Stochastic Programming

Stochastic Programming
Second Edition
Peter Kall
Institute for Operations Research
and Mathematical Methods of Economics
University of Zurich
CH-8044 Zurich
Stein W. Wallace
Molde University College
P.O. Box 2110
N-6402 Molde, Norway
Reference to this text is “Peter Kall and Stein W. Wallace, Stochastic

Programming, John Wiley & Sons, Chichester, 1994”. The text is printed
with permission from the authors. The publisher reverted the rights to the
authors on February 4, 2003. This text is slightly updated from the published
version.
ii STOCHASTIC PROGRAMMING
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
1 Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 A numerical example . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Scenario analysis . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Using the expected value of p . . . . . . . . . . . . . . . 3
1.1.4 Maximizing the expected value of the objective . . . . . 4
1.1.5 The IQ of hindsight . . . . . . . . . . . . . . . . . . . . 5
1.1.6 Options . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 An Illustrative Example . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Stochastic Programs: General Formulation . . . . . . . . . . . . 21
1.4.1 Measures and Integrals . . . . . . . . . . . . . . . . . . . 21
1.4.2 Deterministic Equivalents . . . . . . . . . . . . . . . . . 31
1.5 Properties of Recourse Problems . . . . . . . . . . . . . . . . . 36
1.6 Properties of Probabilistic Constraints . . . . . . . . . . . . . . 46
1.7 Linear Programming . . . . . . . . . . . . . . . . . . . . . . . . 53
1.7.1 The Feasible Set and Solvability . . . . . . . . . . . . . 54
1.7.2 The Simplex Algorithm . . . . . . . . . . . . . . . . . . 64
1.7.3 Duality Statements . . . . . . . . . . . . . . . . . . . . . 70
1.7.4 A Dual Decomposition Method . . . . . . . . . . . . . . 75
1.8 Nonlinear Programming . . . . . . . . . . . . . . . . . . . . . . 80
1.8.1 The Kuhn–Tucker Conditions . . . . . . . . . . . . . . . 83
1.8.2 Solution Techniques . . . . . . . . . . . . . . . . . . . . 89
1.8.2.1 Cutting-plane methods . . . . . . . . . . . . . 90
1.8.2.2 Descent methods . . . . . . . . . . . . . . . . . 93
1.8.2.3 Penalty methods . . . . . . . . . . . . . . . . . 97
1.8.2.4 Lagrangian methods . . . . . . . . . . . . . . . 98
1.9 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . 102
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
iv STOCHASTIC PROGRAMMING
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
2 Dynamic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

2.1 The Bellman Principle . . . . . . . . . . . . . . . . . . . . . . . 110
2.2 Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . 117
2.3 Deterministic Decision Trees . . . . . . . . . . . . . . . . . . . . 121
2.4 Stochastic Decision Trees . . . . . . . . . . . . . . . . . . . . . 124
2.5 Stochastic Dynamic Programming . . . . . . . . . . . . . . . . 130
2.6 Scenario Aggregation . . . . . . . . . . . . . . . . . . . . . . . . 134
2.6.1 Approximate Scenario Solutions . . . . . . . . . . . . . 141
2.7 Financial Models . . . . . . . . . . . . . . . . . . . . . . . . . . 141
2.7.1 The Markowitz’ model . . . . . . . . . . . . . . . . . . . 142
2.7.2 Weak aspects of the model . . . . . . . . . . . . . . . . 143
2.7.3 More advanced models . . . . . . . . . . . . . . . . . . . 145
2.7.3.1 A scenario tree . . . . . . . . . . . . . . . . . . 145
2.7.3.2 The individual scenario problems . . . . . . . . 145
2.7.3.3 Practical considerations . . . . . . . . . . . . . 147
2.8 Hydro power production . . . . . . . . . . . . . . . . . . . . . . 147
2.8.1 A small example . . . . . . . . . . . . . . . . . . . . . . 148
2.8.2 Further developments . . . . . . . . . . . . . . . . . . . 150
2.9 The Value of Using a Stochastic Model . . . . . . . . . . . . . . 151
2.9.1 Comparing the Deterministic and Stochastic Objective
Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
2.9.2 Deterministic Solutions in the Event Tree . . . . . . . . 152
2.9.3 Expected Value of Perfect Information . . . . . . . . . . 154
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
3 Recourse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 159

3.1 Outline of Structure . . . . . . . . . . . . . . . . . . . . . . . . 159
3.2 The L-shaped Decomposition Method . . . . . . . . . . . . . . 161
3.2.1 Feasibility . . . . . . . . . . . . . . . . . . . . . . . . . . 161
3.2.2 Optimality . . . . . . . . . . . . . . . . . . . . . . . . . 168
3.3 Regularized Decomposition . . . . . . . . . . . . . . . . . . . . 173
3.4 Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
3.4.1 The Jensen Lower Bound . . . . . . . . . . . . . . . . . 179
3.4.2 Edmundson–Madansky Upper Bound . . . . . . . . . . 181
3.4.3 Combinations . . . . . . . . . . . . . . . . . . . . . . . . 184
3.4.4 A Piecewise Linear Upper Bound . . . . . . . . . . . . . 185
3.5 Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
3.5.1 Refinements of the bounds on the “Wait-and-See”Solution190
3.5.2 Using the L-shaped Method within Approximation
Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
3.5.3 What is a Good Partition? . . . . . . . . . . . . . . . . 203
CONTENTS v
3.6 Simple Recourse . . . . . . . . . . . . . . . . . . . . . . . . . . 205

3.7 Integer First Stage . . . . . . . . . . . . . . . . . . . . . . . . . 209
3.7.1 Initialization . . . . . . . . . . . . . . . . . . . . . . . . 216
3.7.2 Feasibility Cuts . . . . . . . . . . . . . . . . . . . . . . . 216
3.7.3 Optimality Cuts . . . . . . . . . . . . . . . . . . . . . . 217
3.7.4 Stopping Criteria . . . . . . . . . . . . . . . . . . . . . . 217
3.8 Stochastic Decomposition . . . . . . . . . . . . . . . . . . . . . 217
3.9 Stochastic Quasi-Gradient Methods . . . . . . . . . . . . . . . . 225
3.10 Solving Many Similar Linear Programs . . . . . . . . . . . . . . 229
3.10.1 Randomness in the Objective . . . . . . . . . . . . . . . 232
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
4 Probabilistic Constraints . . . . . . . . . . . . . . . . . . . . . . 243

4.1 Joint Chance Constrained Problems . . . . . . . . . . . . . . . 245
4.2 Separate Chance Constraints . . . . . . . . . . . . . . . . . . . 247
4.3 Bounding Distribution Functions . . . . . . . . . . . . . . . . . 249
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
5 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
5.1 Problem Reduction . . . . . . . . . . . . . . . . . . . . . . . . . 261
5.1.1 Finding a Frame . . . . . . . . . . . . . . . . . . . . . . 262
5.1.2 Removing Unnecessary Columns . . . . . . . . . . . . . 263
5.1.3 Removing Unnecessary Rows . . . . . . . . . . . . . . . 264
5.2 Feasibility in Linear Programs . . . . . . . . . . . . . . . . . . . 265
5.2.1 A Small Example . . . . . . . . . . . . . . . . . . . . . . 271
5.3 Reducing the Complexity of Feasibility Tests . . . . . . . . . . 273
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
6 Network Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 277

6.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
6.2 Feasibility in Networks . . . . . . . . . . . . . . . . . . . . . . . 280
6.2.1 The uncapacitated case . . . . . . . . . . . . . . . . . . 286
6.2.2 Comparing the LP and Network Cases . . . . . . . . . . 287
6.3 Generating Relatively Complete Recourse . . . . . . . . . . . . 288
6.4 An Investment Example . . . . . . . . . . . . . . . . . . . . . . 290
6.5 Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
6.5.1 Piecewise Linear Upper Bounds . . . . . . . . . . . . . . 295
vi STOCHASTIC PROGRAMMING
6.6 Project Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . 301

6.6.1 PERT as a Decision Problem . . . . . . . . . . . . . . . 303
6.6.2 Introduction of Randomness . . . . . . . . . . . . . . . . 303
6.6.3 Bounds on the Expected Project Duration . . . . . . . . 304
6.6.3.1 Series reductions . . . . . . . . . . . . . . . . . 305
6.6.3.2 Parallel reductions . . . . . . . . . . . . . . . . 305
6.6.3.3 Disregarding path dependences . . . . . . . . . 305
6.6.3.4 Arc duplications . . . . . . . . . . . . . . . . . 306
6.6.3.5 Using Jensen’s inequality . . . . . . . . . . . . 306
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
Preface
Over the last few years, both of the authors, and also most others in the field
of stochastic programming, have said that what we need more than anything
just now is a basic textbook—a textbook that makes the area available not
only to mathematicians, but also to students and other interested parties who
cannot or will not try to approach the field via the journals. We also felt
the need to provide an appropriate text for instructors who want to include
the subject in their curriculum. It is probably not possible to write such a
book without assuming some knowledge of mathematics, but it has been our
clear goal to avoid writing a text readable only for mathematicians. We want
the book to be accessible to any quantitatively minded student in business,
economics, computer science and engineering, plus, of course, mathematics.
So what do we mean by a quantitatively minded student? We assume that
the reader of this book has had a basic course in calculus, linear algebra
and probability. Although most readers will have a background in linear
programming (which replaces the need for a specific course in linear algebra),
we provide an outline of all the theory we need from linear and nonlinear
programming. We have chosen to put this material into Chapter 1, so that
the reader who is familiar with the theory can drop it, and the reader who
knows the material, but wonders about the exact definition of some term, or
who is slightly unfamiliar with our terminology, can easily check how we see
things. We hope that instructors will find enough material in Chapter 1 to
cover specific topics that may have been omitted in the standard book on
optimization used in their institution. By putting this material directly into
the running text, we have made the book more readable for those with the
minimal background. But, at the same time, we have found it best to separate
what is new in this book—stochastic programming—from more standard
material of linear and nonlinear programming.
Despite this clear goal concerning the level of mathematics, we must
admit that when treating some of the subjects, like probabilistic constraints
(Section 1.6 and Chapter 4), or particular solution methods for stochastic
programs, like stochastic decomposition (Section 3.8) or quasi-gradient
viii STOCHASTIC PROGRAMMING
methods (Section 3.9), we have had to use a slightly more advanced language
in probability. Although the actual information found in those parts of the
book is made simple, some terminology may here and there not belong to
the basic probability terminology. Hence, for these parts, the instructor must
either provide some basic background in terminology, or the reader should at
least consult carefully Section 1.4.1, where we have tried to put together those
terms and concepts from probability theory used later in this text.
Within the mathematical programming community, it is common to split
the field into topics such as linear programming, nonlinear programming,
network flows, integer and combinatorial optimization, and, finally, stochastic
programming. Convenient as that may be, it is conceptually inappropriate.
It puts forward the idea that stochastic programming is distinct from integer
programming the same way that linear programming is distinct from nonlinear
programming. The counterpart of stochastic programming is, of course,
deterministic programming. We have stochastic and deterministic linear
programming, deterministic and stochastic network flow problems, and so on.
Although this book mostly covers stochastic linear programming (since that is
the best developed topic), we also discuss stochastic nonlinear programming,
integer programming and network flows.
Since we have let subject areas guide the organization of the book, the
chapters are of rather different lengths. Chapter 1 starts out with a simple
example that introduces many of the concepts to be used later on. Tempting as
it may be, we strongly discourage skipping these introductory parts. If these
parts are skipped, stochastic programming will come forward as merely an
algorithmic and mathematical subject, which will serve to limit the usefulness
of the field. In addition to the algorithmic and mathematical facets of the
field, stochastic programming also involves model creation and specification
of solution characteristics. All instructors know that modelling is harder to
teach than are methods. We are sorry to admit that this difficulty persists
in this text as well. That is, we do not provide an in-depth discussion of
modelling stochastic programs. The text is not free from discussions of models
and modelling, however, and it is our strong belief that a course based on this
text is better (and also easier to teach and motivate) when modelling issues
are included in the course.
Chapter 1 contains a formal approach to stochastic programming, with a
discussion of different problem classes and their characteristics. The chapter
ends with linear and nonlinear programming theory that weighs heavily in
stochastic programming. The reader will probably get the feeling that the
parts concerned with chance-constrained programming are mathematically
more complicated than some parts discussing recourse models. There is a
good reason for that: whereas recourse models transform the randomness
contained in a stochastic program into one special parameter of some random
vector’s distribution, namely its expectation, chance constrained models deal
PREFACE ix
more explicitly with the distribution itself. Hence the latter models may
be more difficult, but at the same time they also exhaust more of the
information contained in the probability distribution. However, with respect to
applications, there is no generally valid justification to state that any one of the
two basic model types is “better” or “more relevant”. As a matter of fact, we
know of applications for which the recourse model is very appropriate and of
others for which chance constraints have to be modelled, and even applications
are known for which recourse terms for one part of the stochastic constraints
and chance constraints for another part were designed. Hence, in a first reading
or an introductory course, one or the other proof appearing too complicated
can certainly be skipped without harm. However, to get a valid picture about
stochastic programming, the statements about basic properties of both model
types as well as the ideas underlying the various solution approaches should be
noticed. Although the basic linear and nonlinear programming is put together
in one specific part of the book, the instructor or the reader should pick up
the subjects as they are needed for the understanding of the other chapters.
That way, it will be easier to pick out exactly those parts of the theory that
the students or readers do not know already.
Chapter 2 starts out with a discussion of the Bellman principle for
solving dynamic problems, and then discusses decision trees and dynamic
programming in both deterministic and stochastic settings. There then follows
a discussion of the rather new approach of scenario aggregation. We conclude
the chapter with a discussion of the value of using stochastic models.
Chapter 3 covers recourse problems. We first discuss some topics from
Chapter 1 in more detail. Then we consider decomposition procedures
especially designed for stochastic programs with recourse. We next turn to
the questions of bounds and approximations, outlining some major ideas
and indicating the direction for other approaches. The special case of simple
recourse is then explained, before we show how decomposition procedures for
stochastic programs fit into the framework of branch-and-cut procedures for
integer programs. This makes it possible to develop an approach for stochastic
integer programs. We conclude the chapter with a discussion of Monte-Carlo
based methods, in particular stochastic decomposition and quasi-gradient
methods.
Chapter 4 is devoted to probabilistic constraints. Based on convexity
statements provided in Section 1.6, one particular solution method is described
for the case of joint chance constraints with a multivariate normal distribution
of the right-hand side. For separate probabilistic constraints with a joint
normal distribution of the coefficients, we show how the problem can be
transformed into a deterministic convex nonlinear program. Finally, we
address a problem very relevant in dealing with chance constraints: the
problem of how to construct efficiently lower and upper bounds for a
multivariate distribution function, and give a first sketch of the ideas used
x STOCHASTIC PROGRAMMING
in this area.
Preprocessing is the subject of Chapter 5. “Preprocessing” is any analysis
that is carried out before the actual solution procedure is called. Preprocessing
can be useful for simplifying calculations, but the main purpose is to facilitate
a tool for model evaluation.
We conclude the book with a closer look at networks (Chapter 6). Since
these are nothing else than specially structured linear programs, we can draw
freely from the topics in Chapter 3. However, the added structure of networks
allows many simplifications. We discuss feasibility, preprocessing and bounds.
We conclude the chapter with a closer look at PERT networks.
Each chapter ends with a short discussion of where more literature can be
found, some exercises, and, finally, a list of references.
Writing this book has been both interesting and difficult. Since it is the first
basic textbook totally devoted to stochastic programming, we both enjoyed
and suffered from the fact that there is, so far, no experience to suggest how
such a book should be constructed. Are the chapters in the correct order?
Is the level of difficulty even throughout the book? Have we really captured
the basics of the field? In all cases the answer is probably NO. Therefore,
dear reader, we appreciate all comments you may have, be they regarding
misprints, plain errors, or simply good ideas about how this should have been
done. And also, if you produce suitable exercises, we shall be very happy to
receive them, and if this book ever gets revised, we shall certainly add them,
and allude to the contributor.
About 50% of this text served as a basis for a course in stochastic
programming at The Norwegian Institute of Technology in the fall of 1992. We
wish to thank the students for putting up with a very preliminary text, and
for finding such an astonishing number of errors and misprints. Last but not
least, we owe sincere thanks to Julia Higle (University of Arizona, Tucson),
Diethard Klatte (Univerity of Zurich), Janos Mayer (University of Zurich) and
Pavel Popela (Technical University of Brno) who have read the manuscript1
very carefully and fixed not only linguistic bugs but prevented us from quite a
number of crucial mistakes. Finally we highly appreciate the good cooperation
and very helpful comments provided by our publisher. The remaining errors
are obviously the sole responsibility of the authors.
Zurich and Trondheim, February 1994 P. K. and S.W.W.
1 Written in LATEX
1
Basic Concepts
1.1 Motivation
By reading this introduction, you are certainly already familiar with

deterministic optimization. Most likely, you also have some insights into what
new challenges face you when randomness is (re)introduced into a model.
The interest for studying stochastic programming can come from different
sources. Your interests may concern the algorithmic or mathematical, as well
as modeling and applied aspects of optimization. We hope to provide you with
some insights into the basics of all these areas. In these very first pages we
will demonstrate why it is important, often crucial, that you turn to stochastic
programming when working with decisions affected by uncertainty. And, in in
our view, all decision problems are of this type.
Technically, stochastic programs are much more complicated than the
corresponding deterministic programs. Hence, at least from a practical point
of view, there must be very good reasons to turn to the stochastic models. We
start this book with a small example illustrating that these reasons exist. In
fact, we shall demonstrate that alternative deterministic approaches do not
even look for the best solutions. Deterministic models may certainly produce
good solutions for certain data set in certain models, but there is generally
no way you can conclude that they are good, without comparing them to
solutions of stochastic programs. In many cases, solutions to deterministic
programs are very misleading.
1.1.1 A numerical example

You own two lots of land. Each of them can be developed with necessary
infrastructure and a plant can be built. In fact, there are nine possible
decisions. Eight of them are given in Figure 1, the ninth is to do nothing.
The cost structure is given in the following table. For each lot of land we
give the cost of developing the land and building the plant. The extra column
will be explained shortly.
2 STOCHASTIC PROGRAMMING
2 4
3
1
5 6
7 8
Figure 1 Eight of the nine possible decisions. The area surrounded by thin
lines correspond to Lot 1, the area with thick lines to Lot 2. For example,
Decision 6 is to develop both lots, and build a plant on Lot 1. Decision 9 is to
do nothing.
developing the land building the plant building the plant later
Lot 1 600 200 220
Lot 2 100 600 660
In each of the plants, it is possible to produce one unit of some product. It

can be sold at a price p. The price p is unknown when the land is developed.
Also, if the plant on Lot 1, say, is to be built at its lowest cost, given as 200
in the table, that must take place before p becomes known. However, it is
possible to delay the building of the plant until after p becomes known, but
at a 10% penalty. That is given in the last column of the table. This can only
take place if the lot is already developed. There is not enough time to both
develop the land and build a plant after p has become known.
1.1.2 Scenario analysis

A common way of solving problems of this kind is to perform scenario analysis,
also sometimes referred to as simulation. (Both terms have a broader meaning
than what we use here, of course.) The idea is to construct or sample possible
futures (values of p in our case) and solve the corresponding problem for
these values. After having obtained a number of possible decision this way,
we either pick the best of them (details will be given later), or we try to find
good combinations of the decisions.
In our case it is simple to show that there are only three possible scenario
BASIC CONCEPTS 3
solutions. These are given as follows. Decision numbers refer to Figure 1.
Interval for p Decision number

p < 700 9
700 ≤ p < 800 4
p ≥ 800 7
So whatever scenarios are constructed or sampled, these are the only

possible solutions. Note that in this setting it is never optimal to use delayed
construction. The reason is that each scenario analysis is performed under
certainty, and hence, there is no reason to pay the extra 10% for being allowed
to delay the decision.
Now, assume for simplicity that p can take on only two values, namely 210
and 1250, each with a probability of 0.5. This is a very extreme choice, but
it has been made only for convenience. We could have made the same points
with more complicated (for example continuous) distributions, but nothing
would have been gained by doing that, except make the calculations more
complicated. Hence, the expected price equals 730.
1.1.3 Using the expected value of p

A common solution procedure for stochastic problems is to use the expected
value of all random variables. This is sometimes done very explicitly, but
more often it is done in the following fashion: The modeler collects data,
either by experiments or by checking an existing process over time, and then
calculates the mean, which is then said to be the best available estimate of the
parameter. In this case we would then use 730, and from the list of scenario
solutions above, we see that the optimal solution will Decision 4 with a profit
of
−700 + 730 = 30.
We call this the expected value solution. We can also calculate the expected
value of using the expected value solution. That is, we can use the expected
value solution, and then see how it performs under the possible futures. We
get
1 1
−700 + 210 + 1250 = 30.
2 2
It is not a general result that the expected value of using the expected value
solution equals the scenario solution value corresponding to the expected value
of the parameters (here p). But in this case that happens.
1.1.4 Maximizing the expected value of the objective

We just calculated the expected value of using the expected value solution. It
was 30. We can also calculate the expected value of using any of the possible
scenario solutions. We find that for doing nothing (Decision 9), the expected
value is 0, and for Decision 7 the expected value equals
1 1
−1500 + 420 + 2500 = −40.
2 2
In other words, the expected value solution is the best of the three scenario
solutions in terms of having the best expected performance. But is this the
solution with the best expected performance? Let us answer this question
by simply listing all possible solutions, and calculate their expected value. In
all cases, if the land is developed before p becomes known, we will consider
the option of building the plant at the 10% penalty if that is profitable. The
results are given in Table 1.
Table 1 The expected value of all nine possible solutions. The income is the
value of the product if the plant is already built. If not, it is the value of the
product minus the construction cost at 10% penalty.
Decision Investment Income if Income if Expected

p = 210 p = 1250 profit
1
1 −600 2 1030 −85
1 1
2 −800 2 210 2 1250 −70
1
3 −100 2 590 195
1 1
4 −700 2 210 2 1250 30
1 1
5 −1300 2 210 2 2280 −55
1 1
6 −900 2 210 2 1840 125
1 1
7 −1500 2 420 2 2500 −40
1
8 −700 2 1620 110
9 0 0 0 0
As we see from Table 1, the optimal solution is to develop Lot 2, then wait to
see what the price turns out to be. If the price turns out to be low, do nothing,
if it turns out to be high, build plant 2. The solution that truly maximizes the
expected value of the objective function will be called the stochastic solution.
Note that also two more solutions are substantially better than the expected
value solution.
All three solutions that are better than the expected value solution are
solutions with options in them. That is, they mean that we develop some land
in anticipation of high prices. Of course, there is a chance that the investment
BASIC CONCEPTS 5
will be lost. In scenario analysis, as outlined earlier, options have no value,

and hence, never show up in a solution. It is important to note that the fact
that these solutions did not show up as scenario solutions is not caused by few
scenarios, but by the very nature of a scenario, namely that it is deterministic.
It is incorrect to assume that if you can obtain enough scenarios, you will
eventually come upon the correct solution.
1.1.5 The IQ of hindsight

In hindsight, that is, after the fact, it will always be such that one of the
scenario solutions turn out to be the best choice. In particular, the expected
value solution will be optimal for any 700 < p ≤ 800. (We did not have any
probability mass there in our example, but we could easily have constructed
such a case.) The problem is that it is not the same scenario solution that is
optimal in all cases. In fact, most of them are very bad in all but the situation
where they are best.
The stochastic solution, on the other hand, is normally never optimal after
the fact. But, at the same time, it is also hardly ever really bad.
In our example, with the given probability distribution, the decision of doing
nothing (which has an expected value of zero) and the decision of building
both plants (with an expected value of -40) both have a probability of 50%
of being optimal after p has become known. The stochastic solution, with an
expected value of 195, on the other hand, has zero probability of being optimal
in hindsight.
This is an important observation. If you base your decisions on stochastic
models, you will normally never do things really well. Therefore, people who
prefer to evaluate after the fact can always claim that you made a bad decision.
If you base your decisions on scenario solutions, there is a certain chance that
you will do really well. It is therefore possible to claim that in certain cases
the most risky decision one can make is the one with the highest expected
value, because you will then always be proven wrong after the fact. The IQ of
hindsight is very high.
1.1.6 Options
We have already hinted at it several times, but let us repeat the observation
that the value of a stochastic programming approach to a problem lies in
the explicit evaluation of flexibility. Flexible solutions will always lose in
deterministic evaluations.
Another area where these observations have been made for quite a while is
option theory. This theory is mostly developed for financial models, but the
theory of real options (for example investments) is coming. Let us consider
our extremely simple example in the light of options.
We observed from Table 1 that the expected Net Present Value (NPV)
of Decision 4, i.e. the decision to develop Lot 2 and build a plant, equals
30. Standard theory tells us to invest if a project has a positive NPV, since
that means the project is profitable. And, indeed, Decision 4 represents an
investment which is profitable in terms of expected profits. But as we have
observed, Decision 3 is better, and it is not possible to make both decisions;
they exclude each other. The expected NPV for Decision 3 is 195. The
difference of 165 is the value of an option, namely the option not to build
the plant. Or to put it in a different wording: If your only possibilities were to
develop Lot 2 and build the plant at the same time, or do nothing, and you
were asked how much you were willing to pay in order to be allowed to delay
the building of the plant (at the 10% penalty) the answer is at most 165.
Another possible setting is to assume that the right to develop Lot 2 and
build the plant is for sale. This right can be seen as an option. This option is
worth 195 in the setting where delayed construction of the plant is allowed.
(If delays were not allowed, the right to develop and build would be worth 30,
but that is not an option.)
So what is it that gives an option a value? Its value stems from the right
to do something in the future under certain circumstances, but to drop it in
others if you so wish. And, even more importantly, to evaluate an option you
must model explicitly the future decisions. This is true in our simple model,
but it is equally true in any complex option model. It is not enough to describe
a stochastic future, this stochastic future must contain decisions.
So what are the important aspect of randomness? We may conclude that
there are at least three (all related of course).
1. Randomness is needed to obtain a correct evaluation of the future income
and costs, i.e. to evaluate the objective.
2. Flexibility only has value (and meaning) in a setting of randomness.
3. Only by explicitly evaluating future decisions can decisions containing
flexibility (options) be correctly valued.
BASIC CONCEPTS 7
1.2 Preliminaries
Many practical decision problems—in particular, rather complex ones—can

be modelled as linear programs
⎫
min{c1 x1 + c2 x2 + · · · + cn xn } ⎪
⎪
⎪
⎪
⎪
⎪
subject to ⎪
⎪
a11 x1 + a12 x2 + · · · + a1n xn = b1 ⎪ ⎪
⎬
a21 x1 + a22 x2 + · · · + a2n xn = b2 (2.1)
.. .. ⎪
⎪
⎪
. . ⎪
⎪
⎪
⎪
am1 x1 + am2 x2 + · · · + amn xn = bm ⎪ ⎪
⎪
⎭
x1 , x2 , · · · , xn ≥ 0.
Using matrix–vector notation, the shorthand formulation of problem (2.1)

would read as ⎫
min cT x ⎬
s.t. Ax = b (2.2)
⎭
x ≥ 0.
Typical applications may be found in the areas of industrial production,
transportation, agriculture, energy, ecology, engineering, and many others. In
problem (2.1) the coefficients cj (e.g. factor prices), aij (e.g. productivities)
and bi (e.g. demands or capacities) are assumed to have fixed known real values
and we are left with the task of finding an optimal combination of the values for
the decision variables xj (e.g. factor inputs, activity levels or energy flows) that
have to satisfy the given constraints. Obviously, model (2.1) can only provide
a reasonable representation of a real life problem when the functions involved
(e.g. cost functions or production functions) are fairly linear in the decision
variables. If this condition is substantially violated—for example, because of
increasing marginal costs or decreasing marginal returns of production—we
should use a more general form to model our problem:
⎫
min g0 (x) ⎬
s.t. gi (x) ≤ 0, i = 1, · · · , m (2.3)
⎭
x ∈ X ⊂ IRn .
The form presented in (2.3) is known as a mathematical programming problem.

Here it is understood that the set X as well as the functions gi : IRn → IR, i =
0, · · · , m, are given by the modelling process.
Depending on the properties of the problem defining functions gi and the
set X, program (2.3) is called
(a) linear, if the set X is convex polyhedral and the functions gi , i = 0, · · · , m,
are linear;
(b) nonlinear, if at least one of the functions gi , i = 0, · · · , m, is nonlinear or

X is not a convex polyhedral set; among nonlinear programs, we denote
a program as
(b1) convex, if X ∩ {x | gi (x) ≤ 0, i = 1, · · · , m} is a convex set and g0 is
a convex function (in particular if the functions gi , i = 0, · · · , m are
convex and X is a convex set); and
(b2) nonconvex, if either X ∩{x | gi (x) ≤ 0, i = 1, · · · , m} is not a convex
set or the objective function g0 is not convex.
Case (b2) above is also referred to as global optimization. Another special
class of problems, called (mixed) integer programs, arises if the set X requires
(at least some of) the variables xj , j = 1, · · · , n, to take integer values only.
We shall deal only briefly with discrete (i.e. mixed integer) problems, and
there is a natural interest in avoiding nonconvex programs whenever possible
for a very simple reason revealed by the following example from elementary
calculus.
Example 1.1 Consider the optimization problem
min ϕ(x), (2.4)

x∈IR
where ϕ(x) := 14 x4 − 5x3 + 27x2 − 40x. A necessary condition for solving

problem (2.4) is
ϕ (x) = x3 − 15x2 + 54x − 40 = 0.
Observing that
ϕ (x) = (x − 1)(x − 4)(x − 10),
we see that x1 = 1, x2 = 4 and x3 = 10 are candidates to solve our problem.
Moreover, evaluating the second derivative ϕ (x) = 3x2 − 30x + 54, we get
ϕ (x1 ) = 27,
ϕ (x2 ) = −18,
ϕ (x3 ) = 54,
indicating that x1 and x3 yield a relative minimum whereas in x2 we find
a relative maximum. However, evaluating the two relative minima yields
ϕ(x1 ) = −17.75 and ϕ(x3 ) = −200. Hence, solving our little problem (2.4)
with a numerical procedure that intends to satisfy the first- and second-order
conditions for a minimum, we might (depending on the starting point of the
procedure) end up with x1 as a “solution” without realizing that there exists
a (much) better possibility. 2
As usual, a function ψ is said to attain a relative minimum—also called a

local minimum—at some point x̂ if there is a neighbourhood U of x̂ (e.g. a ball
BASIC CONCEPTS 9
with center x̂ and radius ε > 0) such that ψ(x̂) ≤ ψ(y) ∀y ∈ U . A minimum
ψ(x̄) is called global if ψ(x̄) ≤ ψ(z) ∀z. As we just saw, a local minimum ψ(x̂)
need not be a global minimum.
A situation as in the above example cannot occur with convex programs
because of the following.
Lemma 1.1 If problem (2.3) is a convex program then any local (i.e. relative)
minimum is a global minimum.
Proof If x̄ is a local minimum of problem (2.3) then x̄ belongs to the

feasible set B := X ∩ {x | gi (x) ≤ 0, i = 1, · · · , m}. Further, there is an ε0 > 0
such that for any ball Kε := {x | x − x̄ ≤ ε}, 0 < ε < ε0 , we have that
g0 (x̄) ≤ g0 (x) ∀x ∈ Kε ∩B. Choosing an arbitrary y ∈ B, y = x̄, we may choose
an ε > 0 such that ε < y− x̄ and ε < ε0 . Finally, since, from our assumption,
B is a convex set and the objective g0 is a convex function, the line segment x̄y
intersects the surface of the ball Kε in a point x̂ such that x̂ = αx̄ + (1 − α)y
for some α ∈ (0, 1), yielding g0 (x̄) ≤ g0 (x̂) ≤ αg0 (x̄) + (1 − α)g0 (y), which
implies that g0 (x̄) ≤ g0 (y). 2
During the last four decades, progress in computational methods for solving
mathematical programs has been impressive, and problems of considerable size
may be solved efficiently, and with high reliability.
In many modelling situations it is unreasonable to assume that the
coefficients cj , aij , bi or the functions gi (and the set X) respectively in
problems (2.1) and (2.3) are deterministically fixed. For instance, future
productivities in a production problem, inflows into a reservoir connected
to a hydro power station, demands at various nodes in a transportation
network, and so on, are often appropriately modelled as uncertain parameters,
which are at best characterized by probability distributions. The uncertainty
about the realized values of those parameters cannot always be wiped out
just by inserting their mean values or some other (fixed) estimates during
the modelling process. That is, depending on the practical situation under
consideration, problems (2.1) or (2.3) may not be the appropriate models
for describing the problem we want to solve. In this chapter we emphasize—
and possibly clarify—the need to broaden the scope of modelling real life
decision problems. Furthermore, we shall provide from linear programming
and nonlinear programming the essential ingredients absolutely necessary for
the understanding of the subsequent chapters. Obviously these latter sections
may be skipped—or used as a quick revision—by readers who are already
familiar with the related optimization courses.
Before coming to a more general setting we first derive some typical
stochastic programming models, using a simplified production problem to
illustrate the various model types.
1.3 An Illustrative Example
Let us consider the following problem, idealized for the purpose of easy
presentation. From two raw materials, raw1 and raw2, we may simultaneously
produce two different goods, prod1 and prod2 (as may happen for example in
a refinery). The output of products per unit of the raw materials as well
as the unit costs of the raw materials c = (craw1 , craw2 )T (yielding the
production cost γ), the demands for the products h = (hprod1 , hprod2 )T and
the production capacity b̂, i.e. the maximal total amount of raw materials that
can be processed, are given in Table 2.
According to this formulation of our production problem, we have to deal
with the following linear program:
Table 2 Productivities π(raw i, prod j).
Products
Raws prod1 prod2 c b̂
raw1 2 3 2 1
raw2 6 3 3 1
relation ≥ ≥ = ≤
h 180 162 γ 100
⎫
min(2xraw1 + 3xraw2 ) ⎪
⎪
⎪
⎪
s.t. xraw1 + xraw2 ≤ 100, ⎪
⎪
⎬
2xraw1 + 6xraw2 ≥ 180,
(3.1)
3xraw1 + 3xraw2 ≥ 162, ⎪
⎪
⎪
⎪
xraw1 ≥ 0, ⎪
⎪
⎭
xraw2 ≥ 0.
Due to the simplicity of the example problem, we can give a graphical

representation of the set of feasible production plans (Figure 2).
Given the cost function γ(x) = 2xraw1 + 3xraw2 we easily conclude
(Figure 3) that
x̂raw1 = 36, x̂raw2 = 18, γ(x̂) = 126 (3.2)
is the unique optimal solution to our problem.
Our production problem is properly described by (3.1) and solved by (3.2)
provided the productivities, the unit costs, the demands and the capacity
(Table 2) are fixed data and known to us prior to making our decision on the
production plan. However, this is obviously not always a realistic assumption.
It may happen that at least some of the data—productivities and demands for
BASIC CONCEPTS 11
Figure 2 Deterministic LP: set of feasible production plans.
instance—can vary within certain limits (for our discussion, randomly) and
that we have to make our decision on the production plan before knowing the
exact values of those data.
To be more specific, let us assume that
• our model describes the weekly production process of a refinery relying

on two countries for the supply of crude oil (raw1 and raw2, respectively),
supplying one big company with gasoline (prod1) for its distribution system
of gas stations and another with fuel oil (prod2) for its heating and/or power
plants;
• it is known that the productivities π(raw1, prod1) and π(raw2, prod2), i.e.
the output of gas from raw1 and the output of fuel from raw2 may change
randomly (whereas the other productivities are deterministic);
• simultaneously, the weekly demands of the clients, hprod1 for gas and hprod2
for fuel are varying randomly;
• the weekly production plan (xraw1 , xraw2 ) has to be fixed in advance and
cannot be changed during the week, whereas
• the actual productivities are only observed (measured) during the
production process itself, and
• the clients expect their actual demand to be satisfied during the
corresponding week.
Figure 3 LP: feasible production plans and cost function for γ = 290.
Assume that, owing to statistics, we know that

⎫
hprod1 = 180 + ζ̃1 , ⎪
⎪
⎬
hprod2 = 162 + ζ̃2 ,
(3.3)
π(raw1, prod1) = 2 + η̃1 , ⎪
⎪
⎭
π(raw2, prod2) = 3.4 − η̃2 ,
where the random variables ζ̃j are modelled using normal distributions, and
η̃1 and η̃2 are distributed uniformly and exponentially respectively, with the
following parameters:1
⎫
distr ζ̃1 ∼ N (0, 12), ⎪
⎪
⎬
distr ζ̃2 ∼ N (0, 9),
(3.4)
distr η̃1 ∼ U[−0.8, 0.8], ⎪
⎪
⎭
distr η̃2 ∼ EX P(λ = 2.5).
For simplicity, we assume that these four random variables are mutually
independent. Since the random variables ζ̃1 , ζ̃2 and η̃2 are unbounded,
we restrict our considerations to their respective 99% confidence intervals
1 We use N (µ, σ) to denote the normal distribution with mean µ and variance σ2 .
BASIC CONCEPTS 13
(except for U). So we have for the above random variables’ realizations
⎫
ζ1 ∈ [−30.91, 30.91], ⎪
⎪
⎬
ζ2 ∈ [−23.18, 23.18],
(3.5)
η1 ∈ [−0.8, 0.8], ⎪
⎪
⎭
η2 ∈ [0.0, 1.84].
Hence, instead of the linear program (3.1), we are dealing with the stochastic
linear program
⎫
min(2xraw1 + 3xraw2 ) ⎪
⎪
xraw2 ≤ 100, ⎪
⎪
s.t. xraw1 + ⎪
⎪
⎬
(2 + η̃1 )xraw1 + 6xraw2 ≥ 180 + ζ̃1 ,
(3.6)
3xraw1 + (3.4 − η̃2 )xraw2 ≥ 162 + ζ̃2 , ⎪
⎪
⎪
⎪
xraw1 ≥ 0, ⎪
⎪
⎭
xraw2 ≥ 0.
This is not a well-defined decision problem, since it is not at all clear what
the meaning of “min” can be before knowing a realization (ζ1 , ζ2 , η1 , η2 ) of
(ζ̃1 , ζ̃2 , η̃1 , η̃2 ).
Geometrically, the consequence of our random parameter changes may
be rather complex. The effect of only the right-hand sides ζi varying
over the intervals given in (3.5) corresponds to parallel translations of the
corresponding facets of the feasible set as shown in Figure 4.
We may instead consider the effect of only the ηi changing their values
within the intervals mentioned in (3.5). That results in rotations of the related
facets. Some possible situations are shown in Figure 5, where the centers of
rotation are indicated by small circles.
Allowing for all the possible changes in the demands and in the
productivities simultaneously yields a superposition of the two geometrical
motions, i.e. the translations and the rotations. It is easily seen that the
variation of the feasible set may be substantial, depending on the actual
realizations of the random data. The same is also true for the so-called wait-
and-see solutions, i.e. for those optimal solutions we should choose if we knew
the realizations of the random parameters in advance. In Figure 6 a few
possible situations are indicated. In addition to the deterministic solution
x̂ = (x̂raw1 , x̂raw2 ) = (36, 18), γ = 126,
production plans such as

⎫
ŷ = (ŷraw1 , ŷraw2 ) = (20, 30), γ = 130, ⎬
ẑ = (ẑraw1 , ẑraw2) = (50, 22), γ = 166, (3.7)
⎭
v̂ = (v̂raw1 , v̂raw2 ) = (58, 6), γ = 134
Figure 4 LP: feasible set varying with demands.
Figure 5 LP: feasible set varying with productivities.

BASIC CONCEPTS 15
may be wait-and-see solutions.

Unfortunately, wait-and-see solutions are not what we need. We
have to decide production plans under uncertainty, since we only have
statistical information about the distributions of the random demands and
productivities.
A first possibility would consist in looking for a “safe” production program:
one that will be feasible for all possible realizations of the productivities and
demands. A production program like this is called a fat solution and reflects
total risk aversion of the decision maker. Not surprisingly, fat solutions are
usually rather expensive. In our example we can conclude from Figure 6 that
a fat solution exists at the intersection of the two rightmost constraints for
prod1 and prod2, which is easily computed as
x∗ = (x∗raw1 , x∗raw2 ) = (48.018, 25.548), γ ∗ = 172.681. (3.8)
To introduce another possibility, let us assume that the refinery has made
the following arrangement with its clients. In principle, the clients expect
the refinery to satisfy their weekly demands. However, very likely—according
to the production plan and the unforeseen events determining the clients’
demands and/or the refinery’s productivity—the demands cannot be covered
by the production, which will cause “penalty” costs to the refinery. The
amount of shortage has to be bought from the market. These penalties are
supposed to be proportional to the respective shortage in products, and we
assume that per unit of undeliverable products they amount to
qprod1 = 7, qprod2 = 12. (3.9)
The costs due to shortage of production—or in general due to the amount

of violation in the constraints—are actually determined after the observation
of the random data and are denoted as recourse costs. In a case (like ours) of
repeated execution of the production program it makes sense—according to
what we have learned from statistics—to apply an expected value criterion.
More precisely, we may want to find a production plan that minimizes the
sum of our original first-stage (i.e. production) costs and the expected recourse
costs. To formalize this approach, we abbreviate our notation. Instead of the
four single random variables ζ̃1 , ζ̃2 , η̃1 and η̃2 , it seems convenient to use the
random vector ξ˜ = (ζ̃1 , ζ̃2 , η̃1 , η̃2 )T . Further, we introduce for each of the
two stochastic constraints in (3.6) a recourse variable yi (ξ), ˜ i = 1, 2, which
simply measures the corresponding shortage in production if there is any;
since shortage depends on the realizations of our random vector ξ, ˜ so does the
˜
corresponding recourse variable, i.e. the yi (ξ) are themselves random variables.
Following the approach sketched so far, we now replace the vague stochastic
Figure 6 LP: feasible set varying with productivities and demands; some wait-
and-see solutions.
program (3.6) by the well defined stochastic program with recourse, using
˜ := hprod1 = 180 + ζ̃1 , h2 (ξ)
h1 (ξ) ˜ := hprod2 = 162 + ζ̃2 ,
˜ := π(raw1, prod1) = 2 + η̃1 , β(ξ)

α(ξ) ˜ := π(raw2, prod2) = 3.4 + η̃2 :
⎫
˜ + 12y2 (ξ)]}
min{2xraw1 + 3xraw2 + Eξ̃ [7y1 (ξ) ˜ ⎪
⎪
⎪
⎪
s.t. xraw1 + xraw2 ≤ 100, ⎪ ⎪
⎪
˜ raw1 +
α(ξ)x ˜
6xraw2 + y1 (ξ) ≥ ˜ ⎪
h1 (ξ), ⎪
⎪
⎪
⎬
˜ raw2
3xraw1 + β(ξ)x ˜
+ y2 (ξ) ≥ ˜
h2 (ξ), (3.10)
xraw1 ≥ 0, ⎪
⎪
⎪
⎪
xraw2 ≥ 0, ⎪
⎪
⎪
⎪
˜
y1 (ξ) ≥ 0, ⎪
⎪
⎪
⎭
˜
y2 (ξ) ≥ 0.
In (3.10) Eξ̃ stands for the expected value with respect to the distribution
˜ and in general, it is understood that the stochastic constraints have
of ξ,
to hold almost surely (a.s.) (i.e., they are to be satisfied with probability
1). Note that if ξ˜ has a finite discrete distribution {(ξ i , pi ), i = 1, · · · , r}
(pi > 0 ∀i) then (3.10) is just an ordinary linear program with a so-called
BASIC CONCEPTS 17
dual decomposition structure:

r ⎫
min{2xraw1 + 3xraw2 + i=1 pi [7y1 (ξ i ) + 12y2 (ξ i )]} ⎪
⎪
⎪
⎪
s.t. xraw1 + xraw2 ≤ 100, ⎪
⎪
⎪
i
α(ξ )xraw1 + i
6xraw2 + y1 (ξ ) ≥ h1 (ξ ) ∀i, ⎪
i ⎪
⎪
⎬
i
3xraw1 + β(ξ )xraw2 + y2 (ξ ) ≥ h2 (ξ ) ∀i,
i i
(3.11)
xraw1 ≥ 0, ⎪
⎪
⎪
⎪
xraw2 ≥ 0, ⎪
⎪
⎪
⎪
i
y1 (ξ ) ≥ 0 ∀i, ⎪
⎪
⎭
y2 (ξ ) ≥ 0 ∀i.
i
Depending on the number of realizations of ξ̃, r, this linear program may

become (very) large in scale, but its particular block structure is amenable
to specially designed algorithms. Linear programs with dual decomposition
structure will be introduced in general in Section 1.5 on page 42. A basic
solution method for these problems will be described in Section 1.7.4 (page 75).
To further analyse our refinery problem, let us first assume that only the
˜ i = 1, 2, are changing their values randomly, whereas the
demands, hi (ξ),
productivities are fixed. In this case we are in the situation illustrated in
Figure 4. Even this small idealized problem can present numerical difficulties
if solved as a nonlinear program. The reason for this lies in the fact that
the evaluation of the expected value which appears in the objective function
requires
• multivariate numerical integration;
• implicit definition of the functions ŷi (ξ) (these functions yielding for a fixed
x the optimal solutions of (3.10) for every possible realization ξ of ξ̃),
both of which are rather cumbersome tasks. To avoid these difficulties, we
shall try to approximate the normal distributions by discrete ones. For this
purpose, we
• generate large samples ζiµ , µ = 1, 2, · · · , K, i = 1, 2, restricted to the 99%
intervals of (3.5), sample size K =10 000;
• choose equidistant partitions of the 99% intervals into ri , i = 1, 2,
subintervals (e.g. r1 = r2 = 15);
• calculate for every subinterval Iiν , ν = 1, · · · , ri , i = 1, 2, the arithmetic
mean ζ̄iν of sample values ζiν ∈ Iiν , yielding an estimate for the conditional
expectation of ζ̃i given Iiν ;
• calculate for every subinterval Iiν the relative frequency piν for ζiµ ∈ Iiν
(i.e. piν = kiν /K, where kiν is the number of sample values ζiµ contained
in Iiν ). This yields an estimate for the probability of {ζi ∈ Iiν }.
The discrete distributions {(ζ̄iν , piν ), ν = 1, · · · , ri }, i = 1, 2, are then used
as approximations for the given normal distributions. Figure 7 shows these
discrete distributions for N (0, 12) and N (0, 9), with 15 realizations each.
6
%
br 15
rb br
12
r b b r
r ∼ N (0, 12)
9 b r
r b b ∼ N (0, 9)
r b
6 b r
r b b r
3
r b b r
r b b r
-
-30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30
Figure 7 Discrete distribution generated from N (0, 12), N (0, 9); (r1 , r2 ) =
(15, 15).
Obviously, these discrete distributions with 15 realizations each can

only be rough approximations of the corresponding normal distributions.
Therefore approximating probabilities of particular events using these discrete
distributions can be expected to cause remarkable discretization errors. This
will become evident in the following numerical examples.
Using these latter distributions, with 15 realizations each, we get 152 =
225 realizations for the joint distribution, and hence 225 blocks in our
decomposition problem. This yields as an optimal solution for the linear
program (3.11) (with γ(·) the total objective of (3.11) and γI (x) = 2xraw1 +
3xraw2 )
x̃ = (x̃1 , x̃2 ) = (38.539, 20.539), γ(x̃) = 140.747, (3.12)
with corresponding first-stage costs of
γI (x̃) = 138.694.
Defining ρ(x) as the empirical reliability (i.e. the probability to be feasible) for
any production plan x, we find—with respect to the approximating discrete
distribution—for our solution x̃ that
ρ(x̃) = 0.9541,
whereas using our original linear program’s solution x̂ = (36, 18) would yield
the total expected cost
γ(x̂) = 199.390
BASIC CONCEPTS 19
and an empirical reliability of
ρ(x̂) = 0.3188,
which is clearly overestimated (compared with its theoretical value of 0.25),

which indicates that the crude method of discretization we use here just
for demonstration has to be refined, either by choosing a finer discretization
or preferably—in view of the numerical workload drastically increasing with
the size of the discrete distributions support—by finding a more appropriate
strategy for determining the subintervals of the partition.
Let us now consider the effect of randomness in the productivities. To
this end, we assume that hi (ξ), ˜ i = 1, 2, are fixed at their expected
values and the two productivities α(ξ) ˜ and β(ξ)˜ behave according to their
distributions known from (3.3) and (3.4). Again we discretize the given
distributions confining ourselves to 15 and 18 subintervals for the uniform
and the exponential distributions respectively, yielding 15 × 18 = 270 blocks
in (3.11). Solving the resulting stochastic program with recourse (3.11) as an
ordinary linear program, we get as solution x̄
x̄ = (37.566, 22.141), γ(x̄) = 144.179, γI (x̄) = 141.556,
whereas the solution of our original LP (3.1) would yield as total expected
costs
γ(x̂) = 204.561.
For the reliability, we now get
ρ(x̄) = 0.9497,
in contrast to
ρ(x̂) = 0.2983
for the LP solution x̂.
Finally we consider the most general case of α(ξ), ˜ β(ξ),
˜ h1 (ξ)
˜ and h2 (ξ)
˜
varying randomly where the distributions are discretely approximated by 5-,
9-, 7- and 11-point distributions respectively, in an analogous manner to the
above. This yields a joint discrete distribution of 5 × 9 × 7 × 11 = 3465
realizations and hence equally many blocks in the recourse problem (3.11);
in other words, we have to solve a linear program with 2 × 3465 + 1 = 6931
constraints! The solution x̌ amounts to
x̌ = (37.754, 23.629), γ(x̌) = 150.446, γI (x̌) = 146.396,
with a reliability of
ρ(x̌) = 0.9452,
whereas the LP solution x̂ = (36, 18) would yield
γ(x̂) = 232.492, ρ(x̂) = 0.2499.
So far we have focused on the case where decisions, turning out post festum
to be the wrong ones, imply penalty costs that depend on the magnitude of
constraint violations. Afterwards, we were able to determine the reliability
of the resulting decisions, which represents a measure of feasibility. Note
that the reliability provides no indication of the size of possible constraint
violations and corresponding penalty costs. Nevertheless, there are many real
life decision situations where reliability is considered to be the most important
issue—either because it seems impossible to quantify a penalty or because of
questions of image or ethics. Examples may be found in various areas such as
medical problems as well as technical applications.
For instance, suppose once again that only the demands are random.
Suppose further that the management of our refinery is convinced that it
is absolutely necessary—in order to maintain a client base—to maintain a
reliability of 95% with respect to satisfying their demands. In this case we may
formulate the following stochastic program with joint probabilistic constraints:
min(2xraw1 + 3xraw2 )
s.t. xraw1 + xraw2 ≤ 100,
xraw1 ≥ 0,
xraw2 ≥ 0,

2xraw1 ˜
+ 6xraw2 ≥ h1 (ξ)
P ˜ ≥ 0.95.
3xraw1 + 3xraw2 ≥ h2 (ξ)
This problem can be solved with appropriate methods, one of which will be
presented later in this text. It seems worth mentioning that in this case
using the normal distributions instead of their discrete approximations is
appropriate owing to theoretical properties of probabilistic constraints to be
discussed later on. The solution of the probabilistically constrained program
is
z = (37.758, 21.698), γI (z) = 140.612.
So the costs—i.e. the first-stage costs—are only slightly increased compared
with the LP solution if we observe the drastic increase of reliability. There
seems to be a contradiction on comparing this last result with the solution
(3.12) in that γI (x̃) < γI (z) and ρ(x̃) > 0.95; however, this discrepancy is due
to the discretization error made by replacing the true normal distribution of
(ξ˜1 , ξ̃2 ) by the 15 × 15 discrete distribution used for the computation of the
solution (3.12). Using the correct normal distribution would obviously yield
γI (x̃) = 138.694 (as in (3.12)), but only ρ(x̃) = 0.9115!
BASIC CONCEPTS 21
1.4 Stochastic Programs: General Formulation
In the same way as random parameters in (3.1) led us to the stochastic (linear)
program (3.6), random parameters in (2.3) may lead to the problem
⎫
“min”g0 (x, ξ) ˜ ⎬
˜ ≤ 0, i = 1, · · · , m,
s.t. gi (x, ξ) (4.1)
⎭
x ∈ X ⊂ IRn ,
where ξ˜ is a random vector varying over a set Ξ ⊂ IRk . More precisely, we

assume throughout that a family F of “events”, i.e. subsets of Ξ, and the
probability distribution P on F are given. Hence for every subset A ⊂ Ξ that
is an event, i.e. A ∈ F, the probability P (A) is known. Furthermore, we assume
that the functions gi (x, ·) : Ξ → IR ∀x, i are random variables themselves, and
that the probability distribution P is independent of x.
However, problem (4.1) is not well defined since the meanings of “min” as
well as of the constraints are not clear at all, if we think of taking a decision
on x before knowing the realization of ξ̃. Therefore a revision of the modelling
process is necessary, leading to so-called deterministic equivalents for (4.1),
which can be introduced in various ways, some of which we have seen for
our example in the previous section. Before discussing them, we review some
basic concepts in probability theory, and fix the terminology and notation
used throughout this text.
1.4.1 Measures and Integrals

k
In IR we denote sets of the type
I[a,b) = {x ∈ IRk | ai ≤ xi < bi , i = 1, · · · , k}
as (half-open) intervals. In geometric terms, depending on the dimension k of

IRk , I[a,b) is
• an interval if k = 1,
• a rectangle if k = 2,
• a cube if k = 3,
while for k > 3 there is no common language term for these objects since
geometric imagination obviously ends there.
Sometimes we want to know something about the “size” of a set in IRk , e.g.
the length of a beam, the area of a piece of land or the volume of a building;
in other words, we want to measure it. One possibility to do this is to fix
first how we determine the measure of intervals, and a “natural” choice of a
measure µ would be

1 b−a if a ≤ b,
• in IR : µ(I[a,b) ) =
0 otherwise,

2 (b1 − a1 )(b2 − a2 ) if a ≤ b,
• in IR : µ(I[a,b) ) =
0 otherwise,

(b1 − a1 )(b2 − a2 )(b3 − a3 ) if a ≤ b,
• in IR3 : µ(I[a,b) ) =
0 otherwise.
Analogously, in general for I[a,b) ⊂ IRk with arbitrary k, we have

⎧
⎨
⎪ k
(bi − ai ) if a ≤ b
µ(I[a,b) ) = (4.2)
⎪
⎩ i=1
0 else.
Obviously for a set A that is the disjoint finite union of intervals, i.e.
(n)
A = ∪M n=1 I , I (n) being intervals such that I (n) ∩ I (m) = ∅ for n = m,
M (n)
we define its measure as µ(A) = n=1 µ(I ). In order to measure a set
A that is not just an interval or a finite union of disjoint intervals, we may
proceed as follows.
Any finite collection of pairwise-disjoint intervals contained in A forms
a packing C of A, C being the union of those intervals, with a well-
defined measure µ(C) as mentioned above. Analogously, any finite collection
of pairwise disjoint intervals, with their union containing A, forms a covering
D of A with a well-defined measure µ(D).
Take for example in IR2 the set
Acirc = {(x, y) | x2 + y 2 ≤ 16, y ≥ 0},
i.e. the half-circle illustrated in Figure 8, which also shows a first possible
packing C1 and covering D1 . Obviously we learned in high school that the
area of Acirc is computed as µ(Acirc ) = 12 × π × (radius)2 = 25.1327, whereas
we easily compute µ(C1 ) = 13.8564 and µ(D1 ) = 32. If we forgot all our
wisdom from high school, we would only be able to conclude that the measure
of the half-circle Acirc is between 13.8564 and 32. To obtain a more precise
estimate, we can try to improve the packing and the covering in such a way
that the new packing C2 exhausts more of the set Acirc and the new covering
D2 becomes a tighter outer approximation of Acirc . This is shown in Figure 9,
for which we get µ(C2 ) = 19.9657 and µ(D2 ) = 27.9658.
Hence the measure of Acirc is between 19.9657 and 27.9658. If this is still
not precise enough, we may further improve the packing and covering. For
the half-cirle Acirc , it is easily seen that we may determine its measure in this
way with any arbitrary accuracy.
In general, for any closed bounded set A ⊂ IRk , we may try a similar
procedure to measure A. Denote by CA the set of all packings for A and by
BASIC CONCEPTS 23
Figure 8 Measure of a half-circle: first approximation.
Figure 9 Improved approximate measure of a half-circle.

DA the set of all coverings of A. Then we make the following definition.
The closed bounded set A is measurable if
sup{µ(C) | C ∈ CA } = inf{µ(D) | D ∈ DA },
with the measure µ(A) = sup{µ(C) | C ∈ CA }.
To get rid of the boundedness restriction, we may extend this definition

immediately by saying:
An arbitrary closed set A ⊂ IRk is measurable iff2 for every interval

I[a,b) ⊂ IRk the set A ∩ I[a,b) is measurable (in the sense defined before).
This implies that IRk itself is measurable. Observing that there always exist
collections of countably many pairwise-disjoint intervals I[aν ,bν ) , ν = 1, 2, · · · ,
∞
covering IRk , i.e. ν=1 I[a
ν ,bν ) = IR
k
(e.g. take intervals with all edges having
∞
length 1), we get µ(A) = ν=1 µ(A ∩ I[aν ,bν ) ) as the measure of A. Obviously
µ(A) = ∞ may happen, as it does for instance with A = IR2+ (i.e. the positive
orthant of IR2 ) or with A = {(x, y) ∈ IR2 | x ≥ 1, 0 ≤ y ≤ x1 }. But we also
may find unbounded sets with finite measure as e.g. A = {(x, y) ∈ IR2 | x ≥
0, 0 ≤ y ≤ e−x } (see the exercises at the end of this chapter).
The measure introduced this way for closed sets and based on the
elementary measure for intervals as defined in (4.2) may be extended as a
“natural” measure for the class A of measurable sets in IRk , and will be
denoted throughout by µ. We just add that A is characterized by the following
properties:
if A ∈ A then also IRk − A ∈ A; (4.3 i)

∞

if Ai ∈ A, i = 1, 2, · · · , then also Ai ∈ A. (4.3 ii)
i=1
∞
This implies that with Ai ∈ A, i = 1, 2, · · ·, also i=1 Ai ∈ A.
As a consequence of the above construction, we have, for the natural
measure µ defined in IRk , that
µ(A) ≥ 0 ∀A ∈ A and µ(∅) = 0; (4.4 i)

if Ai ∈ A,
∞i = 1, 2, ·
· · , and Ai ∩ Aj = ∅ for i = j,
∞ (4.4 ii)
then µ( i=1 Ai ) = i=1 µ(Ai ).
In other words, the measure of a countable disjoint union of measurable sets

equals the countable sum of the measures of these sets.
2 “iff” stands for “if and only if”
BASIC CONCEPTS 25
These properties are also familiar from probability theory: there we have
some space Ω of outcomes ω (e.g. the results of random experiments), a
collection F of subsets F ⊂ Ω called events, and a probability measure (or
probability distribution) P assigning to each F ∈ F the probability with
which it occurs. To set up probability theory, it is then required that
(i) Ω is an event, i.e. Ω ∈ F, and, with F ∈ F, it holds that also Ω − F ∈ F,
i.e. if F is an event then so also is its complement (or notF );
(ii) the countable union of events is an event.
Observe that these formally coincide with (4.3) except that Ω can be any
space of objects and need not be IRk .
For the probability measure, it is required that
(i) P (F ) ≥ 0 ∀F ∈ F and P (Ω) = 1;

if Fi ∈ F, i = 1, 2, · · · , and Fi ∩ Fj = ∅ for i = j, then P ( ∞
(ii) i=1 Fi ) =
∞
i=1 P (Fi ).
The only difference with (4.4) is that P is bounded to P (F ) ≤ 1 ∀F ∈ F,
whereas µ is unbounded on IRk . The triple (Ω, F , P ) with the above properties
is called a probability space.
In addition, in probability theory we find random variables and random
vectors ξ̃. With A the collection of naturally measurable sets in IRk , a random
vector is a function (i.e. a single-valued mapping)
ξ˜ : Ω −→ IRk such that, for all A ∈ A, ξ̃ −1 [A] := {ω | ξ̃(ω) ∈ A} ∈ F. (4.5)

˜ of any measurable
This requires the “inverse” (with respect to the function ξ)
k
set in IR to be an event in Ω.
Observe that a random vector ξ˜ : Ω −→ IRk induces a probability measure
Pξ̃ on A according to
Pξ̃ (A) = P ({ω | ξ̃(ω) ∈ A}) ∀A ∈ A.
Example 1.2 At a market hall for the fruit trade you find a particular species
of apples. These apples are traded in certain lots (e.g. of 1000 lb). Buying a lot
involves some risk with respect to the quality of apples contained in it. What
does “quality” mean in this context? Obviously quality is a conglomerate of
criteria described in terms like size, ripeness, flavour, colour and appearance.
Some of the criteria can be expressed through quantitative measurement, while
others cannot (they have to be judged upon by experts). Hence the set Ω of
all possible “qualities” cannot as such be represented as a subset of some IRk .
Having bought a lot, the trader has to sort his apples according to their
“outcomes” (i.e. qualities), which could fall into “events” like “unusable”
(e.g. rotten or too unripe), “cooking apples”and “low (medium, high) quality
eatable apples”. Having sorted out the “unusable” and the “cooking apples”,
for the remaining apples experts could be asked to judge on ripeness, flavour,
colour and appearance, by assigning real values between 0 and 1 to parameters
r, f, c and a respectively, corresponding to the “degree (or percentage) of
achieving” the particular criterion.
Now we can construct a scalar value for any particular outcome (quality)
ω, for instance as
⎧
⎨ 0 if ω ∈ “unusable”,
1
ṽ(ω) := if ω ∈ “cooking apples”,
⎩ 2
(1 + r)(1 + f )(1 + c)(1 + a) otherwise.
Obviously ṽ has the range ṽ[Ω] = {0} ∪ { 12 } ∪ {[1, 16]}. Denoting the events
“unusable” by U and “cooking apples” by C, we may define the collection F
of events as follows. With G denoting the family of all subsets of Ω − (U ∪ C)
let F contain all unions of U, C, ∅ or Ω with any element of G. Assume that
after long series of observations we have a good estimate for the probabilities
P (A), A ∈ F.
According to our scale, we could classify the apples as
• eatable and
– 1st class for ṽ(ω) ∈ [12, 16] (high selling price),
– 2nd class for ṽ(ω) ∈ [8, 12) (medium price),
– 3rd class for ṽ(ω) ∈ [1, 8) (low price);
1
• good for cooking for ṽ(ω) = 2 (cheap);
• waste for ṽ(ω) = 0.
Obviously the probabilities to have 1st-class apples in our lot is
Pṽ ({[12, 16]}) = P (ṽ −1 [{[12, 16]}]), whereas the probability for having 3rd-
class or cooking apples amounts to
Pṽ ({[1, 8)} ∪ { 21 }) = P (ṽ −1 [{[1, 8)} ∪ { 21 }])
= P (ṽ −1 [{[1, 8)}]) + P (C),
using the fact that ṽ is single-valued and {[1, 8)}, { 21 } and hence ṽ −1 [{[1, 8)}],
ṽ −1 [{ 12 }] = C are disjoint. For an illustration, see Figure 10. 2
If it happens that Ω ⊂ IRk and F ⊂ A (i.e. every event is a “naturally”

˜
measurable set) then we may replace ω trivially by ξ(ω) by just applying the
˜
identity mapping ξ(ω) ≡ ω, which preserves the probability measure Pξ̃ on
F , i.e.
Pξ̃ (A) = P (A) for A ∈ F
BASIC CONCEPTS 27
Figure 10 Classification of apples by quality.
since obviously {ω | ξ̃(ω) ∈ A} = A if A ∈ F.

In any case, given a random vector ξ̃ with Ξ ∈ A such that {ω | ξ(ω) ˜ ∈
k
Ξ} = Ω (observe that Ξ = IR always satisfies this, but there may be smaller
sets in A that do so), with F̂ = {B | B = A ∩ Ξ, A ∈ A}, instead of the
abstract probability space (Ω, F , P ) we may equivalently consider the induced
probability space (Ξ, F̂ , Pξ̃ ), which we shall use henceforth and therefore
denote as (Ξ, F , P ). We shall use ξ˜ for the random vector and ξ for the
elements of Ξ (i.e. for the possible realizations of ξ̃).
Sometimes we like to assert a special property (like continuity or
differentiability of some function f : IRk −→ IR) everywhere in IRk . But it
may happen that this property almost always holds except on some particular
points of IRk like N1 = {finitely many isolated points} or (for k ≥ 2) N2 =
{finitely many segments of straight lines}, the examples mentioned being
(‘naturally’) measurable and having the natural measure µ(N1 ) = µ(N2 ) = 0.
In a situation like this, more precisely if there is a set Nδ ∈ A with µ(Nδ ) = 0,
and if our property holds for all x ∈ IRk − Nδ , we say that it holds almost
everywhere (a.e.). In the context of a probability space (Ξ, F , P ), if there is
an event Nδ ∈ F with P (Nδ ) = 0 such that a property holds on Ξ − Nδ , owing
to the practical interpretation of probabilities, we say that the property holds
almost surely (a.s.).
Figure 11 Integrating a simple function.
Next let us briefly review integrals. Consider first IRk with A, its measurable
sets, and the natural measure µ, and choose some bounded measurable set
B ∈ A. Further, let {A1 , · · · , Ar } be a partition
r of B into measurable sets,
i.e. Ai ∈ A, Ai ∩ Aj = ∅ for i = j, and i=1 Ai = B. Given the indicator
functions χAi : B −→ IR defined by

1 if x ∈ Ai ,
χAi (x) =
0 otherwise,
we may introduce a so-called simple function ϕ : B −→ IR given with some

constants ci by

r
ϕ(x) = ci χAi (x)
i=1
= ci for x ∈ Ai .

Then the integral B
ϕ(x)dµ is defined as

r
ϕ(x)dµ = ci µ(Ai ). (4.6)
B i=1
In Figure 11 the integral would result by accumulating the shaded areas with
their respective signs as indicated.
BASIC CONCEPTS 29
Figure 12 Integrating an arbitrary function.
Observe that the sum (or difference) of simple functions ϕ1 and ϕ2 is again
a simple function and that

[ϕ1 (x) + ϕ2 (x)]dµ = ϕ1 (x)dµ + ϕ2 (x)dµ
B B B

ϕ(x)dµ ≤ |ϕ(x)|dµ

B B
from the elementary properties of finite sums. Furthermore, it is easy to see

that for disjoint measurable
sets (i.e. Bj ∈ A, j = 1, · · · , s, and Bj ∩ Bl = ∅
for j = l) such that sj=1 Bj = B, it follows that
s

ϕ(x)dµ = ϕ(x)dµ.
B j=1 Bj
To integrate any other function ψ : B −→ IR that is not a simple function,

we use simple functions to approximate ψ (see Figure 12), and whose integrals
converge.
Any sequence {ϕn } of simple functions on B satisfying

|ϕn (x) − ϕm (x)|dµ −→ 0 for n, m −→ ∞
B
is called mean fundamental. If there exists a sequence {ϕn } such that

ϕn (x) −→ ψ(x) a.e.3 and {ϕn } is mean fundamental
3 The convergence a.e. can be replaced by another type of convergence, which we omit here.

then the integral B ψ(x)dµ is defined by

ψ(x)dµ = lim ϕn (x)dµ
B n→∞ B
and ψ is called integrable.

Observe that

ϕn (x)dµ − ϕm (x)dµ ≤ |ϕn (x) − ϕm (x)|dµ,

B B B

such that { B ϕn (x)dµ} is a Cauchy sequence. Therefore limn→∞ B ϕn (x)dµ
exists. It can be shown that this definition yields a uniquely determined
value for the integral, i.e. it cannot happen that a choice of another mean
fundamental sequence of simple functions converging a.e. to ψ yields a different
value for the integral.
The boundedness of B is not absolutely essential here; with a slight
modification of the assumption “ϕn (x) −→ ψ(x) a.e.” the integrability of
ψ may be defined analogously.
Now it should be obvious that, given a probability space (Ξ, F , P )—
assumed to be introduced by a random vector ξ̃ in IRk —and a function
ψ : Ξ −→ IR, the integral with respect to the probability measure P , denoted
by
˜ =
E ψ(ξ) ψ(ξ)dP,
ξ̃
Ξ
can be derived exactly as above if we simply replace the measure µ by the

probability measure P . Here E refers to expectation and ξ˜ indicates that we
are integrating with respect to the probability measure P induced by the
random vector ξ̃.
Finally, we recall that in probability theory the probability measure P of a
probability space (Ξ, F , P ) in IRk is equivalently described by the distribution
function Fξ̃ defined by
Fξ̃ (x) = P ({ξ | ξ ≤ x}), x ∈ IRk .
If there exists a function fξ̃ : Ξ −→ IR such that the distribution function can
be represented by an integral with respect to the natural measure µ as

Fξ̃ (x̂) = fξ̃ (x)dµ, x̂ ∈ IRk ,
x≤x̂
then fξ̃ is called the density function of P . In this case the distribution function
is called of continuous type. It follows that for any event A ∈ F we have
P (A) = A fξ̃ (x)dµ. This implies in particular that for any A ∈ F such that
BASIC CONCEPTS 31
µ(A) = 0 also P (A) = 0 has to hold. This fact is referred to by saying that
the probability measure P is absolutely continuous with respect to the natural
measure µ. It can be shown that the reverse statement is also true: given a
probability space (Ξ, F , P ) in IRk with P absolutely continuous with respect
to µ (i.e. every event A ∈ F with the natural measure µ(A) = 0 has also a
probability of zero), there exists a density function fξ̃ for P .
1.4.2 Deterministic Equivalents

Let us now come back to deterministic equivalents for (4.1). For instance, in
analogy to the particular stochastic linear program with recourse (3.10), for
problem (4.1) we may proceed as follows. With

0 if gi (x, ξ) ≤ 0,
gi+ (x, ξ) =
gi (x, ξ) otherwise,
the ith constraint of (4.1) is violated if and only if gi+ (x, ξ) > 0 for a given
decision x and realization ξ of ξ̃. Hence we could provide for each constraint a
recourse or second-stage activity yi (ξ) that, after observing the realization ξ,
is chosen such as to compensate its constraint’s violation—if there is one—by
satisfying gi (x, ξ) − yi (ξ) ≤ 0. This extra effort is assumed to cause an extra
cost or penalty of qi per unit, i.e. our additional costs (called the recourse
function) amount to
m

+
Q(x, ξ) = min qi yi (ξ) yi (ξ) ≥ gi (x, ξ), i = 1, · · · , m , (4.7)
y
i=1
yielding a total cost—first-stage and recourse cost—of
f0 (x, ξ) = g0 (x, ξ) + Q(x, ξ). (4.8)
Instead of (4.7), we might think of a more general linear recourse program

with a recourse vector y(ξ) ∈ Y ⊂ IRn (Y is some given polyhedral set, such
as {y | y ≥ 0}), an arbitrary fixed m × n matrix W (the recourse matrix)
and a corresponding unit cost vector q ∈ IRn , yielding for (4.8) the recourse
function
Q(x, ξ) = min{q T y | W y ≥ g + (x, ξ), y ∈ Y }, (4.9)
y
where g + (x, ξ) = (g1+ (x, ξ), · · · , gm

+
(x, ξ))T .
If we think of a factory producing m products, gi (x, ξ) could be understood
as the difference {demand} − {output} of a product i. Then gi+ (x, ξ) >
0 means that there is a shortage in product i, relative to the demand.
Assuming that the factory is committed to cover the demands, problem (4.7)
could for instance be interpreted as buying the shortage of products at the
market. Problem (4.9) instead could result from a second-stage or emergency

production program, carried through with the factor input y and a technology
represented by the matrix W . Choosing W = I, the m × m identity matrix,
(4.7) turns out to be a special case of (4.9).
Finally we also could think of a nonlinear recourse program to define the
recourse function for (4.8); for instance, Q(x, ξ) could be chosen as
Q(x, ξ) = min{q(y) | Hi (y) ≥ gi+ (x, ξ), i = 1, · · · , m; y ∈ Y ⊂ IRn }, (4.10)
where q : IRn → IR and Hi : IRn → IR are supposed to be given.

In any case, if it is meaningful and acceptable to the decision maker to
minimize the expected value of the total costs (i.e. first-stage and recourse
costs), instead of problem (4.1) we could consider its deterministic equivalent,
the (two-stage) stochastic program with recourse
˜ = min E {g0 (x, ξ)

min Eξ̃ f0 (x, ξ) ˜ + Q(x, ξ)}.
˜ (4.11)
x∈X ξ̃x∈X
The above two-stage problem is immediately extended to the multistage

recourse program as follows: instead of the two decisions x and y, to be
taken at stages 1 and 2, we are now faced with K + 1 sequential decisions
x0 , x1 , · · · , xK (xτ ∈ IRn̄τ ), to be taken at the subsequent stages τ =
0, 1, · · · , K. The term “stages” can, but need not, be interpreted as “time
periods”.
Assume for simplicity that the objective of (4.1) is deterministic, i.e.
g0 (x, ξ) ≡ g0 (x0 ). At stage τ (τ ≥ 1) we know the realizations ξ1 , · · · , ξτ of
the random vectors ξ˜1 , · · · , ξ˜τ as well as the previous decisions x0 , · · · , xτ −1 ,
and we have to decide on xτ such that the constraint(s) (with vector valued
constraint functions gτ )
gτ (x0 , · · · , xτ , ξ1 , · · · , ξτ ) ≤ 0
are satisfied, which—as stated—at this stage can only be achieved by the
proper choice of xτ , based on the knowledge of the previous decisions and
realizations. Hence, assuming a cost function qτ (xτ ), at stage τ ≥ 1 we have
a recourse function
Qτ (x0 , x1 , · · · , xτ −1 , ξ1 , · · · , ξτ ) = min{qτ (xτ ) | gτ (x0 , · · · , xτ , ξ1 , · · · , ξτ ) ≤ 0}

xτ
indicating that the optimal recourse action x̂τ at time τ depends on the
previous decisions and the realizations observed until stage τ , i.e.
x̂τ = x̂τ (x0 , · · · , xτ −1 , ξ1 , · · · , ξτ ), τ ≥ 1.

