/ D .15; 11; 15; 13; 5; 14/ is a dual feasible solution for the strong
DW-model. Which is the value of the corresponding lower bound?
(e) Check that the primal solution y12 D y24 D 1 and all the remaining yik D 0
in Example 1.4 (p. 10) is feasible and optimal. Check also that this primal
solution and the dual feasible solution obey the Complementary Slackness
conditions.
4. Given bins of integer capacity L and a set of different item sizes `1 ; : : : ; `m ;
a feasible cutting pattern in a single roll can be modelled as a path in an acyclic
directed graph with L C 1 vertices. Consider a graph G D .V; A/ with V D
f0; 1; 2; : : : ; Lg and A D f.i; j/ W 0  i < j  L and 9d 2 f1; : : : ; mg with j  i D
`d g; meaning that there is an arc between two vertices if there is an item of the
corresponding size. There are additional arcs .k; k C 1/, k D 0; 1; : : : ; L  1,
corresponding to loss. A packing in a single bin is a path between vertices 0 and L:
The 1D-CSP is modelled as the problem of finding the minimum flow between
vertex 0 and vertex L with additional constraints enforcing that the sum of the flows
in the arcs of each order d must be greater than or equal to the corresponding
demand bd , d D 1; : : : ; m. The decision variables xij ; associated with the arcs
defined above, indicate the number of items of size j  i placed in any roll at the
distance of i units from the beginning of the roll. The variable xL0 can be seen as a
feedback arc, from vertex L to vertex 0, and the model is as follows:

min z WD xL0 (1.21)

8
X X < xL0 ; if j D 0
s: to xij  xjk D 0; if j D 1; : : : ; L  1 (1.22)
:
.i;j/2A . j;k/2A xL0 ; if j D L
X
xk;kC`d  bd ; d D 1; 2; : : : ; m (1.23)
.k;kC`d /2A

xij  0; 8.i; j/ 2 A (1.24)

xij integer; 8.i; j/ 2 A: (1.25)

(a) Apply a DW-decomposition to (1.21)–(1.24), keeping (1.22) and (1.24) in

the subproblem and (1.23) in the master problem.
(b) Which are the extreme rays of the subproblem? Is there any extreme point?
(c) Consider a 1D-CSP instance with rolls of length 8 and items of lengths 4,
3 and 2, with order demands of 5, 4 and 8, respectively. The graph that
1.7 Exercises 19

represents the subproblem is shown in the figure (some arcs can be discarded
and are not represented). Build the corresponding Gilmore and Gomory
model.

0 1 2 3 4 5 6 7 8

5. Given rolls of the same size L, each treated separately and indexed by k, k 2 K,
where K is a set of rolls that are sufficient to pack all the items, and clients with
demands of di items of sizes `i , 0 < `i  L, i 2 I, the cutting stock problem can be
modelled using integer decision variables xik , which represent the number of times
item i is cut from roll k, and binary decision variables yk , with yk D 1, if roll k is
used, and 0, otherwise. The model is as follows:
X
min z WD yk (1.26)
k2K
X
s: to xik  di ; 8i 2 I; (1.27)
k2K
X
`i xik  Lyk ; 8k 2 K; (1.28)
i2I

yk 2 f0; 1g; 8k 2 K; (1.29)

xik  0 and integer; 8i 2 I; 8k 2 K: (1.30)

The objective is to cut the minimum number of rolls to satisfy demand, and the
first set of constraints enforces that the demand is satisfied, while the second imposes
that the sum of the lengths of the items placed in one roll cannot exceed a function
that takes the value of the length of the roll when the roll is used, and the value 0,
otherwise.
(a) Apply a DW-decomposition to this model with a block angular structure,
treating each roll as a separate entity, keeping (1.27) in the master problem,
and each knapsack constraint of the set (1.28), together with (1.29)–(1.30),
as a separate subproblem.
(b) Which is the meaning of the convexity constraint for each roll k in the
reformulated model?
(c) Noting that all the rolls have equal size and their cutting stock patterns
are identical, simplify the resulting DW-model, dropping the convexity
constraints, to obtain the Gilmore and Gomory model.
Chapter 2
Classical Dual-Feasible Functions

2.1 Introduction

Dual-feasible functions (DFF) have been used to improve the resolution of different
combinatorial optimization problems with knapsack inequalities, including cutting
and packing, scheduling and network routing problems. They were used mainly to
compute algorithmic lower bounds, but also to generate valid inequalities for integer
programs. During a long time, the literature concerning these two applications
of DFF was somehow disconnected. Functions defined for lower bounding were
often referred to as dual-feasible, whereas the functions used to strengthen integer
programming models were referred to as superadditive and nondecreasing. The
relationship between these two families of functions is that the latter is a dominant
family of DFF. Other designations are also used as for instance “redundant function”
in the context of scheduling problems. These functions are in fact discrete DFF.
A dual-feasible function is defined formally as follows.
Definition 2.1 A function f W Œ0; 1 ! Œ0; 1 is a dual-feasible function, if for any
finite index set I of nonnegative real numbers xi 2 RC , i 2 I, it holds that
X X
xi  1 H) f .xi /  1:
i2I i2I

This implies immediately f .0/ D 0 and f .x/  1=b1=xc for all x 2 .0; 1.
Example 2.1 Figure 2.1 shows a parameter dependent staircase function f W Œ0; 1
! Œ0; 1, defined as

f .x/ WD bCxc=bCc; (2.1)

© Springer International Publishing Switzerland 2016 21

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5_2
22 2 Classical Dual-Feasible Functions

y y y y
1 1 1 1

1 x 1 x 1 x 1 x
˚ 12 12 36 60

Fig. 2.1 Dual-feasible function (2.1) for parameter values C 2 11
; 7
; 7 ; 11

for several parameter values C  1. Black filled circles mean that point belongs
to the graph of the function, while white circles exclude that point. For instance,
f .bCc=C/ D 1, but f .x/ < 1 for all x 2 Œ0; bCc=C/. For the sake of conciseness,
parameters are omitted wherever this is appropriate.
The function f is dual-feasible for any real parameter C  1, as one
P can see as
follows. One gets for any finite set I and numbers xi 2 Œ0; 1 with xi  1 the
i2I
estimation
X X X
s WD bCc  f .xi / D bCxi c  Cxi  C:
i2I i2I i2I

Since bCc  f .xi / 2 N, the sum is also integer, hence s  bCc.

Dual-feasible functions are generally defined in Œ0; 1. However, using discrete
values instead may lead to simpler formulations namely when the original input
data is integer. This alternative formulation yields a so-called discrete dual-feasible
function whose definition is given next.
Definition 2.2 A discrete dual-feasible function f W f0; 1; : : : ; dg ! f0; 1; : : : ; d0 g
with d; d0 2 N n f0g is such that
X X
xi  d H) f .xi /  f .d/ D d0 ;
i2I i2I

for any finite index set I of nonnegative integer numbers xi 2 N, i 2 I.

For every discrete DFF, there is an equivalent DFF defined in Œ0; 1, and vice versa,
as illustrated in the following example. Note, this does not mean a bijection.
Example 2.2 Let g W Œ0; 1 ! Œ0; 1 be a DFF. Setting

f .x/ WD d0  g.x=d/;

for x 2 f0; 1; : : : ; dg yields a discrete DFF. Let for example

g.x/ WD dmaxf0; kx  1ge =.k  1/: (2.2)

2.1 Introduction 23

Fig. 2.2 The function g for

kD6
y
1 g

1 x

This function is a DFF for any k 2 N with k > 1. Hence, d0 WD k  1 yields the
discrete DFF
  
d0 C 1
f .x/ WD max 0; x1 :
d

An illustration of the function g for k D 6 is provided in Fig. 2.2. t

u
Let g W f0; 1; : : : ; dg ! f0; 1; : : : ; d g be a discrete dual-feasible function with
0

d; d0 2 N n f0g and g.d/ D d 0 . A dual-feasible function f W Œ0; 1 ! Œ0; 1 can be

constructed as follows:

f .x/ WD g.bdxc/=d0 : (2.3)

The term dual-feasible comes from the alternative definition of these functions,
which relies on the dual formulation of the well-known Gilmore and Gomory model
for the cutting stock problem. In this context, a function f is said to be dual-feasible
if it maps each x (the size of an item) to its corresponding value in a valid dual
solution of the cutting stock problem. If f is a DFF, then assigning ui WD f .`i =L/
for all i 2 f1; : : : ; mg yields a feasible solution for the dual problem (1.9)–
(1.11), because (1.8) and the definition of a DFF ensure the validity of (1.10)
and (1.11).
Example 2.3 Recall Example 1.6, p. 12, with items of sizes 0.4 and 0.3. Using the
DFF g defined in (2.2), and shown in Fig. 2.3 for different values of the parameter k;
the sizes of the items are mapped into the values indicated in the following
table:

x g2 .x/ g3 .x/ g4 .x/

1 1
0:4 0 2 3
1
0:3 0 0 3
24 2 Classical Dual-Feasible Functions

y y y
1 g2 1 g3 1 g4

1 x 1 x 1 x

Fig. 2.3 DFF gk .x/ WD dmaxf0; kx  1ge =.k  1/; k D 2; 3; 4

Note that the DFF g2 ; g3 and g4 provide the dual feasible solutions
   
1 1 1
O D .0; 0/; D D ;0 and B D ; ;
2 3 3

respectively, shown in Example 1.6.

Each DFF in the family provides a lower bound, whose computation is not
expensive at all. For this instance, the general expression for the values of the lower
bounds is

LBgk WD d.10  gk .0:4/ C 40  gk .0:3//=1e;

for k > 1: t
u
Hence, dual-feasible functions lead to valid lower bounds for the cutting stock
problem, and there is always one such function whose corresponding bound is as
strong as the continuous bound achieved with the Gilmore and Gomory model (1.5)–
P
m
(1.7) presented in Sect. 1.3.1, because there is a DFF f , such that b> u D bi 
iD1
f .`i =L/ equals the optimal objective function value of the continuous relaxation
of (1.5)–(1.7). To see this, assume without loss of generality that `i ¤ `j for all
i ¤ j, and let uO be an optimal solution of (1.9)–(1.11). A suitable DFF f is obtained
by setting f .x/ WD uO i for x D `i =L (i D 1; : : : ; m) and f .x/ WD 0 for all the remaining
points, i.e., f .x/ WD 0 for all x 2 Œ0; 1 n f`i =Lji 2 f1; : : : ; mgg.

2.2 Properties

2.2.1 Maximality

Despite the large number of dual-feasible functions that may be defined, only those
that are non-dominated are interesting since they yield the best lower bounds and
2.2 Properties 25

strongest valid inequalities. To be non-dominated, a dual-feasible function must be

superadditive and nondecreasing.
Definition 2.3 A function f is superadditive if for all x; y with x; y; x C y belonging
to the domain of f , it holds that f .x C y/  f .x/ C f .y/.
Proposition 2.1 If a function f W Œ0; 1 ! Œ0; 1 is superadditive, then f is a DFF.
Proof The range and the superadditivity imply monotonicity and
X X
f .xi /  f . xi /  f .1/  1
i2I i2I

P
for any finite index set I of nonnegative real numbers xi with xi  1. t
u
i2I

Note that a superadditive function may be dominated by another superadditive

function. A non-dominated dual-feasible function is said to be maximal as defined
next.
Definition 2.4 A DFF f W Œ0; 1 ! Œ0; 1 is a maximal dual-feasible function
(MDFF), if there is no other DFF g W Œ0; 1 ! Œ0; 1 with g.x/  f .x/ for all
x 2 Œ0; 1. Similarly, a discrete DFF h W f0; 1; : : : ; dg ! f0; 1; : : : ; d0 g is maximal,
if there is no other discrete DFF k W f0; 1; : : : ; dg ! f0; 1; : : : ; d0 g with k.x/  h.x/
for all x 2 f0; 1; : : : ; dg.
Example 2.4 The following function fBJ;1 .xI C/ is a MDFF for all parameter values
C  1. It is defined as
  
frac.Cx/  frac.C/
fBJ;1 .xI C/ WD bCxc C max 0; =bCc; (2.4)
1  frac.C/

where frac./ denotes the non-integer part of its argument, i.e. frac.C/ C  bCc.
Function fBJ;1 is continuous and piecewise linear for all C, as illustrated in Fig. 2.4.
If C 2 N, then fBJ;1 becomes the identity function. Otherwise, there are dCe many
intervals without a slope, and the slope in the other intervals increases with frac.C/.
Note that, without the max-expression, the non-maximal staircase DFF f W
Œ0; 1 ! Œ0; 1, defined in formula (2.1), would be obtained. In the open intervals

y y y y
1 1 1 1

1 x 1 x 1 x 1 x
˚ 12 12 36 60

Fig. 2.4 Maximal dual-feasible function fBJ;1 for parameter values C 2 11
; 7
; 7 ; 11
26 2 Classical Dual-Feasible Functions

where fBJ;1 is strictly monotone, one has f .x/ < fBJ;1 .x/. Only outside these intervals
f .x/ D fBJ;1 .x/ holds. t
u
Several properties characterize the MDFF. They have to be nondecreasing,
superadditive, and also symmetric, as stated formally in the following theorem.
Theorem 2.1 A function f W Œ0; 1 ! Œ0; 1 is a MDFF if and only if the following
conditions hold:
1. f is monotonely increasing, i.e. f .x/  f .y/ if x  y;
2. f is superadditive;
3. f is symmetric in the sense f .x/ C f .1  x/ D 1, 8x 2 Œ0; 1;
4. f .0/ D 0.
Some of these conditions imply others. Therefore, to prove that a given function is
a MDFF, much weaker prerequisites are in fact sufficient.
Theorem 2.2 A function f W Œ0; 1 ! RC fulfilling the following conditions is a
MDFF:

f .0/ D 0I (2.5)
f .x/ C f .1  x/ D 1; 8x 2 Œ0; 1=2I (2.6)
f .x1 C x2 /  f .x1 / C f .x2 /; 8x1 ; x2 with
0 < x1  x2 < 1=2 and x1 C x2  2=3: (2.7)

Note that to prove that a real function f is a MDFF using this theorem, it is not
necessary to prove that f .x/  1 for all x 2 Œ0; 1. Indeed, this follows from f .x/  0
and the symmetry condition (2.6). Therefore, the range of f is only required to be
part of RC .
When verifying whether a given function f is a MDFF, the main difficulty
in the application of Theorems 2.1 and 2.2 is usually related to the test of its
superadditivity. An approach for this test consists in resorting to the following
function g W .0; 1=2/2 ! R:

g.x1 ; x2 / WD f .x1 C x2 /  f .x1 /  f .x2 /: (2.8)

The function f obeys the superadditivity condition (2.7) if and only if g.x1 ; x2 /  0
for all x1 ; x2 according to (2.7). To check this, the extreme points of g can be sought.
If g is differentiable, then only the critical points, i.e. those with rg.x1 ; x2 / D o,
may be extreme points. If f is differentiable only in some smaller intervals inside
.0; 1/, then analyzing the function (2.8) requires also to check separately points
where f 0 .x1 /, f 0 .x2 / or f 0 .x1 C x2 / does not exist.
Lemma 2.1 If the function f W Œ0; 1 ! Œ0; 1 fulfils the symmetry

condition (2.6),
and if it is differentiable in the interval (0,1), then the point 13 ; 13 is a critical point
for g.x1 ; x2 / in (2.8).
2.2 Properties 27

According to Lemma 2.1, the superadditivity of f should always be checked for

x1 D x2 D 1=3 first. If f is symmetric, then 2  f .1=3/  f .2=3/ holds if and only
if f .1=3/  1=3. A drawback of this approach is that the function g.x1 ; x2 / in (2.8)
may be complicated and have infinitely many critical points. Since every critical
point is potentially an extreme point, checking the superadditivity of the function
f by testing the nonnegativity of g at its extreme points relies on exploring all the
critical points of g.

2.2.2 Maximality of Convex Functions

The particular case of convex functions is addressed next.

Definition 2.5 Let D ¤ ; be a convex set. A function f W D ! R is convex, if for
all x1 ; x2 2 D and  2 .0; 1/, it holds that

f .x1 C .1  /x2 /    f .x1 / C .1  /  f .x2 /:

The superadditivity of these functions (not necessarily bounded to domain and range
Œ0; 1) is established through the following lemma.
Lemma 2.2 Let b > 0 be a constant. If a function f W Œ0; b ! R is convex on Œ0; b
and if f .0/  0, then f is superadditive.
As a corollary, we obtain the following result that allows for the simple identification
of many MDFF.
Lemma 2.3 If f W Œ0; 1 ! Œ0; 1 with f .0/ D 0 fulfills the symmetry condition (2.6),
and if it is convex on Œ0; 1=2, then f is a MDFF.
Using Lemma 2.3, it is easy to see for C  2 that the function fBJ;1 .xI C/ in (2.4) is
a MDFF.

2.2.3 Extremality

To get quickly the strongest bounds and inequalities from dual-feasible functions,
maximality is not enough. Consider for example the case of the 1D-CSP as defined
in Sect. 1.3.1. If we are given three MDFF f ; g; h W Œ0; 1 ! Œ0; 1, such that 2f .x/ D
g.x/ C h.x/ for all x 2 Œ0; 1 (with xi D li =L, i D 1; : : : ; m), then the lower bound
obtained with these dual feasible functions equals

X
m X
m 
bi  g.xi / C bi  h.xi / =2;
iD1 iD1
28 2 Classical Dual-Feasible Functions

and hence
( )
X
m X
m X
m
bi  f .xi /  max bi  g.xi /; bi  h.xi / :
iD1 iD1 iD1

In that case, either g or h leads to a bound, which is not worse than the one
obtained by f . Therefore, in order to avoid unnecessary calculations, the use of f is
superfluous, and it is in practice important, apart from maximality, to know whether
a dual-feasible function leads to solutions, or not, which are always dominated or
achieved by another DFF.
Given a convex set S ¤ ;, a point e 2 S is an extreme point if 2e D s1 C s2 with
s1 ; s2 2 S implies e D s1 D s2 . A similar definition can be adopted for dual-feasible
functions.
Definition 2.6 A MDFF f is an extreme maximal dual-feasible function (EMDFF),
if for any MDFF g; h with

2  f .x/ D g.x/ C h.x/; 8x 2 Œ0; 1;

it follows that f g.
For any non-extreme MDFF f , and any finite set I and xi 2 Œ0; 1, with i 2 I, there
is another MDFF g such that
X X
g.xi /  f .xi /:
i2I i2I

If f is a non-extreme MDFF, then there are other MDFF g; h with

!
X 1 X X
f .xi / D  g.xi / C h.xi / :
i2I
2 i2I i2I

P
and either both summands equal f .xi / or one is larger. Clearly, it makes no sense
i2I
to analyze the case of non-maximal DFF, since they are dominated by at least one
MDFF by definition.
Example 2.5 Consider the following MDFF fCCM;1 .xI C/ illustrated in Fig. 2.5:
8
< bCxc=bCc; if 0  x < 1=2;
fCCM;1 .xI C/ WD 1=2; if x D 1=2; (2.9)
:
1  fCCM;1 .1  xI C/; if 1=2 < x  1:

This function is extreme for all the feasible values of its parameter. As an
example, we provide the proof for 1  C < 3. Let g; h W Œ0; 1 ! Œ0; 1 be MDFF
2.2 Properties 29

y y y
1 1 1

1 x 1 x 1 x
36 60
Fig. 2.5 MDFF fCCM;1 for parameter values 1  C  2, C D 7
and C D 11

with 2fCCM;1 g C h. The conditions for a DFF to be maximal imply

g.0/ D h.0/ D 0 and g.1=2/ D h.1=2/ D 1=2:

Because of the symmetry condition (2.6), we only need to show that g.x/ D h.x/
for all x 2 .0; 1=2/:
 if Cx < 1, then fCCM;1 .x/ D 0, and hence g.x/ D h.x/ D 0 due to the range of g
and h;
 if 2 < C < 3 and 1=C  x < 1=2, then fCCM;1 .x/ D 1=2. Since g.x/; h.x/  1=2
for x  1=2, it follows that g.x/ D h.x/ D 1=2, for 1=C  x < 1=2. t
u
As discussed above, a non-extreme MDFF should not be used to obtain bounds,
because that would yield only a convex combination. However, even an extreme
MDFF needs not to yield corners of the dual polyhedron.
Example 2.6 The identity function is an EMDFF. Consider the one-dimensional
cutting stock instance that b 2 N items of length 5 have to be cut from initial
material of length 9. The identity function yields the bound 59 b according to the dual
variable u WD 5=9. However, any u 2 Œ0; 1 would have been a feasible dual variable,
and 5=9 is in the inner of this region. t
u

2.2.4 Extremality of Convex Functions

Proving or disproving the extremality of a DFF is usually non-trivial. When the

function is convex, we may use the following results to make these proofs easier.
Theorem 2.3 Let f W Œ0; 1 ! Œ0; 1 be a MDFF such that f is convex on Œ0; 1=2.
For given values a; b; c with 0 < a < b < 1=2 and c > 0, if f has a continuous
second derivative on Œa; b and f 00 .x/  c for all x 2 Œa; b, then f is not extreme.
Proof The proof consists of constructing two different MDFF g; h W Œ0; 1 ! Œ0; 1
with 2f g C h. These functions g; h are built by perturbing f inside .a; b/ such
30 2 Classical Dual-Feasible Functions

that g; h remain convex on Œ0; 1=2. We will have to choose an enough small  > 0.
Define the functions l W R ! Œ0; 2 and g; h W Œ0; 1 ! Œ0; 1 according to

1  cos.2x/; if 0 < x < 1;
l.x/ WD
0; otherwiseI

f .x/ C   l. ba
xa
/; if 0  x  1=2;
g.x/ WD
1  g.1  x/; otherwiseI
h.x/ WD 2f .x/  g.x/:

The function l is once continuously differentiable. It follows that f .x/ D g.x/ D h.x/
for all x 2 Œ0; 1 n .a; b/, all these functions are twice continuously differentiable in
.a; b/ and once in an environment U  .0; 1=2/ of the closed interval Œa; b. Since
g and h are symmetric, it remains to show, according to Lemma 2.3, that g0 ; h0 are
monotonely increasing on Œa; b and f 6 g. One obtains inside the interval .a; b/ the
following derivatives:

2  2 
g0 .x/ D f 0 .x/ C   sin .x  a/ I
ba ba
 2 2  2 
g00 .x/ D f 00 .x/ C   cos .x  a/ I
ba ba
h00 .x/ D 2f 00 .x/  g00 .x/

If a < x < aCb2

then g0 .x/ > f 0 .x/. Therefore, f 6 g.
To show the convexity of g in Œ0; 1=2, one may use the estimation g00 .x/ 
2 2
c  . ba /   0 for   . ba 2
/2 c inside the interval .a; b/. Due to g00 .x/  0
in .a; b/ the derivative g is monotonely rising in .a; b/. Because of g0 .a/ D
0

f 0 .a/ and g0 .b/ D f 0 .b/ and the continuous differentiability of g in U the

derivative g0 is also monotone in U. Therefore, g is convex in the needed interval
Œ0; 1=2.
Regarding h, the same considerations are valid. t
u
Because of Theorem 2.3, many functions f W Œ0; 1 ! Œ0; 1, which are convex
on Œ0; 1=2, can be proved to be non-extreme. In practice, while most of the MDFF
that may be built through Lemma 2.3 will be non-extreme, with Theorem 2.3, we
can avoid the search for the “best” MDFF in a large part of the search area. Because
of Theorem 2.3, to build an EMDFF by Lemma 2.3, one must ensure that f is linear
in all intervals where f 00 exists.
The following lemma may simplify the proofs that a certain MDFF is extreme in
the case where the function is piecewise linear. This restriction comes from part (a)
of Lemma 2.4.
Lemma 2.4 Let a; b be given values with 0  a < b  1=2, and a MDFF f W
Œ0; 1 ! Œ0; 1 such that, for all x; y 2 Œa; b, it holds that f .x C y/ D f .x/ C f .y/.
We have:
2.2 Properties 31

(a) f .x/ D f .b/f .a/

ba .xa/Cf .a/, for all x 2 Œa; b, i.e., f is linear in that interval;
(b) if g; h W Œ0; 1 ! Œ0; 1 are MDFF with

2f .x/ D g.x/ C h.x/; 8x 2 Œ0; 1;

and g.a/ D f .a/ and g.b/ D f .b/, then

g.x/ D f .x/; 8x 2 Œa; b:

Proof
(a) Let a0 WD a and b0 WD b. For arbitrary  2 Œa; b it will be shown that f ./ D
f .b/f .a/
ba  .  a/ C f .a/. Clearly, this holds for  2 fa; bg. Assume that for
some x; y 2 Œa; b a similar equation holds. Since f .x C y/ D 2  f . xCy
2 / due to
the prerequisites, it follows that
 
xCy f .x/ C f .y/
f D
2 2
f .b/  f .a/
D  .x  a C y  a/ C f .a/
2.b  a/
 
f .b/  f .a/ xCy
D   a C f .a/:
ba 2

Hence, the proposition is also true for .x C y/=2. To prove it for the given ,
an interval interlock is used. For n D 0; 1; 2; : : : , it is constructed as follows: if
2 > an Cbn then let anC1 WD .an Cbn /=2 and bnC1 WD bn , otherwise anC1 WD an
and bnC1 WD .an C bn /=2. That yields

a0      an            bn      b0

and lim .bn an / D 0, where the desired equation holds for all an and bn . Every
n!1
maximal dual-feasible function is monotonely increasing. Therefore,

f .b/  f .a/
 .an  a/  f ./  f .a/
ba
f .b/  f .a/
  .bn  a/
ba
for all n 2 N. The difference between the right and left part of the inequality
tends to zero. Therefore, f is continuous at , and the proposition (a) is true.
(b) If for certain numbers x; y 2 Œa; b, the equations g.x/ D f .x/ and g.y/ D f .y/
are valid, then the superadditivity condition yields

g.x C y/  f .x C y/ and h.x C y/  f .x C y/;

32 2 Classical Dual-Feasible Functions

 
and consequently, g.x C y/ D f .x C y/ D 2  f . xCy /. Therefore, g xCy

      2    2
f xCy2
and h xCy
2
 f xCy2
. That implies g xCy
2
D f xCy 2
. Choose
any  2 Œa; b. The proposition g./ D f ./ can be shown analogously
to part (a). The monotone sequences .an / and .bn / are defined as above.
Since g.an / D f .an / and g.bn / D f .bn / for all n 2 N, it follows that
g./ D f ./, because g, being a MDFF, is monotone, and f is continuous in
Œa; b. t
u
The following example shows in a simplified way how part (b) of Lemma 2.4 can
be used to prove that a given maximal dual-feasible function is extreme. Moreover,
the example also demonstrates the difficulty in the analysis of a parameter dependent
maximal dual-feasible function that is extreme for some parameter values and non-
extreme for others.
Example 2.7 The function fBJ;1 .xI C/ defined in (2.4) is extreme for C 2 N and for
C  2, but not for 1 < C < 2. Only part of the proof is provided. The remaining
part is left as an exercise.
For C 2 N, the assertion follows almost immediately from Lemma 2.4, because
fBJ;1 becomes the identity function. Suppose, g; h W Œ0; 1 ! Œ0; 1 are MDFF with
2fBJ;1 gCh. Therefore, we have g.0/ D h.0/ D 0 and g.1=2/ D h.1=2/ D 1=2.
Setting a WD 0 and b WD 1=2 in Lemma 2.4(b) closes the proof.
For 1 < C < 2, the function fBJ;1 .xI C/ is a convex combination of fBJ;1 .xI C/
Q and
another continuous MDFF g W Œ0; 1 ! Œ0; 1, where CQ D 2C if 1 < C < 4=3, and
4C
CQ D CC2 if 4=3  C < 2. If 1 < C < 4=3, then
8
ˆ
ˆ 0; if 0  x  1  C1 ;
ˆ
< .CxC1C/.43C/ ; if 1  1  x  1 ;
g.x/ D .32C/.2C/ C 2C
ˆ
ˆ
4CxC23C
42C ;
1
if 2C 1
 x  1  2C ;
:̂ 1
1  g.1  x/; if 1  2C  x  1;

and if 4=3  C < 2 then

8
ˆ
ˆ 0; if 0  x  1  C1 ;
< 2CxC22C
2C if 1  C1  x  3C2
; 4C ;
g.x/ D 1 3C2
ˆ ; if 4C  x  4C ;
CC2
:̂ 2
1  g.1  x/; if CC2
4C  x  1:

t
u
2.3 Generating One-Dimensional Dual-Feasible Functions 33

2.3 Generating One-Dimensional Dual-Feasible Functions

In this section, we show how to build non-trivial dual-feasible functions from simple
superadditive functions. We address first the simple case of linear combination, and
then, we explore the properties of composed dual-feasible functions. We explain
how to define a maximal dual-feasible function from a non-maximal function, while
alternative approaches are explored at the end.

2.3.1 Linear Combination

The simplest way to generate a dual-feasible function is to combine linearly two

functions. First, superadditivity is preserved: if f and g are superadditive functions,
then ˛f C ˇg (with ˛; ˇ 2 RC ) will be superadditive too. Similarly, combining lin-
early two maximal dual-feasible functions does not affect maximality as stated next.
Proposition 2.2 If f and g are discrete MDFF with the same domain f0; 1; : : : ; dg,
d 2 N n f0g, then h WD ˛f C ˇg (˛; ˇ 2 N, ˛ C ˇ > 0) is a discrete MDFF too.
Proof Since ˛; ˇ  0, the range of f and g implies h.x/  0 for all x 2 f0; 1; : : : ; dg.
We show that h is symmetric and superadditive. Since f and g are discrete MDFF,
one has for any x 2 f0; 1; : : : ; dg the following:

0 < f .d/ D f .x/ C f .d  x/ and 0 < g.d/ D g.x/ C g.d  x/

Therefore,

0 < ˛  f .d/ C ˇ  g.d/ D h.d/

D ˛  f .x/ C ˛  f .d  x/ C ˇ  g.x/ C ˇ  g.d  x/
D h.x/ C h.d  x/;

and hence h is symmetric. To verify the superadditivity, choose any x; y 2

f0; 1; : : : ; dg with x C y  d. Since f and g are superadditive, we have

f .x C y/  f .x/ C f .y/ and g.x C y/  g.x/ C g.y/:

That implies

h.x C y/ D ˛  f .x C y/ C ˇ  g.x C y/
 ˛  f .x/ C ˛  f .y/ C ˇ  g.x/ C ˇ  g.y/
D h.x/ C h.y/:

t
u
34 2 Classical Dual-Feasible Functions

Note that, when functions with domain and range Œ0; 1 are considered, ˛f makes
little sense, and ˛f C ˇg (with ˛; ˇ 2 RC ) should be replaced by a convex
combination of f and g. In practice, Proposition 2.2 is meaningful mainly for
discrete functions.
Preserving superadditivity does not imply that maximality is preserved too. For
instance, while b f c and minff ; gg remain superadditive, function minf f ; gg is not
maximal, unless f D g, nor is b f c (x 7! bxc is not even a MDFF).
Extremality was only introduced for maximal dual-feasible functions with
domain and range Œ0; 1. For these functions, and by definition, if one has two
˛
different maximal dual-feasible functions f ; g W Œ0; 1 ! Œ0; 1, then h WD ˛Cˇ fC
ˇ
˛Cˇ
g with ˛; ˇ > 0 cannot yield an extreme maximal dual-feasible function. Only
the trivial combination with ˇ D 0 and f being extreme, or ˛ D 0 and g being
extreme, yields an extreme maximal dual-feasible function.