BASIC CONCEPTS 33
Hence, taking into account the multiple stages, we get as total costs for the
multistage problem

K
f0 (x0 , ξ1 , · · · , ξK ) = g0 (x0 ) + Qτ (x0 , x̂1 , · · · , x̂τ −1 , ξ1 , · · · , ξτ ) (4.12)
τ =1
yielding the deterministic equivalent for the described dynamic decision

problem, the multistage stochastic program with recourse

K
min [g0 (x0 ) + Eξ̃1 ,···,ξ̃τ Qτ (x0 , x̂1 , · · · , x̂τ −1 , ξ̃1 , · · · , ξ˜τ )], (4.13)
x0 ∈X
τ =1
obviously a straight generalization of our former (two-stage) stochastic

program with recourse (4.11).
For the two-stage case, in view of their practical relevance it is worthwile
to describe briefly some variants of recourse problems in the stochastic linear
programming setting. Assume that we are given the following stochastic linear
program
⎫
“min”cT x ⎪
⎪
⎬
s.t. Ax = b,
˜ = h(ξ),˜ ⎪ (4.14)
T (ξ)x ⎪
⎭
x ≥ 0.
Comparing this with the general stochastic program (4.1), we see that the set
X ⊂ IRn is specified as
X = {x ∈ IRn | Ax = b, x ≥ 0},
where the m0 × n matrix A and the vector b are assumed to be deterministic.

In contrast, the m1 × n matrix T (·) and the vector h(·) are allowed to depend
on the random vector ξ,˜ and therefore to have random entries themselves. In
general, we assume that this dependence on ξ ∈ Ξ ⊂ IRk is given as

T (ξ) = T̂ 0 + ξ1 T̂ 1 + · · · + ξk T̂ k ,
(4.15)
h(ξ) = ĥ0 + ξ1 ĥ1 + · · · + ξk ĥk ,
with deterministic matrices T̂ 0 , · · · , T̂ k and vectors ĥ0 , · · · , ĥk . Observing that

the stochastic constraints in (4.14) are equalities (instead of inequalities, as
in the general problem formulation (4.1)), it seems meaningful to equate their
deficiencies, which, using linear recourse and assuming that Y = {y ∈ IRn |
y ≥ 0}, according to (4.9) yields the stochastic linear program with fixed
recourse
⎫
˜
minx Eξ̃ {cT x + Q(x, ξ)} ⎪
⎪
⎪
s.t. Ax = b ⎪
⎬
x ≥ 0, (4.16)
⎪
⎪
where ⎪
⎪
⎭
Q(x, ξ) = min{q T y | W y = h(ξ) − T (ξ)x, y ≥ 0} .
In particular, we speak of complete fixed recourse if the fixed m1 × n recourse
matrix W satisfies
{z | z = W y, y ≥ 0} = IRm1 . (4.17)
This implies that, whatever the first-stage decision x and the realization ξ of
ξ˜ turn out to be, the second-stage program
Q(x, ξ) = min{q T y | W y = h(ξ) − T (ξ)x, y ≥ 0}
will always be feasible. A special case of complete fixed recourse is simple

recourse, where with the identity matrix I of order m1 :
W = (I, −I). (4.18)
Then the second-stage program reads as
Q(x, ξ) = min{(q + )T y + +(q − )T y − | y + −y − = h(ξ)−T (ξ)x, y + ≥ 0, y − ≥ 0},
i.e., for q + + q − ≥ 0, the recourse variables y + and y − can be chosen to

measure (positively) the absolute deficiencies in the stochastic constraints.
Generally, we may put all the above problems into the following form:
⎫
min Eξ̃ f0 (x, ξ)˜ ⎪
⎪
⎪
⎬
˜ ≤ 0, i = 1, · · · , s,
s.t. Eξ̃ fi (x, ξ)
(4.19)
˜ = 0, i = s + 1, · · · , m̄, ⎪
Eξ̃ fi (x, ξ) ⎪
⎪
⎭
x ∈ X ⊂ IRn ,
where the fi are constructed from the objective and the constraints in (4.1)
or (4.14) respectively. So far, f0 represented the total costs (see (4.8) or (4.12))
and f1 , · · · , fm̄ could be used to describe the first-stage feasible set X.
However, depending on the way the functions fi are derived from the problem
functions gj in (4.1), this general formulation also includes other types of
deterministic equivalents for the stochastic program (4.1).
To give just two examples showing how other deterministic equivalent
problems for (4.1) may be generated, let us choose first α ∈ [0, 1] and define
a “payoff” function for all constraints as

1 − α if gi (x, ξ) ≤ 0, i = 1, · · · , m,
ϕ(x, ξ) :=
−α otherwise.
BASIC CONCEPTS 35
Consequently, for x infeasible at ξ we have an absolute loss of α, whereas for

x feasible at ξ we have a return of 1 − α. It seems natural to aim for decisions
on x that, at least in the mean (i.e. on average), avoid an absolute loss. This
is equivalent to the requirement

˜ =
Eξ̃ ϕ(x, ξ) ϕ(x, ξ)dP ≥ 0.
Ξ
Defining f0 (x, ξ) = g0 (x, ξ) and f1 (x, ξ) := −ϕ(x, ξ), we get

⎫
f0 (x, ξ) = g0 (x, ξ), ⎪
⎬

α−1 if gi (x, ξ) ≤ 0, i = 1, · · · , m, (4.20)
f1 (x, ξ) = ⎪
⎭
α otherwise,
implying
˜ = −E ϕ(x, ξ)
Eξ̃ f1 (x, ξ) ˜ ≤ 0,
ξ̃
where, with the vector-valued function g(x, ξ) = (g1 (x, ξ), · · · , gm (x, ξ))T ,

˜ =
Eξ̃ f1 (x, ξ) f1 (x, ξ)dP
Ξ

= (α − 1)dP + αdP
{g(x,ξ)≤0} {g(x,ξ)≤0}
= (α − 1)P ({ξ | g(x, ξ) ≤ 0}) + αP ({ξ | g(x, ξ) ≤ 0})

= α [P ({ξ | g(x, ξ) ≤ 0}) + P ({ξ | g(x, ξ) ≤ 0})]

=1
−P ({ξ | g(x, ξ) ≤ 0}).
˜ ≤ 0 is equivalent to P ({ξ | g(x, ξ) ≤ 0}) ≥

Therefore the constraint Eξ̃ f1 (x, ξ)
α. Hence, under these assumptions, (4.19) reads as

minx∈X Eξ̃ g0 (x, ξ) ˜
(4.21)
s.t. P ({ξ | gi (x, ξ) ≤ 0, i = 1, · · · , m}) ≥ α.
Problem (4.21) is called a probabilistically constrained or chance constrained

program (or a problem with joint probabilistic constraints).
If instead of (4.20) we define αi ∈ [0, 1], i = 1, · · · , m, and analogous
“payoffs” for every single constraint, resulting in
f0 (x, ξ) = g0 (x, ξ)

αi − 1 if gi (x, ξ) ≤ 0,
fi (x, ξ) = i = 1, · · · , m,
αi otherwise,
then we get from (4.19) the problem with single (or separate) probabilistic
constraints:

minx∈X Eξ̃ g0 (x, ξ) ˜
(4.22)
s.t. P ({ξ | gi (x, ξ) ≤ 0}) ≥ αi , i = 1, · · · , m.
If, in particular, we have that the functions gi (x, ξ) are linear in x, and if
furthermore the set X is convex polyhedral, i.e. we have the stochastic linear
program ⎫
“min” cT (ξ)x
˜ ⎪
⎪
⎬
s.t. Ax = b,
˜ ≥ h(ξ),
T (ξ)x ˜ ⎪
⎪
⎭
x ≥ 0,
then problems (4.21) and (4.22) become

minx∈X Eξ̃ cT (ξ)x
˜
(4.23)
s.t. P ({ξ | T (ξ)x ≥ h(ξ)}) ≥ α,
and, with Ti (·) and hi (·) denoting the ith row and ith component of T (·) and
h(·) respectively,

minx∈X Eξ̃ cT (ξ)x˜
(4.24)
s.t. P ({ξ | Ti (ξ)x ≥ hi (ξ)}) ≥ αi , i = 1, · · · , m,
the stochastic linear programs with joint and with single chance constraints
respectively.
Obviously there are many other possibilities to generate types of
deterministic equivalents for (4.1) by constructing the fi in different ways
out of the objective and the constraints of (4.1).
Formally, all problems derived, i.e. all the above deterministic equivalents,
are mathematical programs. The first question is, whether or under which
assumptions do they have properties like convexity and smoothness such that
we have any reasonable chance to deal with them computationally using the
toolkit of mathematical programming methods.
1.5 Properties of Recourse Problems
Convexity may be shown easily for the recourse problem (4.11) under rather
mild assumptions (given the integrability of g0 + Q).
Proposition 1.1 If g0 (·, ξ) and Q(·, ξ) are convex in x ∀ξ ∈ Ξ, and if X is
a convex set, then (4.11) is a convex program.
BASIC CONCEPTS 37
Proof For x̂, x̄ ∈ X, λ ∈ (0, 1) and x̌ := λx̂ + (1 − λ)x̄ we have

g0 (x̌, ξ) + Q(x̌, ξ) ≤ λ[g0 (x̂, ξ) + Q(x̂, ξ)] + (1 − λ)[g0 (x̄, ξ) + Q(x̄, ξ)] ∀ξ ∈ Ξ
implying
˜ ≤ λE {g0 (x̂, ξ̃)+Q(x̂, ξ)}+(1−λ)E
Eξ̃ {g0 (x̌, ξ̃)+Q(x̌, ξ)} ˜ ˜
ξ̃ ξ̃ {g0 (x̄, ξ)+Q(x̄, ξ̃)}.
2
n
Remark 1.1 Observe that for Y = IR+ the convexity of Q(·, ξ) can
immediately be asserted for the linear case (4.16) and that it also holds for the
nonlinear case (4.10) if the functions q(·) and gi (·, ξ) are convex and the Hi (·)
are concave. Just to sketch the argument, assume that ȳ and y̌ solve (4.10)
for x̄ and x̌ respectively, at some realization ξ ∈ Ξ. Then, by the convexity of
gi and the concavity of Hi , i = 1, · · · , m, we have, for any λ ∈ (0, 1),
gi (λx̄ + (1 − λ)x̌, ξ) ≤ λgi (x̄, ξ) + (1 − λ)gi (x̌, ξ)
≤ λHi (ȳ) + (1 − λ)Hi (y̌)
≤ Hi (λȳ + (1 − λ)y̌).
Hence ŷ = λȳ + (1 − λ)y̌ is feasible in (4.10) for x̂ = λx̄ + (1 − λ)x̌, and
therefore, by the convexity of q,
Q(x̂, ξ) ≤ q(ŷ)
≤ λq(ȳ) + (1 − λ)q(y̌)
= λQ(x̄, ξ) + (1 − λ)Q(x̌, ξ).
2

Smoothness (i.e. partial differentiability of Q(x) = Ξ Q(x, ξ) dP ) of
recourse problems may also be asserted under fairly general conditions. For
example, suppose that ϕ : IR2 −→ IR, so that ϕ(x, y) ∈ IR. Recalling that ϕ is
partially differentiable at some point (x̂, ŷ) with respect to x, this means that
∂ϕ(x, y)
there exists a function, called the partial derivative and denoted by ,
∂x
such that
ϕ(x̂ + h, ŷ) − ϕ(x̂, ŷ) ∂ϕ(x̂, ŷ) r(x̂, ŷ; h)
= + ,
h ∂x h
where the “residuum” r satisfies
r(x̂, ŷ; h) h→0
−→ 0.
h
The recourse function is partially differentiable with respect to xj in (x̂, ξ̂) if
there is a function ∂Q(x,ξ)
∂xj such that
Q(x̂ + hej , ξ̂) − Q(x̂, ξ̂) ˆ h)

∂Q(x̂, ξ̂) ρj (x̂, ξ;
= +
h ∂xj h
with
ˆ h) h→0
ρj (x̂, ξ;
−→ 0,
h
∂Q(x,ξ) T
where ej is the jth unit vector. The vector ( ∂Q(x,ξ)
∂x1 , · · · , ∂xn ) is called
the gradient of Q(x, ξ) with respect to x and is denoted by ∇x Q(x, ξ). Now we
are not only interested in the partial differentiability of the recourse function
Q(x, ξ) but also in that of the expected recourse function Q(x). Provided that
˜ is partially differentiable at x̂ a.s., we get
Q(x, ξ)

Q(x̂ + hej ) − Q(x̂) Q(x̂ + hej , ξ) − Q(x̂, ξ)
= dP
h h
Ξ
∂Q(x̂, ξ) ρj (x̂, ξ; h)
= + dP
Ξ−Nδ ∂xj h

∂Q(x̂, ξ) ρj (x̂, ξ; h)
= dP + dP,
Ξ−Nδ ∂x j Ξ−Nδ h
where Nδ ∈ F and P (Nδ ) = 0. Hence, under these assumptions, Q is partially

differentiable if

∂Q(x̂, ξ) 1 h→0
dP exists and ρj (x̂, ξ; h)dP −→ 0.
Ξ−Nδ ∂xj h Ξ−Nδ
This yields the following.
Proposition 1.2 If Q(x, ξ) ˜ is partially differentiable with respect to xj at
some x̂ a.s. (i.e. for all ξ except maybe those belonging to an event with
∂Q(x̂, ξ)
probability zero), if its partial derivative is integrable and if the
∂xj
h→0 ∂Q(x̂)
residuum satisfies (1/h) Ξ ρj (x̂, ξ; h)dP −→ 0 then exists as well
∂xj
and
∂Q(x̂) ∂Q(x̂, ξ)
= dP.
∂xj Ξ ∂xj
Questions arise as a result of the general formulation of the assumptions
of this proposition. It is often possible to decide that the recourse function is
partially differentiable a.s. and the partial derivative is integrable. However,
the requirement that the residuum is integrable and—roughly speaking—its
integral converges to zero faster than h can be difficult to check. Hence we
leave the general case and focus on stochastic linear programs with complete
fixed recourse (4.16) in the following remark.
Remark 1.2 In the linear case (4.16) with complete fixed recourse it is known
from linear programming (see Section 1.7) that the optimal value function
BASIC CONCEPTS 39
Figure 13 Linear affine mapping of a polyhedron.
Q(x, ξ) is continuous and piecewise linear in h(ξ) − T (ξ)x. In other words,

there exist finitely many convex polyhedral cones Bl ⊂ IRm1 with nonempty
interiors such that any two of them have at most boundary points in common
and ∪l Bl = IRm1 , and Q(x, ξ) is given as Q(x, ξ) = dlT (h(ξ) − T (ξ)x) + δl
for h(ξ) − T (ξ)x ∈ Bl . Then, for h(ξ) − T (ξ)x ∈ intBl (i.e. for h(ξ) − T (ξ)x
an interior point of Bl ), the function Q(x, ξ) is partially differentiable with
respect to any component of x. Hence for the gradient with respect to x we
get from the chain rule that ∇x Q(x, ξ) = −T T (ξ)dl for h(ξ) − T (ξ)x ∈ intBl .
Assume for simplicity that Ξ is a bounded interval in IRk and keep x fixed.
Then, by (4.15), we have a linear affine mapping ψ(·) := h(·) − T (·)x : Ξ −→
IRm1 . Therefore the sets
D̂l (x) = ψ −1 [Bl ] := {ξ ∈ Ξ | ψ(ξ) ∈ Bl }

are convex polyhedra (see Figure 13) satisfying l D̂l (x) = Ξ.
Define Dl (x) := intD̂l (x). To get the intended differentiability result, the
following assumption is crucial:
ξ ∈ Dl (x) =⇒ ψ(ξ) = h(ξ) + T (ξ)x ∈ intBl ∀l. (5.1)
By this assumption, we enforce the event {ξ ∈ Ξ | ψ(ξ) ∈ Bl − intBl } to have

the natural measure µ({ξ ∈ Ξ | ψ(ξ) ∈ Bl − intBl }) = 0, which need not be
true in general, as illustrated in Figure 14.
Figure 14 Linear mapping violating assumption (5.1).
Since the Bl are convex polyhedral cones in IRm1 (see Section 1.7) with
nonempty interiors, they may be represented by inequality systems
C l z ≤ 0,
where C l = 0 is an appropriate matrix with no row equal to zero. Fix l and

let ξ ∈ Dl (x) such that, by (5.1), h(ξ) − T (ξ)x ∈ intBl . Then
C l [h(ξ) − T (ξ)x] < 0,
i.e. for any fixed j there exists a τ̂lj > 0 such that
C l [h(ξ) − T (ξ)(x ± τlj ej )] ≤ 0
or, equivalently,
C l [h(ξ) − T (ξ)x] ≤ ∓τlj C l T (ξ)ej ∀τlj ∈ [0, τ̂lj ].

Hence for γ(ξ) = maxi (C l T (ξ)ej )i there is a
tl > 0 : C l [h(ξ) − T (ξ)x] ≤ −|t|γ(ξ)e ∀|t| < tl ,
e = (1, · · · , 1)T . This implies that for γ := max γ(ξ) there exists a t0 > 0 such
ξ∈Ξ
that
C l [h(ξ) − T (ξ)x] ≤ −|t|γe ∀|t| < t0
(choose, for example, t0 = tl /γ). In other words, there exists a t0 > 0 such
that
Dl (x; t) := {ξ | C l [h(ξ) − T (ξ)x] ≤ −|t|γe} = ∅ ∀|t| < t0 ,
and obviously Dl (x; t) ⊂ Dl (x). Furthermore, by elementary geometry, the
natural measure µ satisfies
µ(Dl (x) − Dl (x; t)) ≤ |t|v

BASIC CONCEPTS 41
Figure 15 Difference set Dl (x) − Dl (x; t).
with some constant v (see Figure 15).

For ξ ∈ Dl (x; t) it follows that
C l [h(ξ) − T (ξ)(x + tej )] = C l [h(ξ) − T (ξ)x] − tC l T (ξ)ej

≤ −|t|γe − tC l T (ξ)ej
≤ 0,
owing to the fact that each component of C l T (ξ)ej is absolutely bounded by

γ. Hence in this case we have h(ξ) − T (ξ)(x + tej ) ∈ Bl , and so
Q(x + tej , ξ) − Q(x, ξ) ∂Q(x, ξ)

= −dlT T (ξ)ej = ∀|t| < t0 ,
t ∂xj
i.e. in this case we have the residuum ρj (x, ξ; t) ≡ 0.

For ξ ∈ Dl (x)− Dl (x; t) we have, considering that h(ξ)− T (ξ)(x+ tej ) could
possibly belong to some other Bl̄ , at least the estimate
|Q(x + tej , ξ) − Q(x, ξ)|

≤ max{|dl̄T T (ξ)ej | | ξ ∈ Ξ, ∀l̄} =: β.
|t|
Assuming now that we have a continuous density ϕ(ξ) for P , we know already
from (5.1) that µ({ξ ∈ Ξ | ψ(ξ) ∈ Bl − intBl }) = 0. Hence it follows that

˜
Eξ̃ Q(x, ξ) = Q(x, ξ)ϕ(ξ)dξ
l Dl (x)

= {dlT [h(ξ) − T (ξ)x] − δl }ϕ(ξ)dξ,
l Dl (x)
and, since

Q(x + tej , ξ) − Q(x, ξ)
t→0
ϕ(ξ)dµ ≤ β max ϕ(ξ)|t|v −→ 0,
Dl (x)−Dl (x;t) t ξ∈Ξ

˜ =
∇Eξ̃ Q(x, ξ) ∇x Q(x, ξ)ϕ(ξ)dξ
Dl (x)

l
= T T (ξ)dl ϕ(ξ)dξ.
l Dl (x)
Hence for the linear case—observing (4.15)—we get the differentiability state-
ment of Proposition 1.2 provided that (5.1) is satisfied and P has a continuous
density on Ξ. 2
Summarizing the statements given so far, we see that stochastic programs

with recourse are likely to have such properties as convexity (Proposition 1.1)
and, given continuous-type distributions, differentiability (Proposition 1.2),
which—from the viewpoint of mathematical programming—are appreciated.
On the other hand, if we have a joint finite discrete probability distribution
{(ξ k , pk ), k = 1, · · · , r} of the random data then, for example, problem (4.16)
becomes—similarly to the special example (3.11)—a linear program
⎫
r
⎪
⎪
minx∈X cT x + pk q T y k ⎪
⎪
⎬
k=1 (5.2)
⎪
s.t. T (ξ k )x + W y k = h(ξ k ), k = 1, · · · , r, ⎪
⎪
⎪
⎭
yk ≥ 0
having the so-called dual decomposition structure, as mentioned already for our
special example (3.11) and demonstrated in Figure 16 (see also Section 1.7.4).
cT qT qT ··· qT
T (ξ 1 ) W h(ξ 1 )
T (ξ 2 ) W h(ξ 2 )
.. .. ..
. . .
T (ξ r ) W h(ξ r )
Figure 16 Dual decomposition data structure.

BASIC CONCEPTS 43
However—for finite discrete as well as for continuous distributions—we are

faced with a further problem, which we might discuss for the linear case (i.e.
for stochastic linear programs with fixed recourse (4.16)). By suppP we denote
the support of the probability measure P , i.e. the smallest closed set Ξ ⊂ IRk
such that Pξ̃ (Ξ) = 1. With the practical interpretation of the second-stage
problem as given, for example, in Section 1.3, and assuming that Ξ = suppPξ̃ ,
we should expect that for any first-stage decision x ∈ X the compensation
of deficiencies in the stochastic constraints is possible whatever ξ ∈ Ξ will be
realized for ξ̃. In other words, we expect the program
⎫
Q(x, ξ) = min q T y ⎬
s.t. W y = h(ξ) − T (ξ)x, (5.3)
⎭
y≥0
to be feasible ∀ξ ∈ Ξ. Depending on the defined recourse matrix W and the
given support Ξ, this need not be true for all first-stage decisions x ∈ X.
Hence it may become necessary to impose—in addition to x ∈ X—further
restrictions on our first-stage decisions called induced constraints. To be more
specific, let us assume that Ξ is a (bounded) convex polyhedron, i.e. the convex
hull of finitely many points ξ j ∈ Ξ ⊂ IRk :
conv {ξ 1 , · · · , ξ r }
Ξ=⎧ ⎫
⎨ r
r ⎬

= ξξ = λj ξ j , λj = 1, λj ≥ 0 ∀j .
⎩ ⎭
j=1 j=1
From the definition of a support, it follows that x ∈ IRn allows for a feasible
solution of the second-stage program for all ξ ∈ Ξ if and only if this is true
for all ξ j , j = 1, · · · , r. In other words, the induced first-stage feasibility set
K is given as
K = {x | T (ξ j )x + W y j = h(ξ j ), y j ≥ 0, j = 1, · · · , r}.
From this formulation of K (which obviously also holds if ξ̃ has a finite discrete
distribution, i.e. Ξ = {ξ 1 , · · · , ξ r }), we evidently get the following.
Proposition 1.3 If the support Ξ of the distribution of ξ̃ is either a finite
set or a (bounded) convex polyhedron then the induced first-stage feasibility
set K is a convex polyhedral set. The first-stage decisions are restricted to
x ∈ X K.
Example 1.3 Consider the following first-stage feasible set:
X = {x ∈ IR2+ | x1 − 2x2 ≥ −4, x1 + 2x2 ≤ 8, 2x1 − x2 ≤ 6}.
For the second-stage constraints choose

−1 3 5 2 3
W = , T (ξ) ≡ T =
2 2 2 3 1
and a random vector ξ̃ with the support Ξ = [4, 19] × [13, 21]. Then the
constraints to be satisfied for all ξ ∈ Ξ are
W y = ξ − T x, y ≥ 0.
Observing that the second column W2 of W is a positive linear combination

of W1 and W3 , namely W2 = 13 W1 + 23 W3 , the above second-stage constraints
reduce to the requirement that for all ξ ∈ Ξ the right-hand side ξ − T x can
be written as
ξ − T x = λW1 + µW3 , λ, µ ≥ 0,
or in detail as
ξ1 − 2x1 − 3x2 = −λ + 5µ,
ξ2 − 3x1 − x2 = 2λ + 2µ,
λ, µ ≥ 0.

2 1
Multiplying this system of equations with the regular matrix S = ,
−2 5
which corresponds to adding 2 times the first equation to the second and
adding −2 times the first to 5 times the second, respectively, we get the
equivalent system
2ξ1 + ξ2 − 7x1 − 7x2 = 12µ ≥ 0,

−2ξ1 + 5ξ2 − 11x1 + x2 = 12λ ≥ 0.
Because of the required nonnegativity of λ and µ, this is equivalent to the

system of inequalities
7x1 + 7x2 ≤ 2ξ1 + ξ2 (≥ 21 ∀ξ ∈ Ξ),

11x1 − x2 ≤ −2ξ1 + 5ξ2 (≥ 27 ∀ξ ∈ Ξ).
Since these inequalities have to be satisfied for all ξ ∈ Ξ, choosing the minimal
right-hand sides (for ξ ∈ Ξ) yields the induced constraints as
K = {x | 7x1 + 7x2 ≤ 21, 11x1 − x2 ≤ 27}.
The first-stage feasible set X together with the induced feasible set are illus-
trated in Figure 17. 2

It might happen that X K = ∅; then we should check our model very
carefully to figure out whether we really modelled what we had in mind or
whether we can find further possibilities for compensation that are not yet
contained in our model. On the other hand, we have already mentioned the
case of a complete fixed recourse matrix (see (4.17) on page 34), for which
K = IRn and therefore the problem of induced constraints does not exist.
Hence it seems interesting to recognize complete recourse matrices.
BASIC CONCEPTS 45
Figure 17 Induced constraints K.
Proposition 1.4 An m1 × n matrix W is a complete recourse matrix iff 4

• it has rank rk(W ) = m1 and,
• assuming without loss of generality that its first m1 columns W1 , W2 , · · · , Wm1
are linearly independent, the linear constraints
⎫
Wy = 0 ⎬
yi ≥ 1, i = 1, · · · , m1 , (5.4)
⎭
y≥0
have a feasible solution.
Proof W is a complete recourse matrix iff
{z | z = W y, y ≥ 0} = IRm1 .
From this condition, it follows immediately
m1 that rk(W ) = m1 necessarily has
to hold. In addition, for ẑ = − i=1 Wi ∈ IRm1 the second-stage constraints
W y = ẑ, y ≥ 0 have a feasible solution y̌ such that

m1
n
Wi y̌i + Wi y̌i = ẑ
i=1 i=m1 +1

m1
=− Wi ,
i=1
y̌i ≥ 0, i = 1, · · · , n.
4 We use “iff ” as short-hand for “if and only if ”.
With
y̌i + 1, i = 1, · · · , m1 ,
yi =
y̌i , i > m1 ,
this implies that the constraints (5.4) are necessarily feasible.
To show that the above conditions are also sufficient for complete recourse
let us choose an arbitrary z̄ ∈ IRm1 . Since the columns W1 , · · · , Wm1 are
linearly independent, the system of linear equations

m1
Wi yi = z̄
i=1
has a unique solution ȳ1 , · · · , ȳm1 . If ȳi ≥ 0, i = 1, · · · , m1 , we are finished;

otherwise, we define γ := min{ȳ1 , · · · , ȳm1 }. By assumption, the constraints
(5.4) have a feasible solution y̌. Now it is immediate that ŷ defined by

ȳi − γ y̌i , i = 1, · · · , m1 ,
ŷi =
−γ y̌i , i = m1 + 1, · · · , n,
solves W y = z̄, y ≥ 0. 2
Finally, if (5.3) is feasible for all ξ ∈ Ξ and at least for all x ∈ X = {x |

Ax = b, x ≥ 0} then (4.16) is said to be of relatively complete recourse.
1.6 Properties of Probabilistic Constraints
For chance constrained problems, the situation becomes more difficult, in

general. Consider the constraint of (4.21),
P ({ξ | g(x, ξ) ≤ 0}) ≥ α,
where the gi were replaced by the vector-valued function g defined by

T
g(x, ξ) := g1 (x, ξ), · · · , gm (x, ξ) : a point x̂ is feasible iff the set
S(x̂) = {ξ | g(x̂, ξ) ≤ 0} (6.1)
has a probability measure P (S(x̂)) of at least α. In other words, if G ⊂ F is

the collection of all events of F such that P (G) ≥ α ∀G ∈ G then x̂ is feasible
iff we find at least one event G̃ ∈ G such that for all ξ ∈ G̃, g(x̂, ξ) ≤ 0.
Formally, x̂ is feasible iff ∃G ∈ G:

x̂ ∈ {x | g(x, ξ) ≤ 0}. (6.2)
ξ∈G
BASIC CONCEPTS 47
Hence the feasible set
B(α) = {x | P ({ξ | g(x, ξ) ≤ 0}) ≥ α}
is the union of all those vectors x feasible according to (6.2), and consequently
may be rewritten as

B(α) = {x | g(x, ξ) ≤ 0}. (6.3)
G∈G ξ∈G
Since a union of convex sets need not be convex, this presentation

demonstrates that in general we may not expect B(α) to be convex, even
if {x | g(x, ξ) ≤ 0} are convex ∀ξ ∈ Ξ. Indeed, there are simple examples for
nonconvex feasible sets.
Example 1.4 Assume that in our refinery problem (3.1) the demands are
random with the following discrete joint distribution:

h1 (ξ 1 ) = 160
P = 0.85,
h2 (ξ 1 ) = 135

h1 (ξ 2 ) = 150
P = 0.08,
h2 (ξ 2 ) = 195

h1 (ξ 3 ) = 200
P = 0.07.
h2 (ξ 3 ) = 120
Then the constraints
xraw1 + xraw2 ≤ 100

xraw1 ≥0
xraw2 ≥ 0

˜
P ˜ ≥α
for any α ∈ (0.85, 0.92] require that we
• either satisfy the demands hi (ξ 1 ) and hi (ξ 2 ), i = 1, 2 (enforcing a reliability

of 93%)
and hence choose a production program to cover a demand
160
hA =
195
• or satisfy the demands hi (ξ 1 ) and hi (ξ 3 ), i = 1, 2 (enforcing a reliability of
92%) such
that
our production plan is designed to cope with the demand
200
hB = .
135
Figure 18 Chance constraints: nonconvex feasible set.
It follows that the feasible set for the above constraints is nonconvex, as shown
in Figure 18. 2
As above, define S(x) := {ξ | g(x, ξ) ≤ 0}. If g(·, ·) is jointly convex in (x, ξ)

then, with xi ∈ B(α), i = 1, 2, ξ i ∈ S(xi ) and λ ∈ [0, 1], for (x̄, ξ̄) =
λ(x1 , ξ 1 ) + (1 − λ)(x2 , ξ 2 ) it follows that
¯ ≤ λg(x1 , ξ 1 ) + (1 − λ)g(x2 , ξ 2 ) ≤ 0,
g(x̄, ξ)
i.e. ξ̄ = λξ 1 + (1 − λ)ξ 2 ∈ S(x̄), and hence5
S(x̄) ⊃ [λS(x1 ) + (1 − λ)S(x2 )]
implying
P (S(x̄)) ≥ P (λS(x1 ) + (1 − λ)S(x2 )).
By our assumption on g (joint convexity), any set S(x) is convex. Now we
conclude immediately that B(α) is convex ∀α ∈ [0, 1], if
P (λS1 + (1 − λ)S2 ) ≥ min[P (S1 ), P (S2 )] ∀λ ∈ [0, 1]
for all convex sets Si ∈ F, i = 1, 2, i.e. if P is quasi-concave. Hence we have

proved the following
5 The algebraic sum of sets ρS1 + σS2 := {ξ := ρξ 1 + σξ 2 | ξ 1 ∈ S1 , ξ 2 ∈ S2 }.
BASIC CONCEPTS 49
Figure 19 Convex combination of events involved by distribution functions,

λ = 12 .
Proposition 1.5 If g(·, ·) is jointly convex in (x, ξ) and P is quasi-concave,

then the feasible set B(α) = {x|P ({ξ|g(x, ξ) ≤ 0}) ≥ α} is convex ∀α ∈ [0, 1].
Remark 1.3 The assumption of joint convexity of g(·, ·) is so strong that it
is even not satisfied in the linear case (4.23), in general. However, if in (4.23)
T (ξ) ≡ T (constant) and h(ξ) ≡ ξ then it is satisfied and the constraints
˜ read as
of (4.23), Fξ̃ being the distribution function of ξ,
P ({ξ | T x ≥ ξ}) = Fξ̃ (T x) ≥ α.
Therefore B(α) is convex ∀α ∈ [0, 1] in this particular case if Fξ̃ is a quasi-

concave function, i.e. if Fξ̃ (λξ 1 + (1 − λ)ξ 2 ) ≥ min[Fξ̃ (ξ 1 ), Fξ̃ (ξ 2 )] for any two
ξ 1 , ξ 2 ∈ Ξ and ∀λ ∈ [0, 1]. 2
It seems worthwile to mention the following facts. If the probability measure

P is quasi-concave then the corresponding distribution function Fξ̃ is quasi-
concave. This follows from observing that by the definition of distribution
functions Fξ̃ (ξ i ) = P (Si ) with Si = {ξ | ξ ≤ ξ i }, i = 1, 2, and that for ξˆ =
λξ 1 + (1 − λ)ξ 2 , λ ∈ [0, 1], we have Ŝ = {ξ | ξ ≤ ξ̂} = λS1 + (1 − λ)S2 (see
Figure 19). With P being quasi-concave, this yields
ˆ = P (Ŝ) ≥ min[P (S1 ), P (S2 )] = min[F (ξ 1 ), F (ξ 2 )].
Fξ̃ (ξ) ξ̃ ξ̃
On the other hand, Fξ̃ being a quasi-concave function does not imply in general
that the corresponding probability measure P is quasi-concave. For instance,
Figure 20 P here is not quasi-concave: P (C) = P ( 13 A + 2

3
B) = 0, but
P (A) = P (B) = 12 .
in IR1 every monotone function is easily seen to be quasi-concave, such that

every distribution function of a random variable (always being monotonically
increasing) is quasi-concave.But not every probability measure P on IR is
quasi-concave (see Figure 20 for a counterexample).
Hence we stay with the question of when a probability measure—or its
distribution function—is quasi-concave. This question was answered first for
the subclass of log-concave probability measures, i.e. measures satisfying
P (λS1 + (1 − λ)S2 ) ≥ P λ (S1 ) P 1−λ (S2 )
for all convex Si ∈ F and λ ∈ [0, 1]. That the class of log-concave measures is
really a subclass of the class of quasi-concave measures is easily seen.
Lemma 1.2 If P is a log-concave measure on F then P is quasi-concave.
Proof Let Si ∈ F, i = 1, 2, be convex sets such that P (Si ) > 0, i = 1, 2
(otherwise there is nothing to prove, since P (S) ≥ 0 ∀S ∈ F). By assumption,
for any λ ∈ (0, 1) we have
P (λS1 + (1 − λ)S2 ) ≥ P λ (S1 ) P 1−λ (S2 ).
By the monotonicity of the logarithm, it follows that
ln[P (λS1 + (1 − λ)S2 )] ≥ λ ln[P (S1 )] + (1 − λ) ln[P (S2 )]

≥ min{ln[P (S1 )], ln[P (S2 )]},
BASIC CONCEPTS 51
and hence
P (λS1 + (1 − λ)S2 ) ≥ min[P (S1 ), P (S2 )].
2
As mentioned above, for the log-concave case necessary and sufficient

conditions were derived first, and later corresponding conditions for quasi-
concave measures were found.
Proposition 1.6 Let P on Ξ = IRk be of the continuous type, i.e. have a

density f . Then the following statements hold:
• P is log-concave iff f is log-concave (i.e. if the logarithm of f is a concave

function);
• P is quasi-concave iff f −1/k is convex.
The proof has to be omitted here, since it would require a rather advanced
knowledge of measure theory.
Remark 1.4 Consider
(a) the k-dimensional uniform distribution on a convex body S ⊂ IRk (with

positive natural
measure µ) given by the density
1/µ(S) if x ∈ S,
ϕU (x) :=
0 otherwise
(µ is the natural measure in IRk , see Section 1.4.1);
(b) the exponential
distribution with density
0 if x < 0,
ϕEX P (x) :=
−λx
λe if x ≥ 0
(λ > 0 is constant);
(c) the multivariate normal distribution in IRk described by the density
T
1
Σ−1 (x−m)
ϕN (x) := γe− 2 (x−m)
(γ > 0 is constant, m is the vector of expected values and Σ is the
covariance matrix).
Then we get immediately

k
−k1 µ(S) if x ∈ S,
(a) ϕU (x) =
∞ otherwise,
implying by Proposition 1.6 that the corresponding propability measure
PU is quasi-concave.
(b) Since
−∞ if x < 0,
ln[ϕEX P (x)] =
ln λ − λx if x ≥ 0,
the density of the exponential distribution is obviously log-concave,
implying by Proposition 1.6 that the corresponding measure PEX P is log-
concave and hence, by Lemma 1.2, also quasi-concave.
(c) Taking the logarithm
ln[ϕN (x)] = ln γ − 12 (x − m)T Σ−1 (x − m)
and observing that the covariance matrix Σ and hence its inverse Σ−1 are
positive definite, we see that this density is log-concave, and therefore the
corresponding measure PN is log-concave (by Proposition 1.6) as well as
quasiconcave (by Lemma 1.2).
There are many other classes of widely used continuous type probability
measures, which—according to Proposition 1.6—are either log-concave or at
least quasi-concave. 2
In addition to Proposition 1.5, we have the following statement, which is of

interest because, for mathematical programs in general, we cannot assert the
existence of solutions if the feasible sets are not known to be closed.
Proposition 1.7 If g : IRn × Ξ → IRm is continuous then the feasible set
B(α) is closed.
Proof Consider any sequence {xν } such that xν −→ x̂ and xν ∈ B(α) ∀ν.
To prove the assertion, we have to show that x̂ ∈ B(α). Define A(x) := {ξ |
g(x, ξ) ≤ 0}. Let Vk be the open ball with center x̂ and radius 1/k. Then we
show first that ∞

A(x̂) = cl A(x). (6.4)
k=1 x∈Vk
Here the inclusion “⊂” is obvious since x̂ ∈ Vk ∀k, so we have only to show
that
∞

A(x̂) ⊃ cl A(x).
k=1 x∈Vk
∞
Assume that ξ̂ ∈ k=1 cl x∈Vk A(x). This means that for every k we have

ξˆ ∈ cl x∈Vk A(x); in other words, for every k there exists a ξ k ∈ x∈Vk A(x)
and hence some xk ∈ Vk with ξ k ∈ A(xk ) such that ξ k − ξ ˆ ≤ 1/k (and
obviously x − x̂ ≤ 1/k since x ∈ Vk ). Hence (x , ξ ) −→ (x̂, ξ).
k k k k ˆ Since
ξ k ∈ A(xk ), g(xk , ξ k ) ≤ 0 ∀k and therefore, by the continuity of g(·, ·),
ξˆ ∈ A(x̂), which proves (6.4) to be true.
BASIC CONCEPTS 53
The sequence of sets

K
BK := cl A(x)
k=1 x∈Vk
is monotonically decreasing to the set A(x̂). Since xν −→ x̂, for every K

there exists a νK such that xνK ∈ VK ⊂ VK−1 ⊂ · · · ⊂ V1 , implying that
A(xνK ) ⊂ BK and hence P (BK ) ≥ P (A(xνK )) ≥ α ∀K. Hence, by the well-
known continuity of probability measures on monotonic sequences, we have
P (A(x̂)) ≥ α, i.e. x̂ ∈ B(α). 2
For stochastic programs with joint chance constraints the situation appears
to be more difficult than for stochastic programs with recourse. But, at least
under certain additional assumptions, we may assert convexity and closedness
of the feasible sets as well (Proposition 1.5, Remark 1.3 and Proposition 1.7).
For stochastic linear programs with single chance constraints, convexity
statements have been derived without the joint convexity assumption on
gi (x, ξ) := hi (ξ) − Ti (ξ)x, for special distributions and special intervals for
the values of αi . In particular, if Ti (ξ) ≡ Ti (constant), the situation becomes
rather convenient: with Fi the distribution function of hi (ξ), ˜ we have
P ({ξ | Ti x ≥ hi (ξ)}) = Fi (Ti x) ≥ αi ,
or equivalently
Ti x ≥ Fi−1 (αi ),
where Fi−1 (αi ) is assumed to be the smallest real value η such that Fi (η) ≥ αi .
Hence in this special case any single chance constraint turns out to be just a
linear constraint, and the only additional work to do is to compute Fi−1 (αi ).
1.7 Linear Programming
Throughout this section we shall discuss linear programs in the following

standard form ⎫
min cT x ⎬
s.t. Ax = b, (7.1)
⎭
x ≥ 0,
where the vectors c ∈ IRn , b ∈ IRm and the m × n matrix A are given
and x ∈ IRn is to be determined. Any other LP6 formulation can easily be
6 We use occasionally “LP” as abbreviation for “linear program(ming)”.
transformed to assume the form (7.1). If, for instance, we have the problem
min cT x
s.t. Ax ≥ b
x ≥ 0,
then, by introducing a vector y ∈ IRm
+ of slack variables, we get the problem
min cT x
s.t. Ax − y = b
x≥0
y ≥ 0,
which is of the form (7.1). This LP is equivalent to (7.1) in the sense that
the x part of its solution set and the solution set of (7.1) as well as the two
optimal values obviously coincide. Instead, we may have the problem
min cT x
s.t. Ax ≥ b,
where the decision variables are not required to be nonnegative—so-called free
variables. In this case we may introduce a vector y ∈ IRm + of slack variables
and—observing that any real number may be presented as the difference of two
nonnegative numbers—replace the original decision vector x by the difference
z + − z − of the new decision vectors z + , z − ∈ IRn+ yielding the problem
min{cT z + − cT z − }
s.t. Az + − Az − − y = b,
z+ ≥ 0,
z− ≥ 0,
y ≥ 0,
which is again of the form (7.1). Furthermore, it is easily seen that this
transformed LP and its original formulation are equivalent in the sense that
• given any solution (ẑ + , ẑ − , ŷ) of the transformed LP, x̂ := ẑ + − ẑ − is a
solution of the original version,
• given any solution x̌ of the original LP, the vectors y̌ := Ax̌ − b and
ž + , ž − ∈ IRn+ , chosen such that ž + − ž − = x̌, solve the transformed version,
and the optimal values of both versions of the LP coincide.
1.7.1 The Feasible Set and Solvability

From linear algebra, we know that the system Ax = b of linear equations
in (7.1) is solvable if and only if the rank condition
rk(A, b) = rk(A) (7.2)
BASIC CONCEPTS 55
is satisfied. Given this condition, it may happen that rk(A) < m, but then
we may drop one or more equations from the system without changing its
solution set. Therefore we assume throughout this section that
rk(A) = m, (7.3)
which obviously implies that m ≤ n.
Let us now investigate the feasible set
B := {x | Ax = b, x ≥ 0}
of (7.1). A central concept in linear programming is that of a feasible
basic solution defined as follows: x̂ ∈ B is a feasible basic solution if with
I(x̂) := {i | x̂i > 0} the set {Ai | i ∈ I(x̂)} of columns of A is linearly
independent.7 Hence the components x̂i , i ∈ I(x̂), are the unique solution of
the system of linear equations

Ai xi = b.
i∈I(x̂)
In general, the set I(x̂) and hence also the column set {Ai | i ∈ I(x̂)} may
have less than m elements, which can cause some inconvenience—at least in
formulating the statements we want to present.
Proposition 1.8 Given assumption (7.3), for any basic solution x̂ of B there
exists at least one index set IB (x̂) ⊃ I(x̂) such that the corresponding column
set {Ai | i ∈ IB (x̂)} is a basis of IR . The components x̂i , i ∈ IB (x̂), of
m
x̂ uniquely solve the linear system i∈IB (x̂) Ai xi = b with the nonsingular
matrix (Ai | i ∈ IB (x̂)).
Proof Assume that x̌ ∈ B is a basic solution and that {Ai | i ∈ I(x̌)}
contains k columns, k < m, of A. By (7.3), there exists at least one index
set Jm ⊂ {1, · · · , n} with m elements such that the columns {Ai | i ∈ Jm }
are linearly independent and hence form a basis of IRm . A standard result in
linear algebra asserts that, given a basis of an m-dimensional vector space and
a linear independent subset of k < m vectors, it is possible, by adding m − k
properly chosen vectors from the basis, to complement the subset to become
a basis itself. Hence in our case it is possible to choose m − k indices from Jm
and to add them to I(x̌), yielding IB (x̌) such that {Ai | i ∈ IB (x̌)} is a basis
of IRm . 2
Given a basic solution x̂ ∈ B, by this proposition the matrix A can be

partitioned into two parts (corresponding to x̂): a basic part
B = (Ai | i ∈ IB (x̂))
7 According to this definition, for I(x̂) = ∅, i.e. x̂ = 0 and hence b = 0, it follows that x̂ is
a feasible basic solution as well.
and a nonbasic part
N = (Ai | i ∈ {1, · · · , n} − IB (x̂)).
Introducing the vectors x{B} ∈ IRm —the vector of basic variables—and

{N B}
x ∈ IRn−m —the vector of nonbasic variables—and assigning
{B}
xk = xi , i the kth element of IB (x̂), k = 1, · · · , m,
(7.4)
{N B}
xl = xi , i the lth element of {1, · · · , n} − IB (x̂), l = 1, · · · , n − m,
the linear system Ax = b of (7.1) may be rewritten as
Bx{B} + N x{N B} = b
or equivalently as
x{B} = B −1 b − B −1 N x{N B} , (7.5)
which—using the assignment (7.4) – yields for any choice of the nonbasic
variables x{N B} a solution of our system Ax = b, and in particular for
x{N B} = 0 reproduces our feasible basic solution x̂.
Proposition 1.9 If B = ∅ then there exists at least one feasible basic solution.
Proof Assume that for x̂
Ax̂ = b, x̂ ≥ 0.
If for I(x̂) = {i | x̂i > 0} the column set {Ai | i ∈ I(x̂)} is linearly dependent,
then the linear homogeneous system of equations

i∈I(x̂) Ai yi = 0,
yi = 0, i ∈ I(x̂),
has a solution y̌ = 0 with y̌i < 0 for at least one i ∈ I(x̂)—if this does not
hold for y̌, we could take −y̌, which solves the above homogeneous system as
well. Hence for
λ̄ := max{λ | x̂ + λy̌ ≥ 0}
we have 0 < λ̄ < ∞. Since Ay̌ = 0 obviously holds for y̌, it follows—observing
the definition of λ̄—that for z := x̂ + λ̄y̌
Az = Ax̂ + λ̄Ay̌
= b,
z ≥ 0,
i.e. z ∈ B, and I(z) ⊂ I(x̂), I(z) = I(x̂), such that we have “reduced” our
original feasible solution x̂ to another one with fewer positive components.
BASIC CONCEPTS 57
Now either z is a basic solution or we repeat the above “reduction” with

x̂ := z. Obviously there are only finitely many reductions of the number of
positive components in feasible solutions possible. Hence we have to end up—
after finitely many of these steps—with a feasible basic solution. 2
With an elementary exercise, we see that the feasible set B = {x | Ax =

b, x ≥ 0} of our linear program (7.1) is convex. We want now to point out
that feasible basic solutions play a dominant role in describing feasible sets of
linear programs.
Proposition 1.10 If B is a bounded set and B = ∅ then B is the convex hull
(i.e. the set of all convex linear combinations) of the set of its feasible basic
solutions.
Proof To avoid trivialities or statements on empty sets, we assume that
the right-hand side b = 0. For any feasible solution x ∈ B we again have
the index set I(x) := {i | xi > 0}, and we denote by |I(x)| the number of
elements of I(x). Obviously we have—recalling our assumption that b = 0—
that for any feasible solution 1 ≤ |I(x)| ≤ n. We may prove the proposition
by induction on |I(x)|, the number of positive components of any feasible
solution x. To begin with, we define k0 := minx∈B |I(x)| ≥ 1. For a feasible x
with |I(x)| = k0 it follows that x is a basic solution—otherwise, by the proof
of Proposition 1.9, there would exist a feasible basic solution with less than k0
positive components—and we have x = 1 · x, i.e. a convex linear combination
of itself and hence of the set of feasible basic solutions. Let us now assume
that for some k ≥ k0 and for all feasible solutions x such that |I(x)| ≤ k the
hypothesis is true. Then, given a feasible solution x̂ with |I(x̂)| = k + 1, for
x̂ a basic solution we again have x̂ = 1 · x̂ and thus the hypothesis holds.
Otherwise, i.e. if x̂ is not a basic solution, the homogeneous system

i∈I(x̂) Ai yi = 0
yi = 0, i ∈ I(x̂)
has a solution ỹ = 0, for which at least one component is strictly negative and
another is strictly positive, since otherwise we could assume ỹ ≥ 0, ỹ = 0,
to solve the homogeneous system Ay = 0, implying that x̂ + λỹ ∈ B ∀λ ≥ 0,
which, according to the inequality x̂ + λỹ ≥ λỹ − x̂, contradicts the
assumed boundedness of B. Hence we find for
α := max{λ | x̂ + λỹ ≥ 0},
β := min{λ | x̂ + λỹ ≥ 0}
that 0 < α < ∞ and 0 > β > −∞. Defining v := x̂ + αỹ and w := x̂ + β ỹ,
we have v, w ∈ B and—by the definitions of α and β—that |I(v)| ≤ k
Figure 21 LP: bounded feasible set.
and |I(w)| ≤ k such that, according to our induction assumption, r with

{x{i} ,
i = 1, · · · , r} the set of all feasible basic
r solutions, v =
i=1 λ i x {i}
,
r {i} r
where i=1 λ i = 1, λ i ≥ 0 ∀i, and w = i=1 µ i x , where i=1 µ i =
1, µi ≥ 0 ∀i. As is easily checked, we have x̂ = ρv + (1 − ρ)w with
ρ = −β/(α − β) ∈ (0, 1). This implies immediately that x̂ is a convex linear
combination of {x{i} , i = 1, · · · , r}. 2
The convex hull of finitely many points {x{1} , · · · , x{r} }, formally denoted
by conv{x{1} , · · · , x{r} }, is called a convex polyhedron or a bounded convex
polyhedral set (see Figure 21). Take for instance in IR2 the points z 1 =
(2, 2), z 2 = (8, 1), z 3 = (4, 3), z 4 = (7, 7) and z 5 = (1, 6). In Figure 22 we
have P̃ = conv{z 1 , · · · , z 5 }, and it is obvious that z 3 is not necessary to
generate P̃; in other words, P̃ = conv{z 1 , z 2 , z 3 , z 4 , z 5 } = conv{z 1 , z 2 , z 4 , z 5 }.
Hence we may drop z 3 without any effect on the polyhedron P̃, whereas
omitting any other of the five points would essentially change the shape of
the polyhedron. The points that really count in the definition of a convex
polyhedron are its vertices (z 1 , z 2 , z 4 and z 5 in the example). Whereas in two-
or three-dimensional spaces, we know by intuition what we mean by a vertex,
we need a formal definition for higher-dimensional cases: A vertex of a convex
polyhedron P is a point x̂ ∈ P such that the line segment connecting any two
points in P, both different from x̂, does not contain x̂. Formally,
∃y, z ∈ P, y = x̂ = z, λ ∈ (0, 1), such that x̂ = λy + (1 − λ)z.
It may be easily shown that for an LP with a bounded feasible set B the
feasible basic solutions x{i} , i = 1, · · · , r, coincide with the vertices of B.
By Proposition 1.10, the feasible set of a linear program is a convex
BASIC CONCEPTS 59
Figure 22 Polyhedron generated by its vertices.
polyhedron provided that B is bounded. Hence we have to find out under

what conditions B is bounded or unbounded respectively. For B = ∅ we have
seen already in the proof of Proposition 1.10 that the existence of a ỹ = 0 such
that Aỹ = 0, ỹ ≥ 0, would imply that B is unbounded. Therefore, for B to be
bounded, the condition {y | Ay = 0, y ≥ 0} = {0} is necessary. Moreover, we
have the following.
Proposition 1.11 The feasible set B = ∅ is bounded iff
{y | Ay = 0, y ≥ 0} = {0}.
Proof Given the above observations, it is only left to show that the condition
{y | Ay = 0, y ≥ 0} = {0} is sufficient for the boundedness of B. Assume
in contrast that B is unbounded. This means that we have feasible solutions
arbitrarily large in norm. Hence for any natural number K there exists an
xK ∈ B such that xK ≥ K. Defining
xK
z K := ∀K,
xK
we have ⎫
z K ≥ 0, ⎪
⎪
⎪
⎪
z K = 1, ⎬
Az K = b/xK , ∀K. (7.6)
⎪
⎪
and hence ⎪
⎪
⎭
Az K ≤ b/K
Therefore the sequence {z K , K = 1, 2, · · ·} has an accumulation point ẑ,
for which, according to (7.6), ẑ ≥ 0, ẑ = 1 and Aẑ = 0, and hence
Aẑ = 0, ẑ ≥ 0, ẑ = 0. 2
According to Proposition 1.11, the set C := {y | Ay = 0, y ≥ 0} plays a

decisive role for the boundedness or unboundedness of the feasible set B.
We see immediately that C is a convex cone, which means that for any
two elements y, z ∈ C it follows that λy + µz ∈ C ∀λ, µ ≥ 0. In addition,
we may show that C is a convex polyhedral cone, i.e. there exist finitely
{i}
manyys ∈ C, i = 1, · · · , s, such that any y ∈ C may be represented as
y = i=1 αi y {i} , αi ≥ 0 ∀i. Formally, we also may speak of the positive hull
denoted by pos{y {1} , · · · , y {s} } := {y | y = si=1 αi y {i} , αi ≥ 0 ∀i}.
Proposition 1.12 The set C = {y | Ay = 0, y ≥ 0} is a convex polyhedral
cone.
Proof Since for C = {0} the statement is trivial, we assume that C = {0}.
n
For any arbitrary ŷ ∈ C such that ŷ = 0 and hence i=1 ŷi > 0 we have, with
n n
µ := 1/ i=1 ŷi for ỹ := µŷ, that ỹ ∈ C := {y | Ayn= 0, i=1 yi = 1, y ≥ 0}.
Obviously C ⊂ C and, owing to the constraints i=1 yi = 1, y ≥ 0, the set C
is bounded. Hence, by Proposition 1.10, C is a convex polyhedron generated
{1} {s}
s {i}
s {y , · · · , y } such that ỹ has a representation
by itsfeasible basic solutions
ỹ= i=1 λi y with i=1 λi = 1, λi ≥ 0 ∀i,implying that ŷ = (1/µ)ỹ =
{i}
s
i=1 (λi /µ)y . This shows that C = {y | y = si=1 αi y {i} , αi ≥ 0 ∀i}. 2
In Figure 23 we see a convex polyhedral cone C and its intersection C

with the hyperplane H = {y | eT y = 1} (e = (1, · · · , 1)T ). The vectors
y {1} , y {2} and y {3} are the generating elements (feasible basic solutions) of C,
as discussed in the proof of Proposition 1.12, and therefore they are also the
generating elements of the cone C .
Now we are ready to describe the feasible set B of the linear program (7.1)
in general. Given the convex polyhedron P := conv{x{1} , · · · , x{r} } generated
by the feasible basic solutions {x{1} , · · · , x{r} } ⊂ B and the convex
polyhedral cone C = {y | Ay = 0, y ≥ 0}—given by its generating elements
as pos{y {1} , · · · , y {s} } as discussed in Proposition 1.12—we get the following.
Proposition 1.13 B is the algebraic sum of P and C, formally B = P + C,
meaning that every x̃ ∈ B may be represented as x̃ = z̃ + ỹ, where z̃ ∈ P and
ỹ ∈ C.
Proof Choose an arbitrary x̃ ∈ B. Since {y | Ay = 0, 0 ≤ y ≤ x̃} is compact,
the continuous function ϕ(y) := eT y, where e = (1, · · · , 1)T , attains its
maximum on this set. Hence there exists a ỹ such that
⎫
Aỹ = 0, ⎪
⎪
⎬
ỹ ≤ x̃,
(7.7)
ỹ ≥ 0, ⎪
⎪
⎭
eT ỹ = max{eT y | Ay = 0, 0 ≤ y ≤ x̃}.
BASIC CONCEPTS 61
Figure 23 Polyhedral cone intersecting the hyperplane H = {y | eT y = 1}.
Let x̂ := x̃−ỹ. Then x̂ ∈ B and {y | Ay = 0, 0 ≤ y ≤ x̂} = {0}, since otherwise

we should have a contradiction to (7.7). Hence for I(x̂) = {i | x̂i > 0} we have
{y | Ay = 0, yi = 0, i ∈ I(x̂), y ≥ 0} = {0}
and therefore, by Proposition 1.11, the feasible set
B1 := {x | Ax = b, xi = 0, i ∈ I(x̂), x ≥ 0}
is bounded and, observing that x̂ ∈ B1 , nonempty. From Proposition 1.10, it

follows that x̂ is a convex linear combination of the feasible basic solutions of
Ax = b
xi = 0, i ∈ I(x̂)
x≥0
which are obviously feasible basic solutions of our original constraints
Ax = b
x≥0
as well. It follows that x̂ ∈ P, and, by the above construction, we have ỹ ∈ C

and x̃ = x̂ + ỹ. 2
According to this proposition, the feasible set of any LP is constructed

as follows. First we determine the convex hull P of all feasible basic
solutions, which might look like that in Figure 21, for example; then we
add (algebraically) the convex polyhedral cone C (owing to Proposition 1.10
Figure 24 Adding the polyhedral cone C to the polyhedron P.
associated with the constraints of the LP) to P, which is indicated in

Figure 24.
The result of this operation—for an unbounded feasible set—is shown in
Figure 25; in the bounded case P would remain unchanged (as, for example,
in Figure 21), since then, according to Proposition 1.11, we have C = {0}.
A set given as algebraic sum of a convex polyhedron and a convex polyhedral
cone is called a convex polyhedral set. Observe that this definition contains
the convex polyhedron as well as the convex polyhedral cone as special cases.
We shall see later in this text that it is sometimes of interest to identify
so-called facets of convex polyhedral sets. Consider for instance a pyramid
(in IR3 ). You will certainly agree that this is a three-dimensional convex
polyhedral set. The set of boundary points again consists of different convex
polyhedral sets, namely sides (two-dimensional), edges (one-dimensional) and
vertices (zero-dimensional). The sides are called facets. In general, consider
an arbitrary convex polyhedral set B ⊂ IRn . Without loss of generality,
assume that 0 ∈ B (if not, one could, for any fixed z ∈ B, consider the
transposition B − {z} obviously containing the origin). The dimension of B,
dim B, is the smallest dimension of all linear spaces (in IRn ) containing B.
Therefore dim B ≤ n. For any linear subspace U ∈ IRn and any ẑ ∈ B the
intersection Bẑ,U := [{ẑ} + U] ∩ B = ∅ is again a convex polyhedral set. This
set is called a facet if
• ẑ is a boundary point of B and Bẑ,U does not contain interior points of B;
• dim U = dim Bẑ,U = dim B − 1.
In other words, a facet of B is a (maximal) piece of the boundary of B having
the dimension dim B − 1.
BASIC CONCEPTS 63
Figure 25 LP: unbounded feasible set.
The description of the feasible set of (7.1) given so far enables us to

understand immediately under which conditions the linear program (7.1) is
solvable and how the solution(s) may look.
Proposition 1.14 The linear program (7.1) is solvable iff
B = {x | Ax = b, x ≥ 0} = ∅ (7.8)
and
cT y ≥ 0 ∀y ∈ C = {y | Ay = 0, y ≥ 0}. (7.9)
Given that these two conditions are satisfied, there is at least one feasible basic
solution that is an optimal solution.
Proof Obviously condition (7.9) is necessary for the existence of an optimal

solution. If B = ∅ then we know from Proposition 1.13 that x ∈ B iff
r s
x = i=1 λi x{i} + j=1 µj y {j}
r
with λi ≥ 0 ∀i, µj ≥ 0 ∀j and i=1 λi = 1
where {x{1} , · · · , x{r} } is the set of all feasible basic solutions in B and
{y {1} , · · · , y {s} } is a set of elements generating C, for instance as described
in Proposition 1.12. Hence solving
min cT x
s.t. Ax = b,
x≥0
is equivalent to solving the problem

r s
min{ i=1 λi cT x{i} + j=1 µj cT y {j} }
r
s.t. i=1 λi = 1
λi ≥ 0 ∀i,
µj ≥ 0 ∀j.
The objective value of this latter program can be driven to −∞ if and only
if we have cT y {j} < 0 for at least one j ∈ {1, · · · , s}; otherwise, i.e. if
cT y {j} ≥ 0 ∀j ∈ {1, · · · , s} and hence cT y ≥ 0 ∀y ∈ C, the objective is
minimized by setting µj = 0 ∀j, and choosing λi0 = 1 and λi = 0 ∀i = i0 for
x{i0 } solving min1≤i≤r {cT x{i} }. 2
Observe that in general the solution of a linear program need not be unique.
Given the solvability conditions of Proposition 1.14 and the notation of its
proof, if cT y {j0 } = 0, we may choose µj0 > 0, and x{i0 } + µj0 y {j0 } is a solution
as well; and obviously it also may happen that min1≤i≤r {cT x{i} } is assumed
by more than just one feasible basic solution. In any case, if there is more
than one (different) solution for our linear program then there are infinitely
many owing to the fact that, given the optimal value γ, the set Γ of optimal
solutions is characterized by the linear constraints
Ax = b
cT x ≤ γ
x≥0
and therefore Γ is itself a convex polyhedral set.
1.7.2 The Simplex Algorithm

If we have the task of solving a linear program of the form (7.1) then, by
Proposition 1.14, we may restrict ourselves to feasible basic solutions. Let
x̂ ∈ B be any basic solution and, as before, I(x̂) = {i | x̂i > 0}. Under the
assumption (7.3), the feasible basic solution is called
• nondegenerate if |I(x̂)| = m, and

• degenerate if |I(x̂)| < m.
To avoid lengthy discussions, we assume in this section that for all feasible
basic solutions x{1} , · · · , x{r} of the linear program (7.1) we have
|I(x{i} )| = m, i = 1, · · · , r, (7.10)
i.e. that all feasible basic solutions are nondegenerate. For the case of
degenerate basic solutions, and the adjustments that might be necessary in
BASIC CONCEPTS 65
this case, the reader may consult the wide selection of books devoted to
linear programming in particular. Referring to our former presentation (7.5),
we have, owing to (7.10), that IB (x̂) = I(x̂), and, with the basic part
B = (Ai | i ∈ I(x̂)) and the nonbasic part N = (Ai | i ∈ I(x̂)) of the matrix
A, the constraints of (7.1) may be rewritten—using the basic and nonbasic
variables as introduced in (7.4)—as
⎫
x{B} = B −1 b − B −1 N x{N B} , ⎬
x{B} ≥ 0, (7.11)
{N B} ⎭
x ≥ 0.
Obviously this system yields our feasible basic solution x̂ iff x{N B} = 0, and
then we have, by our assumption (7.10), that x{B} = B −1 b > 0. Rearranging
the components of c analogously to (7.4) into the two vectors
{B}
ck = ci , i the kth element of I(x̂), k = 1, · · · , m,
{N B}
cl = ci , i the lth element of {1, · · · , n} − I(x̂), l = 1, · · · , n − m,
owing to (7.11), the objective may now be expressed as a function of the

nonbasic variables:
cT x = (c{B} )T x{B} + (c{N B} )T x{N B}
(7.12)
= (c{B} )T B −1 b + [(c{N B} )T − (c{B} )T B −1 N ] x{N B} .
This representation of the objective connected to the particular feasible basic

solution x̂ implies the optimality condition for linear programming—the so-
called simplex criterion.
Proposition 1.15 Under the assumption (7.10), the feasible basic solution
resulting from (7.11) for x{N B} = 0 is optimal iff
[(c{N B} )T − (c{B} )T B −1 N ]T ≥ 0. (7.13)
Proof By assumption (7.10), the feasible basic solution given by
x{B} = B −1 b − B −1 N x{N B} ,
{N B}
x =0
{N B}
satisfies x{B} = B −1 b > 0. Therefore any nonbasic variable xl may be
increased to some positive amount without violating the constraints x{B} ≥ 0.
Furthermore, increasing the nonbasic variables is the only feasible change
applicable to them, owing to the constraints x{N B} ≥ 0. From the objective
presentation in (7.12), we see immediately that
cT x̂ = (c{B} )T B −1 b
≤ (c{B} )T B −1 b + [(c{N B} )T − (c{B} )T B −1 N ] x{N B} ∀x{N B} ≥ 0
iff [(c{N B} )T − (c{B} )T B −1 N ]T ≥ 0. 2
Motivated by the above considerations, we call any nonsingular m × m

submatrix B = (Ai | i ∈ IB ) of A a feasible basis for the linear program
(7.1) if B −1 b ≥ 0. Obviously, on rearranging the variables as before into basic
variables x{B} —belonging to B—and nonbasic variables x{N B} —belonging
to N = (Ai | i ∈ IB )—the objective γ and the constraints of (7.1) read
(see (7.11) and (7.12)) as
γ = (c{B} )T B −1 b + [(c{N B} )T − (c{B} )T B −1 N ] x{N B} ,

{B}
x = B −1 b − B −1 N x{N B} ,
{B}
x ≥ 0,
x{N B} ≥ 0,
and x{N B} = 0 corresponds to a feasible basic solution—under our

assumption (7.10), satisfying even x{B} = B −1 b > 0 instead of only x{B} =
B −1 b ≥ 0 in general.
Now we are ready—using the above notation—to formulate the classical
solution procedure of linear programming: the simplex method
Simplex method.
Step 1 Determine a feasible basis B = (Ai | i ∈ IB ) for (7.1) and

N = (Ai | i ∈ IB ).
Step 2 If the simplex criterion (7.13) is satisfied then stop with
x{B} = B −1 b, x{N B} = 0
being an optimal solution; otherwise, there is some ρ ∈ {1, · · · , n − m}

such that for the ρth component of [(c{N B} )T − (c{B} )T B −1 N ]T we
have
[(c{N B} )T − (c{B} )T B −1 N ]T
ρ < 0,
{N B}
and we increase the ρ-th nonbasic variable xρ .
{N B}
If increasing xρ is not “blocked” by the constraints x{B} ≥ 0,
{N B}
i.e. if xρ → ∞ is feasible, then inf B γ = −∞ such that our problem
has no (finite) optimal solution.
{N B}
If, on the other hand, increasing xρ is “blocked” by one of the
{B}
constraints xi ≥ 0, i = 1, · · · , m, such that, for instance, for some
{B}
µ ∈ {1, · · · , m} the basic variable xµ is the first one to become
{B} {N B}
xµ = 0 while increasing xρ , then go to step 3.
BASIC CONCEPTS 67
Step 3 Exchange the µth column of B with the ρth column of N , yielding
new basic and nonbasic parts B̃ and Ñ of A such that B̃ contains
Nρ as its µth column and Ñ contains Bµ as its ρth column. Redefine
B := B̃ and N := Ñ , and rearrange x{B} , x{N B} , c{B} and c{N B}
correspondingly, and then return to step 2.
Remark 1.5 The following comments on the single steps of the simplex
method may be helpful for a better understanding of this procedure:
Step 1 Obviously we assume that B = ∅. The existence of a feasible

basis B follows from Propositions 1.9 and 1.8. Because of our
assumption (7.10), we have B −1 b > 0.
Step 2 (a) If for a feasible basis B we have
[(c{N B} )T − (c{B} )T B −1 N ]T ≥ 0
then by Proposition 1.15 this basis (i.e. the corresponding basic

solution) is optimal.
(b) If the simplex criterion is violated for the feasible basic solution
belonging to B given by x{B} = B −1 b, x{N B} = 0, then
there must be an index ρ ∈ {1, · · · , n − m} such that α0ρ :=
[(c{N B} )T − (c{B} )T B −1 N ]T
ρ < 0, and, keeping all but the ρth
{N B}
nonbasic variables on their present values xj = 0, j = ρ, with
α·ρ := −B −1 Nρ , the objective and the basic variables have the
representations
{N B}
γ = (c{B} )T B −1 b + α0ρ xρ ,
{B} {N B}
x = B −1 b + α·ρ xρ .
According to these formulae, we conclude immediately that for

α·ρ ≥ 0 the nonnegativity of the basic variables would never
{N B}
be violated by increasing xρ arbitrarily such that we had
inf B γ = −∞, whereas for α·ρ ≥ 0 it would follow that the
set of rows {i | αiρ < 0, 1 ≤ i ≤ m} = ∅, and consequently, with
{N B}
β := B −1 b, the constraints x{B} = β + α·ρ xρ ≥ 0 would
{N B}
“block” the increase of xρ at some positive value (remember
that, by the assumption (7.10), we have β > 0). More precisely,
we now have to observe the constraints
βi + αiρ x{N
ρ
B}
≥ 0 for i ∈ {i | αiρ < 0, 1 ≤ i ≤ m}
or equivalently
βi
x{N B}
≤ for i ∈ {i | αiρ < 0, 1 ≤ i ≤ m}.
ρ
−αiρ
Hence, with µ ∈ {i | αiρ < 0, 1 ≤ i ≤ m} denoting a row for

which ! β "
βµ i
= min αiρ < 0, 1 ≤ i ≤ m ,
−αµρ −αiρ
{B} {N B}
xµ is the first basic variable to decrease to zero if xρ is
increased to the value βµ /(−αµρ ), and we observe that at the
same time the objective value is changed to
>0

{B} T β
γ = (c ) β + α0ρ −αµ
µρ
<0 >0
{B} T
< (c ) β
such that we have a strict decrease of the objective.

Step 3 The only point to understand here is that B̃ as constructed in this
step is again a basis. By assumption, B was a basis, i.e. the column
set
(B1 , · · · , Bµ , · · · , Bm ) was linearly independent. Entering step 3
according to step 2 asserts that for α·ρ = −B −1 Nρ we have αµρ < 0,
i.e. in the representation of the column Nρ by the basic columns,
Nρ = − m i=1 Bi αiρ the column Bµ appears with a nonzero coefficient.
In this case it is well known from linear algebra that the column set
(B1 , · · · , Nρ , · · · , Bm ) is linearly independent as well, and hence B̃ is a
basis.
The operation of changing the basis by exchanging one column (step 3) is
usually called a pivot step. 2
Summarizing the above remarks immediately yields the following.

Proposition 1.16 If the linear program (7.1) is feasible then the simplex
method yields—under the assumption of nondegeneracy (7.10)—after finitely
many steps either a solution or else the information that there is no finite
solution, i.e. that inf B γ = −∞.
Proof As mentioned in Remark 1.5, step 3, the objective strictly decreases
in every pivot step. During the cycles (steps 2 and 3) of the method, we only
consider feasible bases. Since there are no more than finitely many feasible
BASIC CONCEPTS 69
bases for any linear program of the form (7.1), the simplex method must end
after finitely many cycles. 2
Remark 1.6 In step 2 of the simplex method it may happen that the simplex
criterion is not satisfied and that we discover that inf B γ = −∞. It is worth
mentioning that in this situation we may easily find a generating element of
the cone C associated with B, as discussed in Proposition 1.12. With the above
notation, we then have a feasible basis B, and for some column Nρ = 0 we
have B −1 Nρ ≤ 0. Then, with e = (1, · · · , 1)T of appropriate dimensions, for
(ŷ {B} , ŷ {N B} ) satisfying
{N B}
ŷ {B} = −B −1 Nρ ŷρ ,
{N B} 1
ŷρ = ,
−eT B −1 Nρ + 1
{N B}
ŷl = 0 for l = ρ
it follows that
B ŷ {B} + N ŷ {N B} = 0
{N B} {N B}
eT ŷ {B} + eT ŷ {N B} = −eT B −1 Nρ ŷρ + ŷρ
{N
= (−eT B −1 Nρ + 1)ŷρ
B}
= 1,
{B}
ŷ ≥ 0,
ŷ {N B} ≥ 0.
Observe that, with B = (B1 , · · · , Bm ) a basis of IRm , owing to
v = B −1 Nρ ≤ 0, and hence 1 − eT v ≥ 1,
we have

B1 · · · Bm 0 B1 · · · Bm 0
rk = rk
1 ··· 1 1 − eT v 1 ··· 1 1

B1 · · · Bm 0
= rk .
0 ··· 0 1
It follows that
B1 B2 · · · Bm Nρ
1 1 ··· 1 1
is a basis of IRm+1 . Hence (ŷ {B} , ŷ {N B} ) is one of the generating elements of
the convex polyhedral cone {(y {B} , y {N B} ) | By {B} + N y {N B} = 0, y {B} ≥
0, y {N B} ≥ 0}, as derived in Proposition 1.12. 2
1.7.3 Duality Statements

Given the linear program (7.1) as so-called primal program
⎫
min cT x ⎬
s.t. Ax = b, (7.14)
⎭
x ≥ 0,
the corresponding dual program is formulated as

max bT u
(7.15)
s.t. AT u ≤ c.
Remark 1.7 Instead of stating a whole bunch of rules on how to assign

the correct dual program to any of the various possible formulations of the
primal linear program, we might recommend transformation of the primal
program to the standard form (7.14), followed by the assignment of the linear
program (7.15) as its dual. Let us just give some examples.
Example 1.5 Assume that our primal program is of the form
min cT x
s.t. Ax ≥ b,
x ≥ 0,
which, by transformation to the standard form, is equivalent to
min cT x
s.t. Ax − Iy = b,
x ≥ 0,
y ≥ 0,
I being the m × m identity matrix, and, according to the above definition,

has the dual program
max bT u
s.t. AT u ≤ c,
−Iu ≤ 0,
or equivalently
max bT u
s.t. AT u ≤ c,
u ≥ 0.
Hence for this case the pair of the primal and its dual program looks like
min cT x max bT u
s.t. Ax ≥ b, s.t. AT u ≤ c,
x ≥ 0; u ≥ 0.
BASIC CONCEPTS 71
Example 1.6 Considering the primal program

min cT x
s.t. Ax ≤ b,
x≥0
in its standard form
min cT x
s.t. Ax + Iy = b,
x ≥ 0,
y≥0
would yield the dual program
max bT u
s.t. AT u ≤ c,
u ≤ 0,
or equivalently, with v := −u,
max (−bT v)
s.t. AT v ≥ −c,
v ≥ 0.
Therefore we now have the following pair of a primal and the corresponding
dual program:
min cT x max (−bT v)
s.t. Ax ≤ b, s.t. AT v ≥ −c,
x ≥ 0; v ≥ 0.
2
Example 1.7 Finally consider the primal program

max g T x
s.t. Dx ≤ f.
This program is of the same form as the dual of our standard linear
program (7.14) and—using the fact that for any function ϕ defined on some set
M we have supx∈M ϕ(x) = − inf x∈M {−ϕ(x)}—its standard form is written
as
− min (−g T x+ + g T x− )
s.t. Dx+ − Dx− + Iy = f,
x+ ≥ 0,
x− ≥ 0,
y ≥ 0,
with the dual program

− max f T z
s.t. DT z ≤ −g,
−DT z ≤ g,
Iz ≤ 0
which is (with w := −z) equivalent to
min f T w
s.t. DT w = g,
w ≥ 0,
such that we have the dual pair
max g T x min f T w
s.t. Dx ≤ f ; s.t. DT w = g,
w ≥ 0.
2
Hence, by comparison with our standard forms of the primal program (7.14)
and the dual program (7.15), it follows that the dual of the dual is the primal
program. 2
There are close relations between a primal linear program and its dual
program. Let us denote the feasible set of the primal program (7.14) by B and
that of its dual program by D. Furthermore, let us introduce the convention
that
inf x∈B cT x = +∞ if B = ∅,
(7.16)
supu∈D bT u = −∞ if D = ∅.
Then we have as a first statement the following so-called weak duality theorem:
Proposition 1.17 For the primal linear program (7.14) and its dual (7.15)
inf cT x ≥ sup bT u.
x∈B u∈D
Proof If either B = ∅ or D = ∅ then the proposition is trivial owing to our

convention (7.16). Assume therefore that both feasible sets are nonempty and
choose arbitrarily an element x̂ ∈ B and an element û ∈ D. Then, from (7.15),
we have
c − AT û ≥ 0,
and, by scalar multiplication with x̂ ≥ 0,
x̂T (c − AT û) ≥ 0,
BASIC CONCEPTS 73
which, observing that Ax̂ = b by (7.14), implies
x̂T c − bT û ≥ 0.
Since x̂ ∈ B and û ∈ D were arbitrarily chosen, we have
cT x ≥ b T u ∀x ∈ B, u ∈ D,
and hence
inf x∈B cT x ≥ supu∈D bT u.
2
In view of this proposition, the question arises as to whether or when it

might happen that
inf cT x > sup bT u.
x∈B u∈D
Example 1.8 Consider the following primal linear program:
min{3x1 + 3x2 − 16x3 }

s.t. 5x1 + 3x2 − 8x3 = 2,
−5x1 + 3x2 − 8x3 = 4,
xi ≥ 0, i = 1, 2, 3,
and its dual program
max{2u1 + 4u2 }
s.t. 5u1 − 5u2 ≤ 3,
3u1 + 3u2 ≤ 3,
−8u1 − 8u2 ≤ −16.
Adding the equations of the primal program, we get
6x2 − 16x3 = 6,
and hence
x2 = 1 + 83 x3 ,
which, on insertion into the first equation, yields
x1 = 15 (2 − 3 − 8x3 + 8x3 )
= − 15 ,
showing that the primal program is not feasible.

Looking at the dual constraints, we get from the second and third
inequalities that
u1 + u2 ≤ 1,
u1 + u2 ≥ 2,
such that also the dual constraints do not allow a feasible solution. Hence, by
our convention (7.16), we have for this dual pair
inf cT x = +∞ > sup bT u = −∞.

x∈B u∈D
However, the so-called duality gap in the above example does not occur
so long as at least one of the two problems is feasible, as is asserted by the
following strong duality theorem of linear programming.
Proposition 1.18 Consider the feasible sets B and D of the dual pair of
linear programs (7.14) and (7.15) respectively. If either B = ∅ or D = ∅ then
it follows that
inf cT x = sup bT u.
x∈B u∈D
If one of these two problems is solvable then so is the other, and we have
min cT x = max bT u.
x∈B u∈D
Proof Assume that B = ∅.

If inf x∈B cT x = −∞ then it follows from the weak duality theorem that
supu∈D bT u = −∞ as well, i.e. that the dual program (7.15) is infeasible.
If the primal program (7.14) is solvable then we know from Proposition 1.14
that there is an optimal feasible basis B such that the primal program may
be rewritten as
min{(c{B} )T x{B} + (c{N B} )T x{N B} }
s.t. Bx{B} + N x{N B} = b,
x{B} ≥ 0,
{N B}
x ≥ 0,
and therefore the dual program reads as
max bT u
s.t. B T u ≤ c{B} ,
N T u ≤ c{N B} .
For B an optimal feasible basis, owing to Proposition 1.15, the simplex

criterion
[(c{N B} )T − (c{B} )T B −1 N ]T ≥ 0
has to hold. Hence it follows immediately that û := (B T )−1 c{B} satisfies the
dual constraints. Additionally, the dual objective value bT û = bT (B T )−1 c{B}
BASIC CONCEPTS 75
is equal to the primal optimal value (c{B} )T B −1 b. In view of Proposition 1.17,

it follows that û is an optimal solution of the dual program. 2
An immediate consequence of the strong duality theorem is Farkas’ lemma,

which yields a necessary and sufficient condition for the feasibility of a system
of linear constraints, and may be stated as follows.
Proposition 1.19 The set
{x | Ax = b, x ≥ 0} = ∅
if and only if
AT u ≥ 0 implies that bT u ≥ 0.
Proof Assume that ũ satisfies AT ũ ≥ 0 and that {x | Ax = b, x ≥ 0} = ∅.
Then let x̂ be a feasible solution, i.e. we have
Ax̂ = b, x̂ ≥ 0,
and, by scalar multiplication with ũ, we get
ũT b =
ũT A
x̂ ≥ 0,
≥0 ≥0
so that the condition is necessary.

Assume now that the following condition holds:
AT u ≥ 0 implies that bT u ≥ 0.
Choosing any û = 0 and defining c := AT û, it follows from Proposition 1.14

that the linear program
min bT u
s.t. AT u ≥ c
is solvable. Then its dual program
max cT x
s.t. Ax = b,
x≥0
is solvable and hence feasible. 2
1.7.4 A Dual Decomposition Method

In Section 1.5 we discussed stochastic linear programs with linear recourse
and mentioned in particular the case of a finite support Ξ of the probability
distribution. We saw that the deterministic equivalent—the linear program

(5.2)—has a dual decomposition structure. We want to sketch a solution
method that makes use of this structure. For simplicity, and just to present
the essential ideas, we restrict ourselves to a support Ξ containing just one
realization such that the problem to discuss is reduced to
⎫
min{cT x + q T y} ⎪
⎪
⎪
s.t. Ax = b, ⎪
⎬
T x + W y = h, (7.17)
⎪
x ≥ 0, ⎪⎪
⎪
⎭
y ≥ 0.
In addition, we assume that the problem is solvable and that the set {x |
Ax = b, x ≥ 0} is bounded. The above problem may be restated as
min{cT x + f (x)}
s.t. Ax = b,
x ≥ 0,
with
f (x) := min{q T y | W y = h − T x, y ≥ 0}.
Our recourse function f (x) is easily seen to be piecewise linear and convex. It
is also immediate that the above problem can be replaced by the equivalent
problem
min{cT x + θ}
s.t. Ax = b
θ − f (x) ≥ 0
x ≥ 0;
however, this would require that we know the function f (x) explicitly in
advance. This will not be the case in general. Therefore we may try to
construct a sequence of new (additional) linear constraints that can be
used to define a monotonically decreasing feasible set B1 of (n + 1)-vectors
(x1 , · · · , xn , θ)T such that finally, with B0 := {(xT , θ)T | Ax = b, x ≥ 0, θ ∈
IR}, the problem min(x,θ)∈B0 ∩B1 {cT x + θ} yields a (first-stage) solution of our
problem (7.17).
After these preparations, we may describe the following particular method.
Dual decomposition method
Step 1 With θ0 a lower bound for
min{q T y | Ax = b, T x + W y = h, x ≥ 0, y ≥ 0},
BASIC CONCEPTS 77
solve the program
min{cT x + θ | Ax = b, θ ≥ θ0 , x ≥ 0}
yielding a solution (x̂, θ̂). Let B1 := {IRn × {θ} | θ ≥ θ0 }.

Step 2 Using the last first-stage solution x̂, evaluate the recourse function
f (x̂) = min{q T y | W y = h − T x̂, y ≥ 0}

= max{(h − T x̂)T u | W T u ≤ q}.
Now we have to distinguish two cases.

(a) If f (x̂) = +∞ then x̂ is not feasible with respect to all constraints
of (7.17) (i.e. x̂ does not satisfy the induced constraints discussed
in Proposition 1.3), and by Proposition 1.14 we have a ũ such
that W T ũ ≤ 0 and (h − T x̂)T ũ > 0. On the other hand, for any
feasible x there must exist a y ≥ 0 such that W y = h − T x. Scalar
multiplication of this equation by ũ yields
ũT (h − T x) = ũTW y ≤ 0,

≤0 ≥0
and hence
ũT h ≤ ũT T x,
which has to hold for any feasible x, and obviously does not hold
for x̂, since ũT (h − T x̂) > 0. Therefore we introduce the feasibility
cut, cutting off the infeasible solution x̂:
ũT (h − T x) ≤ 0.

Then we redefine B1 := B1 {(xT , θ) | ũT (h − T x) ≤ 0} and go on
to step 3.
(b) Otherwise, if f (x̂) is finite, we have for the recourse problem (see
the proof of Proposition 1.18) simultaneously—for x̂—a primal
optimal basic solution ŷ and a dual optimal basic solution û. From
the dual formulation of the recourse problem, it is evident that
f (x̂) = (h − T x̂)T û,
whereas for any x we have
f (x) = sup{(h − T x)T u | W T u ≤ q}

≥ (h − T x)T û
= ûT (h − T x).
Figure 26 Dual decomposition: optimality cuts.
The intended constraint θ ≥ f (x) implies the linear constraint
θ ≥ ûT (h − T x),
which is violated by (x̂T , θ̂)T iff (h − T x̂)T û > θ̂; in this case

we introduce the optimality cut (see Figure 26), cutting off the
nonoptimal solution (x̂T , θ̂)T :
θ ≥ ûT (h − T x).

Correspondingly, we redefine B1 := B1 {(xT , θ) | θ ≥ ûT (h−T x)}
and continue with step 3; otherwise, i.e. if f (x̂) ≤ θ̂, we stop, with
x̂ being an optimal first-stage solution.
Step 3 Solve the updated problem
min{cT x + θ | (xT , θ) ∈ B0 ∩ B1 },
yielding the optimal solution (x̃T , θ̃)T .

With (x̂T , θ̂)T := (x̃T , θ̃)T , we return to step 2.
Remark 1.8 We briefly sketch the arguments regarding the proper

functioning of this method.
Step 1 We have assumed problem (7.17) to be solvable, which implies, by
Proposition 1.14, that
{(x, y) | Ax = b, T x + W y = h, x ≥ 0, y ≥ 0} = ∅,
BASIC CONCEPTS 79
{v | W v = 0, q T v < 0, v ≥ 0} = ∅.
In addition, we have assumed {x | Ax = b, x ≥ 0} to be bounded.

Hence inf{f (x) | Ax = b, x ≥ 0} is finite such that the lower bound
θ0 exists. This (and the boundedness of {x | Ax = b, x ≥ 0}) implies
that
min{cT x + θ | Ax = b, θ ≥ θ0 , x ≥ 0}
is solvable.
Step 2 If f (x̂) = +∞, we know from Proposition 1.14 that {u | W T u ≤
0, (h − T x̂)T u > 0} = ∅, and, according to Remark 1.6, for the convex
polyhedral cone {u | W T u ≤ 0} we may find with the simplex method
one of the generating elements ũ mentioned in Proposition 1.12 that
satisfies (h − T x̂)T ũ > 0. By Proposition 1.12, we have finitely many
generating elements for the cone {u | W T u ≤ 0} such that, after
having used all of them to construct feasibility cuts, for all feasible
x we should have (h − T x)T u ≤ 0 ∀u ∈ {u | W T u ≤ 0} and hence
solvability of the recourse problem. This shows that f (x̂) = +∞ may
appear only finitely many times within this method.
If f (x̂) is finite, the simplex method yields primal and dual optimal
feasible basic solutions ŷ and û respectively. Assume that we already
had the same dual basic solution ũ := û in a previous step to construct
an optimality cut
θ ≥ ũT (h − T x);
then our present θ̂ has to satisfy this constraint for x = x̂ such that
θ̂ ≥ ũT (h − T x̂)
= ûT (h − T x̂)
holds, or equivalently we have f (x̂) ≤ θ̂ and stop the procedure. From

the above inequalities, it follows that
θ̂ ≥ (h − T x̂)T u{i} , i = 1, · · · , k,
if u{1} , · · · , u{k} denote the feasible basic solutions in {u | W T u ≤ q}

used so far for optimality cuts. Observing that in step 3 for any x we
minimize θ with respect to B1 this implies that
θ̂ = max (h − T x̂)T u{i} .

1≤i≤k
Given our stopping rule f (x̂) ≤ θ̂, with the set of all feasible basic
solutions, {u{1} , · · · , u{k} , · · · , u{r} }, of {u | W T u ≤ q}, it follows that

θ̂ = max1≤i≤k (h − T x̂)T u{i}
≤ max1≤i≤r (h − T x̂)T u{i}
= f (x̂)
≤ θ̂
and hence θ̂ = f (x̂), which implies the optimality of x̂.
2
Summarizing the above remarks we have the following.

Proposition 1.20 Provided that the program (7.17) is solvable and {x |
Ax = b, x ≥ 0} is bounded, the dual decomposition method yields an optimal
solution after finitely many steps.
We have described this method for the data structure of the linear program
(7.17) that would result if a stochastic linear program with recourse had just
one realization of the random data. To this end, we introduced the feasibility
and optimality cuts for the recourse function f (x) := min{q T y | W y =
h − T x, y ≥ 0}. The modification for a finite discrete distribution with K
realizations is immediate. From the discussion in Section 1.5, our problem is
of the form
! K "
T
min c x + q iT y i
i=1
s.t. Ax =b
T i x + W y i = hi , i = 1, · · · , K
x ≥ 0,
y i ≥ 0, i = 1, · · · , K.
Thus we may simply introduce feasibility and optimality cuts for all the re-
course functions fi (x) := min{q iT y i | W y i = hi − T i x, y i ≥ 0}, i = 1, · · · , K,
yielding the so-called multicut version of the dual decomposition method. Al-
ternatively, combining the single cuts corresponding to the particular blocks
i = 1, · · · , K with their respective probabilities leads to the so-called L-shaped
method.
1.8 Nonlinear Programming
In this section we summarize some basic facts about nonlinear programming

problems written in the standard form

min f (x)
(8.1)
s.t. gi (x) ≤ 0, i = 1, · · · , m.
BASIC CONCEPTS 81
The feasible set is again denoted by B:
B := {x | gi (x) ≤ 0, i = 1, · · · , m}.
As in the previous section, any other nonlinear program, for instance
min f (x)
s.t. gi (x) ≤ 0, i = 1, · · · , m,
x≥0
or
min f (x)
s.t. gi (x) ≤ 0, i = 1, · · · , m1 ,
gi (x) = 0, i = m1+1 , · · · , m,
x≥0
or
min f (x)
s.t. gi (x) ≥ 0, i = 1, · · · , m,
x ≥ 0,
may be transformed into the standard form (8.1).

We assume throughout this section that the functions f, gi : IRn −→ IR are
given, that at least one of them is not a linear function, and that all of them
∂f ∂gi
are continuously (partially) differentiable (i.e. and are continuous).
∂xj ∂xj
Occasionally we restrict ourselves to the case that the functions are convex,
since we shall not widely deal with nonconvex problems in this book. This
implies, according to Lemma 1.1 that any local minimum of program (8.1) is
a global minimum.
First of all, we have to refer to a well known fact from analysis.
Proposition 1.21 The function ϕ : IRn −→ IR is convex iff for all arbitrarily

chosen x, y ∈ IRn we have
(y − x)T ∇ϕ(x) ≤ ϕ(y) − ϕ(x).
In other words, for a convex function, a tangent (hyperplane) at any arbitrary

point (of its graph) supports the function everywhere from below; a hyperplane
with this property is called a supporting hyperplane for this function (see
Figure 27).
We know from calculus that for some x̂ ∈ IRn to yield a local minimum for
a differentiable function ϕ : IRn −→ IR we have the necessary condition
∇ϕ(x̂) = 0.
Figure 27 Convex function with tangent as supporting hyperplane.
If, moreover, the function ϕ is convex then, owing to Proposition 1.21, this
condition is also sufficient for x̂ to be a global minimum, since then for any
arbitrary x ∈ IRn we have
0 = (x − x̂)T ∇ϕ(x̂) ≤ ϕ(x) − ϕ(x̂)
and hence
ϕ(x̂) ≤ ϕ(x) ∀x ∈ IRn .
Whereas the above optimality condition is necessary for unconstrained
minimization, the situation may become somewhat different for constrained
minimization.
Example 1.9 For x ∈ IR consider the simple problem
min ψ(x) = x2
s.t. x ≥ 1,
with the obvious unique solution

dψ
x̂ = 1, with ∇ψ(x̂) = (x̂) = 2.
dx
Hence we cannot just transfer the optimality conditions for unconstrained op-
timization to the constrained case. 2
Therefore we shall first deal with the necessary and/or sufficient conditions
for some x̂ ∈ IRn to be a local or global solution of the program (8.1).
BASIC CONCEPTS 83
1.8.1 The Kuhn–Tucker Conditions

Remark 1.9 To get an idea of what kind of optimality conditions we may
expect for problems of the type (8.1), let us first—contrary to our general
assumption—consider the case where f, gi , i = 1, · · · , m, are linear functions

f (x) := cT x,
(8.2)
gi (x) := aT
i x − bi , i = 1, · · · , m,
such that we have the gradients

∇f (x) = c,
(8.3)
∇gi (x) = ai ,
and problem (8.1) becomes the linear program

min cT x
(8.4)
s.t. aT
i x ≤ bi , i = 1, · · · , m.
Although we did not explicitly discuss optimality conditions for linear

programs in the previous section, they are implicitly available in the duality
statements discussed there. The dual problem of (8.4) is
⎫
max{−bT u} ⎪
⎪
m ⎪
⎬
s.t. − ai ui = c, (8.5)
⎪
⎪
i=1 ⎪
⎭
u ≥ 0.
Let A be the m × n matrix having aT i , i = 1, · · · , m, as rows. The difference

of the primal and the dual objective functions can then be written as
cT x + bT u = cT x + uT Ax − uT Ax + bT u
= (c + AT u)T x + (b − Ax)T u
m
m (8.6)
= [∇f (x) + ui ∇gi (x)]T x − ui gi (x).
i=1 i=1
From the duality statements for linear programming (Propositions 1.17

and 1.18), we know the following.
(a) If x̂ is an optimal solution of the primal program (8.4) then, by the strong
duality theorem (Proposition 1.18), there exists a solution û of the dual
program (8.5) such that the difference of the primal and dual objective
vanishes. For the pair of dual problems (8.4) and (8.5) this means that
cT x̂ − (−bT û) = cT x̂ + bT û = 0. In view of (8.6) this may also be stated
as the necessary condition

m
∃û ≥ 0 such that ∇f (x̂) + ûi ∇gi (x̂) = 0,
i=1
m
ûi gi (x̂) = 0.
i=1
(b) if we have a primal feasible and a dual feasible solution x̃ and ũ

respectively, such that the difference of the respective objectives is zero
then, by the weak duality theorem (Proposition 1.17), x̃ solves the primal
problem; in other words, given a feasible x̃, the condition

m
∃ũ ≥ 0 such that ∇f (x̃) + ũi ∇gi (x̃) = 0,
i=1
m
ũi gi (x̃) = 0
i=1
is sufficient for x̃ to be a solution of the program (8.4).
Remark 1.10 The optimality condition derived in Remark 1.9 for the linear
case could be formulated as follows:
(1) For the feasible x̂ the negative gradient of the objective f —i.e. the
direction of the greatest (local) descent of f —is equal (with the multipliers
ûi ≥ 0) to a nonnegative linear combination of the gradients of those
constraint functions gi that are active at x̂, i.e. that satisfy gi (x̂) = 0.
(2) This corresponds to the fact that the multipliers satisfy the complemen-
tarity conditions ûi gi (x̂) = 0, i = 1, · · · , m, stating that the multipliers
ûi are zero for those constraints that are not active at x̂, i.e. that satisfy
gi (x̂) < 0.
In conclusion, this optimality condition says that −∇f (x̂) must be contained
in the convex polyhedral cone generated by the gradients ∇gi (x̂) of the con-
straints being active in x̂. This is one possible formulation of the Kuhn–Tucker
conditions illustrated in Figure 28. 2
Let us now return to the more general nonlinear case and consider the
following question. Given that x̂ is a (local) solution, under what assumption
BASIC CONCEPTS 85
Figure 28 Kuhn–Tucker conditions.
does this imply that the above optimality conditions,

m ⎫
⎪
∃û ≥ 0 such that ∇f (x̂) + ûi ∇gi (x̂) = 0, ⎪
⎪
⎬
i=1 (8.7)
m
⎪
ûi gi (x̂) = 0, ⎪
⎪
⎭
i=1
hold? Hence we ask under what assumption are the conditions (8.7) necessary
for x̂ to be a (locally) optimal solution of the program (8.1). To answer this
question, let I(x̂) := {i | gi (x̂) = 0}, such that the optimality conditions (8.7)
are equivalent to
! "

u ui ∇gi (x̂) = −∇f (x̂), ui ≥ 0 for i ∈ I(x̂) = ∅.
i∈I(x̂)
Observing that ∇gi (x̂) and ∇f (x̂) are constant vectors when x is fixed at x̂,
the condition of Farkas’ lemma (Proposition 1.19) is satisfied if and only if
the following regularity condition holds in x̂:
RC 0
z T ∇gi (x̂) ≤ 0, i ∈ I(x̂) implies that z T ∇f (x̂) ≥ 0. (8.8)
Hence we have the rigorous formulation of the Kuhn–Tucker conditions:
Proposition 1.22 Given that x̂ is a (local) solution of the nonlinear

program (8.1), under the assumption that the regularity condition RC 0 is
satisfied in x̂ it necessarily follows that

m
∃û ≥ 0 such that ∇f (x̂) + ûi ∇gi (x̂) = 0,
i=1
m
ûi gi (x̂) = 0.
i=1
Example 1.10 The Kuhn–Tucker conditions need not hold if the regularity
condition cannot be asserted. Consider the following simple problem (x ∈ IR1 ):
min{x | x2 ≤ 0}.
Its unique solution is x̂ = 0. Obviously we have
∇f (x̂) = (1), ∇g(x̂) = (0),
and there is no way to represent ∇f (x̂) as (positive) multiple of ∇g(x̂). (Need-

less to say, the regularity condition RC 0 is not satisfied in x̂.) 2
We just mention that for the case of linear constraints the Kuhn–Tucker
conditions are necessary for optimality, without the addition of any regularity
condition.
Instead of condition RC 0, there are various other regularity conditions
popular in optimization theory, only two of which we shall mention here. The
first is stated as
RC 1
∀z = 0 s.t. z T ∇gi (x̂) ≤ 0, i ∈ I(x̂), ∃{xk | xk = x̂, k = 1, 2, · · ·} ⊂ B
such that
xk − x̂ z
lim xk = x̂, lim = .
k→∞ k→∞ x − x̂
k z
The second—used frequently for the convex case, i.e. if the functions gi are
convex—is the Slater condition
RC 2
∃x̃ ∈ B such that gi (x̃) < 0 ∀i. (8.9)
Observe that there is an essential difference among these regularity conditions:
to verify RC 0 or RC 1, we need to know the (locally) optimal point for which
we want the Kuhn–Tucker conditions (8.7) to be necessary, whereas the Slater
BASIC CONCEPTS 87
Figure 29 The Slater condition implies RC 1.
condition RC 2—for the convex case—requires the existence of an x̃ such that

gi (x̃) < 0 ∀i, but does not refer to any optimal solution. Without proof we
might mention the following.
Proposition 1.23
(a) The regularity condition RC 1 (in any locally optimal solution) implies
the regularity condition RC 0.
(b) For the convex case the Slater condition RC 2 implies the regularity
condition RC 1 (for every feasible solution).
In Figure 29 we indicate how the proof of the implication RC 2 =⇒ RC 1 can

be constructed.
Based on these facts we immediately get the following.
Proposition 1.24
(a) If x̂ (locally) solves problem (8.1) and satisfies RC 0 then the Kuhn–
Tucker conditions (8.7) necessarily hold in x̂.
(b) If the functions f, gi , i = 1, · · · , m, are convex and the Slater condition
RC 2 holds, then x̂ ∈ B (globally) solves problem (8.1) if and only if the
Kuhn–Tucker conditions (8.7) are satisfied for x̂.
Proof: Referring to Proposition 1.23, the necessity of the Kuhn–Tucker

conditions has already been demonstrated. Hence we need only show that in
the convex case the Kuhn–Tucker conditions are also sufficient for optimality.
Assume therefore that we have an x̂ ∈ B and a û ≥ 0 such that

m
∇f (x̂) + ûi ∇gi (x̂) = 0,
i=1

m
ûi gi (x̂) = 0.
i=1
Then, with I(x̂) = {i | gi (x̂) = 0}, we have

∇f (x̂) = − ûi ∇gi (x̂)
i∈I(x̂)
and owing to û ≥ 0 and the convexity of f and gi , it follows from

Proposition 1.21 that for any arbitrary x ∈ B
− x̂)T ∇f (x̂)
f (x) − f (x̂) ≥ (x
=− ûi (x − x̂)T ∇gi (x̂)

i∈I(x̂)
≥− ûi [gi (x) − gi (x̂)]

i∈I(x̂)
≥0 ≤ 0 ∀x ∈ B, i ∈ I(x̂)
≥0
such that f (x) ≥ f (x̂) ∀x ∈ B. 2
Observe that
• to show the necessity of the Kuhn–Tucker conditions we had to use the

regularity condition RC 0 (or one of the other two, being stronger), but we
did not need any convexity assumption;
• to demonstrate that in the convex case the Kuhn–Tucker conditions are
sufficient for optimality we have indeed used the assumed convexity, but we
did not need any regularity condition at all.
Defining the Lagrange function for problem (8.1),

m
L(x, u) := f (x) + ui gi (x)
i=1
we may restate our optimality conditions. With the notation

∂L(x,u) T
∇x L(x, u) := ( ∂L(x,u)
∂x1 , · · · , ∂xn ) ,
T
∇u L(x, u) := ( ∂u1 , · · · , ∂L(x,u)
∂L(x,u)
∂um ) ,
BASIC CONCEPTS 89
and observing that ∇u L(x, u) ≤ 0 simply repeats the constraints gi (x) ≤ 0 ∀i

of our original program (8.1), the Kuhn–Tucker conditions now read as
⎫
∇x L(x̂, û) = 0, ⎪
⎪
⎬
∇u L(x̂, û) ≤ 0,
(8.10)
ûT ∇u L(x̂, û) = 0, ⎪
⎪
⎭
û ≥ 0.
Assume now that the functions f, gi , i = 1, · · · , m, are convex. Then for

any fixed u ≥ 0 the Lagrange function is obviously convex in x. For (x̂, û)
satisfying the Kuhn–Tucker conditions, it follows by Proposition 1.21 that for
any arbitrary x
L(x, û) − L(x̂, û) ≥ (x − x̂)T ∇x L(x̂, û) = 0
and hence
L(x̂, û) ≤ L(x, û) ∀x ∈ IRn .
On the other hand, since ∇u L(x̂, û) ≤ 0 is equivalent to
gmi (x̂) ≤ 0 ∀i, and
the Kuhn–Tucker conditions assert that ûT ∇u L(x̂, û) = i=1 ûi gi (x̂) = 0, it
follows that
L(x̂, u) ≤ L(x̂, û) ∀u ≥ 0.
Hence we have the following.
Proposition 1.25 Given that the functions f, gi , i = 1, · · · , m, in problem
(8.1) are convex, any Kuhn–Tucker point, i.e. any pair (x̂, û) satisfying the
Kuhn–Tucker conditions, is a saddle point of the Lagrange function, i.e. it
satisfies
∀u ≥ 0 L(x̂, u) ≤ L(x̂, û) ≤ L(x, û) ∀x ∈ IRn .
Furthermore, it follows by the complementarity conditions that
L(x̂, û) = f (x̂).
It is an easy exercise to show that for any saddle point (x̂, û), with û ≥ 0,
of the Lagrange function, the Kuhn–Tucker conditions (8.10) are satisfied.
Therefore, if we knew the right multiplier vector û in advance, the task to
solve the constrained optimization problem (8.1) would be equivalent to
that of solving the unconstrained optimization problem minx∈IRn L(x, û). This
observation can be seen as the basic motivation for the development of a class
of solution techniques known in the literature as Lagrangian methods.
1.8.2 Solution Techniques

When solving stochastic programs, we need to use known procedures from
both linear and nonlinear programming, or at least adopt their underlying
ideas. Unlike linear programs, nonlinear programs generally cannot be solved

in finitely many steps. Instead, we shall have to deal with iterative procedures
that we might expect to converge—in some reasonable sense—to a solution
of the nonlinear program under consideration. For better readability of
the subsequent chapters of this book, we sketch the basic ideas of some
types of methods; for detailed technical presentations and convergence proofs
the reader is referred to the extensive specialized literature on nonlinear
programming. We shall discuss
• cutting-plane methods;
• methods of descent;
• penalty methods;
• Lagrangian methods
by presenting one particular variant of each of these methods.
1.8.2.1 Cutting-plane methods

Assume that for problem (8.1) the functions f and gi , i = 1, · · · , m, are convex
and that the—convex—feasible set
B = {x | gi (x) ≤ 0, i = 1, · · · , m}
is bounded. Furthermore, assume that ∃ŷ ∈ int B—which for instance would
be true if the Slater condition (8.9) held. Then, instead of the original problem
min f (x),
x∈B
we could consider the equivalent problem

min θ
s.t. gi (x) ≤ 0, i = 1, · · · , m,
f (x) − θ ≤ 0,
with the feasible set B = {(x, θ) | f (x) − θ ≤ 0, gi (x) ≤ 0, i = 1, · · · , m} ⊂

IRn+1 being obviously convex. With the assumption ŷ ∈ int B, we may further
restrict the feasible solutions in B to satisfy the inequality θ ≤ f (ŷ) without
any effect on the solution set. The resulting problem can be interpreted
as the minimization of the linear objective θ on the bounded convex set
{(x, θ) ∈ B | θ ≤ f (ŷ)}, which is easily seen to contain an interior point
(ỹ, θ̃) as well.
Hence, instead of the nonlinear program (8.1), we may consider—without
loss of generality if the original feasible set B was bounded—the minimization
of a linear objective on a bounded convex set
min{cT x | x ∈ B}, (8.11)
BASIC CONCEPTS 91
where the bounded convex set B is assumed to contain an interior point ŷ.
Under the assumptions mentioned, it is possible to include the feasible set
B of problem (8.11) in a convex polyhedron P, which—after our discussions
in Section 1.7—we may expect to be able to represent by linear constraints.
Observe that the inclusion P ⊃ B implies the inequality
min cT x ≤ min cT x.
x∈P x∈B
The cutting-plane method for problem (8.11) proceeds as follows.

Step 1 Determine a ŷ ∈ int B and a convex polyhedron P0 ⊃ B; let k := 0.
Step 2 Solve the linear program
min{cT x | x ∈ Pk },
yielding the solution x̂k .
If x̂k ∈ B then stop (x̂k solves problem (8.11)); otherwise, i.e. if
x̂ ∈ B, determine
k
λk := min{λ | λŷ + (1 − λ)x̂k ∈ B}

and let
z k := λk ŷ + (1 − λk )x̂k .
(Obviously we have z k ∈ B and moreover z k is a boundary point of B
on the line segment between the interior point ŷ of B and the point
x̂k , which is “external” to B.)
Step 3 Determine a “supporting hyperplane” of B in z k (i.e. a hyperplane
being tangent to B at the boundary point z k ). Let this hyperplane be
given as
Hk := {x | (ak )T x = αk }
such that the inequalities
(ak )T x̂k > αk ≥ (ak )T x ∀x ∈ B
hold. Then define

Pk+1 := Pk {x | (ak )T x ≤ αk },
let k := k + 1, and return to step 2.

In Figure 30 we illustrate one step of the cutting-plane method.
Remark 1.11 By construction—see steps 1 and 3 of the above method—we

have
Pk ⊃ B, k = 0, 1, 2, · · · ,
Figure 30 Cutting-plane method: iteration k.
and hence
cT x̂k ≤ min cT x, k = 0, 1, 2, · · · ,
x∈B
such that as soon as x̂ ∈ B for some k ≥ 0 we would have that x̂k is an

k
optimal solution of problem (8.11), as claimed in step 2. Furthermore, since

z k ∈ B ∀k, we have
cT z k ≥ min cT x ∀k
x∈B
such that cT z k −cT x̂k could be taken after the kth iteration as an upper bound
on the distance of either the feasible (but in general nonoptimal) objective
value cT z k or the optimal (but in general nonfeasible) objective value cT x̂k to
the feasible optimal value minx∈B cT x. Observe that in general the sequence
{cT z k } need not be monotonically decreasing, whereas Pk+1 ⊂ Pk ∀k ensures
that the sequence {cT x̂k } is monotonically increasing. Thus we may enforce
a monotonically decreasing error bound
∆k := cT z lk − cT x̂k , k = 0, 1, 2, · · · ,
by choosing z lk from the boundary points of B constructed in step 2 up to

iteration k such that
cT z lk = min cT z l .
l∈{0,···,k}
Finally we describe briefly how the “supporting hyperplane” of B in z k of

step 3 can be determined. By our assumptions, ŷ ∈ int B and x̂k ∈ B, we
get in step 2 that 0 < λk < 1. Since λk > 0 is minimal under the condition
BASIC CONCEPTS 93
λŷ + (1 − λ)x̂k ∈ B, there is at least one constraint i0 active in z k meaning

that gi0 (z k ) = 0. The convexity of gi0 implies, owing to Proposition 1.21, that
0 > gi0 (ŷ) = gi0 (ŷ) − gi0 (z k ) ≥ (ŷ − z k )T ∇gi0 (z k ),
and therefore that ak := ∇gi0 (z k ) = 0.
Observing that zk = λk ŷ + (1 − λk )x̂k with 0 < λk < 1 is equivalent to
λk
x̂k − z k = − (ŷ − z k ),
1 − λk
we conclude from the last inequality that
(x̂k − z k )T ∇gi0 (z k ) > 0.
On the other hand, for any x ∈ B, gi0 (x) ≤ 0. Again by Proposition 1.21, it
follows that
(x − z k )T ∇gi0 (z k ) ≤ gi0 (x) − gi0 (z k ) = gi0 (x) ≤ 0.
Therefore, with ak := ∇gi0 (z k ) and αk := z kT ∇gi0 (z k ), we may define a sup-
porting hyperplane as required in step 3; this hyperplane is then used in the
definition of Pk+1 to cut off the set {x | (ak )T x > αk }—and hence in partic-
ular the infeasible solution x̂k —from further consideration. 2
1.8.2.2 Descent methods

For the sake of simplicity, we consider the special case of minimizing a convex
function under linear constraints
⎫
min f (x) ⎬
s.t. Ax = b, (8.12)
⎭
x ≥ 0.
Assume that we have a feasible point z ∈ B = {x | Ax = b, x ≥ 0}. Then
there are two possibilities.
(a) If z is optimal then the Kuhn–Tucker conditions have to hold. For (8.12)
these are
∇f (z) + AT u − w = 0,
z T w = 0,
w ≥ 0,
or—with J(z) := {j | zj > 0}—equivalently
AT u − w = −∇f (z),
wj = 0 for j ∈ J(z),
w ≥ 0.
Applying Farkas’ Lemma 1.19 tells us that this system (and hence the
above Kuhn–Tucker system) is feasible if and only if
[∇f (z)]T d ≥ 0 ∀d ∈ {d | Ad = 0, dj ≥ 0 for j ∈ J(z)};
(b) If the feasible point z is not optimal then the Kuhn–Tucker conditions
cannot hold, and, according to (a), there exists a direction d such that
Ad = 0, dj ≥ 0 ∀j : zj = 0 and [∇f (z)]T d < 0. A direction like
this is called a feasible descent direction at z, which has to satisfy the
following two conditions: ∃λ0 > 0 such that z + λd ∈ B ∀λ ∈ [0, λ0 ]
and [∇f (z)]T d < 0. Hence, having at a feasible point z a feasible descent
direction d (for which, by its definition, d = 0 is obvious), it is possible to
move from z in direction d with some positive step length without leaving
B and at the same time at least locally to decrease the objective’s value.
From these brief considerations, we may state the following.
Conceptual method of descent directions

Step 1 Determine a feasible solution z (0) , let k := 0.
Step 2 If there is no feasible descent direction at z (k) then stop (z (k) is
optimal).
Otherwise, choose a feasible descent direction d(k) at z (k) and go to
step 3.
Step 3 Solve the so-called line search problem
min{f (z (k) + λd(k) ) | (z (k) + λd(k) ) ∈ B},

λ
and with its solution λk define z (k+1) := z (k) + λk d(k) . Let k := k + 1

and return to step 2.
Remark 1.12 It is worth mentioning that not every choice of feasible descent
directions would lead to a well-behaved algorithm. By construction we should
get—in any case—a sequence of feasible points {z (k) } with a monotonically
(strictly) decreasing sequence {f (z (k) )} such that for the case that f is
bounded below on B the sequence {f (z (k) )} has to converge to some value
γ. However, there are examples in the literature showing that if we do not
restrict the choice of the feasible descent directions in an appropriate way, it
may happen that γ > inf B f (x), which is certainly not the kind of a result we
want to achieve.
Let us assume that B = ∅ is bounded, implying that our problem (8.12)
is solvable. Then there are various possibilities of determining the feasible
descent direction, each of which defines its own algorithm for which a
BASIC CONCEPTS 95
“reasonable” convergence behaviour can be asserted in the sense that the

sequence {f (z (k) )} converges to the true optimal value and any accumulation
point of the sequence {z (k) } is an optimal solution of our problem (8.12). Let
us just mention two of those algorithms:
(a) The feasible direction method For this algorithm we determine in step 2
the direction d(k) as the solution of the following linear program:
min[∇f (z (k) )]T d
s.t. Ad = 0
(k)
dj ≥ 0 ∀j : zj = 0,
d ≤ e,
d ≥ −e,
with e = (1, · · · , 1)T . Then for [∇f (z (k) )]T d(k) < 0 we have a feasible
descent direction, whereas for [∇f (z (k) )]T d(k) = 0 the point z (k) is an
optimal solution of (8.12).
(b) The reduced gradient method Assume that B is bounded and every feasible
basic solution of (8.12) is nondegenerate. Then for z (k) we find a basis B
in A such that the components of z (k) belonging to B are strictly positive.
Rewriting A—after the necessary rearrangements of columns—as (B, N )
and correspondingly presenting z (k) as (xB , xN B ), we have
BxB + N xN B = b,
or equivalently
xB = B −1 b − B −1 N xN B .
We also may rewrite the gradient ∇f (z (k) ) as (∇B f (z (k) ), ∇N B f (z (k) )).
Then, rearranging d accordingly into (u, v), for a feasible direction we
need to have
Bu + N v = 0,
and hence
u = −B −1 N v.
For the directional derivative [∇f (z (k) )]T d it follows
[∇f (z (k) )]T d = [∇B f (z (k) )]T u + [∇N B f (z (k) )]T v
= [∇B f (z (k) )]T (−B −1 N v) + [∇N B f (z (k) )]T v
= ([∇N B f (z (k) )]T − [∇B f (z (k) )]T B −1 N )v.
Defining the reduced gradient r by
B T
T r
r =
rN B

= [∇B f (z (k) )]T , [∇N B f (z (k) )]T − [∇B f (z (k) )]T B −1 (B, N ),
we have
rB = 0, rN B = ([∇N B f (z (k) )]T − [∇B f (z (k) )]T B −1 N )T ,
and hence

(k) T T T rB
[∇f (z )] d = (u , v )
rN B
= ([∇N B f (z (k) )]T − [∇B f (z (k) )]T B −1 N )v.
Defining v as

−rjN B if rjN B ≤ 0,
vj :=
−xN
j
B NB
rj if rjN B > 0,
and, as above, u := B −1 N v, we have that (uT , v T )T is a feasible direction

(observe that vj ≥ 0 if xN j
B
= 0 and xB > 0 owing to our assumption)
and it is a descent direction if v = 0. Furthermore, z (k) is a solution of
problem (8.12) iff v = 0 (and hence u = 0), since then r ≥ 0, and, with
wT := [∇B f (z (k) )]T B −1 , we have
rB = ∇B f (z (k) ) − B T w = 0,
rNB
= ∇N B f (z (k) ) − N T w ≥ 0,
and
(rB )T xB = 0, (rN B )T xN B = 0,
i.e. v = 0 is equivalent to satisfying the Kuhn–Tucker conditions.
It is known that the reduced gradient method with the above definition
of v may fail to converge to a solution (so-called “zigzagging”). However,
we can perturb v as follows:
⎧
⎨ −rjN B if rjN B ≤ 0,
vj := −xj rj
NB NB
if rjN B > 0 and xN B
≥ ε,
⎩ j
0 if rjN B > 0 and xN
j
B
< ε.
Then a proper control of the perturbation ε > 0 during the procedure can
be shown to enforce convergence.
The feasible direction and the reduced gradient methods have been extended
to the case of nonlinear constraints. We omit the presentation of the general
case here for the sake of better readability.
BASIC CONCEPTS 97
1.8.2.3 Penalty methods

The term “penalty” reflects the following attempt. Replace the original
problem (8.1)
min f (x)
s.t. gi (x) ≤ 0, i = 1, · · · , m,
by appropriate free (i.e. unconstrained) optimization problems
⎫
minx∈IRn Frs (x), ⎪
⎪
⎪
⎬
the function Frs being defined as
1 (8.13)
Frs (x) := f (x) + r ϕ(gi (x)) + ψ(gi (x)), ⎪
⎪
⎪
⎭
s
i∈I i∈J
where I, J ⊂ {1, · · · , m} such that I ∩ J = ∅, I ∪ J = {1, · · · , m}, and the

parameters r, s > 0 are to be chosen or adapted in the course of the procedure.
The role of the functions ϕ and ψ is to inhibit and to penalize respectively
the violation of any one of the constraints. More precisely, for these functions
we assume that
• ϕ, ψ are monotonically increasing and convex;
• the so-called barrier function satisfies
ϕ(η) < +∞ ∀η < 0,

lim ϕ(η) = +∞;
η↑0
• for the so-called loss function we have

= 0 ∀η ≤ 0,
ψ(η)
> 0 ∀η > 0.
Observe that the convexity of f, gi , i = 1, · · · , m, and the convexity and

monotonicity of ϕ, ψ imply the convexity of Frs for any choice of the
parameters r, s > 0. Solving the free optimization problem (8.13) with
parameters r, s > 0 would inhibit the violation of the constraints i ∈ I,
whereas the violation of anyone of the constraints i ∈ J would be penalized
with a positive additive term. Intuitively, we might expect that the solutions
of (8.13) will satisfy the constraints i ∈ I and that for
1
s ↓ 0, or equivalently ↑ +∞,
s
they will eventually satisfy the constraints i ∈ J. Therefore it seems plausible
to control the parameter s in such a way that it tends to zero. Now what
about the parameter r of the barrier term in (8.13)? Imagine that for the
(presumably unique) solution x̂ of problem (8.1) some constraint i0 ∈ I is

active, i.e. gi0 (x̂) = 0. For any fixed r > 0, minimization of (8.13) will not
allow us to approach the solution x̂, since obviously, by the definition of a
barrier function, this would drive the new objective Frs to +∞. Hence it
seems reasonable to drive the parameter r downwards to zero, as well.
With
B1 := {x | gi (x) ≤ 0, i ∈ I}, B2 := {x | gi (x) ≤ 0, i ∈ J}
we have B = B1 ∩ B2 , and for r > 0 we may expect finite values of Frs

only for x ∈ B10 := {x | gi (x) < 0, i ∈ I}. We may close this short
presentation of general penalty methods by a statement showing that, under
mild assumptions, a method of this type may be controlled in such a way that
it results in what we should like to experience.
Proposition 1.26 Let f, gi , i = 1, · · · , m, be convex and assume that
B10 ∩ B2 = ∅
and that B = B1 ∩ B2 is bounded. Then for {rk } and {sk } strictly monotone
sequences decreasing to zero there exists an index k0 such that for all k ≥ k0
the modified objective function Frk sk attains its (free) minimum at some point
x(k) where x(k) ∈ B10 .
The sequence {x(k) | k ≥ k0 } is bounded, and any of its accumulation points
is a solution of the original problem (8.1). With γ the optimal value of (8.1),
the following relations hold:
lim f (x(k) ) = γ,
k→∞
lim rk ϕ(gi (x(k) )) = 0,
k→∞
i∈I
1
lim ψ(gi (x(k) )) = 0.
k→∞ sk
i∈J
1.8.2.4 Lagrangian methods

As mentioned at the end of Section 1.8.1, knowledge of the proper multiplier
m
vector û in the Lagrange function L(x, u) = f (x) + i=1 ui gi (x) for
problem (8.1) would allow us to solve the free optimization problem
min L(x, û)

x∈IRn
instead of the constrained problem
min f (x)
s.t. gi (x) ≤ 0, i = 1, · · · , m.
BASIC CONCEPTS 99
To simplify the description, let us first consider the optimization problem with
equality constraints

min f (x)
(8.14)
s.t. gi (x) = 0, i = 1, · · · , m.
Knowing for this problem the proper multiplier vector û or at least a good
approximate u of it, we should find
min [f (x) + uT g(x)], (8.15)

x∈IRn
m
where uT g(x) = i=1 ui gi (x). However, at the beginning of any solution
procedure we hardly have any knowledge about the numerical size of the
multipliers in a Kuhn–Tucker point of problem (8.14), and using some guess
for u might easily result in an unsolvable problem (inf x L(x, u) = −∞).
On the other hand, we have just introduced penalty methods. Using for
problem (8.14) a quadratic loss function for violating the equality constraints
seems to be reasonable. Hence we could think of a penalty method using as
modified objective
minn [f (x) + 12 λg(x)2 ] (8.16)
x∈IR
and driving the parameter λ towards +∞, with g(x) being the Euclidean
norm of g(x) = (g1 (x), · · · , gm (x))T .
One idea is to combine the two approaches (8.15) and (8.16) such that we are
dealing with the so-called augmented Lagrangian as our modified objective:
min [f (x) + uT g(x) + 12 λg(x)2 ].

x∈IRn
The now obvious intention is to control the parameters u and λ in such a

way that λ → ∞—to eliminate infeasibilities—and that at the same time
u → û, the proper Kuhn–Tucker multiplier vector. Although we are not yet
in a position to appropriately adjust the parameters, we know at least the
skeleton of the algorithm, which usually is referred to as augmented Lagrange
method: With the augmented Lagrangian
Lλ (x, u) := f (x) + uT g(x) + 12 λg(x)2 ,
it may be loosely stated as follows:

⎫
For ⎪
⎪
⎬
• {u(k) } ⊂ IRm bounded,
(8.17)
• {λk } ⊂ IR such that 0 < λk < λk+1 ∀k, λk → ∞, ⎪
⎪
⎭
solve successively minx∈IRn Lλk (x, u(k) ).
Observe that for u(k) = 0 ∀k we should get back the penalty method with
a quadratic loss function, which, according to Proposition 1.26, is known to
“converge” in the sense asserted there.
For the method (8.17) in general the following two statements can be proved,
showing
(a) that we may expect a convergence behaviour as we know it already for
penalty methods; and
(b) how we should successively adjust the multiplier vector u(k) to get the
intended convergence to the proper Kuhn–Tucker multipliers.
Proposition 1.27 If f and gi , i = 1, · · · , m, are continuous and x(k) , k =

1, 2, · · · , are global solutions of
min Lλk (x, u(k) )

x
then any accumulation point x̄ of {x(k) } is a global solution of problem (8.14).
The following statement also shows that it would be sufficient to solve the
free optimization problems minx Lλk (x, u(k) ) only approximately.
Proposition 1.28 Let f and gi , i = 1, · · · , m, be continuously differentiable,
and let the approximate solutions x(k) to the free minimization problems
in (8.17) satisfy
∇x Lλk (x(k) , u(k) ) ≤ k ∀k,
where k ≥ 0 ∀k and k → 0. For some K ⊂ IN let {x(k) , k ∈ K} converge
to some x (i.e. x is an accumulation point of {x(k) , k ∈ IN}), and let
{∇g1 (x ), · · · , ∇gm (x )} be linearly independent. Then ∃u such that
{u(k) + λk g(x(k) ), k ∈ K} −→ u , (8.18)

m
∇f (x ) + ui ∇gi (x ) = 0, g(x ) = 0.
i=1
Choosing the parameters λk according to (8.17), for instance as
λ1 := 1, λk+1 := 1.1λk ∀k ≥ 1,
the above statement suggests, by (8.18), that
u(k+1) := u(k) + λk g(x(k) ) (8.19)
is an appropriate update formula for the multipliers in order to eventually

get—together with x —a Kuhn–Tucker point.
BASIC CONCEPTS 101
Now let us come back to our original nonlinear program (8.1) with inequality
constraints and show how we can make use of the above results for the case
of equality constraints. The key to this is the observation that our problem
with inequality constraints
min f (x)
s.t. gi (x) ≤ 0, i = 1, · · · , m
is equivalent to the following one with equality constraints:
min f (x)
s.t. gi (x) + zi2 = 0, i = 1, · · · , m.
Now applying the augmented Lagrangian method (8.17) to this equality-

constrained problem requires that for

m
# $ % $ %2 &
Lλ (x, z, u) := f (x) + ui gi (x) + zi2 + 12 λ gi (x) + zi2 (8.20)
i=1
we solve successively the problem
min Lλk (x, z, u(k) ).

x, z
The minimization with respect to z included in this problem may be carried

through explicitly, observing that
min Lλk (x, z, u(k) )

z∈IRm
!
m
# (k) $ % $ %2 & "
= minm f (x) + ui gi (x) + zi2 + 12 λk gi (x) + zi2
z∈IR
i=1

m
# (k) $ % $ %2 &
= f (x) + min ui gi (x) + zi2 + 12 λk gi (x) + zi2 .
zi
i=1
Therefore the minimization of L with respect to z requires—with yi := zi2 —

the solution of m problems of the form
# $ % $ %2 &
min ui gi (x) + yi + 12 λ gi (x) + yi , (8.21)
yi ≥0
i.e. the minimization of strictly convex (λ > 0) quadratic functions in yi on

yi ≥ 0. The free minima (i.e. yi ∈ IR) of (8.21) have to satisfy
$ %
ui + λ gi (x) + yi = 0,
yielding 'u (
i
ỹi = − + gi (x) .
λ
Hence we have for the solution of (8.21)

⎫
ỹi if ỹi ≥ 0, ⎪
⎬
yi =
0 ! otherwise,
'u (" (8.22)
i ⎪
⎭
= max 0, − + gi (x)
λ
implying ' ui (
gi (x) + yi = max gi (x), − (8.23)
λ
which, with ẑi2 = yi , after an elementary algebraic manipulation reduces our
extended Lagrangian (8.20) to
L̃λ (x, u) = Lλ (x, ẑ, u)

1 !# &2 "
m
= f (x) + max[0, ui + λgi (x)] − u2i .
2λ i=1
Minimization for some given uk and λk of the Lagrangian (8.20) with respect
to x and z will now be achieved by solving the problem
min L̃λk (x, u(k) ),

x
and, with a solution x(k) of this problem, our update formula (8.19) for the
multipliers—recalling that we now have the equality constraints gi (x)+zi2 = 0
instead of gi (x) = 0 as before—becomes by (8.23)
$ u(k) %
u(k+1) := u(k) + λk “max” g(x(k) ), −
λ (8.24)
$ % k
= “max” 0, u(k) + λk g(x(k) ) ,
where “max” is to be understood componentwise.
1.9 Bibliographical Notes
The observation that some data in real life optimization problems could be
random, i.e. the origin of stochastic programming, dates back to the 1950s.
Without any attempt at completeness, we might mention from the early
contributions to this field Avriel and Williams [3], Beale [5, 6], Bereanu [8],
Dantzig [11], Dantzig and Madansky [13], Tintner [43] and Williams [49].
For more detailed discussions of the situation of the decision maker facing
random parameters in an optimization problem we refer for instance to
Dempster [14], Ermoliev and Wets [16], Frauendorfer [18], Kall [22], Kall and
Prékopa [24], Kolbin [28], Sengupta [42] and Vajda [45].
BASIC CONCEPTS 103
Wait-and-see problems have led to investigations of the distribution of the

optimal value (and the optimal solution); as examples of these efforts, we
mention Bereanu [8] and King [26].
The linear programs resulting as deterministic equivalents in the recourse
case may become (very) large in scale, but their particular block structure
is amenable to specially designed algorithms, which are until now under
investigation and for which further progress is to be expected in view of the
possibilities given with parallel computers (see e.g. Zenios [51]). For those
problems the particular decomposition method QDECOM—which will be
described later—was proposed by Ruszczyński [41].
The idea of approximating stochastic programs with recourse (with a
continuous type distribution) by discretizing the distribution, as mentioned in
Section 1.3, is related to special convergence requirements for the (discretized)
expected recourse functions, as discussed for example by Attouch and Wets [2]
and Kall [23].
More on probabilistically constrained models and corresponding applica-
tions may be found for example in Dupačová et al. [15], Ermoliev and Wets [16]
and Prékopa et al. [36]. The convexity statement of Proposition 1.5 can be
found in Wets [48]. The probalistically constrained program at the end of Sec-
tion 1.3 (page 20) was solved by PROCON. This solution method for problems
with a joint chance constraint (with normally distributed right-hand side) was
described first by Mayer [30], and has its theoretical base in Prékopa [35].
Statements on the induced feasible set K and induced constraints are
found in Rockafellar and Wets [39], Walkup and Wets [46] and Kall [22].
The requirement that the decision on x does not depend on the outcome of
ξ˜ is denoted as nonanticipativity, and was discussed rigorously in Rockafellar
and Wets [40]. The conditions for complete recourse matrices were proved in
Kall [21], and may be found in [22].
Necessary and sufficient conditions for log-concave distributions were
derived first in Prékopa [35]; later corresponding conditions for quasi-concave
measures were derived in Borell [10] and Rinott [37].
More details on stochastic linear programs may be found in Kall [22];
multistage stochastic programs are still under investigation, and were
discussed early by Olsen [31, 32, 33]; useful results on the deterministic
equivalent of recourse problems and for the expectation functionals arising
in Section 1.4 are due to Wets [47, 48].
There is a wide literature on linear programming, which cannot be listed
here to any reasonable degree of completeness. Hence we restrict ourselves to
mentioning the book of Dantzig [12] as a classic reference.
For a rigorous development of measure theory and the foundations of
probability theory we mention the standard reference Halmos [19].
The idea of feasibility and optimality cuts in the dual decomposition method
may be traced back to Benders [7].
There is a great variety of good textbooks on nonlinear programming

(theory and methods) as well. Again we have to restrict ourselves, and just
mention Bazaraa and Shetty [4] and Luenberger [29] as general texts.
Cutting-plane methods have been proposed in various publications, differing
in the way the cuts (separating hyperplanes) are defined. An early version
was published by Kelley [25]; the method we have presented is due to
Kleibohm [27].
The method of feasible directions is due to Zoutendijk [52, 53]; an extension
to nonlinear constraints was proposed by Topkis and Veinott [44].
The reduced gradient method can be found in Wolfe [50], and its extension
to nonlinear constraints was developed by Abadie and Carpentier [1].
A standard reference for penalty methods is the monograph of Fiacco and
McCormick [17].
The update formula (8.19) for the multipliers in the augmented Lagrangian
method for equality constraints motivated by Proposition 1.28 goes back
to Hestenes [20] and Powell [34], whereas the update (8.24) for inequality-
constrained problems is due to Rockafellar [38]. For more about Lagrangian
methods we refer the reader to the book of Bertsekas [9].
Exercises
1. Show that from (4.3) on page 24 it follows that with Ai ∈ A, i = 1, 2, · · · ,

∞

Ai ∈ A and Ai − Aj ∈ A ∀i, j.
i=1
2. Find an example of a two-dimensional discrete probability distribution that

is not quasi-concave.
3. Show that A = {(x, y) ∈ IR2 | x ≥ 1, 0 ≤ y ≤ 1/x} is measurable with
respect to the natural measure µ in IR2 and that µ(A) = ∞ (see
Section 1.4.1, page 21). [Hint: Show first that for
In := {(x, y) | n ≤ x ≤ n + 1, 0 ≤ y < 2}, n ∈ IN,
the sets A ∩ In are measurable. For n ∈ IN the interval
Cn := {(x, y) | n ≤ x < n + 1, 0 ≤ y < 1/(n + 1)}

∞
∞ of A ∩ In with µ(Cn ) = 1/(n + 1). Hence µ(A) = n=1 µ(A ∩
is a packing
In ) ≥ n=1 µ(Cn ) implies µ(A) = ∞.]
4. Show that A := {(x, y) ∈ IR2 | x ≥ 0, 0 ≤ y ≤ e−x } is measurable
and that µ(A) = 1. [Hint: Consider Aα := {(x, y) | 0 ≤ x ≤ α, 0 ≤
BASIC CONCEPTS 105
y ≤ e−x } for arbitrary α > 0. Observe that µ(Aα ), according to its

definition
αin Section 1.4.1, page 21, coincides with the Riemann integral
J(α) = 0 e−x dx. Hence µ(A) = limα→∞ µ(Aα ) = limα→∞ J(α).]
5. Consider the line segment B := {(x, y) ∈ IR2 | 3 ≤ x ≤ 7, y = 5}. Show
that for the natural measure µ in IR2 , µ(B) = 0 (see Section 1.4.1, page 21).
6. Assume that the linear program γ(b) := min{cT x | Ax = b, x ≥ 0} is
solvable for all b ∈ IRm . Show that the optimal value function γ(·) is
piecewise linear and convex in b.
7. In Section 1.8 we discussed various regularity conditions for nonlinear
programs. Let x̂ be a local solution of problem (8.1), page 80. Show that if
RC 1 is satisfied in x̂ then RC 0 also holds true in x̂. (See (8.8) on page 85.)
8. Assume that (x̂, û) is a saddle point of

m
L(x, u) := f (x) + ui gi (x).
i=1
Show that x̂ is a global solution of

min f (x)
s.t. gi (x) ≤ 0, i = 1, · · · , m.
(See Proposition 1.25 for the definition of a saddle point.)
References
[1] Abadie J. and Carpentier J. (1969) Generalization of the Wolfe reduced

gradient method to the case of nonlinear constraints. In Fletcher R. (ed)
Optimization, pages 37–47. Academic Press, London.
[2] Attouch H. and Wets R. J.-B. (1981) Approximation and convergence
in nonlinear optimization. In Mangasarian O. L., Meyer R. M., and
Robinson S. M. (eds) NLP 4, pages 367–394. Academic Press, New York.
[3] Avriel M. and Williams A. (1970) The value of information and stochastic
programming. Oper. Res. 18: 947–954.
[4] Bazaraa M. S. and Shetty C. M. (1979) Nonlinear Programming—Theory
and Algorithms. John Wiley & Sons, New York.
[5] Beale E. M. L. (1955) On minimizing a convex function subject to linear
inequalities. J. R. Stat. Soc. B17: 173–184.
[6] Beale E. M. L. (1961) The use of quadratic programming in stochastic
linear programming. Rand Report P-2404, The RAND Corporation.
[7] Benders J. F. (1962) Partitioning procedures for solving mixed-variables
programming problems. Numer. Math. 4: 238–252.
[8] Bereanu B. (1967) On stochastic linear programming distribution

problems, stochastic technology matrix. Z. Wahrsch. theorie u. verw.
Geb. 8: 148–152.
[9] Bertsekas D. P. (1982) Constrained Optimization and Lagrange Multiplier
Methods. Academic Press, New York.
[10] Borell C. (1975) Convex set functions in d-space. Period. Math. Hungar.
6: 111–136.
[11] Dantzig G. B. (1955) Linear programming under uncertainty. Manage-
ment Sci. 1: 197–206.
[12] Dantzig G. B. (1963) Linear Programming and Extensions. Princeton
University Press, Princeton, New Jersey.
[13] Dantzig G. B. and Madansky A. (1961) On the solution of two-stage
linear programs under uncertainty. In Neyman I. J. (ed) Proc. 4th
Berkeley Symp. Math. Stat. Prob., pages 165–176. Berkeley.
[14] Dempster M. A. H. (ed) (1980) Stochastic Programming. Academic Press,
London.
[15] Dupačová J., Gaivoronski A., Kos Z., and Szántai T. (1991) Stochastic
programming in water resources system planning: A case study and a
comparison of solution techniques. Eur. J. Oper. Res. 52: 28–44.
[16] Ermoliev Y. and Wets R. J.-B. (eds) (1988) Numerical Techniques for
Stochastic Optimization. Springer-Verlag, Berlin.
[17] Fiacco A. V. and McCormick G. P. (1968) Nonlinear Programming:
Sequential Unconstrained Minimization Techniques. John Wiley & Sons,
New York.
[18] Frauendorfer K. (1992) Stochastic Two-Stage Programming, volume 392
of Lecture Notes in Econ. Math. Syst. Springer-Verlag, Berlin.
[19] Halmos P. R. (1950) Measure Theory. D. van Nostrand, Princeton, New
Jersey.
[20] Hestenes M. R. (1969) Multiplier and gradient methods. In Zadeh L. A.,
Neustadt L. W., and Balakrishnan A. V. (eds) Computing Methods in
Optimization Problems—2, pages 143–163. Academic Press, New York.
[21] Kall P. (1966) Qualitative Aussagen zu einigen Problemen der
stochastischen Programmierung. Z. Wahrsch. theorie u. verw. Geb. 6:
246–272.
[22] Kall P. (1976) Stochastic Linear Programming. Springer-Verlag, Berlin.
[23] Kall P. (1986) Approximation to optimization problems: An elementary
review. Math. Oper. Res. 11: 9–18.
[24] Kall P. and Prékopa A. (eds) (1980) Recent Results in Stochastic
Programming, volume 179 of Lecture Notes in Econ. Math. Syst. Springer-
Verlag, Berlin.
[25] Kelley J. E. (1960) The cutting plane method for solving convex
programs. SIAM J. Appl. Math. 11: 703–712.
[26] King A. J. (1986) Asymptotic Behaviour of Solutions in Stochastic
BASIC CONCEPTS 107
Optimization: Nonsmooth Analysis and the Derivation of Non-Normal

Limit Distributions. PhD thesis, University of Washington, Seattle.
[27] Kleibohm K. (1966) Ein Verfahren zur approximativen Lösung von
konvexen Programmen. PhD thesis, Universität Zürich. Mentioned in
C.R. Acad. Sci. Paris 261:306–307 (1965).
[28] Kolbin V. V. (1977) Stochastic Programming. D. Reidel, Dordrecht.
[29] Luenberger D. G. (1973) Introduction to Linear and Nonlinear
Programming. Addison-Wesley, Reading, Massachusetts.
[30] Mayer J. (1988) Probabilistic constrained programming: A reduced
gradient algorithm implemented on pc. Working Paper WP-88-39, IIASA,
Laxenburg.
[31] Olsen P. (1976) Multistage stochastic programming with recourse: The
equivalent deterministic problem. SIAM J. Contr. Opt. 14: 495–517.
[32] Olsen P. (1976) When is a multistage stochastic programming problem
well defined? SIAM J. Contr. Opt. 14: 518–527.
[33] Olsen P. (1976) Discretizations of multistage stochastic programming
problems. Math. Prog. Study 6: 111–124.
[34] Powell M. J. D. (1969) A method for nonlinear constraints in
minimization problems. In Fletcher R. (ed) Optimization, pages 283–298.
Academic Press, London.
[35] Prékopa A. (1971) Logarithmic concave measures with applications to
stochastic programming. Acta Sci. Math. (Szeged) 32: 301–316.
[36] Prékopa A., Ganczer S., Deák I., and Patyi K. (1980) The STABIL
stochastic programming model and its experimental application to the
electricity production in Hungary. In Dempster M. A. H. (ed) Stochastic
Programming, pages 369–385. Academic Press, London.
[37] Rinott Y. (1976) On convexity of measures. Ann. Prob. 4: 1020–1026.
[38] Rockafellar R. T. (1973) The multiplier method of Hestenes and Powell
applied to convex programming. J. Opt. Theory Appl. 12: 555–562.
[39] Rockafellar R. T. and Wets R. J.-B. (1976) Stochastic convex
programming: Relatively complete recourse and induced feasibility. SIAM
J. Contr. Opt. 14: 574–589.
[40] Rockafellar R. T. and Wets R. J.-B. (1976) Nonanticipativity and L1 -
martingales in stochastic optimization problems. Math. Prog. Study 6:
170–187.
[41] Ruszczyński A. (1986) A regularized decomposition method for
minimizing a sum of polyhedral functions. Math. Prog. 35: 309–333.
[42] Sengupta J. K. (1972) Stochastic programming. Methods and applications.
North-Holland, Amsterdam.
[43] Tintner G. (1955) Stochastic linear programming with applications to
agricultural economics. In Antosiewicz H. (ed) Proc. 2nd Symp. Linear
Programming, volume 2, pages 197–228. National Bureau of Standards,
Washington D.C.
[44] Topkis D. M. and Veinott A. F. (1967) On the convergence of some

feasible direction algorithms for nonlinear programming. SIAM J. Contr.
Opt. 5: 268–279.
[45] Vajda S. (1972) Probabilistic Programming. Academic Press, New York.
[46] Walkup D. W. and Wets R. J. B. (1967) Stochastic programs with
recourse. SIAM J. Appl. Math. 15: 1299–1314.
[47] Wets R. (1974) Stochastic programs with fixed recourse: The equivalent
deterministic program. SIAM Rev. 16: 309–339.
[48] Wets R. J.-B. (1989) Stochastic programming. In Nemhauser G. L.
et al. (eds) Handbooks in OR & MS, volume 1, pages 573–629. Elsevier,
Amsterdam.
[49] Williams A. (1966) Approximation formulas for stochastic linear
programming. SIAM J. Appl. Math. 14: 668–877.
[50] Wolfe P. (1963) Methods of nonlinear programming. In Graves R. L. and
Wolfe P. (eds) Recent Advances in Mathematical Programming, pages
67–86. McGraw-Hill, New York.
[51] Zenios S. A. (1992) Progress on the massively parallel solution of network
“mega”-problems. COAL 20: 13–19.
[52] Zoutendijk G. (1960) Methods of Feasible Directions. Elsevier, Amster-
dam/D. Van Nostrand, Princeton, New Jersey.
[53] Zoutendijk G. (1966) Nonlinear programming: A numerical survey. SIAM
J. Contr. Opt. 4: 194–210.
2
Dynamic Systems
2.1 The Bellman Principle
As discussed in Chapter 1, optimization problems can be of various types. The

differences may be found in the goal, i.e. minimization or maximization, in
the constraints, i.e. inequalities or equalities and free or nonnegative variables,
and in the mathematical properties of the functions involved in the objective
or the constraints. We have met linear functions in Section 1.7, nonlinear
functions in Section 1.8 and even integral functions in Section 1.4. Despite
their differences, all these problems may be presented in the unified form
max{F (x1 , · · · , xn ) | x ∈ X}.
Here X is the prescribed feasible set of decisions over which we try to

maximize, or sometimes minimize, the given objective function F .
This general setting also covers a class of somewhat special decision
problems. They are illustrated in Figure 1.
Consider a system that is inspected at finitely many stages. Often stages are
just points in time—the reason for using the term “dynamic”. The example in
Figure 1 has four stages. This is seen from the fact that there are four columns.
Assume that at any stage the system can be in one out of finitely many states.
In Figure 1 there are four possible states in each stage, represented by the four
dots in each column. Also, at any stage (except maybe the last one) a decision
has to be made which possibly will have an influence on the system’s state at
the subsequent stage. Attached to the decision is an immediate return (or else
an immediate cost). In Figure 1 the three arrows in the right part of the figure
indicate that in this example there are three possible decisions: one bringing
us to a lower state in the next stage, one keeping us in the same state, and one
bringing us to a higher state. (We must assume that if we are at the highest
or lowest possible state then only two decisions are possible). Given the initial
state of the system, the overall objective is to maximize (or minimize) some
given function of the immediate returns for all stages and states the system
goes through as a result of our decisions. Formally the problem is described
DYNAMIC SYSTEMS 111
Figure 1 Basic set-up for a dynamic program with four states, four stages
and three possible decisions.
as follows. With
t the stages, t = 1, · · · , T ,
zt the state at stage t,
xt the decision taken at stage t, (in general depending on the state
zt ),
Gt (zt , xt ) the transformation (or transition) of the system from the state zt
and the decision taken at stage t into the state zt+1 at the next
stage, i.e. zt+1 = Gt (zt , xt ),
rt (zt , xt ) the immediate return if at stage t the system is in state zt and the
decision xt is taken,
F the overall objective, given by F (r1 (z1 , x1 ), · · · , rT (zT , xT )), and
Xt (zt ) the set of feasible decisions at stage t, (which may depend on the
state zt ),
our problem can be stated as
max{F (r1 (z1 , x1 ), · · · , rT (zT , xT )) | xt ∈ Xt , t = 1, · · · , T }.
Observe that owing to the relation zt+1 = Gt (zt , xt ), the objective function
can be rewritten in the form Φ(z1 , x1 , x2 , · · · , xT ). To get an idea of the
possible structures we can face, let us revisit the example in Figure 1. The
purpose of the example is not to be realistic, but to illustrate a few points. A

more realistic problem will be discussed in the next section.
Example 2.1 Assume that stages are years, and that the system is inspected
annually, so that the three stages correspond to 1 January of the first, second
and third years, and 31 December of the third year. Assume further that four
different levels are distinguished as states for the system, i.e. at any stage one
may observe the state zt = 1, 2, 3 or 4. Finally, depending on the state of the
system in stages 1, 2 and 3, one of the following decisions may be made:
⎧
⎨ 1, leading to the immediate return rt = 2,
xt = 0, leading to the immediate return rt = 1,
⎩
−1, leading to the immediate return rt = −1.
The transition from one stage to the next is given by
zt+1 = zt + xt .
Note that, since zt ∈ {1, 2, 3, 4} for all t, we have that xt = 1 in state zt = 4
and xt = −1 in state zt = 1 are not feasible, and are therefore excluded.
Finally, assume that there are no decisions in the final stage T = 4. There are
immediate returns, however, given as
⎧
⎪ −2 if zT = 4,
⎪
⎨
−1 if zT = 3,
rT =
⎪
⎪ 1 if zT = 2,
⎩
2 if zT = 1.
To solve max F (r1 , · · · , r4 ), we have to fix the overall objective F as a
function of the immediate returns r1 , r2 , r3 , r4 . To demonstrate possible effects
of properties of F on the solution procedure, we choose two variants.
(a) Let
F (r1 , · · · , r4 ) := r1 + r2 + r3 + r4
and assume that the initial state is z1 = 4. This is illustrated in Figure 2,
which has the same structure as Figure 1. Using the figure, we can check
that an optimal policy (i.e. sequence of decisions), is x1 = x2 = x3 = 0
keeping us in zt = 4 for all t with the optimal value F (r1 , · · · , r4 ) =
1 + 1 + 1 − 2 = 1.
We may determine this optimal policy iteratively as follows. First, we
determine the decision for each of the states in stage 3 by determining
f3∗ (z3 ) := max[r3 (z3 , x3 ) + r4 (z4 )]
x3
for z3 = 1, · · · , 4, and z4 = G3 (z3 , x3 ) = z3 + x3 . For example, if we are

in state 2, i.e. z3 = 2, we have three options, namely −1, 0 and 1. If
DYNAMIC SYSTEMS 113
Figure 2 Dynamic program: additive composition. The solid lines show the
result of the backward recursion.
x3 = 1, we receive an immediate income of 2, and a final value of −1,

since this decision will result in z4 = 2 + 1 = 3. The second option is to
let x3 = 0, yielding an immediate income of 1 and a final value of 1. The
third possibility is to let x3 = −1, yielding incomes of −1 and 1. The total
incomes are therefore 1, 2 and 0 respectively, so the best option is to let
x3 = 0. This is illustrated in the figure by putting an arrow from state 2
in stage 3 to state 2 in stage 4. Letting “(z3 = i) → (x3 , f3∗ )” indicate that
in state z3 = i the optimal decision is x3 and the sum of the immediate
and final income is f3∗ , we can repeat the above procedure for each state
in stage 3 to obtain (z3 = 1) → (0, 3), (z3 = 2) → (0, 2), (z3 = 3) →
(0, 0), (z3 = 4) → (0, −1). This is all illustrated in Figure 2, by adding
the f3∗ -values above the state nodes in stage 3.
Once f3∗ (z3 ) is known for all values of z3 , we can turn to stage 2 and
similarly determine
f2∗ (z2 ) := max[r2 (z2 , x2 ) + f3∗ (z3 )],

x2
where z3 = G2 (z2 , x2 ) = z2 + x2 . This yields f2∗ (1) = 4, f2∗ (2) =

3, f2∗ (3) = 1 and f2∗ (4) = 0. This is again illustrated in Figure 2, together
with the corresponding optimal decisions.
Finally, given that z1 = 4, the problem can be rephrased as
f1∗ (z1 ) := max[r1 (z1 , x1 ) + f2∗ (z2 )],

x1
where z2 = G1 (z1 , x1 ) = z1 + x1 . This immediately yields f1∗ (z1 ) = 1 for

x1 = 0.
In this simple example it is easy to see that f1∗ (z1 ) coincides with
the optimal value of F (r1 , · · · , r4 ), given the initial state z1 , so that the
problem can be solved with the above backward recursion. (The recursion
is called backward because we start in the last period and move backwards
in time, ending up in period 1.)
Note that an alternative way to solve this problem would be to
enumerate all possible sequences of decisions. For this small problem, that
would have been a rather simple task. But for larger problems, both in
terms of states and stages (especially when both are multi-dimensional),
it is easy to see that this will become an impossible task. This reduction
from a full enumeration of all possible sequences of decisions to that of
finding the optimal decision in all states for all stages is the major reason
for being interested in the backward recursion, and more generally, for
being interested in dynamic programming.
(b) As an alternative, let us use multiplication to obtain
F (r1 , · · · , r4 ) := r1 r2 r3 r4
and perform the backward recursion as above, yielding Figure 3. With
ft∗ (zt ) := max[rt (zt , xt )ft+1

∗
(zt+1 )]
xt
for t = 3, 2, 1, where zt+1 = Gt (zt , xt ) = zt + xt and f4∗ (z4 ) = r4 (z4 ),

we should get f1∗ (z1 = 4) = 1 with an “optimal” policy (0, 0, −1).
However, the policy (−1, 1, 0) yields F (r1 , · · · , r4 ) = 4. Hence the
backward recursion does not yield the optimal solution when the returns
are calculated in a multiplicative fashion.
In this example we had
F (r1 (z1 , x1 ), · · · , rT (zT , xT )) = r1 (z1 , x1 ) ⊕ r2 (z2 , x2 ) ⊕ · · · ⊕ rT (zT , xT ),
where the composition operation “⊕” was chosen as addition in case (a) and
multiplication in case (b). For the backward recursion we have made use of
the so-called separability of F . That is, there exist two functions ϕ1 , ψ2 such
that
F (r1 (z1 , x1 ), · · · , rT (zT , xT )) (1.1)

= ϕ1 (r1 (z1 , x1 ), ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT ))).
Furthermore, we proceeded “as if” the following relation held:

DYNAMIC SYSTEMS 115
Figure 3 Dynamic program: multiplicative composition. Solid lines show the

result of the backward recursion (with z1 = 4), whereas the dotted line shows
the optimal sequence of decisions.
max{F (r1 (z1 , x1 ), · · · , rT (zT , xT )) | xt ∈ Xt , t = 1, · · · , T } (1.2)

= max [ϕ1 (r1 (z1 , x1 ), max ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT )))].
x1 ∈X1 x2 ∈X2 ,···,xT ∈XT
This relation is the formal equivalent of the well-known optimality principle,

which was expressed by Bellman as follows (quote).
Proposition 2.1 “An optimal policy has the property that whatever the
initial state and initial decision are, the remaining decisions must constitute
an optimal policy with regard to the state resulting from the first decision.”
As we have seen in Example 2.1, this principle, applied repeatedly in

the backward recursion, gave the optimal solution for case (a) but not for
case (b). The reason for this is that, although the composition operation
“⊕” is separable in the sense of (1.1), this is not enough to guarantee that
the repeated application of the optimality principle (i.e. through backward
recursion) will yield an optimal policy. A sufficient condition under which the
optimality principle holds involves a certain monotonicity of our composition
operation “⊕”. More precisely, we have the following.
Proposition 2.2 If F satisfies the separability condition (1.1) and if ϕ1

is monotonically nondecreasing in ψ2 for every r1 then the optimality
principle (1.2) holds.
Proof The unique meaning of “max” implies that we have that

max ϕ1 (r1 (z1 , x1 ), ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT )))
{xt ∈Xt , t≥1}
≥ ϕ1 (r1 (z1 , x1 ), max [ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT ))]),

{xt ∈Xt , t≥2}
for all x1 . Therefore this also holds when the right-hand side of this inequality
is maximized with respect to x1 .
On the other hand, it is also obvious that
max ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT ))
{xt ∈Xt , t≥2}
≥ ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT )) ∀xt ∈ Xt , t ≥ 2.