2.3.2 Composition

Composing functions is another possible approach to build dual-feasible functions.

As with linear combinations, superadditivity and maximality also remain with
composition.
Proposition 2.3 Let f ; g W Œ0; 1 ! Œ0; 1 be two MDFF. The composed function
f .g.x// is also a MDFF.
Proof Let x; y 2 Œ0; 1 with x C y  1. We have

f .g.0// D f .0/ D 0;
f .g.1  x// D f .1  g.x// D 1  f .g.x//; and
g.x C y/  g.x/ C g.y/;

and hence

f .g.x C y//  f g.x C y/  g.x/  g.y/ C f .g.x// C f .g.y//

 f .g.x// C f .g.y//:

The composition f .g.x// fulfills the sufficient conditions to be a MDFF. Fur-

thermore, since f .g.x// is superadditive and f .g.x//  0, this function is also
non-decreasing. t
u
On the contrary, composing maximal dual-feasible functions that are extreme
does not always yield a function that is extreme as illustrated next.
Proposition 2.4 The composition of extreme MDFF is not necessarily extreme.
2.3 Generating One-Dimensional Dual-Feasible Functions 35

y y y y
1 1 1 1

fFS,1 (·;1) fF S,1 (·;2) fFS,1 (·;3) fFS,1 (·;4)

1 x 1 x 1 x 1 x

Fig. 2.6 MDFF fFS;1 for parameter values k 2 f1; : : : ; 4g

Proof Consider the MDFF fFS;1 .xI k/ W Œ0; 1 ! Œ0; 1


x; if .k C 1/  x 2 N;
fFS;1 .xI k/ WD (2.10)
b.k C 1/  xc=k; otherwise;

with k D 2, and fBJ;1 .xI C/ defined in (2.4) with C D 9=2, and let f WD fFS;1 . fBJ;1 .//.
Both fFS;1 and fBJ;1 are extreme. One has fBJ;1 .10=27/ D 1=3, and fBJ;1 .x/ ¤ 1=3 for
x 2 Œ0; 1 n f10=27g.
The function fFS;1 is illustrated in Fig. 2.6 for several parameter values.
Because of fFS;1 .x/ D 0 for 0  x < 1=3, fFS;1 .1=3/ D 1=3 and fFS;1 .x/ D 1=2
for 1=3 < x < 2=3, it follows that
8
ˆ
ˆ 0; if 0  x < 10=27;
<
1=3; if x D 10=27;
f .x/ D
ˆ 1=2; if 10=27 < x < 17=27;
:̂
1  f .1  x/; if 17=27  x  1:

Let g; h W Œ0; 1 ! Œ0; 1 be two MDFF defined as

8 8
< 0; if 0  x  10=27; < 0; if 0  x < 10=27;
g.x/ WD 1=2; if 10=27 < x < 17=27; and h.x/ WD 1=2; if 10=27  x  17=27;
: :
1; if 17=27  x  1; 1; if 17=27 < x  1:

Then, f .x/ D 13 g.x/ C 23 h.x/, for all x 2 Œ0; 1, and hence f cannot be extreme. t
u

2.3.3 Symmetry

Maximal dual-feasible functions can be built from superadditive functions by

keeping the images of the values smaller than 1=2, and by computing the images
of the values larger than 1=2 by symmetry.
36 2 Classical Dual-Feasible Functions

Theorem 2.4 Let f W Œ0; 1 ! Œ0; 1 be a superadditive function. The following

function fO W Œ0; 1 ! Œ0; 1, defined as
8
< f .x/; if 0  x < 1=2;
fO .x/ WD 1=2; if x D 1=2;
:
1  f .1  x/; if 1=2 < x  1;

is a maximal dual-feasible function dominating f .

Many standard and simple superadditive functions can be used to generate MDFF
using Theorem 2.4. An example is provided and discussed in Sect. 2.4.

2.3.4 Using the Limiting Behaviour of a Function

A maximal dual-feasible function can sometimes be built from a superadditive

function f W Œ0; 1 ! Œ0; 1 when, at some point x where f is not continuous, the
value of f .x/ can be increased without modifying the other values. This approach
yields an improved function with some singular values x such that

lim f .y/ < f .x/ < lim f .y/:

y"x y#x

 
Let fN x .x/ D f .x/ if x ¤ x , and fN x .x / D f .x / C ", with " being a sufficiently

small positive real value. Note that if f is a dual-feasible function, fN x may or may
not be a dual-feasible function. Furthermore, for a given dual-feasible function f ,

let I1 be the set of values x whose images can be increased such that fN x is also a
dual-feasible function. Below, we assume that I1 is a discrete set.
Theorem 2.5 Let f W Œ0; 1 ! Œ0; 1 be a superadditive function. Let I1 be the set of
values x 2 Œ0; 1, for which positive " exist such that fN x is a dual-feasible function.
Assume that f is continuous from the right in the entire set I2 WD Œ0; 1 n I1 of the
remaining values and that the function g W Œ0; 1 ! Œ0; 1 is such that the following
holds:
1. f .x/  g.x/  lim f .y/, for any x in I1 ;
y#x
2. g.x/ C g.y/  g.x C y/, if x; y; x C y 2 I1 ;
3. g.x/ C f .y/  g.x C y/, if x; x C y 2 I1 and y 2 I2 .
Then the function h W Œ0; 1 ! Œ0; 1 with

g.x/; if x 2 I1 ;
h.x/ WD
f .x/; if x 2 I2 ;

is superadditive.
2.3 Generating One-Dimensional Dual-Feasible Functions 37

y y
1 f 1 g

1 x 1 x
Fig. 2.7 Increasing the values of a function at its discontinuities

Example 2.8 The function fFS;1 .I k/ W Œ0; 1 ! Œ0; 1, with k 2 N n f0g, discussed
in Sect. 2.3.2 is a MDFF. Hence, the restriction of fFS;1 to the domain Œ0; C with
1
kC1
 C < 1 is superadditive and non-decreasing. Normalizing this function to
domain and range Œ0; 1 yields the following function f , which depends on k and C
and which is not necessarily symmetric:

Cx=fFS;1 .CI k/; if .k C 1/  Cx 2 N;
f .x/ WD
b.k C 1/  Cxc=.k  fFS;1 .CI k//; otherwise.

1 2
This function is continuous on Œ0; 1 except at the points .kC1/C ; .kC1/C ; : : : Let
for example k WD 5 and C WD 5=8, for which fFS;1 .CI k/ D 3=5. Then,

25x=24; if 15x=4 2 N;
f .x/ D
b15x=4c=3; otherwise;
4
5
8

and f 15 D 18 , f 15 D 59 and f 45 D 56 , while f is equal to 0, 13 , 23 or 1 in

the other points of the interval Œ0; 1. This allows to set g.x/ WD b15x=4c=3 for all
x 2 Œ0; 1, and it yields h g. However, this function is not symmetric, as for
example g.1=2/ D 1=3 < 1=2 shows (Fig. 2.7).
Note that to further obtain a maximal dual-feasible function, we may apply
Theorem 2.4. t
u

2.3.5 Rounding Functions

Another way of generating a dual-feasible function consists in applying two

superadditive functions separately, one to the integer part of a given value, and
the other to its remainder. This approach is valid provided that the conditions of
Lemma 2.5 are satisfied.
38 2 Classical Dual-Feasible Functions

Lemma 2.5 Let f W Œ0; C ! RC and g W Œ0; 1 ! Œ0; 1 be two superadditive

functions, C  1 and

v  WD supfg.y/ C g.z/  g.y C z  1/ j y; z 2 .0; 1 ^ y C z > 1g:

If

f .x C 1/  f .x/  v 

for all x 2 Œ0; C  1, then the function h W Œ0; C ! Œ0; f .bCc/ C 1 defined by

h.x/ WD f .bxc/ C g.frac.x//

is superadditive on Œ0; C.

Proof Choose any x; y 2 Œ0; C with x C y  C. If frac.x/ C frac.y/ < 1, then
frac.x C y/ D frac.x/ C frac.y/ and bx C yc D bxc C byc, such that h.x C y/ 
h.x/ C h.y/ follows immediately from the superadditivity of the functions f and g.
If frac.x/ C frac.y/  1, then frac.x C y/ D frac.x/ C frac.y/  1 and bx C yc D
bxc C byc C 1. Hence, one gets

h.x C y/  h.x/  h.y/ D f .bx C yc/  f .bxc/  f .byc/ C g.frac.x C y//

g.frac.x//  g.frac.y//
 f .bx C yc/  f .bx C yc  1/  v 
 0:

t
u
On the other hand, although the ceiling function is not superadditive, it can lead
to superadditive functions if it is decreased by a suitable value. We now generalize
several results that use this kind of method.
Lemma 2.6 Let f W RC ! RC be a superadditive function. If ˇ  1, then g.x/ WD
maxf0; df .x/e  ˇg is superadditive.
Proof Since f is a superadditive function with domain and range RC , it is
nondecreasing, and hence g is also nondecreasing. Choose any x; y  0 with x  y.
If df .x/e  ˇ, then g.x/ D 0, such that the superadditivity of g follows from its
monotonicity. If df .x/e > ˇ, then

g.x C y/  g.x/  g.y/ D df .x C y/e  ˇ  df .x/e  df .y/e C 2ˇ

 ˇ  1;

since f is superadditive. t
u
2.4 Examples 39

2.4 Examples

In this section, we illustrate the methods discussed above to generate 1-dimensional

dual-feasible functions through different examples of functions that have been
described in the literature. A particular staircase function is also presented to show
the variety of dual-feasible functions that may be further derived.

2.4.1 Applying Symmetry

The function fCCM;1 discussed in Example 2.5 (p. 28) can also be obtained by
applying Theorem 2.4 to the function x 7! bCxc, C  1:
8
< bCxc=bCc; if 0  x < 1=2;
fCCM;1 .xI C/ WD 1=2; if x D 1=2;
:
1  fCCM;1 .1  xI C/; if 1=2 < x  1:

Rounding-down violates the symmetry, such that the obtained function x 7! bCxc
is not maximal.
Another example on how applying symmetry may yield a maximal dual-feasible
function is described next. The function g.x/ described in Sect. 2.1 is a monotone
and superadditive dual-feasible function for every k 2 N n f0; 1g:

g.x/ WD dmaxf0; kx  1ge =.k  1/:

However, it is not maximal (see for example Fig. 2.2). By forcing symmetry as
described in Theorem 2.4, we get the maximal dual-feasible function fVB;2 W Œ0; 1 !
Œ0; 1 with
8
< dmaxf0; kx  1ge=.k  1/; if 0  x < 1=2;
fVB;2 .xI k/ WD 1=2; if x D 1=2; (2.11)
:
1  fVB;2 .1  xI k/; if 1=2 < x  1:

Figure 2.8 illustrates this function for k D 6.

2.4.2 Using Rounding Functions and Applying Symmetry

In this subsection, l m C > 1 and that C is not integer. Let k be an integer

we assume
1
constant with k  frac.C/ . The following function fLL;1 W Œ0; 1 ! Œ0; 1 is based
40 2 Classical Dual-Feasible Functions

Fig. 2.8 Function fVB;2 for

kD6
y
1 fV B,2

1 x

on Lemma 2.5:
n l m o
bCxc C max 0; frac.Cx/frac.C/
1frac.C/  .k  1/ =k
fLL;1 .xI C; k/ WD : (2.12)
bCc

The superadditivity of this function is due to Lemma 2.6. In the following proof,
we use the fact that dxCye D xCdye if x is integer, and that dxeCdye  dxCyeC1
for any x and y. Without the maximum expression, the function fLL;1 would be the
function (2.1). The function fLL;1 is derived from another superadditive function h
by a linear transformation, namely fLL;1 .x/ D h.Cx/=bCc, where the structure of h
is of the kind f .bxc/ C g.frac.x// as in Lemma 2.5.
Proposition 2.5 Function fLL;1 is superadditive.
Proof To show the superadditivity of fLL;1 , Lemmas 2.5 and 2.6 are used. Define the
function g W Œ0; 1 ! Œ0; 1 as
  
x  frac.C/
g.x/ WD max 0;  .k  1/ =k:
1  frac.C/

xfrac.C/
The range of g is indeed part of Œ0; 1, because x 2 Œ0; 1 implies 1frac.C/  1. In the
x
following, the superadditivity of g is proved. The function x 7! 1frac.C/  .k  1/
frac.C/
is linear, and hence it is obviously superadditive. The constant 1frac.C/
 .k  1/ is
at least one, because
 
1 1  frac.C/
k1 1  :
frac.C/ frac.C/

If f is any superadditive function and ˛  0, then x 7! f .x/˛ is also superadditive.

frac.C/
Hence, Lemma 2.6 can be applied, even if the expression 1frac.C/  .k  1/ stands
inside the rounding brackets. The function g is therefore superadditive.
2.4 Examples 41

The identity function is superadditive. To apply Lemma 2.5 with the above
defined function g, we must show that v   1. Choose any y; z 2 .0; 1 with
y C z > 1. Since g.y/, g.z/ and g.y C z  1/ 2 Œ0; 1, the inequality

g.y/ C g.z/  g.y C z  1/  1

is obviously fulfilled for g.y/  g.z/ D 0. Assume g.y/ and g.z/ > 0. Hence, we
have y and z > frac.C/. One gets
 
y C z  1  frac.C/
k  g.y C z  1/   .k  1/
1  frac.C/
  
y  frac.C/ C z  frac.C/
D  1  .k  1/
1  frac.C/
   
y  frac.C/ z  frac.C/
  .k  1/ C  .k  1/  k
1  frac.C/ 1  frac.C/
D k  .g.y/ C g.z/  1/;

as needed, because k 2 N.
Finally, fLL;1 .x/ D h.Cx/=bCc with the function h according to Lemma 2.5. The
linear transformations do not affect the superadditivity. Since h is superadditive, fLL;1
is that too. t
u
The function fLL;1 is not maximal, since there are cases where it is not symmetric.
An improved version of this function can be obtained by applying Theorem 2.4.
Proposition 2.6 The following function fLL;2 .I C; k/ W Œ0; 1 ! Œ0; 1 with C > 1,
C … N and k 2 N, k  d1=frac.C/e is a maximal dual-feasible function, and it
dominates fLL;1 .
8 n l m o
ˆ
ˆ bCxcCmax 0;
frac.Cx/frac.C/
.k1/ =k
< 1frac.C/
; if 0  x < 1=2;
bCc
fLL;2 .xI C; k/ WD 1=2; if x D 1=2; (2.13)
ˆ
:̂ 1  fLL;2 .1  xI C; k/; if 1=2 < x  1:

The graphs of the functions (2.12) and (2.13) are drawn in Fig. 2.9 for C 2 f3:3; 3:4g
and k D d1=frac.C/e.

2.4.3 Improving a Function by Using Its Limiting Behaviour

The following four parameter dependent functions uA ; uB ; uC ; uD W Œ0; 1 !

Œ0; 1 are among the first functions described explicitly in the literature as being
42 2 Classical Dual-Feasible Functions

y y y y
1 1 1 1

fLL,1 (·;3.3, 4) fLL,1 (·;3.4, 3) fLL,2 (·; 3.3, 4) fLL,2 (·; 3.4, 3)

1 x 1 x 1 x 1 x

Fig. 2.9 Original fLL;1 .I C; k/ and improved fLL;2 .I C; k/

dual-feasible functions:
(
1
0; if 0  x  pC1 ;
uA .xI p/ WD 1 1 p 2 N n f0g;
p ; if pC1 < x  1;
8
ˆ
< 0; if 0  x  ap  pC1
p1
;
1 1 2
uB .xI p; a/ WD pC1 C .x  pC1 /=.p  ap  ap /; if ap  pC1 < x < ˇ;
p1

:̂ 1
p; if ˇ  x  1;
p1 1
p 2 N n f0; 1g; a 2 . 2
; /;
p Cp pC1
uC .xI p/ WD maxf0; d.p C 1/x  1e=pg; p 2 N n f0g;
 
x x
uD .xI p; a/ WD =p C uB .x  ˇ  I p; a/;
ˇ ˇ
p1 1
p 2 N n f0; 1; 2g; a 2 . 2
; /;
p Cp pC1
2
where ˇ WD pC1 a for the functions uB and uD . The function uC is equivalent to the
2
function (2.2) with k D pC1. Furthermore, we have for all x 2 Œ0; pC1  that uA .x/ D
uC .x/. Similarly, uB .x/ D uD .x/ holds for all x 2 Œ0; ˇ. For larger x inside the inter-
val Œ0; 1 one gets uC .x/ > uA .x/ or uD .x/ > uB .x/, respectively. The function uA is a
simple staircase function. The dependence of uB and uD on the parameter a is shown
in Fig. 2.10 for p D 3. Since uD is continuous and piecewise linear, it looks like the
function (2.4) (p. 25) for C WD 1=ˇ. Checking this observation is left as an exercise.
The function fFS;1 W Œ0; 1 ! Œ0; 1 depending on the parameter k 2 N n f0g
discussed in Sect. 2.3.2:

x; if .k C 1/  x 2 N;
fFS;1 .xI k/ WD
b.k C 1/  xc=k; otherwise;

is a maximal dual-feasible function. This function was obtained from uC by applying

Theorem 2.5. The values that are equal to 1=.k C1/, 2=.k C1/, : : :, k=.k C1/ remain
unchanged, while the other ones are submitted to a rounding function.
2.4 Examples 43

y y y
1 1 1

uB (·;3, 15 ) uB (·;3, 13
72 ) uB (·;3, 24
5
)
1 x 1 x 1 x

y y y
1 1 1

uD (·;3, 15 ) uD (·;3, 13
72 ) uD (·;3, 24
5
)

1 x 1 x 1 x
˚1 13 5

Fig. 2.10 Dual-feasible functions uB and uD for p D 3 and a 2 5
; ;
72 24

The following function fDG;1 is also built on Theorem 2.5. Since the proof of
superadditivity is complex and long, it is omitted.
1
Proposition 2.7 For every C 2 RnN with C > 1, and any k 2 N with k  d frac.C/ e,
the following function fDG;1 .I C; k/ W Œ0; 1 ! Œ0; 1 is a MDFF:
8 frac.Cx/frac.C/ frac.Cx/frac.C/
bCxc 1 < 1frac.C/ ; if .k  1/  1frac.C/ 2 N;
ˆ
fDG;1 .x/ D C  n l m o
bCc bCc :̂
max 0; .k  1/  frac.Cx/frac.C/
1frac.C/ =k ; otherwise.
(2.14)
This function also dominates fLL;1 , if fLL;1 is not symmetric. In this case, the
functions (2.12), (2.13) and (2.14) differ at some isolated points. The graph of the
latter function is presented for C 2 f3:3; 3:4g and k D d1=frac.C/e in Fig. 2.11.

2.4.4 A Special Case: A Staircase Function

with Infinitely Many Stairs

All maximal dual-feasible functions that were presented until now have a relatively
simple structure. For instance, the function fBJ;1 is Lipschitz-continuous, while all
the other discussed dual-feasible functions have a finite number of discontinuities.
44 2 Classical Dual-Feasible Functions

Fig. 2.11 Function fDG;1

y y
1 1

fDG,1 (·;3.3, 4) fDG,1 (·;3.4, 3)

1 x 1 x

The following example illustrates that maximal dual-feasible functions may be

much more complicated.
Although it has an infinite number of stairs, the function f W Œ0; 1 ! Œ0; 1
defined as follows
81 2 1
< 2 C 1=b 2x1 c; for 2 < x  1;
f .x/ WD 1=2; for x D 1=2; (2.15)
:1 2 1
2
 1=b 12x
c; for 0  x < 2
;

is a maximal dual-feasible function, which is differentiable at x0 D 1=2 with the

derivative f 0 .1=2/ D 1.
Proof Clearly, f is monotonically increasing and symmetric in Œ0; 1, and it has for
all n 2 N with n > 2 discontinuities at 12 ˙ 1n . The stairs have the levels 0, 1 and
1 1
2 ˙ n . Furthermore, f .x/ D 0 for 0  x < 1=6. To show the superadditivity, a small
case distinction is necessary. According to Theorem 2.2, assume 0 < x1  x2 < 1=2
and x1 C x2  2=3.
• If x1 C x2  1=2, then f .x1 C x2 /  x1 C x2  f .x1 / C f .x2 /, because f .x/  x
for 0  x  1=2.
• If 0  x1 < 1=6, then f .x1 / D 0, and therefore, f .x1 C x2 /  f .x1 / C f .x2 /.
• Let 1=6  x1  x2 and x1 C x2 < 1=2. Then, x2 < 1=3 and x1 < 1=4, i.e.,
f .x1 / D 1=6.
– If x2 < 1=4, then f .x2 / D 1=6 and f .x1 C x2 /  f .2  1=6/ D 1=3 D
f .x1 / C f .x2 /.
– If 1=4  x2 < 3=10, then f .x2 / D 1=4 and f .x1 C x2 /  f .5=12/ D 5=12 D
f .x1 / C f .x2 /.
– If 3=10  x2 < 1=3, then f .x2 / D 3=10 and f .x1 C x2 /  f .7=15/ D 7=15 D
f .x1 / C f .x2 /.
2.5 Related Literature 45

Fig. 2.12 Maximal

dual-feasible staircase
y
function (2.15) with infinitely 1 f
many stairs

1 x

To show that f is differentiable with derivative 1 at x0 D 1=2, let x D 1=2C1=.aCb/

with a 2 N, a  2 and 0  b < 1. (The case x < 1=2 can be handled analogously
or by use of the symmetry.) Then,

1 1
f .x/  f .x0 /  1  .x  x0 / D 2

b 1C 2 c aCb
aCb 1

1 1
D 
ba C bc a C b
1 1
D 
a aCb
b
D 2 :
a C ab

Hence, the fraction is nonnegative. When x tends to 1=2, then a ! 1, and therefore
f .x/  f .x0 /  1  .x  x0 / D O.x  x0 /2 . This means that the derivative at x0 D 1=2
exists and has the value 1. t
u
Figure 2.12 shows a simplified version of this staircase function. Around the
point .1=2I 1=2/, there are infinitely many stairs, which cannot be drawn exactly.
Moreover, the usual black and white marks at discontinuities become meaningless,
and therefore, they are omitted.

2.5 Related Literature

The concept of 1-dimensional dual-feasible function was described first by Johnson

(1973). In the same year, a subclass of dual-feasible functions restricted to super-
additive and nondecreasing functions was described to strengthen cuts for integer
programs by Chvátal (1973) [see also Aardal and Weismantel (1997), Nemhauser
46 2 Classical Dual-Feasible Functions

and Wolsey (1998)]. Lueker (1983) introduced the designation. He used the dual-
feasible functions to derive lower bounds for bin-packing problems. The functions
uA .I p/, uB .I p; a/, uC .I p/ and uD .I p; a/ described in Sect. 2.4.3 are due to this
author. The notion of maximality was proposed and described initially in Carlier
and Néron (2007a), while the conditions for maximality were presented by these
authors in Carlier and Néron (2007b). In the latter, the authors defined also the
discrete version of a dual-feasible function that they call redundant functions. They
used these functions to solve scheduling problems. The conditions for maximality
were later reviewed by Rietz et al. (2010) yielding the sufficient conditions described
in Theorem 2.2. The proof of Theorem 2.4 can be found in Clautiaux et al. (2010)
for the case where the function is defined as a discrete dual-feasible function. The
extremality of dual-feasible functions was analyzed first by Rietz et al. (2012a). The
proofs of many of the results concerning extremality can be found in this reference.
Within this chapter, different dual-feasible functions were used as examples.
Most of them were taken from the literature where they were stated either explicitly
or implicitly. The letters used in the index of these functions identify the authors
that proposed them originally. We provide the original source next. The function
fBJ;1 .I C/ is based on the work of Burdett and Johnson (1977). The function
fCCM;1 .I C/ was proposed by Carlier et al. (2007), improving slightly a function
proposed before by Boschetti and Mingozzi (2003). The function fFS;1 .I k/ is due to
Fekete and Schepers (2001). The function fVB;2 .I k/ is based on a non-maximal dual-
feasible function by Vanderbeck (2000). The function fVB;2 .I k/ is maximal, and it
was described first by Clautiaux et al. (2010). The function fDG;1 .I k/ was defined by
Dash and Günlük (2006). The function fLL;1 .I C; k/ was used implicitly by Letchford
and Lodi (2002) to strengthen Chvátal-Gomory cuts (Chvátal 1973) and Gomory
fractional cuts (Gomory 1958) in linear programs. As shown in Sect. 2.4.2, this
function is superadditive but not maximal. Again, the corresponding maximal dual-
feasible function was defined by Clautiaux et al. (2010).

2.6 Exercises

1. Which of the following functions f1 ; f2 W Œ0; 1 ! Œ0; 1 are dual-feasible?


x; if x 2 Q;
f1 .x/ WD
1=d1=xe; otherwise,

0; if x D 0;
f2 .x/ WD
1=b1=xc; otherwise.

2. Show that if f W R ! R is superadditive, then bf c remains superadditive.

2.6 Exercises 47

3. Which of the following functions f3 ; f4 W RC ! RC are superadditive?

f3 .x/ WD bx2 c;

f4 .x/ WD bex c:

4. Let g0 W f0; 1; : : : ; 19g ! f0; 1; : : : ; 10g be the following discrete function:

x 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
g0 .x/ 0 0 0 0 0 2 3 4 5 0 0 0 0 0 0 0 0 0 0 10

(a) Provide a discrete dual-feasible function g1 W f0; 1; : : : ; 19g !

f0; 1; : : : ; 10g with g1 .15/ D 9, which dominates g0 and is maximal.
(b) What is the equivalent dual-feasible function f5 W Œ0; 1 ! Œ0; 1 to g1
according to (2.3)?
(c) Provide a continuous, piecewise linear maximal dual-feasible function f6 W
Œ0; 1 ! Œ0; 1, which dominates f5 .
(d) Define a staircase maximal dual-feasible function f7 W Œ0; 1 ! Œ0; 1, which
dominates f5 .
5. Consider an instance of the 1D-CSP with given nonnegative integer data
m; L; l; b, where 0 < `i  L for i D 1; : : : ; m. Which of the following statements
are true, and which are not?
Hint: The functions fFS;1 ; fVB;2 ; fBJ;1 and fCCM;1 were defined in (2.4), (2.9), (2.10),
(2.11) (pp. 25, 28, 35, 39, respectively).
(a) If f W Œ0; 1 ! Œ0; 1 is a maximal dual-feasible function, then there is no
P
m P
m
other dual-feasible function g W Œ0; 1 ! Œ0; 1 with bi f .`i =L/ < bi 
iD1 iD1
g.`i =L/.
(b) For all k 2 N n f0g, there is always a C  1, C 2 R with fFS;1 .`i =LI k/ D
fBJ;1 .`i =LI C/ for i D 1; : : : ; m.
(c) For all k 2 N n f0; 1g, there is always a C  1, C 2 R with fVB;2 .`i =LI k/ D
fCCM;1 .`i =LI C/ for i D 1; : : : ; m.
(d) For all k 2 N n f0; 1g, there is always a C  1, C 2 R with fVB;2 .`i =LI k/ D
fBJ;1 .`i =LI C/ for i D 1; : : : ; m.
6. Show, using Theorem 2.2, that fBJ;1 .xI C/ defined in (2.4) is maximal for all
C  1.
7. Let p 2 R be a constant with p > 1. Is the function f W Œ0; 1 ! Œ0; 1 with

.2x/p =2; if x  1=2;
f .x/ WD
1  .2  2x/ =2; otherwise,
p

a maximal dual-feasible function? Is it an extreme maximal dual-feasible function?

48 2 Classical Dual-Feasible Functions

Hint: Refer to Lemma 2.3 and Theorem 2.3.

8. Prove that fFS;1 .xI k/ is extreme.
9. Which of the following statements are true and which are not?
(a) The composition f .g.// of two maximal dual-feasible functions f ; g W
Œ0; 1 ! Œ0; 1 is always a non-extreme maximal dual-feasible function if
f is non-extreme and g is surjective, i.e., a mapping onto and not only into
Œ0; 1.
(b) A maximal dual-feasible function f W Œ0; 1 ! Œ0; 1 is surjective if and only
if it is continuous.
(c) If an extreme maximal dual-feasible function f W Œ0; 1 ! Œ0; 1 is convex on
Œ0; 1=2, then f is continuous on Œ0; 1.
10. Let  2 .0; 1=2 be a fixed parameter for the function fMT;0 W Œ0; 1 ! Œ0; 1,
defined by
8
< 0; if x < ;
fMT;0 .xI / WD 1; if x > 1  ; (2.16)
:
x; otherwise.

(a) Show that this function is a maximal dual-feasible function.

(b) Show that this function is extreme for   1=4.
Hint: Refer to Lemma 2.4.
11. Show according to Definition 2.6 (p. 28) that if one has the constants ˛; ˇ 2 RC
with ˛ˇ > 0 and two different maximal dual-feasible functions f ; g W Œ0; 1 ! Œ0; 1,
˛ ˇ
then h WD ˛Cˇ f C ˛Cˇ g is a non-extreme maximal dual-feasible function.
12. Give an example that the composition of two non-symmetric, non-superadditive
dual-feasible functions f ; g W Œ0; 1 ! Œ0; 1 may yield a maximal dual-feasible
function.
13. The function fLL1 W Œ0; 1 ! Œ0; 1 was defined in (2.12) as
    
frac.Cx/  frac.C/
fLL;1 .xI C; k/ WD bCxc C max 0;  .k  1/ =k =bCc:
1  frac.C/

For which parameter values C > 1, C 2 R n N and k 2 N, k  d1=frac.C/e is this

function symmetric?
14. Show that the function fVB;1 W Œ0; 1 ! Œ0; 1, defined in (2.2) as

fVB;1 .xI k/ WD dmaxf0; kx  1ge =.k  1/;

with k 2 N n f0; 1g, is not symmetric for any parameter value k.