Hence, by the assumed monotonicity of ϕ1 with respect to ψ2 , we have that
ϕ1 (r1 (z1 , x1 ), max ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT )))

{xt ∈Xt ,t≥2}
≥ ϕ1 (r1 (z1 , x1 ), ψ2 (r2 (z2 , x2 ), · · · , rT (zT , xT ))) ∀xt ∈ Xt , t ≥ 1.

Taking the maximum with respect to xt , t ≥ 2, on the right-hand side of this
inequality and maximizing afterwards both sides with respect to x1 ∈ X1
shows that the optimality principle (1.2) holds. 2
Needless to say, all problems considered by Bellman in his first book on

dynamic programming satisfied this proposition. In case (b) of our example,
however, the monotonicity does not hold. The reason is that when “⊕” involves
multiplication of possibly negative factors (i.e. negative immediate returns),
the required monotonicity is lost. On the other hand, when “⊕” is summation,
the required monotonicity is always satisfied.
Let us add that the optimality principle applies to a much wider class of
problems than might seem to be the case from this brief sketch. For instance,
if for finitely many states we denote by ρt the vector having as ith component
the immediate return rt (zt = i), and if we define the composition operation
“⊕” such that, with a positive matrix S (i.e. all elements of S nonnegative),
ρt ⊕ ρt+1 = ρt + Sρt+1 , t = 1, · · · , T − 1,
then the monotonicity assumed for Proposition 2.2 follows immediately. This
case is quite common in applications. Then S is the so-called transition matrix,
which means that an element sij represents the probability of entering state
j at stage t + 1, given that the system is in state i at stage t. Iterating the
above composition for T − 1, T − 2, · · · , 1 we get that F (ρ1 , · · · , ρT ) is the
vector of the expected total returns. The ith component gives the expected
overall return if the system starts from state i at stage 1. Putting it this way
we see that multistage stochastic programs with recourse (formula (4.13) in
Chapter 1) belong to this class.
DYNAMIC SYSTEMS 117
2.2 Dynamic Programming
The purpose of this section is to look at certain aspects of the field of dynamic
programming. The example we looked at in the previous section is an example
of a dynamic programming problem. It will not represent a fair description
of the field as a whole, but we shall concentrate on aspects that are useful in
our context. This section will not consider randomness. That will be discussed
later.
We shall be interested in dynamic programming as a means of solving
problems that evolve over time. Typical examples are production planning
under varying demand, capacity expansion to meet an increasing demand
and investment planning in forestry. Dynamic programming can also be used
to solve problems that are not sequential in nature. Such problems will not
be treated in this text.
Important concepts in dynamic programming are the time horizon, state
variables, decision variables, return functions, accumulated return functions,
optimal accumulated returns and transition functions. The time horizon refers
to the number of stages (time periods) in the problem. State variables describe
the state of the system, for example the present production capacity, the
present age and species distribution in a forest or the amount of money
one has in different accounts in a bank. Decision variables are the variables
under one’s control. They can represent decisions to build new plants, to cut
a certain amount of timber, or to move money from one bank account to
another. The transition function shows how the state variables change as a
function of decisions. That is, the transition function dictates the state that
will result from the combination of the present state and the present decisions.
For example, the transition function may show how the forest changes over
the next period as a result of its present state and of cutting decisions, how the
amount of money in the bank increases, or how the production capacity will
change as a result of its present size, investments (and detoriation). A return
function shows the immediate returns (costs or profits) as a result of making
a specific decision in a specific state. Accumulated return functions show the
accumulated effect, from now until the end of the time horizon, associated with
a specific decision in a specific state. Finally, optimal accumulated returns show
the value of making the optimal decision based on an accumulated return
function, or in other words, the best return that can be achieved from the
present state until the end of the time horizon.
Example 2.2 Consider the following simple investment problem, where it is

clear that the Bellman principle holds. We have some money S0 in a bank
account, called account B. We shall need the money two years from now, and
today is the first of January. If we leave the money in the account we will face
an interest rate of 7% in the first year and 5% in the second. You also have
10% 7%
fee 20 fee 20
A A A
fee 10 fee 10
S0 S3
B B B
7% 5%
Figure 4 Graphical description of a simple investment problem.
the option of moving the money to account A. You will there face an interest
rate of 10% the first year and 7% the second year. However, there is a fixed
charge of 20 per year and a charge of 10 each time we withdraw money from
account A. The fixed charge is deducted from the account at the end of a year,
whereas the charges on withdrawals are deducted immediately. The question
is: Should we move our money to account A for the first year, the second year
or both years? In any case, money left in account A at the end of the second
year will be transferred to account B. The goal is to solve the problem for all
initial S0 > 1000. Figure 4 illustrates the example.
Note that all investments will result in a case where the wealth increases,
and that it will never be profitable to split the money between the accounts
(why?).
Let us first define the two-dimensional state variables zt = (zt1 , zt2 ). The
first state variable, zt1 , refers to the account name (A or B); the second
state variable, zt2 , refers to the amount of money St in that account. So
zt = (B, St ) refers to a state where there is an amount St in account B
in stage t. Decisions are where to put the money for the next time period. If
xt is our decision variable then xt ∈ {A, B}. The transition function will be
denoted by Gt (zt , xt ), and is defined via interest rates and charges. It shows
what will happen to the money over one year, based on where the money is
now, how much there is, and where it is put next. Since the state space has
two elements, the function Gt is two-valued. For example

1 1 A 2 2 A
zt+1 = Gt , A = A, zt+1 = Gt , A = St × 1.07 − 20.
St St
Accumulated return functions will be denoted by ft (zt1 , zt2 , xt ). They

describe how the amount zt2 in account zt1 will grow, up to the end of the
time horizon, if the money is put into account xt in the next period, and
optimal decisions are made thereafter. So if f1 (A, S1 , B) = S, we know that
in stage 1 (i.e. at the end of period 1), if we have S1 in account A and then
DYNAMIC SYSTEMS 119
move it to account B, we shall be left with S3 = S in account B at the end of

the time horizon, given that we make optimal decisions at all stages after stage
1. By maximizing over all possible decisions, we find the optimal accumulated
returns ft∗ (zt1 , zt2 ) for a given state. For example,
f1∗ (A, S1 ) = max f1 (A, S1 , x1 ).

x1 ∈{A,B}
The calculations for our example are as follows. Note that we have three
stages, which we shall denote Stage 0, Stage 1 and Stage 2. Stage 2 represents
the point in time (after two years) when all funds must be transferred to
account B. Stage 1 is one year from now, where we, if we so wish, may move
the money from one account to another. Stage 0 is now, where we must decide
if we wish to keep the money in account B or move it to account A.
Stage 2 At Stage 2, all we can do is to transfer whatever money we have in
account A to account B:
f2∗ (A, S2 ) = S2 − 10,
f2∗ (B, S2 ) = S2 ,
indicating that a cost of 10 is incurred if the money is in account A and needs

to be transferred to account B.
Stage 1 Let us first consider account A, and assume that the account contains
S1 . We can keep the money in account A, making S2 = S1 × 1.07 − 20 (this is
the transition function), or move it to B, making S2 = (S1 − 10) × 1.05. This
generates the following two evaluations of the accumulated return function:
f1 (A, S1 , A) = f2∗ (A, S1 × 1.07 − 20) = S1 × 1.07 − 30,

f1 (A, S1 , B) = f2∗ (B, (S1 − 10) × 1.05) = S1 × 1.05 − 10.5.
By comparing these two, we find that, as long as S1 ≥ 975 (which is always

the case since we have assumed that S0 > 1000), account A is best, making
f1∗ (A, S1 ) = S1 × 1.07 − 30.
Next, consider account B. If we transfer the amount S1 to account A, we

get S2 = S1 × 1.07 − 20. If it stays in B, we get S2 = S1 × 1.05. This gives us
f1 (B, S1 , A) = f2∗ (A, S1 × 1.07 − 20) = S1 × 1.07 − 30,

f1 (B, S1 , B) = f2∗ (B, S1 × 1.05) = S1 × 1.05.
By comparing these two, we find that

S1 × 1.07 − 30 if S1 ≥ 1500,
f1∗ (B, S1 ) =
S1 × 1.05 if S1 ≤ 1500.
Stage 0 Since we start out with all our money in account B, we only need
to check that account. Initially we have S0 . If we transfer to A, we get
S1 = S0 × 1.1 − 20, and if we keep it in B, S1 = S0 × 1.07. The accumulated
returns are
f0 (B, S0 , A) = f1∗ (A, S1 ) = f1∗ (A, S0 × 1.1 − 20)
= (S0 × 1.1 − 20) × 1.07 − 30 = 1.177 × S0 − 51.4,
f1∗ (B, S1 ) = f1∗ (B, S0 × 1.07)
f0 (B, S0 , B) =
S0 × 1.1449 − 30 if S0 ≥ 1402,
=
S0 × 1.1235 if S0 ≤ 1402.
Comparing the two options, we see that account A is always best, yielding
f0∗ (B, S0 ) = 1.177 × S0 − 51.4.
So we should move our money to account A and keep it there until the end
of the second period. Then we move it to B as required. We shall be left with
a total interest of 17.7% and fixed charges of 51.4 (including lost interest on
charges).
2
As we can see, the main idea behind dynamic programming is to take one
stage at a time, starting with the last stage. For each stage, find the optimal
decision for all possible states, thereby calculating the optimal accumulated
return from then until the end of the time horizon for all possible states. Then
move one step towards the present, and calculate the returns from that stage
until the end of the time horizon by adding together the immediate returns,
and the returns for all later periods based on the calculations made at the
previous stage. In the example we found that f1∗ (A, S1 ) = S1 × 1.07 − 30.
This shows us that if we end up in stage 1 with S1 in account A, we shall (if
we behave optimally) end up with S1 × 1.07 − 30 in account B at the end of
the time horizon. However, f1∗ does not tell us what to do, since that is not
needed to calculate optimal decisions at stage 0.
Formally speaking, we are trying to solve the following problem, where
x = (x0 , . . . , xT )T :
maxx F (r0 (z0 , x0 ), . . . , rT (zT , xT ), Q(zT +1 ))

s.t. zt+1 = Gt (zt , xt ) for t = 0, . . . , T,
At (zt ) ≤ xt ≤ Bt (zt ) for t = 0, . . . , T,
where F satisfies the requirements of Proposition 2.2. This is to be solved

for one or more values of the initial state z0 . In this set-up, rt is the return
DYNAMIC SYSTEMS 121
function for all but the last stage, Q the return function for the last stage, Gt
the transition function, T the time horizon, zt the (possibly multi-dimensional)
state variable in stage t and xt the (possibly multi-dimensional) decision
variable in stage t. The accumulated return function ft (zt , xt ) and optimal
accumulated returns ft∗ (zt ) are not part of the problem formulation, but
rather part of the solution procedure. The solution procedure, justified by
the Bellman principle, runs as follows.
Find f0∗ (z0 )
by solving recursively
ft∗ (zt ) = max ft (zt , xt )

At (zt )≤xt ≤Bt (zt )
∗
= max ϕt (rt (zt , xt ), ft+1 (zt+1 )) for t = T, . . . , 0
with
zt+1 = Gt (zt , xt ) for t = T, . . . , 0,
fT∗ +1 (zT +1 ) = Q(zT +1 ).
In each case the problem must be solved for all possible values of the state
variable zt , which might be multi-dimensional.
Problems that are not dynamic programming problems (unless rewritten
with a large expansion of the state space) would be problems where, for
example,
zt+1 = Gt (z0 , . . . , zt , x0 , . . . , xt ),
or where the objective function depends in an arbitrary way on the whole
history up til stage t, represented by
rt (z0 , . . . , zt , x0 , . . . , xt ).
Such problems may more easily be solved using other approaches, such as
decision trees, where these complicated functions cause little concern.
2.3 Deterministic Decision Trees
We shall not be overly interested in decision trees in a deterministic setting.

However, since they might be used to analyse sequential decision problems,
we shall mention them. Let us consider our simple investment problem in
Example 2.2. A decision tree for that problem is given in Figure 5. A decision
tree consists of nodes and arcs. The nodes represent states, the arcs decisions.
For each possible state in each stage we must create one arc for each possible
decision. Therefore the number of possible decisions must be very limited for
Stage 0
B,S 0
B
A
Stage 1 A,S1 B,S1
A B A B
Stage 2 A,S2 B, S2 A,S2 B, S2
B B B B
B, S3 B, S3 B, S3 B, S3
Figure 5 Deterministic decision tree for the small investment problem.
this method to be useful, since there is one leaf in the tree for each possible
sequence of decisions.
The tree indicates that at stage 0 we have S0 in account B. We can then
decide to put them into A (go left) or keep them in B (go right). Then at stage
1 we have the same possible decisions. At stage 2 we have to put them into
B, getting S3 , the final amount of money. As before we could have skipped
the last step. To be able to solve this problem, we shall first have to follow
each path in the tree from the root to the bottom (the leaves) to find S3
in all cases. In this way, we enumerate all possible sequences of decisions
that we can possibly make. (Remember that this is exactly what we avoid in
dynamic programming). The optimal sequence, must, of course, be one of these
sequences. Let (AAB) refer to the path in the tree with the corresponding
indices on the arcs. We then get
(ABB) : S3 = ((S0 × 1.1 − 20) − 10) × 1.05 = S0 × 1.155 − 31.5,

(AAB) : S3 = ((S0 × 1.1 − 20) × 1.07 − 20) − 10 = S0 × 1.177 − 51.4,
(BAB) : S3 = ((S0 × 1.07) × 1.07 − 20) − 10 = S0 × 1.1449 − 30,
(BBB) : S3 = S0 × 1.07 × 1.05 = S0 × 1.1235.
We have now achieved numbers in all leaves of the tree (for some reason
decision trees always grow with the root up). We are now going to move back
DYNAMIC SYSTEMS 123
towards the root, using a process called folding back. This implies moving one
step up the tree at a time, finding for each node in the tree the best decision
for that node.
This first step is not really interesting in this case (since we must move the
money to account B), but, even so, let us go through it. We find that the best
we can achieve after two decisions is as follows.
(AB) : S3 = S0 × 1.155 − 31.5,

(AA) : S3 = S0 × 1.177 − 51.4,
(BA) : S3 = S0 × 1.1449 − 30,
(BB) : S3 = S0 × 1.1235.
Then we move up to stage 1 to see what is the best we can achieve if we
presently have S1 in account A. The answer is
max{S0 × 1.155 − 31.5, S0 × 1.177 − 51.4} = S0 × 1.177 − 51.4
so long as S0 > 1000, the given assumption. If we have S1 in account B (the
right node in stage 1), we get

S0 × 1.1449 − 30 if S0 ≥ 1402,
max{S0 × 1.1449 − 30, S0 × 1.1235} =
S0 × 1.1235 if S0 ≤ 1402.
We can then fold back to the top, finding that it is best going left, obtaining
the given S3 = S0 × 1.177 − 51.4. Of course, we recognize most of these
computations from Section 2.2 on dynamic programming.
You might feel that these computations are not very different from those in
the dynamic programming approach. However, they are. For example, assume
that we had 10 periods, rather than just 2. In dynamic programming we would
then have to calculate the optimal accumulated return as a function of St for
both accounts in 10 periods, a total of 2 ∗ 10 = 20 calculations, each involving
a maximization over the two possible decisions. In the decision tree case the
number of such calculations will be 210 + 29 + . . . + 1 = 211 − 1 = 2047. (The
counting depends a little on how we treat the last period.) This shows the
strength of dynamic programming. It investigates many fewer cases. It should
be easy to imagine situations where the use of decision trees is absolutely
impossible due to the mere size of the tree.
On the other hand, the decision tree approach certainly has advantages.
Assume, for example, that we were not to find the optimal investments for
all S0 > 1000, but just for S0 = 1000. That would not help us much in the
dynamic programming approach, except that f1∗ (B, S1 ) = S1 × 1.05, since
S1 < 1500. But that is a minor help. The decision tree case, on the other
hand, would produce numbers in the leaves, not functions of S0 , as shown
above. Then folding back will of course be very simple.
Table 1 Distribution of the interest rate on account A. All outcomes have

probability 0.5, and the outcomes in period 2 are independent of the outcomes
in period 1.
Period Outcome 1 Outcome 2

1 8 12
2 5 9
8% or 12% 5% or 9%
fee 20 fee 20
A A A
fee 10 fee 10
S0 S3
B B B
7% 5%
Figure 6 Simple investment problem with uncertain interest rates.
2.4 Stochastic Decision Trees
We shall now see how decision trees can be used to solve certain classes
of stochastic problems. We shall initiate this with a look at our standard
investment problem in Example 2.2. In addition, let us now assume that the
interest on account B is unchanged, but that the interest rate on account A
is random, with the previously given rates as expected values. Charges on
account A are unchanged. The distribution for the interest rate is given in
Table 1. We assume that the interest rates in the two periods are described
by independent random variables.
Based on this information, we can give an update of Figure 4, where we
show the deterministic and stochastic parameters of the problem. The update
is shown in Figure 6.
Consider the decision tree in Figure 7. As in the deterministic case, square
nodes are decision nodes, from which we have to choose between account A
and B. Circular nodes are called chance nodes, and represent points at which
something happens, in this case that the interest rates become known.
Start at the top. In stage 0, we have to decide whether to put the money
DYNAMIC SYSTEMS 125
Stage 0
B
A
12%
8% 7%
Stage 1
A B A B A B
5% 9% 5% 9%
5% 9% 5% 5%
5%
Stage 2
Figure 7 Stochastic decision tree for the simple investment problem.
into account A or into B. If we choose A, we shall experience an interest rate

of 8% or 12% for the first year. After that we shall have to make a new decision
for the second year. That decision will be allowed to depend on what interest
rate we experienced in the first period. If we choose A, we shall again face
an uncertain interest rate. Whenever we choose B, we shall know the interest
rate with certainty.
Having entered a world of randomness, we need to specify what our decisions
will be based on. In the deterministic setting we maximized the final amount
in account B. That does not make sense in a stochastic setting. A given series
of decisions does not produce a certain amount in account B, but rather
an uncertain amount. In other words, we have to compare distributions. For
example, keeping the money in account A for both periods will result in one
out of four sequences of interest rates, namely (8,5), (8,9), (12,5) or (12,9).
Hence, if we start out with, say 1000, we can end up with (remember the fees)
1083, 1125, 1125 or 1169 (rounded numbers).
An obvious possibility is to look for the decision that produces the highest
expected amount in account B after two periods. However—and this is a very
important point—this does not mean that we are looking for the sequence
of decisions that has the highest expected value. We are only looking for the
best possible first decision. If we decide to put the money in account A in the
first period, we can wait and observe the actual interest rate on the account
before we decide what to do in the next period. (Of course, if we decide to use
B in the first period, we can as well decide what to do in the second period
immediately, since no new information is made available during the first year!)
Stage 0 1126
B
A
1126 1124
12%
8% 7%
Stage 1 1147 1124
1104
A B A B A B
1104 1103 1147 1145 1115 1124

5% 9% 5% 9%
5% 9% 5% 5%
5%
Stage 2 1083 1125 1103 1125 1169 1145 1094 1136 1124
Figure 8 Stochastic decision tree for the investment problem when we
maximize the expected amount in account B at the end of stage 2.
Let us see how this works. First we do as we have done before: we follow each
path down the tree to see what amount we end up with in account B. We
have assumed S0 = 1000. That is shown in the leaves of the tree in Figure 8.
We then fold back. Since the next node is a chance node, we take the
expected value of the two square nodes below. Then for stage 1 we check
which of the two possible decisions has the largest expectation. In the far left
of Figure 8 it is to put the money into account A. We therefore cross out
the other alternative. This process is repeated until we reach the top level.
In stage 0 we see that it is optimal to use account A in the first period and
regardless of the interest rate in the first period, we shall also use account
A in the second period. In general, the second-stage decision depends on the
outcome in the first stage, as we shall see in a moment.
You might have observed that the solution derived here is exactly the same
as we found in the deterministic case. This is caused by two facts. First, the
interest rate in the deterministic case equals the expected interest rate in the
stochastic case, and, secondly, the objective function is linear. In other words,
if ξ˜ is a random variable and a and b are constants then
Eξ̃ (aξ˜ + b) = aE ξ̃ + b.
For the stochastic case we calculated the left-hand side of this expression, and
for the deterministic case the right-hand side.
In many cases it is natural to maximize expected profits, but not always.
One common situation for decision problems under uncertainty is that
the decision is repeated many times, often, in principle, infinitely many.
DYNAMIC SYSTEMS 127
Utility
u(w )
0
w Wealth
0
Figure 9 Example of a typical concave utility function representing risk
aversion.
Investments in shares and bonds, for example, are usually of this kind. The
situation is characterized by long time series of data, and by many minor
decisions. Should we, or should we not, maximize expected profits in such a
case?
Economics provide us with a tool to answer that question, called a utility
function. Although it is not going to be a major point in this book, we should
like to give a brief look into the area of utility functions. It is certainly an
area very relevant to decision making under uncertainty. If you find the topic
interesting, consult the references listed at the end of this chapter. The area
is full of pitfalls and controversies, something you will probably not discover
from our little glimpse into the field. More than anything, we simply want to
give a small taste, and, perhaps, something to think about.
We may think of a utility function as a function that measures our happiness
(utility) from a certain wealth (let us stick to money). It does not measure
utility in any fixed unit, but is only used to compare situations. So we can say
that one situation is preferred to another, but not that one situation is twice
as good as another. An example of a utility function is found in Figure 9.
Note that the utility function is concave. Let us see what that means.
Assume that our wealth is w0 , and we are offered a game. With 50%
probability we shall win δw; with 50% probability we shall lose the same
amount. It costs nothing to take part. We shall therefore, after the game, either
have a wealth of w0 + δw or a wealth of w0 − δw. If the function in Figure 9 is
our utility function, and we calculate the utility of these two possible future
situations, we find that the decrease in utility caused by losing δw is larger

than the increase in utility caused by winning δw. What has happened is that
we do not think that the advantage of possibly increasing our wealth by δw
is good enough to offset our worry about losing the same amount. In other
words, our expected utility after having taken part in the game is smaller
than our certain utility of not taking part. We prefer w0 with certainty to a
distribution of possible wealths having expected value w0 . We are risk-averse.
If we found the two situations equally good, we are risk-neutral. If we prefer
the game to the certain wealth, we are risk seekers or gamblers.
It is generally believed that people are risk-averse, and that they need a
premium to take part in a gamble like the one above. Empirical investigations
of financial markets confirm this idea. The premium must be high enough to
make the expected utility of taking part in the game (including the premium)
equal to the utility of the wealth w0 .
Now, finally, we are coming close to the question we started out with. Should
we maximize expected profit? We have seen above that maximizing expected
profit can be interpreted as maximizing expected utility with a risk-neutral
attitude. In other words, it puts us in a situation where a fair gamble (i.e. one
with expected value zero) is acceptable. When can that be the case?
One reason can be that the project under consideration is very small
compared with the overall wealth of the decision maker, so that risk aversion
is not much of an issue. As an example, consider the purchase of a lottery
ticket. Despite the fact that the expected value of taking part in a lottery
is negative, people buy lottery tickets. This fact can create some theoretical
problems in utility theory, problems that we shall not discuss here. A reference
is given at the end of the chapter.
A more important case arises in public investments. One can argue that
the government should not trade expected values for decreased risks, since
the overall risk facing a government is very small, even if the risk in one
single project is large. The reason behind this argument is that, with a very
large number of projects at hand (which certainly the government has), some
will win, some will lose. Over all, owing to offsetting effects, the government
will face very little risk. (It is like in a life insurance company, where the
death of costumers is not considered a random event. With a large number of
costumers, they “know” how many will die the next year.)
In all, as we see, we must argue in each case whether or not a linear or
concave utility function is appropriate. Clearly, in most cases a linear utility
function creates easier problems to solve. But in some cases risk should indeed
be taken into account.
Let us now continue with our example, and assume we are faced with a
concave utility function
u(s) = ln(s − 1000)

DYNAMIC SYSTEMS 129
Stage 0 4.820
B
A
4.807 4.820
12%
8% 7%
Stage 1 4.634 4.979 4.820
A B A B A B
4.624 4.634 4.979 4.977 4.683 4.820

5% 9% 5% 9%
5% 9% 5% 5%
5%
Stage 2 4.419 4.828 4.634 4.828 5.130 4.977 4.543 4.913 4.820
Figure 10 Stochastic decision tree for the investment problem when we
maximize the expected utility of the amount in account B at the end of period
2.
and that we wish to maximize the expected utility of the final wealth s. In
the deterministic case we found that it would never be profitable to split
the money between the two accounts. The argument is the same when we
simply maximized the expected value of S3 as outlined above. However, when
maximizing expected utility, that might no longer be the case. On the other
hand, the whole set-up used in this chapter assumes implicitly that we do
not split the funding. Hence in what follows we shall assume that all the
money must be in one and only one account. The idea in the decision tree is
to determine which decisions to make, not how to combine them. Figure 10
shows how we fold back with expected utilities. The numbers in the leaves
represent the utility of the numbers in the leaves of Figure 8. For example
u(1083) = ln(1083 − 1000) = ln 83 = 4.419.
We observe that, with this utility function, it is optimal to use account B.
The reason is that we fear the possibility of getting only 8% in the first period
combined with the charges. The result is that we choose to use B, getting
the certain amount S3 = 1124. Note that if we had used account A in the
first period (which is not optimal), the optimal second-stage decision would
depend on the actual outcome of the interest on account A in the first period.
With 8%, we pick B in the second period; with 12%, we pick A.
2.5 Stochastic Dynamic Programming
Looking back at Section 2.2 on dynamic programming, we observe two major

properties of the solution and solution procedure. First, the procedure (i.e.
dynamic programming) produces one solution per possible state in each stage.
These solutions are not stored, since they are not needed in the procedure, but
the extra cost incurred by doing so would be minimal. Secondly, if there is only
one given value for the initial state z0 , we can use these decisions to produce
a series of optimal solutions—one for each stage. In other words, given an
initial state, we can make plans for all later periods. In our small investment
Example 2.2 (to which we added randomness in the interest rates in Section 2.4
we found, in the deterministic case, that with S0 > 1000, x0 = x1 = A and
x2 = B was the optimal solution. That is, we put the money in account A for
the two periods, before we send the money to account B as required at the
end of the time horizon.
When we now move into the area of stochastic dynamic programming, we
shall keep one property of the dynamic programming algorithm, namely that
there will be one decision for each state in each stage, but it will no longer be
possible to plan for the whole period ahead of time. Decisions for all but the
first period will depend on what happens in the mean time. This is the same
as we observed for stochastic decision trees.
Let us turn to the small investment example, keeping the extra requirement
that the money must stay in one account, and using the utility function
u(s) = ln(s − 1000).
Stage 2 As for the deterministic case, we find that
f2∗ (A, S2 ) = ln(S2 − 1010),
f2∗ (B, S2 ) = ln(S2 − 1000),
since we must move the money into account B at the end of the
second year.
Stage 1 We have to consider the two accounts separately.
Account A If we keep the money in account A, we get the following
expected return:
f1 (A, S1 , A) = 0.5[f2∗ (A, S1 × 1.05 − 20) + f2∗ (A, S1 × 1.09 − 20)]
= 0.5 ln[(S1 × 1.05 − 1030) (S1 × 1.09 − 1030)].
If we move the money to account B, we get
f1 (A, S1 , B) = f2∗ (B, (S1 − 10) × 1.05)
= ln(S1 × 1.05 − 1010.5).
To find the best possible solution, we compare these two possibilities
by calculating
DYNAMIC SYSTEMS 131
f1∗ (A, S1 ) = max{f1 (A, S1 , A), f1 (A, S1 , B)}

= max{0.5 ln[(S1 × 1.05 − 1030)(S1 × 1.09 − 1030)],
ln(S1 × 1.05 − 1010.5)},
from which we find, (remembering that S1 > 1000)

⎧
⎨ ln(S1 × 1.05 − 1010.5)
⎪ if S1 < 1077,
∗
f1 (A, S1 ) = 0.5 ln[(S1 × 1.05 − 1030)(S1 × 1.09 − 1030)]
⎪
⎩
if S1 > 1077.
Account B. For account B we can either move the money to

account A to get
f1 (B, S1 , A)
= 0.5[f2∗ (A, S1 × 1.05 − 20) + f2∗ (A, S1 × 1.09 − 20)]
= 0.5 ln[(S1 × 1.05 − 1030)(S1 × 1.09 − 1030)],
or we can keep the money in B to obtain
f1 (B, S1 , B) = f2∗ (B, S1 × 1.05)

= ln(S1 × 1.05 − 1000).
To find the best possible solution, we calculate
f1∗ (B, S1 ) = max{f1 (B, S1 , A), f1 (B, S1 , B)}

= max{0.5 ln[(S1 × 1.05 − 1030)(S1 × 1.09 − 1030)],
ln(S1 × 1.05 − 1000)}.
From this, we find that (remembering that S1 > 1000)

⎧
⎨ ln(S1 × 1.05 − 1000)
⎪ if S1 < 1538
∗
f1 (B, S1 ) = 0.5 ln[(S1 × 1.05 − 1030)(S1 × 1.09 − 1030)]
⎪
⎩
if S1 > 1538.
Stage 0 We here have to consider only the case when the amount S0 > 1000
sits in account B. The basis for these calculations will be the
following two expressions. The first calculates the expected result
of using account A, the second the certain result of using account B.
f0 (B, S0 , A) = 0.5[f1∗ (A, S0 × 1.08 − 20) + f1∗ (A, S0 × 1.12 − 20)],

f0 (B, S0 , B) = f1∗ (B, S0 × 1.07).
Using these two expressions, we then calculate
f0∗ (B, S0 ) = max{f0 (B, S0 , A), f0 (B, S0 , B)}.
To find the value of this expression for f0∗ (B, S0 ), we must make sure
that we use the correct expressions for f1∗ from stage 1. To do that,
we must know how conditions on S1 relate to conditions on S0 . There
are three different ways S0 and S1 can be connected (see e.g. the top
part of Figure 10):
S1 = S0 × 1.08 − 20 ⇒ (S1 = 1077 ⇐⇒ S0 = 1016),

S1 = S0 × 1.12 − 20 ⇒ (S1 = 1077 ⇐⇒ S0 = 979),
S1 = S0 × 1.07 ⇒ (S1 = 1538 ⇐⇒ S0 = 1437).
From this, we see that three different cases must be discussed, namely
1000 < S0 < 1016, 1016 < S0 < 1437 and 1437 < S0 .
Case 1 Here 1000 < S0 < 1016. In this case
f0∗ (B, S0 ) = ln(S0 × 1.1235 − 1000),
which means that we always put the money into account B. (Make
sure you understand this by actually performing the calculations.)
Case 2 Here 1016 < S0 < 1437. In this case
⎧
⎪
⎪ ln(S0 × 1.1235 − 1000) if S0 < 1022,
⎪
⎨
f0∗ (B, S0 ) = 0.25 × ln[(S0 × 1.134 − 1051)
⎪
⎪ ×(S0 × 1.1772 − 1051.8) × (S0 × 1.176 − 1051)
⎪
⎩
×(S0 × 1.2208 − 1051.8)] if S0 > 1022,
which means that we use account B for small amounts and account
A for large amounts within the given interval.
Case 3 Here we have S0 > 1437. In this case
f0∗ (B, S0 ) = 1
4
ln[(S0 × 1.134 − 1051) × (S0 × 1.1772 − 1051.8)
×(S0 × 1.176 − 1051) × (S0 × 1.2208 − 1051.8)],
which means that we should use account A.

Summing up all cases, for stage 0, we get
⎧
⎪
⎪ ln(S0 × 1.1235 − 1000) if S0 < 1022,
⎪
⎨
f0∗ (B, S0 ) = 0.25 × ln[(S0 × 1.134 − 1051)
⎪
⎪ ×(S0 × 1.1772 − 1051.8) × (S0 × 1.176 − 1051)
⎪
⎩
×(S0 × 1.2208 − 1051.8)] if S0 > 1022.
DYNAMIC SYSTEMS 133
S1 >1077
A A A
S 0 >1022 S1<1077 S1 >1538
S3
S0 B B B
S 0 <1022 S 1<1538
Stage 0 Stage 1 Stage 2

Figure 11 Description of the solution to the stochastic investment problem
using stochastic dynamic programming.
If we put these results into Figure 4, we obtain Figure 11. From the latter,
we can easily construct a solution similar to the one in Figure 10 for any
S0 > 1000. Verify that we do indeed get the solution shown in Figure 10 if
S0 = 1000.
But we see more than that from Figure 11. We see that if we choose account
B in the first period, we shall always do the same in the second period. There
is no way we can start out with S0 < 1022 and get S1 > 1538.
Formally, what we are doing is as follows. We use the vocabulary of
Section 2.2. Let the random vector for stage t be given by ξ̃t and let the
return and transition functions become rt (zt , xt , ξt ) and zt+1 = Gt (zt , xt , ξt ).
Given this, the procedure becomes
find f0∗ (z0 )
by recursively calculating
ft∗ (zt ) = min ft (zt , xt )

∗
= min Eξ̃t {ϕt (rt (zt , xt , ξ̃t ), ft+1 (zt+1 ))}, t = T, . . . , 0,
with
zt+1 = Gt (zt , xt , ξt ) for t = 0, . . . , T,
fT∗ +1 (zT +1 ) = Q(zT +1 ),
where the functions satisfy the requirements of Proposition 2.2. In each stage
the problem must be solved for all possible values of the state zt . It is possible
to replace expectations (represented by E above) by other operators with
respect to ξ˜t , such as max or min. In such a case, of course, probability
distributions are uninteresting—only the support matters.
2.6 Scenario Aggregation
So far we have looked at two different methods for formulating and solving
multistage stochastic problems. The first, stochastic decision trees, requires a
tree that branches off for each possible decision xt and each possible realization
of ξ˜t . Therefore these must both have finitely many possible values. The state
zt is not part of the tree, and can therefore safely be continuous. A stochastic
decision tree easily grows out of hand.
The second approach was stochastic dynamic programming. Here we must
make a decision for each possible state zt in each stage t. Therefore, it is clearly
an advantage if there are finitely many possible states. However, the theory is
also developed for a continuous state space. Furthermore, a continuous set of
decisions xt is acceptable, and so is a continuous distribution of ξ˜t , provided
we are able to perform the expectation with respect to ξ˜t .
The method we shall look at in this section is different from those mentioned
above with respect to where the complications occur. We shall now operate
on an event tree (see Figure 12 for an example). This is a tree that branches
off for each possible value of the random variable ξ̃t in each stage t. Therefore,
compared with the stochastic decision tree approach, the new method has
similar requirements in terms of limitations on the number of possible values
of ξ˜t . Both need finite discrete distributions. In terms of xt we must have
finitely many values in the decision tree, the new method prefers continuous
variables. Neither of them has any special requirements on zt .
The second approach we have discussed so far for stochastic problems
is stochastic dynamic programming. The new method we are about to
outline is called scenario aggregation. We shall see that stochastic dynamic
programming is more flexible than scenario aggregation in terms of
distributions of ξ˜t , is similar with respect to xt , but is much more restrictive
with respect to the state variable zt , in the sense that the state space is hardly
of any concern in scenario aggregation.
If we have T time periods and ξt is a vector describing what happens in
time period t (i.e. a realization of ξ˜t ) then we call
s = (ξ0s , ξ1s , . . . , ξTs )
a scenario. It represents one possible future. So assume we have a set of

scenarios S describing all (or at least the most interesting) possible futures.
What do we do? Assume our “world” can be described by state variables zt
and decision variables xt and that the cost (i.e. the return function) in time
period t is given by rt (zt , xt , ξt ). Furthermore, as before, the state variables
can be calculated from
zt+1 = Gt (zt , xt , ξt ),
DYNAMIC SYSTEMS 135
with z0 given. Let α be a discount factor. What is often done in this case is
to solve for each s ∈ S the following problem
⎫

T
⎪
⎪
min t
α rt (zt , xt , ξts ) T +1
+α Q(zT +1 ) ⎪
⎬
t=0 (6.1)
s.t. zt+1 = Gt (zt , xt , ξts ) for t = 0, . . . , T with z0 given, ⎪
⎪
⎪
⎭
where Q(z) represents the value of ending the problem in state z, yielding
an optimal solution xs = (xs0 , xs1 , . . . , xsT ). Now what? We have a number of
different solutions—one for each s ∈ S. Shall we take the average and calculate
for each t

xt = ps xst ,
s∈S
where ps is the probability that we end up on scenario s? This is very often

done, either by explicit probabilities or by more subjective methods based on
“looking at the solutions”. However, several things can go wrong. First, if x is
chosen as our policy, there might be cases (values of s) for which it is not even
feasible. We should not like to suggest to our superiors a solution that might
be infeasible (infeasible probably means “going broke”, “breaking down” or
something like that). But even if feasibility is no problem, is using x a good
idea?
In an attempt to answer this, let us again turn to event trees. In Figure 12 we
have T = 1. The top node represents “today”. Then one out of three things can
happen, or, in other words, we have a random variable with three outcomes.
The second row of nodes represents “tomorrow”, and after tomorrow a varying
number of things can happen, depending on what happens today. The bottom
row of nodes takes care of the rest of the time—the future.
This tree represents six scenarios, since the tree has six leaves. In the setting
of optimization that we have discussed, there will be two decisions to be made,
namely one “today” and one “tomorrow”. However, note that what we do
tomorrow will depend on what happens today, so there is not one decision for
tomorrow, but rather one for each of the three nodes in the second row. Hence
x0 works as a suggested first decision, but x1 isn’t very interesting. However, if
we are in the leftmost node representing tomorrow, we can talk about an x1 for
the two scenarios going through that node. We can therefore calculate, for each
version of “tomorrow”, an average x1 , where the expectation is conditional
upon being on one of the scenarios that goes through the node.
Hence we see that the nodes in the event tree are decision points and the arcs
are realizations of random variables. From our scenario solutions xs we can
therefore calculate decisions for each node in the tree, and these will all make
Today
First random variable
Tomorrow
Second
random
variable
The future
Figure 12 Example of an event tree for T = 1.
sense, because they are all possible decisions, or what are called implementable
decisions.
For each time period t let {s}t be the set of all scenarios having ξ0s , . . . , ξt−1
s
in common with scenario s. In Figure 12, {s}0 = S, whereas each {s}2 contains
only one scenario. There are three sets {s}1 . Let p({s}t ) be the sum of the
probabilities of all scenarios in {s}t . Hence, after solving (6.1) for all s, we
calculate for all {s}t

ps xst

x({s}t ) = .
p({s}t )
s ∈{s}t
So what does this solution mean? It has the advantage that it is

implementable, but is it the optimal solution to any problem we might want
to solve? Let us now turn to a formal mathematical description of a multistage
problem that lives on an event tree, to see how x({s}t ) may be used. In this
description we are assuming that we have finite discrete distributions.
) T *

s t s s s T +1 s
min p α rt (zt , xt , ξt ) + α Q(zT +1 )
s∈S t=0
subject to
s
zt+1 = Gt (zts , xst , ξts ) for t = 0, . . . , T with z0s = z0 given,
At (zt ) ≤ xst ≤ Bt (zts ) for t = 0, . . . , T,
s
ps xs (6.2)
s t
xt = for t = 0, . . . , T and all s.

p({s}t )
s ∈{s}t
Note that only (6.2), the implementability constraints, connect the

scenarios. As discussed in Section 1.8, a common approach in nonlinear
DYNAMIC SYSTEMS 137
optimization is to move constraints that are seen as complicated into the

objective function, and penalize deviations. We outlined a number of different
approaches. For scenario aggregation, the appropriate one is the augmented
Lagrangian method. Its properties, when used with equality constraints such
as (6.2), were given in Propositions 1.27 and 1.28. Note that if we move the
implementability constraints into the objective, the remaining constraints are
separable in the scenarios (meaning that there are no constraints containing
information from more than one scenario). Our objective then becomes

T '
p(s) αt rt (zts , xst , ξts )
s∈S t=0 ⎫ (6.3)
ps xst (
⎬
+wts (xst − ) + αT +1 Q(zTs +1 )
p({s}t ) ⎭
s ∈{s}t
where wts is the multiplier for implementability for scenario s in period t.

If we add an augmented Lagrangian term, this problem can, in principle,
be solved by an approach where we first fix w, then solve the overall problem,
then update w and so on until convergence, as outlined in Section 1.8.2.4.
However, a practical problem (and a severe one as well) results from the fact
that the augmented Lagrangian term will change the objective function from
one where the different variables are separate, to one where products between
variables occur. Hence, although this approach is acceptable in principle, it
does not work well numerically, since we have one large problem instead of
many scenario problems that can be solved separately. What we then do is to
replace

ps xst

p({s}t )
s ∈{s}t
with

ps xst

x({s}t ) =
p({s}t )
s ∈{s}t
from the previous iteration. Hence, we get

T
' (
p(s) αt rt (zts , xst , ξts ) + wts [xst − x({s}t )] + αT +1 Q(zTs +1 ) .
s∈S t=0
But since, for a fixed w, the terms wts x({s}t ) are fixed, we can as well drop
them. If we then add an augmented Lagrangian term, we are left with
T ' (
p(s) αt rt (zts , xst , ξts ) + wts xst + 12 ρ[xst − x({s}t )]2 + αT +1 Q(zTs +1 ) .
s∈S t=0
procedure scenario(s, x, xs );
begin
Solve the problem

T
2 T +1
min t
α [rt (zt , xt , ξts ) + wts xt + ρ(xt − x) ] + α
1
2
Q(zT +1 )
t=0
s.t. zt+1 = Gt (zt , xt , ξts ) for t = 0, . . . , T, with z0 given,

to obtain xs = (xs0 , . . . , xsT ) and z s = (z0s , . . . , zTs +1 );
end;
Figure 13 Procedure for solving individual scenario problems.
Our problem is now totally separable in the scenarios. That is what we need
to define the scenario aggregation method. See the algorithms in Figures 13
and 14 for details. A few comments are in place. First, to find an initial
x({s}t ), we can solve (6.1) using expected values for all random variables.
Finding the correct value of ρ, and knowing how to update it, is very hard.
We discussed that to some extent in Chapter 1: see in particular (8.17). This is
a general problem for augmented Lagrange methods, and will not be discussed
here. Also, we shall not go into the discussion of stopping criteria, since the
details are beyond the scope of the book. Roughly speaking, though, the goal
is to have the scenario problems produce implementable solutions, so that xs
equals x({s}t ).
Example 2.3 This small example concerns a very simple fisheries

management model. For each time period we have one state variable, one
decision variable, and one random variable. Let zt be the state variable,
representing the biomass of a fish stock in time period t, and assume that
z0 is known. Furthermore, let xt be a decision variable, describing the portion
of the fish stock caught in a given year. The implicit assumption made here
is that it requires a fixed effort (measured, for example, in the number of
participating vessels) to catch a fixed portion of the stock. This seems to be
a fairly correct description of demersal fisheries, such as for example the cod
fisheries. The catch in a given year is hence zt xt .
During a year, fish grow, some die, and there is a certain recruitment. A
common model for the total effect of these factors is the so called Schaefer
model, where the total change in the stock, due to natural effects listed above,
DYNAMIC SYSTEMS 139
procedure scen-agg;
begin
for all s and t do wts := 0;
Find initial x({s}t );
Initiate ρ > 0;
repeat
for all s ∈ S do scenario(s, x({s}t ), xs );
for all x({s}t ) do
ps xs
t
x({s}t ) = ;

p({s} t)
s ∈{s}t
Update ρ if needed;
for all s and t do
wts := wts + ρ [xst − x({s}t )];
until result good enough;
end;
Figure 14 Principal set-up of the scenario aggregation method.
is given by + zt ,
szt 1 − ,
K
where s is a growth ratio and K is the carrying capacity of the environment.
Note that if zt = K there is no net change in the stock size. Also note that
if zt > K, then there is a negative net effect, decreasing the size of the stock,
and if zt < K, then there is a positive net effect. Hence zt = K is a stable
situation (as zt = 0 is), and the fish stock will, according to the model, stabilize
at z = K if no fishing takes place.
If fish are caught, the catch has to be subtracted from the existing stock,
giving us the following transition function:
+ zt ,
zt+1 = zt − xt zt + szt 1 − .
K
This transition function is clearly nonlinear, with both a zt xt term and a zt2
term. If the goal is to catch as much as possible, we might choose to maximize
∞

αt zt xt ,
t=0
where α is a discount factor. (For infinite horizons we need 0 ≤ α < 1, but

for finite problems we can choose to let α = 1.) In addition to this, we have
the natural constraint

0 ≤ xt ≤ 1.
So far, this is a deterministic control problem. It is known, however, that
predicting the net effects of growth, natural mortality and recruitment is very
difficult. In particular, the recruitment is not well understood. Therefore, it
seems unreasonable to use a deterministic model to describe recruitment, as
we have in fact done above. Let us therefore assume that the growth ratio s
is not known, but rather given by a random vector ξ˜t in time period t.
To fit into the framework of scenario aggregation, let us assume that we are
able to cut the problem after T periods, giving it a finite horizon. Furthermore,
assume that we have found a reasonable finite discretization of ξ˜t for all t ≤ T .
It can be hard to do that, but we shall offer some discussion in Section 3.4.
A final issue when making an infinite horizon problem finite is to construct a
function Q(zT +1 ) that, in a reasonable way, approximates the value of ending
up in state zT +1 at time T + 1. Finding Q can be difficult. However, let us
briefly show how one approximation can be found for our problem.
Let us assume that all ξ˜t are independent and identically distributed with
expected value ξ. Furthermore, let us simply replace all random variables with
their means, and assume that each year we catch exactly the net recruitment,
i.e. we let
+ zt ,
xt = ξ 1 − .
K
But since this leaves zt = zT +1 for all t ≥ T + 1, and therefore all xt for
t ≥ T + 1 equal, we can let
∞
ξzT +1 (1 − zT +1 /K)
Q(zT +1 ) = αt−T −1 xt zt = .
1−α
t=T +1
With these assumptions on the horizon, the existence of Q(zT +1 ) and a finite
discretization of the random variables, we arrive at the following optimization
problem, (the objective function amounts to the expected catch, discounted
over the horizon of the problem; of course, it is easy to bring this into monetary
terms):
' (
T t s s T +1
max s∈S p(s) t=0 α zt xt + α Q(zTs +1 )
' + ,(
zs
s
s.t. zt+1 = zts 1 − xξt + ξts 1 − Kt , with z0s = z0 given,
0 ≤ xst ≤ 1,

ps xst

xst = s ∈{s}t p({s}t ) .
We can then apply scenario aggregation as outlined before.

DYNAMIC SYSTEMS 141
2.6.1 Approximate Scenario Solutions

Consider the algorithm just presented. If the problem being solved is genuinely
a stochastic problem (in the sense that the optimal decisions change compared
with the optimal decisions in the deterministic—or expected value—setting),
we should expect scenario solutions xs to be very different initially, before the
dual variables ws obtain their correct values. Therefore, particularly in early
iterations, it seems a waste of energy to solve scenario problems to optimality.
What will typically happen is that we see a sort of “fight” between the scenario
solutions xs and the implementable solution x({s}t ). The scenario solutions
try to pull away from the implementable solutions, and only when the penalty
(in terms of wts ) becomes properly adjusted will the scenario solutions agree
with the implementable solutions. In fact, the convergence criterion, vaguely
stated, is exactly that the scenario solutions and the implementable solutions
agree.
From this observation, it seems reasonable to solve scenario problems only
approximately, but precisely enough to capture the direction in which the
scenario problem moves relative to the implementable solution. Of course, as
the iterations progress, and the dual variables wts adjust to their correct values,
the scenario solutions and the implementable solutions agree more and more.
In the end, if things are properly organized, the overall set-up converges. It
must be noted that the convergence proof for the scenario aggregation method
does indeed allow for approximate scenario solutions. From an algorithmic
point of view, this would mean that we replaced the solution procedure in
Figure 13 by one that found only an approximate solution.
It has been observed that by solving scenario problems only very
approximately, instead of solving them to optimality, one obtains a method
that converges much faster, also in terms of the number of outer iterations. It
simply is not wise to solve scenario problems to optimality. Not only can one
solve scenario problems approximately, one should solve them approximately.
2.7 Financial Models
Optimization models involving uncertainty have been used for a long time.
One of the best known models is the mean-variance model of Markowitz, for
which he later got the Nobel economics prize. In this section, we shall first
discuss the main principles behind Markowitz’ model. We shall then discuss
some of the weaknesses of the model, mostly in light of the subjects of this
book, before we proceed to outline later developments in financial modeling.
2.7.1 The Markowitz’ model

The purpose of the Markowitz model is to help investors distribute their funds
in a way that does not represent a waste of money. It is quite clear that
when you invest, there is a tradeoff between the expected payoff from your
investment, and the risk associated with it. Normally, the higher the expected
payoff, the higher the risk. However, for a given payoff you would normally
want as little risk as possible, and for a given risk level, you would want the
expected payoff to be as large as possible. If you, for example, have a higher
risk than necessary for a given expected payoff, you are wasting money, and
this is what the Markowitz model is constructed to help you avoid.
But what is risk? It clearly has something to do with the spread of the
possible payoffs. A portfolio (collection of investments) is riskier the higher the
spread, all other aspects equal. In the Markowitz model, the risk is measured
by the variance of the (random) payoffs from the investment. The model will
not tell us in what way we should combine expected payoffs with variance,
only make sure that we do not waste money. How to actually pick a portfolio
is left to other theories, such as for example utility theory, as discussed briefly
in Section 2.4.
Financial instruments such as for example bonds, stocks, options and
bank deposits all have random payoffs, although the uncertainty vary a lot.
Furthermore, the instruments are not statistically independent, but rather
strongly correlated. It is obvious that if the value of a 3 year bond increases,
so will normally the value of, say, a 5 year bond. The correlation is almost, but
not quite, perfect. In the same way, stocks from companies in similar sectors
of the economy often move together. On the other hand, if energy prices rise
internationally, the value of an oil company may increase, whereas the value of
an aluminum producer may decrease. If the interest rates increase, bonds will
normally decrease in value. In other words, we must in this setting operate
with dependent random variables.
Assume we have n possible investments instruments. Let xi be the
proportion of our funds invested in intrument i. Hence, xi = 1. Let the
payoff of instrument i be ξ̃i (with ξ˜ = (ξ˜1 , . . . , ξ̃n )), and let V be the variance-
covariance matrix for the investment, i.e.
! "
Vij = E (ξ˜i − E ξ̃i )(ξ˜j − E ξ̃j ) .
The variance of a portfolio is now xT V x, and the mean payoff is xT E ξ̃. We

now solve (letting e be a vector of 1’s).
DYNAMIC SYSTEMS 143
Mean Infeasible
Efficient
region
frontier
Inefficient
region
Variance
Figure 15 Efficient frontier generated by the Markowitz model.
⎫
min xT V x ⎪
⎪
⎬
s.t. xT E ξ̃ = v
(7.4)
xT e = 1 ⎪
⎪
⎭
x ≥ 0.
By solving (7.4) parametrically in v we obtain a curve called the efficient
frontier. An example is shown in Figure 15.
We note that the curve is flat on the top, indicating that there is a maximal
possible expected payoff, and also that there is a minimal possible variance.
Also, note that the curve bends backwards, showing that if you want to
achieve a lower mean than the one that corresponds to the minimal variance,
you will have to accept an increased variance. The points below the curve
are achievable, but represent a waste of money, the points above are not
achievable. Hence, we wish to be on the curve, the efficient frontier. There is
nothing in the model that tells us where on the curve we ought to be.
2.7.2 Weak aspects of the model

The above model, despite the fact that for most investors it is considered
advanced, has a number of shortcomings which relate to the subject of this
book. The first we note is that this is a two-stage model; We make decisions
under uncertainty (the investments), we then observe what happens, and
finally we obtain payoffs according to what happened. An important question
is therefore to what extent the problem is well modeled as a two-stage problem.
More and more people, both in industry and academia, tend to think that this
is not the case. The reasons are many, here follows a list of some of them. We
list them as they represent a valid way of thinking for any decision problem
under uncertainty.
• A two-stage (one period) model cannot treat instruments with different

time horizons correctly. If the length of the period is chosen to be short,
some instruments will look worse than they are, if it is long, others will
suffer.
• A one-period model cannot capture correctly the long term tradeoffs
between risk and expectation. Stocks, for example, are very risky in the
short run, and will suffer from a short time period in the model. If the time
period is long, stocks will look very good due to their high expected payoffs,
despite the risk. This is something we know well from daily life. If you have
money, and need to put them away for, say, three months, you do not buy
stocks. But if you need them in ten years, stocks will be a good alternative.
And this difference is not just caused by transaction costs.
• A one-period model cannot capture transaction costs properly. The way we
have presented the model, there are no transaction costs at all. A result of
this is that users observe that when the model is applied, it suggests far
too much trading because there is no penalty on changing the portfolio. We
could make a new model in Markowitz’ spirit with transaction costs (putting
a penalty on changes from the present portfolio), but it would be difficult
to formulate properly. The reason is that the transaction costs would have
to be offset by a better payoff over just one period. In reality, some (but
not all) reinvestments have years and years to pay off.
• A one-period model cannot (by definition) capture dynamic aspects (or
trends) in the development of the payoffs over time.
In addition to these observations, the model contains a number of implicit

assumption. Let us mention just one. By construction the model assumes that
only the first two moments (mean and variance) are relevant for a portfolio.
In other words, skewness and higher moments are disregarded. Most people
would argue that this is unrealistic for most instruments. However, for some,
there cannot be any doubt that just two moments is insufficient. Consider an
option, say, the (European) option to buy a share one year from now at price
p. Since this is a right but not a duty, the distribution of payoffs from the
option will consist of two parts: A point mass at zero, and a continuous part.
The point mass comes from all cases where p is higher than the market value
one year from now. If p is lower than the market value, then the payoff equals
the market value minus p. Hence, even if the distribution for the share was
fully described by two moments, this cannot be true for the option. Similar
statements can be made about other instruments such as forwards, futures,
swaps etc.
The result of using Markowitz’ model is the efficient frontier. Clearly, if
the assumptions behind the model are not correct, then the frontier is not so
efficient after all.
DYNAMIC SYSTEMS 145
2.7.3 More advanced models

In the previous subsection we discussed weak aspects of the mean-variance
model. There can be no doubt that the arguments are valid, and that there
exist other arguments as well. However, all models are exactly that, they are
models, and hence, they disregard certain aspects of the real phenomenon
they are made to represent. A good model is not one that captures every
aspect of a problem, but rather one that captures the essential aspects. Hence,
the fact that we can argue against the model is in itself not enough to say
that the model is not good. More than that is needed. For example, we may
demonstrate by using it (or maybe by simulation) that it gives bad results, or
we may demonstrate that there is indeed a better alternative. However, weak
as a model may be in some respects, it may still be the best available tool in
a given situation.
When we now turn to discuss new models for portfolio selection, we do it
because users and scientists have observed that it is indeed possible to obtain
results that both theoretically and in practice give better decisions.
An investment problem of the type discussed by Markowitz normally has
either a very long (but finite) time horizon, or it is in fact an infinite time
horizon problem. From a practical point of view, we would not know how to
solve it as an infinite horizon problem, so let us assume that the problem has
many, but finitely many, time periods. At least in principle, decisions can be
made in all periods, so that the problem is both multistage and multiperiod.
2.7.3.1 A scenario tree

A possibility in this situation is to represent the scenarios in terms of a scenario
or event tree. This will allow for arbitrary dependencies in the payoffs, thereby
taking care of the objections we made about the Markowitz model. We can
capture the fact that different instruments have different time horizons, and
we can allow for trends. In fact, this is a major advantage of using event trees;
we can allow for any kind of depenencies.
2.7.3.2 The individual scenario problems

A major advantage of using scenario aggregation on this problem (and many
others) is that the individual scenario problems become simple to interpret,
and, if the underlying optimization problem has structure, that this structure
is maintained. In its simplest form the scenario problems are in this case
generalized networks, a problem class for which there are efficient codes. We
refer to Chapter 6 for a more detailed look at networks.
Figure 16 shows an example of what a scenario problem may look like.
This is one of several ways to represent the problem. Periods (stages) run
Bonds
Stocks
Real estate
Cash
Stage 1 Stage 2 Stage 3
Figure 16 Network describing possible investments for a single scenario.
horizontally. For each stage we first have a column with one node for each
instrument. In the example these are four investment categories. The arcs
entering from the left brings the initial portfolio into the model, measured in
the amount of money held in each category. The arcs that run horizontally
between nodes of the same category represent investments held from one
period to the next. The node which is alone in a column represents trading.
Arcs that run to or from this cash node, represent the selling or buying of
instruments. A stage consists of one column of 4 nodes plus the single cash
node.
We mentioned that this is a generalized network. That means that the
amount of money that enters an arc is not the same as the amount that leaves
it. For example, if you put money in the bank, and the interest rate is 10%,
then the flow into the arc is multiplied by 1.1 to produce the flow out. This
parameter is called the multiplier of the arc. This way the investment generally
increases over time. For most categories this parameter is uncertain, normally
with a mean greater than one. This is how we represent uncertainty in the
model. For a given scenario, these multipliers are known.
Arcs going to the cash trading node from all nodes but the cash node to its
left, will have multiplyers less than 1 to represent variable transaction costs.
The arcs that leave the cash trading node have the same multipliers as the
horizontal arcs for the same investment categories, reduced (deterministically)
for variable transaction costs. Fixed transaction costs are not hard to model,
but they would produce very difficult models to solve.
We can also have arcs going backwards in time. They represent borrowings.
Since you must pay interest on borrowings (deterministic or stochastic), these
arcs have multipliers less than one, meaning that if you want 100 USD now,
you must pay back more than 100 USD in a later time period.
DYNAMIC SYSTEMS 147
If we transform all investments into cash in the last period (maybe without
transaction costs) a natural objective is to maximize this value. This way
we have set the scene for using scenario aggregation on a financial model. It
appears that these models are very promising.
2.7.3.3 Practical considerations

First, it is important to observe that most likely, a realistically sized model
will be far too large to solve with today’s algorithms and computers. Hence,
it must be reduced in size, or simplified in some other way. For problem size
reductions, we refer to Section 3.4, where different approaches are discussed.
An other possibility is to resort to sampling schemes, ending up with statistical
convergence statements. This is discussed to some extent in Sections 3.8
and 3.9.
The model we have presented for the flow of funds is also, most likely,
too simple. Usually, legal considerations as well as investment policies of the
company will have to be introduced. However, that will be problem dependent,
and cannot be discussed here.
2.8 Hydro power production
The production of electric power from rivers and reservoirs represents an area
where stochastic programming methodology has been used for a long time.
The reason is simply that the environment in which planning must take place
is very uncertain. In particular, the inflow of water to the rivers and reservoirs
vary a lot both in the short and long term. This is caused by variation in
rainfall, but even more by the uncertainty related to the time of snow melting
in the spring. Furthermore, the demand for power is also random, depending
on such as temperature, the price of oil, and general economic conditions. The
actual setting for the planners will vary a lot from country to country. Norway,
with close to 100% of her electricity coming from hydro, is in a very different
situation from for example France with a high dependency on nuclear power,
or the US with a more mixed system. In addition, general market regulations
will affect modeling. Some countries have strictly regulated markets, others
have full deregulation and competition.
We shall now present a very simple model for electricity production. As an
example, we shall assume that electricity can be sold at fixed prices, which
could be interpreted as if we were a small producer in a competitive market.
It is worth noting that in many contexts it is necessary to consider also price
as random. In still other contexts, the goal is not at all to maximize profit,
but rather to satisfy demand. So, there are many variations of this problem.
We shall present one in order to illustrate the basics.
2.8.1 A small example

Let us look at a rather simple version of the problem. Let there by two
reservoirs, named A and B. The reservoirs are connected by a river, with
A being upstream of B. We shall assume that the periods are long enough
for water released from reservoir A in a period to reach reservoir B in the
same period. This implies either that the reservoirs are close or that the time
periods are long. It will be easy to change the model if it is more reasonable
to let the water arrive in reservoir B in the next period. We shall also assume
that both water released for production and water spilled from reservoir A
(purposely or as a result of a full reservoir) will reach reservoir B. Sometimes
spilled water is lost.
There are three sets of variables. Let
vij be the volumes of water in reservoir i, (i ∈ {A,B}) at the beginning of

period j, (j ∈ {0, 1, 2, . . . , T }) (here vi0 is given to be the initial volume
in each reservoir), and
uij be the volume of water in reservoir i released to power station i during

period j, and
rij be the amount of water spilling out of reservoir i in period j.
There is one major set of parameters for the constraints, which we

eventually will interpret as random variables. Let
qij be the volume of water flowing into reservoir i during period j.
Bounds on the variables uij and vij are also given. They typically represent
such as reservoir size, production capacity and legal restrictions.
uij ∈ [ui , ui ] and

vij ∈ [v ij , v ij ] for i ∈ {A, B} and j ∈ {1, 2, . . . , T }.
The reservoir balance equations in terms of volumes then become for

i = 1, . . . , T :
vAi = vA,i−1 + qAi − uAi − rAi

vBi = vB,i−1 + qBi + uAi + rAi − uBi − rBi .
What we now lack is a description of the objective function plus the end
effects. To facilitate that, let
DYNAMIC SYSTEMS 149
inflow
inflow
Figure 17 Simple river system with two reservoirs and two plants.
cj denote the value of the electricity generated from one unit of water in
period j,
Φ(vAT , vBT ) denote the value function for water at the final stage, and
φi denote the marginal values of water at the final stage (the partial
derivatives of Φ).
The objective we wish to maximize then has the form (assuming discounting
is contained in cj )

T
cj (uAj + uBj ) + Φ(vAT , vBT ).
j=1
The function Φ is very important in this model. The reason is that a major
feature of the model is that it distributes water between the periods covered
by the model, and all later periods. Hence, if Φ underestimates the future
value of water, the model will most likely suggest an empty reservoir after
stage T , and if it is set too high, the model will almost only save water. The
estimation of Φ, which is normally done in a model with a very long time
horizon (often infinite) has been subjected to research for several decades.
Very often it is the partial derivatives φi that are estimated, rather than the
function itself.
Now, if the inflow is random, we can set up an event tree. Most likely, the
inflows to the reservoirs are dependent, and if the periods are short, there
may also be dependence over time. The model we are left with can, at least
in principle, be solved with scenario aggregation.
2.8.2 Further developments

The model shown above is very simplified. Modeling of real systems must take
into account a number of other aspects as well. In this section, we list some
of them to give you a feeling for what may happen.
First, these models are traditionally set in a context where the major goal is
to meet demand, rather than maximize profit. In a pure hydro based system,
the goal is then to obtain as much energy as possible from the available water
(which of course is still uncertain). In a system with other sources for energy
as well, we also have to take into account the cost of these sources, for example
natural gas, oil or nuclear power.
Obviously, in a model as simple as ours, maximizing the amount of energy
obtained from the available water resources makes little sense, as we have
(implicitly) assumed that the amount of energy we get from 1m3 of water
is fixed. The reality is normally different. First, the turbines are not equally
efficient at all production levels. They have some optimal (below maximal)
production levels where the amount of energy per m3 water is optimized.
Generally, the function describing energy production as a result of water usage
in a power plant with several turbines is neither convex nor monotone. In
particular, the non-convexity is serious. It stems from physical properties of
the turbines.
But there is more than that. The energy production also depends on the
head (hydrostatic pressure) that applies at a station during a period. It is
common to measure water pressure as the height of the water column having
the given pressure at the bottom. This is particularly complicated if the
water released from one power plant is submerged in the reservoir of the
downstream power plant. In this case the head of the upper station will depend
on the reservoir level of the lower station, generating another source of non-
convexities.
Traditionally, these models have been solved using stochastic dynamic
programming. This can work reasonably well as long as the dimension of
the state space is small. A requirement in stochastic dynamic programming
is that there is independence between periods. Hence, if water inflow in one
period (stage) is correlated to that of the previous period(s), the state space
must be expanded to contain the inflow in these previous period(s). If this
happens, SDP is soon out of business.
Furthermore, in deregulated markets it may be necessary to include price
as a random variable. Price is correlated to inflow in the present period, but
even more to inflow in earlier periods through the reservoir levels. This creates
dependencies which are very hard to tackle in SDP.
Hence, researchers have turned to other methods, for example scenario
aggregation, where dependencies are of no concern. So far, it is not clear
how successful this will be.
DYNAMIC SYSTEMS 151
2.9 The Value of Using a Stochastic Model
We have so far embarked on formulating and solving stochastic programming

models, without much concern about whether or not that is a worthwhile
thing to do. Most decision problems are certainly affected by randomness,
but that is not the same as saying that the randomness should be introduced
into a model. We all know that the art of modelling amounts to describing
the important aspects of a problem, and dropping the unimportant ones. We
must remember that randomness, although present in the situation, may turn
out to be one of the unimportant issues.
We shall now, briefly, outline a few approaches for evaluating the importance
of randomness. We shall see that randomness can be (un)important in several
different ways.
2.9.1 Comparing the Deterministic and Stochastic Objective Values

The most straightforward way to check if randomness is unimportant is
to compare the optimal objective value of the stochastic model with the
corresponding optimal value of the deterministic model (probably produced
by replacing all random variables by their means).
When we compare the optimal objective values (and also the solutions) in
these two cases, we must be aware that what we are observing is composed of
several elements. First, while the deterministic solution has one decision for
each time period, the stochastic solution “lives”on a tree, as we have discussed
in this chapter. The major point here is that the deterministic model has lost
all elements of dynamics (it has several time periods, but all decisions are
made here and now). Therefore decisions that have elements of options in
them will never be of any use. In a deterministic world there is never a need
to do something just in case.
Secondly, replacing random variables by their means will in itself have an
effect, as we shall discuss in much more detail in the next chapter.
Therefore, even if these two models come out with about the same optimal
objective value, one does not really know much about whether or not it is
wise to work with a stochastic model. These models are simply too different
to say much in most situations.
From this short discussion, you may have observed that there are really two
major issues when solving a model. One is the optimal objective value, the
other the optimal solution. It depends on the situation which of these is more
important. Sometimes one’s major concern is if one should do something or
not; in other cases the question is not if one should do something, but what
one should do.
When we continue, we shall be careful, and try to distinguish these cases.
2.9.2 Deterministic Solutions in the Event Tree

To illustrate this idea we shall use the following example.
Example 2.4 Assume that we have a container that can take up to 10 units,
and that we have two possible items that can be put into the container. The
items are called A and B, and some of their properties are given in Table 2.
Table 2 Properties of the two items A and B.
Item Value Minimum size Maximum size

A 6 5 8
B 4 3 6
The goal is to fill the container with as valuable items as possible. However,
the size of an item is uncertain. For simplicity, we assume that each item can
have two different sizes, as given in Table 2. All sizes occur with the same
probability of 0.5. As is always the case with a stochastic model, we must
decide on how the stages are defined. We shall assume that we must pick an
item before we learn its size, and that once it is picked, it must be put into
the container. If the container becomes overfull, we obtain a penalty of 2 per
unit in excess of 10. We have the choice of picking only one item, and they
can be picked in any order.
A stochastic decision tree for the problem is given in Figure 18, where we
have already folded back and crossed out nonoptimal decisions. We see that
the expected value is 7.5. That is obtained by first picking item A, and then,
if item A turns out to be small, also pick item B. If item A turns out to be
large, we choose not to pick item B.
2
If we assume that the event tree (or the stochastic part of the stochastic
decision tree) is a fair description of the randomness of a model, the following
simple approach gives a reasonable measure of how good the deterministic
model really is. Start in the root of the event tree, and solve the deterministic
model. (Probably this means replacing random variables by their means.
However, this approach can be used for any competing deterministic model.)
Take that part of the deterministic solution that corresponds to the first stage
of the stochastic model, and let it represent an implementable solution in the
root of the event tree. Then go to each node at level two of the event tree and
repeat the process. Taking into consideration what has happened in stage 1
(which is different for each node), solve the deterministic model from stage
DYNAMIC SYSTEMS 153
Figure 18 Stochastic decision tree for the container problem.
2 onwards, and use that part of the solution that corresponds to stage 2 as
an implementable solution. Continue until you have reached the leaves of the
event tree.
This is a fair comparison, since even people who prefer deterministic models
resolve them as new information becomes available (represented by the event
tree). In this setting we can compare both decisions and (expected) optimal
objective values. What we may observe is that although the solutions are
different, the optimal values are almost the same. If that is the case, we are
observing flat objective functions with many (almost) optimal solutions. If we
observe large differences in objective values, we have a clear indication that
solving a stochastic model is important.
Let us return to Example 2.4. Let the following simple deterministic
algorithm be an alternative to the stochastic programming approach in
Figure 18. Consider all items not put into the container so far. For each item,
calculate the value of adding it to the container, given that it has its expected
size. If at least one item adds a positive value to the content of the container,
pick the one with the highest added value. Then put it in, and repeat.
This is not meant to be a specially efficient algorithm—it is only presented
for its simplicity to help us make a few points. If we apply this algorithm to
our case, we see that with an empty container, item A will add 6 to the value
of the container and item B will add 4. Hence we pick item A. The algorithm
will next determine if B should be picked or not. However, for the comparison
between the deterministic and stochastic approach, it suffices to observe that

item A is picked first. This coincides with the solution in Figure 18.
Next we observe the size of A. If it is small, there is still room for 5 units
in the container. Since B has an expected size of 4.5, it will add 4 to the
value of the container, and will therefore be picked. On the other hand, if A
turns out to be large, there is only room for 2 more units, and B will add
4 − 2.5 × 2 = −1 to the value, and it will therefore not be picked. Again, we
get exactly the same solution as in Figure 18.
So what have we found out? We have seen that for this problem, with its
structure and data, the deterministic approach was as good as the stochastic
approach. However, it is not possible to draw any general conclusions from
this. In fact, it illustrates a very important point: it is extremely difficult
to know if randomness is important before we have solved the problem and
checked the results. But, in this special case, anyone claiming that using
stochastic decision trees on this problem was like shooting sparrows with
cannons will be proved correct.
2.9.3 Expected Value of Perfect Information

For simplicity, assume that we have a two-stage model. Now compare the
optimal objective value of the stochastic model with the expected value of
the wait-and-see solutions. The latter is calculated by finding the optimal
solution for each possible realization of the random variables. Clearly, it is
better to know the value of the random variable before making a decision than
having to make the decision before knowing. The difference between these two
expected objective values is called the expected value of perfect information
(EVPI), since it shows how much one could expect to win if one were told
what would happen before making one’s decisions. Another interpretation is
that this difference is what one would be willing to pay for that information.
What does it mean to have a large EVPI? Does it mean that it is important
to solve a stochastic model? The answer is no! It shows that randomness plays
an important role in the problem, but it does not necessarily show that a
deterministic model cannot function well. By resorting to the set-up of the
previous subsection, we may be able to find that out. We can be quite sure,
however, that a small EVPI means that randomness plays a minor role in the
model.
In the multistage case the situation is basically the same. It is, however,
possible to have a very low EVPI, but at the same time have a node far down
in the tree with a very high EVPI (but low probability.)
Let us again turn to Example 2.4. Table 3 shows the optimal solutions for
the four cases that can occur, if we make the decisions after the true values
have become known. Please check that you agree with the numbers.
With each case in Table 3 equally probable, the expected value of the wait-
DYNAMIC SYSTEMS 155
Table 3 The four possible wait-and-see solutions for the container problem in
Example 2.4.
Size of A Size of B Solution Value

5 3 A, B 10
5 6 A, B 8
8 3 A, B 8
8 6 A 6
and-see solution is 8, which is 0.5 more than what we found in Figure 18.
Hence EVPI equals 0.5; The value of knowing the true sizes of the items
before making decisions is 0.5. This is therefore also the maximal price one
would pay to know this.
What if we were offered to pay for knowing the value of A or B before
making our first pick? In other words, does it help to know the size of for
example item B before choosing what to do? This is illustrated in Figure 19.
Figure 19 Stochastic decision tree for the container problem when we know
the size of B before making decisions.
We see that the EVPI for knowing the size of item B is 0.5, which is the
same as that for knowing both A and B. The calculation for item A is left as
an exercise.
Example 2.5 Let us conclude this section with another similar example. You
are to throw a die twice, and you will win 1 if you can guess the total number
of eyes from these two throws. The optimal guess is 7 (if you did not know
that already, check it out!), and that gives you a chance of winning of 16 . So
the expected win is also 16 .
Now, you are offered to pay for knowing the result of the first throw. How
much will you pay (or alternatively, what is the EVPI for the first throw)? A
close examination shows that knowing the result of the first throw does not
help at all. Even if you knew, guessing a total of 7 is still optimal (but that
is no longer a unique optimal solution), and the probability that that will
happen is still 16 . Hence, the EVPI for the first stage is zero.
Alternatively, you are offered to pay for learning the value of both throws
before “guessing”. In that case you will of course make a correct guess, and
be certain of winning one. Therefore the expected gain has increased from 16
to 1, so the EVPI for knowing the value of both random variables is 56 . 2
As you see, EVPI is not one number for a stochastic program, but can
be calculated for any combination of random variables. If only one number is
given, it usually means the value of learning everything, in contrast to knowing
nothing.
References
[1] Bellman R. (1957) Dynamic Programming. Princeton University Press,

Princeton, New Jersey.
[2] Helgason T. and Wallace S. W. (1991) Approximate scenario solutions in
the progressive hedging algorithm. Ann. Oper. Res. 31: 425–444.
[3] Howard R. A. (1960) Dynamic Programming and Markov Processes. MIT
Press, Cambridge, Massachusetts.
[4] Nemhauser G. L. (1966) Dynamic Programming. John Wiley & Sons, New
York.
[5] Rockafellar R. T. and Wets R. J.-B. (1991) Scenarios and policy aggregation
in optimization under uncertainty. Math. Oper. Res. 16: 119–147.
[6] Schaefer M. B. (1954) Some aspects of the dynamics of populations
important to the management of the commercial marine fisheries. Inter-
Am. Trop. Tuna Comm. Bull. 1: 27–56.
[7] Wallace S. W. and Helgason T. (1991) Structural properties of the
progressive hedging algorithm. Ann. Oper. Res. 31: 445–456.
[8] Watson S. R. and Buede D. M. (1987) Decision Synthesis. The Principles
DYNAMIC SYSTEMS 157
and Practice of Decision Analysis. Cambridge University Press, Cambridge,

UK.
[9] Wets R. J.-B. (1989) The aggregation principle in scenario analysis and
stochastic optimization. In Wallace S. W. (ed) Algorithms and Model
Formulations in Mathematical Programming, pages 91–113. Springer-
Verlag, Berlin.
3
Recourse Problems
The purpose of this chapter is to discuss principal questions of linear recourse

problems. We shall cover general formulations, solution procedures and
bounds and approximations.
Figure 1 shows a simple example from the fisheries area. The assumption is
that we know the position of the fishing grounds, and potential locations for
plants. The cost of building a plant is known, and so are the distances between
grounds and potential plants. The fleet capacity is also known, but quotas,
and therefore catches, are only known in terms of distributions. Where should
the plants be built, and how large should they be?
This is a typical two-stage problem. In the first stage we determine which
plants to build (and how big they should be), and in the second stage we catch
and transport the fish when the quotas for a given year are known. Typically,
quotas can vary as much as 50% from one year to the next.
3.1 Outline of Structure
Let us formulate a two-stage stochastic linear program. This formulation differs

from (4.16) of Chapter 1 only in the randomness in the objective of the
recourse problem.
min cT x + Q(x)
s.t. Ax = b, x ≥ 0,
where
Q(x) = pj Q(x, ξ j )
j
and
Q(x, ξ) = min{q(ξ)T y | W (ξ)y = h(ξ) − T (ξ)x, y ≥ 0},
where pj is the probability
that ξ˜ = ξ j ,the jth realization of ξ,
˜ h(ξ) =

h0 + Hξ = h0 + i hi ξi , T (ξ) = T0 + i Ti ξi and q(ξ) = q0 + i qi ξi .
Nord-
Trøndelag
Sør-Trøndelag
Møre og Romsdal
Hedmark
Sogn og Fjordane
Oppland
Hordaland
Buskerud
Akershus
OSLO
Telemark
Østfold
Vest-
fold
Roga- Aust-
land Agder
Vest-
Agder
Figure 1 A map showing potential plant sites and actual fishing grounds for
Southern Norway and the North Sea.
The function Q(x, ξ) is called the recourse function, and Q(x) therefore the
expected recourse function.
In this chapter we shall look at only the case with fixed recourse, i.e. the
case where W (ξ) ≡ W . Let us repeat a few terms from Section 1.4, in order
to prepare for the next section. The cone pos W , mentioned in (4.17) of
Chapter 1, is defined by
pos W = {t | t = W y, y ≥ 0}.
The cone pos W is illustrated in Figure 2. Note that
W y = h, y ≥ 0 is feasible ⇐⇒ h ∈ pos W.
Recall that a problem has complete recourse if
pos W = Rm .
Among other things, this implies that
h(ξ) − T (ξ)x ∈ pos W for all ξ and all x.
But that is definitely more than we need in most cases. Usually, it is more
than enough to know that
RECOURSE PROBLEMS 161
W1 W2
W3
W4
Figure 2 The cone pos W for a case where W has three rows and four
columns.
h(ξ) − T (ξ)x ∈ pos W for all ξ and all x ≥ 0 satisfying Ax = b.

If this is true, we have relatively complete recourse. Of course, complete
recourse implies relatively complete recourse.
3.2 The L-shaped Decomposition Method
This section contains a much more detailed version of the material found
in Section 1.7.4. In addition to adding more details, we have now added
randomness more explicitly, and have also chosen to view some of the aspects
from a different perspective. It is our hope that a new perspective will increase
the understanding.
3.2.1 Feasibility
The material treated here coincides with step 2(a) in the dual decomposition
method of Section 1.7.4. Let the second-stage problem be given by
Q(x, ξ) = min{q(ξ)T y | W y = h(ξ) − T (ξ)x, y ≥ 0},

where W is fixed. Assume we are given an x̂ and should like to know if that x̂
˜ We assume
yields a feasible second-stage problem for all possible values of ξ.
Hξ
pos W
h0-T0 x^
Figure 3 Illustration showing that if infeasibility is to occur for a fixed x̂, it
must occur for an extreme point of the support of H ξ̃, and hence of ξ.˜ In this
example T (ξ) is assumed to be equal to T0 .
that ξ̃ has a rectangular and bounded support. Consider Figure 3. We have

there drawn pos W plus a parallelogram that represents all possible values of
h0 + H ξ̃ − T0 x̂. We have assumed that T (ξ) ≡ T0 , only to make the illustration
simpler.
Figure 3 should be interpreted as representing a case where H is a 2 × 2
matrix, so that the extreme points of the parallelogram correspond to the
extreme points of the support Ξ of ξ̃. This is a known result from linear
algebra, namely that if one polyhedron is a linear transformation of another
polyhedron, then the extreme points of the latter are maps of extreme points
in the first.
What is important to note from Figure 3 is that if the second-stage problem
is to be infeasible for some realizations of ξ̃ then at least one of these
realizations will correspond to an extreme point of the support. The figure
shows such a case. And conversely, if all extreme points of the support produce
feasible problems, all other possible realizations of ξ̃ will also produce feasible
problems. Therefore, to check feasibility, we shall in the worst case have to
check all extreme points of the support. With k random variables, and Ξ a k-
dimensional rectangle, we get 2k points. Let us define A to be a set containing
these points. In Chapter 5 we shall discuss how we can often reduce the number
of points in A without removing the property that if all points in A yield a
feasible second-stage problem, so will all other points in the support.
We shall next turn to another aspect of feasibility, namely the question

of how to decide if a given x = x̂ will yield feasible second-stage problems
for all possible values of ξ̃ in a setting where we are not aware of relatively
complete recourse. What we shall outline now corresponds to Farkas’ lemma
(Proposition 1.19, page 75). Farkas’ lemma states that
{y | W y = h, y ≥ 0} = ∅
if and only if
W T u ≥ 0 implies that hT u ≥ 0.
The first of these equivalent statements is just an alternative way of saying
that h ∈ pos W , which we now know means that h represents a feasible
problem.
By changing the sign of u, the second of the equivalent statements can be
rewritten as
W T u ≤ 0 implies that hT u ≤ 0.
or equivalently
hT t ≤ 0 whenever t ∈ {u | W T u ≤ 0}.
However, this may be reformulated as
{u | W T u ≤ 0} = {u | uT W y ≤ 0 for all y ≥ 0}
= {u | uT h ≤ 0 for all h ∈ pos W }.
The last expression defines the polar cone of pos W as
pol pos W = {u | uT h ≤ 0 for all h ∈ pos W }.
Using Figure 4, we can now restate Farkas’ lemma the following way. The
system W y = h, y ≥ 0, is feasible if and only if the right-hand side h
has a non-positive inner product with all vectors in the cone pol pos W , in
particular with its generators. Generators were discussed in Chapter 1 (see e.g.
Remark 1.6, page 69). The matrix W ∗ , containing as columns all generators
of pol pos W , is denoted the polar matrix of W .
We shall see in Chapter 5 how this understanding can be used to generate
relatively complete recourse in a problem that does not possess that property.
For now, we are satisfied by understanding that if we knew all the generators
of pol pos W , that is the polar matrix W ∗ , then we could check feasibility of a
second-stage problem by performing a number of inner products (one for each
generator), and if at least one of them gave a positive value then we could
conclude that the problem was indeed infeasible.
If we do not know all the generators of pol pos W , and we are not aware
of relatively complete recourse, for a given x̂ and all ξ ∈ A we must check
pos W
pol pos W
Figure 4 The polar of a cone.

for feasibility. We should like to check for feasibility in such a way that if the
given problem is not feasible, we automatically come up with a generator of
pol pos W . For the discussion, we shall use Figure 5.
We should like to find a σ such that
σ T t ≤ 0 for all t ∈ pos W.
This is equivalent to requiring that σ T W ≤ 0. In other words, σ should be

in the cone pol pos W . But, assuming that the right-hand side h(ξ) − T (ξ)x̂
produces an infeasible problem, we should at the same time require that
σ T [h(ξ) − T (ξ)x̂] > 0,
because if we later add the constraint σ T [h(ξ) − T (ξ)x] ≤ 0 to our problem,

we shall exclude the infeasible right-hand side h(ξ) − T (ξ)x̂ without leaving
out any feasible solutions. Hence we should like to solve
max{σ T (h(ξ) − T (ξ)x̂) | σ T W ≤ 0, σ ≤ 1},

σ
where the last constraint has been added to bound σ. We can do that, because
otherwise the maximal value will be +∞, and that does not interest us since
we are looking for the direction defined by σ. If we had chosen the 2 norm, the
maximization would have made sure that σ came as close to h(ξ) − T (ξ)x̂ as
possible (see Figure 5). Computationally, however, we should not like to work
with quadratic constraints. Let us therefore see what happens if we choose
the 1 norm. Let us write our problem differently to see the details better. To
do that, we need to let the unconstrained σ be replaced by σ 1 − σ 2 , where
σ 1 , σ 2 ≥ 0. We then get the following:
max{(σ 1 −σ 2 )T (h(ξ)−T (ξ)x̂) | (σ 1 −σ 2 )T W ≤ 0, eT (σ 1 +σ 2 ) ≤ 1, σ 1 , σ 2 ≥ 0},
where e is a vector of ones. To more easily find the dual of this problem, let
us write it down in a more standard format:
max(σ 1 − σ 2 )T (h(ξ) − T (ξ)x̂) dual variables
W T σ1 − W T σ2 ≤ 0 y
eT σ 1 + eT σ 2 ≤ 1 t
σ1 , σ2 ≥ 0
From this, we find the dual linear program to be
min{t | W y + et ≥ (h(ξ) − T (ξ)x̂), −W y + et ≥ −(h(ξ) − T (ξ)x̂), y, t ≥ 0}.
Note that if the optimal value in this problem is zero, we have W y =

h(ξ) − T (ξ)x̂, so that we do indeed have h(ξ) − T (ξ)x̂ ∈ pos W , contrary
pos W
h ( x ) - T ( x ) x$
s
Figure 5 Generation of feasibility cuts.
to our assumption. We also see that if t gets large enough, the problem is
always feasible. This is what we solve for all ξ ∈ A. If for some ξ we find a
positive optimal value, we have found a ξ for which h(ξ) − T (ξ)x̂ ∈ pos W ,
and we create the cut
σ T (h(ξ) − T (ξ)x) ≤ 0 ⇐⇒ σ T T (ξ)x ≥ σ T h(ξ). (2.1)

The σ used here is a generator of pol pos W , but it is not in general as close
to h(ξ) − T (ξ)x̂ as possible. This is in contrast to what would have happened
had we used the 2 norm. (See Example 3.1 below for an illustration of this
point.)
Note that if T (ξ) ≡ T0 , the expression σ T T0 x in (2.1) does not depend on
ξ. Since at the same time (2.1) must be true for all ξ, we can for this special
case strengthen the inequality by calculating

σ T T0 x ≥ σ T h0 + max σ T H t.
t∈Ξ
Since σ T T0 is a vector and the right-hand side is a scalar, this can conveniently
be written as −γ T x ≥ δ. The x̂ we started out with will not satisfy this
constraint.
Example 3.1 We present this little example to indicate why the 1 and 2
norms give different results when we generate feasibility cuts. The important
point is how the two norms limit the possible σ values. The 1 norm is given
in the left part of Figure 6, the 2 norm in the right part.
Figure 6 Illustration of the difference between the 1 and 2 norms when

generating feasibility cuts.
For simplicity, we have assumed that pol pos W equals the positive
quadrant, so that the constraints σ T W ≤ 0 reduce to σ ≥ 0. Since at the
same time σ ≤ 1, we get that σ must be within the shaded part of the two
figures.
For convenience, let us denote the right-hand side by h, and let σ =
(σ x , σ y )T , to reflect the x and y parts of the vector. In this example h =
(4, 2)T . For the 1 norm the problem now becomes.
max{4σ x + 2σ y σ x + σ y ≤ 1, σ ≥ 0}.
σ
The optimal solution here is σ = (1, 0)T . Graphically this can be seen from
the figure from the fact that an inner product equals the length of one vector
multiplied by the length of the projection of the second vector on the first. If
we take the h vector as the fixed first vector, the feasible σ vector with the
largest projection on h is σ = (1, 0)T .
For the 2 norm the problem becomes
max{4σ x + 2σ y (σ x )2 + (σ y )2 ≤ 1, σ ≥ 0}.

σ
-
The optimal solution here is σ = 15 (2, 1)T , which is a vector in the same
direction as h.
In this example we see that if σ is found using the 1 norm, it becomes a
generator of pol pos W , but it is not as close to h as possible. With the 2
norm, we did not get a generator, but we got a vector as close to h as possible.
procedure LP(W :matrix; b, q, y:vectors; feasible:boolean);

begin
if min{q T y |W y = b, y ≥ 0} is feasible then begin
feasible := true;
y is the optimal y’;
end
else feasible := f alse;
end;
Figure 7 LP solver.
3.2.2 Optimality
The material discussed here concerns step 1(b) of the dual decomposition
method in Section 1.7.4. Let us first note that if we have relatively complete
recourse, or if we have checked that h(ξ) − T (ξ)x ∈ pos W for all ξ ∈ A, then
the second-stage problem
min{q(ξ)T y | W y = h(ξ) − T (ξ)x, y ≥ 0}
is feasible. Its dual formulation is given by
max{π T (h(ξ) − T (ξ)x) | π T W ≤ q(ξ)T }.
As long as q(ξ) ≡ q0 , the dual is either feasible or infeasible for all x and ξ, since
x and ξ do not enter the constraints. We see that this is more complicated
if q is also affected by randomness. But even when ξ enters the objective
function, we can at least say that if the dual is feasible for one x and a
given ξ then it is feasible for all x for that value of ξ, since x enters only
the objective function. Therefore, from standard linear programming duality,
since the primal is feasible, the primal must be unbounded if and only if the
dual is infeasible, and that would happen for all x for a given ξ, if randomness
affects the objective function. If q(ξ) ≡ q0 then it would happen for all x
and ξ. Therefore we can check in advance for unboundedness, and this is
particularly easy if randomness does not affect the objective function. Note
that this discussion relates to Proposition 1.18. Assume we know that our
problem is bounded.
procedure master(K, L:integer;x, θ:real;feasible:boolean);

begin
if L > 0 then begin
⎛ ⎛ c ⎞ ⎛ x̂ ⎞ ⎞
) * ) * +
⎜ A 0 0 0 0 b ⎜ 1 ⎟ ⎜ θ− ⎟ ⎟
LP⎜ ⎜ ⎟ ⎜ ⎟
⎝ −Γ 0 0 −I 0 , ∆ , ⎝ −1 ⎠ , ⎝ θ ⎠ , feasible⎠;
⎟
−β e −e 0 −I α 0 s1
0 s2
if (feasible) then θ̂ := θ+ − θ− ;
end
else begin

A 0 b c x̂
LP , , , , feasible ;
−Γ −I ∆ 0 s
if feasible then θ̂ := −∞;
end;
end;
Figure 8 Master problem solver for the L-shaped decomposition method.
Now consider
Q(x) = pj Q(x, ξ j ),
j
with
Q(x, ξ) = min{q(ξ)T y | W y = h(ξ) − T (ξ)x, y ≥ 0}.
It is clear from standard linear programming theory that Q(x, ξ) is piecewise
linear and convex in x (for fixed ξ). Provided that q(ξ) ≡ q0 , Q(x, ξ) is
also piecewise
linear and convex in ξ (for fixed x). (Remember that T (ξ) =
T0 + Ti ξi .) Similarly, if h(ξ) − T (ξ)x ≡ h0 − T0 x, while q(ξ) = q0 + i qi ξi ,
then, from duality, Q(x, ξ) is piecewise linear and concave in ξ. Each linear
piece corresponds to a basis (possibly several in the case of degeneracy).
Therefore Q(x), being a finite sum of such functions, will also be convex and
piecewise linear in x. If, instead of minimizing, we were maximizing, convexity
and concavity would change places in the statements.
In order to arrive at an algorithm for our problem, let us now reformulate
the latter by introducing a new variable θ:
min cT x + θ
s.t. Ax = b,
θ ≥ Q(x),
−γkT x ≥ δk for k = 1, . . . , K,
x ≥ 0,
procedure feascut(A:set; x̂:real; newcut:boolean; K:integer);

begin
A := A; newcut := false;
while A = ∅ and not (newcut) do begin

⎛ , ξ); A := A \ {ξ};
pickξ(A ⎛ ⎞ ⎛ ⎞ ⎞
0 ŷ
⎜ W e −I 0 h(ξ) − T (ξ)x̂ ⎜ 1 ⎟ ⎜ t̂ ⎟ ⎟
LP⎝ , , ⎝ ⎠ , ⎝ ⎠ , feasible⎠;
−W e 0 −I −h(ξ) + T (ξ)x̂ 0 s1
0 s2
newcut := (t̂ > 0);
if newcut then begin
(* Create a feasibility cut—see page 161. *)
K := K + 1;
T
Construct the cut −γK x ≥ δK ;
end;
end;
end;
Figure 9 Procedure used to find feasibility cuts.
where, as before,
Q(x) = pj Q(x, ξ j )
j
and
Q(x, ξ) = min{q(ξ)T y | W y = h(ξ) + T (ξ)x, y ≥ 0}.
Of course, computationally we cannot use θ ≥ Q(x) as a constraint since
Q(x) is only defined implicitly by a large number of optimization problems.
Instead, let us for the moment drop it, and solve the above problem without
it, simply hoping it will be satisfied (assuming so far that all feasibility cuts
−γkT x ≥ δk are there, or that we have relatively complete recourse). We then
get some x̂ and θ̂ (the first time θ̂ = −∞). Now we calculate Q(x̂), and then
check if θ̂ ≥ Q(x̂). If it is, we are done. If not, our x̂ is not optimal—dropping
θ ≥ Q(x) was not acceptable.
Now
Q(x̂) = pj Q(x̂, ξ j ) = pj q(ξ j )T y j
j j
where y is the optimal second-stage solution yielding Q(x̂, ξ j ). But, owing to

j
linear programming duality, we also have

pj q(ξ j )T y j = pj (π̂ j )T [h(ξ j ) − T (ξ j )x̂],
j j
procedure L-shaped;
begin
K := 0, L := 0;
θ̂ := −∞
LP(A, b, c, x̂, feasible);
stop := not (feasible);
while not (stop) do begin
feascut(A, x̂,newcut,K);
if not (newcut) then begin
Find Q(x̂);
stop := (θ̂ ≥ Q(x̂));
if not (stop) then begin
(* Create an optimality cut—see page 168. *)
L := L + 1;
T
Construct the cut −βL x + θ ≥ αL ;
end;
end;
if not (stop) then begin
master(K, L, x̂, θ̂,feasible);
end;
end;
end;
Figure 10 The L-shaped decomposition algorithm.
where π̂ j is the optimal dual solution yielding Q(x̂, ξ j ). The constraints in the
dual problem are, as mentioned before, π T W ≤ q(ξ j )T , which are independent
of x. Therefore, for a general x, and corresponding optimal vectors π j (x), we
have

Q(x) = pj (π j (x))T [h(ξ j ) − T (ξ j )x] ≥ pj (π̂ j )T [h(ξ j ) − T (ξ j )x],
j j
since π̂ is feasible but not necessarily optimal, and the dual problem is a
maximization problem. Since what we dropped from the constraint set was
θ ≥ Q(x), we now add in its place

θ≥ pj (π̂ j )T [h(ξ j ) − T (ξ j )x] = α + β T x,
j
or
−β T x + θ ≥ α.
Q(x)
x
cx+θ
cut 5
0 x3 x1
x2 cut 3 cut 1
cut 2
(x , θ )
5 5
cut 4
(x4 ,θ )
4
Figure 11 Example of the progress of the L-shaped decomposition algorithm.
Since there are finitely many feasible bases coming from the matrix W , there
are only finitely many such cuts.
We are now ready to present the basic setting of the L-shaped decomposition
algorithm. It is shown in Figure 10. To use it, we shall need a procedure
that solves LPs. It can be found in Figure 7. Also, to avoid too complicated
expressions, we shall define a special procedure for solving the master problem;
see Figure 8. Furthermore, we refer to procedure pickξ(A, ξ), which simply
picks an element ξ from the set A, and, finally, we use procedure feascut
which is given in Figure 9. The set A was defined on page 162.
In the algorithms to follow, let −Γx ≥ ∆ represent the K feasibility
cuts −γkT x ≥ δk , and let −βx + Iθ ≥ α represent the L optimality cuts
−βlT x + θ ≥ αl . Furthermore, let e be a column of 1s of appropriate size.
The example in Figure 11 can be useful in understanding the L-shaped

decomposition algorithm. The five first solutions and cuts are shown. The
initial x̂1 was chosen arbitrarily. Cuts 1 and 2 are feasibility cuts, and the rest
optimality cuts. θ̂1 = θ̂2 = θ̂3 = −∞. To see if you understand this, try to
find (x̂6 , θ̂6 ), cut 6 and then the final optimal solution.
3.3 Regularized Decomposition
As mentioned at the end of Section 1.7.4, the recourse problem (for a discrete
distribution) looks like
K ⎫
min{cT x + i=1 pi (q i )T y i } ⎪
⎪
⎪
⎪
s.t. Ax =b ⎬
T i x + W y i = hi , i = 1, · · · , K (3.1)
⎪
⎪
x ≥ 0, ⎪
⎪
⎭
y i ≥ 0, i = 1, · · · , K.
To use the multicut method mentioned in Section 1.7.4, we simply have

to introduce feasibility and optimality cuts for all the recourse functions
fi (x) := min{(q i )T y i | W y i = hi − T i x, y i ≥ 0}, i = 1, · · · , K, until the
overall procedure has converged.
In general, with the notation of the previous section, these cuts have the
form
γ T x + δ ≤ 0, where γ = −T T σ, δ = hT σ, (3.2)
γ T x + δ ≤ θ, where γ = −T T π̂, δ = f (x̂) − γ T x̂, (3.3)
where (3.2) denotes a feasibility cut and (3.3) denotes an optimality cut, the σ
and π̂ resulting from step 2 of the dual decomposition method of Section 1.7.4,
as explained further in Section 3.2. Of course, the matrix T and the right-hand
side vector h will vary, depending on the block i for which the cut is derived.
One cycle of a multicut solution procedure for problem (3.1) looks as follows:
Let B1i = {(x, θ1 , · · · , θK ) | · · ·}, i = 1, · · · , K, be feasible for the cuts

generated so far for block i (obviously for block i restricting only (x, θi )).
Given B0 = {(x, θ) | Ax = b, x ≥ 0, θ ∈ IRK } and the sets B1i , solve the
master program

K +
K ,
T i
min c x + p θi (x, θ1 , · · · , θK ) ∈ B0 ∩ B1i , (3.4)

i=1 i=1
yielding (x̂, θ̂1 , · · · , θ̂K ) as a solution. With this solution try to construct
further cuts for the blocks.
• If there are no further cuts to generate, then stop (optimal solution);

• otherwise repeat the cycle.
The advantage of a method like this lies in the fact that we obviously make
use of the particular structure of problem (3.1) in that we have to deal in the
master program only with n + K variables instead of n + i ni , if y i ∈ IRni .
The drawback is easy to see as well: we may have to add very many cuts,
and so far we have no reliable criterion to drop cuts that are obsolete for
further iterations. Moreover, initial iterations are often inefficient. This is not
surprising, since in the master (3.4) we deal only with
θi ≥ max[(γ ij )T x + δij ]
j∈Ji
for Ji denoting the set of optimality cuts generated so far for block i with the
related dual basic solutions π̂ ij according to (3.3), and not, as we intend to,
with
θi ≥ fi (x) = max[(γ ij )T x + δij ]
j∈Jî
where Jî enumerates all dual feasible basic solutions for block i. Hence
we are working in the beginning with a piecewise linear convex function
(maxj∈Ji [(γ ij )T x + δij ]) supporting fi (x) that does not sufficiently reflect the
shape of fi (see e.g. Figure 26 of Chapter 1, page 78). The effect may be—and
often is—that even if we start a cycle with an (almost) optimal first-stage
solution x of (3.1), the first-stage solution x̂ of the master (3.4) may be far
away from x , and it may take many further cycles to come back towards x .
The reason for this is now obvious: if the set of available optimality cuts, Ji , is
a small subset of the collection Jî then the piecewise linear approximation of
fi (x) may be inadequate near x . Therefore it seems desirable to modify the
master program in such a way that, when starting with some overall feasible
first-stage iterate z k , its solution xk does not move too far away from z k .
Thereby we can expect to improve the approximation of fi (x) by an optimality
cut for block i at xk . This can be achieved by introducing into the objective of
the master the term x− z k 2 , yielding a so-called regularized master program

K +K ,
1 k 2 T i
min x − z + c x + p θi (x, θ1 , · · · , θK ) ∈ B0 ∩ B1i ,
2ρ
i=1 i=1
(3.5)
with a control parameter ρ > 0. To avoid too many constraints in (3.5), let
us start with some z 0 ∈ B0 such that fi (z 0 ) < ∞ ∀i and G0 being the feasible
set defined by the first-stage equations Ax = b and all optimality cuts at z 0 .
Hence we start (for k = 0) with the reduced regularized master program

K
1
min x − z k 2 + cT x + pi θi (x, θ1 , · · · , θK ) ∈ Gk . (3.6)
2ρ
i=1
Observe that the objective of (3.6) implicitly contains the function1
F̂ (x) = cT x + min{pT θ | (x, θ) ∈ Gk },

θ
which, according to the above discussion, is a piecewise linear convex function

supporting from below our original piecewise linear objective
F (x) = cT x + p T
f (x)
= cT x + pi fi (x).
i
Excluding by assumption degeneracy in the constraints defining Gk , a point

(x, θ) ∈ IRn+K is a vertex, i.e. a basic solution, of Gk iff (including the first-
stage equations Ax = b) exactly n + K constraints are active (i.e. satisfied
as equalities), owing to the simple fact that a point in IRn+K is uniquely
determined by the intersection of n + K independent hyperplanes.2 In the
following we sometimes want to check whether at a certain overall feasible
x̂ ∈ IRn the support function F̂ has a kink, which in turn implies that for
θ̂ ∈ arg minθ {pT θ | (x̂, θ) ∈ Gk } at (x̂, θ̂) we have a vertex of Gk . Hence we
have to check whether at (x̂, θ̂) exactly n + K constraints are active.
Having solved (3.6) with a solution xk , and xk not being overall feasible,
we just add the violated constraints (either xi ≥ 0 from the first-stage or
the necessary feasibility cuts from the second stage) and resolve (3.6). If xk is
overall feasible, we have to decide whether we maintain the candidate solution
z k or whether we replace it by xk . As shown in Figure 12, there are essentially
three possibilities:
• F (xk ) = F̂ (xk ), i.e. the supporting function coincides at xk with the true
objective function (see x1 in Figure 12);
• F̂ (xk ) < F (xk ), but at xk there is a kink of F̂ and the decrease of the
true objective from z k to xk is ‘substantial’ as compared with the decrease
F̂ (xk ) − F̂ (z k ) = F̂ (xk ) − F (z k ) (< 0) (we have F̂ (z k ) = F (z k ) in view
of the overall feasibility of z k ); more precisely, for some fixed µ ∈ (0, 1),
F (xk )− F (z k ) ≤ (1 − µ)[F̂ (xk )− F (z k )] (see x2 in Figure 12 with µ = 0.75);
• neither of the two above situations arises (see x3 in Figure 12).
1With p = (p1 , · · · , pK )T .
2 Recall that in IRn+K never more than n + K independent hyperplanes intersect at one
point.
Figure 12 Keeping or changing the candidate solutions in QDECOM.
In these cases we should decide respectively

• z 2 := x1 , observing that no cut was added, and therefore keeping z 1
unchanged would block the procedure;
• z 3 := x2 , realizing that x2 is “substantially” better than z 2 —in terms of
the original objective—and that at the same time F̂ has a kink at x2 such
that we might intuitively expect—thus clearly making use of a heuristic
argument—to make a good step forward towards the optimal kink of the
true objective;
• z 4 := z 3 , since—neither rationally nor heuristically—can we see any
convincing reason to change the candidate solution. Hence it seems
preferable to first improve the approximation of F̂ to F by introducing
the necessary optimality cuts.
After these considerations, motivating the measures to be taken in the
various steps, we want to formulate precisely one cycle of the regularized
decomposition method (RD), which with

K
F (x) := cT x + pi fi (x)
i=1
for µ ∈ (0, 1), is described as follows.

Step 1 Solve (3.6) at z k , getting xk as first-stage solution and θk =
k T
(θ1k , · · · , θK ) as recourse approximates. If, for F̂k := cT xk + pT θk ,
F̂k = F (z ) then stop (z k is an optimal solution of (3.1)). Otherwise,
k
go to step 2.
Step 2 Delete from (3.6) some constraints that are inactive at (xk , θk ) such
that no more than n + K constraints remain.
Step 3 If xk satisfies the first-stage constraints (i.e. xk ≥ 0) then go to
step 4; otherwise add to (3.6) no more than K violated (first-stage)
constraints, yielding the feasible set Gk+1 , put z k+1 := z k , k := k + 1,
and go to step 1.
Step 4 For i = 1, · · · , K solve the second-stage problems at xk and
(a) if fi (xk ) = ∞ then add to (3.6) a feasibility cut;
(b) otherwise, if fi (xk ) > θik then add to (3.6) an optimality cut.
Step 5 If fi (xk ) = ∞ for at least one i then put z k+1 := z k and go to step 7.
Otherwise, go to step 6.
Step 6 If F (xk ) = F̂k , or else if F (xk ) ≤ µF (z k ) + (1 − µ)F̂k and if exactly
n + K constraints were active at (xk , θk ), then put z k+1 := xk ;
otherwise, put z k+1 := z k .
Step 7 Determine Gk+1 as resulting from Gk after deleting and adding
constraints due to step 2 and step 4 respectively. With k := k + 1,
go to step 1.
It can be shown that this algorithm converges in finitely many steps.

The parameter ρ can be controlled during the procedure so as to increase
it whenever steps (i.e. xk − z k ) seem too short, and decrease it when
F (xk ) > F (z k ).
3.4 Bounds
Section 3.2 was devoted to the L-shaped decomposition method. We note that
the deterministic methods very quickly run into dimensionality problems with
respect to the number of random variables. With much more than 10 random
variables, we are in trouble.
This section discusses bounds on stochastic problems. These bounds can be
useful and interesting in their own right, or they can be used as subproblems
in larger settings. An example of where we might need to bound a problem,
and where this problem is not a subproblem, is the following. Assume that a
company is facing a decision problem. The decision itself will be made next
year, and at that time all parameters describing the problem will be known.
However, today a large number of relevant parameters are unknown, so it
is difficult to predict how profitable the operation described by the decision

problem will actually be. It is desired to know the expected profitability of the
operation. The reason is that, for planning purposes, the firm needs to know
the expected activities and profits for the next year. Given the large number of
uncertain parameters, it is not possible to calculate the exact expected value.
However, using bounding techniques it may be possible to identify an interval
that contains the expected value. Technically speaking, one needs to find the
expected value of the “wait-and-see” solution discussed in Chapter 1, and also
in Example 2.4. Another example, which we shall see later in Section 6.6, is
that of calculating the expected project duration time in a project consisting
of activities with random durations.
Bounding methods are also useful if we wish to use deterministic
decomposition methods (such as the L-shaped decomposition method or
scenario aggregation), on problems with a large number of random variables.
That will be discussed later in Section 3.5.2. One alternative to bounding
involves the development of approximations using stochastic methods. We
shall outline two of them later, they are called stochastic decomposition
(Section 3.8) and stochastic quasi-gradient methods (Section 3.9).
As discussed above, bounds can be used either to approximate the expected
value of some linear program or to bound the second-stage problem in a
two-stage problem. These two settings are principally the same, and we
shall therefore consider the problem of finding the expected value of a linear
program. We shall discuss this in terms of a function φ(ξ), which in the two-
stage case represents Q(x̂, ξ) for a fixed x̂. To illustrate, we shall look at the
refinery example of Section 1.3. The problem is repeated here for convenience:
⎫
φ(ξ) = “ min ” {2xraw1 + 3xraw2 } ⎪ ⎪
⎪
⎪
s.t. xraw1 + xraw2 ≤ 100, ⎪
⎪
⎬
2xraw1 + 6xraw2 ≥ 180 + ξ1 ,
(4.1)
3xraw1 + 3xraw2 ≥ 162 + ξ2 , ⎪ ⎪
⎪
⎪
xraw1 ≥ 0, ⎪
⎪
⎭
xraw2 ≥ 0.
where both ξ1 and ξ2 are normally distributed with mean 0. As discussed in
Section 1.3, we shall look at the 99% intervals for both (as if that was the
support). This gives us
ξ1 ∈ [−30.91, 30.91], ξ2 ∈ [−23.18, 23.18].
The interpretation is that 100 is the production limit of a refinery, which

refines crude oil from two countries. The variable xraw1 represents the amount
of crude oil from Country 1 and xraw2 the amount from Country 2. The
qualities of the crude oils are different, so one unit of crude oil from Country
1 gives two units of Product 1 and three units of Product 2, whereas the crude
φ(ξ)
1
ξ)
2
ξ)
ξ
Figure 13 Two possible lower bounding functions.
oil from the second country gives 6 and 3 units of the same products. Company
1 wants at least 180 + ξ1 units of Product 1 and Company 2 at least 162 + ξ2
units of Product 2. The goal now is to find the expected value of φ(ξ); in other
words, we seek the expected value of the “wait-and-see” solution. Note that
this interpretation is not the one we adopted in Section 1.3.
3.4.1 The Jensen Lower Bound

Assume that q(ξ) ≡ q0 , so that randomness affects only the right-hand side.
The purpose of this section is to find a lower bound on Q(x̂, ξ), for fixed x̂,
and for that purpose we shall, as just mentioned, use φ(ξ) ≡ Q(x̂, ξ) for a
fixed x̂.
Since φ(ξ) is a convex function, we can bound it from below by a linear
function L(ξ) = cξ + d. Since the goal will always be to find a lower bound
that is as large as possible, we shall require that the linear lower bound be
tangent to φ(ξ) at some point ξ̂. Figure 13 shows two examples of such lower-
bounding functions. But the question is which one should we pick. Is L1 (ξ)
or L2 (ξ) the better?
ˆ the slope
If we let the lower bounding function L(ξ) be tangent to φ(ξ) at ξ,
ˆ
must be φ (ξ), and we must have
ˆ = φ (ξ)
φ(ξ) ˆ ξˆ + b,
ˆ = L(ξ).
since φ(ξ) ˆ Hence, in total, the lower-bounding function is given by
ˆ + φ (ξ)(ξ
L(ξ) = φ(ξ) ˆ − ξ).
ˆ
Since this is a linear function, we easily calculate the expected value of the
lower-bounding function:
EL(ξ) ˆ + φ (ξ)(E
˜ = φ(ξ) ˆ ξ̃ − ξ̂) = L(E ξ̃).
In other words, we find the expected lower bound by evaluating the lower
bounding function in E ξ̃. From this, it is easy to see that we obtain the best
(largest) lower bound by letting ξ̂ = E ξ̃. This can be seen not only from the
fact that no linear function that supports φ(ξ) can have a value larger than
φ(E ξ̃) in E ξ̃, but also from the following simple differentiation:
d
L(E ξ̃) = φ (ξ)
ˆ − φ (ξ)
ˆ + φ (ξ)(E
ˆ ξ̃ − ξ̂).
dξˆ
If we set this equal to zero we find that ξˆ = E ξ̃. What we have developed is
the so-called Jensen lower bound, or the Jensen inequality.
Proposition 3.1 If φ(ξ) is convex over the support of ξ˜ then

˜ ≥ φ(E ξ̃)
Eφ(ξ)
This best lower bound is illustrated in Figure 14. We can see that the
Jensen lower bound can be viewed two different ways. First, it can be seen as
a bound where a distribution is replaced by its mean and the problem itself
is unchanged. This is when we calculate φ(E ξ̃). Secondly, it can be viewed as
a bound where the distribution is left unchanged and the function is replaced
by a linear affine function, represented by a straight line. This is when we
integrate L(ξ) over the support of ξ̃. Depending on the given situation, both
these views can be useful.
There is even a third interpretation. We shall see it used later in the
stochastic decomposition method. Assume we first solve the dual of φ(E ξ̃) to
obtain an optimal basis B. This basis, since ξ does not enter the constraints of
the dual of φ, is dual feasible for all possible values of ξ. Assume now that we
solve the dual version of φ(ξ) for all ξ, but constrain our optimization so that
we are allowed to use only the given basis B. In such a setting, we might claim
that we use the correct function, the correct distribution, but optimize only in
an approximate way. (In stochastic decomposition we use not one, but a finite
number of bases.) The Jensen lower bound can in this setting be interpreted as
representing approximate optimization using the correct problem and correct
distribution, but only one dual feasible basis.
It is worth pointing out that these interpretations of the Jensen lower bound
are put forward to help you see how a bound can be interpreted in different
ways, and that these interpretations can lead you in different directions when
trying to strengthen the bound. An interpretation is not necessarily motivated
by computational efficiency.
Looking back at our example in (4.1), we find the Jensen lower bound by
calculating φ(E ξ̃) = φ(0). That has been solved already in Section 1.3, where
we found that φ(0) = 126.
3.4.2 Edmundson–Madansky Upper Bound

Again let ξ̃ be a random variable. Let the support Ξ = [a, b], and assume that
q(ξ) ≡ q0 . As in the previous section, we define φ(ξ) = Q(x̂, ξ). (Remember
that x is fixed at x̂.) Consider Figure 14, where we have drawn a linear function
U (ξ) between the two points (a, φ(a)) and (b, φ(b)). The line is clearly above
φ(ξ) for all ξ ∈ Ξ. Also this straight line has the formula cξ + d, and since we
know two points, we can calculate
φ(b) − φ(a) b a
c= , d= φ(a) − φ(b).
b−a b−a b−a
We can now integrate, and find (using the linearity of U (ξ))
˜ = φ(b) − φ(a) E ξ̃ + b φ(a) − a φ(b)

EU (ξ)
b−a b−a b−a
b − E ξ̃ E ξ̃ − a
= φ(a) + φ(b) .
b−a b−a
In other words, if we have a function that is convex in ξ over a bounded
support Ξ = [a, b], it is possible to replace an arbitrary distribution by a
two point distribution, such that we obtain an upper bound. The important
parameter is
E ξ̃ − a
p= ,
b−a
so that we can replace the original distribution with
P {ξ˜ = a} = 1 − p, P {ξ˜ = b} = p. (4.2)
As for the Jensen lower bound, we have now shown that the Edmundson–
Madansky upper bound can be seen as either changing the distribution and
keeping the problem, or changing the problem and keeping the distribution.
Looking back at our example in (4.1), we have two independent random
variables. Hence we have 22 = 4 LPs to solve to find the Edmundson–
Madansky upper bound. Since both distributions are symmetric, the
probabilities attached to these four points will all be 0.25. Calculating this
we find an upper bound of
1
4
(106.6825 + 129.8625 + 122.1375 + 145.3175) = 126.
Edmundson-Madansky U(ξ)
φ(ξ)
Jensen L(ξ)
a Eξ b ξ
Figure 14 The Jensen lower bound and the Edmundson–Madansky upper
bound in a minimization problem. Note that x is fixed.
This is exactly the same as the lower bound, and hence it is the true value
˜ We shall shortly comment on this situation where the bounds turn
of Eφ(ξ).
out to be equal.
In higher dimensions, the Jensen lower bound corresponds to a hyperplane,
while the Edmundson–Madansky bound corresponds to a more general
polynomial. A two-dimensional illustration of the Edmundson–Madansky
bound is given in Figure 15. Note that if we fix the value of all but one of the
variables, we get a linear function. This polynomial is therefore generated by
straight lines. From the viewpoint of computations, we do not have to relate to
this general polynomial. Instead, we take one (independent) random variable
at a time, and calculate (4.2). This way we end up with 2 possible values for
each random variable, and hence, 2k possible values of ξ for which we have to
evaluate the recourse function.
Assume that the function φ(ξ) in Figure 14 is linear. Then it appears from
the figure that both the Jensen lower bound and the Edmundson–Madansky
upper bound are exact. This is indeed a correct observation: both bounds are
exact whenever the function is linear. And, in particular, this means that if
the function is linear, the error is zero. In the example (4.1) used to illustrate
the Jensen and Edmundson–Madansky bounds we observed that the bounds
where equal. This shows that the function φ(ξ) is linear over the support we
used.
One special use of the Jensen lower bound and Edmundson–Madansky
upper bound is worth mentioning. Assume we have a random vector,
containing a number of independent random variables, and a function that is
convex with respect to that random vector, but the random vector either has
a continuous distribution, or a discrete distribution with a very large number
of outcomes. In both cases we might have to simplify the distribution before
Figure 15 Illustration of the Edmundson–Madansky upper bound in two

dimensions. The function itself is not drawn. The Jensen lower bound, which is
simply a plane, is also not drawn.
making any attempts to attack the problem.

The principle we are going to use is as follows. Take one random variable
at a time. First partition the support of the variable into a finite number of
intervals. Then apply the principle of the Edmundson–Madansky bound on
one interval at a time. Since we are inside an interval, we use conditional
distributions, rather than the original one. This will in effect replace the
distribution over the interval by a distribution that has probability mass only
at the end points. This is illustrated in Figure 16, where we have shown the
case for one random variable. The support of ξ̃ has been partitioned into
two parts, called cells. For each of these cells, we have drawn the straight
lines corresponding to the Jensen lower bound and the Edmundson–Madansky
upper bound. Corresponding to each cell, there is a one-point distribution that
gives a lower bound, and a two-point distribution that gives an upper bound,
just as we have outlined earlier.
If the random variables have continuous (but bounded) distributions, we
use these conditional bounds to replace the original distribution with discrete
distributions. If the distribution is already discrete, we can remove some of
the outcomes by using the Edmundson–Madansky inequality conditionally on
parts of the support, again pushing probability mass to the end points of
the intervals. Of course, the Jensen inequality can be used in the same way
to construct conditional lower bounds. The point with these changes is not
to create bounds per se, but to simplify distributions in such a way that we
have control over what we have done to the problem when simplifying. The
Edmundson-Madansky
φ(ξ)
Jensen
a b ξ
cell 1 cell 2
Figure 16 Illustration of the effect on the Jensen lower bound and the
Edmundson–Madansky upper bound of partitioning the support into two cells.
Figure 17 Simplifying distributions by using Jensen and Edmundson–

Madansky on subintervals of the support. The stars represent conditional
expectations, and hence a distribution resulting in a lower bound. The bars
are endpoints of intervals, representing a distribution yielding an upper bound.
idea is outlined in Figure 17. Whatever the original distribution was, we now
have two distributions: one giving an overall lower bound, the other an overall
upper bound.
Since the random variables in the vector were assumed to be independent,
this operation has produced discrete distributions for the random vector as
well.
3.4.3 Combinations
If we have randomness in the objective function, but not in the right-hand
side (so h(ξ)−T (ξ)x ≡ h0 −T0 x), then, by simple linear programming duality,
we can obtain the dual of Q(x, ξ) with all randomness again in the right-hand
side, but now in a setting of maximization. In such a setting the Jensen bound
is an upper bound and the Edmundson–Madansky bound a lower bound.
If we have randomness in both the objective and the right-hand side, and the
random variables affecting these two positions are different and independent,
then we get a lower bound by applying the Jensen rule on the right-hand side
random variables and the Edmundson–Madansky rule in the objective. If we
do it the other way around, we get an overall upper bound.
3.4.4 A Piecewise Linear Upper Bound

Although the Edmundson–Madansky distribution is very useful, it still
requires that Q(x, ξ) be evaluated at an exponential number of points. That is,
if there are k random variables, we must work with 2k points. This means that
with more than about 10 random variables we are not in business. In order to
facilitate upper bounds, a number of approaches that are not of exponential
complexity in terms of the number of random variables have been designed.
In what follows we shall briefly demonstrate how to obtain a piecewise linear
upper bound that does not exhibit this exponential characterization.
The idea behind the development is as follows. The recourse function Q(x̂, ξ)
is convex in ξ (for a fixed x̂). We might envisage it as a bowl. The Jensen lower
bound represents a supporting hyperplane below the recourse function, like a
table on which the bowl sits. Any supporting hyperplane would give a lower
bound, but, as we have seen, the one that touches Q(x̂, ξ) in E ξ̃ gives the
highest lower bound. The Edmundson–Madansky upper bound, on the other
hand, is much like a lid on the bowl. They are both illustrated in Figure 14.
The purpose of the piecewise linear upper bound is to find another bowl that
fits inside the bowl Q(x̂, ξ), but at the same time has more curvature than
the Edmundson–Madansky lid. Also, this new bowl must represent a function
that is easy to integrate.
The piecewise linear upper bound has exactly these properties. It should
be noted that it is impossible to compare the piecewise linear upper bound
with the Edmundson–Madansky bound, in the sense that either one can be
best in a given example. In particular, the new bound may be +∞ even if
the problem is feasible (meaning that Q(x̂, ξ) < ∞ for all possible ξ). This
can never happen to the Edmundson–Madansky upper bound. It seems that
the new bound is reasonably good on “loose” problems, i.e. problems that are
very far from being infeasible, such as problems with complete recourse. The
Edmundson–Madansky bound is better on “tight” problems.
Let us illustrate the method in a simplified setting. We shall only consider
randomness in the right-hand side of W y = b, and leave the discussion of
randomness in the upper bound c to Chapter 6. Define φ(ξ) by
φ(ξ) = min{q T y | W y = b + ξ, 0 ≤ y ≤ c}
y
where all components in the random vector ξ˜T = (ξ˜1 , ξ˜2 , . . .) are mutually
independent. Furthermore, let the support be given by Ξ(ξ) ˜ = [A, B]. For
convenience, but without any loss of generality, we shall assume that E ξ̃ = 0.
The goal is to create a piecewise linear, separable and convex function in ξ:

d+ ξ if ξ ≥ E ξ̃ = 0
U (ξ) = φ(0) + i i i i
(4.3)
−
i
di ξi if ξi < E ξ̃i = 0.
There is a very good reason for such a choice. Note how U (ξ) is separable in
its components ξi . Therefore, for almost all distribution functions, U is simple
to integrate.
To appreciate the bound, we must understand its basic motivation. If we
take some minimization problem, like the one here, and add extra constraints,
the resulting problem will bound the original problem from above. What we
shall do is to add restrictions with respect to the upper bounds c. We shall do
this by viewing φ(ξ) as a parametric problem in ξ, and reserve portions of the
upper bound c for the individual random variables ξi . We may, for example,
end up by saying that two units of cj are reserved for variable ξi , meaning that
these two units can be used in the parametric analysis, only when we consider
ξi . For all other variables ξk these two units will be viewed as nonexisting.
The clue of the bound is to introduce the best possible set of such constraints,
such that the resulting problem is easy to solve (and gives a good bound).
First, let us calculate φ(E ξ̃) = φ(0) by finding
φ(0) = min{q T y | W y = b, 0 ≤ y ≤ c} = q T y 0 .
y
This can be interpreted as the basic setting, and all other values of ξ will be
seen as deviations from E ξ̃ = 0. (Of course, any other starting point will also
do—for example solving Q(A), where, as stated before, A is the lowest possible
value of ξ.) Note that since y 0 is “always” there, we can in the following operate
with bounds −y 0 ≤ y ≤ c − y 0 . For this purpose, we define α1 = −y 0 and
β 1 = c − y 0 . Let ei be a unit vector of appropriate size with a +1 in position
i.
Next, define a counter r and let r := 1. Now check out the case when ξr > 0
by solving (remembering that Br is the maximal value of ξr )
min{q T y | W y = er Br , αr ≤ y ≤ β r } = q T y r+ = d+
r Br . (4.4)
y
Note that d+r represents the per unit cost of increasing the right-hand side
from 0 to er Br . Similarly, check out the case with ξr < 0 by solving
min{q T y | W y = er Ar , αr ≤ y ≤ β r } = q T y r− = d−
r Ar . (4.5)
y
Now, based on y r± , we shall assign portions of the bounds to the random

variable ξ̃r . These portions of the bounds will be given to ξ˜r and left unused
by other random variables, even when ξ̃r does not need them. That is done by
means of the following problem, where we calculate what is left for the next
random variable:
αr+1
i = αri − min{yir+ , yir− , 0}. (4.6)
What we are doing here is to find, for each variable, how much ξ̃r , in the worst
case, uses of the bound on variable i in the negative direction. That is then
subtracted off what we had before. There are three possibilities. Both (4.4) and
(4.5) may yield non-negative values for the variable yi . In that case nothing
is used of the available “negative bound” αri . Then αr+1i = αri . Alternatively,
if (4.4) has yi < 0, then it will in the worst case use yir+ of the available
r+
“negative bound”. Finally, if (4.5) has yir− < 0 then in the worst case we use
yir− of the bound. Therefore αr+1
i is what is left for the next random variable.
Similarly, we find
βir+1 = βir − max{yir+ , yir− , 0}, (4.7)

where βir+1 shows how much is still available of bound i in the forward
(positive) direction.
We next increase the counter r by one and repeat (4.4)–(4.7). This takes
care of the piecewise linear functions in ξ.
Note that it is possible to solve (4.4) and (4.5) by parametric linear
programming, thereby getting not just one linear piece above E ξ̃ and one
below, but rather piecewise linearity on both sides. Then (4.6) and (4.7) must
be updated to “worst case” analysis of bound usage. That is simple to do.
Let us turn to our example (4.1). Since we have developed the piecewise
linear upper bound for equality constraints, we shall repeat the problem with
slack variables added explicitly.
φ(ξ1 , ξ2 ) = min{2xraw1 + 3xraw2 }
s.t. xraw1 + xraw2 + s1 = 100,
2xraw1 + 6xraw2 − s2 = 180 + ξ1 ,
3xraw1 + 3xraw2 − s3 = 162 + ξ2 ,
xraw1 ≥ 0,
xraw2 ≥ 0,
s1 ≥ 0,
s2 ≥ 0,
s3 ≥ 0.
In this setting, what we need to develop is the following:
+ +
d1 ξ1 if ξ1 ≥ 0, d2 ξ2 if ξ2 ≥ 0,
U (ξ1 , ξ2 ) = φ(0, 0) + +
d−
1 ξ1 if ξ1 < 0, d−
2 ξ2 if ξ2 < 0.
First, we have already calculated φ(0, 0) = 126 with xraw1 = 36, xraw2 =
18 and s1 = 46. Next, let us try to find d± 1
1 . To do that, we need α ,
which equals (−36, −18, −46, 0, 0). We must then formulate (4.4), using ξ1 ∈
[−30.91, 30.91]:
min{2xraw1 + 3xraw2 }
s.t. xraw1 + xraw2 + s1 = 0,
2xraw1 + 6xraw2 − s2 = 30.91,
3xraw1 + 3xraw2 − s3 = 0,
xraw1 ≥ −36,
xraw2 ≥ −18,
s1 ≥ −46,
s2 ≥ 0,
s3 ≥ 0.
The solution to this is y 1+ = (−7.7275, 7.7275, 0, 0, 0)T, with a total cost of

7.7275. This gives us
(2, 3, 0, 0, 0)y 1+
d+
1 = = 0.25.
30.91
Next, we solve the same problem, just with 30.91 replaced by −30.91.
This amounts to problem (4.5), and gives us the solution is y 1− =
(7.7275, −7.7275, 0, 0, 0)T, with a total cost of −7.7275. Hence, we get
(2, 3, 0, 0, 0)y 1−
d−
1 = = 0.25.
−30.91
The next step is to update α according to (4.6) to find out how much is left
of the negative bounds on the variables. For xraw1 we get
α2raw1 = −36 − min{−7.7275, 7.7275, 0} = −28.2725.
For xraw2 we get in a similar manner
α2raw2 = −18 − min{7.7275, −7.7275, 0} = −10.2725.
For the three other variables, α2i equals α1i . We can now turn to (4.4) for
random variable 2. The problem to solve is as follows, when we remember the
ξ2 ∈ [−23.18, 23.18].
s.t. xraw1 + xraw2 + s1 = 0,
2xraw1 + 6xraw2 − s2 = 0,
3xraw1 + 3xraw2 − s3 = 23.18,
xraw1 ≥ −28.2725,
xraw2 ≥ −10.2725,
s1 ≥ −46,
s2 ≥ 0,
s3 ≥ 0.
The solution to this is y 2+ = (11.59, −3.863, −7.727, 0, 0)T, with a total cost
of 11.59. This gives us
(2, 3, 0, 0, 0)y 2+
d+
2 = = 0.5.
23.18
Next, we solve the same problem, just with 23.18 replaced by −23.18.
This amounts to problem (4.5), and gives us the solution y 2− =
(−11.59, 3.863, 7.727, 0, 0)T, with a total cost of −11.59. Hence we get
(2, 3, 0, 0, 0)y 2−
d−
2 = = 0.5.
−23.18
This finishes the calculation of the (piecewise) linear functions in the upper
bound. What we have now found is that

1 1
4 ξ1 if ξ1 ≥ 0, ξ2 if ξ2 ≥ 0,
U (ξ1 , ξ2 ) = 126 + 1 + 21
4 ξ1 if ξ1 < 0, 2 ξ2 if ξ2 < 0,
which we easily see can be written as
U (ξ1 , ξ2 ) = 126 + 14 ξ1 + 12 ξ2 .
In other words, as we already knew from calculating the Edmundson–

Madansky upper bound and Jensen lower bound, the recourse function is
linear in this example. Let us, for illustration, integrate with respect to ξ1 .
30.91
1
ξ f (ξ1 ) dξ1 = 14 E ξ̃1 = 0.
4 1
−30.91
This is how it should be for linearity, the contribution from a random variable
over which U (and therefore φ) is linear is zero. We should of course get the
same result with respect to ξ2 , and therefore the upper bound is 126, which
equals the Jensen lower bound.
Now that we have seen how things go in the linear case, let us try to see
how the results will be when linearity is not present. Hence assume that we
have now developed the necessary parameters d± i for (4.3). Let us integrate
with respect to the random variable ξ˜i , assuming that Ξi = [Ai , Bi ]:
0 Bi
d−
1 ξi f (ξi )dξi + d+
1 ξi f (ξi )dξi
Ai 0
= d− ˜ ˜ + ˜ ˜
i E{ξi | ξ̃i ≤ 0}P {ξi ≤ 0} + di E{ξi | ξ̃i > 0}P {ξi > 0}.
This result should not come as much of a surprise. When one integrates a
linear function, one gets the function evaluated at the expected value of the
random variable. We recognize this integration from the Jensen calculations.

From this, we also see, as we have already claimed a few times, that if d+i = di
−
for all i, then the contribution to the upper bound from ξ̃ equals φ(E ξ̃), which
equals the contribution to the Jensen lower bound.
Let us repeat why this is an upper bound. What we have done is to
distribute the bounds c on the variables among the different random variables.
They have been given separate pieces, which they will not share with others,
˜ do not need the capacities themselves.
even if they, for a given realization of ξ,
This partitioning of the bounds among the random variables represents
a set of extra constraints on the problem, and hence, since we have a
minimization problem, the extra constraints yield an upper bound. If we
run out of capacities before all random variables have received their parts,
we must conclude that the upper bound is +∞. This cannot happen with
the Edmundson–Madansky upper bound. If φ(ξ) is feasible for all ξ then the
Edmundson–Madansky bound is always finite. However, as for the Jensen and
Edmundson–Madansky bounds, the piecewise linear upper bound is also exact
when the recourse function turns out to be linear.
As mentioned before, we shall consider random upper bounds in Chapter 6,
in the setting of networks.
3.5 Approximations
3.5.1 Refinements of the bounds on the “Wait-and-See”Solution

Let us, also in this section, assume that x = x̂, and as before define
φ(ξ) = Q(x̂, ξ). Using any of the above (or other) methods we can find bounds
on the recourse function. Assume we have calculated L and U such that
˜ ≤ U.
L ≤ Eφ(ξ)
We can now look at U − L to see if we are happy with the result or not.
If we are not, there are basically two approaches that can be used. Either
we might resort to a better bounding procedure (probably more expensive in
terms of CPU time) or we might start using the old bounding methods on a
partition of the support, thereby making the bounds tighter. Since we know
only finitely many different methods, we shall eventually be left with only the
second option.
The set-up for such an approach to bounding will be as follows. First,
partition the support of the random variables into an arbitrary selection of
cells—possibly only one cell initially. We shall only consider cells that are
rectangles, so that they can be described by intervals on the individual random
variables. Figure 18 shows an example in two dimensions with five cells. Now,
apply the bounding procedures on each of the cells, and add up the results.
Cell2
Cell 1
Cell 3
Cell 4 Cell 5
Figure 18 Partitioning of cells.
For example, in Figure 18, we need to find five conditional expectations, so

that we can calculate the Jensen lower bound on each cell. Adding these
up, using as weights the probability of being in the cell, we get an overall
lower bound. In the same way, an upper bound can be calculated on each cell
and added up to produce an overall upper bound. If the error U − L is too
large, one or more of the cells must be partitioned. It is natural to chose the
cell(s) with the largest error(s), but along which coordinate(s) should it/they
be partitioned, and through which point(s) in the cell(s)? Note that this use
of conditional expectations is very similar to the way we created discrete
distributions towards the end of Section 3.4.2. In particular, check out the
discussion of Figure 16 on page 184.
Not much is known about what is a good point for partitioning. Obvious
possibilities are the middle of the support and the (conditional) mean or
median. Our experience is that the middle of the support is good, so we shall
use that. However, this subject is clearly open for discussion. Let us therefore
turn to the problem of picking the correct coordinate (random variable). For
example, if we have picked Cell 1 in Figure 18 to be partitioned, should we
draw a vertical or horizontal line? This might seem like a minor question at
first sight. However, this is not at all the case. To see why, assume there is
a random variable that is never of any importance, such as a random upper
bound on a variable that, because of its high cost, is never used. Hence the
realized value of this random variable is totally uninteresting. Assume that,
for some reason, we pick this random variable for partitioning. The effect
will be that when we calculate the bounds again on the two new cells and
add them up, we have exactly the same error as before. But—and that is
crucial—we now have two cells instead of one. From a practical point of view,
these cells are exactly equal. They are only different with respect to a random
variable that could as well have been dropped. Hence, in effect, we now have
increased our work load. It is now harder to achieve a given error bound than
it was before the partition. And note, we shall never recover from the error,
in the sense that intelligent choices later on will not counteract this one bad
choice. Each time we make a bad partition, the workload from there onwards
basically doubles for the cell from which we started. Since we do not want to
unnecessarily increase the workload too often, we must be careful with how
we partition.
Now that we know that bad choices can increase the workload, what should
we do? The first observation is that chosing at random is not a good idea,
because, every now and then, we shall make bad choices. On the other hand,
it is clear that the partitioning procedure will have to be a heuristic. Hence,
we must make sure that we have a heuristic rule that we hope never makes
really bad choices.
By knowing our problem well, we may be able to order the random variables
according to their importance in the problem. Such an ordering could be used
as is, or in combination with other ideas. For some network problems, such as
the PERT problem (see Section 6.6), the network structure may present us
with such a list. If we can compile the list, it seems reasonable to ask, from a
modelling point of view, if the random variables last on the list should really
have been there in the first place. They do not appear to be important.
Over the years, some attempts to understand the problem of partitioning
have been made. Most of them are based on the assumption that the
Edmundson–Madansky bound was used to calculate the upper bound. The
reason is that the dual variables associated with the solution of the recourse
function tell us something about its curvature. With the Edmundson–
Madansky bound, we solve the recourse problem at all extreme points of
the support, and thus get a reasonably good idea of what the function looks
like.
To introduce some formality, assume we have only one random variable
˜ with support Ξ = [A, B]. When finding the Edmundson–Madansky upper
ξ,
bound, we calculated φ(A) = Q(x̂, A) and φ(B) = Q(x̂, B), obtaining dual
solutions π A and π B . We know from duality that
φ(A) = (π A )T [h(A) − T (A)x̂],

φ(B) = (π B )T [h(B) − T (B)x̂].
We also know that, as long as q(ξ) ≡ q0 (which we are assuming in this

section), a π that is dual feasible for one ξ is dual feasible for all ξ, since ξ
does not enter the dual constraints. Hence, we know that
α = φ(A) − (π B )T [h(A) − T (A)x̂] ≥ 0
and
β = φ(B) − (π A )T [h(B) − T (B)x̂] ≥ 0.
φ φ
α β
ξ ξ
α
β
Figure 19 An illustration of a situation where both α and β give good

information about curvature.
The parameters α and β contain information about the curvature of φ(ξ).