2.6 Exercises 49

15. Let C 2 R with C  1 be a constant. Which of the following statements are

true and which are not?
(a) For every maximal dual-feasible function g W Œ0; 1 ! Œ0; 1, the function
f W Œ0; 1 ! RC , defined as f .x/ WD bCxc C g.frac.Cx//, is superadditive.
(b) For every superadditive function g W R ! R, the function f W R ! RC ,
defined as f .x/ WD maxf0; dg.x/e  Cg, is superadditive.
(c) If C 2 N, then for every dual-feasible function f W Œ0; 1 ! Œ0; 1, and any
finite index set I of nonnegative real numbers xi with i 2 I, the following
implication holds:
X 1 X 1
xi  H) f .xi /  :
i2I
C i2I
C
Chapter 3
General Dual-Feasible Functions

3.1 Introduction

Classical dual-feasible functions are defined only for nonnegative arguments thus
limiting their applicability. In this chapter, we explore the extension of dual-feasible
functions to more general domains with a focus on real numbers. Other attempts
of generalizing the concept of dual-feasible function will be done later in the book.
In Chap. 4, we will discuss for instance an extension to multidimensional domains
yielding the so-called vector packing dual-feasible functions, which may be used to
compute bounds for vector packing problems.
Extending the principles of dual-feasible functions to the domain of real numbers
is not trivial. The properties that apply to dual-feasible functions, and which have
been reviewed in the previous chapter, are affected in this exercise, and some of
them are even lost. This makes the task of deriving good non-dominated functions
much more difficult. In the sequel, we will explore in depth the new properties of
general dual-feasible functions. Different examples will be brought to discussion to
illustrate the main and new ideas behind these functions.
Given the hardness in deriving dual-feasible functions that apply to the domain
of real numbers, we will devote the second part of the chapter to the presentation
of general construction principles that lead to specific instances of general dual-
feasible functions. The defining characteristics of these principles will be described
first, and followed by the analysis of specific examples for each case.

© Springer International Publishing Switzerland 2016 51

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5_3
52 3 General Dual-Feasible Functions

3.2 Extension of Dual-Feasible Functions to General

Domains

3.2.1 Definition

A general dual-feasible function is a generalization of the classical dual-feasible

functions to any real arguments, i.e. arguments that are not restricted to nonnegative
values. The definition of a general dual-feasible function states formally as follows.
Definition 3.1 A function f W R ! R is a general dual-feasible function, if for any
finite index set I of real numbers xi 2 R, i 2 I, it holds that
X X
xi  1 H) f .xi /  1: (3.1)
i2I i2I

Extending the domain and range of dual-feasible functions to R instead of

Œ0; 1 has several facets. As we will see, this extension is far from trivial, because
important known properties of the bounded classical case are lost. For instance,
any maximal dual-feasible function with domain and range Œ0; 1 obeys a certain
symmetry rule. It was possible to characterize these functions by an equivalence,
i.e. they were maximal dual-feasible functions if and only if they fulfilled a certain
set of rules. However, while we extend the domain and range to the field of real
numbers, the mentioned symmetry rule is not longer a necessary property of the
maximal dual-feasible functions. Some similar rules as in the classical case remain
sufficient, but it is not longer possible to formulate an analogous equivalence in the
generalized case.
Example 3.1 The classical dual-feasible function fBJ;1 W Œ0; 1 ! Œ0; 1 was defined
in Chap. 2 for any real parameter C  1 as
  
frac.Cx/  frac.C/
fBJ;1 .xI C/ WD bCxc C max 0; =bCc;
1  frac.C/

where frac./ denotes the non-integer part of an expression, i.e. frac.C/ C  bCc.
This function can be extended to a general dual-feasible function using the same
formula. Note however that this does not work for many other functions. Figure 3.1
illustrates this function fBJ;1 for C 2 f6=5; 8=5g.
t
u
Further conditions for a function f W R ! R to be a general dual-feasible
function have already been identified. The following proposition describes two of
them. The proof of their validity is left as an exercise (see Exercise 1 at the end of
the chapter).
Proposition 3.1 Let f W R ! R be any function. If f is a general dual-feasible
function, then it has the following two properties.
3.2 Extension of Dual-Feasible Functions to General Domains 53

y y

1 1

fBJ,1 ·; 65 fBJ,1 ·; 85

−1 1 x −1 1 x

−1 −1

Fig. 3.1 fBJ;1 as a maximal general dual-feasible function

1. For any x 2 .0; 1, it holds that f .x/  1=b1=xc.

2. For any finite set fxi 2 R W i 2 Ig of real numbers, the following holds:
X X
xi  0 H) f .xi /  0 (3.2)
i2I i2I

The converse is generally false.

These properties are necessary but not sufficient as the following example illustrates.
Example 3.2 Let c > 1 be any real constant. The following function f shows that
even when the properties (1.) and (2.) in Proposition 3.1 are fulfilled, the function
may not be a general dual-feasible function:

x; if 0  x  1;
f .x/ WD
cx; otherwise:

The property (1.) is obviously fulfilled. The same happens with (2.) because f .x/ 
cx for all x 2 R. However, we have that f .1/ C f .2/ D c > 1, in contradiction to
1 C 2  1, and thus a violation of the defining condition (3.1). t
u
On the other hand, the composition of general dual-feasible functions leads to
functions that are still general dual-feasible functions.
Lemma 3.1 The composition of general dual-feasible functions f ; g W R ! R is a
general dual-feasible function.
P
Proof Let I be any finite index set of real numbers xi with xi  1. One gets
P i2I
g.xi /  1, because g is a general dual-feasible function. Therefore, Definition 3.1
i2I
54 3 General Dual-Feasible Functions

yields
X
f .g.xi //  1;
i2I

because f is also a general dual-feasible function. t

u

3.2.2 Maximality

As happens with classical dual-feasible functions, only non-dominated functions

are of interest. In this section, similarly to the bounded case, the notion of maximal
general dual-feasible function is introduced. We will also provide characterising
conditions for these functions and point out the possible loss of symmetry.
Definition 3.2 A general dual-feasible function f is maximal, if there is no other
general dual-feasible function g with f .x/  g.x/ for all x 2 R.
The simplest maximal general dual-feasible functions are the linear ones according
to the following proposition, but even these functions raise some exceptions.
Proposition 3.2 For every c 2 Œ0; 1, the linear function f W R ! R with f .x/ WD cx
is a maximal general dual-feasible function.
Proof The function f is a general dual-feasible function according to Definition 3.1,
Pn
because for any n 2 N n f0g and numbers x1 ; : : : ; xn 2 R with xi  1, it holds
iD1
that

X
n X
n
f .xi / D c  xi  c:
iD1 iD1

Suppose there is a general dual-feasible function g W R ! R with g.x/  cx for all

x 2 R, and g.y/ > cy for a certain y 2 R. Definition 3.1 implies g.y/ C g.y/  0.
Since g.y/  f .y/, the contradiction

0  g.y/ C g.y/ > cy  cy D 0

follows. Since f is not dominated by another general dual-feasible function, it is

maximal. u
t
The previous proposition shows that, in the case of general dual-feasible
functions, the symmetry rule

f .x/ C f .1  x/ D 1; for all x  1=2; (3.3)

3.2 Extension of Dual-Feasible Functions to General Domains 55

is no more a necessary condition for the function to be maximal. The conditions for
maximality are restated as follows.
Theorem 3.1 Let f W R ! R be a given function.
(a) If f satisfies the following conditions, then f is a general MDFF:
1. f .0/ D 0;
2. f is superadditive, i.e. for all x; y 2 R, it holds that

f .x C y/  f .x/ C f .y/I (3.4)

3. there is an " > 0, such that f .x/  0 for all x 2 .0; "/;
4. f obeys the symmetry rule (3.3);
(b) If f is a general MDFF, then the above properties (1.)–(3.) hold for f , but not
necessarily (4.);
(c) If f satisfies the above conditions (1.)–(3.), then f is monotonely increasing;
(d) If the symmetry rule (3.3) holds and f obeys the inequality (3.4) for all x; y 2 R
with x  y  1x 2 , then f is superadditive.

Unlike for classical dual-feasible functions, where the range was Œ0; 1, here the
nonnegativity of the function values for nonnegative arguments must explicitly be
demanded.
Example 3.3 The following example of a continuous and piecewise linear function
f W R ! R (Fig. 3.2) shows that Theorem 3.1 would become false, if the
prerequisite (3.) in part (a) was not considered.
8
ˆ
ˆ 5x; if x  0;
ˆ
ˆ
ˆ
< x; if 0  x  1=3;
f .x/ WD 5x  2; if 1=3  x  2=3;
ˆ
ˆ
ˆ
ˆ 2  x; if 2=3  x  1;
:̂
5x  4; if x  1:

The conditions (1.) and (4.) in Theorem 3.1 can easily be checked. To verify the
superadditivity, choose any x; y 2 R with

1x
xy
2

according to part (d) of Theorem 3.1. Hence, we have x  1=3 and

1Cx 2
xCy  :
2 3

Let d WD f .x C y/  f .x/  f .y/. A small case distinction is necessary to show

d  0.
56 3 General Dual-Feasible Functions

y
f

1x

Fig. 3.2 Necessity of the monotonicity

1. x > 0 yields f .x/ D x and 0 < y < 1=2. If y  1=3, then f .y/ D y and

f .x C y/  x  y;

and hence d  0. If y > 1=3, then f .y/ D 5y  2 and

f .x C y/ D 5x C 5y  2;

and hence d D 6x > 0.

2. x  0 yields f .x/ D 5x, and therefore

d D f .x C y/  f .y/  5x  0:

The details of the proof in this case are omitted. t

u

3.2.3 Extremality

The notion of extremality remains essentially the same as for classical dual-feasible
functions.
3.2 Extension of Dual-Feasible Functions to General Domains 57

Definition 3.3 A maximal general dual-feasible function f W R ! R is extreme, if

any maximal general dual-feasible functions g; h W R ! R with 2f .x/ D g.x/Ch.x/
for all x 2 R are necessarily identical to f .
Given that little is known about the set of maximal general dual-feasible
functions, analysing the extremality of maximal dual-feasible functions becomes
much more difficult, although it has been done for some functions as we show in the
following example.
Example 3.4 Given the constant c 2 f0; 1g, the linear function f W R ! R with

f .x/ WD cx

is an extreme maximal general dual-feasible function.

This is obvious for c D 0, because for any general dual-feasible functions g; h W
R ! R, it holds necessarily that g.x/  0 and h.x/  0 for x  0. If g and h are
maximal, then additionally g.x/  0 and h.x/  0 for all x  0 hold, such that
f g.
If c D 1, then let g; h W R ! R be any maximal general dual-feasible functions
with

g.x/ C h.x/ D 2x

for all x 2 R. The defining condition for dual-feasible functions implies that

g.1=q/  1=q and h.1=q/  1=q

for any q 2 N n f0g, and hence

g.1=q/ D 1=q:

The superadditivity yields

g. p=q /  p=q and h. p=q /  p=q

for all p; q 2 N n f0g, and hence g.x/ D x for all x 2 QC . The monotonicity implies
g.x/ D x for all x 2 RC . If there would be an x < 0 with g.x/ > x, then g would
not be a general dual-feasible function. Therefore, g.x/  x and finally

g.x/ D x D f .x/

for all x 2 R. t
u
The set of maximal dual-feasible functions is convex in the bounded case of
domain and range Œ0; 1, while in the generalized case, since the symmetry needs
not to hold, the set of maximal general dual-feasible functions is not convex, as
58 3 General Dual-Feasible Functions

counter examples show. Until now, the properties of this set have not been explored
in depth, and issues as if it is at least connected, way connected or even star shaped,
remain to be determined. A first result concerning the set of general dual-feasible
functions is stated in the following proposition.
Proposition 3.3 The set of general dual-feasible functions f W R ! R is closed, i.e.
any converging sequence of general dual-feasible functions converges to a general
dual-feasible function.
Proof Let I be any finite index set, and .fn / be a converging sequence of general
dual-feasible functions, i.e. for each x 2 R, the limit

f .x/ WD lim fn .x/

n!1

P
exists. For any xi 2 R (i 2 I) with xi  1, we have
i2I

X X X
f .xi / D lim fn .xi / D lim fn .xi /  1:
n!1 n!1
i2I i2I i2I

Therefore, f is a general dual-feasible function. t

u
Proposition 3.3 shows that if someone replaces a non-maximal general dual-feasible
function by a dominating one and repeats this process again and again, then a general
dual-feasible function will still be obtained.

3.3 Applications

In the previous chapters, we already mentioned that classical dual-feasible functions

generate solutions that are feasible for the dual of the continuous relaxation
of instances of the 1-dimensional cutting stock problem. General dual-feasible
functions do the same for the case where the sizes of the items and the variables
of the dual problem are unrestricted in sign. Negative sizes may occur, if a certain
fixed quantity of extra space in containers, which can be seen as items of negative
size, may be used in limited number.
Balance constraints can also lead to negative item sizes. Consider for example a
process assignment problem, where p processes have to be assigned to a minimum
number of identical machines, which are equipped with D units of RAM and C
units of CPU capacity each. Every process i has given demands ci in CPU and di in
RAM, respectively. After the assignment, the processors should be well-balanced,
i.e. CPU and RAM should be approximately equally exhausted, otherwise new
processes could not be assigned to the machine, such that the unused resource would
be wasted.
3.3 Applications 59

p
Column vectors a j 2 ZC represent configurations, i.e. processes scheduled for a
single processor. The balance constraint can be written as c> a  d> a  T, where
T > 0 is a threshold above which the machine is not considered well balanced.
Setting l WD T1 .c  d/, the constraint becomes l> a  1, where the elements of c  d
can be positive or negative.
p
This variant can be modelled similarly to the cutting-stock problem. Let b 2 RC
and l 2 R be some fixed vectors denoting respectively the demands and the sizes of
p
p
a set of p items. Each configuration may be represented by a column vector aj 2 ZC
and is feasible only if it obeys the capacity constraint l> aj  1. The patterns form
the matrix A. The continuous relaxation of this variant of the cutting stock problem
is given by

X
n
min xj s.to Ax D b; x 2 ZnC : (3.5)
jD1

If one item has a negative size, overproduction would make a trivial solution
with only one bin possible. Therefore, the demand constraints must be satisfied as
equalities, and the dual of (3.5) has variables unrestricted in sign

max b> u s.to u 2 Rp ; u> aj  1; j D 1; : : : ; n;

whose solutions can be obtained through general dual-feasible functions f W R ! R

by

ui WD f .`i /; i D 1; : : : ; p:

Example 3.5 Suppose that C D D D 21, and that seven processes with the
following resource consumption have to be scheduled:

ci 13 9 7 6 5 4 3
di 1 5 7 8 9 10 11

Let T WD 4. Besides the constraints that the processes do not overload the machines,
the balance constraint yields

. 12 4 0 2 4 6 8 /a  4:

Therefore, the first process must be combined with the last one, while the single
balance constraint does not restrict the other processes. One may get a solution with
four machines as illustrated in Fig. 3.3. The horizontal direction shows the CPU
usage, the vertical direction the memory consumption on the machines. Here, one
has a more-dimensional vector packing problem. In the continuous relaxation, one
could use the following patterns in the quantity 1=2:
60 3 General Dual-Feasible Functions

Fig. 3.3 Vector packing problem with a balance constraint

a1 D .1; 0; 0; 0; 0; 2; 0/>
a2 D .1; 0; 0; 0; 1; 0; 1/>
a3 D .0; 1; 0; 2; 0; 0; 0/>
a4 D .0; 1; 1; 0; 1; 0; 0/>
a5 D .0; 0; 1; 0; 0; 0; 1/>

where only the last pattern does not occupy the whole machine. Therefore, the
continuous relaxation would yield a lower bound of two and a half needed
machines. For this problem, lower bounds may be obtained by general dual-feasible
functions. t
u

3.4 Properties of Maximal General Dual-Feasible Functions

In this section, we explore the properties of maximal general dual-feasible functions.

This analysis implies frequently to prove superadditivity. Since these proofs are
usually not straightforward, we will resort to intermediary results to simplify
them. These results are recalled below in the form of lemmas. The first one is an
extension of Lemma 2.2. The second relies on the fact that showing nonnegativity
for nonnegative arguments is usually simpler than proving monotonicity.
Lemma 3.2 Let f W R ! R and b 2 R with b > 0 and f .0/ D 0 be given. If
f is convex on Œ0; b, then the superadditivity (3.4) holds for all x; y 2 Œ0; b with
x C y  b. If f is convex on Œb; 0, then the superadditivity (3.4) holds for all
x; y 2 Œb; 0 with x C y  b.
This lemma remains valid for b ! 1, i.e. if f is convex on RC and f .0/ D 0,
then f is superadditive for all nonnegative arguments. Nevertheless, such a function
will usually not be a maximal general dual-feasible function. Figure 3.4 shows two
real functions f ; g, which are strict convex on Œb; 0 and Œ0; b, respectively. Since
f .0/ D g.0/ D 0, both functions are superadditive in the specified intervals.
3.4 Properties of Maximal General Dual-Feasible Functions 61

y y g

−b x b x

Fig. 3.4 Strict convex functions

3.4.1 Structure

A first impression about the structure of maximal general dual-feasible functions is

given in the following. One will see that these functions are affine-linearly bounded.
The next proposition is motivated by the following functions f1 ; f2 ; f3 W R ! R,
which are not maximal general dual-feasible functions:
8 2 8 p
< x ; if x < 0; <  x; if x < 0;
f1 .x/ WD x; if 0  x  1; f2 .x/ WD x; if 0  x  1;
: 2 : p
x ; if x > 1; x; if x > 1;

.1 C ln 2/  x; if x < 0;
f3 .x/ WD
.1 C ln 2/  x  ln.1 C x/; otherwise:

The function f1 violates the superadditivity condition, because

f1 .2/ D 4 < 2 D 2  f1 .1/;

while f2 is obviously not a general dual-feasible function because of

5  1 C .4/  1;

such that the defining condition on general dual-feasible functions can be applied,
but

5  f2 .1/ C f2 .4/ D 5  2 > 1:

62 3 General Dual-Feasible Functions

Regarding f3 , one has f3 .0/ D 0 and f3 .1/ D 1. Moreover,

1
f30 .x/ D 1 C ln 2  > 0;
xC1

for x > 0, such that f3 is nondecreasing. The monotonicity of f30 shows that f3 is
strictly convex for positive arguments, and therefore superadditive in RC . It is easy
to see that the superadditivity of f3 holds on entire R. Therefore, f3 is a general dual-
feasible function fulfilling all the necessary conditions of Theorem 3.1, but it is not
maximal as the next proposition states.
Proposition 3.4 Let f W R ! R be a maximal general dual-feasible function and

t WD supf f .x/=x W x > 0g:

Then, we have

f .x/
lim D t  f .1/;
x!1 x
and for any x 2 R, it holds that

tx  maxf0; t  1g  f .x/  tx;

i.e. f is the sum of a linear and a bounded function.

C
Example 3.6 The function fBJ;1 considered in Example 3.1 yields t D bCc and

frac.C/ frac.C/
tx  fBJ;1 .x/  t  D :
C bCc

Here, the mapping x 7! tx  fBJ;1 .x/ is a periodic, piecewise linear function, as

illustrated in Fig. 3.5. u
t
The following lemma can be used to prove that a symmetric function f , bounded
by a maximal general dual-feasible function g for arguments less than 1=2, is
a maximal general dual-feasible function if it further satisfies some additional
conditions.
Lemma 3.3 Let f ; g W R ! R be two functions with f .1/ D 1, and both fulfilling
the conditions (3.) and (4.) of part (a) of Theorem 3.1 (p. 55). If g is a maximal
general dual-feasible function, and if

f .x/  g.x/ for all x < 1=2;

3.4 Properties of Maximal General Dual-Feasible Functions 63

y
tx

fBJ,1 ·; 65

−1 1 x

−1

Fig. 3.5 fBJ;1 compared to its asymptote

and

f .x C y/  f .x/ C f .y/ for all x; y 2 R with x C y < 1=2; (3.6)

then f is a maximal general dual-feasible function.

For classical symmetric dual-feasible functions g, it was enough to have f .x/  g.x/,
for a given function f W Œ0; 1 ! Œ0; 1 and for all x < 1=2, to conclude that
f is a maximal dual-feasible function. For general dual-feasible functions, the
situation changes. Nevertheless, Lemma 3.3 gives the possibility to demand the
superadditivity of the symmetric function f only for x C y < 1=2, instead of all
x; y with

1x
xy ;
2
as it would be necessary according to part (d) of Theorem 3.1. The following
example shows that the prerequisite on superadditivity must not be dropped entirely.
Example 3.7 Let g be the identity function and
8
< x3 ; if x  1;
f .x/ WD x; if 1  x  2;
:
.x  1/3 C 1; if x  2:

The functions f and g fulfill all prerequisites except the superadditivity condition.
One has f .1/ D 1, f .x/  g.x/ for all x  2, both f and g are symmetric, and g is a
64 3 General Dual-Feasible Functions

maximal general dual-feasible function. However, f violates the last prerequisite of

Lemma 3.3. Since f is not a general dual-feasible function, this example shows the
need of the restricted superadditivity (3.6). t
u

3.4.2 Behaviour at Given Points

Every classical maximal dual-feasible function f is symmetric, and therefore it

fulfills the conditions

f .1/ D 1 and f .1=2/ D 1=2:

Even if former proofs for the latter condition cannot be used for maximal general
dual-feasible functions, the two conditions hold for many maximal general dual-
feasible functions. Furthermore, they are not only necessary conditions, but they
allow also immediately the lower bound

f .x/  b2xc=2

for x > 1 for these functions due to the superadditivity. An insight concerning the
reasons why not all approaches for classical maximal dual-feasible functions work
is given in Exercise 3.
Proposition 3.5 If f W R ! R is a maximal general dual-feasible function and not
of the kind x 7! tx with 0  t < 1, then f .1/ D 1 and f .1=2/ D 1=2.
Proof Let t WD supff .x/=x W x > 0g  0. If t < 1, then Proposition 3.4 implies
f .x/ D tx for all x 2 R, such that the proof is complete. Otherwise t  1 and
f .x/  tx C 1  t for all x 2 R, and f .1/  1. Therefore, we have f .1/ D 1. Since f
is a maximal general dual-feasible function, it is nondecreasing and superadditive.
Suppose f .1=2/ < 1=2 (clearly, f .1=2/ > 1=2 is impossible). Define g W R ! R by

f .x/; if x ¤ 1=2;
g.x/ WD
1=2; otherwise:

Due to the assumption, g cannot be a general dual-feasible function, i.e. there are
P
n P
n
values n 2 N n f0g and x1 ; : : : ; xn 2 R with xi  1, but g.xi / > 1. Without
iD1 iD1
loss of generality, assume xi D 1=2 for i  k and xi ¤ 1=2 for i > k with a certain
k 2 N, k  n. That yields

X
n
xi  1  k=2
iDkC1
3.4 Properties of Maximal General Dual-Feasible Functions 65

and
X
n X
n
1< g.xi / D k=2 C f .xi /
iD1 iDkC1
!
X
n
 k=2 C f xi
iDkC1

 k=2 C f .1  k=2/;

where empty sums equal zero. Since f .1/ D 1, one gets further

1 < frac.k=2/ C bk=2c  1 C f .1  k=2/  frac.k=2/ C f .1  k=2 C bk=2c  1/

D frac.k=2/ C f .1  frac.k=2//:

If k is even, then the contradiction 1 < 0 C f .1/ arises. If k is odd, then

1 < 1=2 C f .1=2/

yields also a contradiction. t

u
The following proposition characterises the behaviour of a maximal general dual-
feasible function at the point zero, and it can help to prove or disprove that a given
real function is a maximal general dual-feasible function.
Proposition 3.6 If f W R ! R is a maximal general dual-feasible function, then,
for any y 2 R, the left and right limits lim f .x/ and lim f .x/ exist, and it holds that
x"y x#y

 
lim f .x/ D inf lim f .x/  lim f .x/ :
x"0 y2R x"y x#y

The idea is that if f has a gap at the point y, then the superadditivity requires for
negative x
0 that f .x/ is small enough. Meanwhile, f .x/ must also not be too
small, otherwise f will not be maximal.
Example 3.8 Define the function f W R ! R by
8
< b  b2xc; if x < 1=2;
f .x/ WD 1=2; if x D 1=2; (3.7)
:
1  b  b2  2xc; if x > 1=2;

which is a maximal general dual-feasible function for every b  1, as can be proved

by Theorem 3.1. This discontinuous function has as its largest gap b, namely at
points y, where 2y 2 Z. That fits to

lim f .x/ D b  .1/ D b: t

u
x"0
66 3 General Dual-Feasible Functions

3.4.3 Limits of Possible Convexity

The following proposition demonstrates the limits of possible convexity for maxi-
mal general dual-feasible functions.
Proposition 3.7 If f W R ! R is a maximal general dual-feasible function, then
it cannot be strict convex in an environment of the point zero and also not concave
(except linear) in an interval Œ0; b or Œb; 0 with b > 0.
Example 3.9 Let b WD 1=2 and

f .x/ WD 4x3

for x 2 Œb; b. Then, f is strictly concave in the interval Œb; 0, and hence it is not
superadditive. For instance, we have that

f .1=4/ D 1=16;

and also

f .1=2/ D 1=2 < 2  f .1=4/:

t
u

3.4.4 Composition and Convex Combinations

While the composition of general dual-feasible functions yields a general dual-

feasible function, maximality may be lost in this process as shown in the following
proposition. The same happens with the convex combination of functions.
Proposition 3.8 The composition or convex combination of maximal general dual-
feasible functions is not necessarily a maximal general dual-feasible function.
Proof Let b; c 2 R be any constants with 0 < c < 1 and bc  1. The function
f W R ! R defined as f .x/ WD cx is a maximal general dual-feasible function (see
Proposition 3.2). Let g be the function (3.7). That yields the composition f .g.// D
c  g./, i.e.
8
< bc  b2xc; if x < 1=2;
f .g.x// D c=2; if x D 1=2;
:
c  bc  b2  2xc; if x > 1=2;
3.5 Examples 67

8
< bc  b2xc; if x < 1=2;
 1=2; if x D 1=2;
:
1  bc  b2  2xc; if x > 1=2;

and for x  1=2 the inequality is strict. The latter function is of the same type as
the function (3.7) with parameter bc  1, and hence it is a maximal general dual-
feasible function. The function f .g.// can also be seen as a convex combination of
the constant zero-function f 0, which is a maximal general dual-feasible function
according to Proposition 3.2, with factor 1  c and g with factor c. t
u

3.5 Examples

As already mentioned, many classical dual-feasible functions cannot be extended

easily to general dual-feasible functions or at least not with the same formula. Such
an example is given next.
Example 3.10 The function fCCM;1 .I C/ already described in Chap. 2 (see p. 28) is
defined for any real parameter C  1 as follows:
8
<
bCxc=bCc; if x < 1=2;
fCCM;1 .xI C/ D 1=2; if x D 1=2;
:
1  bC  Cxc=bCc; if x > 1=2:

This function was a maximal dual-feasible function for the domain Œ0; 1. However,
for domain R, it is not even a general dual-feasible function. To see this, let
 
1
n WD C1
1  frac.C/

and choose an enough small " > 0. One gets

1
fCCM;1 .1 C "/ D 1 C
bCc

and
 
1  n  n  bCc 1n
fCCM;1 D  n:
C bCc

Hence,
 
1  n  n  bCc 1
n  fCCM;1 .1 C "/ C fCCM;1 D > 0:
C bCc
68 3 General Dual-Feasible Functions

Since

1  n  n  bCc 1
n  .1 C "/ C D n" C  .Cn C 1  n  n  bCc/
C C
1  n  .1  frac.C//
D n" C
C
1  1 C frac.C/  1
 n" C
C
<0

for sufficiently small ", a contradiction to the point (2.) of Proposition 3.1 arises. u
t
On the other hand, we explore in the sequel a different case where the defining
formulation of a classical maximal dual-feasible function leads without any change
to a maximal general dual-feasible function. That happens in particular with the
function fDG;1 discussed in Chap. 2 (see (2.14) at page 43).
Proposition 3.9 For any C 2 R n N, C > 1 and k 2 N with k  frac1.C/ , the
following function fDG;1 W R ! R is a maximal general dual-feasible function:
8 frac.Cx/frac.C/ .Cx/frac.C/
ˆ
< 1frac .C/
; if .k  1/  frac1 frac .C/
2 N;
bCxc 1
fDG;1 .x/ WD C  n o
bCc bCc :̂ .Cx/frac.C/
max 0; d.k  1/  frac1 frac .C/
e=k ; otherwise.

Proof Define the auxiliary function h W R ! R as


x; if .k  1/  x 2 N;
h.x/ WD (3.8)
maxf0; d.k  1/  xe=kg; otherwise.

Then, we have
 
frac.Cx/frac.C/
bCxc C h 1frac.C/
fDG;1 .x/ D :
bCc

First, some properties of h are derived, which will be used later to prove the
sufficient conditions of Theorem 3.1. Clearly, one obtains h.x/ D 0 for x  0
and h.1/ D 1. Moreover, h rises monotonely in the closed interval Œ0; 1, because h
is piecewise constant and one gets for any p 2 f1; : : : ; k  1g the estimations

p p  p  pC1
limp h.x/ D < Dh  D limp h.x/:
x" k1 k k1 k1 k x# k1
3.5 Examples 69

Additionally, we have

d.k  1/xe b.k  1/xc C 1

 h.x/  ; for all x 2 Œ0; 1; (3.9)
k k

due to the definition of h, because either h.x/ D d.k  1/xe=k or h.x/ D x. In the
latter case, it holds that d.k  1/xe D .k  1/x, such that both inequalities become
equivalent to x  0 or x  1, respectively. The function h is also symmetric inside
the interval Œ0; 1, i.e.

h.x/ C h.1  x/ D 1; for all x 2 Œ0; 1; (3.10)

because

.k  1/  x 2 N ” .k  1/  .1  x/ 2 N;

and if .k  1/  x … N, then

h.x/ C h.1  x/ D d.k  1/  xe=k C d.k  1/  .1  x/e=k

D d.k  1/  x C .k  1/  .1  x/ C 1e=k
D 1:

Clearly, the function fDG;1 fulfills the conditions (1.) and (3.) of part (a) of
Theorem 3.1. To show the symmetry, choose any x 2 R. It must be verified that
   
frac.Cx/  frac.C/ frac.C  Cx/  frac.C/
bCxcCbCCxcCh Ch D bCc:
1  frac.C/ 1  frac.C/

This is obvious for frac.Cx/ D frac.C/. If frac.Cx/ < frac.C/, then

0 < frac.C  Cx/  frac.C/;

and hence

bCxc C bC  Cxc
fDG;1 .x/ C fDG;1 .1  x/ D
bCc
D 1:

If frac.Cx/ > frac.C/, then

bCxc C bC  Cxc D bCc  1

70 3 General Dual-Feasible Functions

and
   
frac.C  Cx/  frac.C/ frac.C/  frac.Cx/ C 1  frac.C/
h Dh
1  frac.C/ 1  frac.C/
 
frac.Cx/  frac.C/
D h 1
1  frac.C/
 
frac.Cx/  frac.C/
D 1h ;
1  frac.C/

such that fDG;1 is symmetric also in this case.