In particular, note that if, for example, α = 0 then π B is an optimal dual
solution corresponding to ξ = A. If π A = π B in such a case, we are simply
facing dual degeneracy. In line with this argument, a small α (or β) should
mean little curvature. But we may, for example, have α large and β small. So
what is going on?
Figure 19 shows two different cases (both in one dimension), where both α
and β are good indicators of how important a random variable is. Intuitively, it
seems reasonable to say that the left part of the figure indicates an important
random variable, and the right part an unimportant one. And, indeed, in the
left part both α and β will be large, whereas in the right part both will be
small.
But then consider Figure 20. Intuitively, the random variable is
unimportant, but, in fact, the slopes at the end points are the same as in
the left part of Figure 19, and the slopes describe how the objective changes
as a function of the random variable. However, in this case α is very small,
whereas β is large. What is happening is that α and β pick up two properties
of the recourse function. If the function is very flat (as in the right part of
Figure 19) then both parameters will be small. If the function is very non-
linear (as in the left part of Figure 19), both parameters will be large. But if
α
ξ
Figure 20 An illustration of a case where α is small and β is large.
we have much curvature in the sense of the slopes of φ at the end point, but
still almost linearity (as in Figure 20), then the smaller of the two parameters
will be small. Hence the conclusion seems to be to calculate both α and β,
pick the smaller of the two, and use that as a measure of nonlinearity.
Using α and β, we have a good measure of nonlinearity in one dimension.
However, with more than one dimension, we must again be careful. We can
certainly perform tests corresponding to those illustrated in Figures 19 and 20,
for one random variable at a time. But the question is what value should
we give the other random variables during the test. If we have k random
variables, and have the Edmundson–Madansky calculations available, there
are 2k−1 different ways we can fix all but one variable and then compare dual
solutions. There are at least two possible approaches.
A first possibility is to calculate α and β for all neighbouring pairs of
extreme points in the support, and pick the one for which the minimum of
α and β is the largest. We then have a random variable for which φ is very
nonlinear, at least in parts of the support. We may, of course, have picked a
variable for which φ is linear most of the time, and this will certainly happen
once in a while, but the idea is tested and found sound.
An alternative, which tries to check average nonlinearity rather than
maximal nonlinearity, is to use all 2k−1 pairs of neighbouring extreme points
involving variation in only one random variable, find the minimum of α and
β for each such pair, and then calculate the average of these minima. Then
we pick the random variable for which this average is maximal.
The number of pairs of neighbouring extreme points is fairly large. With k
random variables, we have k2k−1 pairs to compare. Each comparison requires
the calculation of two inner products. We have earlier indicated that the
Edmundson–Madansky upper bound cannot be used for much more than 10
random variables. In such a case we must perform 5120 pairwise comparisons.
Looking back at Figures 19 and 20, we note that an important feature of

the recourse function is its slope, as a function of the random variable. We
alluded to the slope when discussing the parameters α and β, but we did not
really show how to find the slopes.
We know from linear programming duality that the optimal value of
a dual variable shows how the objective function will change (locally) as
the corresponding right-hand side element increases. Given that we use
the Edmundson–Madansky upper bound, these optimal dual solutions are
available to us at all extreme points of the support. If ξ j is the value of ξ˜ at
such an extreme point, we have
φ(ξ j ) = Q(x̂, ξ j ) = q T y(ξ j ) = π(ξ j )T [h(ξ j ) − T (ξ j )x̂].
What we need to know to utilize this information is how the right-hand side
changes as a given random variable ξ˜i changes. This is easy to calculate, since
all we have to do is to find the derivative of
+ ,
h(ξ) − T (ξ)x̂ = h0 + hj ξj − T0 + Tj ξj x̂
j j
with respect to ξi . This is easily found to be
hi − Ti x̂ ≡ δi .
Note that this expression is independent of the value of ξ̃, and hence it is the
same at all extreme points of the support. Now, if π(ξ j ) is the optimal dual
solution at an extreme point of the support, represented by ξ j , then the slope
of φ(ξ) = Q(x̂, ξ) with respect to ξi is given by
π(ξ j )T δi .
And, more generally, if we let δ T = (δ1 , δ2 , . . .), the vector
π(ξ j )T δ ≡ (π j )T (5.1)
characterizes how φ(ξ) = Q(x̂, ξ) changes with respect to all random variables.
Since these calculations are performed at each extreme point of the support,
and each extreme point has a probability according to the Edmundson–
Madansky calculations, we can interpret the vectors π j as outcomes of
a random vector π̃ that has 2k possible values and the corresponding
Edmundson–Madansky probabilities. If, for example, the random variable π̃i
has only one possible value, we know that φ(ξ) is linear in ξi . If π̃i has several
possible values, its variance will tell us quite a bit about how the slope varies
over the support. Since the random variables ξ̃i may have very different units,
and the dual variables measure changes in the objective function per unit
change in a right-hand side element, it seems reasonable to try to account

for differences in units. A possible (heuristic) approach is to multiply the
outcomes of π̃i by the length of the support of ξ˜i , before calculating means
and variances. (Assume, for example, that we are selling apples and bananas,
and the demands are uncertain. For some reason, however, we are measuring
bananas in tons and apples in kilograms. Now, if π̃1 refers to bananas and π̃2
to apples, would you see these products as equally important if π̃1 and π̃2 had
the same variance?)
Computationally, this is easier than the approach based on α and β, because
it requires that only 2k inner products are made. The distribution of π̃ can
be calculated as we visit the extreme points of the support (for finding the
Edmundson–Madansky bound), and we never have to store the inner products.
All the above ideas are based on information from the Edmundson–
Madansky upper bound, and therefore the solution of 2k linear programs. As
we have pointed out several times, for much more than 10 random variables,
we are not able to find the Edmundson–Madansky upper bound. And if so,
we shall not be able to use the partitioning ideas above either. Therefore we
should have ideas of how to partition that do not depend on which upper
bound we use. This does not imply, though, that the ideas that are to follow
cannot be used with success for the Edmundson–Madansky upper bound as
well.
One idea, which at first sight looks rather stupid, is the following: perform
all possible bi-partitions (i.e. with k random variables, perform all k, one at a
time) and pick the one that is best. By “best”, we here mean with the smallest
error in the next step. More formally, let
Ui − Li
be the error on the “left” cell if we partition random variable i, and let
Uir − Lri
be the error on the “right” cell. If pi is the probability of being in the left
cell, given that we are in the original cell, when we partition coordinate i, we
chose to partition the random variable i that minimizes
(Ui − Li )pi + (Uir − Lri )(1 − pi ). (5.2)

In other words, we perform all possible partitions, keep the best, and discard
the remaining information. If the upper bound we are using is expensive in
terms of CPU time, such an idea of “look-ahead” has two effects, which pull
in different directions. On one hand, the information we are throwing away
has cost a lot, and that seems like a waste. On the other hand, the very
fact that the upper bound is costly makes it crucial to have few cells in the
end. With a cheap (in terms of CPU time) upper bound, the approach seems
more reasonable, since checking all possibilities is not particularly costly, but,
even so, bad partitions will still double the work load locally. Numerical tests
indicate that this approach is very good even with the Edmundson–Madansky
upper bound, and the reason seems to be that it produces so few cells. Of
course, without Edmundson–Madansky, we cannot calculate α, β and π̃, so if
we do not like the look-ahead, we are in need of a new heuristic.
We have pointed out before that the piecewise linear upper bound can
obtain the value +∞. That happens if one of the problems (4.4) or (4.5)
becomes infeasible. If that takes place, the random variable being treated
when it happens is clearly a candidate for partitioning.
So far we have not really defined what constitutes a good partition. We shall
return to that after the next subsection. But first let us look at an example
illustrating the partitioning ideas.
Example 3.2 Consider the following function:
φ(ξ1 , ξ2 ) = max{x + 2y}

s.t. −x + y ≤ 6,
2x − 3y ≤ 21,
−3x + 7y ≤ 49,
x + 12y ≤ 120,
2x + 3y ≤ 45,
x ≤ ξ1 ,
y ≤ ξ2 .
Let us assume that Ξ1 = [0, 20] and Ξ2 = [0, 10]. For simplicity, we shall
assume uniform and independent distributions. We do that because the form
of the distribution is rather unimportant for the heuristics we are to explain.
The feasible set for the problem, except the upper bounds, is given in
Figure 21. The circled numbers refer to the numbering of the inequalities.
For all problems we have to solve (for varying values of ξ), it is reasonably
easy to read the solution directly from the figure.
Since we are maximizing, the Jensen bound is an upper bound, and the
Edmundson–Madansky bound a lower bound. We easily find the Jensen upper
bound from
φ(10, 5) = 20.
To find a lower bound, and also to calculate some of the information

needed to use the heuristics, we first calculate φ at all extreme points of the
support. Note that in what follows we view the upper bounds on the variables
as ordinary constraints. The results for the extreme-point calculations are
summed up in Table 1.
Figure 21 Set of feasible solutions for Example 3.2.
Table 1 Important characteristics of the solution of φ(ξ1 , ξ2 ) at the four

extreme points of the support.
(ξ1 , ξ2 ) x y φ Optimal dual solution (π)
(0, 0) = (L, L) 0.000 0.000 0.000 (0, 0, 0, 0, 0, 1, 2)

(20, 0) = (U, L) 10.500 0.000 10.500 (0, 12 , 0, 0, 0, 0, 72 )
(0, 10) = (L, U ) 0.000 6.000 12.000 (2, 0, 0, 0, 0, 3, 0)
(20, 10) = (U, U ) 8.571 9.286 27.143 (0, 0, 0, 0.0476, 0.476, 0, 0)
The first idea we wish to test is based on comparing pairs of extreme points,
to see how well the optimal dual solution (which is dual feasible for all right-
hand sides) at one extreme-point works at a neighbouring extreme point. We
use the indexing L and U to indicate Low and Up of the support.
LL:UL We first must test the optimal dual solution π LL together with the
right-hand side bUL . We get
α = (π LL )T bUL − φ(U, L)
= (0, 0, 0, 0, 0, 1, 2)(6, 21, 49, 120, 45, 20, 0)T − φ(U, L)
= 20 − 10.5 = 9.5.
We then do the opposite, to find
β = (π UL )T bLL − φ(L, L)
= (0, 12 , 0, 0, 0, 0, 72 )(6, 21, 49, 120, 45, 0, 0)T − φ(L, L)
= 10.5 − 0 = 10.5.
The minimum is therefore 9.5 for the pair LL:UL.
LL:LU Following a similar logic, we get the following:
α = (π LL )T bLU − φ(L, U )
= (0, 0, 0, 0, 0, 1, 2)(6, 21, 49, 120, 45, 0, 10)T − φ(L, U )
= 20 − 12 = 8,
β = (π LU )T bLL − φ(L, L)
= (2, 0, 0, 0, 0, 3, 0)(6, 21, 49, 120, 45, 0, 0)T − φ(L, L)
= 12 − 0 = 12.
The minimal value for the pair LL:LU is therefore 8.
LU:UU For this pair we get the following:
α = (π UU )T bLU − φ(L, U )
= (0, 0, 0, 0.0476, 0.476, 0, 0)(6, 21, 49, 120, 45, 0, 10)T − φ(L, U )
= 27.143 − 12 = 15.143
β = (π LU )T bUU − φ(L, L)
= (2, 0, 0, 0, 0, 3, 0)(6, 21, 49, 120, 45, 20, 10)T − φ(U, U )
= 72 − 27.143 = 44.857.
The minimal value for the pair LU:UU is therefore 15.143.
UL:UU For the final pair the results are given by
α = (π UU )T bUL − φ(U, L)
= (0, 0, 0, 0.0476, 0.476, 0, 0)(6, 21, 49, 120, 45, 20, 0)T − φ(U, L)
= 27.143 − 10.5 = 16.643,
β = (π UL )T bUU − φ(U, U )
= (0, 12 , 0, 0, 0, 0, 72 )(6, 21, 49, 120, 45, 20, 10)T − φ(U, U )
= 46.5 − 27.143 = 18.357.
The minimal value for the pair UL:UU is therefore 16.643.
If we were to pick the pair with the largest minimum of α and β, we should
pick the pair UL:UU, over which it is ξ2 that varies. In such a case we have
tried to find that part of the function that is the most nonlinear. When we
look at Figure 21, we see that as ξ2 increases (with ξ1 = 20), the optimal
solution moves from F to E and then to D, where it stays when ξ2 comes
above the y coordinate in D. It is perhaps not so surprising that this is the
most serious nonlinearity in φ.
If we try to find the random variable with the highest average nonlinearity,
by summing the errors over those pairs for which the given random variable
varies, we find that for ξ˜1 the sum is 9.5 + 15.143 = 24.643, and for ξ˜2 it is
8 + 16.643, which also equals 24.643. In other words, we have no conclusion.
The next approach we suggested was to look at the dual variables as in
(5.1). The right-hand side structure is very simple in our example, so it is
easy to find the connections. We define two random variables: π̃1 for the row
constraining x, and π̃2 for the row constraining y. With the simple kind of
uniform distributions we have assumed, each of the four values for π̃1 and π̃2
will have probability 0.25. Using Table 1, we see that the possible values for
π̃1 are 0, 1 and 3 (with 0 appearing twice), while for π̃2 they are 0, 2 and 3.5
(also with 0 appearing twice). There are different ideas we can follow.
1. We can find out how the dual variables vary between the extreme points.
The largest individual change is that π̃2 falls from 3.5 to 0 as we go from UL
to UU. This should again confirm that ξ˜2 is a candidate for partitioning.
2. We can calculate E π̃ = (1, 11
8 ), and the individual variances to 1.5 and
2.17. If we choose based on variance, we pick ξ˜2 .
3. We also argued earlier that the size of the support was of some importance.
A way of accommodating that is to multiply all outcomes with the length
of the support. (That way, all dual variables are, in a sense, a measure of
change per total support.) That should make the dual variables comparable.
The calculations are left to the reader. We now end up with π̃1 having the
largest variance. (And if we now look at the biggest change in dual variable
over pairs of neighboring extreme points, ξ˜1 will be the one to partition.)
No conclusions should be made based on these numbers in terms of what

is a good heuristic. We have presented these numbers to illustrate the
computations and to indicate how it is possible to make arguments about
partitioning. Before we conclude, let us consider the “look-ahead” strategy
(5.2). In this case there are two possibilities: either we split at ξ1 = 10 or we
split at ξ2 = 5. If we check what we need to compute in this case, we will
find that some calculations are required in addition to those in Table 1, and
Table 2 Function values needed for the “look-ahead”strategy.
(5,5) 15 (10,7.5) 25
(15,5) 25 (10,2.5) 15
(10,10) 27.143 (0,5) 10
(20,5) 25 (10,0) 10
φ(E ξ̃) = φ(10, 5) = 20, which we have already found. The additional numbers
are presented in Table 2.
Based on this, we can find the total error after splitting to be about 4.5
both for ξ˜1 and for ξ˜2 . Therefore, based on “look-ahead”, we cannot decide
what to do.
2
3.5.2 Using the L-shaped Method within Approximation Schemes

We have now investigated how to bound Q(x) for a fixed x. We have done
that by combining upper and lower bounding procedures with partitioning of
the support of ξ̃. On the other hand, we have earlier discussed (exact) solution
procedures, such as the L-shaped decomposition method (Section 3.2) and the
scenario aggregation (Section 2.6). These methods take a full event/scenario
tree as input and solve this (at least in principle) to optimality. We shall now
see how these methods can be combined.
The starting point is a set-up like Figure 18. We set up an initial partition
of the support, possibly containing only one cell. We then find all conditional
expectations (in the example there are five), and give each of them a
probability equal to that of being in their cell, and we view this as our
“true” distribution. The L-shaped method is then applied. Let ξ i denote the
conditional expectation of ξ,˜ given that ξ̃ is contained in the ith cell. Then
the partition gives us the support {ξ 1 , . . . ξ }. We then solve
⎫
min cT x + L(x) ⎬
s.t. Ax = b, (5.3)
⎭
x ≥ 0,
where

L(x) = pj Q(x, ξ j ),
j=1
cx + U1 ( x )
cx + U 2 ( x )
cx + Q( x )
cx + L2 ( x )
cx + L1 ( x )
cx$ + U1 ( x$ )
cx$ + Q ( x$ )
cx$ + L1 ( x$ )
x$
Figure 22 Example illustrating the use of bounds in the L-shaped
decomposition method. An initial partition corresponds to the lower bound-
ing function L1 (x) and the upper bounding function U1 (x). For all x we have
L1 (x) ≤ Q(x) ≤ U1 (x). We minimize cx + L1 (x) to obtain x̂. We find the error
U1 (x̂) − L1 (x̂), and we decide to refine the partition. This will cause L1 to be
replaced by L2 and U1 by U2 . Then the process can be repeated.
with pj being the probability of being in cell j. Let x̂ be the optimal solution
to (5.3). Clearly if x is the optimal solution to the original problem then
cT x̂ + L(x̂) ≤ cT x + L(x) ≤ cT x + Q(x),
so that the optimal value we found by solving (5.3) is really a lower bound
on min cT x + Q(x). The first inequality follows from the observation that x̂
minimizes cT x + L(x). The second inequality holds because L(x) ≤ Q(x) for
all x (Jensen’s inequality). Next, we use some method to calculate U(x̂), for
example the Edmundson–Madansky or piecewise linear upper bound. Note
that
cT x + Q(x) ≤ cT x̂ + Q(x̂) ≤ cT x̂ + U(x̂),
so cT x̂+U(x̂) is indeed an upper bound on cT x+Q(x). Here the first inequality

holds because x minimizes cT x + Q(x), and the second because, for all x,
Q(x) ≤ U(x).
We then have a solution x̂ and an error U(x̂) − L(x̂). If we are not satisfied
with the precision, we refine the partition of the support, and repeat the use of
the L-shaped method. It is worth noting that the old optimality cuts generated
in the L-shaped method are still valid, but generally not tight. The reason is
that, with more cells, and hence a larger , the function L(x) is now closer to
Q(x). Feasibility cuts are still valid and tight. Figure 22 illustrates how the
approximating functions L(x) and U(x) change as the partition is refined.
In total, this gives us the procedure in Figure 23. The procedure refine(Ξ)
will not be detailed, since there are so many options. We refer to our earlier
discussion of the subject in Section 3.5.1. Note that, for simplicity, we have
assumed that, after a partitioning, the procedure starts all over again in the
repeat loop. That is of course not needed, since we already have checked the
present x̂ for feasibility. If we replace the set A by Ξ in the call to procedure
feascut, the procedure Bounding L-shaped must stay as it is. In many cases
this may be a useful change, since A might be very large. (In this case old
feasibility cuts might no longer be tight.)
3.5.3 What is a Good Partition?

We have now seen partitioning used in two different settings. In the first we
just wanted to bound a one-stage stochastic program, while in the second we
used it in combination with the L-shaped decomposition method. The major
difference is that in the latter case we solve a two-stage stochastic program
between each time we partition. Therefore, in contrast to the one-stage setting,
the same partition (more and more refined) is used over and over again.
In the two-stage setting a new question arises. How many partitions should
we make between each new call to the L-shaped decomposition method? If we
make only one, the overall CPU time will probably be very large because a
new LP (only slightly changed from last time) must be solved each time we
make a new cell. On the other hand, if we make many partitions per call to
L-shaped, we might partition extensively in an area where it later turns out
that partitioning is not needed (remember that x enters the right-hand side
of the second-stage constraints, moving the set of possible right-hand sides
around). We must therefore strike a balance between getting enough cells and
not getting them in the wrong places.
This brings us to the question of what is a good partitioning strategy.
It should clearly be one that minimizes CPU time for solving the problem
at hand. Tests indicate that for the one-stage setting, using the idea of the
variance of the (random) dual variables on page 195, is a good idea. It creates
quite a number of cells, but because it is cheap (given that we already use
the Edmundson–Madansky upper bound) it is quite good overall. But, in
the setting of the L-shaped decomposition method, this large number of cells
become something of a problem. We have to carry them along from iteration
to iteration, repeatedly finding upper and lower bounds on each of them. Here
it is much more important to have few cells for a given error level. And that
procedure Bounding L-shaped(1 , 2 :real);

begin
K := 0, L := 0; Ξ̂ := {E ξ̃};
θ̂ := −∞, LP(A, b, c, x̂, feasible);
while not (stop) do begin
feascut(A, x̂,newcut);
Find L(x̂);
newcut := (L(x̂) − θ̂ > 1 );
(* Create an optimality cut—see page 168 *)
L := L + 1;
T
Construct the cut −βL x + θ ≥ αL ;
end;
end;
end
else begin
Find U(x̂);
stop := (U(x̂) − L(x̂) ≤ 2 );
if not (stop) then refine(Ξ̂);
end;
end;
end;
Figure 23 The L-shaped decomposition algorithm in a setting of approxima-

tions and bounds. The procedures that we refer to start on page 168, and the
set A was defined on page 162.
is best achieved by looking ahead using (5.2). Our general advice is therefore
that in the setting of two (or more) stages one should seek a strategy that
minimizes the final number of cells, and that it is worthwhile to pay quite a
lot per iteration to achieve this goal.
3.6 Simple Recourse
Let us consider the particular simple recourse problem
min{cT x + Eξ̃ Q(x, ξ)

˜ | Ax = b, x ≥ 0}, (6.1)
where
Q(x, ξ) = min{q +T y + + q −T y − | y + − y − = ξ − T x, y + ≥ 0, y − ≥ 0}.
Hence we assume
W = (I, −I),
T (ξ) ≡ T (constant),
h(ξ) ≡ ξ,
and in addition
q = q + + q − ≥ 0.
In other words, we consider the case where only the right-hand side is
random, andwe shall see that in this case, using our former presentation
h(ξ) = h0 + i hi ξi , we only need to know the marginal distributions of the
components hj (ξ) of h(ξ). However, stochastic dependence or independence
of these components does not matter at all. This justifies the above setting
h(ξ) ≡ ξ.
By linear programming duality, we have for the recourse function
Q(x, ξ)
= min{q +T y + + q −T y − | y + − y − = ξ − T x, y + ≥ 0, y − ≥ 0}
= max{(ξ − T x)T π | −q − ≤ π ≤ q + }. (6.2)
Observe that our assumption q ≥ 0 is equivalent to solvability of the second-

stage problem. Defining
χ := T x,
the dual solution π of (6.2) is obvious:

qi+ if ξi − χi > 0,
πi =
−qi− if ξi − χi ≤ 0.
Hence, with
(ξi − χi )qi+ if χi < ξi ,
Q̂i (χi , ξi ) =
−(ξi − χi )qi− if χi ≥ ξi ,
we have
Q(x, ξ) = Q̂i (χi , ξi ) with χ = T x.
i
The expected recourse follows immediately:

˜ =
Eξ̃ Q(x, ξ) Q(x, ξ)Pξ̃ (dξ)
Ξ

= Q̂i (χi , ξi )Pξ̃ (dξ)
i Ξ

+ −
= qi (ξi − χi )Pξ̃ (dξ) − qi (ξi − χi )Pξ̃ (dξ) .
i ξi >χi ξi ≤χi
The last expression shows that knowledge of the marginal distributions of the
ξ˜i is sufficient to evaluate the expected recourse. Moreover, Eξ̃ Q(x, ξ) ˜ is a
˜ = m
so-called separable function in (χ1 , · · · , χm1 ), i.e. Eξ̃ Q(x, ξ) i=1 Qi (χi ),
1
+ −
where, owing to q + q = q,
⎫
Qi (χi ) = qi+ ξi >χi (ξi − χi )Pξ̃ (dξ) − qi− ξi ≤χi (ξi − χi )Pξ̃ (dξ) ⎪
⎬
= qi+ Ξ (ξi − χi )Pξ̃ (dξ) − (qi+ + qi− ) ξi ≤χi (ξi − χi )Pξ̃ (dξ) (6.3)
⎪
⎭
+ +
= qi ξ i − qi χi − q i ξi ≤χi (ξi − χi )Pξ̃ (dξ)
with ξ i = Eξ̃ ξ˜i .

The reformulation (6.3) reveals the shape of the functions Qi (χi ). Assume
that Ξ is bounded such that αi < ξi ≤ βi ∀i, ξ ∈ Ξ. Then we have
⎧ +
⎪
⎪ q ξ − qi+ χi if χi ≤ αi ,
⎨ i i
+ +
Qi (χi ) = qi ξ i − qi χi − q i (ξi − χi )Pξ̃ (dξ) if αi < χi < βi , (6.4)
⎪
⎪ ξi ≤χi
⎩
−qi− ξ i + qi− χi if χi ≥ βi ,
showing that for χi < αi and χi > βi the functions Qi (χi ) are linear (see
Figure 24). In particular, we have
Qi (χi ) = Q̂i (χi , ξ i ) if χi ≤ αi or χi ≥ βi . (6.5)
Following the approximation scheme described in Section 3.5.1, the
relation (6.5) allows us to determine an error bound without computing the
E–M bound.3 To see this, consider any fixed χ̂i . If χ̂i ≤ αi or χ̂i ≥ βi then,
by (6.5),
Qi (χ̂i ) = Q̂i (χ̂i , ξ i ).
3 By “E–M” we mean the Edmundson–Madansky bound described in Section 3.4.2.
Figure 24 Simple recourse: supporting Qi (χi ) by Q̂i (χi , ξ i ).
If, on the other hand, αi < χ̂i < βi , we partition the interval (αi , βi ] into the
two subintervals (αi , χ̂i ] and (χ̂i , βi ] with the conditional expectations
1 2
ξ i = Eξ̃ (ξ˜i | ξi ∈ (αi , χ̂i ]), ξ i = Eξ̃ (ξ˜i | ξi ∈ (χ̂i , βi ]).
Obviously relation (6.5) also applies analogously to the conditional

expectations
Q1i (χ̂i ) = Eξ̃ (Q̂i (χ̂i , ξ̃i ) | ξi ∈ (αi , χ̂i ])
and
Q2i (χ̂i ) = Eξ̃ (Q̂i (χ̂i , ξ˜i ) | ξi ∈ (χ̂i , βi ]).
Therefore
1 2
Q1i (χ̂i ) = Q̂i (χ̂i , ξ i ), Q2i (χ̂i ) = Q̂i (χ̂i , ξ i ),
and, with p1i = P (ξi ∈ (αi , χ̂i ]) and p2i = P (ξi ∈ (χ̂i , βi ]),
Qi (χ̂i ) = p1i Q1i (χ̂i ) + p2i Q2i (χ̂i )

1 2
= p1i Q̂i (χ̂i , ξ i ) + p2i Q̂i (χ̂i , ξ i ).
Hence, instead of using the E–M upper bound, we can easily determine the
1 2 1 2
exact value Qi (χ̂i ). With Q̂i (χi , ξ i , ξ i ) := p1i Q̂i (χi , ξ i ) + p2i Q̂i (χi , ξ i ), the
resulting situation is demonstrated in Figure 25.
Assume now that for a partition of the intervals (αi , βi ] into subintervals
Iiν := (δiν , δiν+1 ], ν = 0, · · · , Ni − 1, with αi = δi0 < δi1 < · · · < δiNi = βi ,
we have minimized the Jensen lower bound (see Section 3.4.1), letting piν =
1 2
Figure 25 Simple recourse: supporting Qi (χi ) by Q̂i (χi , ξ i , ξ i ).
P (ξi ∈ Iiν ), ξ iν = Eξ̃ (ξ˜i | ξi ∈ Iiν ):
'
k N i −1 (
minx,χ cT x + piν Q̂i (χi , ξ iν )
i=1 ν=0
s.t. Ax = b,
T x − χ = 0,
x ≥ 0,
yielding the solution x̂ and χ̂ = T x̂. Obviously relation (6.5) holds for
conditional expectations Qiν (χ̂i ) (with respect to Iiν ) as well. Then for each
component of χ̂ there are three possibilities.
(a) If χ̂i ≤ αi , then
Q̂i (χ̂i , ξ iν ) = Qiν (χ̂i ) = Eξ̃ (Q̂i (χ̂i , ξ˜i ) | ξi ∈ Iiν ), ν = 0, · · · , Ni − 1,
and hence

N i −1
Qi (χ̂i ) = piν Q̂i (χ̂i , ξ iν ),

ν=0
i.e. there is no error with respect to this component.
(b) If χ̂i ≥ βi , then it again follows from (6.5) that

N i −1
Qi (χ̂i ) = piν Q̂i (χ̂i , ξ iν ).

ν=0
(c) If χ̂i ∈ Iiµ for exactly one µ, with 0 ≤ µ < Ni , then there are two cases.
First, if δiµ < χ̂i < δiµ+1 , partition Iiµ = (δiµ , δiµ+1 ] into
1 2
Jiµ = (δiµ , χ̂i ] and Jiµ = (χ̂i , δiµ+1 ].
Now, again owing to (6.5), it follows that
2
ρ
Qi (χ̂i ) = piν Q̂i (χ̂i , ξ iν ) + pρiµ Q̂i (χ̂i , ξ iµ ),
ν=µ ρ=1
where
ρ
pρiµ = P (ξi ∈ Jiµ
ρ ρ
), ξ iµ = Eξ̃ (ξ˜i | ξi ∈ Jiµ ), ρ = 1, 2.
If, on the other hand, χ̂i = δiµ+1 , we again have

N i −1
Qi (χ̂i ) = piν Q̂i (χ̂i , ξ iν ).

ν=0
In conclusion, having determined the minimal point χ̂ for the Jensen lower
bound, we immediately get the exact expected recourse at this point and
decide whether for all components the relative error fits into a prescribed
tolerance, or in which component the refinement (partitioning the subinterval
containing χ̂i by dividing it exactly at χ̂i ) seems appropriate for a further
improvement of the approximate solution of (6.1). Many empirical tests have
shown this approach to be very efficient. In particular, for this special problem
type higher dimensions of ξ˜ do not cause severe computational difficulties,
as they did for general stochastic programs with recourse, as discussed in
Section 3.5 .
3.7 Integer First Stage
This book deals almost exclusively with convex problems. The only exception
is this section, where we discuss, very briefly, some aspects of integer
programming. The main reason for doing so is that some solution procedures
for integer programming fit very well with some decomposition procedures
for (continuous) stochastic programming. Because of that we can achieve
two goals: we can explain some connections between stochastic and integer
programming, and we can combine the two subject areas. This allows us to
arrive at a method for stochastic integer programming. Note that talking
about stochastic and integer programming as two distinct areas is really
meaningless, since stochastic programs can contain integrality constraints, and
integer programs can be stochastic. But we still do it, with some hesitation,
since the splitting is fairly common within the mathematical programming

community.
To get started, let us first formulate a deterministic integer programming
problem in a very simple format, and then outline a common solution
procedure, namely branch-and-bound. An integer program can be formulated
as
⎫
min cT x ⎬
s.t. Ax = b (7.1)
⎭
xi ∈ {ai , ai + 1, . . . , bi − 1, bi } for all i,
where {ai , ai + 1, . . . , bi − 1, bi } is the set of all integers from ai to bi .
The branch-and-bound procedure is based on replacing xi ∈ {ai , ai +
1, . . . , bi − 1, bi } by ai ≤ xi ≤ bi for all i, and solving the corresponding
relaxed linear program to obtain x̂. If x̂ happens to be integral, we are done,
since integrality is satisfied without being enforced. If x̂ is not integral, we
have obtained a lower bound z = cT x̂ on the true optimal objective, since
dropping constraints in a minimization problem yields a lower bound.
To continue from here, we pick one variable xj , called the branching variable,
and one integer dj . Normally dj is chosen as the largest integer less than
the value of xj in the LP solution, and xj is normally a variable that was
nonintegral in the LP solution. We then replace our original problem (7.1) by
two similar problems:
⎫ ⎫
min cT x ⎪ ⎪
⎬ ⎪
⎪ ⎪
⎪
s.t. Ax = b, ⎪
⎪
⎪
⎪
xi ∈ {ai , . . . , bi } for all i = j, ⎪ ⎪
⎭ ⎪
⎪ ⎪
⎪
xj ∈ {aj , . . . , dj }, ⎪
⎪
⎬
and (7.2)
⎫ ⎪
⎪
⎪
⎪
min cT x ⎪
⎪ ⎪
⎬ ⎪
⎪
⎪
s.t. Ax = b, ⎪
⎪
⎪
⎪
xi ∈ {ai , . . . , bi } for all i = j, ⎪
⎪ ⎪
⎪
⎭ ⎭
xj ∈ {dj + 1, . . . , bj }.
What we have done is to branch. We have replaced the original problem by
two similar problems that each investigate their part of the solution space.
The two problems are now put into a collection of waiting nodes. The term
“waiting node” is used because the branching can be seen as building up a
tree, where the original problem sits in the root and the new problems are
stored in child nodes. Waiting nodes are then leaves in the tree, waiting to be
analysed. Leaves can also be fathomed or bounded, as we shall see shortly.
We next continue to work with the problem in one of these waiting nodes.
We shall call this problem the present problem. When doing so, a number of
different situations can occur.
Figure 26 The situation after three branchings in a branch-and-bound tree.

One waiting node is left.
1. The present problem may be infeasible, in which case it is simply dropped,

or fathomed.
2. The present problem might turn out to have an integral optimal solution
x̂, in other words a solution that is truly feasible. If so, we compare cT x̂
with the best-so-far objective value z (we initiate z at +∞). If the new
objective value is better, we keep x̂ and update z so that z = cT x̂. We then
fathom the present problem.
3. The present problem might have a nonintegral solution x̂ with cT x̂ ≥ z. In
this case the present problem cannot possibly contain an optimal integral
solution, and it is therefore dropped, or bounded. (This is the process that
gives half of the name of the method.)
4. The present problem has a solution x̂ that does not satisfy any of the above
criteria. If so, we branch as we did in (7.2), creating two child nodes. We
then add them to the tree, making them waiting nodes.
An example of an intermediate stage for a branch-and-bound tree can be

found in Figure 26. Three branchings have taken place, and we are left with
two fathomed, one bounded and one waiting node. The next step will now be
to branch on the waiting node.
Note that as branching proceeds, the interval over which we solve the
continuous version must eventually contain only one point. Therefore, sooner
or later, we come to a situation where the problem is either infeasible, or we

are faced with an integral solution. We cannot go on branching forever. Hence
the algorithm will eventually stop, either telling us that no integral solution
exists, or giving us such a solution.
Much research in integer programming concerns how to pick the correct
variable xj for branching, how to pick the branching value dj , how to formulate
the problem so that branching becomes simpler, and how to obtain a good
initial (integer) solution so as to have a z < ∞ to start out with. To have
a good integer programming algorithm, these subjects are crucial. We shall
not, however, discuss those subjects here. Instead, we should like to draw
attention to some analogies between the branch-and-bound algorithm for
integer programs and the problem of bounding a stochastic (continuous)
program.
• In the integer case we partition the solution space, and in the stochastic
case the input data (support of random variables).
• In the integer case we must find a branching variable, and in the stochastic
case a random variable for partitioning.
• In the integer case we must find a value dj of xj (see (7.2)) for the branching,
and in the stochastic case we must determine a point dj in the support
through which we want to partition.
• Both methods therefore operate with a situation as depicted in Figure 18,
but in one case the rectangle is the solution space, while in the other it is
the support of the random variables.
• Both problems can be seen as building up a tree. For integer programming
we build a branch-and-bound tree. For stochastic programming we build a
splitting tree. The branch-and-bound-tree in Figure 26 could have been a
splitting tree as well. In that case we should store the error rather than the
objective value.
• In the integer case we fathom a problem (corresponding to a cell in
Figure 18, or a leaf in the tree) when it has nothing more to tell us, in
the stochastic case we do this when the bounds (in the cell or leaf) are close
enough.
From this, it should be obvious that anyone who understands the ins and
outs of integer programming, will also have a lot to say about bounding
stochastic programs. Of course there are differences, but they are smaller
than one might think.
So far, what we have compared is really the problem of bounding the
recourse function with the problem of solving an integer program by branch-
and-bound. Next, let us consider the cutting-plane methods for integer
programs, and compare them with methods like the L-shaped decomposition
method for (continuous) stochastic programs. It must be noted that cutting-

plane methods are hardly ever used in their pure form for solving integer
programs. They are usually combined with other methods. For the sake of
exposition, however, we shall biefly sketch some of the ideas.
When we solve the relaxed linear programming version of (7.1), we
have difficulties because we have increased the solution space. However,
all points that we have added are non-integral. In principle, it is possible
to add extra constraints to the linear programming relaxation to cut off
some of these noninteger solutions, namely those that are not convex
combinations of feasible integral points. These cuts will normally be added in
an iterative manner, very similarly to the way we added cuts in the L-shaped
decomposition method. In fact, the L-shaped decomposition method is known
as Benders’ decomposition in other areas of mathematical programming, and
its original goal was to solve (mixed) integer programming problems. However,
it was not cast in the way we are presenting cuts below.
So, in all its simplicity, a cutting-plane method will run through two major
steps. The first is to solve a relaxed linear program; the second is to evaluate
the solution, and if it is not integral, add cuts that cut away nonintegral
points (including the present solution). These cuts are then added to the
relaxed linear program, and the cycle is repeated. Cuts can be of different
types. Some come from straightforward arithmetic operations based on the
LP solution and the LP constraints. These are not necessarily very tight.
Others are based on structure. For a growing number of problems, knowledge
about some or all facets of the (integer) solution space is becoming available.
By a facet in this case, we understand the following. The solution space of the
relaxed linear program contains all integral feasible points, and none extra. If
we add a minimal number of new inequalities, such that no integral points are
cut off, and such that all extreme points of the new feasible set are integers,
then the intersection between a hyperplane representing such an inequality
and the new set of feasible solutions is called a facet. Facets are sometimes
added as they are found to be violated, and sometimes before the procedure
is started.
How does this relate to the L-shaped decomposition procedure? Let us be
a bit formal. If all costs in a recourse problem are zero, and we choose to use
the L-shaped decomposition method, there will be no optimality cuts, only
feasibility cuts. Such a stochastic linear program could be written as
⎫
min cT x ⎪
⎪
⎬
s.t. Ax = b,
(7.3)
x ≥ 0, ⎪
⎪
⎭
W y(ξ) = h(ξ) − T (ξ)x, y(ξ) ≥ 0.
To use the L-shaped method to solve (7.3), we should begin solving the
problem
min cT x
s.t. Ax = b,
x ≥ 0,
i.e. (7.3) without the last set of constraints added. Then, if the resulting x̂
makes the last set of constraints in (7.3) feasible for all ξ, we are done. If not,
an implied feasibility cut is added.
An integer program, on the other hand, could be written as
⎫
min cT x ⎬
s.t. Ax = b, (7.4)
⎭
xi ∈ {ai , . . . , bi } for all xi .
A cutting-plane procedure for (7.4) will solve the problem with the constraints
a ≤ x ≤ b so that the integrality requirement is relaxed. Then, if the resulting
x̂ is integral in all its elements, we are done. If not, an integrality cut is added.
This cut will, if possible, be a facet of the solution space with all extreme points
integer.
By now, realizing that integrality cuts are also feasibility cuts, the
connection should be clear. Integrality cuts in integer programming are just
a special type of feasibility cuts.
For the bounding version of the L-shaped decomposition method we
combined bounding (with partitioning of the support) with cuts. In the same
way, we can combine branching and cuts in the branch-and-cut algorithm for
integer programs (still deterministic). The idea is fairly simple (but requires
a lot of details to be efficient). For all waiting nodes, before or after we
have solved the relaxed LP, we add an appropriate number of cuts, before
we (re)solve the LP. How many cuts we add will often depend on how well we
know the facets of the (integer) solution space. This new LP will have a smaller
(continuous) solution space, and is therefore likely to give a better result—
either in terms of a nonintegral optimal solution with a higher objective value
(increasing the probability of bounding), or in terms of an integer solution.
So, finally, we have reached the ultimate question. How can all of this be
used to solve integer stochastic programs? Given the simplification that we
have integrality only in the first-stage problem, the procedure is given in
Figure 27. In the procedure we operate with a set of waiting nodes P. These
are nodes in the cut-and-branch tree that are not yet fathomed or bounded.
The procedure feascut was presented earlier in Figure 9, whereas the new
procedure intcut is outlined in Figure 28. Let us try to compare the L-shaped
integer programming method with the continuous one presented in Figure 10.
procedure L-shaped Integer;

begin
Let z := ∞, the best solution so far;
Let P := {initial problem with K := L := 0};
while P = ∅ do begin
Pickproblem(P, P ); P := P \ {P };
repeat (∗ for problem P ∗)
fathom := not (feasible) or (cT x̂ + θ̂ > z);
if not (fathom) then begin
feascut(A, x̂,newcut);
if not (newcut) then intcut(x̂, newcut);
if x̂ integral then begin
Find Q(x̂);
z := min{z, cT x̂ + Q(x̂)};
fathom := (θ̂ ≥ Q(x̂));
if not (fathom) then begin
L := L + 1;
T
Create the cut −βL x + θ ≥ αL ;
end;
end
else begin
Use branching to create 2 new problems P1 and P2 ;
Let P := P ∪ {P1 , P2 };
end;
end;
end;
until fathom;
end; (∗ while ∗)
end;
Figure 27 The L-shaped decomposition method when the first-stage problem

contains integers.
procedure intcut(x̂:real; newcut:boolean);

begin
if violated integrality constraints found then begin
K := K + 1;
T
Create a cut −γK x + θ ≥ δK ;
newcut := true;
end
else newcut := false;
end;
Figure 28 Procedure for generating cuts based on integrality.
3.7.1 Initialization
In the continuous case we started by assuming the existence of an x̂, feasible
in the first stage. It can be found, for example, by solving the expected value
problem. This is not how we start in the integer case. The reason is partly
that finding a feasible solution is more complicated in that setting. On the
other hand, it might be argued that if we hope to solve the integer stochastic
problem, we should be able to solve the expected value problem (or at least
find a feasible solution to the master problem), thereby being able to start out
with a feasible solution (and a z better than ∞). But, even in this case, we shall
not normally be calling procedure master with a feasible solution at hand. If
we have just created a feasibility cut, the present x̂ is not feasible. Therefore
the difference in initialization is natural. This also affects the generation of
feasibility cuts.
3.7.2 Feasibility Cuts

Both approaches operate with feasibility cuts. In the continuous case these
are all implied constraints, needed to make the second-stage problem feasible
for all possible realizations of the random variables. For the integer case,
we still use these, and we add any cuts that are commonly used in branch-
and-cut procedures in integer programming, preferably facets of the solution
space with integral extreme points. To reflect all possible kinds of such cuts
(some concerning second-stage feasibility, some integrality), we use a call
to procedure feascut plus the new procedure intcut. Typically, implied
constraints are based on an x̂ that is nonintegral, and therefore infeasible.
In the end, though, integrality will be there, based on the branching part of
the algorithm, and then the cuts will indeed be based on a feasible (integral)
solution.
3.7.3 Optimality Cuts

The creation of optimality cuts is the same in both cases, since in the integer
case we create such cuts only for feasible (integer) solutions.
3.7.4 Stopping Criteria

The stopping criteria are basically the same, except that what halts the whole
procedure in the continuous case just fathoms a node in the integer case.
3.8 Stochastic Decomposition
Throughout this book we are trying to reserve superscripts on variables and

parameters for outcomes/realizations, and subscripts for time and components
of vectors. This creates difficulties in this section. Since whatever we do will
be wrong compared with our general rules, we have chosen to use the indexing
of the original authors of papers on stochastic decomposition.
The L-shaped decomposition method, outlined in Section 3.2, is a
deterministic method. By that, we mean that if the algorithm is repeated with
the same input data, it will give the same results each time. In contrast to this,
we have what are called stochastic methods. These are methods that ideally
will not give the same results in two runs, even with the same input data. We
say “ideally” because it is impossible in the real world to create truly random
numbers, and hence, in practice, it is possible to repeat a run. Furthermore,
these methods have stopping criteria that are statistical in nature. Normally,
they converge with probability 1.
The reason for calling these methods random is that they are guided by
some random effects, for example samples. In this section we are presenting
the method called stochastic decomposition (SD). The approach, as we present
it, requires relatively complete recourse.
We have until now described the part of the right-hand side in the
recourse problem that does not depend on x by h0 + Hξ. This was done
to combine two different effects, namely to allow certain right-hand side
elements to be dependent, but at the same time to be allowed to work on
independent random variables. SD does not require independence, and hence
we shall replace h0 + Hξ by just ξ, since we no longer make any assumptions
about independence between components of ξ. We do assume, however, that
q(ξ) ≡ q0 , so all randomness is in the right-hand side. The problem to solve
is therefore the following:
min{φ(x) ≡ cT x + Q(x)}
s.t. Ax = b,
x ≥ 0,
where
Q(x) = Q(x, ξ)f (ξ) dξ
with f being the density function for ξ̃ and
Q(x, ξ) = min{q0T y | W y = ξ − T (ξ)x, y ≥ 0}.
Using duality, we get the following alternative formulation of Q(x, ξ):
Q(x, ξ) = max{π T [ξ − T (ξ)x] | π T W ≤ q0T }.
Again we note that ξ and x do not enter the constraints of the dual
formulation, so that if a given ξ and x produce a solvable problem, the problem
is dual feasible for all ξ and x. Furthermore, if π 0 is a dual feasible solution
then
Q(x, ξ) ≥ (π 0 )T [ξ − T (ξ)x]
for any ξ and x, since π 0 is feasible but not necessarily optimal in a
maximization problem. This observation is a central part of SD. Refer back
to our discussion of how to interpret the Jensen lower bound in Section 3.4.1,
where we gave three different interpretations, one of which was approximate
optimization using a finite number of dual feasible bases, rather than all
possible dual feasible bases. In SD we shall build up a collection of dual
feasible bases, and in some of the optimizations use this subset rather than
all possible bases. In itself, this will produce a lower-bounding solution.
But SD is also a sampling technique. By ξk , we shall understand the sample
made in iteration k. At the same time, xk will refer to the iterate (i.e. the
presently best guess of the optimal solution) in iteration k. The first thing to
do after a new sample has been made available is to evaluate Q(xk , ξj ) for
the new iterate and all samples ξj found so far. First we solve for the newest
sample ξk ,
Q(xk , ξk ) = max{π T [ξk − T (ξk )xk ] | π T W ≤ q0T },
to obtain an optimal dual solution πk . Note that this optimization, being the
first involving ξk , is exact. If we let V be the collection of all dual feasible
solutions obtained so far, we now add πk to V . Next, instead of evaluating
Q(xk , ξj ) for j = 1, . . . , k − 1 (i.e. for the old samples) exactly, we simply solve
max{π T (ξj − T (ξj )xk ) | π ∈ V }

π
to obtain πjk . Since V contains a finite number of vectors, this operation is very
simple. Note that for all samples but the new one we perform approximate
optimization using a limited set of dual feasible bases. The situation is
ξ)
ξ
ξ ξ ξ
Figure 29 Illustration of how stochastic decomposition performs exact
optimization for the latest (third) sample point, but inexact optimization for
the two old points.
illustrated in Figure 29. There we see the situation for the third sample point.
We first make an exact optimization for the new sample point, ξ3 , obtaining
a true optimal dual solution π3 . This is represented in Figure 29 by the
supporting hyperplane through ξ3 , Q(x3 , ξ3 ). Afterwards, we solve inexactly
for the two old sample points. There are three bases available for the inexact
optimization. These bases are represented by the three thin lines. As we see,
neither of the two old sample points find their true optimal basis.
˜ = {ξ1 , ξ2 , ξ3 }, with each outcome having the same probability 1 , we
If Ξ(ξ) 3
could now calculate a lower bound on Q(x3 ) by computing
3
1 3 T
L(x3 ) = (π ) (ξj − T (ξj )x3 ).
3 j=1 j
This would be a lower bound because of the inexact optimization performed

for the old sample points. However, the three sample points probably do not
represent the true distribution well, and hence what we have is only something
that in expectation is a lower bound. Since, eventually, this term will converge
towards Q(x), we shall in what follows write
1 k T
k
Q(xk ) = (π ) (ξj − T (ξj )xk ).
k j=1 j
Remember, however, that this is not the true value of Q(xk )—just an estimate.
In other words, we have now observed two major differences from the
exact L-shaped method (page 171). First, we operate on a sample rather
than on all outcomes, and, secondly, what we calculate is an estimate of
a lower bound on Q(xk ) rather than Q(xk ) itself. Hence, since we have a
lower bound, what we are doing is more similar to what we did when we
used the L-shaped decomposition method within approximation schemes, (see
page 204). However, the reason for the lower bound is somewhat different. In
the bounding version of L-shaped, the lower bound was based on conditional
expectations, whereas here it is based on inexact optimization. On the other
hand, we have earlier pointed out that the Jensen lower bound has three
different interpretations, one of which is to use conditional expectations (as
in procedure Bounding L-shaped) and another that is inexact optimization
(as in SD). So what is actually the principal difference?
For the three interpretations of the Jensen bound to be equivalent, the
limited set of bases must come from solving the recourse problem in the points
of conditional expectations. That is not the case in SD. Here the points are
random (according to the sample ξj ). Using a limited number of bases still
produces a lower bound, but not the Jensen lower bound.
Therefore SD and the bounding version of L-shaped are really quite
different. The reason for the lower bound is different, and the objective value
in SD is only a lower bound in terms of expectations (due to sampling). One
method picks the limited number of points in a very careful way, the other
at random. One method has an exact stopping criteria (error bound), the
other has a statistically based stopping rule. So, more than anything else,
they are alternative approaches. If one cannot solve the exact problem, one
either resorts to bounds or to sample-based methods.
In the L-shaped method we demonstrated how to find optimality cuts. We
can now find a cut corresponding to xk (which is not binding and might even
not be a lower bound, although it represents an estimate of a lower bound).
As for the L-shaped method, we shall replace Q(x) in the objective by θ, and
then add constraints. The cut generated in iteration k is given by
1 k T
k
θ≥ (π ) [ξj − T (ξj )x] = αkk + (βkk )T x.
k j=1 j
The double set of indices on α and β indicate that the cut was generated in
iteration k (the subscript) and that it has been updated in iteration k (the
superscript).
In contrast to the L-shaped decomposition method, we must now also look
at the old cuts. The reason is that, although we expect these cuts to be loose
(since we use inexact optimization), they may in fact be far too tight (since
they are based on a sample). Also, being old, they are based on a sample that
is smaller than the present one, and hence, probably not too good. We shall
therefore want to phase them out, but not by throwing them away. Assume
that there exists a lower bound on Q(x, ξ) such that Q(x, ξ) ≥ Q for all x and
ξ. Then the old cuts
θ ≥ αk−1
j + (βjk−1 )T x for j = 1, . . . , k − 1
will be replaced by
⎫
k − 1 k−1 1 ⎬
θ≥ [αj + (βjk−1 )T x] + Q
k k (8.1)
= αkj + (βjk )T x for j = 1, . . . , k − 1. ⎭
For technical reasons, Q = 0 is to be preferred. This inequality is looser than

the previous one, since Q ≤ Q(x, ξ). The master problem now becomes
⎫
min cT x + θ ⎪
⎪
⎬
s.t. Ax = b,
k T (8.2)
−(βj ) x + θ ≥ αj for j = 1, . . . , k, ⎪
k
⎪
⎭
x ≥ 0,
yielding the next iterate xk+1 . Note that, since we assume relatively complete
recourse, there are no feasibility cuts. The above format is the one to be used
for computations. To understand the method better, however, let us show an
alternative version of (8.2) that is less useful computationally but is more
illustrative (see Figure 30 for an illustration):
# &
min φk (x) ≡ cT x + maxj∈{1,···,k} [αkj + (βjk )T x]
s.t. Ax = b, x ≥ 0.
This defines the function φk (x) and shows more clearly than (8.2) that we do
indeed have a function in x that we are minimizing. Also φk (x) is the present
estimate of φ(x) = cT x + Q(x).
The above set-up has one major shortcoming: it might be difficult to
extract a converging subsequence from the sequence xk . A number of changes
therefore have to be made. These make the algorithm look more messy, but
the principles are not lost. To make it simpler (empirically) to extract a
converging subsequence, we shall introduce a sequence of incumbent solutions
xk . Following the incumbent, there will be an index ik that shows in which
iteration the current xk was found.
We initiate the method by setting the counter k := 0, choose an r ∈ (0, 1)
(to be explained later), and let ξ0 := E ξ̃. Thus we solve
min cT x + q0T y
s.t. Ax = b,
W y = ξ0 − T (ξ0 )x, x, y ≥ 0,
φ()
Figure 30 Representation of cT x + Q(x) by a piecewise linear function.
to obtain an initial x1 . We initiate the incumbent x0 = x1 and show that it

was found in iteration 1 by letting i0 = 1.
Next, let us see what is done in a general iteration of the algorithm. First
the counter is increased by letting k := k + 1, and a sample ξk is found. We
now need to find a new cut k as outlined before, and we need to update the
cut that corresponds to the current incumbent. First, we solve
max{π T [ξk − T (ξk )xk ] | π T W ≤ q0T }
to obtain πk . Next, we solve
max{π T (ξk − T (ξk )xk−1 ) | π T W ≤ q0T }
to obtain π k . As before, we then update the set of dual feasible bases by

letting V := V ∪ {πk , π k }.
We then need to make one new cut and update the old cuts. First, the new
cut is made exactly as before. We solve
max{π T [ξj − T (ξj )xk ] | π ∈ V }
to obtain πj for j = 1, . . . , k − 1, and then create the kth cut as
1
k
θ≥ (πj )T [ξj − T (ξj )x] = αkk + (βkk )T x.
k j=1
In addition, we need to update the incumbent cut ik . This is done just the
way we found cut k. We solve
max{π T [ξj − T (ξj )xk−1 ] | π ∈ V }
to obtain π j , and replace the old cut ik by
1
k
θ≥ (π j )T [ξj − T (ξj )x] = αkik−1 + (βikk−1 )T x.
k j=1
The remaining cuts are updated as before by letting
k − 1 k−1 1
θ≥ [αj +(βjk−1 )T x]+ Q = αkj +(βjk )T x for j = 1, . . . , k−1, j = ik−1 .
k k
Now, it is time to check if the incumbent should be changed. We shall use

Figure 31 for illustration, and we shall use the function φk (x) defined earlier.
In the figure we have k = 3. When we entered iteration k, our approximation
of φ(x) was given by φk−1 (x). Our incumbent solution was xk−1 and our
iterate was xk . We show this in the top part of Figure 31 as x2 and x3 .
The position of x2 is somewhat arbitrary, since we cannot know how things
looked in the previous iteration. Therefore φk−1 (xk )−φk−1 (xk−1 ) ≤ 0 was our
approximation of how much we should gain by making xk our new incumbent.
However, xk might be in an area where φk−1 (x) is a bad approximation of
φ(x). The function φk (x), on the other hand, was developed around xk , and
should therefore be good in that area (in addition to being approximately
as good as φk−1 (x) around xk−1 ). This can be seen in the bottom part of
Figure 31, where φ3 (x) is given. The function φ3 (x) is based on three cuts.
One is new, the other two are updates of the two cuts in the top part of the
figure, according to (8.1). Hence φk (xk ) − φk (xk−1 ) (a negative number) is a
measure of how much we actually gained. If
φk (xk ) − φk (xk−1 ) < r[φk−1 (xk ) − φk−1 (xk−1 )],
we gained at least a portion r of what we hoped for, and we let xk := xk and

ik := k. If not, we were not happy with the change, and we let xk := xk−1
and ik := ik−1 . When we have updated the incumbent, we solve a new master
problem to obtain xk+1 and repeat the process.
The stopping criterion for SD is of a statistical nature, and its complexity
is beyond the scope of this book. For a reference, see the end of the chapter.
cx + Q( x)
f2 ( x2 )
f2 ( x )
x
x2 x3
b a
cx + Q( x)
f3 ( x 2 ) f3 ( x )
f3 ( x 3 )
x3 x
x2 c
b a
Figure 31 Calculations to find out if the incumbent should be changed.
3.9 Stochastic Quasi-Gradient Methods
We are still dealing with recourse problems stated in the somewhat more
general form
min f (x) + Q(x, ξ) Pξ̃ (dξ) . (9.1)
x∈X Ξ
This formulation also includes the stochastic linear program with recourse,
letting
X = {x | Ax = b, x ≥ 0},
f (x) = cT x,
Q(x, ξ) = min{(q(ξ))T y | W y = h(ξ) − T (ξ)x, y ≥ 0}.
To describe the so-called stochastic quasi-gradient method (SQG), we

simplify the notation by defining
F (x, ξ) := f (x) + Q(x, ξ)
and hence considering the problem

˜
min Eξ̃ F (x, ξ), (9.2)
x∈X
for which we assume that

˜ is finite and convex in x,
Eξ̃ F (x, ξ) (9.3 i)
X is convex and compact. (9.3 ii)
Observe that for stochastic linear programs with recourse the assump-
tions (9.3) are satisfied if, for instance,
• we have relatively complete recourse, the recourse function Q(x, ξ) is a.s.

finite ∀x, and the components of ξ˜ are square-integrable (i.e. their second
moments exist);
• X = {x | Ax = b, x ≥ 0} is bounded.
Then, starting from some feasible point x0 ∈ X, we may define an iterative

process by
xν+1 = ΠX (xν − ρν v ν ), (9.4)
where v ν is a random vector, ρν ≥ 0 is some step size and ΠX is the projection
onto X, i.e. for y ∈ IRn , with · · · the Euclidean norm,
ΠX (y) = arg min y − x. (9.5)

x∈X
By assumption (9.3 i), ϕ(x) := Eξ̃ F (x, ξ)˜ is convex in x. If this function is
also differentiable with respect to x at any arbitrary point z with the gradient
g := ∇ϕ(z) = ∇x Eξ̃ F (z, ξ̃), then −g is the direction of steepest descent of
˜ in z, and we should probably like to choose −g as the
ϕ(x) = Eξ̃ F (x, ξ)
search direction to decrease our objective. However, this does not seem to be
a practical approach, since, as we know already, evaluating ϕ(x) = Eξ̃ F (x, ξ),˜
as well as ∇ϕ(z) = ∇x Eξ̃ F (z, ξ̃), is a rather cumbersome task.
In the differentiable case we know from Proposition 1.21 on page 81 that,
for a convex function ϕ,
(x − z)T ∇ϕ(z) ≤ ϕ(x) − ϕ(z) (9.6)
has to hold ∀x, z (see Figure 27 in Chapter 1). But, even if the convex function
ϕ is not differentiable at some point z, e.g. if it has a kink there, it is shown
in convex analysis that there exists at least one vector g such that
(x − z)T g ≤ ϕ(x) − ϕ(z) ∀x. (9.7)
Any vector g satisfying (9.7) is called a subgradient of ϕ at z, and the set of all
vectors satisfying (9.7) is called the subdifferential of ϕ at z and is denoted by
∂ϕ(z). If ϕ is differentiable at z then ∂ϕ(z) = {∇ϕ(z)}; otherwise, i.e. in the
nondifferentiable case, ∂ϕ(z) may contain more than one element as shown
for instance in Figure 32. Furthermore, in view of (9.7), it is easily seen that
∂ϕ(z) is a convex set.
If ϕ is convex and g = 0 is a subgradient of ϕ at z then, by (9.7) for λ > 0,
it follows that
ϕ(z + λg) ≥ ϕ(z) + g T (x − z)
= ϕ(z) + g T (λg)
= ϕ(z) + λg2
> ϕ(z).
Hence any subgradient, g ∈ ∂ϕ, such that g = 0 is a direction of ascent,
although not necessarily the direction of steepest ascent as the gradient would
be if ϕ were differentiable in z. However, in contrast to the differentiable case,
−g need not be a direction of strict descent for ϕ in z. Consider for example
the convex function in two variables
ψ(u, v) := |u| + |v|.
Then for ẑ = (0, 3) we have g = (1, 1)T ∈ ∂ψ(ẑ), since for all ε > 0 the
T
gradient ∇ψ(ε, 3) exists and is equal to g. Hence, by (9.6), we have, for all
(u, v),
T T
u ε u−ε 1
− g=
v 3 v−3 1
=u−ε+v−3
≤ |u| + |v| − |ε| − |3|,
Figure 32 Nondifferentiable convex function: subgradients.
which is obviously true ∀ε ≥ 0, such that g is a subgradient in (0, 3)T . Then

for 0 < λ < 3 and ẑ − λg = (−λ, 3 − λ)T it follows that
ψ(ẑ − λg) = 3 = ψ(ẑ),
and therefore, in this particular case, −g is not a strict descent direction for
ψ in ẑ. Nevertheless, as we see in Figure 33, moving from ẑ along the ray
ẑ − λg, λ > 0, for any λ < 3 we would come closer—with respect to the
Euclidean norm—to arg min ψ = {(0, 0)T } than we are at ẑ.
It is worth noting that this property of a subgradient of a convex function
holds in general, and not only for our particular example. Let ϕ be a convex
function and assume that g ∈ ∂ϕ(z), g = 0. Assume further that z ∈ arg min ϕ
and x ∈ arg min ϕ. Then we have for ρ > 0 with the Euclidean norm,
using (9.7),
(z − ρg) − x 2 = (z − x ) − ρg2

= z − x 2 + ρ2 g2 − 2ρg T (z − x )
≤ z − x 2 + ρ2 g2 − 2ρ[ϕ(z) − ϕ(x )].
Since, by our assumption, ϕ(z) − ϕ(x ) > 0, we may choose a step size
ρ = ρ̄ > 0 such that
ρ̄2 g2 − 2ρ̄[ϕ(z) − ϕ(x ] < 0,
implying that z − ρ̄g is closer to x ∈ arg min ϕ than z. This property provides
Figure 33 Decreasing the distance to arg min ψ using a subgradient.
the motivation for the iterative procedures known as subgradient methods,

which minimize convex functions even in the nondifferentiable case.
Obviously for the above procedure (9.4) we may not expect any reasonable
convergence statement without further assumptions on the search direction
v ν and on the step size ρν . Therefore let v ν be a so-called stochastic quasi-
gradient, i.e. assume that
E(v ν | x0 , · · · , xν ) ∈ ∂x Eξ̃ F (xν , ξ̃) + bν , (9.8)
where ∂x denotes the subdifferential with respect to x, as mentioned above,
coinciding with the gradient in the differentiable case.
Let us recall what we are doing here. Starting with some xν , we choose
for (9.4) a random vector v ν . It seems plausible to assume that v ν depends in
˜
some way on ξ˜ (e.g. on an observation ξ ν or on a sample {ξ ν1 , · · · , ξ νNν } of ξ)
and on xν . Then, after the choice of the step size ρν , by (9.4) the next iterate
xν+1 depends on xν . It follows that v ν is itself random. This implies that the
tuples (x0 , x1 , · · · , xν ) are random ∀ν ≥ 1. Hence (9.8) is not yet much of a
requirement. It just says that the expected value of v ν , under the condition
of the path of iterates generated so far, (x0 , · · · , xν ), is to be written as the
sum of a subgradient g ν ∈ ∂x Eξ̃ F (xν , ξ̃) and some vector bν .
Since, by the convexity according to (9.3 i) and applying (9.7),
Eξ̃ F (x∗ , ξ̃) − Eξ̃ F (xν , ξ)
˜ ≥ g νT (x∗ − xν ) (9.9)
for any solution x∗ of (9.2) and any g ν ∈ ∂x Eξ̃ F (xν , ξ),

˜ we have from (9.8)
that
0 ≥ Eξ̃ F (x∗ , ξ)
˜ − E F (xν , ξ̃) ≥ E(v ν | x0 , · · · , xν )T (x∗ − xν ) + γν ,
ξ̃ (9.10)
where
γν = −bνT (x∗ − xν ). (9.11)
Intuitively, if we assume that {xν } converges to x and all v ν are uniformly
ν→∞
bounded, i.e. |v ν | ≤ α for some constant α, we should require that bν −→ 0,
ν→∞
implying γν −→ 0 as well. Observe that the particular choice of a stochastic
subgradient
v ν ∈ ∂x F (xν , ξ ν ), (9.12)
or more generally
1
Nν
vν = wµ , wµ ∈ ∂x F (xν , ξ νµ ), (9.13)
Nν µ=1
where the ξ ν or ξ νµ are independent samples of ξ̃, would yield bν = 0,

γν = 0 ∀ν, provided that the operations of integration and differentiation may
be exchanged, as asserted for example by Proposition 1.2 for the differentiable
case.
Finally, assume that for the step size ρν together with v ν and γν we have
∞
∞

ρν ≥ 0, ρν = ∞, Eξ̃ (ρν |γν | + ρ2ν v ν 2 ) < ∞. (9.14)
ν=0 ν=0
With the choices (9.12) or (9.13), for uniformly bounded v ν this assumption
could obviously be replaced by the step size assumption
∞
∞

ρν ≥ 0, ρν = ∞, ρ2ν < ∞. (9.15)
ν=0 ν=0
With these prerequisites, it can be shown that, under the assumptions (9.3),
(9.8) and (9.14) (or (9.3), (9.12) or (9.13), and (9.15)) the iterative
method (9.4) converges almost surely (a.s.) to a solution of (9.2).
3.10 Solving Many Similar Linear Programs
In both the L-shaped (continuous and integer) and stochastic decomposition

methods we are faced with the problem of solving many similar LPs. This is
most obvious in the L-shaped method: cut formation requires the solution of
many LPs that differ only in the right-hand side and objective. This amount of
work, which is typically enormous, must be performed in each major iteration.

For stochastic decomposition, it is perhaps less obvious that we are facing such
a large workload, but, added over all iterations, we still end up with a large
number of similar LPs.
The problem of solving a large number of similar LPs has attracted attention
for quite a while, in particular when there is only right-hand side randomness.
Therefore let us proceed under the assumption that q(ξ) ≡ q0 .
The major idea is that of bunching. This is a simple idea. If we refer back
to the discussion of the L-shaped decomposition method, we observed that
the dual formulation of the recourse problem was given by
max{π T (h(ξ) − T (ξ)x) | π T W ≤ q0T }. (10.1)

π
What we observe here is that the part that varies, h(ξ) − T (ξ)x, appears only
in the objective. As a consequence, if (10.1) is feasible for one value of x and
ξ, it is feasible for all values of x and ξ. Of course, the problem might be
unbounded (meaning that the primal is infeasible) for some x and ξ. For the
moment we shall assume that that does not occur. (But if it does, it simply
shows that we need a feasibility cut, not an optimality cut).
In a given iteration of the L-shaped decomposition method, x will be fixed,
and all we are interested in is the selection of right-hand sides resulting from
all possible values of ξ. Let us therefore simplify notation, and assume that
we have a selection of right-hand sides B, so that, instead of (10.1), we solve
max{π T h | π T W ≤ q0T } (10.2)

π
for all h ∈ B. Assume (10.2) is solved for one value of h ∈ B with optimal
basis B. Then B is a dual feasible basis for all h ∈ B. Therefore, for all
h ∈ B for which B −1 h ≥ 0, the basis B is also primal feasible, and hence
optimal. The idea behind bunching is simply to start out with some h ∈ B,
find the optimal basis B, and then check B −1 h for all other h ∈ B. Whenever
B −1 h ≥ 0, we have found the optimal solution for that h, and these right-
hand sides are bunched together. We then remove these right-hand sides from
B, and repeat the process, of course with a warm start from B, using the dual
simplex method, for one of the remaining right-hand sides in B. We continue
until all right-hand sides are bunched. That gives us all information needed
to find Q and the necessary optimality cut.
This procedure has been followed up in several directions. An important
one is called trickling down. Again, we start out with B, and we solve (10.2)
for some right-hand side to obtain a dual feasible basis B. This basis is stored
in the root of a search tree that we are about to make. Now, for one h ∈ B at
a time do the following. Start in the root of the tree, and calculate B −1 h. If
B −1 h ≥ 0, register that this right-hand side belongs to the bunch associated
B
1
4
B B
2 3
B B
B
5
B
Figure 34 Example of a bunching tree.
with B, and go to the next h ∈ B. If B −1 h ≥ 0, pick a row for which primal

feasibility is not satisfied. Perform a dual pivot step to obtain a new basis
B (still dual feasible). Create a new node in the search tree associated with
this new B . If the pivot was made in row i, we let the new node be the ith
child of the node containing the previous basis. Continue until optimality is
found. This situation is illustrated in Figure 34, where a total of seven bases
are stored. The numbers on the arc refer to the row where pivoting took place,
the B in the nodes illustrate that there is a basis stored in each node.
This might not seem efficient. However, the real purpose comes after some
iterations. If a right-hand side h is such that B −1 h ≥ 0, and one of the negative
primal variables corresponds to a row index i such that the ith child of the
given node in the search tree already exists, we simply move to that child
without having to price. This is why we use the term trickling down. We try
to trickle a given h as far down in the tree as possible, and only when there
is no negative primal variable that corresponds to a child node of the present
node do we price and pivot explicitly, thereby creating a new branch in the
tree.
Attempts have been made to first create the tree, and then trickle down the
right-hand sides in the finished tree. This was not successful for two reasons.
If we try to enumerate all dual feasible bases, then the tree grows out of hand
(this corresponds to extreme point enumeration), and if we try to find the
correct selection of such bases, then that in itself becomes an overwhelming
problem. Therefore a pre-defined tree does not seem to be a good idea.

It is worth noting that the idea of storing a selection of dual feasible bases,
as was done in the stochastic decomposition method, is also related to the
above approach. In that case the result is a lower bound on Q(x).
A variant of these methods is as follows. Start out with one dual feasible
basis B as in the trickling down procedure. Pick a leading right-hand side.
Now solve the problem corresponding to this leading right-hand side using
the dual simplex method. On pivoting along, create a branch of the search
tree just as for trickling down. The difference is as follows. For each basis
B encountered, check B −1 h for all h ∈ B. Then split the right-hand sides
remaining in B into three sets. Those that have B −1 h ≥ 0 are bunched with
that B, and removed from B. Those that have a primal infeasibility in the
same row as the one chosen to be the pivot row for the leading problem are
kept in B and hence carried along at least one more dual pivot step. The
remaining right-hand sides are left behind in the given node, to be picked up
later on.
When the leading problem has been solved to optimality, and bunching has
been performed with respect to its optimal basis, check if there are any right-
hand sides left in B. If there are, let one of them be the leading right-hand
side, and continue the process. Eventually, when a leading problem has been
solved to optimality, B = ∅. At that time, start backtracking the search tree.
Whenever a selection of right-hand sides left behind is encountered, pick one
of them as the leading problem, and repeat the process. On returning to the
root, and finding there are no right-hand sides left behind there, the process
is finished. All right-hand sides are bunched. Technically, what has now been
done is to traverse the search tree in pre-order.
What remains to be discussed is what to store in the search tree. We have
already seen that the minimal amount to store at any arc in the tree is
the index of the leaving basic column (represented by the numbering of the
children), and the entering column. If that is all we store, we have to pivot, but
not price out, in each step of the trickling down. If we have enough storage, it
is more efficient to store for example the eta-vector (from the revised simplex
method) or the Schur complement (it is not important here if you do not
know what the eta-vector or the Schur complement is). Of course, we could in
principle store B −1 , but for all practical problems that is too much to store.
3.10.1 Randomness in the Objective

The discussion of trickling down etc. was carried out in a setting of right-hand
side randomness only. However, as with many other problems we have faced
in this book, pure objective function randomness can be changed into pure
right-hand side randomness by using linear programming duality. Therefore
the discussions of right-hand side randomness apply to objective function
randomness as well.
Then, one may ask what happens if there is randomness in both the
objective and the right-hand side. Trickling down cannot be performed the
way we have outlined it in that case. This is because a basis that was optimal
for one ξ will, in general, be neither primal nor dual feasible for some other ξ.
On the other hand, the basis may be good, not far from the optimal one. Hence
warm starts based on an old basis, performing a combination of primal and
dual simplex steps, will almost surely be better than solving the individual
LPs from scratch.
Benders’ [1] decomposition is the basis for all decomposition methods in this
chapter. In stochastic programming, as we have seen, it is more common to
refer to Benders’ decomposition as the L-shaped decomposition method. That
approach is outlined in detail in Van Slyke and Wets [63]. An implementation
of the L-shaped decomposition method, called MSLiP, is presented in
Gassmann [31]. It solves multistage problems based on nested decomposition.
Alternative computational methods are also discussed in Kall [44].
The regularized decomposition method has been implemented under the
name QDECOM. For further details on the method and QDECOM, in
particular for a special technique to solve the master (3.6), we refer to the
original publication of Ruszczyński [61]; the presentation in this chapter is
close to the description in his recent paper [62].
Some attempts have also been made to use interior point methods. As
examples consider Birge and Qi [7], Birge and Holmes [6], Mulvey and
Ruszczyński [60] and Lustig, Mulvey and Carpenter [55]. The latter two
combine interior point methods with parallel processing.
Parallel techniques have been tried by others as well; see e.g. Berland [2]
and Jessup, Yang and Zenios [42]. We shall mention some others in Chapter 6.
The idea of combining branch-and-cut from integer programming with
primal decomposition in stochastic programming was developed by Laporte
and Louveaux [53]. Although the method is set in a strict setting of
integrality only in the first stage, it can be expanded to cover (via a
reformulation) multistage problems that possess the so-called block-separable
recourse property, see Louveaux [54] for details.
Stochastic quasi-gradient methods were developed by Ermoliev [20, 21],
and implemented by, among others, Gaivoronski [27, 28]. Besides stochastic
quasi-gradients several other possibilities for constructing stochastic descent
directions have been investigated, e.g. in Marti [57] and in Marti and
Fuchs [58, 59].
The Jensen lower bound was developed in 1906 [41]. The Edmundson–
Madansky upper bound is based on work by Edmundson [19] and

Madansky [56]. It has been extended to the multidimensional case by
Gassmann and Ziemba [33]; see also Hausch and Ziemba [36] and Edirisinghe
and Ziemba [17, 18]. Other references in this area include Huang, Vertinsky
and Ziemba [39] and Huang, Ziemba and Ben-Tal [40]. The Edmundson–
Madansky bound was generalized to the case of stochastically dependent
components by Frauendorfer [23].
The piecewise linear upper bound is based on two independent approaches,
namely those of Birge and Wets [11] and Wallace [66]. These were later
combined and strengthened in Birge and Wallace [8].
There is a large collection of bounds based on extreme measures (see e.g.
Dulá [12, 13], Hausch and Ziemba [36], Huang, Ziemba and Ben-Tal [40] and
Kall [48]). Both the Jensen and Edmundson–Madansky bounds can be put
into this category. For a fuller description of these methods, consult Birge and
Wets [10], Dupačová [14, 15, 16] and Kall [47]; more on extreme measures
may be found in Karr [51] and Kemperman [52].
Bounds can also be found when limited information is available. Consult
e.g. Birge and Dulá [5]. An upper bound based on structure can be found in
Wallace and Yan [68].
Stochastic decomposition was developed by Higle and Sen [37, 38].
The ideas presented about trickling down and similar methods come from
different authors, in particular Wets [70, 72], Haugland and Wallace [35],
Wallace [65, 64] and Gassmann and Wallace [32]. A related approach is that
of Gartska and Rutenberg [29], which is based on parametric optimization.
Partitioning has been discussed several times during the years. Some general
ideas are presented in Birge and Wets [9]. More detailed discussions (with
numerical results), on which the discussions in this book are based, can be
found in Frauendorfer and Kall [26] and Berland and Wallace [3, 4]. Other
texts about approximation by discretization include for example those of
Kall [43, 45, 46], and Kall, Ruszczyński and Frauendorfer [49].
When partitioning the support to tighten bounds, it is possible to use more
complicated cells than we have done. For example, Frauendorfer [24, 25] uses
simplices. It is also possible to use more general polyhedra.
For simple recourse, the separability of the objective, which facilitates
computations substantially, was discovered by Wets [69]. The ability to replace
the Edmundson–Madansky upper bound by the true objective’s value was
discussed in Kall and Stoyan [50]. Wets [71] has derived a special pivoting
scheme that avoids the tremendous increase of the problem size known from
general recourse problems according to the number of blocks (i.e. realizations).
See also discussions by Everitt and Ziemba [22] and Hansotia [34].
The fisheries example in the beginning of the chapter comes from
Wallace [67]. Another application concerning natural resources is presented
by Gassmann [30].
Exercises
1. The second-stage constraints of a two-stage problem look as follows:

1 3 −1 0 −6 5 −1 0
y= ξ+ x
2 −1 2 1 −4 0 2 4
y≥0
where ξ̃ is a random variable with support Ξ = [0, 1]. Write down the LP
(both primal and dual formulation) needed to check if a given x produces
a feasible second-stage problem. Do it in such a way that if the problem
is not feasible, you obtain an inequality in x that cuts off the given x. If
you have access to an LP code, perform the computations, and find the
inequality explicitly for x̂ = (1, 1, 1)T .
2. Look back at problem (4.1) we used to illustrate the bounds. Add one extra
constraint, namely
xraw1 ≤ 40.
(a) Find the Jensen lower bound after this constraint has been added.
(b) Find the Edmundson–Madansky upper bound.
(c) Find the piecewise linear upper bound.
(d) Try to find a good variable for partitioning.
3. Assume that you are facing a decision problem where randomness is

involved. You have no idea about the distribution of the random variables
involved. However, you can obtain samples from the distribution by running
an expensive experiment. You have decided to use stochastic decomposition
to solve the problem, but are concerned that you may not be able to
perform enough experiments for convergence to take place. The cost of a
single experiment is much higher than the costs involved in the arithmetic
operations of the algorithm.
(a) Argue why (or why not) it is reasonable to use stochastic decomposition
under the assumptions given. (You can assume that all necessary
convexity is there.)
(b) What changes could you suggest in stochastic decomposition in order
to (at least partially) overcome the fact that samples are so expensive?
4. Let ϕ be a convex function. Show that

x ∈ arg min ϕ iff 0 ∈ ∂ϕ(x ).
(See the definition following (9.7).
5. Show that for a convex function ϕ and any arbitrary z the subdifferential
∂ϕ(z) is a convex set. [Hint: For any subgradient (9.7) has to hold.]
6. Assume that you are faced with a large number of linear programs that
you need to solve. They represent all recourse problems in a two-stage
stochastic program. There is randomness in both the objective function
and the right-hand side, but the random variables affecting the objective
are different from, and independent of, the random variables affecting the
right-hand side.
(a) Argue why (or why not) it is a good idea to use some version of
bunching or trickling down to solve the linear programs.
(b) Given that you must use bunching or trickling down in some version,
how would you organize the computations?
7. First consider the following integer programming problem:
min{cx | Ax ≤ h, xi ∈ {0, . . . , bi } ∀i}.

x
Next, consider the problem of finding Eφ(x̃), with
φ(x) = min{cy | Ay ≤ h, 0 ≤ y ≤ x}.

y
(a) Assume that you solve the integer program with branch-and-bound.
Your first step is then to solve the integer program above, but with
xi ∈ {0, . . . , bi } ∀i replaced by 0 ≤ x ≤ b. Assume that you get x̂.
Explain why x̂ can be a good partitioning point if you wanted to find
Eφ(x̃) by repeatedly partitioning the support, and finding bounds on
each cell. [Hint: It may help to draw a little picture.]
(b) We have earlier referred to Figure 18, stating that it can be seen
as both the partitioning of the support for the stochastic program,
and partitioning the solution space for the integer program. Will the
number of cells be largest for the integer or the stochastic program
above? Note that there is not necessarily a clear answer here, but you
should be able make arguments on the subject. Question (a) may be
of some help.
8. Look back at Figure 17. There we replaced one distribution by two others:
one yielding an upper bound, and one a lower bound. The possible values
for these two new distributions were not the same. How would you use the
ideas of Jensen and Edmundson–Madansky to achieve, as far as possible,
the same points? You can assume that the distribution is bounded. [Hint:
The Edmundson–Madansky distribution will have two more points than
the Jensen distribution.]
References
[1] Benders J. F. (1962) Partitioning procedures for solving mixed-variables

programming problems. Numer. Math. 4: 238–252.
[2] Berland N. J. (1993) Stochastic optimization and parallel processing. PhD
thesis, Department of Informatics, University of Bergen.
[3] Berland N. J. and Wallace S. W. (1993) Partitioning of the support to
tighten bounds on stochastic PERT problems. Working paper, Department
of Managerial Economics and Operations Research, Norwegian Institute of
Technology, Trondheim.
[4] Berland N. J. and Wallace S. W. (1993) Partitioning the support to
tighten bounds on stochastic linear programs. Working paper, Department
of Managerial Economics and Operations Research, Norwegian Institute of
Technology, Trondheim.
[5] Birge J. R. and Dulá J. H. (1991) Bounding separable recourse functions
with limited distribution information. Ann. Oper. Res. 30: 277–298.
[6] Birge J. R. and Holmes D. (1992) Efficient solution of two stage stochastic
linear programs using interior point methods. Comp. Opt. Appl. 1: 245–276.
[7] Birge J. R. and Qi L. (1988) Computing block-angular Karmarkar
projections with applications to stochastic programming. Management Sci.
pages 1472–1479.
[8] Birge J. R. and Wallace S. W. (1988) A separable piecewise linear upper
bound for stochastic linear programs. SIAM J. Control and Optimization
26: 725–739.
[9] Birge J. R. and Wets R. J.-B. (1986) Designing approximation schemes
for stochastic optimization problems, in particular for stochastic programs
with recourse. Math. Prog. Study 27: 54–102.
[10] Birge J. R. and Wets R. J.-B. (1987) Computing bounds for stochastic
programming problems by means of a generalized moment problem. Math.
Oper. Res. 12: 149–162.
[11] Birge J. R. and Wets R. J.-B. (1989) Sublinear upper bounds for stochastic
programs with recourse. Math. Prog. 43: 131–149.
[12] Dulá J. H. (1987) An upper bound on the expectation of sublinear functions
of multivariate random variables. Preprint, CORE.
[13] Dulá J. H. (1992) An upper bound on the expectation of simplicial functions
of multivariate random variables. Math. Prog. 55: 69–80.
[14] Dupačová J. (1976) Minimax stochastic programs with nonconvex
nonseparable penalty functions. In Prékopa A. (ed) Progress in Operations
Research, pages 303–316. North-Holland, Amsterdam.
[15] Dupačová J. (1980) Minimax stochastic programs with nonseparable
penalties. In Iracki K., Malanowski K., and Walukiewicz S. (eds)
Optimization Techniques, Part I, volume 22 of Lecture Notes in Contr.
Inf. Sci., pages 157–163. Springer-Verlag, Berlin.
[16] Dupačová J. (1987) The minimax approach to stochastic programming and

an illustrative application. Stochastics 20: 73–88.
[17] Edirisinghe N. C. P. and Ziemba W. T. (1994) Bounding the expectation of
a saddle function, with application to stochastic programming. Math. Oper.
Res. 19: 314–340.
[18] Edirisinghe N. C. P. and Ziemba W. T. (1994) Bounds for two-stage
stochastic programs with fixed recourse. Math. Oper. Res. 19: 292–313.
[19] Edmundson H. P. (1956) Bounds on the expectation of a convex function of
a random variable. Technical Report Paper 982, The RAND Corporation.
[20] Ermoliev Y. (1983) Stochastic quasigradient methods and their application
to systems optimization. Stochastics 9: 1–36.
[21] Ermoliev Y. (1988) Stochastic quasigradient methods. In Ermoliev Y. and
Wets R. J.-B. (eds) Numerical Techniques for Stochastic Optimization,
pages 143–185. Springer-Verlag.
[22] Everitt R. and Ziemba W. T. (1979) Two-period stochastic programs with
simple recourse. Oper. Res. 27: 485–502.
[23] Frauendorfer K. (1988) Solving SLP recourse problems with arbitrary
multivariate distributions—the dependent case. Math. Oper. Res. 13: 377–
394.
[24] Frauendorfer K. (1989) A simplicial approximation scheme for convex
two-stage stochastic programming problems. Manuscript, Inst. Oper. Res.,
University of Zurich.
[25] Frauendorfer K. (1992) Stochastic Two-Stage Programming, volume 392 of
Lecture Notes in Econ. Math. Syst. Springer-Verlag, Berlin.
[26] Frauendorfer K. and Kall P. (1988) A solution method for SLP recourse
problems with arbitrary multivariate distributions – the independent case.
Probl. Contr. Inf. Theory 17: 177–205.
[27] Gaivoronski A. (1988) Interactive program SQG-PC for solving stochastic
programming problems on IBM PC/XT/AT compatibles—user guide.
Working Paper WP-88-11, IIASA, Laxenburg.
[28] Gaivoronski A. (1988) Stochastic quasigradient methods and their
implementation. In Ermoliev Y. and Wets R. J.-B. (eds) Numerical
Techniques for Stochastic Optimization, pages 313–351. Springer-Verlag.
[29] Gartska S. J. and Rutenberg D. P. (1973) Computation in discrete
stochastic programs with recourse. Oper. Res. 21: 112–122.
[30] Gassmann H. I. (1989) Optimal harvest of a forest in the presence of
uncertainty. Can. J. Forest Res. 19: 1267–1274.
[31] Gassmann H. I. (1990) MSLiP: A computer code for the multistage
stochastic linear programming problem. Math. Prog. 47: 407–423.
[32] Gassmann H. I. and Wallace S. W. (1993) Solving linear programs with
multiple right hand sides: Pivoting and ordering schemes. Working paper,
Department of Economics, Norwegian Institute of Technology, Trondheim.
[33] Gassmann H. and Ziemba W. T. (1986) A tight upper bound for the
expectation of a convex function of a multivariate random variable. Math.