Showing the superadditivity requires for any x; y 2 R to verify that

d WD bCc  .fDG;1 .x C y/  fDG;1 .x/  fDG;1 .y//

is not negative. For this purpose, let

 
frac.Cx/  frac.C/
I1 WD x 2 R W .k  1/  2N :
1  frac.C/
Without loss of generality, suppose that frac.Cx/  frac.Cy/. Four cases arise:
1. If frac.Cy/  frac.C/, then

d  bCx C Cyc  bCxc  bCyc

 0;

because the rounding-down is superadditive.

2. The case where frac.Cx/  frac.C/ < frac.Cy/ yields two subcases. If
frac.Cx/ C frac.Cy/ < 1, then

frac.Cx C Cy/  frac.C/ frac.Cy/  frac.C/

d D h. /  h. /
1  frac.C/ 1  frac.C/
 0;

because of the monotonicity of h in the interval .0; 1/. If frac.Cx/Cfrac.Cy/  1,

then
 
frac.Cy/  frac.C/
d  1h
1  frac.C/
> 0:

3. The case frac.Cx/ > frac.C/ and frac.Cx/ C frac.Cy/ < 1 can only happen if
frac.C/ < 1=2. One gets

bCx C Cyc D bCxc C bCyc;

3.5 Examples 71

and hence
   
frac.Cx/ C frac.Cy/  frac.C/ frac.Cx/  frac.C/
dDh h
1  frac.C/ 1  frac.C/
 
frac.Cy/  frac.C/
h :
1  frac.C/

The estimation (3.9) yields

 
frac.Cx/ C frac.Cy/  frac.C/
dk  .k  1/ 
1  frac.C/

frac.Cx/  frac.C/
 .k  1/ 
1  frac.C/

frac.Cy/  frac.C/
 1  .k  1/  1 (3.11)
1  frac.C/
k1
  .frac.Cx/ C frac.Cy/  frac.C/  frac.Cx/ C frac.C/
1  frac.C/
 frac.Cy/ C frac.C//  2
k  frac.C/  frac.C/
D 2
1  frac.C/
1  frac.C/ 1
 2 (because of k  )
1  frac.C/ frac.C/

D 1:

The right-hand side of (3.11) is integer. It may be equal to 1 only, if the

rounding-brackets had no influence. If x … I1 or y … I1 , then the rounding changed
the values, such that the right-hand side of (3.11) is at least zero, and hence d  0.
If x; y 2 I1 , then
   
frac.Cx/  frac.C/ frac.Cy/  frac.C/
h Ch
1  frac.C/ 1  frac.C/
frac.Cx/ C frac.Cy/  2frac.C/
D
1  frac.C/
 
frac.Cx C Cy/  2frac.C/
Dh
1  frac.C/
 
frac.Cx C Cy/  frac.C/
h
1  frac.C/

due to the monotonicity of h inside the interval Œ0; 1, and hence, we have again
d  0.
72 3 General Dual-Feasible Functions

4. When frac.Cx/ > frac.C/ and frac.Cx/ C frac.Cy/  1, we get

   
frac.Cx C Cy/  frac.C/ frac.Cx/  frac.C/
d D bCx C Cyc C h  bCxc  h
1  frac.C/ 1  frac.C/
 
frac.Cy/  frac.C/
 bCyc  h
1  frac.C/
   
frac.Cx/ C frac.Cy/  1  frac.C/ frac.Cx/  frac.C/
D 1Ch h
1  frac.C/ 1  frac.C/
 
frac.Cy/  frac.C/
h :
1  frac.C/

If frac.Cx/frac.C/
1frac.C/
C frac.Cy/frac.C/
1frac.C/
 1, then the monotonicity of h inside the
interval Œ0; 1 and (3.10) yield
   
frac.Cx/  frac.C/ frac.Cy/  frac.C/
h Ch  1;
1  frac.C/ 1  frac.C/

such that
 
frac.Cx/ C frac.Cy/  1  frac.C/
dh
1  frac.C/
 0:

Therefore assume for the rest of the proof that

frac.Cx/  frac.C/ frac.Cy/  frac.C/

C > 1:
1  frac.C/ 1  frac.C/

One has
   
frac.Cx/ C frac.Cy/  2frac.C/ frac.Cx/  frac.C/
d D 1Ch 1 h
1  frac.C/ 1  frac.C/
 
frac.Cy/  frac.C/
h :
1  frac.C/

Analogously to the previous case, (3.9) yields

 
frac.Cx/ C frac.Cy/  2frac.C/
dk  k C .k  1/  kC1
1  frac.C/
 
frac.Cx/  frac.C/ frac.Cy/  frac.C/
 .k  1/   1  .k  1/  1
1  frac.C/ 1  frac.C/
(3.12)
3.6 Building Maximal General Dual-Feasible Functions 73

k1
  .frac.Cx/ C frac.Cy/  2frac.C/  frac.Cx/ C frac.C/
1  frac.C/
 frac.Cy/ C frac.C//  1
D 1;

and the right-hand side in (3.12) is integer. If it equals 1, then the rounding
brackets will not change anything, and hence x; y 2 I1 . In this case, one gets
 
frac.Cx/  frac.C/
h C
1  frac.C/
 
frac.Cy/  frac.C/ frac.Cx/  frac.C/ frac.Cy/  frac.C/
Ch 1 D C 1
1  frac.C/ 1  frac.C/ 1  frac.C/
frac.Cx/ C frac.Cy/  2frac.C/
D 1
1  frac.C/
 
frac.Cx/ C frac.Cy/  2frac.C/
Dh 1 ;
1  frac.C/

such that again d  0.

Since in all cases it holds that d  0, the proof is complete. t
u

3.6 Building Maximal General Dual-Feasible Functions

In this section, we explore three different methods to build maximal general dual-
feasible functions by extending a given classical dual-feasible function to domain
and range R. The second and third approach simply use affine-linear expressions
outside the interval Œ0; 1 where the extended function has the form

x 7! tx C c;

with t and c being constants, and t a sufficiently large value. The additive constants
are generally different for x < 0 and x > 1.

3.6.1 Method I

Classical maximal dual-feasible functions can be extended to domain and range R

by considering the sum of a periodic with a monotone staircase function whose
average slope must be large enough. The feasible lower bound of this average slope
can be expressed in different ways. The classical maximal dual-feasible function is
74 3 General Dual-Feasible Functions

applied to the non-integer part of the given real argument, while a suitable multiple
of the integer part of the argument is added. These ideas are stated formally in the
following proposition.
Proposition 3.10 Let g W Œ0; 1 ! Œ0; 1 be a maximal dual-feasible function,

h.x; y/ WD g.x C y/  g.x/  g.y/ with x; y 2 Œ0; 1 and x C y  1; (3.13)

and

b0 WD supfg.x/ C g.y/  g.x C y  1/ j x; y 2 Œ0; 1 ^ x C y  1g:

Then, we have that

b0 D 1 C supfh.x; y/ W 0 < x  y < 1=2 and x C y  2=3g

and

1  b0  2:

For b  b0 , g can be extended to a maximal general dual-feasible function f W R !

R as follows

g.frac.x// C b  bxc; if x < 1;
f .x/ WD
1  f .1  x/; otherwise:

For b > b0 , f becomes a non-trivial convex combination of maximal general

dual-feasible functions. In that case, f is clearly not extreme.
To see this, choose any  2 .0; b  b0  and set b1 WD b   and b2 WD b C . We
may use the proposition with b1 and b2 to get the non-identical functions f ; f1 ; f2 .
One obtains for any x < 1 that

2f .x/  f1 .x/  f2 .x/ D 2  .g.frac.x// C b  bxc/  g.frac.x//  b1  bxc

 g.frac.x//  b2  bxc
D bxc  .2b  b1  b2 /
D 0:

If x  1, then the symmetry leads to the same result.

Example 3.11 For any k 2 N n f0g, the function fFS;1 .I k/ W Œ0; 1 ! Œ0; 1 (already
described in Chap. 2, see (2.10), page 35) is a classical maximal dual feasible
function, which is defined as follows

x; if .k C 1/  x 2 N;
fFS;1 .xI k/ D (3.14)
b.k C 1/  xc=k; otherwise:
3.6 Building Maximal General Dual-Feasible Functions 75

Let g be this function fFS;1 in Proposition 3.10. Formula (3.13) yields

     
1 1 2 1
h  ";  " D fFS;1  2"  2fFS;1 "
kC1 kC1 kC1 kC1
1
D
k
for enough small " > 0. For the chosen function g, greater values of h are not
possible, i.e.

max fh.x; y/ j 0  x; y  1 and x C y  1g D 1=k:

kC1
Therefore, one gets b0 D k
, and according to Proposition 3.10 with b WD b0 the
function f W R ! R with
8
< frac.x/ C kC1
k
 bxc; if .k C 1/  frac.x/ 2 N ^ x < 1;
f .x/ WD b.kC1/frac.x/c
C kC1  bxc; if .k C 1/  frac.x/ … N ^ x < 1;
: k k
1  f .1  x/; if x  1;

is a maximal general dual-feasible function. This function is illustrated in Fig. 3.6

for k 2 f1; 2g. u
t

y f y f

1 1

1x 1 x

Fig. 3.6 Extending fFS;1 .I k/ to a general MDFF f for k 2 f1; 2g by Proposition 3.10
76 3 General Dual-Feasible Functions

3.6.2 Method II

The next proposition allows the extension of a Lipschitz-continuous maximal dual-

feasible function g W Œ0; 1 ! Œ0; 1 to the domain and range R by using affine-linear
terms for the arguments that are outside Œ0; 1. A function f W X  R ! R is
Lipschitz-continuous, if there is a constant L  0 with

j f .x/  f .y/j  L  jx  yj for all x; y 2 X:

Lipschitz-continuity is a strong property that can be generalized to higher-

dimensional Euclidean spaces without difficulty. Let X  Rn be an open set,
n 2 N. According to Rademacher’s theorem, every Lipschitz-continuous function
f W X ! Rm , m 2 N, is almost everywhere totally differentiable. Furthermore,
a differentiable function is Lipschitz-continuous if and only if its derivative is
bounded. Moreover, we have the following implications, where the converse is
false:
– every Lipschitz-continuous function f is Hölder-continuous, i.e. there are con-
stants c > 0 and ˛ 2 .0; 1 with

jf .x/  f .y/j  c  jx  yj˛ for all x; yI

– every Hölder-continuous function f is uniformly continuous, i.e., for all " > 0,
there is a ı > 0 depending on " only, such that jx  yj < ı implies

jf .x/  f .y/j < "I

– every uniform continuous function is also continuous in the usual sense, where ı
may depend on x and y. Here, the converse holds under an additional prerequisite.
Every continuous function on a compact set is uniformly bounded according to
Heine’s theorem.
Examples of Lipschitz-continuous
p functions are x 7! jxj and x 7! ejxj on a bounded
set, while the function x 7! maxf0; xg, for example, is not Lipschitz-continuous.
Proposition 3.11 Let p; t 2 R and g W Œ0; 1 ! Œ0; 1 be a maximal dual-feasible
function with

jg.x/  g.y/j  t  jx  yj
3.6 Building Maximal General Dual-Feasible Functions 77

for all x; y 2 Œ0; 1, i.e. the function g is Lipschitz-continuous with L WD t. Then, we
have t  1, and, for 1  p  t, the following function f W R ! R is a maximal
general dual-feasible function:
8
< tx C 1  p; if x < 0;
f .x/ WD tx C p  t; if x > 1;
:
g.x/; otherwise:

Exercise 8 is an example that weakening the Lipschitz-continuity to Hölder-

continuity can cause the loss of superadditivity.
Example 3.12 The following function g W Œ0; 1 ! Œ0; 1 is symmetric, strict convex
in Œ0; 12  and a Lipschitz-continuous maximal dual-feasible function:

2x2 ; if 0  x  12 ;
g.x/ WD
1  g.1  x/; otherwise:

To calculate the smallest valid Lipschitz-constant for g, the largest slope is needed.
Since g is differentiable in .0; 1/, the supremum of the derivative is sought. One gets

g0 .x/ D 4x

inside the interval .0; 12 /. Because of the symmetry of g, it holds that

g0 .1  x/ D g0 .x/;

such that g0 is continuous at 1=2, and hence it is enough to analyze g (or 0

g ) for
1 0 1
x  2 only. The derivative g takes its maximum at x D 1=2 with g 2 D 2.
0

Hence, the smallest valid Lipschitz-constant is 2. The extension of g to a maximal

general dual-feasible function f W R ! R for t D 2 and p 2 f1; tg according to
Proposition 3.11 is illustrated in Fig. 3.7.
t
u

3.6.3 Method III

Most of the classical maximal dual-feasible functions are not Lipschitz-continuous,

such that Proposition 3.11 cannot be used in these cases. In this section, we
introduce a similar proposition that does not need this strong prerequisite. Some
generality is lost in the construction, because instead of two parameters only one
will apply. Here again, affine-linear expressions are used outside the interval Œ0; 1
to extend a given classical maximal dual-feasible function to a general one.
78 3 General Dual-Feasible Functions

y y

f f

1 1

1 x 1 x

Fig. 3.7 Application of Proposition 3.11 to g with t D 2 and p 2 f1; tg

Proposition 3.12 Let t 2 R and g W Œ0; 1 ! Œ0; 1 be a maximal dual-feasible

function with g.x/  tx for all x 2 Œ0; 1. Then, we have t  1, and the following
function f W R ! R is a maximal general dual-feasible function:
8
< tx C 1  t; if x < 0;
f .x/ WD tx; if x > 1;
:
g.x/; otherwise:

Example 3.13 Let us again use the function (3.14) as example. One gets for p 2 N
and p  k that
p
limp fFS;1 .x/ D :
x# kC1 k

Hence, the smallest possible t for Proposition 3.12 is

kC1
tD ;
k
3.6 Building Maximal General Dual-Feasible Functions 79

because if .k C 1/  x 2 N for a certain x 2 Œ0; 1, then fFS;1 .x/ D x. With these
choices, one obtains the following maximal general dual-feasible function f W R !
R:
8
ˆ
ˆ tx C 1  t; if x < 0;
<
tx; if x > 1;
f .x/ WD
ˆ x; if .k C 1/  x 2 f0; 1; : : : ; k C 1g;
:̂
b.k C 1/  xc=k; otherwise:

Figure 3.8 shows this function for k 2 f1; 2g.

t
u

y y
f
f
2 2

1 1

1 x 1 x

Fig. 3.8 Extending fFS;1 .I k/ to a general MDFF f for k 2 f1; 2g by Proposition 3.12
80 3 General Dual-Feasible Functions

3.6.4 Examples
3.6.4.1 Based on Method I

Proposition 3.10 provides an approach to construct maximal general dual-feasible

functions from classical ones. Since the use of b > b0 never yields extreme maximal
dual-feasible functions, one should know the mentioned parameter b0 , whenever it
is possible.
If one is given a function g W Œ0; 1 ! Œ0; 1 and shall check if g is a maximal dual-
feasible function, particularly if g is superadditive, then often the function (3.13)
(p. 74) must be tested for nonnegativity, sometimes with long or complex case
distinctions. While in this analysis the infimum of h is explored, one can often also
find the supremum without much more effort. Hence, the value of b0 is found almost
directly as a side effect of the superadditivity proof.
As an example, we will explore the function fDG;1 with the parameters C 2 RnN,
1
C > 1 and k 2 N and k  frac.C/ , such that k  2. However, the restrictions
x; y 2 Œ0; 1 and x C y  1 are dropped, because they would not simplify
the analysis and fDG;1 is a maximal general dual-feasible function, such that all
expressions will remain defined. The only negative effect of this relaxation might
be an overestimation of b0 . To avoid a confusion with the proof of Proposition 3.9
(p. 68), the identifiers h and d of that proof are kept, and hence h refers here to the
auxiliary function (3.8) and not to the function (3.13).
Without loss of generality, we assume that frac.Cx/  frac.Cy/. The first case
in the proof of Proposition 3.9 was frac.Cy/  frac.C/, yielding
 
frac.Cx C Cy/  frac.C/
d D bCx C Cyc  bCxc  bCyc C h :
1  frac.C/

If frac.Cx/ C frac.Cy/ < 1, then

 
frac.Cx C Cy/  frac.C/
dDh < 1;
1  frac.C/

because the argument of h is below 1. If frac.Cx/ C frac.Cy/  1, then

frac.Cx C Cy/  frac.C/ D frac.Cx/ C frac.Cy/  1  frac.C/ < 0

and d D 1. This latter situation requires frac.C/  12 . Hence, if frac.C/  12 , then

setting x WD y WD frac.C/
C will yield d D 1. Since x; y 2 Œ0; 1 and

2frac.C/ 2frac.C/
xCyD D < 1;
C bCc C frac.C/
3.6 Building Maximal General Dual-Feasible Functions 81

1
the value d D 1 is not an overestimation. It yields b0 D 1 C bCc for the case
1
frac.C/  2 .
Assume in the following that frac.C/ < 12 . In that case, we have k  3. We
want to explore the maximal possible value of d under this condition. The following
further cases arise:
1. frac.Cx/  frac.C/ < frac.Cy/ yields two subcases.
If frac.Cx/ C frac.Cy/ < 1, then
   
frac.Cx C Cy/  frac.C/ frac.Cy/  frac.C/
dDh h
1  frac.C/ 1  frac.C/
   
frac.Cy/ frac.Cy/  frac.C/
h h
1  frac.C/ 1  frac.C/
   
frac.Cy/ frac.Cy/  frac.C/
 .k  1/  C 1  .k  1/  =k
1  frac.C/ 1  frac.C/
 
frac.C/
 1 C .k  1/  =k
1  frac.C/
<1

due to the monotonicity of h, the inequalities (3.9) and frac.C/ < 12 .

If frac.Cx/ C frac.Cy/  1, then
   
frac.Cx/ C frac.Cy/  1  frac.C/ frac.Cy/  frac.C/
d D1Ch h :
1  frac.C/ 1  frac.C/

Since frac.Cx/  frac.C/, it follows that

frac.Cx/ C frac.Cy/  1  frac.C/  frac.Cy/  1 < 0;

and hence
 
frac.Cx/ C frac.Cy/  1  frac.C/
h D 0:
1  frac.C/

Therefore, in this case d becomes maximal, if frac.Cy/ is minimized, i.e. for

frac.Cy/ D 1  frac.C/, yielding
 
frac.C/
d D 1h 1
1  frac.C/
 
frac.C/
Dh
1  frac.C/
1
 1 :
k
82 3 General Dual-Feasible Functions

2. The case where frac.Cx/ > frac.C/ and frac.Cx/ C frac.Cy/ < 1 yields
   
frac.Cx/ C frac.Cy/  frac.C/ frac.Cx/  frac.C/
dDh h
1  frac.C/ 1  frac.C/
 
frac.Cy/  frac.C/
h :
1  frac.C/

Hence,

frac.Cx/ C frac.Cy/  frac.C/
dk  .k  1/  C1
1  frac.C/
   
frac.Cx/  frac.C/ frac.Cy/  frac.C/
 .k  1/   .k  1/ 
1  frac.C/ 1  frac.C/
k1
  .frac.Cx/ C frac.Cy/  frac.C/  frac.Cx/ C frac.C/
1  frac.C/
 frac.Cy/ C frac.C// C 1
frac.C/
D .k  1/  C 1;
1  frac.C/

and therefore again d  1  1k .

3. The case where frac.Cx/ > frac.C/ and frac.Cx/ C frac.Cy/  1 yields
   
frac.Cx/ C frac.Cy/  1  frac.C/ frac.Cx/  frac.C/
d D 1Ch h
1  frac.C/ 1  frac.C/
 
frac.Cy/  frac.C/
h :
1  frac.C/

If frac.Cx/ C frac.Cy/ > 1 C frac.C/, then


frac.Cx/ C frac.Cy/  1  frac.C/
dk  k C .k  1/  C1
1  frac.C/
   
frac.Cx/  frac.C/ frac.Cy/  frac.C/
 .k  1/   .k  1/ 
1  frac.C/ 1  frac.C/
k1
 kC1C  .frac.Cx/ C frac.Cy/  1  frac.C/  frac.Cx/
1  frac.C/
C frac.C/  frac.Cy/ C frac.C//
D k C 1  .k  1/
D 2;
3.6 Building Maximal General Dual-Feasible Functions 83

and hence d  2k .
If frac.Cx/ C frac.Cy/  1 C frac.C/, then
 
frac.Cx/ C frac.Cy/  1  frac.C/
h D 0;
1  frac.C/

and hence
   
frac.Cx/  frac.C/ frac.Cy/  frac.C/
d D 1h h
1  frac.C/ 1  frac.C/
k2
 :
k
For the sake of shortness, further details are omitted.
The final result
1
b0 D 1 C
bCc

for fDG;1 .I C; k/ with frac.C/  12 , and a smaller b0 in the case 0 < frac.C/ < 12 can
be compared with the according b0 for the function fBJ;1 .I C/, which was defined
as a classical maximal dual-feasible function in Chap. 2 (see (2.4) at page 25). The
latter is also a general maximal dual-feasible function as described in Example 3.1
on page 52. One gets similarly for this function the result
 
frac.C/
b0 D 1 C min 1; =bCc;
1  frac.C/

1
which is in the case frac.C/  2
the same value as for fDG;1 .

3.6.4.2 Based on Method II

Proposition 3.11 requires a Lipschitz-continuous maximal dual-feasible function.

This prerequisite is fulfilled by fBJ;1 .I C/ for all parameter values C  1. This
function is one of the strongest known classical dual-feasible functions. Its largest
slope, namely

C
t1 WD ;
bCc  .1  frac.C//

1 frac.Cx/
is the smallest valid Lipschitz-constant. That can be seen in the term bCc  1frac.C/
in the definition of fBJ;1 in the intervals with slope, i.e. where frac.Cx/ > frac.C/.
The additional additive terms in the definition of the function do not influence the
84 3 General Dual-Feasible Functions

y y
f
f

1 1

−1 1 x −1 1 x

−1 −1

Fig. 3.9 Function f obtained from fBJ;1 according to Proposition 3.11 for p D 1 and p D t

slope. Hence, for any t  t1 and p 2 Œ1; t, a general maximal dual-feasible function
f W R ! R can be obtained by Proposition 3.11, namely

8
ˆ
< tx C 1  p; if x < 0;
f .x/ WD  ntx C p  t;o if x > 1;
:̂ bCxc C max 0I frac.Cx/frac.C/ =bCc; otherwise:
1frac.C/

Example 3.14 Figure 3.9 shows the resulting functions, when Proposition 3.11 is
used with fBJ;1 .I 65 /, t WD t1 D 32 and p 2 f1; tg.
In the left case, one has for all x > 0 the strict inequality f .x/ < tx with

f .x/ f .x/
t D sup D lim ;
x>0 x x!1 x

while in the right case it holds that

t < f .1/

(see also Proposition 3.4, p. 62).

t
u
3.6 Building Maximal General Dual-Feasible Functions 85

3.6.4.3 Based on Method III

Proposition 3.12 can be used for every maximal dual-feasible function g W Œ0; 1 !
Œ0; 1 due to the weaker prerequisites. It is only necessary to choose

t  t0 WD supfg.x/=x W 0 < x < 1g:

The value of t0 is calculated in the sequel for fDG;1 .I C; k/ with C 2 R n N, C > 1
1
and k 2 N with k  frac.C/ , such that k  2.
Since fDG;1 is a staircase function, only the right limits

fDG;1 .y/
lim
y#x y

at discontinuities may play a role. Discontinuities arise if Cx 2 N n f0g or

frac.Cx/  frac.C/
.k  1/  2 N:
1  frac.C/

Let a WD bCxc and

frac.Cx/  frac.C/
b WD .k  1/  :
1  frac.C/

In the first case, one gets

fDG;1 .x/ D a=bCc and x D a=C;

such that
C
t : (3.15)
bCc

The second case is Cx … N and b 2 N, yielding

 
bC1
lim fDG;1 .y/ D a C =bCc;
y#x k

b 2 f0; : : : ; k  2g, a < Cx < a C 1 and

k1
bD  .Cx  a  frac.C//:
1  frac.C/
86 3 General Dual-Feasible Functions

Hence, we have
 
b  .1  frac.C//
xD C a C frac.C/ =C
k1

and

.a C bC1
/=bCc
t  k
 : (3.16)
b.1frac.C//
k1 C a C frac.C/ =C

Between the two lower bounds (3.15) and (3.16) for t, we keep the largest one. Since
frac.C/  k  1 and b  k  2, the following equivalent inequalities are obtained:

frac.C/  .k2  k  bk/  k  b  1

bk  bk  frac.C/ C frac.C/  .k2  k/  .k  1/  .b C 1/
b  .1  frac.C// bC1
C frac.C/ 
k1 k
b  .1  frac.C// bC1
C a C frac.C/  a C
k1 k
bC1
aC
1 b.1frac.C//
k

k1
C a C frac.C/

Finally, the bound (3.15) is at least as large as (3.16), such that the result is

C
t0 D :
bCc

The same result holds forlfBJ;1 . The

m application of Proposition 3.12 to fDG;1 and
fBJ;1 for C D 12 5
(and k WD 1
frac.C/
D 3 in the case of fDG;1 ) with t D t0 D 6
5
is
illustrated in Fig. 3.10.

3.7 Related Literature

The extension of dual-feasible functions to more general domains is recent. The first
contributions in this field are due to Rietz et al. (2012b); Rietz et al (2014); Rietz
et al. (2015). In Rietz et al. (2012b), the authors discussed the first results of the
extension of dual-feasible functions to the domain of real numbers. In Rietz et al
(2014), they explored the properties of these general dual-feasible functions. The
proof of Lemma 3.2 can be found in this paper, while Lemma 3.3 is an extension
of Lemma 3 of Rietz et al. (2010) to domain and range R. The proofs of the other
3.8 Exercises 87

y y
f f

1 1

1 x 1 x

12 12 6
Fig. 3.10 Applying Proposition 3.12 to fDG;1 .I 5
; 3/ and fBJ;1 .I 5
/ for t D 5

properties described in Sect. 3.4 can also be found in Rietz et al (2014). The methods
for constructing maximal general dual-feasible functions were first discussed in
Rietz et al. (2015).

3.8 Exercises

1. Prove the validity of the necessary conditions of Proposition 3.1.

2. Let f W R ! R be a general dual-feasible function with f .x0 / > 0 for a certain
x0 2 R. Show that f .x/ < 0 for all x < 0.
3. Find the error in the following reasoning.
Suppose that someone states (falsely) that every maximal general dual-feasible
function f W R ! R obeys the symmetry rule (3.3):
“The symmetry (3.3) holds, if and only if f .x/ C f .1  x/ D 1 for all x. Due to the
defining condition on dual-feasible functions, we have f .x/ C f .1  x/  1, for all x.
To show f .x/ C f .1  x/  1, for all x ¤ 1=2, we define the function h W R ! R as

f .x/; if x ¤ x1 ;
h.x/ WD
1  f .1  x1 /; otherwise:

Suppose there is an x1 > 1=2 with f .x1 / C f .1  x1 / < 1. Then, h must hurt the
defining condition on dual-feasible functions, since f is a maximal dual-feasible
function and h.x/  f .x/, for all x. That requires that there are n 2 N and
x2 ; : : : ; xn 2 R with

X
n X
n
xi  1 and h.xi / > 1;
iD1 iD1
88 3 General Dual-Feasible Functions

i.e.

X
n
f .xi / > f .1  x1 /;
iD2

but f is superadditive and monotone, implying

X
n X
n
f .xi /  f . xi /  f .1  x1 /:
iD2 iD2

That contradiction proves that the assumed x1 > 1=2 with f .x1 / C f .1  x1 / < 1
does not exist. Similar considerations yield f .1=2/ D 1=2.”
4. Consider an instance of the one-dimensional cutting stock problem with a stock
length L equal to 122, and six different item lengths l D .62; 61; 50; 30; 20; 12/>
with corresponding demands equal to b D .2; 1; 1; 1; 2; 4/>.
>
(a) Calculate the material bound zM WD l Lb .
(b) Provide a feasible integer solution such that the items are cut from no more
than dzM e C 1 pieces of the initial material.
(c) Suppose that we may use once one unit length more in the initial material,
i.e. one pattern may use the length 123 instead of 122. Provide an optimal
integer solution for this case.
(d) Calculate lower bounds for the given instance (both without and with the
extra length naccording
o to task (c)) using the function fBJ;1 for all parameter
`i ; L`i
L L
values C 2 . What is the maximum among these bounds for the
two cases?
5. Which of the following statements is true?
(a) Any general dual-feasible function is the sum of a linear and a bounded
function.
(b) Any general dual-feasible functions is bounded from above by a linear
function.
6. Let F be the set of all functions f W R ! R with the following two properties:
i 2 R W i 2 Ig of real numbers, the implication (3.2) holds,
P finite set fxP
– For any
i.e., xi  0 H) f .xi /  0.
i2I i2I
– Any function g W R ! R with g.x/  f .x/ for all x 2 R, which obeys a similar
implication as (3.2), is necessarily identical to f , such that f is maximal in this
sense.
Let L be the set of linear functions with nonnegative slope, i.e. L is the set of
functions f W R ! R with f .x/ D cx, where c  0 is a constant. Show that
F D L.
3.8 Exercises 89

7. Let k 2 N n f0g be a constant. Define the following functions f0 ; : : : ; f3 W R ! R

by

f0 .x/ WD .1 C tanh 1/  x  maxf0; tanh xgI


x; if .k C 1/  x 2 N;
f1 .x/ WD
b.k C 1/  xc=k; otherwiseI

x; if .k C 1/  x 2 Z;
f2 .x/ WD
b.k C 1/  xc=k; otherwiseI
8
< d.k C 1/  x  1e=k; if x  1;
f3 .x/ WD x; if .k C 1/  x 2 f0; 1; : : : ; kg;
:
b.k C 1/  xc=k; otherwise:

Which of these functions are a general dual-feasible function? Which are a maximal
general dual-feasible function? Which properties of part (a) of Theorem 3.1 do these
functions possess?
8. Show using the following function g W Œ0; 1 ! Œ0; 1 that Hölder-continuity
instead of Lipschitz-continuity is not enough for the construction method II
described in Sect. 3.6.2:
 p
.1  p1  2x/=2; if x  1=2;
g.x/ WD
.1 C 2x  1/=2; otherwise.
Chapter 4
Applications for Cutting and Packing Problems

4.1 Introduction

Dual-feasible functions have been designed specifically for the cutting-stock prob-
lem. As shown in Chap. 1, they arise naturally from the dual of the classical
formulation of Gilmore and Gomory for this problem. Since many problems can
be modeled using a similar formulation, it makes sense to explore the concept of
dual-feasible function within a more general class of applications. A first approach
is to consider multi-dimensional dual-feasible functions, which can be used to derive
lower bounds for the vector packing problem. Here, we also consider different
packing problems with more complicated subproblems such as multi-dimensional
orthogonal packing and packing with conflicts. Dual-feasible functions can still be
derived in these cases.

4.2 Set-Covering Dual-Feasible Functions

In Chap. 2, dual-feasible functions were introduced as functions f W Œ0; 1 ! Œ0; 1,

such that for every finite index set I of real numbers xi 2 Œ0; 1 with i 2 I, the
following implication holds:
X X
xi  1 H) f .xi /  1 (4.1)
i2I i2I

From a linear programming point of view, it means that the function must
produce the values related to a dual solution that is valid for any cutting-stock
instance. Clearly, when one particular instance is considered, one may want to relax

© Springer International Publishing Switzerland 2016 91

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5_4
92 4 Applications for Cutting and Packing Problems

this strong constraint, and generate a dual solution that is valid for this particular
instance.
The data dependency can be introduced by choosing a finite ground set I0 , fixing
the numbers xi 2 Œ0; 1 for all i 2 I0 , and demanding the implication (4.1) only for
any subset I  I0 . In this way, the order demands of bin-packing instances may be
considered also in the function f as illustrated in the following example.
Example 4.1 Consider an instance of the 1-dimensional bin-packing problem with
four item lengths `1 D 10, `2 D 9, `3 D 6 and `4 D 4 and the respective order
demands b WD .1; 1; 2; 1/>. The length 6 is demanded twice, while the other lengths
are needed only once. The items have to be packed into the minimal number of bins
of length L D 18.
One wants to define the function f , such that

4
X
bi  f .`i =L/
iD1

is a valid lower bound for the optimal objective function value. If f should be a
classical dual-feasible function, then among others

f .1=2/  1=2 and f .2=9/  1=4

would necessarily hold. Since we are discussing data dependent dual-feasible

functions f , we set

I0 WD f1; : : : ; 5g;

and

x1 WD 5=9; x2 WD 1=2; x3 WD x4 WD 1=3; x5 WD 2=9:

Additionally, setting now

f .x1 / WD f .x2 / WD 2=3 and f .x3 / WD f .x4 / WD f .x5 / WD 1=3

is feasible for the data dependent function, as it can be checked easily. That yields
the lower bound 7=3. This bound could not be found by a classical dual-feasible
function, because the continuous relaxation of the given instance allows to use the
patterns

.1; 0; 0; 2/>;
.0; 2; 0; 0/>;
.0; 0; 3; 0/>;
.1; 0; 1; 0/>;
4.2 Set-Covering Dual-Feasible Functions 93

each in the quantity 1=2, such that the bound becomes 2. Only the last pattern was
proper, and even this pattern did not use the given length L completely. t
u
In the next section, we define formally this concept of data-dependent set-covering
dual-feasible function, i.e. functions that are designed for a specific instance.

4.2.1 Data-Dependent Dual-Feasible Functions

We consider a ground set I D f1; : : : ; ng of elements to cover bi times each. Let

P be a given polyhedral subproblem, and for a given instance DP of P, let C.DP /
be the finite set of extremal solutions of subproblem P applied to this instance.
For any extreme solution (pattern) c of C.DP /, aic 2 N (0  aic  bi ) is the
number of times element i is covered in subproblem solution c, and vc the cost
of subproblem solution c. For most bin-packing problems, where the goal is to
minimize the number of bins used, the value vc is equal to 1 for all patterns c.
We define c the variables indicating the number of times subproblem solution c is
used in the solution.

X
min vc c (4.2)
c2C.DP /
X
s.t. aic c  bi ; 8i 2 I (4.3)
c2C.DP /

c  0; 8c 2 C.DP / (4.4)

Now we give a definition of data-dependent set-covering dual-feasible function.

Definition 4.1 Let I be a set of items and DP an instance of the problem P. Let
C.DP / be the set of extremal solutions of subproblem P related to the instance
DP . A data-dependent set-covering dual-feasible function (SC-DDFF) related to the
instance DP is a mapping g defined from I to RC such that
X
c 2 C.DP / H) aic  g.i/  vc
i2I

For two different instances of the same problem, a SC-DDFF can be valid for
one and not for the other. Unlike classical dual-feasible functions, set-covering dual-
feasible functions apply to indices i instead of the sizes `i . This is due to the fact
that in the cutting-stock problem, the size `i of an item is sufficient to characterize
the element, whereas in a more general context, elements may be more complex
(several dimensions, a vertex in a graph, for example). In this formalism, geometric
constraints of packing applications are modeled as a set of feasible patterns (the set
94 4 Applications for Cutting and Packing Problems

of instance vectors). Practically speaking, being able to characterize this (possibly

exponential size) set without enumerating all its elements is crucial.

4.2.2 Data-Independent Dual-Feasible Functions

In certain specific cases, such as the cutting-stock problem, it is possible to define

functions that can be applied to any instance. This can be defined properly when the
elements i to cover in model (4.2)–(4.4) are characterized in the subproblem by a
unique real vector yi 2 Rk for a given k. We thus define the notion of set-covering
dual-feasible function.
Definition 4.2 Let k be an integer value, P a subproblem, I a set of elements, each
element i characterized by a unique real vector yi 2 Rk , and C.DP / the set of all
possible extremal solutions c related to any data of the given subproblem P and
vc defined as above. A set-covering dual-feasible function (SC-DFF) for P is an
application f defined from Rk to RC such that
X
c 2 C.DP / H) aic f .yi /  vc :
i2I

Note that any SC-DFF is a data-dependent SC-DFF for any instance.

Obviously, classical dual-feasible functions designed for the cutting-stock prob-
lem are special cases of set-covering dual-feasible functions, where k D 1, vc D 1,
and subproblem P is the knapsack problem. From now on, we will name cutting
stock dual-feasible functions (CS-DFF) the classical dual-feasible functions, while
the acronym CS-MDFF will be used for maximal CS-DFF.

4.2.3 General Properties

If one is able to compute set-covering dual-feasible functions for a given problem,

a fast lower bound is directly obtained.
Proposition 4.1 (Lower Bound Property for SC-DFF) Let DP be an instance of
problem P that is modeled as a set-covering problem, I the ground set of elements
to cover, and bi ; i 2 I the number of times each element has to be covered,
P and yi
the description of element i in the subproblem. If f is a SC-DFF, then i2I bi  f .yi /
is a valid lower bound for model (4.2)–(4.4) applied to instance DP .
Proposition 4.2 (Lower Bound Property for SC-DDFF) Let DP be an instance
of problem P that is modeled as a set-covering problem, I the ground set of elements
to cover, and bi ; i 2 I the number of times each element has to be covered. If f is a
4.3 Vector Packing Dual-Feasible Functions 95

P
SC-DDFF dependent on instance DP , then i2I bi  f .i/ is a valid lower bound for
model (4.2)–(4.4) applied to instance DP .
Once a set-covering dual-feasible function is designed, it is possible to generate
a large number of other set-covering by applying a cutting-stock dual-feasible
function to the values obtained.
Proposition 4.3 (Composition of SC-DFF and CS-DFF) The composition of a
SC-DFF g and a CS-DFF f is a SC-DFF, i.e. f .g.// is a SC-DFF.
In the remainder of this chapter we give several examples of dual-feasible functions
for different hard combinatorial problems.

4.3 Vector Packing Dual-Feasible Functions

4.3.1 Basic Definition

We now consider the m-dimensional vector packing problem (mD-VPP). That is

a multi-dimensional version of the cutting-stock problem where all dimensions
are independent, for instance one dimension is a geometric length and another
dimension a time or a weight. In order to simplify the presentation, all sizes are
assumed to be normalized, such that the bins become m-dimensional unit cubes.
The relation signs , , < and > will be used for vectors if the relation is
componentwise true. For example, s  t means si  ti , i D 1; : : : ; m. Furthermore,
let w WD .1; 1; : : : ; 1/> 2 Rm , and o WD .0; 0; : : : ; 0/> 2 Rm .
Problem 4.1 (Vector Packing, mD-VPP) An instance D WD .II LI b/ of the mD-
VPP consists in a set I D f1; 2; : : : ; ng of n items, whose sizes are given in the
matrix L, and the vector b D .b1 ; : : : ; bn /> 2 .N n f0g/n of order demands. In
L D .l11 ; l12 ; : : : ; l1m I : : : I ln1 ; ln2 ; : : : ; lnm / 2 Œ0; 1nm , the ith row-vector is li .
The mD-VPP consists in partitioning the set of items into a minimum number
of subsets such that the items in each subset fit into a bin, i.e. in no dimension the
sum of the sizes exceeds 1 for any subset. Hence, a pattern a 2 Nn is feasible, if the
capacity constraints hold on all the m dimensions, i.e.

L> a  w:

Example 4.2 Consider an instance of the 2-dimensional vector packing problem

with n D 3 items of sizes
 
7=8 1=4 1=16
L> D ;
3=8 3=8 5=16
96 4 Applications for Cutting and Packing Problems

Fig. 4.1 Optimal solution of a 2-dimensional vector packing problem instance (Example 4.2)

and demands b D .1; 2; 1/> , such that the second item is demanded twice. The only
optimal solution consists in using the two feasible patterns

a1 D .1; 0; 1/> and a2 D .0; 2; 0/>;

each in the quantity 1. This solution is illustrated in Fig. 4.1. The remaining space
in the two bins is

.1=16; 5=16/> and .1=2; 1=4/>;

such that one additional instance of the third item would have fit into the first bin.
This example demonstrates also the following difference between the vector
packing problem and the 1-dimensional cutting-stock problem. In the latter, the
material bound is generally weak, but it is always above half of the optimal objective
function value, and it can be used for some theoretical estimations. However, in the
vector packing problem, the material bound becomes even weaker, because it is only
subadditive and no longer additive, when more items are added. The material bound
would yield in this example 1516 usage of the first bin in the first dimension, and 4
3

for the second bin in the second dimension, while the total material bound is only
23 15 3
16 < 16 C 4 . t
u
A straightforward way of computing lower bounds for the mD-VPP (m > 1)
would be to consider each dimension of the multi-dimensional problem indepen-
dently and to compute a lower bound for each of the related m instances of the
1-dimensional bin-packing problem separately, for example by applying a dual-
feasible function to the obtained data. However, this may lead to arbitrarily bad
results. Consider as an example an instance, where each item i has a size equal to 1
on dimension i and " on the other dimensions, " > 0 sufficiently small. Any bound
based on that decomposition into m independent 1-dimensional problems yields the
optimal value 2 for all the m problems, such that finally at most two is obtained as
lower bound, although one needs m bins to pack all the items. By increasing the
value of m, the ratio between the optimal objective function value and the obtained
lower bound can become arbitrarily bad.
4.3 Vector Packing Dual-Feasible Functions 97

To avoid this bad behaviour, the m dimensions must be considered simulta-

neously. That applies also to the functions to be defined. Therefore, we apply
the concept of set-covering dual-feasible function, where the subproblem is the
following multi-dimensional knapsack problem.
Problem 4.2 (Multi-Dimensional Knapsack, mD-KP) An instance D WD
.II LI bI / of the mD-KP consists in a set I of n items whose sizes are given
in the matrix L, and whose profits are given in the vector , and the vector b of
upper bounds. Pn
PnThe mD-KP can be stated as follows. mD-KP.II LI bI / D fmax iD1 i xi :
iD1 `ij xi  1, j D 1; : : : ; m, 0  xi  bi , i D 1; : : : ; ng.

We now introduce the concept of vector packing dual-feasible functions, where

the 1-dimensional domain is now replaced by an m-dimensional one. These
functions can be applied directly to mD-VPP instances without separating the m
dimensions. Therefore, they may lead to much stronger lower bounds for the mD-
VPP. Our aim consists in getting results that are similar to those obtained for the
1-dimensional case.
First, we introduce formally the definitions of (maximal) vector packing dual-
feasible functions.
Definition 4.3 A function f W Œ0; 1m ! Œ0; 1 is a vector packing dual-feasible
P if for all instances of the mD-VPP and all feasible patterns a 2
function (VP-DFF),
Nn satisfying niD1 ai `id  wd , d D 1; : : : ; m, the following inequality holds:

X
n
ai  f .l>
i /  1:
iD1

A VP-DFF is called maximal if it is not dominated by another VP-DFF, as stated

in the following definition.
Definition 4.4 A VP-DFF f is maximal (VP-MDFF), if there is no other VP-DFF
g with g.x/  f .x/ for all x 2 Œ0; 1m .
Examples of some simple VP-DFF are the projections to the jth coordinate of
the argument-vector ( j D 1; : : : ; m), i.e. fj .x/ D xj . These functions lead to lower
bounds for the mD-VPP which are already known from the 1-dimensional bin-
packing problem for the separated m dimensions.
Example 4.3 Consider the instance of Example 4.2 with fourfold demand vector,
i.e., now let b WD .4; 8; 4/> . The simple material bound would yield 23=4 < 6.
Separating the two dimensions would lead to the bound 6 in the first dimension and
7 in the second direction, because one could use twice the pattern . 1; 0; 2/>, such
that ten items with size 3=8 would remain, which would fill five more bins. Finally,
the maximum of the bounds is to be taken, yielding the value 7. However, the first
item is incompatible with the second one due to the first dimension, such that the
remaining demand .2; 8; 0/> requires 2 C d8=2e D 6 more bins, i.e. 8.
98 4 Applications for Cutting and Packing Problems

A bound 8 could be achieved, if we would have e.g. a VP-DFF f W Œ0; 12 ! Œ0; 1
with
     
7 3 1 3 1 1 5
f ; D 1; f ; D ; f ; D 0;
8 8 4 8 2 16 16

because
1
b1  1 C b2  C b3  0 D 8:
2
Note that such a function f can be constructed according to Proposition 4.9 (p. 106),
which will be introduced later as a general construction principle, with u WD . 14 ; 0/>
and with g being the function (4.8) (p. 100) with parameter C D 8=3. t
u

4.3.2 General Properties of VP-MDFF

We now show that some properties for the 1-dimensional case can be generalized
to the multidimensional case, and we give a complete characterization of maximal
functions for the m-dimensional case. Finally, we show how to build such maximal
functions from non-maximal superadditive ones by forcing symmetry.
The necessary conditions from the 1-dimensional case for a function to be
maximal are still valid for the higher-dimensional case. However, it has to be
checked how the higher-dimensional case can be described and if stronger sufficient
conditions are needed. These ideas led to the following theorems.
Theorem 4.1 Any VP-MDFF f W Œ0; 1m ! Œ0; 1 has necessarily the following
properties:
1. f is superadditive, i.e. for all x; y 2 Œ0; 1m with x C y  w, it holds that

f .x C y/  f .x/ C f .y/I (4.5)

2. f is non-decreasing:

f .x/  f .y/; if o  x  y  wI

3. f is symmetric, i.e. for all x 2 Œ0; 1m , it holds that

f .x/ C f .w  x/ D 1; (4.6)

and especially f .w/ D 1 and f . 12 w/ D 1=2.

Properties (4.5)–(4.6) of Theorem 4.1 are also sufficient conditions for a function
f W Œ0; 1m ! Œ0; 1 to be a VP-MDFF. However, more restricted sufficient conditions
4.3 Vector Packing Dual-Feasible Functions 99

can be derived too, as stated in Theorem 4.2. This theorem may help to simplify the
proofs of maximality in the next sections. Before introducing these new sufficient
conditions, first in Lemma 4.1 an additional assertion is described that is useful to
prove the maximality of a VP-DFF.
Lemma 4.1 If a VP-DFF f W Œ0; 1m ! Œ0; 1 satisfies the symmetry condition (4.6),
then f is a VP-MDFF.
This is obvious since for such a symmetric VP-DFF f , if there is a VP-DFF g
such that g.x/ > f .x/ for a given x, then g.w  x/ < f .w  x/ must hold (otherwise
g.x/ C g.w  x/ > 1).
Example 4.4 Consider the 2-dimensional case, and let f W Œ0; 12 ! Œ0; 1 be the
following function:
8
ˆ 1 1
< 0; if x  2 w ^ x ¤ 2 w;
f .x/ WD 1; if x  2 w ^ x ¤ 12 w;
1 (4.7)
:̂ 1 ; otherwise:
2

This function is a VP-DFF, because one has, for any finite set of vectors x1 ; : : : ; xn 2
P
n P
n
Œ0; 12 with xi  w; that f .xi /  1, as it can be checked easily. Additionally,
iD1 iD1
since the symmetry condition holds, f is maximal, and hence it is a VP-MDFF. t
u
The following theorem restricts the needed sufficient conditions for maximality
proofs. The idea came from the 1-dimensional case, and here twice a certain dimen-
sion can be chosen, where the weaker sufficient conditions of the 1-dimensional
cutting stock problem are applied.
Theorem 4.2 Given two constants r; s 2 f1; : : : ; mg and a function f W Œ0; 1m !
Œ0; 1, the following conditions are sufficient for f to be a VP-MDFF:
1. Equation (4.6) is true for all x 2 Œ0; 1m with xr  1=2;
2. Inequality (4.5) holds for all x; y 2 Œ0; 1m with x C y  w, and xs  ys  1=2
and xs C ys  2=3.
The following propositions state that the functions resulting from the convex
combination of VP-MDFF or from the composition of a VP-MDFF with a maximal
dual-feasible function remain maximal.
Proposition 4.4 Any convex combination of VP-MDFF is a VP-MDFF.
Example 4.5 Consider again the 2-dimensional case. The following function g W
Œ0; 12 ! Œ0; 1 is a VP-MDFF:

x1 C x2
g.x/ WD :
2
100 4 Applications for Cutting and Packing Problems

The convex combination of g with the function (4.7) with the weights  2 Œ0; 1 and
1  , respectively, yields another VP-MDFF h W Œ0; 12 ! Œ0; 1, namely
8

ˆ
< 2
 .x1 C x2 /; if x  12 w ^ x ¤ 12 w;

h.x/ D 2
 .x1 C x2 / C 1  ; if x  12 w ^ x ¤ 12 w;
:̂ 
 .x1 C x2 / C 1 ; otherwise:
2 2

t
u
Proposition 4.5 The composition of a VP-MDFF f with a MDFF g, i.e. g.f .//, is
a VP-MDFF.
Example 4.6 Let f W Œ0; 1m ! Œ0; 1 be simply

kxk1 x1 C    C xm
f .x/ WD D
m m

and let g W Œ0; 1 ! Œ0; 1 be the parameter dependent function

8
< bCxc=bCc; if x < 1=2;
g.x/ WD 1=2; if x D 1=2; (4.8)
:
1  g.1  x/; otherwise;

with C 2 R, C  1. That yields the composed function

8C ˘
< m  kxk1 =bCc; if kxk1 < 2 ;
m

g.f .x// D 1=2; if kxk1 D m2 ;

:
1  g.1  f .x//; otherwise:

t
u
We now describe how a VP-MDFF can be built from a superadditive m-
dimensional vector function by forcing symmetry. This result generalizes Theo-
rem 2.4.
Proposition 4.6 Let f W Œ0; 1m ! Œ0; 1 be a superadditive function, and M be any
subset of Œ0; 1m n f 21 wg such that:
1. for all x 2 Œ0; 1m n f 21 wg, the following equivalence holds:

x 2 M ” w  x … MI

2. for any x; y 2 M, it holds that

x C y 6 w: (4.9)
4.3 Vector Packing Dual-Feasible Functions 101

The following function g W Œ0; 1m ! Œ0; 1 built from f is a VP-MDFF:

8
< 1=2; if 2x D w;
g.x/ WD 1  f .w  x/; if x 2 M;
:
f .x/; otherwise.

Example 4.7 There are various ways to choose the set M in Proposition 4.6, for
instance as the union of m parts according to
 
1
M WD Œ0; 1  Œ0; 1      Œ0; 1  ;1 [
2
   
1 1
[ Œ0; 1      Œ0; 1  ;1  [ [
2 2
     
1 1 1
[ ;1   :
2 2 2
  ˚ 
For m D 2, this set becomes Œ0; 1  12 ; 1 [ 12 ; 1  12 , i.e. the upper half of the
unit square, where the border belongs only partially to M. t
u
Example 4.8 Additionally, M could be chosen for example as follows, where the
inner of M is a triangle:

1
M WD fx 2 Œ0; 12 W x1 C x2 > 1g [ fx 2 . ; 1  Œ0; 1 W x1 C x2 D 1g:
2
The resulting function obtained by applying Proposition 4.6 to a random superaddi-
tive VP-DFF is depicted in Fig. 4.2. t
u

x2 x2 x2
1 1 1
0.8 f (x) = 1
2

0.5 0.5
0.4 f (x) = 1
0.2 0.2
x1 x1 x1 f (x) = 0
0.5 1 1/2 1 0.2 0.5 0.8 1
Set M f (x) g(x)

Fig. 4.2 Using Proposition 4.6 to build a VP-DFF. For set M, the dashed line and the point
.1=2; 1=2/ do not belong to M, while the solid line does
102 4 Applications for Cutting and Packing Problems

4.3.3 General Classes of VP-MDFF

In this section, several general classes of VP-MDFF are described. When general
schemes for generating VP-MDFF are proposed, some specific functions that can
be obtained from these schemes are described and analyzed.

4.3.3.1 Class I: Functions Based on Projections into 1-Dimensional

Domains

The first set of VP-MDFF is built from the projection of the m-dimensional data
into 1-dimensional domains. A formal definition of these VP-MDFF is given in
Proposition 4.7.

C with u w D 1.
Proposition 4.7 Let g W Œ0; 1 ! Œ0; 1 be a MDFF and u 2 Rm >

The function fI .I g; u/ W Œ0; 1 ! Œ0; 1 with

m

fI .xI g; u/ WD g.u> x/

is a VP-MDFF.
Using the MDFF fFS;1 in Proposition 4.7 yields the function described in
Corollary 4.1.
Corollary 4.1 Let v 2 Rm C such that v w 2 N n f0; 1g. The following function
>

fI;FS;1 .I v/ W Œ0; 1 ! RC is a VP-MDFF:

m

(
v> x
v> w
; if v> x 2 N;
fI;FS;1 .x; v/ WD bv> xc
v> w1
; otherwise:

Applying Proposition 4.7 with the MDFF fBJ;1 leads to the following function.
Corollary 4.2 Let v 2 Rm C be any vector with v w  1. Then, the function
>

fI;BJ;1 .I v/ W Œ0; 1 ! RC with

m

  
frac.v> x/  frac.v> w/
fI;BJ;1 .xI v/ WD bv> xc C max 0; =bv>wc
1  frac.v> w/

is a VP-MDFF.
Note that the function fI;BJ;1 .I v/ of Corollary 4.2 is only a convex combination of
the projections f1 ; : : : ; fm if v> w 2 N.
Example 4.9 To demonstrate how a function f of Class I can be built, consider
the simple function fMT;0 .I 12 /, defined in Formula (2.16), p. 48, applied to each
dimension, with different vectors u. We restrict the example to two dimensions.
4.3 Vector Packing Dual-Feasible Functions 103

If u D .1; 0/> , then

1
f .x/ D 0; if x1 < ;
2
1 1
f .x/ D ; if x1 D ;
2 2
1
f .x/ D 1; if x1 > :
2
3

1 >
If u D 4; 4 , then

3 1 1
f .x/ D 0; if x1 C x2 < ;
4 4 2
1 3 1 1
f .x/ D ; if x1 C x2 D ;
2 4 4 2
3 1 1
f .x/ D 1; if x1 C x2 > :
4 4 2
1

1 >
If u D 2; 2 , then

f .x/ D 0; if x1 C x2 < 1;
1
f .x/ D ; if x1 C x2 D 1;
2
f .x/ D 1; if x1 C x2 > 1:

The behaviour of these functions, and other examples of parameters, are illustrated
in Fig. 4.3. t
u

x2 x2 x2 x2
1 1 1 1
f (x) = 1
2

1/2 f (x) = 1

f (x) = 0
x1 x1 x1 x1
0.5 1 1/3 2/3 1 1 1

Fig. 4.3 Behaviour of class I functions for respectively u D .1; 0/> ; . 34 ; 14 /> ; . 12 ; 12 /> and
.0; 1/> when function fMT;0 .LI 12 / is used
104 4 Applications for Cutting and Packing Problems

Example 4.10 Recall Example 1.7, p. 14. After scaling the values of the weights
and volumes, the 2-dimensional items have the following sizes:

Item 1 2 3 4 Vehicle
Volume .x1 / 0.5 0.75 0.25 0.5 1
Weight .x2 / 0.6 0.4 0.8 0.2 1

Using the function fMT;0 .I 12 / for different values of u; the sizes of the items are
mapped into the values indicated in the following table:

.x1 ; x2 / u D .1; 0/> u D . 34 ; 14 /> u D . 12 ; 12 /> u D . 14 ; 34 /> u D .0; 1/>

.0:5; 0:6/ 12 1 1 1 1
.0:75; 0:4/ 1 1 1 0 0
.0:25; 0:8/ 0 0 1 1 1
.0:5; 0:2/ 12 0 0 0 0

Note that the several DFF in the family provide the dual feasible solutions uO 1 D
. 12 ; 1; 0; 12 /> ; uO 2 D .1; 1; 0; 0/>; uO 3 D .1; 1; 1; 0/> and uO 4 D .1; 0; 1; 0/>; indicated
in Example 1.7.
As all the demands are equal to 1, the value of the lower bound is equal to the
sum of the elements of u: O The function with u D . 12 ; 12 /> that provides the dual
feasible solution uO 3 D .1; 1; 1; 0/> is the one that yields the best lower bound, equal
to 3. t
u

4.3.3.2 Class II

Some of the ideas of the 1-dimensional MDFF can be adapted for the m-dimensional
vector packing problem, for instance, the function which maps small items to zero
and large ones to 1, while the other items remain unchanged, can be generalized
in the following way. Note that the difficulty in this generalization lies in finding a
suitable definition of small and large items when vectors are involved.
Proposition 4.8 Let h W Œ0; 1m ! R be non-decreasing with h.x/ C h.w  x/ > 0
for all x 2 Œ0; 1m , and let g W Œ0; 1m ! Œ0; 1 be a VP-MDFF. The following
functions fII1 ; fII2 W Œ0; 1m ! Œ0; 1 are VP-MDFF:
8 8
< 0; if h.x/  0 < 0; if h.x/ < 0
fII1 .x/ WD 1; if h.w  x/  0 ; fII2 .x/ WD 1; if h.w  x/ < 0 :
: :
g.x/; otherwise g.x/; otherwise
4.3 Vector Packing Dual-Feasible Functions 105

 m
Corollary 4.3 Let u 2 0; 12 , and let g W Œ0; 1m ! Œ0; 1 be a VP-MDFF. The
following functions fII3 .I g; u/; fII4 .I g; u/ W Œ0; 1m ! Œ0; 1 are also VP-MDFF:
8 1
< 0; if x  u and x ¤ 2 w;
fII3 .xI g; u/ WD 1; if x  w  u and x ¤ 12 w;
:
g.x/; otherwiseI
8
< 0; if x < u;
fII4 .xI g; u/ WD 1; if x > w  u;
:
g.x/; otherwise:

Corollary 4.4 Let k  kp be an L p -norm in Rm with 1  p  1, i.e.

v
u m
uX
kxk1 D max jxr j and kxkp D t
p
jxr jp for p < 1: (4.10)
rD1;:::;m
rD1

Let g W Œ0; 1 ! Œ0; 1 be a VP-MDFF and " 2 .0; kwkp =2/. The following function
fII5 W Œ0; 1m ! Œ0; 1 is a VP-MDFF:
8
< 0; if kxkp  ";
fII5 .x/ WD 1; if kw  xkp  ";
:
g.x/; otherwise:

Corollary 4.5 Let g W Œ0; 1m ! R be any non-decreasing function, and let r 2
f1; : : : ; mg. The following function fII6 W Œ0; 1m ! Œ0; 1 is a VP-MDFF:
8
< 0; if 2w> x < m and g.x/ < 0;
fII6 .x/ WD 1; if 2w> x > m and g.w  x/ < 0;
:
xr ; otherwise:

Example 4.11 A 2-dimensional example of function fII3 .I g; u/ obtained by using

Corollary 4.3 is given next. The VP-MDFF g W Œ0; 12 ! Œ0; 1 in the left-hand side
of Fig. 4.4 is defined by
8
< 0; if x1 < 1=4;
g.x/ WD 1; if x1 > 3=4;
:
x1 ; otherwise.