Prog. Study 27: 39–53.
[34] Hansotia B. J. (1980) Stochastic linear programs with simple recourse: The
equivalent deterministic convex program for the normal, exponential and
Erlang cases. Naval. Res. Logist. Quart. 27: 257–272.
[35] Haugland D. and Wallace S. W. (1988) Solving many linear programs that
differ only in the righthand side. Eur. J. Oper. Res. 37: 318–324.
[36] Hausch D. B. and Ziemba W. T. (1983) Bounds on the value of information
in uncertain decision problems, II. Stochastics 10: 181–217.
[37] Higle J. L. and Sen S. (1991) Stochastic decomposition: An algorithm for
two stage stochastic linear programs with recourse. Math. Oper. Res. 16:
650–669.
[38] Higle J. L. and Sen S. (1991) Statistical verification of optimality conditions
for stochastic programs with recourse. Ann. Oper. Res. 30: 215–240.
[39] Huang C. C., Vertinsky I., and Ziemba W. T. (1977) Sharp bounds on the
value of perfect information. Oper. Res. 25: 128–139.
[40] Huang C. C., Ziemba W. T., and Ben-Tal A. (1977) Bounds on the
expectation of a convex function of a random variable: With applications
to stochastic programming. Oper. Res. 25: 315–325.
[41] Jensen J. L. (1906) Sur les fonctions convexes et les inégalités entre les
valeurs moyennes. Acta Math. 30: 173–177.
[42] Jessup E. R., Yang D., and Zenios S. A. (1993) Parallel factorization
of structured matrices arising in stochastic programming. Report 93-02,
Department of Public and Business Administration, University of Cyprus,
Nicosia, Cyprus.
[43] Kall P. (1974) Approximations to stochastic programs with complete fixed
recourse. Numer. Math. 22: 333–339.
[44] Kall P. (1979) Computational methods for solving two-stage stochastic
linear programming problems. Z. Angew. Math. Phys. 30: 261–271.
[45] Kall P. (1986) Approximation to optimization problems: An elementary
review. Math. Oper. Res. 11: 9–18.
[46] Kall P. (1987) On approximations and stability in stochastic programming.
In Guddat J., Jongen H. T., Kummer B., and Nožička F. (eds) Parametric
Optimization and Related Topics, pages 387–407. Akademie-Verlag, Berlin.
[47] Kall P. (1988) Stochastic programming with recourse: Upper bounds and
moment problems—a review. In Guddat J., Bank B., Hollatz H., Kall
P., Klatte D., Kummer B., Lommatzsch K., Tammer K., Vlach M., and
Zimmermann K. (eds) Advances in Mathematical Optimization (Dedicated
to Prof. Dr. Dr. hc. F. Nožička), pages 86–103. Akademie-Verlag, Berlin.
[48] Kall P. (1991) An upper bound for SLP using first and total second
moments. Ann. Oper. Res. 30: 267–276.
[49] Kall P., Ruszczyński A., and Frauendorfer K. (1988) Approximation
techniques in stochastic programming. In Ermoliev Y. M. and Wets R.
J.-B. (eds) Numerical Techniques for Stochastic Optimization, pages 33–64.

Springer-Verlag, Berlin.
[50] Kall P. and Stoyan D. (1982) Solving stochastic programming problems
with recourse including error bounds. Math. Operationsforsch. Statist., Ser.
Opt. 13: 431–447.
[51] Karr A. F. (1983) Extreme points of certain sets of probability measures,
with applications. Math. Oper. Res. 8: 74–85.
[52] Kemperman J. M. B. (1968) The general moment problem, a geometric
approach. Ann. Math. Statist. 39: 93–122.
[53] Laporte G. and Louveaux F. V. (1993) The integer L-shaped method for
stochastic integer programs. Oper. Res. Lett. 13: 133–142.
[54] Louveaux F. V. (1986) Multistage stochastic linear programs with block
separable recourse. Math. Prog. Study 28: 48–62.
[55] Lustig I. J., Mulvey J. M., and Carpenter T. J. (1991) Formulating two-
stage stochastic programs for interior point methods. Oper. Res. 39: 757–
770.
[56] Madansky A. (1959) Bounds on the expectation of a convex function of a
multivariate random variable. Ann. Math. Statist. 30: 743–746.
[57] Marti K. (1988) Descent Directions and Efficient Solutions in Discretely
Distributed Stochastic Programs, volume 299 of Lecture Notes in Econ.
Math. Syst. Springer-Verlag, Berlin.
[58] Marti K. and Fuchs E. (1986) Computation of descent directions
and efficient points in stochastic optimization problems without using
derivatives. Math. Prog. Study 28: 132–156.
[59] Marti K. and Fuchs E. (1986) Rates of convergence of semi-stochastic
approximation procedures for solving stochastic optimization problems.
Optimization 17: 243–265.
[60] Mulvey J. M. and Ruszczyński A. (1992) A new scenario decomposition
method for large-scale stochastic optimization. Technical Report SOR-91-
19, Princeton University, Princeton, New Jersey.
[61] Ruszczyński A. (1986) A regularized decomposition method for minimizing
a sum of polyhedral functions. Math. Prog. 35: 309–333.
[62] Ruszczyński A. (1993) Regularized decomposition of stochastic programs:
Algorithmic techniques and numerical results. Working Paper WP-93-21,
IIASA, Laxenburg.
[63] Van Slyke R. and Wets R. J.-B. (1969) L-shaped linear programs with
applications to optimal control and stochastic linear programs. SIAM J.
Appl. Math. 17: 638–663.
[64] Wallace S. W. (1986) Decomposing the requirement space of a
transportation problem into polyhedral cones. Math. Prog. Study 28: 29–47.
[65] Wallace S. W. (1986) Solving stochastic programs with network recourse.
Networks 16: 295–317.
[66] Wallace S. W. (1987) A piecewise linear upper bound on the network
recourse function. Math. Prog. 38: 133–146.

[67] Wallace S. W. (1988) A two-stage stochastic facility location problem with
time-dependent supply. In Ermoliev Y. and Wets R. J.-B. (eds) Numerical
Techniques in Stochastic Optimization, pages 489–514. Springer-Verlag,
Berlin.
[68] Wallace S. W. and Yan T. (1993) Bounding multistage stochastic linear
programs from above. Math. Prog. 61: 111–130.
[69] Wets R. (1966) Programming under uncertainty: The complete problem. Z.
Wahrsch. theorie u. verw. Geb. 4: 316–339.
[70] Wets R. (1983) Stochastic programming: Solution techniques and
approximation schemes. In Bachem A., Grötschel M., and Korte B. (eds)
Mathematical Programming: The State-of-the-Art, Bonn 1982, pages 566–
603. Springer-Verlag, Berlin.
[71] Wets R. J.-B. (1983) Solving stochastic programs with simple recourse.
Stochastics 10: 219–242.
[72] Wets R. J.-B. (1988) Large scale linear programming techniques. In
Ermoliev Y. and Wets R. J.-B. (eds) Numerical Techniques for Stochastic
Optimization, pages 65–93. Springer-Verlag.
4
Probabilistic Constraints
As we have seen in Sections 1.5 and 1.6, at least under appropriate assump-
tions, chance-constrained problems such as (4.21), or particularly (4.23), as
well as recourse problems such as (4.11), or particularly (4.16), (all from Chap-
ter 1), appear as ordinary convex smooth mathematical programming prob-
lems. This might suggest that these problems may be solved using known
nonlinear programming methods. However, this viewpoint disregards the fact
that in the direct application of those methods to problems like
˜
minx∈X Eξ̃ cT (ξ)x
s.t. P ({ξ | T (ξ)x ≥ h(ξ)}) ≥ α
or
˜
min Eξ̃ {cT x + Q(x, ξ)}
x∈X
where
Q(x, ξ) = min{q T y | W y ≥ h(ξ) − T (ξ)x, y ∈ Y },
we had repeatedly to obtain gradients and evaluations for functions like
P ({ξ | T (ξ)x ≥ h(ξ)})
or
˜
Eξ̃ {cT x + Q(x, ξ)}.
Each of these evaluations requires multivariate numerical integration, so that
up to now this seems to be outside of the set of efficiently solvable problems.
Hence we may try to follow the basic ideas of some of the known nonlinear
programming methods, but at the same time we have to find ways to evade
the exact evaluation of the integral functions contained in these problems.
On the other hand we also know from the example illustrated in Figure 18
of Chapter 1 that chance constraints may easily define nonconvex feasible
sets. This leads to severe computational problems if we intend to find a global
optimum. There is one exception to this general problem worth mentioning.
Proposition 4.1 The feasible set
B(1) := {x | P ({ξ | T (ξ)x ≥ h(ξ)}) ≥ 1}
is convex.
Proof Assume that x, y ∈ B(1) and that λ ∈ (0, 1). Then for Ξx := {ξ |
T (ξ)x ≥ h(ξ)} and Ξy := {ξ | T (ξ)y ≥ h(ξ)} we have P (Ξx ) = P (Ξy ) = 1.
As is easily shown, this implies for Ξ∩ := Ξx ∩ Ξy that P (Ξ∩ ) = 1.
Obviously, for z := λx + (1 − λ)y we have T (ξ)z ≥ h(ξ) ∀ξ ∈ Ξ∩ such that
{ξ | T (ξ)z ≥ h(ξ)} ⊃ Ξ∩ . Hence we have z ∈ B(1). 2
Considering once again the example illustrated in Figure 18 in Section 1.6,

we observe that if we had required a reliability α > 93%, the feasible set
would have been convex. This is a consequence of Proposition 4.1 for discrete
distributions, and may be stated as follows.
Proposition 4.2 Let ξ˜ have a finite discrete distribution described by P (ξ =
ξ j ) = pj , j = 1, · · · , r (pj > 0 ∀j). Then for α > 1 − minj∈{1,···,r} pj the
feasible set
B(α) := {x | P ({ξ | T (ξ)x ≥ h(ξ)}) ≥ α}
is convex.
Proof: The assumption on α implies that B(α) = B(1) (see Exercises at the
end of this chapter). 2
In conclusion, for discrete distributions and reliability levels chosen “high

enough” we have a convex problem. Replacing Eξ̃ c(ξ) ˜ by c, we then simply
have to solve the linear program (provided that X is convex polyhedral)
minx∈X cT x
s.t. T (ξ j )x ≥ h(ξ j ), j = 1, · · · , r.
This observation may be helpful for some particular chance-constrained
problems with discrete distributions. However, it also tells us that for chance-
constrained problems stated with continuous-type distributions and requiring
a reliability level α < 1, we cannot expect—as discussed in Section 3.5 for the
recourse problem—approximating the continuous distribution by successively
refined discrete ones to be a successful approach. The reason should now be
obvious: refining the discrete (approximating) distributions would imply at
some stage that minj pj < 1−α such that the “approximating” problems were
likely to become nonconvex—even if the original problem with its continuous
distribution were convex. And approximating convex problems by nonconvex
ones should certainly not be our aim!
In the next two sections we shall describe under special assumptions
(multivariate normal distributions) how chance-constrained programs can
PROBABILISTIC CONSTRAINTS 245
be treated computationally. In particular, we shall verify that, under our

assumptions, a program with joint chance constraints becomes a convex
program and that programs with separate chance contraints may be
reformulated to become a deterministic convex program amenable to standard
nonlinear programming algorithms.
4.1 Joint Chance Constrained Problems
Let us concentrate on the particular stochastic linear program

⎫
min cT x ⎪
⎪
⎬
s.t. P ({ξ | T x ≥ ξ}) ≥ α
(1.1)
Dx = d, ⎪
⎪
⎭
x ≥ 0.
For this problem we know from Propositions 1.5–1.7 in Section 1.6 that if the
distribution function F is quasi-concave then the feasible set B(α) is a closed
convex set.
Under the assumption that ξ˜ has a (multivariate) normal distribution,
we know that F is even log-concave. We therefore have a smooth convex
program. For this particular case there have been attempts to adapt penalty
and cutting-plane methods to solve (1.1). Further, variants of the reduced
gradient method as sketched in Section 1.8.2 have been designed.
These approaches all attempt to avoid the “exact” numerical integration
associated with the evaluation of F (T x) = P ({ξ | T x ≥ ξ}) and its gradient
∇x F (T x) by relaxing the probabilistic constraint
P ({ξ | T x ≥ ξ}) ≥ α.
To see how this may be realized, let us briefly sketch one iteration of the
reduced gradient method’s variant implemented in PROCON, a computer
program for minimizing a function under PRObabilistic CONstraints.
With the notation
G(x) := P ({ξ | T x ≥ ξ}),
let x be feasible in
⎫
min cT x ⎪
⎪
⎬
s.t. G(x) ≥ α,
(1.2)
Dx = d, ⎪
⎪
⎭
x ≥ 0,
and—assuming D to have full row rank—let D be partitioned as D = (B, N )
into basic and nonbasic parts and accordingly partition xT = (y T , z T ), cT =
(f T , g T ) and a descent direction wT = (uT , v T ). Assume further that for some

tolerance ε > 0,
yj > ε ∀j (strict nondegeneracy). (1.3)

T T T
Then the search direction w = (u , v ) is determined by the linear program
⎫
max τ ⎪
⎪
⎪
⎪
s.t. f Tu + g T v ≤ −τ, ⎪
⎪
⎬
∇y G(x)T u + ∇z G(x)T v ≥ θτ if G(x) ≤ α + ε,
(1.4)
Bu + N v = 0, ⎪
⎪
⎪
⎪
vj ≥ 0 if zj ≤ ε, ⎪
⎪
⎭
v∞ ≤ 1,
where θ > 0 is a fixed parameter as a weight for the directional derivatives
of G and v∞ = maxj {|vj |}. According to the above assumption, we have
from (1.4)
u = −B −1 N v,
which renders (1.4) into the linear program

⎫
max τ ⎪
⎪
⎪
⎪
s.t. rT v ≤ −τ, ⎬
sT v ≥ θτ if G(x) ≤ α + ε, (1.5)
⎪
⎪
vj ≥ 0 if zj ≤ ε, ⎪
⎪
⎭
v∞ ≤ 1,
where obviously
rT = g T − f T B −1 N,
sT = ∇z G(x)T − ∇y G(x)T B −1 N
are the reduced gradients of the objective and the probabilistic constraint
function. Problem (1.5)—and hence (1.4)—is always solvable owing to its
nonempty and bounded feasible set. Depending on the obtained solution
(τ ∗ , u∗T , v ∗T ) the method proceeds as follows.
Case 1 When τ ∗ = 0, ε is replaced by 0 and (1.5) is solved again. If

τ ∗ = 0 again, the feasible solution xT = (y T , z T ) is obviously optimal.
Otherwise the steps of case 2 below are carried out, starting with
the original ε > 0.
Case 2 When 0 < τ ∗ ≤ ε, the following cycle is entered:
Step 1 Set ε := 0.5ε.
Step 2 Solve (1.5). If still τ ∗ ≤ ε, go to step 1; otherwise, case 3

applies.
Case 3 When τ ∗ > ε, w∗T = (u∗T , v ∗T ) is accepted as search direction.
If a search direction w∗T = (u∗T , v ∗T ) has been found, a line search follows
using bisection. Since the line search in this case amounts to determining the
intersection of the ray x + µw∗ , µ ≥ 0 with the boundary bdB(α) within the
tolerance ε, the evaluation of G(x) becomes important. For this purpose a
special Monte Carlo technique is used, which allows efficient computation of
upper and lower bounds of G(x) as well as the gradient ∇G(x).
If the next iterate x̌, resulting from the line search, still satisfies strict
nondegeneracy, the whole step is repeated with the same partition of D into
basic and nonbasic parts; otherwise, a basis exchange is attempted to reinstall
strict nondegeneracy for a new basis.
4.2 Separate Chance Constraints
Let us now consider stochastic linear programs with separate (or single) chance
constraints as introduced at the end of Section 1.4. Using the formulation given
there we are dealing with the problem

minx∈X Eξ̃ cT (ξ)x˜
(2.1)
s.t. P ({ξ | Ti (ξ)x ≥ hi (ξ)}) ≥ αi , i = 1, · · · , m,
where Ti (ξ) is the ith row of T (ξ). The main question is whether or under
what assumptions the feasibility set defined by any one of the constraints
in (2.1),
{x | P ({ξ | Ti (ξ)x ≥ hi (ξ)} ≥ αi },
is convex. As we know from Section 1.6, this question is very simple to answer
˜
for the special case where Ti (ξ) ≡ Ti , i.e. where only the right-hand side hi (ξ)
is random. That is, with Fi the distribution function of hi (ξ),˜
{x | P ({ξ | Ti x ≥ hi (ξ)}) ≥ αi } = {x | Fi (Ti x) ≥ αi }

= {x | Ti x ≥ Fi−1 (αi )}.
It follows that the feasibility set for this particular chance constraint is just
the feasibility set of an ordinary linear constraint.
For the general case let us first simplify the notation as follows. Let
Bi (αi ) := {x | P ({(tT , h)T | tT x ≥ h}) ≥ αi },

where (t̃T , h̃)T is a random vector. Assume now that (t̃T , h̃)T has a joint
normal distribution with expectation µ ∈ IRn+1 and (n + 1) × (n + 1)
covariance matrix S. For any fixed x, let ζ̃(x) := xT t̃ − h̃. It follows that
our feasible set may be rewritten in terms of the random variable ζ̃(x) as
Bi (αi ) = {x | P (ζ(x) ≥ 0) ≥ αi }. From probability theory, we know
that, because ζ̃(x) is a linear combination of jointly normally distributed
random variables, it has a (one-dimensional)
n normal distribution function
Fζ̃ with expectation mζ̃ (x) = µ x
j=1 j j − µn+1 , and, using the (n + 1)-
vector z(x) := (x1 , · · · , xn , −1) , the variance σζ̃2 (x) = z(x)T Sz(x). Since the
T
covariance matrix S of a (nondegenerate) multivariate normal distribution

is positive-definite, it follows that the variance σζ̃2 (x) and, as can be easily
shown, the standard deviation σζ̃ (x) are convex in x (and σζ̃ (x) > 0 ∀x in
view of zn+1 (x) = −1). Hence we have
Bi (αi ) = {x | P (ζ(x) ≥ 0) ≥ αi }
) *
ζ(x) − mζ̃ (x) −mζ̃ (x)

= x P ≥ ≥ αi .
σζ̃ (x) σζ̃ (x)
Observing that for the normally distributed random variable ζ̃(x) the random
variable [ζ̃(x) − mζ̃ (x)]/σζ̃ (x) has the standard normal distribution function
Φ, it follows that
! + −m (x) , "
ζ̃
Bi (αi ) = x1 − Φ ≥ αi .
σζ̃ (x)
Hence
! + −m (x) , "
ζ̃
Bi (αi ) = x1 − Φ ≥ αi
σζ̃ (x)
! + −m (x) , "
ζ̃
= xΦ ≤ 1 − αi
σζ̃ (x)
! −m (x) "

≤ Φ−1 (1 − αi )
ζ̃
= x
σ (x)
! ζ̃ "

= x − Φ−1 (1 − αi )σζ̃ (x) − mζ̃ (x) ≤ 0 .
Here mζ̃ (x) is linear affine in x and σζ̃ (x) is convex in x. Therefore the left-
hand side of the constraint
−Φ−1 (1 − αi )σζ̃ (x) − mζ̃ (x) ≤ 0
is convex iff Φ−1 (1 − αi ) ≤ 0, which is exactly the case iff αi ≥ 0.5. Hence
we have, under the assumption of normal distributions and αi ≥ 0.5, instead
of (2.1) a deterministic convex program with constraints of the type
−Φ−1 (1 − αi )σζ̃ (x) − mζ̃ (x) ≤ 0,
which can be solved with standard tools of nonlinear programming.
4.3 Bounding Distribution Functions
In Section 4.1 we mentioned that particular methods have been developed to

compute lower and upper bounds for the function
G(x) := P ({ξ | T x ≥ ξ}) = Fξ̃ (T x)
contained in the constraints of problem (1.1). Here Fξ̃ (·) denotes the
˜ In the following we sketch some
distribution function of the random vector ξ.
ideas underlying these bounding methods. For a more technical presentation,
the reader should consult the references provided below.
To simplify the notation,let us assume that ξ̃ is a random vector with a
support Ξ ⊂ IRn . For any z ∈ IRn , we have
Fξ̃ (z) = P ({ξ | ξ1 ≤ z1 , · · · , ξn ≤ zn }).
Defining the events Ai := {ξ | ξi ≤ zi }, i = 1, · · · , n, it follows that
Fξ̃ (z) = P (A1 ∩ · · · ∩ An ).
Denoting the complements of the events Ai by
Bi := Aci = {ξ | ξi > zi },
we know from elementary probability theory that
A1 ∩ · · · ∩ An = (B1 ∪ · · · ∪ Bn )c ,
and consequently
Fξ̃ (z) = P (A1 ∩ · · · ∩ An )
= P ((B1 ∪ · · · ∪ Bn )c )
= 1 − P (B1 ∪ · · · ∪ Bn ).
Therefore asking for the value of Fξ̃ (z) is equivalent to looking for the
probability that at least one of the events B1 , · · · , Bn occurs. Defining the
counter ν̃ : Ξ −→ IN by
ν̃(ξ) := {number of events out of B1 , · · · , Bn that occur at ξ},
ν̃ is clearly a random variable having the range of integers {0, 1, · · · , n}.

Observing that P (B1 ∪ · · · ∪ Bn ) = P (ν̃ ≥ 1), we have
Fξ̃ (z) = 1 − P (ν̃ ≥ 1).
Hence finding a good approximation for P (ν̃ ≥ 1) yields at the same time a
satisfactory approximation of Fξ̃ (z).
With the binomial coefficients for µ, k ∈ IN defined for µ ≥ k as

µ µ!
=
k k!(µ − k)!

µ
(where 0! = 1 and = 0 for µ < k) the binomial moments of ν̃ are
k
introduced as
n
ν̃ i
Sk,n := Eξ̃ = P ({ξ | ν̃(ξ) = i}), k = 0, 1, · · · , n. (3.1)
k k
i=0

i
Since = 1, i = 0, 1, · · · , n, it follows that S0,n = 1. Furthermore,
0
choosing v ∈ IRn+1 according to vi := P ({ξ | ν̃(ξ) = i}), i = 0, 1, · · · , n, it is
obvious from (3.1) that v solves the system of linear equations
⎫
v0 + v1 + v2 + · · · + vn = S0,n , ⎪
⎪
v1 + 2v2 + · · · + nvn = S1,n , ⎪
⎪
⎪
⎪
⎪
⎪
v2 + · · · +
n ⎪
⎪
2
vn = S2,n , ⎬
.. (3.2)
.. ⎪
⎪
. . ⎪
⎪
⎪
⎪
.. .. ⎪
⎪
. . ⎪
⎪
⎭
vn = Sn,n .
The coefficient matrix of (3.2) is upper-triangular, with all main diagonal

elements equal to 1, and hence with a determinant of 1, such that vi = P ({ξ |
ν̃(ξ) = i}), i = 0, 1, · · · , n, is the unique solution of this system of linear
equations.
n However, solving the complete system (3.2) to get P (ν̃ ≥ 1) =
i=1 vi would require the computation of all binomial moments. This would
be a cumbersome task again.
Instead, we could proceed as follows. Observing that our unique solution,
representing probabilities, is nonnegative, it is no restriction to add the
conditions vi ≥ 0, ∀i to (3.2). In turn, we relax the system by dropping
some of the equations (also the first one), in that way getting rid of the
need to determine the corresponding binomial moments. Obviously, the above
(formerly unique) solution is still feasible to the relaxed system, but no
longer unique in general. Hence we get a lower or upper bound n on P (ν̃ ≥ 1)
by minimizing or maximizing, respectively, the objective i=1 vi under the
relaxed constraints.
To be more specific, let us consider the following relaxation as an example.
For the lower bound we choose

⎫
min{v1 + v2 + · · · + vn } ⎪
⎪
⎪
⎪
s.t. v1 + 2v2 + · · · + nvn = S1,n , ⎬
n (3.3)
v2 + · · · + vn = S2,n , ⎪
⎪
2 ⎪
⎪
⎭
vi ≥ 0, i = 1, · · · , n.
and correspondingly for the upper bound we formulate

⎫
max{v1 + v2 + · · · + vn } ⎪
⎪
⎪
⎪
s.t. v1 + 2v2 + · · · + nvn = S1,n , ⎬
n (3.4)
v2 + · · · + vn = S2,n , ⎪
⎪
2 ⎪
⎪
⎭
vi ≥ 0, i = 1, · · · , n.
These linear programs are feasible and bounded, and therefore solvable. So,
there exist optimal feasible 2 × 2 bases B.
Consider an arbitrary 2 × 2 matrix of the form
⎛ ⎞
i i+r
B=⎝ i i+r ⎠,
2 2
where 1 ≤ i < n and 1 ≤ r ≤ n − i. Computing the determinant of B, we get

i+r i
det B = i − (i + r)
2 2
= 12 [i(i + r)(i + r − 1) − (i + r)i(i − 1)]
= 12 i(i + r)r
>0
for all i and r such that 1 ≤ i < n and 1 ≤ r ≤ n − i. Hence any two
columns of the coefficient matrix of (3.3) (or equivalently of (3.4)) form a
basis. The question is which one is feasible and optimal. Let us consider the
second property first. According to Proposition 1.15, Section 1.7 (page 65), a
basis B of (3.3) satisfies the optimality condition if
1 − eT B −1 Nj ≥ 0 ∀j = i, i + r,
where eT = (1, 1) and Nj is the jth column of the coefficient matrix of (3.3).
Obviously, for (3.4) we have the reverse inequality as optimality condition:
1 − eT B −1 Nj ≤ 0 ∀j = i, i + r.
It is straithforward to check1 that

⎛ ⎞
i+r−1 2
⎜ −
ir ir ⎟
B −1 = ⎜
⎝
⎟.
⎠
i−1 2
−
(i + r)r (i + r)r
⎛ ⎞
j
For Nj = ⎝ j ⎠ we get
2
2i + r − j
eT B −1 Nj = j . (3.5)
i(i + r)
Proposition 4.3 The basis
⎛ ⎞
i i+r
B=⎝ i i+r ⎠
2 2
satisfies the optimality condition
(a) for (3.3) if and only if r = 1 (i arbitrary);
(b) for (3.4) if and only if i = 1 and i + r = n.
Proof
(a) If r ≥ 2, we get from (3.5) for j = i + 1
2i + r − j
eT B −1 Ni+1 = j
i(i + r)
i(i + r) + r − 1
=
i(i + r)
> 1,
so that the optimality condition for (3.3) is not satisfied for r > 1, showing
that r = 1 is necessary.
Now let r = 1. Then for j < i we have, according to (3.5),
2i + 1 − j
eT B −1 Nj = j
i(i + 1)
j + i2 − (j − i)2
=
i(i + 1)
i(i + 1) − (j − i)2
<
i(i + 1)
< 1,
1 BB −1 = I, the identity matrix!
whereas for j > i + 1 we get

2i + 1 − j
eT B −1 Nj = j
i(i + 1)
j(i + 1) + j(i − j)
=
i(i + 1)
< 1,
the last inequality resulting from the fact that subtracting the
denominator from the numerator yields
j(i + 1) + j(i − j) − i(i + 1) = (j − i) [(i + 1) − j] < 0.

>1 <0
Hence in both cases the optimality condition for (3.3) is strictly satisfied.
(b) If i + r < n then we get from (3.5) for j = n
n(i + r) + n(i − n)
eT B −1 Nn =
i(i + r)
<1
since
{numerator} − {denominator} = n(i + r) + n(i − n) − i(i + r)
= (n − i)(i + r − n)
< 0.
Finally, if i > 1 then, with (3.5), we have for j = 1

2i + r − 1
eT B −1 Nn =
i(i + r)
(i − 1) + (i + r)
=
i(i + r)
i−1 1
= +
i(i + r) i
1 1
< 3+2
< 1.
Hence the only possible choice for a basis satisfying the optimality
condition for problem (3.4) is i = 1, r = n − 1.
2
As can be seen from the simplex method, a basis that satisfies the optimality
condition strictly does determine a unique optimal solution if it is feasible.
Hence we now have to find from the optimal bases

⎛ ⎞
i i+1
B=⎝ i i+1 ⎠
2 2
the one that is ⎛
feasible for (3.3). ⎞
i i+1
A basis B = ⎝ i i + 1 ⎠ is feasible for (3.3) if and only if
2 2
⎛ ⎞
2
1 −
S1,n ⎜ i ⎟ S1,n
B −1 ⎜
=⎝ ⎟
S2,n i−1 2 ⎠ S2,n
−
i+1 i+1
⎛ ⎞
2
⎜ S1,n − S2,n ⎟
i
=⎜
⎝ i−1
⎟
⎠
2
− S1,n + S2,n
i+1 i+1
≥ 0,
or, equivalently, if
(i − 1)S1,n ≤ 2S2,n ≤ iS1,n .
Hence we have to choose i such that i − 1 = 2S2,n /S1,n , where α is the
integer part of α (i.e. the greatest integer less than or equal to α). With this
particular i the optimal value of (3.3) amounts to
2 i−1 2 2 2
S1,n − S2,n − S1,n + S2,n = S1,n − S2,n .
i i+1 i+1 i+1 i(i + 1)
Thus we have found a lower bound for P (ν̃ ≥ 1) as
4 5
2 2 2S2,n
P (ν̃ ≥ 1) ≥ S1,n − S2,n , with i − 1 = . (3.6)
i+1 i(i + 1) S1,n
For the optimal basis of (3.4)
⎛ ⎞
1 n
B=⎝ n ⎠
0
2
we have ⎛ ⎞
2
1 −
⎜ n−1 ⎟
B −1 = ⎜
⎝
⎟
⎠
2
0
n(n − 1)
and hence ⎛ ⎞
2
S1,n − S2,n
S1,n ⎜ n−1 ⎟
B −1 =⎜
⎝
⎟.
⎠
S2,n 2
S2,n
n(n − 1)
The last vector is nonnegative since the definition of the binomial moments
implies (n − 1)S1,n − 2S2,n ≥ 0 and S2,n ≥ 0. This yields for (3.4) the optimal
value S1,n − (2/n)S2,n . Therefore we finally get an upper bound for P (ν̃ ≥ 1)
as
2
P (ν̃ ≥ 1) ≤ S1,n − S2,n . (3.7)
n
In conclusion, recalling that
Fξ̃ (z) = 1 − P (ν̃ ≥ 1),
we have shown the following.
Proposition 4.4 The distribution function Fξ̃ (z) is bounded according to

2
Fξ̃ (z) ≥ 1 − S1,n − S2,n
n
and
4 5
2 2 2S2,n
Fξ̃ (z) ≤ 1 − S1,n − S2,n , with i − 1 = .
i+1 i(i + 1) S1,n
Example 4.1 We defined in (3.1) the binomial moments of ν̃ as

n
ν̃ i
Sk,n := Eξ̃ = P ({ξ | ν̃(ξ) = i}), k = 0, 1, · · · , n.
k k
i=0
Another way to introduce these moments is the following. With the same
notation as at the beginning of this section, let us define new random variables
χ̃i : Ξ −→ IR, i = 1, · · · , n, as the indicator functions

1 if ξ ∈ Bi ,
χ̃i (ξ) :=
0 otherwise.
n
Then clearly ν̃ = i=1 χ̃i , and

ν̃ χ̃1 + · · · χ̃n
= = χ̃i1 χ̃i2 · · · χ̃ik .
k k
1≤i1 ≤···≤ik ≤n
Taking the expectation on both sides yields for the binomial moments Sk,n

ν̃
Eξ̃ = Eξ̃ (χ̃i1 χ̃i2 · · · χ̃ik )
k
1≤i1 ≤···≤ik ≤n

= P (Bi1 ∩ · · · ∩ Bik ) .
1≤i1 ≤···≤ik ≤n
This formulation indicates the possibility of estimating the binomial moments

from large samples through the relation

Sk,n = Eξ̃ (χ̃i1 χ̃i2 · · · χ̃ik )
1≤i1 ≤···≤ik ≤n
if they are difficult to compute directly.

Consider now the following example. Assume that we have a four-
dimensional random vector ξ˜ with mutually independent components. Let
z ∈ IR4 be chosen such that with pi = P (Ai ), i = 1, 2, 3, 4, we have
pT = (0.9, 0.95, 0.99, 0.92).
Consequently, for qi = P (Bi ) = 1 − pi we get
q T = (0.1, 0.05, 0.01, 0.08).

64
Obviously we get Fξ̃ (z) = i=1 pi = 0.778734. From the above representation
of the binomial moments, we have
4

S1,n = qi = 0.24
i=1
3 4

S2,n = qi qj = 0.0193
i=1 j=i+1
such that we get from (3.7) for P (ν̃ ≥ 1) the upper bound
2
PU = 0.24 − × 0.0193 = 0.23035.
4
7 8
According to (3.6), we find i−1 = 2×0.0193
0.24 = 0 and hence i = 1, so that (3.6)
yields the lower bound
2 2
PL = × 0.24 − × 0.0193 = 0.1757.
2 2
In conclusion, we get for Fξ̃ (z) = 0.778734 the bounds 1 − PU ≤ Fξ̃ (z) ≤
1 − PL , and hence
0.76965 ≤ Fξ̃ (z) ≤ 0.8243.
Observe that these bounds could be derived without any specific informa-
tion about the type of the underlying probability distribution (except the
assumption of independent components made only for the sake of a simple
presentation). 2
Further bounds have been derived for P (ν̃ ≥ 1) using binomial moments up
to the order m, 2 < m < n, as well as for P (ν̃ ≥ r), r > 1. For some of them
explicit formulae could also be derived, while others require the computational
solution of optimization problems with algorithms especially designed for the
particular problem structures.
One of the first attempts to state deterministic equivalent formulations for

chance-constrained programs can be found in Charnes and Cooper [4].
The discussion of convexity of joint chance constraints with stochastic
right-hand sides was initiated by Prékopa [7, 8, 11], investigating log-concave
measures, and could be extended to quasi-concave measures through the
results of Borell [1], Brascamp and Lieb [3] and Rinott [16]. Marti [5] derived
convexity statements in particular for separate chance constraints, for various
distribution functions and probability levels, including the one mentioned first
by van de Panne and Popp [19] for the multivariate normal distribution and
described in Section 4.2.
Prékopa [9] proposed an extension of Zoutendijk’s method of feasible
directions for the solution of (1.1), which was implemented under the name
STABIL by Prékopa et al. [15]. For more general types of chance-contrained
problems solution, approaches have also been considered by Prékopa [10, 12].
After all, the case of joint chance constraints with nondiscrete random matrix
is considered to be a hard problem.
As described in Section 4.1, Mayer developed a special reduced gradient
method for (1.1) and implemented it as PROCON [6]. For the evaluation of
the probability function G(x) = P ({ξ | T x ≥ ξ}) and its gradient ∇G(x), an
efficient Monte-Carlo technique due to Szántai [17] was used.
An alternative method following the lines of Veinott’s supporting
hyperplane algorithm was implemented by Szántai [18].
There has been for some time great interest in getting (sharp) bounds for
distribution functions and, more generally, for probabilities of certain events
in complex systems (e.g. reliabilities of special technical installations). In
Section 4.3 we only sketch the direction of thoughts in this field. Among the
wide range of literature on the subject, we just refer to the more recent papers
of Prékopa [13, 14] and of Boros and Prékopa [2], from which the interested
reader may trace back to earlier original work.
Exercises
1. Given a random vector ξ˜ with support Ξ in IRk , assume that for A ⊂ Ξ

and B ⊂ Ξ we have P (A) = P (B) = 1. Show that then also P (A ∩ B) = 1.
2. Under the assumptions of Proposition 4.2, the support of the distribution
is Ξ = {ξ 1 , · · · , ξ r }, with P (ξ = ξ j ) = pj > 0 ∀j. Show that for
α > 1 − minj∈{1,···,r} pj the only event A ⊂ Ξ satisfying P (A) ≥ α is
A = Ξ.
3. Show that for the random variable ζ̃(x) introduced in Section 4.2 with
σζ̃2 (x), σζ̃ (x) is also a convex function in x.
4. In Section 4.3 we saw that the binomial moments S0,n , S1,n , · · · , Sn,n
determine uniquely the probabilities vi = P (ν̃ = i), i = 0, 1, · · · , n, as
the solution
n of (3.2). From the first equation, it follows, owing to S0,n = 1,
that i=0 vi = 1. To get lower and upper bounds for P (ν̃ ≥ 1), we derived
the linear programs (3.3) and (3.4) by omitting, among others, the first
equation.
(a) Show that in any case (provided that S1,n and S2,n are binomial
n
moments) for the optimal solution v̂ of (3.3), i=1 v̂i ≤ 1.
n
(b) If for theoptimal solution v̂ of (3.3) i=1 v̂i < 1 then we have
n
v0 = 1 − i=1 v̂i > 0. What does this mean with respect to Fξ̃ (z)?

(c) Solving (3.4) can result in ni=1 v̂i > 1. To what extent does this result
improve your knowledge about Fξ̃ (z)?
References
[1] Borell C. (1975) Convex set functions in d-space. Period. Math. Hungar.
6: 111–136.
[2] Boros E. and Prékopa A. (1989) Closed-form two-sided bounds for
probabilities that at least r and exactly r out of n events occur. Math.
Oper. Res. 14: 317–342.
[3] Brascamp H. J. and Lieb E. H. (1976) On extensions of the Brunn–
Minkowski and Prekopa–Leindler theorems, including inequalities for log
concave functions, and with an application to the diffusion euation. J.
Funct. Anal. 22: 366–389.
[4] Charnes A. and Cooper W. W. (1959) Chance-constrained programming.
Management Sci. 5: 73–79.
[5] Marti K. (1971) Konvexitätsaussagen zum linearen stochastischen
Optimierungsproblem. Z. Wahrsch. theorie u. verw. Geb. 18: 159–166.

[6] Mayer J. (1988) Probabilistic constrained programming: A reduced
gradient algorithm implemented on pc. Working Paper WP-88-39, IIASA,
Laxenburg.
[7] Prékopa A. (1970) On probabilistic constrained programming. In Kuhn
H. W. (ed) Proc. of the Princeton Symposioum on Math. Programming,
pages 113–138. Princeton University Press, Princeton, New Jersey.
[8] Prékopa A. (1971) Logarithmic concave measures with applications to
stochastic programming. Acta Sci. Math. (Szeged) 32: 301–316.
[9] Prékopa A. (1974) Eine Erweiterung der sogenannten Methode der
zulässigen Richtungen der nichtlinearen Optimierung auf den Fall
quasikonkaver Restriktionen. Math. Operationsforsch. Statist., Ser. Opt.
5: 281–293.
[10] Prékopa A. (1974) Programming under probabilistic constraints with a
random technology matrix. Math. Operationsforsch. Statist., Ser. Opt. 5:
109–116.
[11] Prékopa A. (1980) Logarithmic concave measures and related topics. In
Dempster M. A. H. (ed) Stochastic Programming, pages 63–82. Academic
Press, London.
[12] Prékopa A. (1988) Numerical solution of probabilistic constrained
programming problems. In Ermoliev Y. and Wets R. J.-B. (eds)
Numerical Techniques for Stochastic Optimization, pages 123–139.
Springer-Verlag, Berlin.
[13] Prékopa A. (1988) Boole-bonferroni inequalities and linear programming.
Oper. Res. 36: 145–162.
[14] Prékopa A. (1990) Sharp bounds on probabilities using linear
programming. Oper. Res. 38: 227–239.
[15] Prékopa A., Ganczer S., Deák I., and Patyi K. (1980) The STABIL
stochastic programming model and its experimental application to the
electricity production in Hungary. In Dempster M. A. H. (ed) Stochastic
Programming, pages 369–385. Academic Press, London.
[16] Rinott Y. (1976) On convexity of measures. Ann. Prob. 4: 1020–1026.
[17] Szántai T. (1987) Calculation of the multivariate probability distribution
function values and their gradient vectors. Working Paper WP-87-82,
IIASA, Laxenburg.
[18] Szántai T. (1988) A computer code for solution of probabilistic-
constrained stochastic programming problems. In Ermoliev Y. M. and
Wets R. J.-B. (eds) Numerical Techniques for Stochastic Optimization,
pages 229–235. Springer-Verlag, Berlin.
[19] van de Panne C. and Popp W. (1963) Minimum cost cattle feed under
probabilistic problem constraint. Management Sci. 9: 405–430.
5
Preprocessing
The purpose of this chapter is to discuss different aspects of preprocessing

the data associated with a stochastic program. The term “preprocessing”
is rather vague, but whatever it could possibly mean, our intention here
is to discuss anything that will enhance the model understanding and/or
simplify the solution procedures. Thus “preprocessing” refers to any analysis
of a problem that takes place before the final solution of the problem. Some
tools will focus on the issue of model understanding, while others will focus
on issues related to choice of solution procedures. For example, if it can be
shown that a problem has (relatively) complete recourse, we can apply solution
procedures where that is required. At the same time, the fact that a problem
has complete recourse is of value to the modeller, since it says something about
the underlying problem (or at least the model of the underlying problem).
5.1 Problem Reduction
Reducing the problem size can be of importance in a setting of stochastic

programming. Of course, it is always useful to remove unnecessary rows and
columns. In the setting of a single deterministic linear programming problem
it may not pay off to remove rows and columns. That is, it may cost more to
figure out which columns and rows are not needed than it costs to solve the
overall problem with the extra data in it. In the stochastic setting, the same
coefficient matrix is used again and again, so it definitely pays to reduce the
problem size. The problem itself becomes smaller, and, even more importantly,
the number of possible bases can be substantially reduced (especially if we
are able to remove rows). This can be particularly important when using
the stochastic decomposition method (where we build up a collection of dual
feasible bases) and in trickling down within the setting of the L-shaped
decomposition method. Let us start by defining a frame and showing how
to compute it.
procedure framebylp(W :(m × n) matrix);

begin
n1 := n;
q := 0;
for i := n1 downto 1 do begin
LP(W \ Wi , Wi , q, y,feasible);
if feasible then begin
Wi := Wn ;
n := n − 1;
end;
end;
end;
Figure 1 Finding a frame.
5.1.1 Finding a Frame

Let us repeat the definition of pos W :
pos W = {t | t = W y, y ≥ 0}.
In words, pos W is the set of all positive (nonnegative) linear combinations

of columns of the matrix W . A subset of the columns, determining a matrix
W , is called a frame if pos W = pos W , and equality is not preserved if
any one column is removed from W . So, by finding the frame of a given
matrix, we remove all columns that are not needed to describe the pointed
cone pos W . As an example, if we use a two-phase simplex method to solve a
linear programming problem, only the columns of W are needed in phase 1.
If W is a matrix, and j is an index, let W \ Wj be the matrix W with
column j removed.
A simple approach for finding a frame is outlined in Figure 1. To do that, we
need a procedure that solves LPs. It can be found in Figure 7 in Chapter 3. The
matrix W in procedure framebylp in Figure 1 is both input and output. On
entry, it contains the matrix for which we seek the frame; on exit, it contains
those columns that were in the frame. The number of columns, n, is changed
accordingly.
To summarize, the effect of the frame algorithm is that as many columns
as possible are removed from a matrix W without changing the pointed cone
spanned by the columns. We have earlier discussed generators of cones. In this
case we may say that the columns in W , after the application of procedure
framebylp, are generators of pos W . Let us now turn to the use of this
algorithm.
PREPROCESSING 263
W1
W2
pos W
W3
W4
Figure 2 Illustration of the frame algorithm.
5.1.2 Removing Unnecessary Columns

This can be useful in a couple of different settings. Let us first see what
happens if we simply apply the frame algorithm to the recourse matrix W .
We shall then remove columns that are not needed to describe feasibility. This
is illustrated in Figure 2. Given the matrix W = (W1 , W2 , W3 , W4 ), we find
that the shaded region represents pos W and the output of a frame algorithm
is either W = (W1 , W2 , W4 ) or W = (W1 , W3 , W4 ). The procedure framebylp
will produce the first of these two cases.
Removing columns not needed for feasibility can be of use when verifying
feasibility in the L-shaped decomposition method (see page 171). We are there
to solve a given LP for all ξ ∈ A. If we apply frame to W before checking
feasibility, we get a simpler problem to look at, without losing information,
since the removed columns add nothing in terms of feasibility. If we are willing
to live with two version of the recourse matrix, we can therefore reduce work
while computing.
From the modelling perspective, note that columns thrown out are only
needed if the cost of the corresponding linear combination is higher than that
of the column itself. The variable represented by the column does not add to
our production possibilities—only, possibly, to lower our costs. In what follows
in this subsection let us assume that we have only right-hand side randomness,
and let us, for simplicity, denote the cost vector by q. To see if a column can
reduce our costs, we define

q 1
W := ,
W 0
that is, a matrix containing the coefficient matrix, the cost vector and an extra
column. To see the importance of the extra column, consider the following
interpretation of pos W (remember that pos W equals the set of all positive
linear combinations of columns from W ):
⎧ ⎫
⎨ ⎬
q1 · · · qn 1 q
pos = W = λ W , q ≥ λ q .
W1 · · · Wn 0 ⎩ W
k k k k
⎭
λk ≥0 λk ≥0
In other words, finding a frame of W means removing all columns

qj
with Wj = λk Wk , and qj ≥ λk qk
Wj
λk ≥0 λk ≥0
in a sequential manner until we are left with a minimal (but not necessarily
unique) set of columns. A column thrown out in this process will never be
part of an optimal solution, and is hence not needed. It can be dropped. From
a modelling point of view, this means that the modeller has added an activity
that is clearly inferior. Knowing that it is inferior should add to the modeller’s
understanding of his model.
A column that is not a part of the frame of pos W , but is a part of the
frame of pos W , is one that does not add to our production possibilities, but
its existence might add to our profit.
5.1.3 Removing Unnecessary Rows

There is a large amount of research on the topic of eliminating redundant
constraints. In this section we shall focus on the use of frames in removing
unnecessary rows. Not very surprisingly, this problem has a dual relationship
to that of removing columns. Let us first look at it from a general point of
view, and then see how we can apply the results in stochastic programming.
Assume we have the system
W y ≤ h, y ≥ 0.
Let W j be the jth row of W , such that the jth inequality is given by
W j y ≤ hj . A row j is not needed if there exists a vector α ≥ 0 such that

αi W i = W j
i=j
and
αi hi ≤ hj .
i=j
PREPROCESSING 265
Finding which rows satisfy this is equivalent to finding the frame of

T
h 1
pos
WT 0
where T indicates the transpose. Of course, if we have ≥ or = in the original
setting, we can easily transform that into a setting with only ≤.
The next question is where we can use this in our setting. The first,
and obvious, answer is to apply it to the first-stage (deterministic) set of
constraints Ax = b. (On the other hand, note that we may not apply frame
to the first-stage coefficient matrix in order to remove unnecessary columns;
these columns may be necessary after feasibility and optimality cuts have been
added.)
It is more difficult to apply these results to the recourse problem. In
principle, we have to check if a given row is unnecessary with all possible
combinations of x and ξ. ˜ This may happen with inequality constraints, but it
is not very likely with equality constraints. With inequalities, we should have
to check if an inequality W j y ≤ hj was implied by the others, even when the
jth inequality was at its tightest and the others as loose as possible. This is
possible, but not within the frame setting.
We have now discussed how the problem can be reduced in size. Let us
now assume that all possible reductions have been performed, and let us start
discussing feasibility. This will clearly be related to topics we have seen in
earlier chapters, but our focus will now be more specifically directed towards
preprocessing.
5.2 Feasibility in Linear Programs
The tool for understanding feasibility in linear programs is the cone

pol pos W . We have discussed it before, and it is illustrated in Figure 3. The
important aspect of Figure 3 is that a right-hand side h represents a feasible
recourse problem if and only if h ∈ pos W . But this is equivalent to requiring
that hT y ≤ 0 for all y ∈ pol pos W . In particular, it is equivalent to requiring
that hT y ≤ 0 for all y that are generators of pol pos W . In the figure there
are two generators. You should convince yourself that a vector is in pos W
if and only if it has a nonpositive inner product with the two generators of
pol pos W .
Therefore what we shall need to find is a matrix W ∗ , to be referred to as
the polar matrix of W , whose columns are the generators of pol pos W , so
that we get
pos W ∗ = pol pos W.
Assume that we know a column w∗ from W ∗ . For h to represent a feasible
recourse problem, it must satisfy hT w∗ ≤ 0.
pos W
pol pos W
Figure 3 Finding the generators of pol pos W .
There is another important aspect of the polar cone pos W ∗ that we have
not yet discussed. It is indicated in Figure 3 by showing that the generators
are pairwise normals. However, that is slightly misleading, so we have to turn
to a three-dimensional figure to understand it better. We shall also need the
term facet. Let a cone pos W have dimension k. Then every cone K positively
spanned by k−1 generators from pos W , such that K belongs to the boundary
of pos W , is called a facet. Consider Figure 4.
What we note in Figure 4 is that the generators are not pairwise normals,
but that the facets of one cone have generators of the other as normals. This
goes in both directions. Therefore, when we state that h ∈ pos W if and only
if hT y ≤ 0 for all generators of pol pos W , we are in fact saying that either h
represents a feasible problem because it is a linear combination of columns in
W or because it satisfies the inequality implied by the facets of pos W . In still
other words, the point of finding W ∗ is not so much to describe a new cone,
but to replace the description of pos W in terms of generators with another
in terms of inequalities.
This is useful if the number of facets is not too large. Generally speaking,
performing an inner product of the form bT y is very cheap. In parallel
processing, an inner product can be pipelined on a vector processor and the
different inner products can be done in parallel. And, of course, as soon as we
find one positive inner product, we can stop—the given recourse problem is
infeasible.
Readers familiar with extreme point enumeration will see that going from
PREPROCESSING 267
pos W
pol pos W
Figure 4 Three-dimensional picture of pos W and pol pos W = pos W ∗ .
a generator to a facet representation of pos W is indeed extreme point

enumeration. As such, it is a problem with exponential complexity. Therefore
we cannot in general expect to find W ∗ in reasonable time. However, taking
a practical view of the matter, it is our suggestion that an attempt is made.
The results are generally only interesting if there are relatively few facets,
and those cases are the easiest. Figure 5 presents a procedure for finding the
facets. It is called procedure support because it finds a minimal selection of
supporting hyperplanes (not necessarily unique) of pos W , such that pos W is
fully described. In practice, it has been shown to possess the desired property
that it solves quickly if there are few facets. An example is presented shortly
to help in understanding this procedure support.
The procedure support finds the polar matrix W ∗ , and thereby the
support of pos W . The matrix W is reduced by the application of procedure
framebylp, but is otherwise unchanged on exit. The process is initialized with
a matrix W ∗ that spans the entire column (range) space. We typically do this
by letting
procedure support(W, W ∗ :matrices);

begin
framebylp(W );
done := f alse;
for i := 1 to n do if not done then begin
α := WiT W ∗ ;
I+ := {k|α[k] > 0};
I− := {k|α[k] < 0};
I0 := {k|α[k] = 0};
done := (I− ∪ I0 = ∅);
if done then W ∗ := 0;
if I+ = ∅ and not done then begin
if I− = ∅ then W ∗ := WI∗0 ;
else begin
for all k ∈ I+ do
for all j ∈ I− do
Ckj := Wk∗ − (α[k]/α[j])Wj∗ ;
W := WI∗0 ∪ WI∗− ∪kj Ckj ;
∗
framebylp(W ∗ );
end; (* else *)
end; (* if *)
end; (* for *)
end;
Figure 5 Finding the support.
⎛ ⎞
1 0 · · · 0 −1
⎜ 0 1 · · · 0 −1 ⎟
W ∗ := ⎜
⎝ .. ...
. . .. ⎟
. ⎠
..
. ..
0 0 · · · 1 −1
or ⎛ ⎞
1 0 · · · 0 −1 0 ··· 0
⎜ 0 1 · · · 0 0 −1 ··· 0 ⎟
W ∗ := ⎜
⎝ ... ... .. . .. .. .. .. ⎟
. .. . . . . ⎠
0 0 ··· 1 0 0 · · · −1
∗
On exit W is the polar matrix. We initiate support by a call to framebylp in
order to remove all columns from W that are not needed to describe pos W .
Example 5.1 Let us turn to a small example to see how procedure support
PREPROCESSING 269
Figure 6 The cones pos W and pol pos W before any column has been added
to W .
progresses. Since pos W and pol pos W live in the same dimension, we can
draw them side by side.
Let us initially assume that

3 1 −1 −2
W = .
1 1 2 1
The first thing to do, according to procedure support, is to subject W to a
frame finding algorithm, to see if some columns are not needed. If we do that
(check it to see that you understand frames) we end up with

3 −2
W = .
1 1
Having reduced W , we then initialize W ∗ to span the whole space. Consult
Figure 6 for details. We see there that

1 0 −1 0
W∗ = .
0 1 0 −1
Consult procedure support. From there, it can be seen that the approach
is to take one column from W at a time, and with it perform some calculations.
Figure 6 shows the situation before we consider the first column of W . Calling
it pos W is therefore not quite correct. The main point, however, is that the
left and right parts correspond. If W has no columns then pol pos W spans
the whole space.
Now, let us take the first column from W . It is given by W1 = (3, 1)T . We
next find the inner products between W1 and all four columns of W ∗ . We get
α = (3, 1, −3, −1)T.
In other words, the sets I+ = {1, 2} and I− = {3, 4} have two members
each, while I0 = ∅. What this means is that two of the columns must be
Figure 7 The cones pos W and pol pos W after one column has been added
to W .
removed, namely those in I + , and two kept, namely those in I − . But to avoid
losing parts of the space, we now calculate four columns Ckj . First, we get
C13 = C24 = 0. They are not interesting. But the other two are useful:
1
1 0 1 0 1 −1 −3
C14 = +3 = , C23 = +3 = .
0 −1 −3 1 0 1
Since our only interests are directions, we scale the latter to (−1, 3)T . This
brings us into Figure 7. Note that one of the columns in pos W ∗ is drawn
with dots. This is done to indicate that if procedure framebylp is applied to
W ∗ , that column will disappear. (However, that is not a unique choice.)
Note that if W had had only this one column then W ∗ , as it appears in
Figure 7, is the polar matrix of that one-column W . This is a general property
of procedure support. At any iteration, the present W ∗ is the polar matrix
of the matrix containing those columns we have so far looked at.
Now let us turn to the second column of W . We find

T ∗ −1 1 −1
α = (−2, 1)W = (−2, 1) = (5, −5, 2)
3 −3 0
We must now calculate two extra columns, namely C12 and C32 . The first
gives 0, so it is not of interest. For the latter we get
) 3*
−1 2 1 −5
C32 = +5 = ,
0 −3 − 65
which we scale to (−1, −2)T . This gives us Figure 8. To the left we have pos W ,
with W being the matrix we started out with, and to the right its polar cone.
A column represents a feasible problem if it is inside pos W , or equivalently, if
it has a nonpositive inner product with all generators of pos W ∗ = pol pos W .
2
PREPROCESSING 271
Figure 8 The cones pos W and pol pos W after two columns have been added
to W .
Assume we could indeed find W ∗ using procedure support. Let w∗ be

some column in W ∗ . For feasibility, we must have
(w∗ )T [h0 + Hξ − T (ξ)x] ≤ 0 for all ξ.
Hence
(w∗ )T T (ξ)x ≥ (w∗ )T (h0 + Hξ) for all ξ.
If randomness affects both h and T , as indicated above, we must, at least in
principle, create one inequality per ξ for each column from W ∗ . However, if
T (ξ) ≡ T0 , we get a much easier set-up by calculating
$ %
(w∗ )T T0 x ≥ (w∗ )T h0 + max (w∗ )T H t,
t∈Ξ
where Ξ is the support of ξ. ˜ If we do this for all columns of W ∗ and add the
resulting inequalities in terms of x to Ax = b, we achieve relatively complete
recourse. Hence we see that relatively complete recourse can be generated.
This is why the term is useful. It is very hard to test for relatively complete
recourse. With relatively complete recourse we should never have to worry
about feasibility.
Since the inequalities resulting from the columns of W ∗ can be dominated
by others (in particular, if T (ξ) is truly random), the new rows, together with
those in Ax = b, should be subjected to row removal, as outlined earlier in
this chapter.
5.2.1 A Small Example

Let us return to the example we discussed in Section 1.3. We have now named
the right-hand side elements b1 , b2 and b3 , since they are the focus of the
discussion here (in the numerical example they had the values 100, 180 and
162):
s. t. xraw1 + xraw2 ≤ b1 ,
2xraw1 + 6xraw2 ≥ b2 ,
3xraw1 + 3xraw2 ≥ b3 ,
xraw1 ≥ 0,
xraw2 ≥ 0.
The interpretation is that b1 is the production limit of a refinery, which refines

crude oil from two countries. The variable xraw1 represents the amount of
crude oil from Country 1 and xraw2 the amount from Country 2. The quality
of the crudes is different, so one unit of crudes from Country 1 gives two units
of Product 1 and three units of Product 2, whereas the crudes from the second
country gives 6 and 3 units of the same products. Company 1 wants at least
b2 units of Product 1 and Company 2 at least b3 units of Product 2.
If we now calculate the inequalities describing pos W , or alternatively the
generators of pol pos W , we find that there are three of them:
b1 ≥0
6b1 − b2 ≥0
3b1 − b3 ≥ 0.
The first should be easy to interpret, and it says something that is not very
surprising: the production capacity must not be negative. That we already
knew. The second one is more informative. Given appropriate units on crudes
and products, it says that the demand of Company 1 must not exceed six times
the production capacity of the refinery. Similarly, the third inequality says
that the demand of Company 2 must not exceed three times the production
capacity of the refinery. (The inequalities are not as meaningless as they
might appear at first sight: remember that the units for refinery capacity and
finished products are not the same.) These three inequalities, one of which
was obvious, are examples of constraints that are not explicitly written down
by the modeller, but still are implied by him or her. And they should give the
modeller extra information about the problem.
In case you wonder where the feasibility constraints are, what we have just
discussed was a one-stage deterministic model, and what we obtained was
three inequalities that can be used to check feasibility of certain instances of
that model. For example, the numbers used in Section 1.3 satisfy all three
constraints, and hence that problem was feasible. (In the example b1 = 100,
b2 = 180 and b3 = 162.)
PREPROCESSING 273
h3
H3
pos W
h1H
h2 H
2
Figure 9 Illustration of feasibility.
5.3 Reducing the Complexity of Feasibility Tests
In Chapter 3, (page 162), we discussed the set A that is a set of ξ values such
that if h0 + Hξ − T (ξ)x produces a feasible second-stage problem for all ξ ∈ A
˜ We pointed out
then the problem will be feasible for all possible values of ξ.
that in the worst case A had to contain all extreme points in the support of
˜
ξ.
Assume that the second stage is given by
Q(x, ξ) = min{q(ξ)T y | W y = h0 + Hξ − T0 x, y ≥ 0},
where W is fixed and T (ξ) ≡ T0 . This covers many situations. In R2 consider
the example in Figure 9, where ξ˜ = (ξ˜1 , ξ̃2 , ξ˜3 ).
Since h1 ∈ pos W , we can safely fix ξ̃1 at its lowest possible value, since if
things are going to go wrong, then they must go wrong for ξ1min . Or, in other
words, if h0 + H ξˆ − T0 x ∈ pos W for ξˆ = (ξ1min , ξˆ2 , ξ̂3 ) then so is any other
vector with ξ̃2 = ξ̂2 and ξ̃3 = ξˆ3 , regardless of the value of ξ˜1 . Similarly, since
−h2 ∈ pos W , we can fix ξ˜2 at its largest possible value. Neither h3 nor −h3
are in pos W , so there is nothing to do with ξ̃3 .
Hence to check if x yields a feasible solution, we must check if
h0 +Hξ−T0 x ∈ pos W for ξ = (ξ1min , ξ2max , ξ3min )T and ξ = (ξ1min , ξ2max , ξ3max )T
Hence in this case A will contain only two points instead of 23 = 8. In general,
we see that whenever a column from H, in either its positive or negative
direction, is found to be in pos W , we can halve the number of points in A.

In some cases we may therefore reduce the testing to one single problem.
It is of importance to understand that the reduction in the size of A has
two positive aspects. First, if we do not have (or do not know that we have)
relatively complete recourse, the test for feasibility, and therefore generation of
feasibility cuts, becomes much easier. But equally important is the fact that
it tells us something about our problem. If a column from H is in pos W ,
we have found a direction in which we can move as far as we want without
running into feasibility problems. This will, in a real setting, say something
about the random effect we have modelled using that column.
Preprocessing and similar procedures have been used in contexts totally

different from ours. This is natural, since questions of model formulations and
infeasibilities are equally important in all areas of mathematical programming.
For further reading, consult e.g. Roodman [7], Greenberg [3, 4, 5] or Chinneck
and Dravnieks [1].
An advanced algorithm for finding frames can be found in Wets and
Witzgall [12]. Later developments include the work of Rosen et al. [8] and Dulá
et al. [2]. The algorithm for finding a support was described by Tschernikov[9],
and later also by Wets [11]. For computational tests using the procedure see
Wallace and Wets [10]. Similar procedures for networks will be discussed in
Chapter 6.
For an overview of methods for extreme point enumeration see e.g.
Mattheiss and Rubin [6].
Exercises
1. Let W be the coefficient matrix for the following set of linear equations:
x + 12 y − z + s1 = 0,
2x + z + s2 = 0,
x, y, z, s1 , s2 ≥ 0.
(a) Find a frame of pos W .

(b) Draw a picture of pos W , and find the generators of pol pos W by
simple geometric arguments.
(c) Find the generators of pol pos W by using procedure support in
Figure 5. Make sure you draw the cones pos W and pol pos W after
each iteration of the algorithm, so that you see how it proceeds.
PREPROCESSING 275
2. Let the following set of equations be given:
x+y +z ≤ 4,
2x +z ≤ 5,
y +z ≤ 8,
x, y, z ≥ 0.
(a) Are there any columns that are not needed for feasibility? (Remember
the slack variables!)
(b) Let W contain the columns that were needed from question (a),
including the slacks. Try to find the generators of pol pos W by
geometric arguments, i.e. draw a picture.
3. Consider the following recourse problem constraints:

1 3 2 2 −1 0 −1 −4 5 1
y= + ξ+ x
3 1 7 2 −2 −1 1 −1 3 2
with y ≥ 0. Assume that all random variables are independent, with

support [0, 1]. Look back at Section 5.3, where we discussed how we could
simplify the feasibility test if we were not aware of relatively complete
recourse. We there defined a set A that was such that if the recourse
problem was feasible for all ξ ∈ A then it was feasible for all ξ. In the
worst case A has, in our case, 25 = 32 elements. By whatever method you
find useful (what about a picture?), reduce this number to six, and list the
six elements.
References
[1] Chinneck J. W. and Dravnieks E. W. (1991) Locating minimal infeasible

constraint sets in linear programs. ORSA J.Comp. 3: 157–168.
[2] Dulá J. H., Helgason R. V., and Hickman B. L. (1992) Preprocessing
schemes and a solution method for the convex hull problem in
multidimensional space. In Balci O. (ed) Computer Science and
Operations Research: New Developments in their Interfaces, pages 59–
70. Pergamon Press, Oxford.
[3] Greenberg H. J. (1982) A tutorial on computer-assisted analysis. In
Greenberg H. J., Murphy F. H., and Shaw S. H. (eds) Advanced
Techniques in the Practice of Operations Research. Elsevier, New York.
[4] Greenberg H. J. (1983) A functional description of ANALYZE: A
computer-assisted analysis. ACM Trans. Math. Software 9: 18–56.
[5] Greenberg H. J. (1987) Computer-assisted analysis for diagnosing

infeasible or unbounded linear programs. Math. Prog. Study 31: 79–97.
[6] Mattheiss T. H. and Rubin D. S. (1980) A survey and comparison of
methods for finding all vertices of convex polyhedral sets. Math. Oper.
Res. 5: 167–185.
[7] Roodman G. M. (1979) Post-infeasibility analysis in linear programming.
Management Sci. 9: 916–922.
[8] Rosen J. B., Xue G. L., and Phillips A. T. (1992) Efficient computation
of extreme points of convex hulls in IRd . In Pardalos P. M. (ed) Advances
in Optimization and Parallel Computing, pages 267–292. North-Holland,
Amsterdam.
[9] Tschernikow S. N. (1971) Lineare Ungleichungen. VEB Deutscher Verlag
der Wissenschaften, Berlin. (Translated from Russian).
[10] Wallace S. W. and Wets R. J.-B. (1992) Preprocessing in stochastic
programming: The case of linear programs. ORSA Journal on Computing
4: 45–59.
[11] Wets R. J.-B. (1990) Elementary, constructive proofs of the theorems
of Farkas, Minkowski and Weyl. In Gabszewicz J., Richard J.-F., and
Wolsey L. (eds) Economic Decision Making: Games, Econometrics and
Optimization: Contributions in Honour of Jacques Dreze, pages 427–432.
North-Holland, Amsterdam.
[12] Wets R. J.-B. and Witzgall C. (1967) Algorithms for frames and lineality
spaces of cones. J. Res. Nat. Bur. Stand. 71B: 1–7.
6
Network Problems
The purpose of this chapter is to look more specifically at networks. There are
several reasons for doing this. First, networks are often easier to understand.
Some of the results we have outlined earlier will be repeated here in a network
setting, and that might add to understanding of the results. Secondly, some
results that are stronger than the corresponding LP results can be obtained
by utilizing the network structure. Finally, some results can be obtained that
do not have corresponding LP results to go with them. For example, we shall
spend a section on PERT problems, since they provide us with the possibility
of discussing many important issues.
The overall setting will be as before. We shall be interested in two-
or multistage problems, and the overall solution procedures will be the
same. Since network flow problems are nothing but specially structured LPs,
everything we have said before about LPs still hold. The bounds we have
outlined can be used, and the L-shaped decomposition method, with and
without bounds, can be applied as before. We should like to point out,
though, that there exists one special case where scenario aggregation looks
more promising for networks than for general LPs: that is the situation where
the overall problem is a network. This may require some more explanation.
When we discuss networks in this chapter, we refer to a situation in which
the second stage (or the last stage in a multistage setting) is a network.
We shall mostly allow the first stage to be a general linear program. This
rather limited view of a network problem is caused by properties of the L-
shaped decomposition method (see page 171). The computational burden in
that algorithm is the calculation of Q(x̂), the expected recourse cost, and to
some extent the check of feasibility. Both those calculations concern only the
recourse problem. Therefore, if that problem is a network, network algorithms
can be used to speed up the L-shaped algorithm.
What if the first-stage problem is also a network? Example 2.2 (page 117)
was such an example. If we apply the L-shaped decomposition method to
that problem, the network structure of the master problem is lost as soon as
feasibility and optimality cuts are added. This is where scenario aggregation,
outlined in Section 2.6, can be of some use. The reason is that, throughout the
calculations, individual scenarios remain unchanged in terms of constraints, so
that structure is not lost. A nonlinear term is added to the objective function,
however, so if the original problem was linear, we are now in a setting of
quadratic objectives and linear (network) constraints. If the original problem
was a nonlinear network, the added terms will not increase complexity at all.
6.1 Terminology
Consider a network with arcs E = {1, . . . , m} and nodes N = {1, . . . , n}. An

arc k ∈ E will be denoted by k ∼ (i, j), indicating that it starts at i and ends
at j. The capacity of k will be denoted by γ(k) and the cost by q(k). For each
node i ∈ N , let β(i) be the external flow. We let β(i) > 0 denote supply and
β(i) < 0 demand.
We say that a network flow problem is capacitated if all arcs k have
γ(k) < ∞. If all arcs are uncapacitated (logically that γ(k) = ∞), we say
that the network is uncapacitated. Most networks have arcs of both types,
and their properties will then be mixtures of what we discuss for the two
cases in this chapter.
By G(Y ),we understand a network consisting of the nodes in Y ⊆ N and
all arcs in E connecting nodes in Y . Of course, G(N ) is the original network.
For two arbitrary sets Y, Y ⊂ N , let {k ∼ (i, j) | i ∈ Y, j ∈ Y } ⊆ E be
denoted by [Y, Y ]+ and let {k ∼ (i, j) | j ∈ Y, i ∈ Y } ⊆ E be denoted by
[Y, Y ]− .
For Y ⊂ N define Q+ = [Y, N \ Y ]+ and Q− = [Y, N \ Y ]− . We call
Q = Q+ ∪ Q− = [Y, N \ Y ] a cut. Whenever we refer to Y and Q, without
stating their relationship, we are assuming that Q = [Y, N \ Y ]. For each
Y ⊆ N , let b(Y ) ∈ {0, 1}n be an index vector for the set Y , i.e. b(Y, i) = 1 if
i ∈ Y , and 0 otherwise. Similarly, for each Q ⊆ E, let a(Q+ ) ∈ {0, 1}m be an
index vector for the set Q+ , i.e. a(Q+ , k) = 1 if k ∈ Q+ and 0 otherwise.
The node–arc incidence matrix for a network will be denoted by W , and
is defined by
1 if k ∼ (i, j) for some j,
W (i, k) = −1 if k ∼ (j, i) for some j,
0 otherwise.
The rows in the node–arc incidence matrix are linearly dependent. For
the system W y = b to have a solution, we know from Chapter 1 that
rk W = rk (W | b). In a network this requirement means that there must
be one node where the external flow equals exactly the negative sum of the
external flows in the other nodes. This node is called the slack node. It is
customary not to include a row for that node in W . Hence W has only n − 1
rows, and it has full rank provided the network is connected. A network is
NETWORK PROBLEMS 279
Figure 1 Network used to demonstrate definitions.
connected if for all Y ⊂ N we have Q = [Y, N \ Y ] = ∅.