Using u WD .1=3; 1=3/> results in the function fII3 .I g; u/ depicted in the right part
of Fig. 4.4. t
u
106 4 Applications for Cutting and Packing Problems

1 1
f(x)=0

2/3

f(x)=1
1/3

f(x)=x1
1/4 3/4 1 1/3 2/3 1

Fig. 4.4 Applying Corollary 4.3 to the left-hand VP-MDFF, using u D . 13 ; 13 />

4.3.3.3 Class III

The following proposition describes another VP-MDFF. First the general function
for the m-dimensional case is defined, then a special case for m D 2 is given in
Corollary 4.6. The rationale behind this function consists in assigning the value
0 (or 1) to items that are very small (respectively large) on some dimensions,
unless they are very large (or small) on another dimension. The remaining items
are mapped via any other VP-MDFF of appropriate dimension.
Proposition 4.9 Let g W Œ0; 1m ! Œ0; 1 be a VP-MDFF and u 2 Œ0; 1=2m . The
following function fIII1 W Œ0; 1m ! Œ0; 1
8
< 0; if 9i 2 f1; : : : ; mg W xi < ui ^ 6 9j 2 f1; : : : ; ig with xj > 1  uj ;
fIII1 .x/ WD 1; if 9i 2 f1; : : : ; mg W xi > 1  ui ^ 6 9j 2 f1; : : : ; ig with xj < uj ;
:
g.x/; otherwise,

is a VP-MDFF.
Corollary 4.6 The function fIII2 .I u1 ; u2 ; r; q/ W Œ0; 12 ! Œ0; 1 with 0  u1 ; u2 
1=2 and r; q 2 f1; 2g, which is defined as
8
< 1; if xq > 1  u1 or .xq  u1 and x3q > 1  u2 /;
fIII2 .xI u1 ; u2 ; r; q/ WD xr ; if 1  u1  xq  u1 and 1  u2  x3q  u2 ;
:
0; otherwise,

is a VP-MDFF.
Example 4.12 Consider an example in dimension 2 and choose r D 1, q D 2,
u1 D 1=4 and u2 D 1=3.
8
  < 1; if x2 > 3=4 or .x2  1=4 and x1 > 2=3/;
1 1
fIII2 xI ; ; 1; 2 WD x1 ; if 3=4  x2  1=4 and 2=3  x1  1=3;
4 3 :
0; otherwise,
4.3 Vector Packing Dual-Feasible Functions 107

Fig. 4.5 Applying x2

Corollary 4.6 with r D 1,
q D 2, u1 D 1=4 and 1
u2 D 1=3 f (x) = x1
3/4

f (x) = 1

1/4 f (x) = 0

x1
1/3 2/3 1

The obtained function is depicted in Fig. 4.5. t

u

4.3.3.4 Class IV

In this subsection, another class of VP-MDFF in its general form and a special-
ization for the 2-dimensional case are described. This VP-MDFF depends on a
non-decreasing function g whose properties are stated in the next proposition. How
g should be chosen is explained at the end of the section.
Proposition 4.10 Let m 2 N n f0g, k 2 .1=3; 1=2, g W Œ0; 1m ! Œ0; 1 be a non-
decreasing function with

x; y 2 Œ0; 1m ; x C y  w H) g.x/ C g.y/  1;

˚ 
and let M be a subset of Œ0; 1m n 12 w such that x 2 M ” w  x … M for all
˚ 
x 2 Œ0; 1m n 12 w .
The function fIV1 .I g; k; M/ W Œ0; 1mC1 ! Œ0; 1 defined as
8
ˆ
ˆ 0; if xmC1 < k;
ˆ
ˆ
ˆ
ˆ 1; if xmC1 > 1  k;
ˆ
ˆ
ˆ
< 1=2; if x1 D x2 D    D xmC1 D 1=2;
fIV1 .x; xmC1 I g; k; M/ WD g.x/; if 1=2 < xmC1  1  k
ˆ
ˆ
ˆ
ˆ or .xmC1 D 1=2 and x 2 M/;
ˆ
ˆ
ˆ
ˆ 1  g.w  x/; if k  xmC1 < 1=2
:̂
or .xmC1 D 1=2 and w  x 2 M/;
(4.11)
is a VP-MDFF.
108 4 Applications for Cutting and Packing Problems

Note, here the condition (4.9), p. 100, to M needs not to be demanded.

The function (4.11), i.e. fIV1 .; I g; k; M/, becomes for m D 1 and M D . 12 ; 1 as
follows:
8
ˆ
ˆ 0; if x2 < k;
ˆ
ˆ
ˆ
ˆ 1; if x2 > 1  k;
ˆ
ˆ
ˆ
< 1=2; if x1 D x2 D 1=2;
fIV2 .x1 ; x2 I g; k; .1=2; 1/ D g.x1 /; if 1=2 < x2  1  k
ˆ
ˆ
ˆ
ˆ or .x2 D 1=2 and x1 > 1=2/;
ˆ
ˆ
ˆ
ˆ 1  g.1  x1 /; if k  x2 < 1=2
:̂
or .x2 D 1=2 and x1 < 1=2/:

The following VP-MDFF is similar, but it differs for x1 2 f0; 1g, i.e. in these
cases it may get other function values.
Proposition 4.11 Let g be a non-decreasing function defined from Œ0; 1 to Œ0; 1,
such that

g.y/ C g.1  y/  1; for all y 2 Œ0; 1:

1 1

The function fIV3 .I g; k/ W Œ0; 12 ! Œ0; 1 with k 2 3; 2 and defined as
8
ˆ
ˆ 0; if x1 D 0 or .x1 < 1 and x2 < k/;
ˆ
ˆ
ˆ
ˆ 1; if x1 D 1 or .x1 > 0 and x2 > 1  k/;
ˆ
ˆ
ˆ
< 1=2; if x1 D x2 D 1=2;
fIV3 .xI g; k/ WD g.x1 /; if .1=2 < x2  1  k and 0 < x1 < 1/
ˆ
ˆ
ˆ
ˆ or .x2 D 1=2 and 1=2 < x1 < 1/;
ˆ
ˆ
ˆ
ˆ 1  g.1  x1 /; if .k  x2 < 1=2 and 0 < x1 < 1/
:̂
or .x2 D 1=2 and 0 < x1 < 1=2/;

is a VP-MDFF.
In the following proposition, we show how the g function in Proposition 4.11
should be defined so as to get the best lower bound that can be obtained from the
corresponding VP-MDFF fIV3 .I g; k/ for the 2D-VPP.
Proposition 4.12 Given an instance of the 2D-VPP, the best lower bound based on
the function f˚IV3 .I g;k/ of Proposition 4.11 can be found with the following function
g W Œ0; 1 ! 0; 12 ; 1 that depends on the parameters s; t 2 R with 0  s  12 and
s  t  1  s:
8
< 0; if x < s;
g.x/ WD 1=2; if s  x  t;
:
1; if x > t:
4.3 Vector Packing Dual-Feasible Functions 109

Fig. 4.6 Applying x2

Proposition 4.11 to construct
a VP-MDFF 1

f (x) = 1
2

3/5
1/2 f (x) = 1
2/5

f (x) = 0

x1
1/4 1/2 3/4 1

Example 4.13 For a 2-dimensional example let us choose s D 1=4, t D 1=2 and
k D 2=5. The obtained function is depicted in Fig. 4.6. t
u

4.3.3.5 Class V

Let s; t 2 .0; 1m be two constant vectors. Let u1 ; u2 be feasible (but not necessarily
optimal) dual values for an instance of the m-dimensional vector packing problem
with two items of sizes s and t, each demanded at least once. The following proposi-
tion describes a new family of superadditive VP-DFF. Recall that these functions can
be transformed into VP-MDFF by enforcing symmetry via Proposition 4.6, p. 100.
Proposition 4.13 The function fV .I s; t; u1 ; u2 / W Œ0; 1m ! Œ0; 1 is a superadditive
VP-DFF:

fV .xI s; t; u1 ; u2 / WD maxfa1 u1 C a2 u2 ja1 ; a2 2 N; a1 s C a2 t  xg:

Calculating the function requires solving an integer optimization problem for

every argument x. However, the complexity remains low, if the possible values
a1 ; a2 2 N are bounded by a small constant.
Assume without loss of generality that

max si  max ti :
i2f1;:::;mg i2f1;:::;mg

Since one has only two items s; t > o, the function value can be easily calculated by
trying all possible numbers a2 , i.e. a2 2 N and

a2  minfxi =ti W i 2 f1; : : : ; mgg;

110 4 Applications for Cutting and Packing Problems

Fig. 4.7 Class V function x2

obtained from Example 4.14.
The bold line corresponds to 1
the segment Œ.1; 0:4/; .1; 1/ f (x) = 1
2
associated to value 1

f (x) = 1
0.4
f (x) = 0
0.2

x1
0.5 1

and setting

a1 WD min .xi  a2  ti /=si :
i2f1;:::;mg

These calculations have the same effort as dynamic optimization with exactly two
different items. Hence, the complexity to calculate the appropriate a1 and a2 is
pseudo-polynomial.
Example 4.14 Consider the 2-dimensional example with s D .0:5; 0:2/> and t D
.0:5; 0:6/> . Pick the feasible dual values u1 D 0:5 and u2 D 0:5.
The optimization problem is

f .x/ D maxf0:5a1 C 0:5a2 W 0:5a1 C 0:5a2  x1 ; 0:2a1 C 0:6a2  x2 ; a1 ; a2 2 Ng:

The resulting superadditive but not maximal function is depicted in Fig. 4.7. It
can be formulated as follows.
8
ˆ
<1;
ˆ if x1 D 1 and x2  0:4
f .x/ D 0; if x1 < 0:5 or x2 < 0:2
ˆ
:̂1=2; otherwise
t
u

4.4 Orthogonal Packing

Here, we consider the m-dimensional bin-packing problem with and without

rotation. Classical Cutting-Stock-DFF (CS-DFF) can be used to derive dual-feasible
functions for this problem. Actually, most of the results in this section were initially
4.4 Orthogonal Packing 111

proposed without the general formalism of dual-feasible functions. Our presentation

allows to gather different results from the literature under the same formalism.
We first define properly the m-dimensional Orthogonal Bin-Packing Problem,
where one tries to pack m-dimensional rectangular bricks into large m-dimensional
rectangular bricks. When rotation is allowed, since the bins are not always cubes, it
is not possible to normalize the sizes in Œ0; 1m .
Problem 4.3 (m-Dimensional Orthogonal Bin-Packing Problem, m-OPP) An
instance D D .II LI w/ of the m-OPP consists in a set I D f1; 2; : : : ; ng of n items,
whose sizes are given in the matrix L D .l11 ; l12 ; : : : ; l1m I : : : I ln1 ; ln2 ; : : : ; lnm / 2
Nnm (with li being the ith row-vector of L), and a rectangular brick described by
its dimensions w. The m-OPP consists in finding a partition of the set of items into a
minimum number of subsets such that the items in each subset fit into a bin (i.e. the
rectangular bricks fit into the boundary of the large rectangular brick, no two items
overlap, and the edges of the items are parallel to the edges of the bin). If the rotation
of the items is allowed, we have a m-OPP with rotation (m-OPP-R), otherwise, we
have a m-OPP with fixed orientation (m-OPP-O).
The application of the concept of set-covering dual-feasible function to the
orthogonal packing problem with and without rotation will be called m-OPP-R-DFF
and m-OPP-O-DFF.

4.4.1 DFF for the Oriented Case (m-OPP-O-DFF)

One dual-feasible function can be applied to each dimension of an instance of m-

OPP-O to obtain a lower bound. Similarly to the vector packing problem, the shape
of each item i is described by a size vector l. Thus, it is possible to design dual-
feasible functions for this case that are independent of the data. Using the formalism
introduced at the beginning of the section, the result can be written as follows.
Proposition 4.14 Let fj W j D 1; : : : ; m be CS-DFF. The following function g W
Rm 7! Œ0; 1 is a m-OPP-O-DFF.

Y
m
g.x/ WD fj .xj =wj /
jD1

Example 4.15 Let m D 2, w D .10; 10/ and consider four items of size .6; 6/. The
trivial lower bound is equal to d4  .36=100/e D 2. By taking f1 D fMT;0 .I 12 / and
f2 D fMT;0 .I 12 /, one obtains a bound equal to d4  .1  1/e D 4. t
u
If the identity function is used for fj ; j D 1; : : : ; m, the classical bound based on
the surface/volume of the bins is obtained. No actual m-dimensional dual-feasible
functions were derived in the literature (i.e. dual-feasible functions that would not
112 4 Applications for Cutting and Packing Problems

consider the problem dimension by dimension). This can be explained by the fact
that characterizing the set of feasible patterns is hard. Even verifying that a pattern
is feasible is NP-complete.

4.4.2 DFF for the Case with Rotation (m-OPP-R-DFF)

The first result derives from a simple fact. For a set of CS-DFF f1 ; : : : ; fm , if a
lower bound for the oriented case based on these functions is run for all possible
orientations of the items, and if the minimum is recorded, a valid lower bound is
obtained. Of course, the bound obtained would need an exponential time, since it
would take .mŠ/n lower bounds to compute. Nevertheless a lower bound can be
computed by considering the following relaxation: for each item i, keep the smallest
image that it can have for its possible orientations. This leads to n  mŠ values to
compute, which can lead to a practical method for lower bounding. Let S be the set
of the mŠ permutations  D .1/; : : : ; .m/ representing all possible orientations in
m dimensions.
Proposition 4.15 Let fj W j D 1; : : : ; m be CS-DFF. The following function '1 W
Rm 7! Œ0; 1 is a m-OPP-R-DFF.
8 9
<Ym =
'1 .x/ WD min fj .x .j/ =wj /
 2SWx .j/ wj ;jD1;:::;m : ;
jD1

A better m-OPP-R-DFF, that dominates the previous one if the fj are increasing
and superadditive, and if the container has equal size on each dimension, is now
described.
Proposition 4.16 Let fj W j D 1 : : : ; m be m CS-DFF. If the instance of m-OPP-R is
such that w1 D w2 D : : : D wm then the following function '2 W Rm 7! Œ0; 1 is a
m-OPP-R-DFF.
( Qm )
X jD1 fj .x .j/ =wj /
'2 .x/ WD
 2S
mŠ

The result is not intuitive, but it becomes obvious when the following relaxation
is considered. From a m-OPP-R instance, construct a m-OPP-O instance I 0 of size
mŠ  n where each item is repeated once for each of its orientations. Clearly, the
value of an optimal solution for this new problem cannot be more than mŠ times the
value of an optimal solution for the original m-OPP-R instance (see Fig. 4.8).
Example 4.16 Take m D 2, w D .10; 10/ and four identical items .8; 3/, and choose
the identity function for f1 and f2 D fMT;0 .I 0:25/, where the latter function was
4.5 Bin-Packing 113

7
3 4
4 6
2 6 2 7
1
1 5 3 5

7
3 4
2 6
1 5

Fig. 4.8 An example of the relaxation of the OPP-R for m D 2. Note that for any solution for the
OPP-R using z bins, there is always a solution of the duplicated OPP-O problem using 2z bins

defined in Formula (2.16), p. 48, as

8
< 0; if x < ;
fMT;0 .xI / WD 1; if x > 1  ;
:
x; otherwise.

'1 ..8; 3// D minf0:8fMT;0 .0:3I 0:25/; 0:3fMT;0 .0:8I 0:25/g D minf0:80:3; 0:3
1g D 0:24. The bound obtained is equal to d4  24=100e D 1.
'2 ..8; 3// D .0:8  fMT;0 .0:3I 0:25/ C 0:3  fMT;0 .0:8I 0:25//=2 D .0:24 C
0:3/=2 D 0:27. The bound obtained is equal to d4  0:27e D 2. t
u

4.5 Bin-Packing

In this section, we present data-dependent dual-feasible functions designed for the

bin-packing problem. For any item i, it is assumed that 0 < `i  1. For our study, the
only difference between bin-packing and cutting-stock is the fact that the number
of items for each type is small and thus packing b1=`i c instances of a given item i
may not be allowed. This means that the set of valid patterns is smaller. In the data-
dependent dual-feasible functions described in this section, the number of times an
item is repeated in the instance can be taken into account.
Problem 4.4 (Bin-Packing Problem, BP) An instance D D .I; l/ of the OPP
consists in a set I D f1; 2; : : : ; ng of n items, whose sizes are given in the vector
l 2 Œ0; 1nm . The Bin-Packing problem consists in finding a partition of the set of
items into a minimum number of subsets such that the items in each subset fit into a
bin, i.e. the sum of the sizes in each dimension does not exceed 1 for any subset.
In what follows, we use the classical 1-dimensional knapsack problem.
114 4 Applications for Cutting and Packing Problems

Problem 4.5 (Binary Knapsack Problem, KP–01) An instance D D .J; l; W; ˛/

of the binary knapsack problem is composed of a set J of items i, a size vector l 2
Œ0; 1jJj jJj
C , a bin size W 2 RC and a profit vector ˛ 2 RC . Formally, KP–01.J; l; W; ˛/
can be stated as follows.
( )
X X
KP–01.J; l; W; ˛/ D max ˛i xi W `i xi  W; xi 2 f0; 1g; 8i 2 J
i2J i2J

Similarly to Cutting-Stock DFF (CS-DFF) that lead to lower bounds for the
cutting-stock problem, one can define the notion of Bin-Packing DDFF (BP-
DDFF) by specifying the subproblem of Definition 4.1 as a binary knapsack
problem. The difference between the two classes of functions is that the polyhedral
subproblem used in the definition is not the same (unbounded knapsack problem
for the cutting-stock, and binary knapsack problem for bin-packing). Actually, any
CS-DFF is a BP-DDFF, but the converse is generally not true.
Proposition 4.17 Let D D .I; l/ be BP instance, J  I a set of pairwise
incompatible items (`i C `j > 1; 8i; j 2 J) and let ˛ 2 RnC . The following function
g1 W I ! Œ0; 1 is a BP-DDFF defined for D.
8
ˆ
ˆ 1 if i 2 J and KP–01.1; I n J; l; ˛/ D 0
ˆ
ˆ
ˆ
ˆ
<1  KP–01.1  `i ; I n J; l; ˛/=
ˆ
g1 .i/ WD KP–01.1; I n J; l; ˛/ if i 2 J and KP–01.1; I n J; l; ˛/ ¤ 0
ˆ
ˆ
ˆ
ˆ
ˆ0
ˆ if i 2 I n J and ˛i D 0
:̂˛ =KP–01.1; I n J; l; ˛/ if i 2 I n J and ˛i ¤ 0
i

The percentage of the bin taken by each small item i is equal to ˛i , and the sizes
of the large items are computed by solving the knapsack problem described above.
Note that in some degenerate cases, the image of an item in J may be smaller than
the image of an item in J n I.
Example 4.17 Consider a BP-instance .I; l/, with I D f1; : : : ; 10g and

l D .0:1; 0:1; 0:1; 0:2; 0:3; 0:3; 0:3; 0:3; 0:6; 0:6/:

Choose J D f9; 10g and ˛ D .0; 0; 0; 1; 1; 1; 1; 1; 6; 6/.

KP–01.f1; : : : ; 8g; l; ˛/ D 3:
g1 .1/ D g1 .2/ D g1 .3/ D 0;
g1 .4/ D g1 .5/ D g1 .6/ D g1 .7/ D g1 .8/ D 1=3;
g1 .9/ D g1 .10/ D 1  1=3 D 2=3:
4.6 Bin-Packing Problem with Conflicts 115

Note that the obtained values cannot be computed by the means of a dual-feasible
function, since g1 .4/ D 1=3, which would not be possible because `4 D 0:2 and
for any dual-feasible function g, g.0:2/  1=5 (otherwise the dual-feasible function
condition would not hold for five items of size 0:2). t
u
The knapsack problems involved are NP-hard in the general case. However, they
can be solved in pseudo-polynomial time using dynamic programming. When the
size of the bin is large, it may entail a large computing time. In this case, the set of
parameters ˛ should be chosen in a way to re-enable the resolution of the knapsack
problem in a polynomial time (for example ˛i D 1; 8i 2 J).

4.6 Bin-Packing Problem with Conflicts

We now describe data-dependent dual-feasible functions for the bin-packing with

conflicts. In this problem, some pairs of items cannot be packed in the same bin. A
natural way of modelling these constraints is to use a graph. Note that we consider
the generic case without specifying the geometric constraint applied to the problem.
Similarly to BP-DDFF, we will name BPC-DDFF the DDFF designed for the bin-
packing problem with conflicts. Only data-dependent functions are useful here,
since it makes no sense to define a function that would be valid for any graph.
Actually the class of BPC-DFF would be equivalent to the class of CS-DFF, because
the instances with a complete compatibility graph have to be considered.
Problem 4.6 (Bin-Packing Problem with Conflicts, BPC) An instance D D
.II lI E/ of the m-OPP consists in a set I D f1; 2; : : : ; ng of n items, whose sizes
are given in the vector l 2 Œ0; 1n , and E  I  I a set of compatibility edges. The
Bin-Packing with conflicts consists in finding a partition of the set of items into a
minimum number of subsets such that the items in each subset fit into a bin, i.e. the
sum of the sizes in each dimension does not exceed 1 for any subset, and for any
two items i; j in the same partition .i; j/ 2 E.
Both techniques described below are based on graph-theoretical concepts,
namely graph triangulation, and tree-decomposition, which are defined next.

4.6.1 BPC-DDFF Based on a Knapsack Subproblem

The first BPC-DDFF is a generalization of function g1 defined in Proposition 4.17

for BP. It has to solve knapsack problems with conflicts instead of a classical
knapsack problem.
Problem 4.7 (Binary Knapsack Problem with Conflicts, KPC) An instance D D
.J; l; W; ˛; E/ of Binary Knapsack Problem with Conflicts consists of an item set
J D f1; : : : ; ng, whose sizes are given in the vector l 2 Œ0; 1n , a profit vector ˛ 2 Rn ,
116 4 Applications for Cutting and Packing Problems

and E a set of compatibility edges. Formally, KPC.J; l; W; ˛; E/ can be stated as

follows.

KPC.J; l; ˛; E/
( )
X X
D max ˛i xi W `i xi  W; xi C xj  1; 8.i; j/ 62 E; xi 2 f0; 1g; 8i 2 I
i2J i2J

In the following proposition, and in the remainder of the book, for a vertex i,
N.i/ D fig [ f j W .i; j/ 2 Eg is the neighbourhood of i.
Proposition 4.18 Let D D .II lI E/ be a BPC instance, J a set of pairwise
incompatible items, and f˛i 2 RC ; i 2 I n Jg a list of coefficients. Function
g1 .I J; ˛/ W I ! Œ0; 1 is a BPC-DDFF.
(
1  KPC.1  `i ; N.i/; l; ˛/=KPC.1; I n J; l; ˛/ if i 2 J
g1 .i/ WD
˛i =KPC.1; I n J; l; ˛/ if i 2 I n J

Example 4.18 Consider an instance D D .II lI E/ where I D f1; 2; 3; 4g,

l D .0:7; 0:6; 0:3; 0:2/

and

E D f.2; 3/; .2; 4/; .3; 4/g:

Let J D f1; 2g, and ˛3 D 1, ˛4 D 1.

KPC.W; J; l; ˛/ D 2:
g1 .3I J; ˛/ D g1 .4I J; ˛/ D 1=2;
g1 .1I J; ˛/ D 1  0 D 1;
g1 .2I J; ˛/ D 1  1=2 D 1=2:

The obtained bound is equal to d1 C 1=2 C 1=2 C 1=2e D 3.

Any bound based on a cutting-stock dual-feasible function would have been at
most two, because the optimum of the related cutting-stock problem is 2 according
to the partition I D f1; 3g [ f2; 4g. Graph colouring would also yield the bound
2 only, since two colours are enough for colouring the complementary graph
according to f1g; f2; 3; 4g. Both bounds are less than the obtained bound 3. t
u
To the best of our knowledge, no dynamic programming scheme exists for
the disjunctive knapsack problem with general graphs. When a conflict graph G
is considered, only cliques of G can be solutions of the knapsack problem with
conflicts. Thus a (possibly not practically tractable) solution for the latter is to
4.6 Bin-Packing Problem with Conflicts 117

compute all maximal cliques of the conflict graph, and then to solve for each clique
the associated knapsack problem. The maximum value obtained for all cliques is
the optimal value for the knapsack problem with conflicts. This solution is tractable
only if the cliques are in small number, and they can be computed with a small
complexity. Neither of the two conditions are fulfilled when a random graph is
considered. For this method to be tractable, the problem can be relaxed by adding
edges to the compatibility graph in such a way that it becomes triangulated. A graph
G is triangulated if for every cycle of length k > 3, there is a chord joining two non-
consecutive vertices. Any triangulated graph G has at most n maximal cliques. In
addition, they can be computed in linear time. Finding the minimum set of edges
to add in order to obtain a triangulated graph is a NP-hard problem, so a heuristic
should be used.

4.6.2 A BPC-DDFF Based on Graph Decomposition

Suppose the set I of items can be decomposed into two sets I1 and I2 of pairwise
incompatible items. In this case, two different dual-feasible functions f and g can
be applied to I1 and I2 , since the instance can be decomposed into two distinct sub-
instances. Now, if there is a third set I3 where each item is compatible with some
items of I1 and I2 , each item of I3 will be packed either with items of I1 , items of
I2 , or neither of these items, but not both. This leads to the following BPC-DDFF,
which depends on two CS-DFF f and g.
Proposition 4.19 Let D D .II lI E/ be an instance of BPC, and let also .I1 ; I2 ; I3 / be
a partition of I such that E \ f.i1 ; i2 / W i1 2 I1 ; i2 2 I2 g D ;. Let also f and g be two
CS-DFF. Function h.I f ; g; I1; I2 / W I ! Œ0; 1 defined as follows is a BPC-DDFF.
8
ˆ
<f .`i /
ˆ if i 2 I1
i 7! g.` / if i 2 I2
ˆ i
:̂minff .` /; g.` /g otherwise
i i

This technique can be generalized by decomposing the graph into different

intersecting subsets. Function h2 is based on the concept of tree-decomposition,
which captures the possible associations of items.
Definition 4.5 A tree-decomposition of G D .I; E/ is a pair .X; T/, where T D
.V; A/ is a tree with node set V and edge set A, and X D fXv W v 2 Vg, is a family
of subsets of I such that:
1. [v2V Xv D I
2. 8Œv; w 2 A, there is a Xv , v 2 V with v 2 Xv and w 2 Xv
3. 8i; j; k 2 V, if j is on the path from i to k in T, then Xi \ Xk  Xj
118 4 Applications for Cutting and Packing Problems

Let .X; T/ be a tree-decomposition of G. The basic idea of h2 is to assign a

given DFF fs to each subset s 2 X. Let F be a list of CS-DFF f1 ; : : : ; fjSj , one for
each node of the tree decomposition. Recall that S is the set of all permutations
representing the possible orientations. For each vertex i in the graph we define Si
the set of nodes of the tree decomposition containing i. Clearly there is always a set
of functions f1 ; : : : ; fjSj that allows to dominate the application of a single DFF (e.g.
f1 D f2 D : : : D fjSj ).
Proposition 4.20 Let D D .II lI E/ be an instance of BPC, and .X; T/ be a tree-
decomposition of G. Let also F D f1 ; : : : ; fjSj be a list of CS-DFF, and for each
vertex i 2 I, let Si be the set of nodes of the tree decomposition containing i. The
following function h2 .I F; .X; T// W I ! Œ0; 1 is a BPC-DDFF.

i 7! minffs .`i /g
s2Si

An issue is to choose a suitable set of functions to be applied to the nodes of the

tree decomposition.
Example 4.19 Let I D f1; : : : ; 7g,

l D .0:2; 0:2; 0:3; 0:4; 0:8; 0:5; 0:5/;

and

E D f.1; 2/; .1; 3/; .1; 4/; .2; 3/; .2; 4/; .3; 4/; .3; 5/; .4; 5/; .4; 6/; .5; 7/; .6; 7/g

(see Fig. 4.9).

1 3 3
5
2 4 4

1 3
5
5
4 7
2 4
7 6

Fig. 4.9 Compatibility graph for Example 4.19 and a possible tree-decomposition for this graph
4.7 Related Literature 119

Table 4.1 Values of the Items

functions obtained for the
Function 1 2 3 4 5 6 7
different nodes of the tree
decomposition f1 0.2 0.2 0.3 0.4
f2 0.3 0.4 1
f3 0.4 1 0.5 0.5
min 0.2 0.2 0.3 0.4 1 0.5 0.5

Let .X; T/ be a tree decomposition of .I; E/ such that

V1 D f1; 2; 3; 4g; V2 D f3; 4; 5g; and V3 D f4; 5; 6; 7g:

Let f1 D id and f2 D f3 D fMT;0 .I 0:25/. The values obtained are reported in
Table 4.1.
The obtained bound is equal to

d0:2 C 0:2 C 0:3 C 0:4 C 1 C 0:5 C 0:5e D 4:

Note that this bound is better than :

• any that would be designed for bin-packing only, since the optimal solution is 3
bins: f1; 5g; f2; 3; 4g; f6; 7g;
• any graph-colouring bound, since there is a proper colouring of the complemen-
tary graph using 3 colours only: f1; 2; 3; 4g; f5g; f6; 7g.
t
u

4.7 Related Literature

Dual-feasible functions for vector packing problems were proposed by Alves et al.
(2014). Applying dual-feasible functions to orthogonal packing problem has been
done for the first time by Fekete and Schepers (2004). It has also been done
implicitly by Boschetti and Mingozzi (2003). In Carlier et al. (2007), the functions
implicitly used by Boschetti and Mingozzi (2003) were described and slightly
improved. Caprara et al. (2005) show that applying a dual-feasible function on each
dimension of an 2-dimensional orthogonal packing problem often leads to a bound
of excellent quality. The notion of data-dependent functions was proposed by Carlier
et al. (2007). Clautiaux et al. (2007) show that dual-feasible functions can be used
to produce bounds for the orthogonal packing problem with rotation. Dual-feasible
functions for the case with conflicts were introduced by Khanafer et al. (2010).
Everything the reader needs to know about treewidth and graph triangulation is
respectively available in Rose et al. (1976) and Robertson and Seymour (1986).
120 4 Applications for Cutting and Packing Problems

The general framework for set-covering dual-feasible functions has been proposed
by Clautiaux (2010).

4.8 Exercises

1. For each of the two following functions f1 and f2 defined from Œ0; 12 to Œ0; 1,
indicate whether or not the function is a VP-DFF, and if it is a VP-MDFF. If the
function is a VP-DFF and not a VP-MDFF, propose a VP-MDFF that dominates it.

1 if x1 > 1=2 and x2 > 1=2
f1 .x/ WD
0 otherwise

1 if x1 > 1=2 or x2 > 1=2
f2 .x/ WD
0 otherwise

2. Is function f3 a 2-OPP-O-DFF?
8  ˘
ˆ
ˆ 1  w2 x 2
; if x1 > 2w1 =3 and x2 > w2 =2;
ˆ
ˆ 4
ˆ
ˆ
ˆ
ˆ 1=2; if x1 > 2w1 =3 and x2 D w2 =2;
ˆ
ˆ  ˘
ˆ
ˆ x42 ; if x1 > 2w1 =3 and x2 < w2 =2;
<  ˘

f3 .x/ WD 1  w2 x 2
; if 2w1 =3  x1  w1 =3 and x2 > w2 =2;
ˆ
ˆ 4
ˆ
ˆ =2 if 2w1 =3  x1  w1 =3 and x2 D w2 =2;
ˆ
ˆ x 1
ˆ
ˆ  x2 ˘
ˆ
ˆ if 2w1 =3  x1  w1 =3 and x2 < w2 =2;
ˆ 1 4
x
:̂
0; if x1 < w1 =3:

3. Describe the 2-OPP-O-DFF obtained using Proposition 4.14 with functions

fCCM;1 .I 4/ and fCCM;1 .I 2/. Apply it to the following instance: W D .10; 10/,
I D f1; : : : ; 6g and

L D ..5; 4/; .5; 4/; .6; 4/; .6; 4/; .8; 8/; .2; 2//:

4. Compute the 2-OPP-R-DFF obtained using Proposition 4.15 and functions

fMT;0 .I 0:3/ and fCCM;1 .I 3/. Apply it to instance W D .10; 10/, I D f1; : : : ; 6g,
and

L D ..7; 7/; .8; 8/; .4; 7/; .4; 7/; .7; 3/; .7; 3//:

Do the same using Proposition 4.16.