We shall also need the following sets:
F + (Y ) = {nodes j | k ∼ (i, j) for i ∈ Y } ∪ Y,
B + (Y ) = {nodes j | k ∼ (j, i) for i ∈ Y } ∪ Y.
The set F + (Y ) contains Y itself plus all nodes that can be reached directly
(i.e. in one step) from a node in Y . Similarly B + (Y ) contains Y and all nodes
from which Y can be reached directly.
Two other sets that are very similar to F + (Y ) and B + (Y ) are
F ∗ (Y ) = {nodes j | ∃ a directed path from some node i ∈ Y to node j} ∪ Y,
B ∗ (Y ) = {nodes j | ∃ a directed path from node j to some node i ∈ Y } ∪ Y.
Thus the sets F + and B + pick up immediate successors and predecessors,
whereas F ∗ and B ∗ pick up all successors and predecessors.
Example 6.1 Let us consider Figure 1 to briefly illustrate most of the

concepts we have introduced.
The node set N = {1, 2, 3, 4}, and the arc set E = {1, 2, 3, 4, 5}. An example
of an arc is 5 ∼ (2, 3), since arc 5 starts at node 2 and ends at node 3. Let
Y = {1, 3} and Y = {2}. The network G(Y ) consists of nodes 1 and 3, and
arc 2, since that is the only arc connecting nodes in Y . Furthermore, for the
same Y and Y , we have [Y, Y ]+ = {1}, since arc 1 is the only arc going from
nodes 1 or 3 to node 2. Similarly [Y, Y ]− = {5}. If we define Q = [Y, N \ Y ]
then Q+ = {1, 4} and Q− = {5}. Therefore Q = {1, 4, 5} is a cut.
Again, with the same definition of Y , we have
b(Y ) = (1, 0, 1, 0)T , a(Q+ ) = (1, 0, 0, 1, 0)T.
Furthermore, we have
F + ({1}) = {1, 2, 3}, F ∗ ({1}) = {1, 2, 3, 4},
since we can reach nodes 2 and 3 in one step, but we need two steps to reach
node 4. Node 1 itself is in both sets by definition.
Two examples of predecessors of a node are
B + ({1}) = {1}, B ∗ ({2, 3}) = {1, 2, 3},
since node 1 has no predecessors, and nodes 2 and 3 can be reached from node
1.
A common problem in network flows is the min cost network flow problem.
It is given as follows.
min q(1)y(1) + q(2)y(2) + q(3)y(3) + q(4)y(4) + q(5)y(5)

s.t. y(1) + y(2) = β(1),
−y(1) + y(3) + y(5) = β(2),
− y(2) + y(4) − y(5) = β(3),
− y(3) − y(4) = β(4),
y(k) ≤ γ(k), k = 1, . . . , 5,
y(k) ≥ 0, k = 1, . . . , 5.
The coefficient matrix for this problem has rank 3. Therefore the node–arc
incidence matrix has three rows, and is given by
⎛ ⎞
1 1 0 0 0
W = ⎝ −1 0 1 0 1 ⎠ .
0 −1 0 1 −1
6.2 Feasibility in Networks
In Section 3.2 and Chapter 5 we discussed feasibility in linear programs.

As will become apparent shortly, it is easier to obtain feasibility results for
networks than for LPs. Let us first run through the development, and then
later see how this fits in with the LP results.
A well-known result concerning feasibility in networks states that if the net
flow across every cut in a network is less than or equal to the capacity of that
cut, then the problem is feasible. More formally, this can be stated as follows,
using β T = (β(1), . . . , β(n)) and γ T = (γ(1), . . . , γ(m)).
Proposition 6.1 A capacitated network flow problem with total supply equal
to total demand is feasible iff for every cut Q = [Y, N \Y ], b(Y )T β ≤ a(Q+ )T γ.
function Connected(W : set of nodes) : boolean;

begin
PickNode(i, W );
Qlist := {i};
V isited := {i};
while Qlist = ∅ do begin
PickNode(i, Qlist);
Qlist := Qlist \ {i};
s := (B ∗ (i) ∪ F ∗ (i)) ∩ (W \ V isited);
Qlist := Qlist ∪ s;
V isited := V isited ∪ s;
end;
Connected := (V isited = W );
end;
Figure 2 Function checking network connectedness.
The above proposition is very simple in nature. However, from a

computational point of view, it is not very useful. It requires that we look
at all subsets Y of N , in other words 2n subsets. For reasonably large n it is
not computationally feasible to try to enumerate subsets this way. Another
problem that might not be that obvious when reading the proposition is that it
is not an “if and only if” statement in a very useful sense. There is no guarantee
that inequalities arising from the proposition are indeed needed. We might—
and most probably will—end up with inequalities that are implied by other
inequalities. A key issue in this respect is the connectedness of a network. We
defined earlier that a network was connected if for all Y ⊂ N we have that
Q = [Y, N \ Y ] = ∅. It is reasonably easy to check connectedness of a network.
Details are given in function Connected in Figure 2. Note that we use F ∗
and B ∗ . If they are not available, we can also use F + and B + , or calculate
F ∗ and B ∗ , which is quite simple.
Using the property of connectedness, it is possible to prove the following
stronger result.
Proposition 6.2 Let Q = [Y, N \ Y ]. For capacitated networks the
inequalities
b(Y )T β ≤ a(Q+ )T γ, b(N \ Y )T β ≤ a(Q− )T γ
are both needed if and only if G(Y ) and G(N \ Y ) are both connected.
Otherwise, none of the inequalities are needed.
Example 6.2 Let us look at the small example network in Figure 3 to at

2
a f
d
b
1 4 5
g
c e
3
Figure 3 Example network 1.
least partially see the relevance of the last proposition. The following three
inequalities are examples of inequalities describing feasibility for the example
network:
β(2) ≤ γ(d) + γ(f ),
β(3) ≤ γ(e),
β(2) + β(3) ≤ γ(d) + γ(e) + γ(f ).
Proposition 6.2 states that the latter inequality is not needed, because
G({2, 3}) is not connected. From the inequalities themselves, we easily see
that if the first two are satisfied, then the third is automatically true. It is
perhaps slightly less obvious that, for the very same reason, the inequality
β(1) + β(4) + β(5) ≤ γ(a) + γ(c)
is also not needed. It is implied by the requirement that total supply must
equal total demand plus the companions of the first two inequalities above.
(Remember that each node set gives rise to two inequalities). More specifically,
the inequality can be obtained by adding the following two inequalities and
one equality (representing supply equals demand):
β(1) + β(2) + β(4) + β(5) ≤ γ(c),

β(1) + β(3) + β(4) + β(5) ≤ γ(a),
− β(1) − β(2) − β(3) − β(4) − β(5) = 0.
2
Once you have looked at this for a while, you will probably realize that the
part of Proposition 6.2 that says that if G(Y ) or G(N \ Y ) is disconnected
then we do not need any of the inequalities is fairly obvious. The other part of
the proposition is much harder to prove, namely that if G(Y ) and G(N \ Y )
are both connected then the inequalities corresponding to Y and N \ Y are
both needed. We shall not try to outline the proof here.
Proposition 6.2 might not seem very useful. A straightforward use could
still require the enumeration of all subsets of N , and for each such subset
a check to see if G(Y ) and G(N \ Y ) are both connected. However, we can
obtain more than that.
The first important observation is that the result refers to the connectedness
of two networks—both the one generated by Y and the one generated by N \Y .
Let Y1 = N \ Y . If both networks are connected, we have two inequalities that
we need, namely
b(Y )T β ≤ a(Q+ )T γ
and
b(Y1 )T β = b(N \ Y )T β ≤ a(Q− )T γ.
On the other hand, if at least one of the networks is disconnected, neither
inequality will be needed. Therefore, checking each subset of N means doing
twice as much work as needed. If we are considering Y and discover that both
G(Y ) and G(Y1 = N \ Y ) are connected, we write down both inequalities at
the same time. An easy way to achieve this is to disregard some node (say
node n) from consideration in a full enumeration. This way, we will achieve
n ∈ N \ Y for all Y we investigate. Then for each cut where the connectedness
requirement is satisfied we write down two inequalities. This will halve the
number of subsets to be checked.
In some cases it is possible to reduce the complexity of a calculation by
collapsing nodes. By this, we understand the process of replacing a set of
nodes by one new node. Any other node that had an arc to or from one of
the collapsed nodes will afterwards have an arc to or from the new node: one
for each original arc. If Y is a set of nodes, we let A(Y ) be the set of original
nodes represented by the present node set Y as a result of collapsing.
To simplify statements later on we shall also need a way to simply state
which inequalities we want to write down. An algorithm for the capacitated
case is given in Figure 4.
Note that we allow the procedure to be called with Y = ∅. This is a
technical devise to ensure consistent results, but you should not let that
confuse you at the present time. Based on Proposition 6.2, it is possible to
develop procedures that in some cases circumvent the exponential complexity
arising from checking all subsets of N . We shall use Figure 5 to illustrate some
of our points.
Proposition 6.3 If B + (i)∪F + (i) = {i, j} then nodes i and j can be collapsed
after the inequalities generated by CreateIneq({i}) have been created.
procedure CreateIneq(Y : set of nodes);

begin
if A(Y ) = ∅ then begin
create the inequality b(A(Y ))T β ≤ a(Q+ )T γ;
create the inequality b(A(N \ Y ))T β ≤ a(Q− )T γ;
end
else begin
create the inequality b(A(N ))T β ≤ 0;
create the inequality −b(A(N ))T β ≤ 0;
end;
end;
Figure 4 Algorithm for generating inequalities—capacitated case.
Figure 5 Example network used to illustrate Proposition 6.3.
The only set Y where i ∈ Y but j ∈ Y , at the same time as both G(Y )
and G(N \ Y ) are connected, is the set where Y = {i}. The reason is that
node j blocks node i’s connections to all other nodes. Therefore, after calling
CreateIneq({i}), we can safely collapse node i into node j. Examples of this
can be found in Figure 5, (see e.g. nodes 4 and 5). This result is easy to
implement, since all we have to do is run through all nodes, one at a time,
and look for nodes satisfying B + (i) ∪ F + (i) = {i, j}. Whenever collapses take
place, F + and B + (or, alternatively, F ∗ and B ∗ ) must be updated for the
remaining nodes.
By repeatedly using this proposition, we can remove from the network all
trees (and trees include “double arcs” like those between nodes 2 and 5). We
are then left with circuits and paths connecting circuits. The circuits can be
both directed and undirected. In the example in Figure 5 we are left with
procedure AllFacets;
begin
TreeRemoval;
CreateIneq(∅);
Y := ∅;
W := N \ {n};
Facets(Y, W );
end;
Figure 6 Main program for full enumeration of inequalities satisfying

Proposition 6.2.
procedure Facets(Y, W : set of nodes);

begin
PickNode(Y, W, i);
if i = 0 then begin
W := W \ {i};
Facets(Y, W );
Y := Y ∪ {i};
Facets(Y, W );
if Connected(N \ Y ) then CreateIneq(Y );
end;
end;
Figure 7 Recursive algorithm for generating facets.
nodes 1, 2 and 3. We shall assume that there is a procedure TreeRemoval that

takes care of this reduction.
There is one final remark to be made based on Proposition 6.2. For each
set Y we must check the connectedness of both G(Y ) and G(N \ Y ). We
can skip the first if we simply make sure that G(Y ) is always connected.
This can easily be achieved by building up Y (in the enumeration) such that
it is always connected. We shall assume that we have available a procedure
PickNode(Y, W, i) that picks a node i from W provided that node is reachable
from Y in one step. Otherwise, it returns i := 0.
We now present a main program and a main procedure for the full
enumeration. They are listed in Figures 6 and 7.
6.2.1 The uncapacitated case

The corresponding results for the uncapacitated networks can be found by
checking what happens when we put γ(k) = ∞ in all previous results. The
result corresponding to Proposition 6.1 is as follows.
Proposition 6.4 An uncapacitated network flow problem with total supply

equal to total demand is feasible iff for every cut Q = [Y, N \ Y ], with Q+ = ∅,
b(Y )T β ≤ 0.
This result is developed by observing that the inequality in Proposition 6.1

becomes
b(Y )T β ≤ ∞
for all cuts but those with Q+ = ∅. And this is, of course, always
true. Similarly, a connectedness result can be obtained that corresponds to
Proposition 6.2.
Proposition 6.5 For an uncapacitated network a cut Q = [Y, N \ Y ] with

Q+ = ∅ is needed if and only if G(Y ) and G(N \ Y ) are both connected.
Collapsing nodes was discussed for the capacitated case. Those results apply
here as well, in particular Proposition 6.3. But for the uncapacitated case we
can make a few extra observations.
Proposition 6.6 For an uncapacitated network, if Q+ = [Y, N \ Y ]+ = ∅

then F ∗ (i) ⊆ Y if i ∈ Y .
From this it easily follows.
Proposition 6.7 If j1 , j2 , . . . , jK is a set of arcs in an uncapacitated network

such that jk ∼ (ik , ik+1 ) and i1 = iK+1 then the nodes i1 , . . . , iK will always
be on the same side of a cut Q if Q+ = ∅.
We utilize this by collapsing all directed circuits in the network. As an

example, consider Figure 8, which is almost like Figure 3, except that arc b
has been turned around.
Since arcs a, d and b, as well as arcs b, c and e, constitute directed circuits,
we can collapse these circuits and arrive at the network in Figure 9. Of course,
it is now much easier to investigate all possible subsets of N .
If a network has both capacitated and uncapacitated arcs, we must
apply the results for capacitated networks, but drop any inequality which
corresponds to a cut where Q+ contains an uncapacitated arc.
2
a f
d
b
1 4 5
g
c e
3
Figure 8 Example network 2, assumed to be uncapacitated.
f,g
1,2,3,4 5
Figure 9 Example network 2 after collapsing the directed circuit.
6.2.2 Comparing the LP and Network Cases

We used Section 5.2 to discuss feasibility in linear programs. Since network
flow problems are just special cases of linear programs, those results apply
here as well, of course. On the other hand, we have just discussed feasibility
in networks more specifically, and apparently the setting was very different.
The purpose of this section is to show in some detail how these results relate
to each other.
Let us first repeat the major discussions from Section 5.2. Using the cone
pos W = {t | t = W y, y ≥ 0}, we defined the polar cone
pos W ∗ = pol pos W = {t | tT y ≤ 0 for all y ∈ pos W }.
The interesting property of the cone pos W ∗ is that the recourse problem is
feasible if and only if a given right-hand side has a non-positive inner product
with all generators of the cone. And if there are not too many generators, it
is much easier to perform inner products than to check if a linear program is
feasible. Refer to Figure 4 for an illustration in three dimensions.1 To find the
polar cone, we used procedure support in Figure 5. The major computational
burden in that procedure is the call to procedure framebylp, outlined in
Figure 1. In principle, to determine if a column is part of the frame, we must
remove the column from the matrix, put it as a right-hand side, and see if
the corresponding system of linear equations has a solution or not. If it has
1 Figures and procedures referred to in this Subsection are contained in Chapter 5
a solution, the column is not part of the frame, and can be removed. An
important property of this procedure is that to determine if a column can be
discarded, we have to use all other columns in the test. This is a major reason
why procedure framebylp is so slow when the number of columns gets very
large.
So, a generator w∗ of the cone pos W ∗ has the property that a right-hand
side h must satisfy hT w∗ ≤ 0 to be feasible. In the uncapacitated network
case we saw that a right-hand side β had to satisfy b(Y )T β ≤ 0 to represent a
feasible problem. Therefore the index vector b(Y ) corresponds exactly to the
column w∗ . And calling procedure framebylp to remove those columns that
are not in the frame of the cone pos W ∗ corresponds to using Proposition 6.5.
Therefore the index vector of a node set from Proposition 6.5 corresponds to
the columns in W ∗ .
Computationally there are major differences, though. First, to find a
candidate for W ∗ , we had to start out with W , and use procedure support,
which is an iterative procedure. The network inequalities, on the other hand,
are produced more directly by looking at all subsets of nodes. But the
most important difference is that, while the use of procedure framebylp,
as just explained, requires all columns to be available in order to determine if
one should be discarded, Proposition 6.5 is totally local. We can pick up
an inequality and determine if it is needed without looking at any other
inequalities. With possibly millions of candidates, this difference is crucial.
We did not develop the LP case for explicit bounds on variables. If such
bounds exist, they can, however, be put in as explicit constraints.
If so, a
b(Y )
column w∗ from W ∗ corresponds to the index vector .
−a(Q+ )
6.3 Generating Relatively Complete Recourse
Let us now discuss how the results obtained in the previous section can help
us, and how they can be used in a setting that deserves the term preprocessing.
Let us first repeat some of our terminology, in order to see how this fits in
with our discussions in the LP setting.
A two-stage stochastic linear programming problem where the second-stage
problem is a directed capacitated network flow problem can be formulated as
follows:
$ %
minx cT x + Q(x)
s.t. Ax = b, x ≥ 0,
where

Q(x) = Q(x, ξ j )pj
and
Q(x, ξ) =
miny1 {(q 1 )T y 1 | W y 1 = h10 + H 1 ξ − T 1 (ξ)x, 0 ≤ y 1 ≤ h20 + H 2 ξ − T 2 (ξ)x},
where W is the node–arc incidence matrix for the network. To fit into a more
general setting, let
W 0
W =
I I
so that Q(x, ξ) can also be written as
Q(x, ξ) = min{q T y | W y = h0 + Hξ − T (ξ)x, y ≥ 0}
y
1 1
y1 2 1 q h0
where y = , y is the slack of y , q = , h 0 = , T (ξ) =
y2 0 h20
1 1
T (ξ) H
2 and H = 2 . Given our definition of β and γ, we have, for a
T (ξ) H
given x̂,
β
= h0 + Hξ − T (ξ)x̂ = h0 + hi ξi − T (ξ)x̂.
γ
i
Using the inequalities derived in the previous section, we can proceed to
transform these inequalities into inequalities in terms of x. By adding these
inequalities to the first-stage constraints Ax = b, we get relatively complete
recourse, i.e. we guarantee that any x satisfying the (expanded) first-stage
˜ An
constraints will yield a feasible second-stage problem for any value of ξ.
inequality has the form

b[A(Y )]T β = β(i) ≤ γ(k) = a(Q+ )T γ.
i∈A(Y ) k∈Q+
Let us replace β and γ with their expressions in terms of x and ξ̃. An

inequality then says that the following must be true for all values of x and all
˜
realizations ξ of ξ:
' ( ' (
b[A(Y )]T h10 + h1i ξi − T 1 (ξ)x ≤ a(Q+ )T h20 + h2i ξi − T 2 (ξ)x .
i i
Collecting all x terms on the left-hand side and all other terms on the right-
hand side we get the following expression:
' + , + ,(
− b(A(Y ))T T01 + Tj1 ξj + a(Q+ )T T02 + Tj2 ξj x
j j
# &
≤ − b[A(Y )]T h1i + a(Q+ )T h2i ξi − b[A(Y )]T h10 + a(Q+ )T h20 .
i
Since this must be true for all possible values of ξ̃, we get one such inequality
for each ξ. If T (ξ) ≡ T0 , we can make this more efficient by calculating only
one cut, given by the following inequality:
# &
− b[A(Y )]T T01 + a(Q+ )T T02 x
# &
≤ min − b[A(Y )]T h1i + a(Q+ )T h2i ξi − b[A(Y )]T h10 + a(Q+ )T h20 .
ξ∈Ξ
i
The minimization is of course very simple in the independent case, since

the minimization can be moved inside the sum. When facets have been
transformed into inequalities in terms of x, we might find that they are linearly
dependent. We should therefore subject them, together with the constraints
Ax = b, to a procedure that removes redundant constraints. We have discussed
this subject in Chapter 5.
The above results have two applications. Both are related to preprocessing.
Let us first repeat the one we briefly mentioned above, namely that, after the
inequalities have been added to Ax = b, we have relatively complete recourse,
i.e. any x satisfying the (expanded) first-stage constraints will automatically
produce a feasible recourse problem for all values of ξ. ˜ This opens up the
avenue to methods that require this property, and it can help in others where
this is really not needed. For example, we can use the L-shaped decomposition
method (page 171) without concern about feasibility cuts, or apply the
stochastic decomposition method as outlined in Section 3.8.
Another—and in our view more important—use of these inequalities is in
model understanding. As expressions in x, they represent implicit assumptions
made by the modeller in terms of the first-stage decisions. They are implicit
because they were never written down, but they are there because otherwise
the recourse problem can become infeasible. And, as part of the model, the
modeller has made the requirements expressed in these implicit constraints. If
there are not too many implicit assumptions, the modeller can relate to them,
and either learn about his or her own model, or might decide that he or she
did not want to make these assumption. If so, there is need for a revision of
the model.
It is worth noting that the inequalities in terms of β and γ are also
interesting in their own right. They show the modeller how the external
flow and arc capacities must combine in order to produce a feasible recourse
problem. Also, this can lead to understanding and/or model reformulation.
6.4 An Investment Example
Consider the simple network in Figure 10. It represents the flow of sewage (or
some other waste) from three cities, represented by nodes 1, 2 and 3.
Figure 10 Transportation network for sewage, used for the example in

Section 6.4.
All three cities produce sewage, and they have local treatment plants to take
care of some of it. Both the amount of sewage from a city and its treatment
capacity vary, and the net variation from a city is given next to the node
representing the city. For example, City 1 always produces more than it can
treat, and the surplus varies between 10 and 20 units per unit time. City 2,
on the other hand, sometimes can treat up to 5 units of sewage from other
cities, but at other times has as much as 15 units it cannot itself treat. City
3 always has extra capacity, and that varies between 5 and 15 units per unit
time.
The solid lines in Figure 10 represent pipes through which sewage can be
pumped (at a cost). Assume all pipes have a capacity of up to 5 units per unit
time. Node 4 is a common treatment site for the whole area, and its capacity
is so large that for practical purposes we can view it as being infinite. Until
now, whenever a city had sewage that it could not treat itself, it first tried to
send it to other cities, or site 4, but if that was not possible, the sewage was
simply dumped in the ocean. (It is easy to see that that can happen. When
City 1 has more than 10 units of untreated sewage, it must dump some of it.)
New rules are being introduced, and within a short period of time dumping
sewage will not be allowed. Four projects have been suggested.
• Increase the capacity of the pipe from City 1 (via City 2) to site 4 with x1
units (per unit time).
• Increase the capacity of the pipe from City 2 to City 3 with x2 units (per
unit time).
• Increase the capacity of the pipe from City 1 (via City 3) to site 4 with x3
units (per unit time).
• Build a new treatment plant in City 1 with a capacity of x4 units (per unit
time).
It is not quite clear if capacity increases can take on any values, or just some
predefined ones. Also, the cost structure of the possible investments are not
yet clear. Even so, we are asked to analyse the problem, and create a better
basis for decisions.
The first thing we must do, to use the procedures of this chapter, is to
make sure that, technically speaking, we have a network (as defined at the
start of the chapter). A close look will reveal that a network must have equality
constraints at the node, i.e. flow in must equal flow out. That is not the case
in our little network. If City 3 has spare capacity, we do not have to send
extra sewage to the city, we simply leave the capacity unused if we do not
need it. The simplest way to take care of this is to introduce some new arcs
in the network. They are shown with dotted lines in Figure 10. Finally, to
have supply equal to demand in the network (remember from Proposition 6.1
that this is needed for feasibility), we let the external flow in node 4 be the
negative of the sum of external flows in the other three nodes.
You may wonder if this rewriting makes sense. What does it mean when
“sewage” is sent along a dotted line in the figure? The simple answer is that
the amount exactly equals the unused capacity in the city to which the arc
goes. (Of course, with the given numbers, we realize that no arc will be needed
from node 4 to node 1, but we have chosen to add it for completeness.)
Now, to learn something about our problem, let us apply Proposition 6.2
to arrive at a number of inequalities. You may find it useful to try to write
them down. We shall write down only some of them. The reason for leaving
out some is the following observation: any node set Y that is such that Q+
contains a dotted arc from Figure 10 will be uninteresting, because
a(Q+ )T γ = ∞,
so that the inequality says nothing interesting. The remaining inequalities are
as follows (where we have used that all existing pipes have a capacity of 5 per
unit time).
⎫
β1 ≤ 10 + x1 + x3 + x4 , ⎪
⎪
⎪
⎪
β2 ≤ 10 + x1 + x2 , ⎪
⎪
⎪
⎪
β3 ≤ 5 + x3 , ⎬
β1 + β2 + β3 ≤ 10 + x1 + x3 + x4 , (4.1)
⎪
β1 + β 2 ≤ 15 + x1 + x2 + x3 + x4 , ⎪ ⎪
⎪
⎪
β1 + β3 ≤ 10 + x1 + x3 + x4 , ⎪
⎪
⎪
⎭
β2 + β3 ≤ 10 + x1 + x3 .
Let us first note that if we set all xi = 0 in (4.1), we end up with a number
of constraints that are not satisfied for all possible values of β. Hence, as we
already know, there is presently a chance that sewage will be dumped.
However, our interest is mainly to find out about which investments to
make. Let us therefore rewrite (4.1) in terms of xi rather than βi :
⎫
x1 +x3 +x4 ≥ β1 − 10 ≥ 10, ⎪
⎪
⎪
⎪
x1 +x2 ≥ β2 − 10 ≥ 5, ⎪
⎪
⎪
⎪
+x3 ≥ β3 − 5 ≥ −10, ⎬
x1 +x3 +x4 ≥ β1 +β2 +β3 − 10 ≥ 20, (4.2)
⎪
⎪
x1 +x2 +x3 +x4 ≥ β1 +β2 − 15 ≥ 20, ⎪
⎪
⎪
⎪
x1 +x3 +x4 ≥ β1 +β3 − 10 ≥ 5, ⎪
⎪
⎭
x1 +x3 ≥ β2 +β3 − 10 ≥ 0.
The last inequality in each constraint of (4.2) is obtained by simply

maximizing over the possible values of β, since what is written down must be
true for all values of β. We can now start to remove some of the constraints
because they do not say anything, or because they are implied by others.
When this cannot be done manually, we can use the methods outlined in
Section 5.1.3. In the arguments that follow, remember that xi ≥ 0.
First, we can remove inequalities 1 and 6, because they are weaker than
inequality 4. But inequality 4, having the same right-hand side as number 5,
but fewer variables on the left-hand side, implies number 5, and the latter can
therefore be dropped. Inequality number 3 is uninteresting, and so is number
7 (since we clearly do not plan to make negative investments). This leaves us
with only two inequalities, which we shall repeat:

x1 + x2 ≥ 5,
(4.3)
x1 + x3 + x4 ≥ 20.
Even though we know nothing so far about investment costs and pumping
costs through the pipes, we know a lot about what limits the options.
Investments of at least five units must be made on a combination of x1 and x2 .
What this seems to say is that the capacity out of City 2 must be increased by
at least 5 units. It is slightly more difficult to interpret the second inequality. If
we see both building pipes and a new plant in City 1 as increases in treatment
capacity (although they are of different types), the second inequality seems to
say that a total of 20 units must be built to facilitate City 1. However, a closer
look at which cut generated the inequality reveals that a more appropriate
interpretation is to say that the three cities, when they are seen as a whole,
must obtain extra capacity of 20 units. It was the node set Y = {1, 2, 3} that
generated the cut.
The two constraints (4.3) are all we need to pass on to the planners. If these
two, very simple, constraints are taken care of, sewage will never have to be
dumped. Of course, if the investment problem is later formulated as a linear
program, the two constraints can be added, thereby guaranteeing feasibility,
and, from a technical point of view, relatively complete recourse.
[-1,1]
2
6 (0,[
) 4 2,6
[2 ,4] ])
(0
(0, 2 ,[2
1
,4 1
])
1
(0,[2,4]) 3 (0,[6,8])
[1,3] 1 2 4 7 5 Slack
2
]) 5 [-2,0]
1 ,8
(0, ,[4 ])
[2, (0 0,2
6]) 3
4 (0,[
8 2
3
[3,3]
Figure 11 Network illustrating the different bounds.
6.5 Bounds
We discussed some bounds for general LPs in Chapter 3. These of course

also apply to networks, since networks are nothing but special cases of linear
programs. The Jensen lower bound can be found by replacing each random
variable (external flow or arc capacity) by its mean and solving the resulting
deterministic network flow problem. The Edmundson–Madansky upper bound
is found by evaluating the network flow problem at all extreme points of the
support. (If the randomness sits in the objective function, the methods give
opposite bounds, just as we discussed for the LP case.)
Figure 11 shows an example that will be used in this section to illustrate
bounds. The terminology is as follows. Square brackets, for example [a, b], are
used to denote supports of random variables. Placed next to a node, they
show the size of the random external flow. Placed in a setting like (c, [a, b]),
the square bracket shows the support of the upper bound on the arc flow for
the arc next to which it is placed. In this setting, c is the lower bound on
the flow. It can become negative in some of the methods. The circled number
next to an arc is the unit arc cost, and the number in a square on the arc is
the arc number. For simplicity, we shall assume that all random variables are
independent and uniformly distributed.
Figure 12 shows the set-up for the Jensen lower bound for the example from
Figure 11. We have now replaced each random variable by its mean, assuming
that the distributions are symmetric. The optimal flow is
f = (2, 0, 0, 0, 2, 2, 1, 1)T,
with a cost of 18.
Although the Edmundson–Madansky distribution is very useful, it still has
the problem that the objective function must be evaluated in an exponential
2
) 4 6 (0,4
(0,3 )
1
2 (0
,3) 1
1
(0,3) 3 (0,7)
2 1 2 4 7 5 Slack
2
) 5 -1
1 ,6
(0 (0 )
,4) (0,1
4
3
8 2
3
3
Figure 12 Example network with arc capacities and external flows
corresponding to the Jensen lower bound.
number of points. If there are k random variables, we must work with 2k

points. This means that with more than about 10 random variables we are
not in business. Thus, since there are 11 random variable in the example, we
must solve 211 problems to find the upper bound. We have not done that here.
In what follows, we shall demonstrate how to obtain a piecewise linear upper
bound that does not exhibit this exponential characterization. A weakness of
this bound is that it may be +∞ even if the problem is feasible. That may
not happen to the Edmundson–Madansky upper bound. We shall continue to
use the network in Figure 11 to illustrate the ideas.
6.5.1 Piecewise Linear Upper Bounds

Let us illustrate the method in a simplified setting. Define φ(ξ, η) by
φ(ξ, η) = min{q T y | W y = b + ξ, 0 ≤ y ≤ c + η},

y
where all elements of the random vectors ξ˜ = (ξ˜1T , ξ˜2T , . . .)T and η̃ =
(η̃1T , η̃2T , . . .)T are mutually independent. Furthermore, let the supports be
given by Ξ(ξ) ˜ = [A, B] and Ξ(η̃) = [0, C]. The matrix W is the node–arc
incidence matrix for a network, with one row removed. That row represents
the slack node. The external flow in the slack node equals the negative sum of
the external flows in the other nodes. The goal is to create an upper bounding
function U (ξ, η) that is piecewise linear, separable and convex in ξ, as well as
easily integrable in η:
d+ (ξ − E ξ̃ ) if ξ ≥ E ξ̃ ,
i i i i i
U (ξ, η) = φ(E ξ̃, 0) + H(η) + −
i
di (E ξ̃i − ξi ) if ξi < E ξ̃i,
for some parameters d± i . The principles of the ξ part of this bound were
outlined in Section 3.4.4 and will not be repeated in all details here. We shall
use the developments from that section here, simply by letting η = 0 while
developing the ξ part. Because this is a restriction (constraint) on the original
problem, it produces an upper bound. Then, afterwards, we shall develop
H(η). In Section 3.4.4 we assumed that E ξ̃ = 0. We shall now drop that
assumption, just to illustrate that it was not needed, and to show how many
parameters can be varied in this method.
Let us first see how we can find the ξ part of the function, leaving η = 0.
First, let us calculate
φ(E ξ̃, 0) = min{q T y | W y = b + E ξ̃, 0 ≤ y ≤ c} = q T y 0 .

y
This is our basic setting, and all other values of ξ will be seen as deviations
from E ξ̃. Note that since y 0 is “always” there, we shall update the arc
capacities to become −y 0 ≤ y ≤ c − y 0 . For this purpose, we define α1 = −y 0
and β 1 = c − y 0 . Let ei be a unit vector of appropriate dimension with a +1
in position i.
Next, define a counter r and let r := 1. Now, check out the case when
ξ1 > E ξ̃1 by solving
min{q T y | W y = er (Br − E ξ̃r ), αr ≤ y ≤ β r } = q T y r+ = d+

r (Br − E ξ̃r ).
y
(5.1)
Similarly, check out the case with ξ1 < E ξ̃1 by solving
min{q T y | W y = er (Ar − E ξ̃r ), αr ≤ y ≤ β r } = q T y r− = d−

r (Ar − E ξ̃r ).
y
(5.2)
Now, based on y r± , we shall assign portions of the arc capacities to the random
variable ξ̃r . These portions will be given to ξ̃r and left unused by other random
variables, even when ξ˜r does not need them. The portions will correspond
to paths in the network connecting node r to the slack node (node 5 in the
example). That is done by means of the following problem, where we calculate
what is left for the next random variable:
αr+1
i = αri − min{yir+ , yir− , 0}. (5.3)
What we are doing here is to find, for each variable, how much ξ˜r , in the
worst case, uses of arc i in the negative direction. That is then subtracted
from what we had before. There are three possibilities. We may have both
(5.1) and (5.2) yielding nonnegative values for the variable i. Then nothing is
used of the available “negative capacity” αri . Then αr+1
i = αri . Alternatively,
2
) 4 6 (0,2
(0,2 )
1
2 (0
,2) 1
1
(0,2) 3 (0,6)
2 1 2 4 7 5 Slack
2
) 5 -1
1 ,4
(0 (0 )
,2) (0,0
4
3
8 2
3
3
Figure 13 ˜ 0) for the network in Figure 11.

Network needed to calculate φ(E ξ,
when (5.1) has yir+ < 0, it will in the worst case use yir+ of the available
“negative capacity”. Finally, when (5.2) has yir− < 0, in the worst case we
use yir− of the capacity. Therefore, αr+1
i is what is left for the next random
variable. Similarly, we find
βir+1 = βir − max{yir+ , yir− , 0}, (5.4)

where βir+1 shows how much is still available of the capacity on arc i in the
forward (positive) direction.
We next increase the counter r by one and repeat (5.1)–(5.4). This takes
care of the piecewise linear functions in ξ.
Let us now look at our example in Figure 11. To calculate the ξ part of the
bound, we put all arc capacities at their lowest possible value and external
flows at their means. This is shown in Figure 13.
The optimal solution in Figure 13 is given by
y 0 = (2, 0, 0, 0, 3, 2, 2, 0)T,
with a cost of 22. The next step is update the arc capacities in Figure 13 to
account for this solution. The result is shown in Figure 14.
Since the external flow in node 1 varies between 1 and 3, and we have so
far solved the problem for a supply of 2, we must now find the cost associated
with a supply of 1 and a demand of 1 in node 1. For a supply of 1 we get the
solution
y 1+ = (0, 1, 0, 0, 0, 0, 1, 0)T,
with a cost of 5. Hence d+
1 = 5. For a demand of 1 we get
y 1− = (−1, 0, 0, 0, 0, −1, 0, 0)T,

2
4 6 (-2,
,0) 0)
(-2 2 1
(0
,2) 1
1
(0,2) 3 (-2,4)
+/- 1
1 2 4 7 5
2 -/+ 1
1 ) 5
(0 1
(-3, )
,2) (0,0
4
3
8 2
3
Figure 14 ˜ 0).
Arc capacities after the update based on φ(E ξ,
+/- 1
2
4 6 (-1,
,0) 0)
(-1 2 1
(0
,2) 1
1
(0,1) 3 (-2,3)
1 2 4 7 5
2 -/+ 1
) 5
1 ,1
(0 (-3 )
,2) (0,0
4
3
8 2
3
Figure 15 ˜ 0) and node 1.

with a cost of −3, so that d−1 = 3. Hence we have used one unit of the forward
capacity of arcs 2 and 7, and one unit of the reverse capacity of arcs 1 and
6. Note that both solutions correspond to paths between node 1 and node 5
(the slack node). We update to get Figure 15.
For node 2 the external flow varies between −1 and 1, so we shall now check
the supply of 1 and demand of 1 based on the arc capacities of Figure 15. For
supply we get
y 2+ = (0, 0, 0, 1, 0, 0, 1, 0)T,
with a cost of 3. For the demand of 1 we obtain
y 2− = (0, 0, 0, 0, 0, −1, 0, 0)T,
with a cost of −1. Hence d+ −

2 = 3 and d2 = 1. Node 3 had deterministic
external flow, so we turn to node 4. Node 4 had a demand between 0 and 2
2
4 6 (0,0
,0) )
(-1 2 1
(0
,1) 1
1
(0,1) 3 (-2,2)
1 2 4 7 5
2 -/+ 1
) 5 +/- 1
1
(0 1
(-3, )
,2) (0,0
4
3
8 2
3
Figure 16 ˜ 0) and nodes 1 and
2.
units, and we have so far solved for a demand of 1. Therefore we must now
look at a demand of 1 and a supply of 1 in node 4, based on the arc capacities
in Figure 16. In that figure we have updated the capacities from Figure 15
based on the solutions for node 2.
A supply in node 4 gives us the solution
y 4+ = (0, 0, 0, 0, 0, 0, 1, 0)T,
with a cost of 2. One unit demand, on the other hand, gives us
y 4− = (0, 0, 0, 0, 0, −1, 0)T,
with a cost of −2. The parameters are therefore d+ −

4 = 2 = d4 . This leaves the
arc capacities in Figure 17.
What we have found so far is as follows:
φ(ξ, η) = 22+ H(η)

5(ξ1 − 2) if ξ1 ≥ 2,
+
3(ξ1 − 2) if ξ1 < 2,

3ξ2 if ξ2 ≥ 0,
+
ξ2 if ξ2 < 0,

2(ξ4 + 1) if ξ4 ≥ −1,
+
2(ξ4 + 1) if ξ4 < −1.
If, for simplicity, we assume that all distributions are uniform, we easily
2
4 6 (0,0
,0) )
(-1 2 1
(0
,1) 1
1
(0,1) 3 (-1,1)
1 2 4 7 5
2
) 5
1 ,1
(0 (-3 )
,2) (0,0
4
3
8 2
3
˜ 0) and external flow
Figure 17 Arc capacities after the update based on φ(E ξ,
in all nodes.
integrate the upper-bounding function U (ξ, η) to obtain
U = 22 + H(η)
2 3
+ 1 3(ξ1 − 2) 12 dξ1 + 2 5(ξ1 − 2) 12 dξ1
0 1
+ −1 ξ2 12 dξ2 + 0 3ξ2 12 dξ2
−1 0
+ −2 2(ξ4 + 1) 12 dξ4 + −1 2(ξ4 + 1) 21 dξ4
= 22 + H(η) − 3 × 1
4
+5× 1
4
−1× 1
4
+3× 1
4
−2× 1
4
+2× 1
4
= 23 + H(η).
Note that there is no contribution from ξ4 to the upper bound. The reason
is that the recourse function φ(ξ, η) is linear in ξ4 . This property of discovering
that the recourse function is linear in some random variable is shared with
the Jensen and Edmundson–Madansky bounds.
We then turn to the η part of the bound. Note that if (5.3) and (5.4) were
calculated after the final y r± had been found, the α and β show what is left
of the deterministic arc capacities after all random variable ξ˜i have received
their shares. Let us call these α∗ and β ∗ . If we add to each upper bound in
Figure 17 the value C (remember that the support of the upper arc capacities
was Ξ = [0, C]), we get the arc capacities of Figure 18. Now we solve the
problem
min{q T y | W y = 0, α∗ ≤ y ≤ β ∗ + C} = q T y ∗ . (5.5)
y
With zero external flow in Figure 18, we get the optimal solution
y ∗ = (0, 0, 0, 0, −1, 0, −1, 1)T,

2
4 6 (0,4
,2) )
(-1 2 1
(0
,3) 1
1
(0,3) 3 (-1,3)
1 2 4 7 5
2
) 5
1 ,5
(0 (-3 )
,6) (0,2
4
3
8 2
3
Figure 18 Arc capacities used to calculate H(η) for the example in Figure 11.
with a cost of −4. This represents cycle flow with negative costs. The cycle
became available as a result of making arc 8 having a positive arc capacity.
If, again for simplicity, we assume that η8 is uniformly distributed over [0, 2],
we find that the capacity of that cycle has a probability of being 1 equal to
0.5. The remaining probability mass is uniformly distributed over [0, 1]. We
therefore get
1
EH(η) = −4 × 1 × 12 − 4 1
2
x dx = −2 − 1 = −3.
0
The total upper bound for this example is thus 23 − 3 = 20, compared with
the Jensen lower bound of 18.
In this example the solution y ∗ of (5.5) contained only one cycle. In general,
y ∗ may consist of several cycles, possibly sharing arcs. It is then necessary to
pick y ∗ apart into individual cycles. This can be done in such a way that all
cycles have nonpositive costs (those with zero costs can then be discarded),
and such that all cycles that use a common arc use it in the same direction.
We shall not go into details of that here.
6.6 Project Scheduling
We shall spend a whole section on the subject of project scheduling, and we

shall do so in a setting of PERT (project evaluation and review technique)
networks. There are several reasons for looking specifically at this class
of problems. First, project scheduling is widely used, and therefore known
to many people. Even though it seems that the setting of CPM (critical
path method) is more popular among industrial users, the difference is
not important from a principle point of view. Secondly, PERT provides us
with a genuine opportunity to discuss some modelling issues related to the

relationship between time periods and stages. We shall see that PERT has
sometimes been cast in a two-stage setting, but that it can be hard to interpret
that in a useful way. Thirdly, the more structure a problem has, the better
bounds can often be found. PERT networks provide us with a tool for showing
how much structure provide tight bounds.
Before we continue, we should like to point out a possible confusion in
terms. When PERT was introduced in 1959, it was seen as a method for
analysing projects with stochastic activity durations. However, the way in
which randomness was treated was quite primitive (in fact, it is closely
related to the Jensen lower bound that we discussed in Section 3.4.1).
Therefore, despite the historical setting, many people today view PERT as
a deterministic approach, simply disregarding what the original authors said
about randomness. When we use the term PERT in the following, we shall
refer to the mathematical formulation with its corresponding deterministic
solution procedure, and not to its original random setting.
There are many ways to formulate the PERT problem. For our purpose,
the following will do. A PERT network is a network where arcs correspond
to activities, and nodes to events. If arc k ∼ (i, j) then activity k can start
when event i has taken place, and event j can take place when all activities k ,
with k ∼ (i , j), have finished. A PERT network must be acyclic, otherwise
an activity must finish before it can start—a meaningless situation. Because
of acyclicity, we can number nodes, such that if k ∼ (i, j) then i < j. As a
consequence, node 1 represents the event “We are ready to start” and n the
event “The project is finished”. Let πi be the time event i takes place, and
let us define π1 := 0. Furthermore, let qk be the duration of activity k. Since
an event can take place only after all activities preceding it have finished, we
must have
πj ≥ πi + qk for all k ∼ (i, j).
Since πn is the time at which the project finishes, we can calculate the
minimal project completion time by solving
⎫
min πn ⎬
s.t. πj − πi ≥ qk for all k ∼ (i, j), (6.1)
⎭
π1 = 0.
It is worth noting that (6.1) is not really a decision problem. There are namely
no decisions. We are only calculating consequences of an existing setting of
relations and durations.
6.6.1 PERT as a Decision Problem

As pointed out, (6.1) is not a decision problem, since there are no decisions
to be made. Very often, activity durations are not given by nature, but can
be affected by how much resources we put into them. For example, it takes
longer to build a house with one carpenter than with two. Assume we have
available a budget of B units of resources, and that if we spend one unit on
activity k, its duration will decrease by ak time units. A possible decision
problem is then to spend the budget in such a way that the project duration
is minimized. This can be achieved by solving the following problem:
⎫
min πn ⎪
⎪
⎪
s.t. πj − πi ≥ qk − ak xk for all k ∼ (i, j), ⎪
⎪
⎪
⎪
⎬
xk ≤ B, (6.2)
⎪
⎪
k
⎪
⎪
π1 = 0, ⎪
⎪
⎪
⎭
xk ≥ 0.
Of course, there might be other constraints, such as xk ≤ ck , but they can be
added to (6.2) as needed.
6.6.2 Introduction of Randomness

It seems natural to assume that activity durations are random. If so, the
project duration is also random, and we can no longer talk about finding the
minimal project duration time. However, a natural alternative seems to be
to look for the expected (minimal) project duration time. In (6.1) and (6.2)
the goal would then be to minimize Eπn . However, we must now be careful
about how we interpret the problems. Problem (6.1) is simple enough. There
are still no decisions, so we are only trying to calculate when, on expectation,
the project will finish, if all activities start as soon as they can. But when
we turn to (6.2) we must be careful. In what order do things happen? Do we
first decide on x, and then simply sit back (as we did with (6.1)) and observe
what happens? Or do we first observe what happens, and then make decisions
on x? These are substantially different situations. It is of importance that
you understand the modelling aspects of this difference. (There are solution
differences as well, but they are less interesting now.) In a sense, the two
interpretations bound the correct problem from above and below.
If we interpret (6.2) as a problem where x is determined before the activity
durations are known, we have in fact a standard two-stage stochastic program.
The first-stage decision is to find x, and the second-stage “decision” to find the
project duration given x and a realization of q(ξ).˜ (We put q(ξ)˜ to show that
q is indeed a random variable.) But—and this is perhaps the most important
question to ask in this section—is this a good model? What does it mean?
First, it is implicit in the model that, while the original activity durations
are random, the changes ak xk are not. In terms of probability distributions,
therefore what we have done is to reduce the means of the distributions
describing activity durations, but without altering the variances. This might
or might not be a reasonable model. Clearly, if we find this unreasonable, we
could perhaps let ak be a random variable as well, thereby making also the
effect of the investment xk uncertain.
The above discussion is more than anything a warning that whenever we
introduce randomness in a model, we must make sure we know what the
randomness means. But there is a much more serious model interpretation if
we see (6.2) as a two-stage problem. It means that we think we are facing
a project where, before it is started, we can make investments, but where
afterwards, however badly things go, we shall never interfere in order to fix
shortcomings. Also, even if we are far ahead of schedule, we shall not cut back
on investments to save money. We may ask whether such projects exist—
projects where we are free to invest initially, but where afterwards we just sit
back and watch, whatever happens.
From this discussion you may realize (as you have before—we hope) that the
definition of stages is important when making models with stochasticity. In
our view, project scheduling with uncertainty is a multistage problem, where
decisions are made each time new information becomes available. This makes
the problem extremely hard to solve (and even formulate—just try!) But this
complexity cannot prevent us from pointing out the difficulties facing anyone
trying to formulate PERT problems with only two stages.
We said earlier that there were two ways of interpreting (6.2) in a setting
of uncertainty. We have just discussed one. The other is different, but has
similar problems. We could interpret (6.2) with uncertainties as if we first
observed the values of q and then made investments. This is the “wait-and-
see solution”. It represents a situation where we presently face uncertainty, but
where all uncertainty will be resolved before decisions have to be made. What
does that mean in our context? It means that before the project starts, all
uncertainty related to activities disappears, everything becomes known, and
we are faced with investments of the type (6.2). If the previous interpretation
of our problem was odd, this one is probably even worse. In what sort of
project will we have initial uncertainty, but before the first activity starts,
everything, up to the finish of the project, becomes known? This seems almost
as unrealistic as having a deterministic model of the project in the first place.
6.6.3 Bounds on the Expected Project Duration

Despite our own warnings in the previous subsection, we shall now show
how the extra structure of PERT problems allows us to find bounds on the
expected project duration time if activity durations are random. Technically
speaking, we are looking for the expected value of the objective function in
˜ There is a very large collection
(6.1) with respect to the random variables q(ξ).
of different methods for bounding PERT problems. Some papers are listed at
the end of this chapter. However, most, if not all, of them can be categorized
as belonging to one or more of the following groups.
6.6.3.1 Series reductions

If there is a node with only one incoming and one outgoing arc, the node is
removed, and the arcs replaced by one arc with a duration equal to the sum
of the two arc durations. This is an exact reformulation.
6.6.3.2 Parallel reductions

If two arcs run in parallel with durations ξ̃1 and ξ˜2 then they are replaced
with one arc having duration max{ξ˜1 , ξ̃2 }. This is also an exact reformulation.
6.6.3.3 Disregarding path dependences

Let π̃i be a random variable describing when event i takes place. Then we can
calculate
π̃j = max ˜
{π̃i + qk (ξ)}, with k ∼ (i, j),
i∈B + (i)\{i}
as if all these random variables were independent. However, in a PERT

network, the π̃’s will normally be dependent (even if the q’s are independent),
since the paths leading up to the nodes usually share some arcs. Not only will
they be dependent, but the correlation will always be positive, never negative.
Hence viewing the random variables as independent will result in an upper
bound on the project duration. The reason is that E max{ξ˜1 , ξ˜2 } is smaller
if the nonnegative ξ˜1 and ξ˜2 are (positively) correlated than if they are not
correlated. A small example illustrates this
Example 6.3 Assume we have two random variables ξ˜1 and ξ̃2 , with joint
distribution as in Table 1. Note that both random variables have the same
marginal distributions; namely, each of them can take on the values 1 or 2,
each with a probability 0.5. Therefore E max{ξ˜1 , ξ̃2 } = 1.7 from Table 1, but
0.25(1 + 2 + 2 + 2) = 1.75 if we use the marginal distributions as independent
distributions. Therefore, if ξ̃1 and ξ˜2 represent two paths with some joint arc,
disregarding the dependences will create an upper bound.
2
Table 1 Joint distribution for ξ˜1 and ξ˜2 , plus the calculation of max{ξ˜1 , ξ̃2 }.
ξ̃1 ξ̃2 Prob. max

1 1 0.3 1
1 2 0.2 2
2 1 0.2 2
2 2 0.3 2
Figure 19 Arc duplication.
6.6.3.4 Arc duplications

If there is a node i with B + (i ) = {i, i }, so that the node has only one
incoming arc k ∼ (i, i ), remove node i , and for each j ∈ F + (i )\ {i } replace
k ∼ (i , j) by k ∼ (i, j). The new arc has associated with it the random
duration qk (ξ) ˜ := qk (ξ) ˜ If arc k had a deterministic duration,
˜ + qk (ξ).
this is an exact reformulation. If not, we get an upper bound based on the
previous principle of disregarding path dependences. (This method is called
arc duplication because we duplicate arc k and use one copy for each arc k .)
An exactly equal result applies if there is only one outgoing arc. This result
is illustrated in Figure 19, where F + (i ) = {i , 1, 2, 3}.
If there are several incoming and several outgoing arcs, we may pair up all
incoming arcs with all outgoing arcs. This always produces an upper bound
based on the principle of disregarding path dependences.
6.6.3.5 Using Jensen’s inequality

˜ or a
Since our problem is convex in ξ, we get a lower bound whenever a qk (ξ)
π̃i (as defined above) is replaced by its mean.
Note that if we have a node and choose to apply arc duplication, we
get an exact reformulation if all incoming arcs and all outgoing arcs have
deterministic durations, an upper bound if they do not, and a lower bound if

we first replace the random variables on the incoming and outgoing arcs by
their means and then apply arc duplication. If there is only one arc in or one
arc out, we take the expectation for that arc, and then apply arc duplication,
observing an overall lower bound.
The vocabulary in this chapter is mostly taken from Rockafellar [25], which
also contains an extremely good overview of deterministic network problems.
A detailed look at network recourse problem is found in Wallace [28].
The original feasibility results for networks were developed by Gale [10]
and Hoffman [13]. The stronger versions using connectedness were developed
by Wallace and Wets. The uncapacitated case is given in [31], while the
capacitated case is outlined in [33] (with a proof in [32]). More details of the
algorithms in Figures 6 and 7 can also be found in these papers. Similar results
were developed by Prékopa and Boros [23]. See also Kall and Prékopa [14].
As for the LP case, model formulations and infeasibility tests have of course
been performed in many contexts apart from ours. In addition to the references
given in Chapter 5, we refer to Greenberg [11, 12] and Chinneck [3].
The piecewise linear upper bound is taken from Wallace [30]. At the very
end of our discussion of the piecewise linear upper bound, we pointed out that
the solution y ∗ to (5.5) could consist of several cycles sharing arcs. A detailed
discussion of how to pick y ∗ apart, to obtain a conformal realization can be
found in Rockafellar [25], page 476. How to use it in the bound is detailed
in [30]. The bound has been strengthened for pure arc capacity uncertainty
by Frantzeskakis and Powell [8].
Special algorithms for stochastic network problems have also been
developed; see e.g. Qi [24] and Sun et al. [27].
We pointed out at the beginning of this chapter that scenario aggregation
(Section 2.6) could be particularly well suited to problems that have network
structure in all periods. This has been utilized by Mulvey and Vladimirou
for financial problems, which can be formulated in a setting of generalized
networks. For details see [19, 20]. For a selection of papers on financial
problems (not all utilizing network structures), consult Zenios [36, 37], and,
for a specific application, see Dempster and Ireland [5].
The above methods are well suited for parallel processing. This has been
done in Mulvey and Vladimirou [18] and Nielsen and Zenios [21].
Another use of network structure to achieve efficient methods is described
in Powell [22] for the vehicle routing problem.
The PERT formulation was introduced by Malcolm et al. [17]. An overview
of project scheduling methods can be found in Elmaghraby [7]. A selection of
Figure 20 Example network for calculating bounds.
bounding procedures based on the different ideas listed above can be found in
the following: Fulkerson [9], Kleindorfer [16], Shogan [26], Kamburowski [15]
and Dodin [6]. The PERT problem as an investment problem is discussed in
Wollmer [34].
The max flow problem is another special network flow problem that is much
studied in terms of randomness. We refer to the following papers, which discuss
both bounds and a two-stage setting: Cleef and Gaul [4], Wollmer [35], Aneja
and Nair [1], Carey and Hendrickson [2] and Wallace [29].
Exercises
1. Consider the network in Figure 20. The interpretation is as for Figure 11

regarding parameters, except that we for the arc capacities simply have
written a number next to the arc. All lower bounds on flow are zero.
Calculate the Jensen lower bound, the Edmundson-Madansky upper
bound, and the piecewise linear upper bound for the expected minimal
cost in the network.
2. When outlining the piecewise linear upper bound, we found a function that
was linear both above and below the expected value of the random variable.
Show how (5.1) and (5.2) can be replaced by a parametric linear program
to get not just one linear piece above the expectation and one below, but
rather piecewise linearity on both sides. Also, show how (5.3) and (5.4)
must then be updated to account for the change.
3. The max flow problem is the problem of finding the maximal amount of
flow that can be sent from node 1 to node n in a capacitated network. This
problem is very similar to the PERT problem, in that paths in the latter
correspond to cuts in the max flow problem. Use the bounding ideas listed
in Section 6.6.3 to find bounds on the expected max flow in a network with
random arc capacities.
4. In our example about sewage treatment in Section 6.4 we introduced four
investment options.
(a) Assume that a fifth investment is suggested, namely to build a pipe
with capacity x5 directly from City 1 to site 4. What are the constraints
on xi for i = 1, . . . , 5 that must now be satisfied for the problem to be
feasible?
(b) Disregard the suggestion in question (a). Instead, it is suggested to
see the earlier investment 1, i.e. increasing the pipe capacity from City
1 to cite 4 via City 2 as two different investment. Now let x1 be the
increased capacity from City 1 to City 2, and x5 the increased capacity
from City 2 to cite 4 (the dump). What are now the constraints on xi
for i = 1, . . . , 5 that must be satisfied for the problem to be feasible?
Make sure you interpret the constraints.
5. Develop procedures for uncpacitated networks corresponding to those in

Figures 4, 6 and 7.
References
[1] Aneja Y. P. and Nair K. P. K. (1980) Maximal expected flow in a network

subject to arc failures. Networks 10: 45–57.
[2] Carey M. and Hendrickson C. (1984) Bounds on expected performance
of networks with links subject to failure. Networks 14: 439–456.
[3] Chinneck J. W. (1990) Localizing and diagnosing infeasibilities in
networks. Working paper, Systems and Computer Engineering, Carleton
University, Ottawa, Ontario.
[4] Cleef H. J. and Gaul W. (1980) A stochastic flow problem. J. Inf. Opt.
Sci. 1: 229–270.
[5] Dempster M. A. H. and Ireland A. M. (1988) A financial expert decision
support system. In Mitra G. (ed) Mathematical Methods for Decision
Support, pages 415–440. Springer-Verlag, Berlin.
[6] Dodin B. (1985) Reducability of stochastic networks. OMEGA Int. J.
Management 13: 223–232.
[7] Elmaghraby S. (1977) Activity Networks: Project Planning and Control
by Network Models. John Wiley & Sons, New York.
[8] Frantzeskakis L. F. and Powell W. B. (1989) An improved polynomial
bound for the expected network recourse function. Technical report,
Statistics and Operations Research Series, SOR-89-23, Princeton
University, Princeton, New Jersey.

[9] Fulkerson D. R. (1962) Expected critical path lengths in PERT networks.
Oper. Res. 10: 808–817.
[10] Gale D. (1957) A theorem of flows in networks. Pac. J. Math. 7: 1073–
1082.
[11] Greenberg H. J. (1987) Diagnosing infeasibility in min-cost network flow
problems. part I: Dual infeasibility. IMA J. Math. in Management 1:
99–109.
[12] Greenberg H. J. (1988/9) Diagnosing infeasibility in min-cost network
flow problems. part II: Primal infeasibility. IMA J. Math. in Management
2: 39–50.
[13] Hoffman A. J. (1960) Some recent applications of the theory of a
multivariate random variable. Proc. Symp. Appl. Math. 10: 113–128.
[14] Kall P. and Prékopa A. (eds) (1980) Recent Results in Stochastic
Programming, volume 179 of Lecture Notes in Econ. Math. Syst. Springer-
Verlag, Berlin.
[15] Kamburowski J. (1985) Bounds in temporal analysis of stochastic
networks. Found. Contr. Eng. 10: 177–185.
[16] Kleindorfer G. B. (1971) Bounding distributions for a stochastic acyclic
network. Oper. Res. 19: 1586–1601.
[17] Malcolm D. G., Roseboom J. H., Clark C. E., and Fazar W. (1959)
Applications of a technique for R&D program evaluation. Oper. Res.
7: 646–696.
[18] Mulvey J. M. and Vladimirou H. (1989) Evaluation of a parallel hedging
algorithm for stochastic network programming. In Sharda R., Golden
B. L., Wasil E., Balci O., and Stewart W. (eds) Impact of Recent
Computer Advances on Operations Research, pages 106–119. North-
Holland, New York.
[19] Mulvey J. M. and Vladimirou H. (1989) Stochastic network optimization
models for investment planning. Ann. Oper. Res. 20: 187–217.
[20] Mulvey J. M. and Vladimirou H. (1991) Applying the progressive hedging
algorithm to stochastic generalized networks. Ann. Oper. Res. 31: 399–
424.
[21] Nielsen S. and Zenios S. A. (1993) A massively parallel algorithm for
nonlinear stochastic network problems. Oper. Res. 41: 319–337.
[22] Powell W. B. (1988) A comparative review of alternative algorithms for
the dynamic vehicle allocation problem. In Golden B. and Assad A. (eds)
Vehicle Routing: Methods and Studies, pages 249–291. North-Holland,
Amsterdam.
[23] Prékopa A. and Boros E. (1991) On the existence of a feasible flow in a
stochastic transportation network. Oper. Res. 39: 119–129.
[24] Qi L. (1985) Forest iteration method for stochastic transportation
problem. Math. Prog. Study 25: 142–163.
[25] Rockafellar R. T. (1984) Network Flows and Monotropic Optimization.

John Wiley & Sons, New York.
[26] Shogan A. W. (1977) Bounding distributions for a stochastic PERT
network. Networks 7: 359–381.
[27] Sun J., Tsai K. H., and Qi L. (1993) A simplex method for network
programs with convex separable piecewise linear costs and its application
to stochastic transshipment problems. In Du D. Z. and Pardalos P. M.
(eds) Network Optimization Problems: Algorithms, Applications and
Complexity, pages 283–300. World Scientific, Singapore.
[28] Wallace S. W. (1986) Solving stochastic programs with network recourse.
Networks 16: 295–317.
[29] Wallace S. W. (1987) Investing in arcs in a network to maximize the
expected max flow. Networks 17: 87–103.
[30] Wallace S. W. (1987) A piecewise linear upper bound on the network
recourse function. Math. Prog. 38: 133–146.
programming: The case of uncapacitated networks. ORSA J.Comp. 1:
252–270.
[32] Wallace S. W. and Wets R. J.-B. (1993) The facets of the polyhedral set
determined by the Gale–Hoffman inequalities. Math. Prog. 62: 215–222.
programming: The case of capacitated networks. ORSA J.Comp. 7: 44–
62.
[34] Wollmer R. D. (1985) Critical path planning under uncertainty. Math.
Prog. Study 25: 164–171.
[35] Wollmer R. D. (1991) Investments in stochastic maximum flow problems.
Ann. Oper. Res. 31: 459–467.
[36] Zenios S. A. (ed) (1993) Financial Optimization. Cambridge University
Press, Cambridge, UK.
[37] Zenios S. A. (1993) A model for portfolio management with mortgage-
backed securities. Ann. Oper. Res. 43: 337–356.
Index
absolutely continuous, 31 complete recourse, see stochastic pro-
accumulated return function, see dy- gram with
namic programming cone, see convex
almost everywhere (a.e.), 27 connected network, see networks
almost surely (a.s.), 16, 28 convex
approximate optimization, 218 cone, 60
augmented Lagrangian, see Lagrange polar, 163
function polyhedral cone, 39, 60, 69, 160
generating elements, 60, 69, 79,
backward recursion, see dynamic pro- 163, 166
gramming polyhedral set, 62
barrier function, 97 polyhedron, 58, 91, 234
basic solution, see feasible vertex, 58, 175
basic variables, 56, 65 convex hull, 43, 57
Bellman, 110, 115–117, 121 convex linear combination, 57
optimality principle, 115 cross out, see decision tree
cut, see networks
solution procedure, 121
cutting plane method, see methods
Benders’ decomposition, 213, 233
(nlp)
block-separable recourse, 233
bounds decision node, see decision tree
Edmundson–Madansky upper bound, decision tree
181–185, 192, 194, 203, 234 chance node, 124
Jensen lower bound, 179–182, 184, cross out, 126
185, 218, 220, 233 decision node, 124
limited information, 234 deterministic, 121–123
piecewise linear upper bound, 185– definition, 121
190, 234 folding back, 123
example, 187–189 stochastic, 124–129
stopping criterion, 212 density function, 30, 51
bunching, 230 descent direction, 84, 226
deterministic equivalent, 21, 31–36, 103
cell, 183, 190, 196, 201, 203, 212, 234 deterministic method, 217
chance constraints, see stochastic pro- distribution function, 30
gram with dual decomposition
chance node, see decision tree data structure, 17, 42
complementarity conditions, 84, 89 master program, 173
method, 75–80, 161, 168, 173
1 Italic page numbers (e.g. 531) indicate to dual program, see linear program
literature. duality gap, 74
duality theorem transaction costs, 146

strong, 74 first-stage
weak, 72 costs, 15, 31
dynamic programming fishery model, 138, 159, 234
accumulated return function, 117 forestry model, 234
backward recursion, 114 free variables, 54
deterministic, 117–121 function
solution procedure, 121 differentiable, 37, 81
immediate return, 110 integrable, 30
monotonicity, 115 separable, 206
return function, 117 simple, 28
separability, 114
stage, 110 gamblers, 128
state, 110, 117 generators, see convex
stochastic, 130–133 global optimization, 8
solution procedure, 133 gradient, 38, 84, 226
time horizon, 117
transition function, 117 here-and-now, 151
dynamic systems, 110–116 hindsight, 5
hydro power production, 147–150
Edmundson-Madansky upper bound, additional details, 150
see bounds numerical example, 148–149
event, 25
event tree, 134, 135 immediate return, see dynamic pro-
EVPI, 154–156 gramming
expectation, 30 implementable decision, 136, 141
expected profit, 126, 128 indicator function, 28
expected value of perfect information, induced
see EVPI constraints, 43, 214
expected value solution, 3 feasible set, 43
integer programming, see program
facet, 62, 213, 216 integral, 28, 30
Farkas’ lemma, 75, 163 interior point method, 233
fat solution, 15
feasibility cut, 77, 103, 161–168, 173, Jensen inequality, 180, 202
177, 203, 214 Jensen lower bound, see bounds
example, 166–167
feasible Kuhn–Tucker conditions, 83–89
basic solution, 55
degenerate, 64 L-shaped method, 80, 161–173, 213,
nondegenerate, 64 217, 220, 229, 233
basis, 55 algorithms, 168–170
set, 55 example, 172–173
feasible direction MSLiP, 233
method, see methods (nlp) within approximation scheme, 201–
financial models, 141–147 203
efficient frontier, 143, 144 algorithm, 203
Markowitz’ mean-variance, 142–143 Lagrange function, 88
weak aspects, 143–144 augmented, 99
multistage, 145–147 multipliers, 84
portfolio, 142 saddle point, 89
Index 315
Lagrangian methods, see methods (nlp) financial model, 145

linear program PERT, see PERT
dual program, 70 node–arc incidence matrix, see net-
multiple right-hand sides, 229–233 works
primal program, 70 nonbasic variables, 56, 65
standard form, 53 nonlinear programming, 80–102, 104
linear programming, 53–80, 103
parametric, 187 optimality condition, 65
log-concave necessary, 84
measure, 50, 51, 103 sufficient, 84
loss function, 97 optimality cut, 78, 103, 168–173, 177,
203
Markowitz’ model, 142–143 optimality principle, see Bellman
options, 144 option, 4–6
transaction costs, 144 options, 144
weak aspects, 143–144 outcome, 25
maximum
relative, 8 parallel processing, 233
mean fundamental, 29 partition, 28, 234
measurable set, 24 curvature of function, 192
measure, 22 example, 197–201
“natural”, 24 look ahead, 196, 205
extreme, 234 look-ahead, 196
probability, 25 point of, 212
measure theory, 103 quality of, 203–205
methods (nlp), 89–102 refinement of, 190–201, 212
augmented Lagrangian, 99, 104, 137 penalty method, see methods (nlp)
update, 100 piecewise linear upper bound, see bounds
cutting planes, 90–93, 104 pivot step, 68
descent directions, 93–96 polar cone, see convex
feasible directions, 95, 104 polar matrix, 163
Lagrangian, 98–102 generating elements, 163
penalties, 97–98, 104 polyhedral cone, see convex
reduced gradients, 95, 104 polyhedral set, see convex
minimization polyhedron, see convex
constrained, 82 positive hull, 60
unconstrained, 82 preprocessing
minimum induced constraints, see induced con-
global, 9 straints
local, 8 simplified feasibility test, see feasibil-
relative, 8 ity cut
model understanding, 145 probabilistic constraints, see stochastic
monotonicity, see dynamic program- program with
ming probability
MSLiP, 233 distribution, 25
multicut method, 80 space, 25
multipliers, see Lagrange function theory, 103
multistage, see stochastic program with PROCON, 103
program
nested decomposition, 233 convex, 8
networks integer, 8, 209
best-so-far, 211 return function, see dynamic program-

bounding, 210, 211 ming
branch, 210 risk averse, 128
branch-and-bound, 210–212 risk-neutral, 128
branch-and-cut, 214, 216
branching variable, 210, 212 saddle point, see Lagrange function
cutting-plane, 212, 214 sampling, see stochastic decomposition
facet, 213, 216 scenario, see scenario aggregation
fathom, 210–212 scenario aggregation, 134–141
partition, 212 approximate solution, 141
relaxed linear program, 213 event tree, 134, 135, 145
waiting node, 210, 211 scenario, 134
linear, 7 scenario solution, 141
mathematical, 7 scenario analysis, 2
nonconvex, 8, 209 Schur complement, 232
nonlinear, 8 second-stage
progressive hedging, see scenario aggre- activity, 31
gation program, 32
project scheduling, see PERT separability, see dynamic programming
simplex
QDECOM, 103, 175, 233 criterion, 65
quasi-concave method, 64–69
function, 49 eta-vector, 232
measure, 48, 51, 103 slack node, see networks
slack variables, 54
random Slater condition, 86
variable, 25 stage, see dynamic programming
vector, 25 state, see dynamic programming
recourse stochastic decomposition, 217–223, 229,
activity, 31 232, 234
costs, 15, 31 cut, 220–222
expected, 15 estimate of lower bound, 220
function, 31, 160 incumbent, 221, 223
expected, 160 relatively complete recourse, 217, 221
matrix, 31 sampling, 218, 220
complete, 45, 160 stopping 223
fixed, 160 stopping criterion, 217, 220
relatively complete, 161 stochastic method, 217
simple, 34 stochastic program, 13
program, 31 approximation schemes, 103
variable, 15 general formulation, 21–36
vector, 31 linear, 13, 33, 36, 159–161
reduced gradient, 95 nonlinear, 32
method, see methods (nlp) value of, 151–156
regularity condition, 85, 86 stochastic program with
regularized decomposition, 173–177, 233 chance constraints, see probabilistic
master program, 174 constraints
method, 176 complete recourse, 34, 160
reliability, 18, 47 fixed recourse, 34, 160
removing columns, see preprocessing integer first stage, 209–217, 233
removing rows, see preprocessing algorithms, 214
Index 317
feasibility cut, 216 wait-and-see solution, 13, 103, 178, 190

initialization, 216
optimality cut, 217
stopping criterion, 217
probabilistic constraints
applications, 103
closedness, 52
convexity, 49
joint, 20, 35, 36, 46
methods, 103
models, 103
properties, 46–53
separate, 36
single, 36
recourse, 16, 32, 159–236
approximations, 190–205
bounds, 177–190, 212
convexity, 36
differentiability, 38
methods, 103
multistage, 32, 33, 103, 145
nonanticipativity, 103
properties, 36–46
smoothness, 37
relatively complete recourse, 46, 161,
217
simple recourse, 34, 205–209, 234
stochastic programming
models, 9, 21–36
stochastic quasi-gradient, 228, 233
methods, 225–229
stochastic solution, 4, 5
stochastic subgradient, 229
subdifferential, 226
subgradient, 226
methods, 228
support of probability measure, 43
supporting hyperplane
of convex function, 81
of convex set, 91
time horizon, see dynamic program-

ming
transition function, see dynamic pro-
gramming
trickling down, 230, 232, 234
unbounded solution, 168

utility function, 127
vertex, see convex

Stochastic Programming

Uploaded by

Stochastic Programming

Uploaded by

Stochastic Programming

Reference to this text is “Peter Kall and Stein W. Wallace, Stochastic

2 Dynamic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

3 Recourse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 159

3.6 Simple Recourse . . . . . . . . . . . . . . . . . . . . . . . . . . 205

4 Probabilistic Constraints . . . . . . . . . . . . . . . . . . . . . . 243

6 Network Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 277

6.6 Project Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . 301

Zurich and Trondheim, February 1994 P. K. and S.W.W.

By reading this introduction, you are certainly already familiar with

1.1.1 A numerical example

In each of the plants, it is possible to produce one unit of some product. It

1.1.2 Scenario analysis

solutions. These are given as follows. Decision numbers refer to Figure 1.

Interval for p Decision number

So whatever scenarios are constructed or sampled, these are the only

1.1.3 Using the expected value of p

−700 + 730 = 30.

1.1.4 Maximizing the expected value of the objective

Decision Investment Income if Income if Expected

will be lost. In scenario analysis, as outlined earlier, options have no value,

1.1.5 The IQ of hindsight

Many practical decision problems—in particular, rather complex ones—can

Using matrix–vector notation, the shorthand formulation of problem (2.1)

The form presented in (2.3) is known as a mathematical programming problem.

(b) nonlinear, if at least one of the functions gi , i = 0, · · · , m, is nonlinear or

Example 1.1 Consider the optimization problem

min ϕ(x), (2.4)

where ϕ(x) := 14 x4 − 5x3 + 27x2 − 40x. A necessary condition for solving

As usual, a function ψ is said to attain a relative minimum—also called a

Proof If x̄ is a local minimum of problem (2.3) then x̄ belongs to the

1.3 An Illustrative Example

Table 2 Productivities π(raw i, prod j).

Due to the simplicity of the example problem, we can give a graphical

Figure 2 Deterministic LP: set of feasible production plans.

• our model describes the weekly production process of a reﬁnery relying

Assume that, owing to statistics, we know that

x̂ = (x̂raw1 , x̂raw2 ) = (36, 18), γ = 126,

production plans such as

Figure 4 LP: feasible set varying with demands.

Figure 5 LP: feasible set varying with productivities.

may be wait-and-see solutions.

x∗ = (x∗raw1 , x∗raw2 ) = (48.018, 25.548), γ ∗ = 172.681. (3.8)

qprod1 = 7, qprod2 = 12. (3.9)

The costs due to shortage of production—or in general due to the amount

˜ := π(raw1, prod1) = 2 + η̃1 , β(ξ)

dual decomposition structure:

Depending on the number of realizations of ξ̃, r, this linear program may

Obviously, these discrete distributions with 15 realizations each can

and an empirical reliability of

which is clearly overestimated (compared with its theoretical value of 0.25),

x̄ = (37.566, 22.141), γ(x̄) = 144.179, γI (x̄) = 141.556,

x̌ = (37.754, 23.629), γ(x̌) = 150.446, γI (x̌) = 146.396,

whereas the LP solution x̂ = (36, 18) would yield

γ(x̂) = 232.492, ρ(x̂) = 0.2499.

1.4 Stochastic Programs: General Formulation

where ξ˜ is a random vector varying over a set Ξ ⊂ IRk . More precisely, we

1.4.1 Measures and Integrals

I[a,b) = {x ∈ IRk | ai ≤ xi < bi , i = 1, · · · , k}

as (half-open) intervals. In geometric terms, depending on the dimension k of

Analogously, in general for I[a,b) ⊂ IRk with arbitrary k, we have

Acirc = {(x, y) | x2 + y 2 ≤ 16, y ≥ 0},

Figure 8 Measure of a half-circle: ﬁrst approximation.

Figure 9 Improved approximate measure of a half-circle.

DA the set of all coverings of A. Then we make the following deﬁnition.

The closed bounded set A is measurable if

with the measure µ(A) = sup{µ(C) | C ∈ CA }.

To get rid of the boundedness restriction, we may extend this deﬁnition

An arbitrary closed set A ⊂ IRk is measurable iﬀ2 for every interval

if A ∈ A then also IRk − A ∈ A; (4.3 i)

µ(A) ≥ 0 ∀A ∈ A and µ(∅) = 0; (4.4 i)

In other words, the measure of a countable disjoint union of measurable sets

ξ˜ : Ω −→ IRk such that, for all A ∈ A, ξ̃ −1 [A] := {ω | ξ̃(ω) ∈ A} ∈ F. (4.5)

Pξ̃ (A) = P ({ω | ξ̃(ω) ∈ A}) ∀A ∈ A.

If it happens that Ω ⊂ IRk and F ⊂ A (i.e. every event is a “naturally”

Figure 10 Classiﬁcation of apples by quality.