4.8 Exercises 121

5. Consider the 2-OPP-R. Let f and g be two CS-MDFF. For w > h, let

F f ;g .w; h/ WD bw=hcf .h=W/  g.h=H/ C F f ;g .h; w mod h/I

F f ;g .w; 0/ WD 0:

1. Prove that if w; h 2 N; w  h, and if f and g can be computed in finite time,

then F f ;g .w; h/ can be computed in finite time too.
2. Prove that '3 W i 7! F f ;g .wi ; hi / is a 2-OPP-R-DFF.
3. Show that if two CS-MDFF f and g are used, if the bin is a square, and if each
item i is such that 9k; wi D k  hi , then

'3 .i/  '1 .i/  '2 .i/;

8i 2 I (with '1 ; '2 from Propositions 4.15 and 4.16).

4. Show that if two CS-MDFF f and g are used, if the bin is a square, then

'3 .i/  '2 .i/

8i 2 I (with '2 from Proposition 4.16).

6. Let us introduce a new variant of the bin-packing problem: the multi-mode bin-
packing. In this problem, each item can have several “modes”, i.e. it has a finite
number of possible sizes. The choice of the mode is a part of the optimization.
Propose a MMBP-DFF.
7. Indicate whether the mappings f1 , f2 and f3 are data-dependent dual-feasible
functions for the instance l D .0:2; 0:2; 0:3; 0:3; 0:3; 0:6; 0:7/.

i 1 2 3 4 5 6 7
li 0.2 0.2 0.3 0.3 0.3 0.6 0.7
f1 .i/ 0.3 0.3 0.2 0.2 0.2 0.4 0.7
f2 .i/ 0.2 0.3 0.25 0.25 0.25 0.5 0.7
f3 .i/ 0.2 0.3 0.2 0.25 0.25 0.5 0.7

8. Propose a strictly better data-dependent dual-feasible function than function f

for the instance W D 1; I D f1; : : : ; 7g, l D f0:3; 0:3; 0:6; 0:6; 0:8; 0:8; 0:9g.

i 1 2 3 4 5 6 7
li 0:3 0:3 0:6 0:6 0:8 0:8 0:9
f .i/ 0:33 0:33 0:5 0:5 1 1 1
122 4 Applications for Cutting and Packing Problems

9. Show that increasing the size of an item may decrease the value of the bound
obtained using the BP-DDFF defined in Proposition 4.17. Can it happen with a
CS-MDFF?
10. Apply Proposition 4.18 to obtain a BPC-DDFF for the following instance of
BPC: W D 10, I D f1; : : : ; 14g,

l D .0:8; 0:8; 0:8; 0:8; 0:2; 0:7; 0:3; 0:6; 0:3; 0:6; 0:3; 0:5; 0:5; 0:4/

are the sizes and

E D I  I n f.8; 14/; .9; 14/; .10; 14/; .11; 14/; .12; 13/; .12; 14/; .13; 14/g

the arcs. Choose your coefficients ˛i in such a way that the lower bound produced
is optimal.
11. Apply Proposition 4.20 to obtain a BPC-DDFF for the following instance of
BPC: W D 1, I D f1; : : : ; 12g,

l D .0:4; 0:3; 0:3; 0:3; 0:3; 0:3; 0:7; 0:7; 0:7; 0:7; 0:3; 0:3/;

and the compatibility edge set is

E D f.1; 2/; .1; 3/; .1; 4/; .2; 3/; .2; 4/; .3; 4/; .3; 5/; .3; 8/; .2; 10/;
.1; 7/; .1; 9/; .9; 10/; .5; 6/.6; 11/; .11; 12/; .4; 12/; .7; 8/g

(see below). Choose your CS-DFF in such a way that the lower bound produced is
optimal.

7 8

9 1 3 6

10 2 4 11

12

Hint: use a tree-decomposition with four nodes.

4.8 Exercises 123

P
12. Consider an instance of BPC such that i2I `i  1, and G D .I; E/ is co-
interval. Propose a BPC-DDFF for this problem that always leads to an optimal
solution.
Hint: the problem is a well-known polynomial case of one of the most famous hard
combinatorial problems.
Chapter 5
Other Applications in General Integer
Programming

In this chapter, we briefly review an alternative application of dual-feasible functions

in general integer programming. We explore these functions in particular to derive
valid inequalities for integer programs. Since the notion of superadditivity is
essential for this purpose, we start by reviewing superadditivity in the scope of
valid inequalities. Different examples are provided with alternative families of dual-
feasible functions. We discuss also the difference between the valid inequalities
derived by dual-feasible functions and the well-known Chvátal-Gomory cuts.

5.1 Superadditive Functions in Integer Programming

When good dual-feasible functions are sought, they are frequently characterized by
superadditivity and monotonicity. For the sake of clarity, these properties are briefly
recalled in the sequel for general domains. Given X  Rm , a function F W X ! R is
superadditive if for all x; y 2 X with x C y 2 X, it holds that

F.x/ C F.y/  F.x C y/:

The function F W X ! R is nondecreasing if for all x; y 2 X, one has that

x  y H) F.x/  F.y/:

Analogously to the duality theory of linear optimization (without integer con-

straints), there is also a strong duality theorem in the discrete counterpart. Here
superadditive and nondecreasing functions are essential. Given A 2 Qmn , b 2 Qm
and c 2 Qn and an integer linear optimization problem

© Springer International Publishing Switzerland 2016 125

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5_5
126 5 Other Applications in General Integer Programming

max c> x (5.1)

s: to Ax  b (5.2)
x2 ZnC ; (5.3)

the dual consists in determining a nondecreasing and superadditive function F W

Rm ! R such that

min F.b/ (5.4)

s: to F.o/ D 0 (5.5)
F.a /  cj for j D 1; : : : ; n;
j
(5.6)

where aj is the j-th column of the matrix A, and o denotes as usual the zero vector.
As mentioned before, the relationship between the two problems is given by the
strong duality theorem. If the problem (5.1)–(5.3) is solvable, i.e. a feasible x exists
and the objective function value is bounded from above, then the optimal objective
function values of the primal and the dual problem are equal. If F.b/ is unbounded
from below (1), then there is no x 2 ZnC fulfilling (5.2), and if (5.1)–(5.3) yields
an unbounded objective function value (C1) then F does not exist. Furthermore,
if xO is an optimal solution of (5.1)–(5.3) then any optimal F in the dual problem
necessarily obeys the equation

F.Ax/ D c> x D F.b/  F.b  Ax/

for all x 2 ZnC with x  xO .

Since the problem (5.1)–(5.3) is NP-hard, solving it exactly or finding an optimal
superadditive and nondecreasing function F W Rm ! R according to the above
conditions (5.4)–(5.6) may be very difficult. However, contributing to the resolution
of (5.1)–(5.3) is possible by deriving valid inequalities using a superadditive and
nondecreasing function F W Rm ! R fulfilling the constraints (5.5). These functions
lead to the following valid inequalities:

X
n
F.aj /  xj  F.b/:
jD1

5.2 Valid Inequalities for Integer Programs

General dual-feasible functions can be used not only to compute fast lower bounds,
but also to generate valid inequalities for integer problems defined over sets of the
kind fx 2 ZnC W Ax  bg, as stated formally in the sequel.
5.3 Examples 127

Proposition 5.1 If f is a maximal general dual-feasible function P (and hence a

superadditive function with f .0/ D 0) and S D fx 2 ZnC W njD1 aij xj  bi ; i D
P
1; : : : ; mg, then for any i, njD1 f .aij /xj  f .bi / is a valid inequality for S.
In Chaps. 2 and 3 it turned out that (general) dual feasible functions should be
superadditive to get good lower bounds, because otherwise another dominating
function exists. In the context of valid inequalities for integer problems, the same
applies. Any valid inequality for S can be obtained either through a superadditive
function or it is dominated by an inequality that can be computed in this way. Cuts
generated by a superadditive function are commonly referred to as superadditive
inequalities. Among these cuts, those that are not dominated by any other valid
inequality are called maximal. This applies particularly to the facets of the integer
hull of S. Maximal valid inequalities are necessarily superadditive. The same
properties characterize the dominant families of dual-feasible functions.
Any maximal inequality for S can be obtained through the Gomory procedure
based on recursive linear combinations and rounding of other inequalities for S.
However, in order to get these maximal cuts, it might be necessary to use a very
long recursion. Other authors assumed that other superadditive functions, eventually
more complex ones, might generate these maximal cuts using shorter recursions,
demonstrating the relevance of research on dual-feasible functions as tools to
compute valid inequalities for integer programs. Other works propose alternative
characterizations of the integer hull of S in terms of a finite set of superadditive
inequalities, but the cardinality of this set may be very large, such that many cuts
are required.

5.3 Examples

In this section, we show through different examples how to apply several generalized
dual-feasible functions to derive valid inequalities, which may be better than the
well-known Chvátal-Gomory cuts.
Given the inequality system Ax  b, x 2 Nn , any nonnegative linear
combination of the inequalities may be used to derive cuts. Choosing u  o yields
u> Ax  u> b, hence the following scalar inequality

d> x  r; (5.7)

and finally the Chvátal-Gomory cut

X
n
bdj c  xj  brc: (5.8)
jD1
128 5 Other Applications in General Integer Programming

If r > 0, then dividing the inequality (5.7) by r and applying a maximal general
dual-feasible function f W R ! R with f .1/ D 1 leads to the valid inequality
Xn  
dj
f  xj  1: (5.9)
jD1
r

In some situations it may happen that the inequality (5.9) is stronger than (5.8),
but not if 0 < r < 1, as the right-hand sides show immediately. The following
example demonstrates the strength of a maximal general dual-feasible function for
the construction of valid inequalities, even in the case r D 1.
Example 5.1 We use the function (3.7), p. 65, with the parameter b WD 1, hence
8
< b2dc; if d < 1=2;
f .d/ D 1=2; if d D 1=2; (5.10)
:
d2de  1; if d > 1=2:

Let be given the inequality

1:6x1  0:4x2  1:

Of course, if there would be no negative coefficient then a coefficient dj > 1 would

immediately imply xj D 0 for that j. The Chvátal-Gomory cut (5.8) transforms the
given inequality to x1  x2  1, but the function (5.10) used in (5.9) leads to the
stronger inequality 3x1  x2  1. This inequality would be obtained by the Chvátal-
Gomory procedure only after multiplying the given inequality by a suitable number
like 1.9. t
u
Example 5.2 The dual-feasible function (2.10) is used for deriving valid inequali-
ties. It was defined as

x; if .k C 1/  x 2 N;
fFS;1 .xI k/ WD
b.k C 1/  xc=k; otherwise:

and is again illustrated in Fig. 5.1. Next, we show the result of applying the function
fFS;1 with parameters k 2 f1; 2; 3g to a given knapsack inequality (after dividing it
by the right-hand side):

9x1 C 7x2 C 6x3 C 4x4 C 2x5  12

ai =b 0:75 0:58 0:5 0:33 0:17

k D 1 1x1 C 1x2 C 12 x3  1
1
k D 2 1x1 C x
2 2
C 12 x3 C 1
x
3 4
 1
3 2
kD3 x
4 1
C x
3 2
C 12 x3 C 1
x
3 4
 1

t
u
5.3 Examples 129

fFS,1 (x;1) fFS,1 (x;2)

1 x 1 2 x
2 3 3
fF S,1 (x;3) fFS,1 (x;4)

1 1 3 x 1 2 3 4 x
4 2 4 5 5 5 5
Fig. 5.1 MDFF fFS;1 .I k/ for k 2 f1; : : : ; 4g

Given multiple knapsack constraints like in the vector packing problem, one may
use a VP-MDFF to get cuts.
Example 5.3 Neglecting the integrality constraints in

8x1 C 5x2 C 5x3  10

5x1 C 8x2 C 5x3  10
x1 ; x2 ; x3 2 f0; 1g

would allow the fractional solution x D .0:5; 0:8; 0:2/>, because

8  0:5 C 5  0:8 C 5  0:2 D 4 C 4 C 1 D 9  10

and

5  0:5 C 8  0:8 C 5  0:2 D 2:5 C 6:4 C 1 D 9:9  10:

130 5 Other Applications in General Integer Programming

However, taking e.g. the VP-MDFF (4.9), p. 106, with the parameter u WD . 12 ; 12 /> ,
which considers all constraints simultaneously, yields the valid inequality

x1 C x2 C 0:5x3  1;

which is violated by the fractional solution, since

0:5 C 0:8 C 0:5  0:2 D 1:4 > 1:

t
u
The following example illustrates the generation of valid inequalities with dual-
feasible functions obtained from bin packing problems with conflicts.
Example 5.4 Consider the following system of inequalities:

4x1 C 3x2 C 3x3  10

x2 C x3  1
x1 ; x2 ; x3 2 f0; 1g

The first inequality does not restrict anything. However, a BPC-DFF yields the
valid inequality

0:9x1 C 0:1x2 C 0:1x3  1:

The DFF used is the one defined in Proposition 4.17, with J D f2; 3g, ˛2 D 0:1 and
˛3 D 0:1. t
u

5.4 Related Literature

In Nemhauser and Wolsey (1998), the relationship between the integer linear opti-
mization problem (5.1)–(5.3) and its dual (5.4)–(5.6) and a revision of superadditive
valid inequalities are provided, together with the basic function underlying the
Chvátal-Gomory procedure. Previous results on the (explicit or implicit) use of dual-
feasible functions to generate valid inequalities for integer programs were reported
by Vanderbeck (2000), Alves (2005), Rietz et al (2014), Letchford and Lodi (2002),
and Dash and Günlük (2006).
5.5 Exercises 131

5.5 Exercises
P
n
1. Let be given the feasible region aij xj  bi , i D 1; : : : ; m, x 2 Nn for an
jD1
integer linear optimization problem. Which of the following assertions are true?
Justify your answer.
P
n
(a) f .aij /xj  f .bi / is a valid inequality for every general DFF f W R ! R.
iD1
Pn
(b) f .aij /xj  f .bi / is a valid inequality for every superadditive function
iD1
f W R ! R.
2. Consider the set of all ordered triplets .x1 ; x2 ; x3 / 2 Z3C , for which

5x1 C 4x2 C 3x3  11

3x1 C 4x2 C 2x3  8

holds. Derive the valid inequality

2x1 C 2x2 C x3  4

using the VP-MDFF fI;FS;1 of Corollary 4.1, p. 102, with the parameter choice v WD
.3; 2/> .
3. Let w WD .1; : : : ; 1/> 2 Rm . Given the inequality system

X
n
aj xj  wI x  o;
jD1

where aj 2 Œ0; 1m for all j, suppose that a VP-MDFF f W Œ0; 1m ! Œ0; 1 yields
f .aj0 / D 1 for a certain j0 2 f1; : : : ; ng. Why does this imply f .aj / D 0 for all those
j, for which aj0 C aj  w? What is the conclusion for the possible usage of the
functions of Classes II, III and IV?
Appendix A
Hints and Solutions to Selected Exercises

Chapter 1

1.1. As stated in the text, the matrix A is just the vector .1; 3/ and the vector c D
.1; 1/> .
(a) The extreme points of X are x1 D .0; 0/> , x2 D . 52 ; 0/> , x3 D .0; 4/>
with X D fx 2 R2 W x D 1 .0; 0/> C 2 . 52 ; 0/> C 3 .0; 4/> ; 1 C
2 C 3 D 1; 1 ; 2 ; 3  0g. Therefore, Œc> x1 ; c> x2 ; c> x3  D Œ0; 52 ; 4 and
ŒAx1 ; Ax2 ; Ax3  D Œ0; 52 ; 12. The DW-model is: zDW WD maxf01 C 52 2 C43 W
01 C 52 2 C 123  6; 1 C 2 C 3 D 1; 1 ; 2 ; 3  0g. The optimal
solution is .1 ; 2 ; 3 / D .0; 12 ; 7 /, which maps to the solution in the original
19 19
30 28 >
space x D . 19 ; 19 / .


(b) zIP D 2; zDWI D 14 58

5 D 2:8; zDW D zLP D 19 D 3:053.
 

1.2. Check Fig. A.1 to identify the extreme points of the sets X 1 and X 2 , respec-
tively. The matrices and vectors in the reformulated model are as follows:
 
1> 1 1>
 1 1 1 1   0220  
c X Dc x1 x2 x3 x4 D 3 5  D 0 6 11 10
0012
     
 1 1 1 1 12 0220 0244
A1 X1 1
D A x1 x2 x3 x4 D  D
32 0012 0684
 
> >     010  
c 2 X2 D c2 x21 x22 x23 D 1 2  D 016
003
     
  21 010 023
A2 X2 D A2 x21 x22 x23 D  D :
11 003 013

© Springer International Publishing Switzerland 2016 133

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5
134 Appendix A

x12 x22

x23 =
0
3

x14 =
0
2

x13 = 2
1

X1 X2

x11 = 0
x12 = 2 x11 x21 = 0
x22 = 1 x21
0 0 0 0

Fig. A.1 Sets X 1 and X 2

The original model with a block angular structure is reformulated into the
following DW-model:

max zDW WD 612 C 1113 C 1014 C 122 C 623

s: to 212 C 413 C 414 C 222 C 323  6
612 C 813 C 414 C 122 C 323  8
11 C 112
1
C 113 C 114 D1
121 C 122 C 123 D 1
11 ; 12 ; 13 ; 14 ; 21 ; 22 ; 23  0:

The optimal solution is .11 ; 12 ; 13 ; 14 ; 21 ; 22 ; 23 / D .0; 0; 12 ; 12 ; 13 ; 0; 23 /;
and zDW D 29 2 : In terms of the original variables, the optimal solution
is .x11 ; x12 ; x21 ; x22 / D .1; 32 ; 0; 2/; and z D 29
2
:
1.3. First, we address the decomposition with the constraints (1.18) in the subprob-
lem.
(a) Let X i D fxi1 ; : : : ; xiki g be the set of all feasible assignments of jobs to
machine i; 8i 2 I: A feasible assignment is a vector xik D .xi1k ; xi2k ; : : : ; xijJjk / 2
BjJj ; whose elements indicate if the job is in the plan. The variables yik ; k 2
Ki ; 8i 2 I in the reformulated model are defined as:

1; if feasible assignment xik is used in machine i
yik D
0; otherwise:

The reformulated model is as follows:

X X
max z WD . pij xijk /yik
i2I;k2Ki j2J
Appendix A 135

X
s: to xijk yik D 1; j 2 J
i2I;k2Ki
X
yik  1; i 2 I
k2Ki

yik 2 f0; 1g; k 2 Ki ; i 2 I

The block angular structure of the model leads to m knapsack subprob-

lems, one for each machine. Concerning the other decomposition, with the
constraints (1.19) in the subproblem, the reader may check that the reformulated
model is the LP relaxation of the model (1.17)–(1.20).
(c) It is the model presented in Example 1.4.
(d) The lower bound is equal to 35. The reader may check that the optimal solution
of the LP relaxation of the original is equal to 28.75.
1.4.
(b) The solution set of the subproblem is the set of all feasible circulation flows that
are solutions of the set of constraints (1.22). A circulation flow is a flow along
a directed cycle composed of arcs xij and the arc xL0 :
The only extreme point is the null solution, with xij D 0; 8.i; j/ 2 A; and xL0 D
0:
Each extreme ray corresponds to the set of all feasible flows along a single path
and the feedback arc xL0 : For instance, starting at the null solution, the points
with x04 D x48 D x80 D ; for   0; define one of the extreme rays of the
solution set.
(c) It is the model in Example 1.3.
1.5. The knapsack constraint for roll k; which defines a bounded set, is replaced
by the convex hull of the integer extreme solutions to the knapsack polytope. These
p p p
solutions are described by a vector .x1k ; : : : ; xik ; : : : ; xjIjk /T , 8k 2 K, 8p 2 P, where
P is the set of feasible cutting patterns.
The reformulated model is as follows:
XX p
min z WD k (A.1)
k2K p2P
XX p p
s: to xik k  di ; 8i 2 I (A.2)
k2K p2P
X p
k  1; 8k 2 K (A.3)
p2P
X p p
xik k  0; and integer; 8i 2 I; 8k 2 K (A.4)
p2P
136 Appendix A

X p
k 2 f0; 1g; 8k 2 K (A.5)
p2P
p
k  0; 8p 2 P; 8k 2 K; (A.6)
p
where the decision variables k denote the number of times pattern p is cut in roll k:
Constraints (A.2) guarantee that the demand for each item is satisfied by the
patterns selected for the jKj rolls. The convexity constraint for roll k (A.3) enforces
that the solution is a convex combination of the extreme points of the knapsack
polytope. The null solution is also an integer extreme point. As the corresponding
column has the structure of a slack column, the convexity constraint is of the type  :
This is consistent with the model, because some rolls may not be cut. When applying
the decomposition, the integrality requirements on the yk ; 8k 2 K; in the original
model, which are translated into constraints (A.5), cannot be dropped, as otherwise
the integrality requirements (A.4) may not be sufficient to provide integer solutions
to the cutting stock problem [see Valério de Carvalho (2002)].
The integrality requirements ensure that we pick a feasible cutting pattern for
each roll k; and, as the rolls have equal length, the index k can be dropped and the
p p p
solutions described by a vector .x1 ; : : : ; xi ; : : : ; xjIj /T ; 8p 2 P; leading to Gilmore
and Gomory model [see Vance (1998) for further details].

Chapter 2

2.1. The function f1 is a dual-feasible function, because 0  f1 .x/  x for all

x 2 Œ0; 1. f2 is not a DFF, because f2 .1=3/ D 1=3 and f2 .2=3/ D 1, hence f2 .1=3/ C
f2 .2=3/ > 1 in spite of 1=3 C 2=3  1.
2.2. Choose any x; y 2 R. One gets bf .x/c C bf .y/c  bf .x/ C f .y/c  bf .x C y/c,
because bc is non-decreasing.
2.3. The function f3 is superadditive, while f4 is not. The square of nonnegative
numbers is superadditive, because x; y  0 implies .x C y/2  x2  y2 D 2xy  0.
The superadditivity of x 7! x2 is preserved by the rounding-down. The function f4
is not superadditive, because f4 .0/ > 0.
2.4.
x 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
(a)
g1 .x/ 080 0 0 1 2 3 4 5 5 5 5 6 7 8 9 10 10 10 10
ˆ
ˆ 0; if 0  x < 4=19;
ˆ
ˆ
ˆ
< b19x  3c=10; if 4=19  x  8=19;
(b) f5 .x/ D 1=2; if 8=19 < x < 11=19;
ˆ
ˆ
ˆ
ˆ b19x  6c=10; if 11=19  x  16=19;
:̂
1; if 16=19 < x  1;
Appendix A 137

8
ˆ
ˆ 0; if 0  x < 3=19;
ˆ
ˆ
ˆ
< .19x  3/=10; if 3=19  x  8=19;
(c) f6 .x/ D 1=2; if 8=19 < x < 11=19;
ˆ
ˆ
ˆ
ˆ .19x  6/=10; if 11=19  x  16=19;
:̂
8 1; if 16=19 < x  1;
ˆ
ˆ 0; if 0  x < 4=19;
ˆ
ˆ
ˆ
< b19x  3c=10; if 4=19  x  8=19;
(d) f7 .x/ D 1=2; if 8=19 < x < 11=19;
ˆ
ˆ
ˆ d19x  6e=10; if 11=19  x  16=19;
ˆ
:̂
1; if 16=19 < x  1:
2.5.
(a) False.
There are many counter examples. For instance, if f fFS;1 .I 2/ and `i < L=2
P
m
for i D 1; : : : ; m and b > o, then bi  f .`i =L/ D 0, which is less than the
iD1
P
m
continuous bound bi  `i =L, which is obtained by the identity function for g.
iD1
(b) True.
.kC1/kL
Setting C WD kLC1 leads to fBJ;1 . Ln I C/ D fFS;1 . Ln I k/ for n D 0; 1; : : : ; L.
Proof Both functions fFS;1 and fBJ;1 are maximal dual-feasible functions. Hence, the
assertion needs to be verified for 0 < Ln < 12 only. Since L; n 2 N n f0g, that allows
the assumption L  3.
p
Let p 2 N with p  k and Ln
kC1 kLk
. Since C D k C kLC1 2 .k; k C 1/, we obtain

 ( )!
n .k C 1/  kn .kL C 1/  frac. .kC1/kn /  kL C k
kLC1
fBJ;1 D C max 0; =k:
L kL C 1 1Ck

Three cases have now to be distinguished:

1. n D Lp
kC1
> 0 yields

.k C 1/  kn kLp kL C 1  p
D D p1C 2 .p  1; p/;
kL C 1 kL C 1 kL C 1

and hence
n   
kL C 1  p  kL C k
fBJ;1 D p  1 C max 0; =k
L kC1
 
kC1p
D p1C =k
kC1
138 Appendix A

kp  k C p  1 C k C 1  p
D
.k C 1/  k
p
D
kC1
 
p
D fFS;1
kC1
n
D fFS;1 :
L

2. n < Lp
kC1
: Since k; L; n; p are integers, it follows that n  Lp1
kC1
. We get

Lp  1 k  .Lp  1/ kL C 1  k  p
C D D p1C 2 .p  1; p/;
.k C 1/  L kL C 1 kL C 1

because the nominator of the last fraction equals

k  .L  1/ C 1  p  2k C 1  p > 0:

Therefore,
n  
Lp  1
fBJ;1  fBJ;1
L .k C 1/  L
  
kL C 1  k  p  kL C k
D p  1 C max 0; =k
kC1
  
1p
D p  1 C max 0; =k
kC1
p1
D
k
 
Lp  1
D fFS;1 :
.k C 1/  L
Lp LpC1
3. n > kC1 : we get n  kC1 and

Lp C 1 k  .Lp C 1/
C D
.k C 1/  L kL C 1
kp
D pC 2 Œp; p C 1/;
kL C 1
Appendix A 139

and hence
n  
Lp C 1
fBJ;1  fBJ;1
L .k C 1/  L
  
k  p  kL C k
D p C max 0; =k
kC1
  
k  .2  L/  p
D p C max 0; =k
kC1
p
D
k
 
Lp C 1
D fFS;1 :
.k C 1/  L

Lp
We get for kC1 < n < L.pC1/ from the combination of the second and third case
p pC11
kC1

k  fBJ;1 . L / 
n
k , and hence fBJ;1 Ln D pk and analogously pk D fFS;1 Ln
because of the monotonicity. t
u
(c) True.
1
k
2 N with k  2
and any L 2 N n f0g, the use of C WD k 
For any L leads to
fVB;2 Ln I k D fCCM;1 Ln I C for all n 2 N with n  L.
Proof Since fVB;2 and fCCM;1 are maximal dual-feasible functions, it is enough to
verify the proposition for 0 < n < L=2. Then, it follows that
n  
kn
fVB;2 D  1 =.k  1/
L L

and
n 
Cn
fCCM;1 D =.k  1/:
L L

For any x 2 R n Z, it holds that dx  1e D bxc. One has

 
Cn kn n kn  1 kn
D  2 2 ; ;
L L L L L

because 0 < n < L2 .

If Ljkn, then
n  
kn
fVB;2 D  1 =.k  1/
L L
140 Appendix A

and

Cn kn
D  1;
L L

i.e.,
n   n
kn
fCCM;1 D  1 =.k  1/ D fVB;2 :
L L L
n
 kn ˘  kn1 ˘ n

If L 6 jkn, then fVB;2 L

D L
=.k  1/ D L
=.k  1/ D fCCM;1 L
. t
u
(d) False.
Choose any p 2 N with p  3. Let k WD 3, L WD 3p, m  2, `1 WD p and
`2 WD p C 1. Suppose there is a C 2 R with C  1 and
   
pC1 pC1
fVB;2 I 3 D fBJ;1 IC
3p 3p

and
   
p 1
fVB;2 I 3 D fBJ;1 IC :
3p 3

The latter yields

  
1 C
fVB;2 I3 D 0 D ;
3 3

i.e.
C
1  C < 3; and  frac.C/:
3

The case 2  C < 3 leads to frac.C/ D C2  C3 , and hence 23 C  2, implying

the contradiction C  3. Therefore, 1  C < 2 and frac.C/ D C  1  C3 such
that 32  C < 2. One has

   ( )
pC1 pC1 frac. pC1
3p  C/  C C 1
fBJ;1 IC D  C C max 0;
3p 3p 2C
( pC1 )
3p
CCC1
D 0 C max 0; :
2C
Appendix A 141

Since p  3 and 32  C < 2, one gets .p  2/  .3  2C/  0, and hence

p  .3  2C/  6  4C. Therefore, C  .1  2p/ C 3p  6  3C and
pC1
3p
CCC1 C  .p C 1/  3Cp C 3p
D
2C 3p  .2  C/
C  .1  2p/ C 3p
D
p  .6  3C/
1

p
 
in contradiction to fVB;2 pC1
3p I 3 D 12 .

2.6. The proof is given next.

Proof One has obviously fBJ;1 .0/ D 0 and fBJ;1 .x/  0 for all x 2 Œ0; 1. To show
the symmetry, choose any x 2 Œ0; 1. We get either frac.Cx/  frac.C/ and hence

frac.C  Cx/ D frac.C/  frac.Cx/  frac.C/;

fBJ;1 .x/ D bCxc=bCc;
fBJ;1 .1  x/ D bC  Cxc=bCc
D .bCc  bCxc/ =bCc;

or one has frac.Cx/ > frac.C/, and consequently frac.C  Cx/ D frac.C/ C 1 
frac.Cx/ > frac.C/ and

fBJ;1 .x/ C fBJ;1 .1  x/ D

 
frac.Cx/  frac.C/ frac.C  Cx/  frac.C/
D bCxc C bC  Cxc C C =bCc
1  frac.C/ 1  frac.C/

D bCxc C bCc  bCxc  1 C

frac.Cx/  frac.C/ C frac.C/ C 1  frac.Cx/  frac.C/
=bCc
1  frac.C/
 
1  frac.C/
D 1C  1 =bCc
1  frac.C/
D 1:

To show the superadditivity, take any x; y 2 Œ0; 1 with x C y  1 and, without

loss of generality, take also frac.x/  frac.y/. Let

d WD .fBJ;1 .x C y/  fBJ;1 .x/  fBJ;1 .y//  bCc:

142 Appendix A

We have to show d  0. Two cases are distinguished.

• If frac.Cx/Cfrac.Cy/ < 1, then bCxcCbCyc D bCxCCyc and frac.CxCCy/ D
frac.Cx/ C frac.Cy/, such that
   
frac.Cx/ C frac.Cy/  frac.C/ frac.Cx/  frac.C/
d D max 0;  max 0; 
1  frac.C/ 1  frac.C/
 
frac.Cy/  frac.C/
max 0; :
1  frac.C/

If frac.Cx/  frac.C/, then d  0 becomes obvious. Otherwise

frac.Cx/ C frac.Cy/  frac.C/ frac.Cx/  frac.C/ frac.Cy/  frac.C/

dD  
1  frac.C/ 1  frac.C/ 1  frac.C/
frac.C/
D
1  frac.C/
 0:

• If frac.Cx/ C frac.Cy/  1, then bCxc C bCyc D bCx C Cyc  1 and frac.Cx C

Cy/ D frac.Cx/ C frac.Cy/  1. Hence, we have
 
frac.Cx/ C frac.Cy/  1  frac.C/
d D bCc C max 0;
1  frac.C/
   
frac.Cx/  frac.C/ frac.Cy/  frac.C/
 max 0;  max 0; :
1  frac.C/ 1  frac.C/

frac.Cy/frac.C/
If frac.Cx/  frac.C/, then d > bCc  1  0, because 1frac.C/
<
1frac.C/
1frac.C/
D 1. If frac.Cx/ > frac.C/, then

frac.Cx/ C frac.Cy/  1  frac.C/

d  bCc C
1  frac.C/
frac.Cx/  frac.C/ frac.Cy/  frac.C/
 
1  frac.C/ 1  frac.C/
frac.C/  1
D bCc C
1  frac.C/
D bCc  1
 0:

t
u
Appendix A 143

2.7. The function f is a maximal dual-feasible function, but not extreme, because
it is symmetric and strict convex on Œ0; 1=2 and also repeatedly differentiable in
.0; 1=2/.
2.8. The proof is given next.
Proof Let g; h W Œ0; 1 ! Œ0; 1 be maximal dual-feasible functions with 2 
fFS;1 .x/ D g.x/ C h.x/ for all x 2 Œ0; 1. One has to show that fFS;1 .x/ D g.x/
for all x 2 .0; 1=2/ with 0 < fFS;1 .x/ < 1=2. Therefore assume k > 1. Since g; h are
dual-feasible functions, it follows that
 
1
.k C 1/  g 1
kC1

and
 
1 1
h  :
kC1 kC1
1
1
1
1
p

Since fFS;1 kC1 D kC1 , one gets g kC1 D kC1 , implying g kC1  p for any
p
kC1 p
p 2 f1; 2; : : : ; kg due to the monotonicity of g. Analogously h kC1  kC1 , and
p p
hence g. kC1 / D f . kC1 /. It remains to show g.x/ D f .x/ if .k C 1/  x … N. There
is a p 2 N with

0 < p < x  .k C 1/ < p C 1

and p < k=2, such that fFS;1 .x/ D p=k. Let x0 WD p=k. Because of 0 < p < k=2 < k,
it holds that

p < p  .k C 1/=k D x0  .k C 1/ < p C 1;

i.e. fFS;1 .x0 / D x0 . Similarly to the other case g 1k D 1k D fFS;1 1k and g.x0 / D x0
follows. Since x0 is in the inner of an open interval on which fFS;1 is constant, g must
be constant in the same interval. Otherwise the monotonicity of g and h would yield
a contradiction. This implies

g.x/ D g.x0 / D x0 D fFS;1 .x0 / D fFS;1 .x/

p
for all x 2 . kC1 ; pC1
kC1 /, and hence g fFS;1 . t
u
2.9.
(a) True.
If f is not extreme, then there are different maximal dual-feasible functions
f1 ; f2 W Œ0; 1 ! Œ0; 1 with 2f f1 C f2 . There is an x0 2 Œ0; 1 with
f1 .x0 / ¤ f2 .x0 /. Since g is surjective, there is an x1 2 Œ0; 1 with g.x1 / D x0
that yields f1 .g.x1 // D f1 .x0 / ¤ f2 .x0 / D f2 .g.x1 //. Hence, f1 .g.// and
144 Appendix A

f2 .g.// are different functions. Moreover, for all x 2 Œ0; 1, it holds that
2f .g.x//  f1 .g.x//  f2 .g.x// D 0.

(b) True.
If f is continuous, then the image of the interval Œ0; 1, which is a connected
set, must be connected, and hence an interval. Since f .0/ D 0 and f .1/ D 1,
this image is the interval Œ0; 1, such that f is surjective. The opposite direction
consists in showing the continuity of f , provided that f is surjective. The
function f is continuous at the point zero, because for all x 2 .0; 1, it holds
that

0  f .x/  1=b1=xc:

Hence, lim f .x/ D 0. The symmetry of f implies also the continuity at 1. Let
x#0
xN 2 .0; 1/ be an arbitrary constant. Let .xn / and .yn / be any sequences with

0  x1  x2      xN      y2  y1  1

and

lim xn D lim yn D xN :
n!1 n!1

The monotonicity of f implies

0  f .x1 /  f .x2 /      f .Nx/      f .y2 /  f .y1 /  1:

Every monotone and bounded sequence converges, hence the left and right
limits of f at xN exist, and it holds that

lim f .x/  f .Nx/  lim f .x/:

x"Nx x#Nx

Since f is surjective, it cannot happen that lim f .x/ < f .Nx/ or lim f .x/ > f .Nx/.
x"Nx x#Nx

(c) False.
The function fFS;1 .I 1/ defined in (2.10) (p. 35), is a counter example. This
function is an extreme maximal dual-feasible function and convex on [0,1/2],
but not continuous.
2.10.
(a) The function fMT;0 is obviously symmetric, monotone and nonnegative. Accord-
ing to Theorem 2.2, only the superadditivity remains to be proved. Choose
any x; y 2 .0; 1=2/ with x  y. If x < , then fMT;0 .x/ D 0, such that the
Appendix A 145

superadditivity follows from the monotonicity. If x  , then

fMT;0 .x/ C fMT;0 .y/ D x C y  fMT;0 .x C y/:

(b) Let g; h W Œ0; 1 ! Œ0; 1 be any maximal dual-feasible functions with 2fMT;0
g C h. We show g.x/ D h.x/ for all x. If x < , then fMT;0 .x/ D 0 implies
immediately g.x/ D h.x/ D 0 due to the nonnegativity of g and h. It holds for
all x; y 2 Œ; 1
2  that

fMT;0 .x C y/ D x C y D fMT;0 .x/ C fMT;0 .y/:

Since fMT;0 .1=4/ D 1=4 and fMT;0 .1=3/ D 1=3, it follows that g.1=4/ D
h.1=4/ D 1=4 and g.1=3/ D h.1=3/ D 1=3, because larger function values
would violate the definition of a dual-feasible function. Now Lemma 2.4 can
be applied with a WD 1=4 and b WD 1=3, yielding g.x/ D h.x/ D x for all
x 2 Π14 ; 13 . The superadditivity of g and h implies g.2x/  2g.x/ D 2x and
h.2x/  2x for all these x, andhence  D h.2x/ D 2x. The symmetryyields
 g.2x/ 
g.x/ D h.x/ D x for all x 2 13 ; 12 [ 23 ; 34 too, and hence for all x 2 14 ; 34 .
3 
Moreover, for x 2 4 ; 1   , one obtains g.x=2/ D h.x=2/ D x=2 and again
due to the superadditivity g.x/ D h.x/ D x. Finally the symmetry yields
g h. t
u
2.11. The proof is given next.
Proof Assume without loss of generality ˛  ˇ. If ˛ D ˇ, then, according to
Definition 2.6, the different maximal dual-feasible functions f ; g immediately yield
that h is not extreme. Otherwise define the function h1 W Œ0; 1 ! Œ0; 1 by

2˛ ˇ˛
h1 .x/ WD 2h.x/  g.x/ D f .x/ C g.x/:
˛Cˇ ˛Cˇ

That yields 2h g C h1 , i.e. if h1 is proved to be a maximal dual-feasible function

different from g, then the proof is complete. Since 0 < ˛ < ˇ, both factors
2˛ ˇ˛
˛Cˇ
and ˛Cˇ are positive, such that h1 is a maximal dual-feasible function due
to Proposition 2.2. Since f 6 g, there is an x0 2 Œ0; 1 with f .x0 / ¤ g.x0 /. Suppose
h1 .x0 / D g.x0 /. That assumption would imply

2˛ 2˛
f .x0 / D g.x0 /
˛Cˇ ˛Cˇ

in contradiction to f .x0 / ¤ g.x0 / and ˛ > 0. t

u
146 Appendix A

2.12. Consider the following functions

8
< 1=2; if x D 2=5;
f .x/ WD 1; if x D 2=3;
:
0; otherwise,

and
8
< 2=5; if x D 1=2;
g.x/ WD 2=3; if 1=2 < x  1;
:
0; otherwise.

Both f and g are dual-feasible functions, but neither symmetric nor superadditive.
Their composition yields
8
< 0; if 0  x < 1=2;
f .g.x// D fFS;1 .xI 1/ D 1=2; if x D 1=2;
:
1; if 1=2 < x  1:

for any x 2 Œ0; 1.

2.13. The function fLL;1 is symmetric if and only if k D 2.
Proof If k D 2, then choose any x 2 Œ0; 1. If frac.Cx/  frac.C/, then frac.C 
Cx/  frac.C/, and hence

bCxc C bC  Cxc
fLL;1 .x/ C fLL;1 .1  x/ D
bCc
bCxc C bCc  bCxc
D
bCc
D 1:

If frac.Cx/ > frac.C/, then frac.C  Cx/ > frac.C/ and

  
frac.Cx/  frac.C/
fLL;1 .x/ C fLL;1 .1  x/ D bCxc C bC  Cxc C =2
1  frac.C/
  
frac.C  Cx/  frac.C/
C =2 =bCc
1  frac.C/
D .bCxc C bCc  1  bCxc C 1=2 C 1=2/ =bCc
D 1:
Appendix A 147

 
1frac.C/
If k > 2, then let x WD frac.C/ C k1 =C. Then x > 0 is obvious. It holds
1frac.C/
also that k1
< 1, and hence x < .frac.C/ C 1/=C  1. We show

fLL;1 .x/ C fLL;1 .1  x/ < 1:

First, one gets

1  frac.C/ 1  frac.C/
Cx D frac.C/ C  frac.C/ C
k1 2
1 C frac.C/
D <1
2

and frac.Cx/ > frac.C/. That implies

1  frac.C/
0 < frac.C  Cx/  frac.C/ D 1  frac.Cx/ D 1  frac.C/ 
k1
1  frac.C/
D  .k  2/
k1
and
  
frac.Cx/  frac.C/
fLL;1 .x/ C fLL;1 .1  x/ D bCxc C bC  Cxc C  .k  1/ =k
1  frac.C/
  
frac.C  Cx/  frac.C/
C  .k  1/ =k =bCc
1  frac.C/
 
k2
D bCxc C bCc  1  bCxc C 1=k C =bCc
k
1
D 1 < 1:
k  bCc

t
u
2.14. The proof is given next.
Proof Let x WD 1=k. Then, x 2 Œ0; 1 and fVB;1 .x/ D 0, and hence fVB;1 .1  x/ D
k2
k1
< 1. t
u
2.15.
(a) False.
Let for example C WD 1:9, x WD 0:3 and g be the maximal dual-feasible
function (2.10) with k D 2. Then, f .x/ D g.x/ D 1, f .2x/ D 1 < 2  f .x/.

(b) False.
148 Appendix A

Let e.g. C WD 1, x WD 8, y WD 2 and g be the identity function. One gets

f .x/ D 0, f .y/ D 1, f .x C y/ D f .6/ D 0 < f .x/ C f .y/.
(c) True.
Pn
If x1 ; : : : ; xn 2 RC with xi  C1 are given, then using each summand C
iD1
times, i.e. setting xnpCi WD xi for i D 1; : : : ; n and p D 1; : : : ; C  1, yields

X
Cn X
n
xi D C  xi  1:
iD1 iD1

Hence, according to Definition 2.1, we have

X
Cn X
n
f .xi / D C  xi  1:
iD1 iD1

Chapter 3

3.1. The proof is given next.

Proof If f is a general dual-feasible function and x 2 .0; 1, then let n WD b1=xc 2
N n f0g. Setting

x1 WD x2 WD    WD xn WD x

yields

X
n
xi D nx  1
iD1

and

X
n
f .xi / D n  f .x/  1
iD1

due to Definition 3.1. Hence, we have f .x/  1=n.

P
n
Given n 2 N and xi 2 R (i D 1; : : : ; n) with xi  0, it cannot happen that
iD1

X
n
f .xi / D " > 0;
iD1
Appendix A 149

because using each xi in the quantity p WD b1="c C 1, i.e. setting

xnC1 WD x1 ; : : : ; xnp WD xn ;

one gets

X
np
X
n
xi D p  xi  0;
iD1 iD1

but

X
np
f .xi / D p  " > 1
iD1

in contradiction to Definition 3.1, and the fact that f is a dual-feasible function. t

u
3.2. Suppose, there is an x1 < 0 with f .x1 /  0. Let n WD b1=f .x0 /c C 1. Then,
n 2 N n f0g and n  f .x0 / > 1. Taking n times the summand x0 and m times the
summand x1 , where m 2 N is chosen appropriately, one gets n  x0 C m  x1  1,
but n  f .x0 / C m  f .x1 / > 1 in contradiction to Definition 3.1.
3.3. The sequence x2 ; : : : ; xn can contain x1 repeatedly, even if x1 > 1=2, because
negative summands may occur too. Suppose that x1 D x2 D    D xk , with k 2 N
and 0 < k  n. Instead of

X
n
f .xi / > f .1  x1 /;
iD2

one gets

X
n
1< h.xi /
iD1

X
n
D k  h.x1 / C h.xi /
iDkC1

X
n
D k  k  f .1  x1 / C f .xi /;
iDkC1

but this does not imply a contradiction.

150 Appendix A

3.4.
(a) 353=122.
(b) Consider for instance the following four patterns that are cut once each:

.1; 0; 1; 0; 0; 0/>;
.1; 0; 0; 1; 0; 2/>;
.0; 1; 0; 0; 2; 1/>;
.0; 0; 0; 0; 0; 1/>:

The corresponding waste lengths are equal to 10, 6, 9 and 110.

(c) The pattern

.1; 1; 0; 0; 0; 0/>

needs the length 123. Use this and for example the following two patterns once:

.1; 0; 1; 0; 0; 0/> .waste10/; and

.0; 0; 0; 1; 2; 4/ >
.waste4/:
61 61 61 61 61 61 61 61 61 61
(d) C 55 51 46 36 31 2 30 25 15 10 6
221 165 115 35 5 353 163 15 131 77 143
Without extra length 98 82 62 22 2 122 58 7 56 27 50
110 57 17 176 325 17 261 307 57
With extra length 49
2 31 11
2 61 116 8 112 108 20

Without extra space, the largest lower bound was 353

122

2:89. The extra space
counts like an item of length 1. Its contribution decreases the obtained lower
bound only slightly to 176
61
.
Remark Here, the optimal objective function value of the continuous relaxation is
32
a bit larger than the material bound, namely 2 35
2:91 (without extra space).
3.5.
(a) False.
Non-maximal general dual-feasible functions need not to have this structure.
An example is

x 7! ejxj :

(b) True.
Let f W R ! R be any general dual-feasible function and g W R ! R a
maximal general dual-feasible function dominating f . (If f is already maximal,
Appendix A 151

then g D f .) Proposition 3.4 (p. 62) implies

f .x/  g.x/  tx

for all x 2 R, where t WD sup f .x/

x
remains finite.
x>0

3.6. The proof is given next.

Proof We show first that L  F similarly to Proposition 3.2, p. 54. Let f 2 L and
therefore the constant c be given. One gets for any finite index set I of real numbers
X X X
xi  0 H) f .xi / D c  xi  0;
i2I i2I i2I

such that f fulfills the first condition. Assume that f is dominated by a real function
g, i.e. f .x/  g.x/ for all x 2 R and there is an y 2 R with g.y/ > f .y/. Then,

g.y/  f .y/

and

g.y/ C g.y/ > f .y/ C f .y/ D 0;

such that g does not fulfill the implication (3.2).

It remains to show that F  L . Let f 2 F be given. One gets f .0/ D 0,

f .x/  0  f .x/

for all x 2 RC , and also, analogously to part (b) of Theorem 3.1, that f is
superadditive. Moreover,

f .x/
c WD lim
x!1 x

must be finite, because

f .x/ C dxe  f .1/  0

for all x > 0. Of course, we have c  0. It follows that f .x/  cx for all x > 0,
otherwise the superadditivity of f would contradict the definition of c. It can also
not happen that f .x/ > cx for a certain x < 0. A detailed proof could be done like in
Proposition 3.4. Therefore, the linear function x 7! cx dominates f , but f is maximal
in the sense of the condition (2.), hence f .x/ D cx for all x 2 R. t
u
3.7. The functions f0 ; : : : ; f3 fulfill obviously the conditions (1.) and (3.) of
Theorem 3.1. Moreover, f0 is a superadditive general dual-feasible function, but not
152 Appendix A

symmetric, and f0 is not a maximal general dual-feasible functions. The reasons are
the following:

f0 .1/ D 1;

such that the superadditivity of f0 will imply that f0 is a general dual-feasible

function. One gets for x > 0 that f0 is differentiable and has the derivative

f00 .x/ D 1 C tanh 1  .1  tanh2 x/ D tanh 1 C tanh2 x;

which rises strictly monotonely. Therefore, f0 is strict convex for x > 0 and hence
superadditive for positive arguments. The superadditivity holds without restriction,
because if x < 0  y, then

f0 .x C y/  f0 .x/  f0 .y/ D tanh y  maxf0; tanh.x C y/g  0;

and if x; y < 0, then f0 .x C y/ D f0 .x/ C f0 .y/. The strict convexity implies also
f0 .x/ < x for 0 < x < 1, such that f0 cannot be symmetric and is dominated by
g W R ! R with
8
< .1 C tanh 1/  x; if x  0;
g.x/ WD x; if 0  x  1;
:
.1 C tanh 1/  x  tanh 1; if x  1;

which is a maximal general dual-feasible function according to Proposition 3.11

(with p WD 1 and t D 1 C tanh 1).
Let " > 0 be sufficiently small. The functions f1 and f2 are not monotone, hence
not superadditive, because one gets

f1 .2k/ D f2 .2k/ D 2k;

but

f1 .2k  "/ D f2 .2k  "/ D ..k C 1/  2k  1/=k D 2k C 2  1=k > 2k:

The function f2 is not even a general dual-feasible function, because

 
1 1
f2 D
kC1 kC1

and
1
f2 .1 C "/ D 1 C ;
k
Appendix A 153

yielding

1
f2 . / C f2 .1 C "/ > 1
kC1

in spite of

1
1C" < 1:
kC1

The functions f2 and f3 are symmetric, but not f1 . One has for example f1 .2/ D 2,
but f1 .1/ D 1  1=k, and hence

f1 .1/ C f1 .2/ D 1  1=k < 1:

Regarding f2 , if .k C 1/  x … Z, then

f2 .x/ C f2 .1  x/ D b.k C 1/  xc=k C b.k C 1/  .1  x/c=k

D b.k C 1/  .x C 1  x/  1c=k
D 1:

We show that f1 is a generalP dual-feasible function. Let any finite index set I
of real numbers xi (i 2 I) with xi  1 be given. If all xi are non-positive, then
P i2I
f1 .xi /  0 is immediately clear. Otherwise, one obtains
i2I

X X
f1 .xi / < .k C 1/  xi =k  1 C 1=k
i2I i2I

and also
X X
k f1 .xi /  b.k C 1/  xi c:
i2I i2I

Since the right-hand side is an integer with

X
k f1 .xi / < k C 1;
i2I

it follows that
X
k f1 .xi /  k:
i2I
154 Appendix A

The monotone function f3 differs from f1 only in points x > 1 where

.k C 1/  x 2 N. One gets for these x that

f3 .x/ D ..k C 1/  x  1/=k

D f3 .x  "/
D f1 .x  "/
> f1 .x/:

Since f1 is a general dual-feasible function, and f3 .x/ D lim f1 .y/ for any x 2 R,
y"x
the function f3 is also a general dual-feasible function. Since f3 dominates f1 , the
function f1 is not a maximal general dual-feasible function. f3 is symmetric and can
therefore not be dominated by another general dual-feasible function. Hence, f3 is a
maximal general dual-feasible function.
3.8. The given function g is a Hölder continuous classical maximal dual-feasible
function, because there is a constant c > 0, such that it holds for every x; y 2 Œ0; 1
that
p
jg.x/  g.y/j  c  jx  yj;

and g is strict convex on Œ0; 1=2, symmetric and therefore superadditive, and g.0/ D
0. Let p WD 1, y WDp 1=2 and x
0 with x < 0. Then, g.y/ D 1=2, f .x/ D tx and
g.x C y/ D .1  2x/=2. Hence, we have
p
f .x C y/  f .x/  f .y/ D tx  x=2 < 0

for x > 1
2t2
.

Chapter 4

4.1. Check the different classes of VP-MDFF, or show that there are two elements
x and y such that x C y  w and f .x/ C f .y/ > 1.
4.2. Check the different classes of VP-MDFF.
4.3. Immediate.
4.4. Immediate.
4.5. First, note that applying the function to each item is equivalent to cutting
the pieces into squares and applying two maximal dual-feasible functions on the
resulting 2-OPP-O instance.
Appendix A 155

By writing the corresponding values of function, and using the superadditivity of

f and g, question 3 is directly answered. Question 4 is similar, and can be answered
by induction on w and h and using question 3.
4.6. One can use the same methods used for the m-OPP-R.
4.7. Check that for all feasible patterns, the sum of the images is less than or equal
to 1. Note that 1 and 2 are two different elements, although they have the same size.
4.8. The values of f .1/ and f .2/ can be increased.
4.9. A solution is almost given in the next question. This cannot happen with a
CS-MDFF because such a function only depends on the sizes and is increasing.
4.10. Choose J as large as possible, and remark that the presence of item 5 does not
modify the optimal solution.
4.11. Choose the node set f1; 2; 9; 10g, f1; 2; 3; 4g, f1; 3; 7; 8g, f3; 4; 5; 6; 11; 12g.
4.12. The problem is equivalent to the special case of graph colouring, where the
graph is an interval graph.

Chapter 5

5.1.
(a) False.
Without superadditivity, counter examples exist like the following one: 1=2x 
1 is given, and hence x WD 2 is feasible. Suppose f .1=2/ D 1=2 and f .1/ D 0.
That yields the contradiction f .1=2/  2  f .1/ or 1  0.
(b) False.
A superadditive function f with f .0/ < 0 may also yield contradictions. For
instance, applying such a function to the inequality x  0, which allows x WD 0,
could yield the false conclusion x < 0.
5 3
> 4 1
>
5.2. This VP-MDFF maps the vectors 11 ; 8 and 11 ; 2 to 1=2 and the vector
3 1
>
;
11 4
to 1=4. That yields the valid inequality

x1 x2 x3
C C  1;
2 2 4
which is equivalent to the demanded one.
(Remark: The demanded inequality could also be obtained by adding the two
inequalities, dividing by 4 and applying the rounding procedure due to Chvátal and
Gomory.)
156 Appendix A

5.3. The definition of the VP-DFF implies f .aj0 / C f .aj /  1, and hence f .aj /  0.
Because of the range of f , it follows that f .aj / D 0. That implies also that if a column
vector aj0 gets the special mapping to 1 due to the property “large argument vector”
then every “small argument vector” aj with aj0 Caj  w will get the special mapping
to 0, and the other VP-MDFF in the considered construction principles will not be
applied to these vectors. Moreover, the special mapping to 1 can be applied only to
vectors aj with 2aj 6 w.
References

Aardal K, Weismantel R (1997) Polyhedral combinatorics. Wiley, New York

Alves C (2005) Cutting and packing: problems, models and exact algorithms. PhD thesis,
Universidade do Minho, Guimaraes
Alves C, Valério de Carvalho J, Clautiaux F, Rietz J (2014) Multidimensional dual-feasible
functions and fast lower bounds for the vector packing problem. Eur J Oper Res 233:43–63
Bazaraa M, Jarvis J, Sherali H (2010) Linear programming and network flows. Wiley, New York
Boschetti M, Mingozzi A (2003) The two-dimensional finite bin packing problem. Part I: new
lower bounds for the oriented case. 4 OR 1:27–42
Burdett C, Johnson E (1977) A subadditive approach to solve linear integer programs. Ann Discret
Math 1:117–144
Caprara A, Locatelli M, Monaci M (2005) Bilinear packing by bilinear programming. In: Jünger M,
Kaibel V (eds) Integer programming and combinatorial optimization, 11th international IPCO
conference, Berlin, 8–10 June 2005. Lecture notes in computer science, vol 3509. Springer,
Berlin, pp 377–391
Carlier J, Néron E (2007a) Computing redundant resources for cumulative scheduling problems.
Eur J Oper Res 176(3):1452–1463
Carlier J, Néron E (2007b) Computing redundant resources for the resource constrained project
scheduling problem. Eur J Oper Res 176(3):1452–1463
Carlier J, Clautiaux F, Moukrim A (2007) New reduction procedures and lower bounds for the
two-dimensional bin-packing problem with fixed orientation. Comput Oper Res 34:2223–2250
Chvátal V (1973) Edmonds polytopes and a hierarchy of combinatorial problems. Discret Math
4:305–337
Clautiaux F (2010) New collaborative approaches for bin-packing problems. Habilitation à Diriger
des Recherches, Université de Lille 1, France
Clautiaux F, Jouglet A, Hayek J (2007) A new lower bound for the non-oriented two-dimensional
bin-packing problem. Oper Res Lett 35:365–373
Clautiaux F, Alves C, Valério de Carvalho J (2010) A survey of dual-feasible and superadditive
functions. Ann Oper Res 179:317–342
Dantzig GB, Wolfe P (1960) Decomposition principle for linear programs. Oper Res 8:101–111
Dash S, Günlük O (2006) Valid inequalities based on simple mixed-integer sets. Math Program
105:29–53
Fekete S, Schepers J (2001) New classes of fast lower bounds for bin packing problems. Math
Program 91:11–31

© Springer International Publishing Switzerland 2016 157

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5
158 References

Fekete S, Schepers J (2004) A general framework for bounds for higher-dimensional orthogonal
packing problems. Math Meth Oper Res 60:311–329
Geoffrion A (1974) Lagrangian relaxation and its uses in integer programming. Math Program
Study 2:82–114
Gilmore P, Gomory R (1961) A linear programming approach to the cutting stock problem (part I).
Oper Res 9:849–859
Gomory R (1958) Outline of an algorithm for integer solutions to linear programs. Bull Am Math
Soc 64:275–278
Johnson D (1973) Near optimal bin packing algorithms. Dissertation, Massachussetts Institute of
Technology, Cambridge, MA
Khanafer A, Clautiaux F, Talbi E (2010) New lower bounds for bin packing problems with
conflicts. Eur J Oper Res 206:281–288
Letchford A, Lodi A (2002) Strengthening Chvával-Gomory cuts and Gomory fractional cuts. Oper
Res Lett 30:74–82
Lueker G (1983) Bin packing with items uniformly distributed over intervals [a,b]. In: Proceedings
of the 24th annual symposium on foundations of computer science (FOCS 83). IEEE Computer
Society, Silver Spring, MD, pp 289–297
Martello S, Toth P (1990) Knapsack problems - algorithms and computer implementation. Wiley,
Chichester
Nemhauser G, Wolsey L (1998) Integer and combinatorial optimization. Wiley, New York
Rietz J, Alves C, Valério de Carvalho J (2010) Theoretical investigations on maximal dual feasible
functions. Oper Res Lett 38:174–178
Rietz J, Alves C, Valério de Carvalho J (2012a) On the extremality of maximal dual feasible
functions. Oper Res Lett 40:25–30
Rietz J, Alves C, Valério de Carvalho J, Clautiaux F (2012b) Computing valid inequalities for
general integer programs using an extension of maximal dual-feasible functions to negative
arguments. In: Proceedings of the 1st international conference on operations research and
enterprise systems (ICORES 2012)
Rietz J, Alves C, Valério de Carvalho J, Clautiaux F (2014) On the properties of general
dual-feasible functions. In: Murgante B, Misra S, Rocha AMAC, Torre C, Rocha JG, Falcão MI,
Taniar D, Apduhan BO, Gervasi O (eds) Computational science and its applications – ICCSA
2014. Lecture notes on computer science, vol 8580. Springer, pp 180–194. doi:10.1007/978-3-
319-09129-7_14. http://dx.doi.org/10.1007/978-3-319-09129-7_14
Rietz J, Alves C, Valério de Carvalho J, Clautiaux F (2015) Constructing general dual-feasible
functions. Oper Res Lett 43:427–431
Robertson N, Seymour P (1986) Graph minors. II algorithmic aspects of tree-width. J Algorithms
7:309–322
Rose D, Tarjan E, Lueker G (1976) Algorithmic aspects of vertex elimination on graphs. SIAM J
Comput 5:146–160
Spieksma F (1994) A branch-and-bound algorithm for the two-dimensional vector packing
problem. Comput Oper Res 21:19–25
Valério de Carvalho J (1999) Exact solution of bin packing problems using column generation and
branch-and-bound. Ann Oper Res 86:629–659
Valério de Carvalho J (2002) A note on branch-and-price algorithms for the one-dimensional
cutting stock problems. Comput Optim Appl 21:339–340
Vance P (1998) Branch-and-Price algorithms for the one-dimensional cutting stock problem.
Comput Optim Appl 9:211–228
Vanderbeck F (2000) Exact algorithm for minimizing the number of setups in the one-dimensional
cutting stock problem. Oper Res 46(6):915–926
Index

Bin-packing problem (BP), 113 Integrality property, 4

Bin-packing problem with conflicts (BPC), 115
Bin-packing with conflicts DFF, 115
Binary knapsack problem with conflicts LP relaxation, 1
(KPC), 115
Binary knapsack problem, KP–01, 114
Block angular structure, 8 Maximality, 25, 54
m-dimensional orthogonal bin-packing
problem (m-OPP), 111
Column generation, 10 Minkowski’s theorem, 3
master problem, 10 Multi-dimensional knapsack, mD-KP, 97
subproblem, 10
Composition, 34, 53, 66
Cutting Stock Problem, 7, 23 Orthogonal packing DFF (m-OPP-R-DFF,
m-OPP-O-DFF), 111
Dantzig-Wolfe decomposition, 3
Data-dependent set-covering dual-feasible Set-covering dual-feasible function (SC-DFF),
function (SC-DDFF), 93 94
Discrete dual-feasible function, 22 Superadditivity, 25
Dual-feasible function, 21 Symmetry, 26, 35

Extremality, 28, 56
Tree-decomposition (of a graph), 117
Triangulated (graph), 117
General dual-feasible function, 52
Gilmore and Gomory model, 7, 23
Vector packing (mD-VPP), 95
Vector packing dual-feasible function
Integrality constraints, 1 (VP-DFF), 97

© Springer International Publishing Switzerland 2016 159

C. Alves et al., Dual-Feasible Functions for Integer Programming
and Combinatorial Optimization, EURO Advanced Tutorials on Operational
Research, DOI 10.1007/978-3-319-27604-5